1 /* 2 * Copyright (c) 1998, 2017, 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 "ci/ciMethodData.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "libadt/vectset.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/connode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/idealGraphPrinter.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/mulnode.hpp" 37 #include "opto/rootnode.hpp" 38 #include "opto/superword.hpp" 39 40 //============================================================================= 41 //------------------------------is_loop_iv------------------------------------- 42 // Determine if a node is Counted loop induction variable. 43 // The method is declared in node.hpp. 44 const Node* Node::is_loop_iv() const { 45 if (this->is_Phi() && !this->as_Phi()->is_copy() && 46 this->as_Phi()->region()->is_CountedLoop() && 47 this->as_Phi()->region()->as_CountedLoop()->phi() == this) { 48 return this; 49 } else { 50 return NULL; 51 } 52 } 53 54 //============================================================================= 55 //------------------------------dump_spec-------------------------------------- 56 // Dump special per-node info 57 #ifndef PRODUCT 58 void LoopNode::dump_spec(outputStream *st) const { 59 if (is_inner_loop()) st->print( "inner " ); 60 if (is_partial_peel_loop()) st->print( "partial_peel " ); 61 if (partial_peel_has_failed()) st->print( "partial_peel_failed " ); 62 } 63 #endif 64 65 //------------------------------is_valid_counted_loop------------------------- 66 bool LoopNode::is_valid_counted_loop() const { 67 if (is_CountedLoop()) { 68 CountedLoopNode* l = as_CountedLoop(); 69 CountedLoopEndNode* le = l->loopexit(); 70 if (le != NULL && 71 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) { 72 Node* phi = l->phi(); 73 Node* exit = le->proj_out(0 /* false */); 74 if (exit != NULL && exit->Opcode() == Op_IfFalse && 75 phi != NULL && phi->is_Phi() && 76 phi->in(LoopNode::LoopBackControl) == l->incr() && 77 le->loopnode() == l && le->stride_is_con()) { 78 return true; 79 } 80 } 81 } 82 return false; 83 } 84 85 //------------------------------get_early_ctrl--------------------------------- 86 // Compute earliest legal control 87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { 88 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); 89 uint i; 90 Node *early; 91 if (n->in(0) && !n->is_expensive()) { 92 early = n->in(0); 93 if (!early->is_CFG()) // Might be a non-CFG multi-def 94 early = get_ctrl(early); // So treat input as a straight data input 95 i = 1; 96 } else { 97 early = get_ctrl(n->in(1)); 98 i = 2; 99 } 100 uint e_d = dom_depth(early); 101 assert( early, "" ); 102 for (; i < n->req(); i++) { 103 Node *cin = get_ctrl(n->in(i)); 104 assert( cin, "" ); 105 // Keep deepest dominator depth 106 uint c_d = dom_depth(cin); 107 if (c_d > e_d) { // Deeper guy? 108 early = cin; // Keep deepest found so far 109 e_d = c_d; 110 } else if (c_d == e_d && // Same depth? 111 early != cin) { // If not equal, must use slower algorithm 112 // If same depth but not equal, one _must_ dominate the other 113 // and we want the deeper (i.e., dominated) guy. 114 Node *n1 = early; 115 Node *n2 = cin; 116 while (1) { 117 n1 = idom(n1); // Walk up until break cycle 118 n2 = idom(n2); 119 if (n1 == cin || // Walked early up to cin 120 dom_depth(n2) < c_d) 121 break; // early is deeper; keep him 122 if (n2 == early || // Walked cin up to early 123 dom_depth(n1) < c_d) { 124 early = cin; // cin is deeper; keep him 125 break; 126 } 127 } 128 e_d = dom_depth(early); // Reset depth register cache 129 } 130 } 131 132 // Return earliest legal location 133 assert(early == find_non_split_ctrl(early), "unexpected early control"); 134 135 if (n->is_expensive()) { 136 assert(n->in(0), "should have control input"); 137 early = get_early_ctrl_for_expensive(n, early); 138 } 139 140 return early; 141 } 142 143 //------------------------------get_early_ctrl_for_expensive--------------------------------- 144 // Move node up the dominator tree as high as legal while still beneficial 145 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) { 146 assert(n->in(0) && n->is_expensive(), "expensive node with control input here"); 147 assert(OptimizeExpensiveOps, "optimization off?"); 148 149 Node* ctl = n->in(0); 150 assert(ctl->is_CFG(), "expensive input 0 must be cfg"); 151 uint min_dom_depth = dom_depth(earliest); 152 #ifdef ASSERT 153 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) { 154 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl); 155 assert(false, "Bad graph detected in get_early_ctrl_for_expensive"); 156 } 157 #endif 158 if (dom_depth(ctl) < min_dom_depth) { 159 return earliest; 160 } 161 162 while (1) { 163 Node *next = ctl; 164 // Moving the node out of a loop on the projection of a If 165 // confuses loop predication. So once we hit a Loop in a If branch 166 // that doesn't branch to an UNC, we stop. The code that process 167 // expensive nodes will notice the loop and skip over it to try to 168 // move the node further up. 169 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) { 170 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 171 break; 172 } 173 next = idom(ctl->in(1)->in(0)); 174 } else if (ctl->is_Proj()) { 175 // We only move it up along a projection if the projection is 176 // the single control projection for its parent: same code path, 177 // if it's a If with UNC or fallthrough of a call. 178 Node* parent_ctl = ctl->in(0); 179 if (parent_ctl == NULL) { 180 break; 181 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) { 182 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control(); 183 } else if (parent_ctl->is_If()) { 184 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 185 break; 186 } 187 assert(idom(ctl) == parent_ctl, "strange"); 188 next = idom(parent_ctl); 189 } else if (ctl->is_CatchProj()) { 190 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) { 191 break; 192 } 193 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph"); 194 next = parent_ctl->in(0)->in(0)->in(0); 195 } else { 196 // Check if parent control has a single projection (this 197 // control is the only possible successor of the parent 198 // control). If so, we can try to move the node above the 199 // parent control. 200 int nb_ctl_proj = 0; 201 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) { 202 Node *p = parent_ctl->fast_out(i); 203 if (p->is_Proj() && p->is_CFG()) { 204 nb_ctl_proj++; 205 if (nb_ctl_proj > 1) { 206 break; 207 } 208 } 209 } 210 211 if (nb_ctl_proj > 1) { 212 break; 213 } 214 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node"); 215 assert(idom(ctl) == parent_ctl, "strange"); 216 next = idom(parent_ctl); 217 } 218 } else { 219 next = idom(ctl); 220 } 221 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) { 222 break; 223 } 224 ctl = next; 225 } 226 227 if (ctl != n->in(0)) { 228 _igvn.hash_delete(n); 229 n->set_req(0, ctl); 230 _igvn.hash_insert(n); 231 } 232 233 return ctl; 234 } 235 236 237 //------------------------------set_early_ctrl--------------------------------- 238 // Set earliest legal control 239 void PhaseIdealLoop::set_early_ctrl( Node *n ) { 240 Node *early = get_early_ctrl(n); 241 242 // Record earliest legal location 243 set_ctrl(n, early); 244 } 245 246 //------------------------------set_subtree_ctrl------------------------------- 247 // set missing _ctrl entries on new nodes 248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { 249 // Already set? Get out. 250 if( _nodes[n->_idx] ) return; 251 // Recursively set _nodes array to indicate where the Node goes 252 uint i; 253 for( i = 0; i < n->req(); ++i ) { 254 Node *m = n->in(i); 255 if( m && m != C->root() ) 256 set_subtree_ctrl( m ); 257 } 258 259 // Fixup self 260 set_early_ctrl( n ); 261 } 262 263 //------------------------------is_counted_loop-------------------------------- 264 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) { 265 PhaseGVN *gvn = &_igvn; 266 267 // Counted loop head must be a good RegionNode with only 3 not NULL 268 // control input edges: Self, Entry, LoopBack. 269 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) { 270 return false; 271 } 272 Node *init_control = x->in(LoopNode::EntryControl); 273 Node *back_control = x->in(LoopNode::LoopBackControl); 274 if (init_control == NULL || back_control == NULL) // Partially dead 275 return false; 276 // Must also check for TOP when looking for a dead loop 277 if (init_control->is_top() || back_control->is_top()) 278 return false; 279 280 // Allow funny placement of Safepoint 281 if (back_control->Opcode() == Op_SafePoint) { 282 if (UseCountedLoopSafepoints) { 283 // Leaving the safepoint on the backedge and creating a 284 // CountedLoop will confuse optimizations. We can't move the 285 // safepoint around because its jvm state wouldn't match a new 286 // location. Give up on that loop. 287 return false; 288 } 289 back_control = back_control->in(TypeFunc::Control); 290 } 291 292 // Controlling test for loop 293 Node *iftrue = back_control; 294 uint iftrue_op = iftrue->Opcode(); 295 if (iftrue_op != Op_IfTrue && 296 iftrue_op != Op_IfFalse) 297 // I have a weird back-control. Probably the loop-exit test is in 298 // the middle of the loop and I am looking at some trailing control-flow 299 // merge point. To fix this I would have to partially peel the loop. 300 return false; // Obscure back-control 301 302 // Get boolean guarding loop-back test 303 Node *iff = iftrue->in(0); 304 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) 305 return false; 306 BoolNode *test = iff->in(1)->as_Bool(); 307 BoolTest::mask bt = test->_test._test; 308 float cl_prob = iff->as_If()->_prob; 309 if (iftrue_op == Op_IfFalse) { 310 bt = BoolTest(bt).negate(); 311 cl_prob = 1.0 - cl_prob; 312 } 313 // Get backedge compare 314 Node *cmp = test->in(1); 315 int cmp_op = cmp->Opcode(); 316 if (cmp_op != Op_CmpI) 317 return false; // Avoid pointer & float compares 318 319 // Find the trip-counter increment & limit. Limit must be loop invariant. 320 Node *incr = cmp->in(1); 321 Node *limit = cmp->in(2); 322 323 // --------- 324 // need 'loop()' test to tell if limit is loop invariant 325 // --------- 326 327 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit? 328 Node *tmp = incr; // Then reverse order into the CmpI 329 incr = limit; 330 limit = tmp; 331 bt = BoolTest(bt).commute(); // And commute the exit test 332 } 333 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant 334 return false; 335 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 336 return false; 337 338 Node* phi_incr = NULL; 339 // Trip-counter increment must be commutative & associative. 340 if (incr->is_Phi()) { 341 if (incr->as_Phi()->region() != x || incr->req() != 3) 342 return false; // Not simple trip counter expression 343 phi_incr = incr; 344 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi 345 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 346 return false; 347 } 348 349 Node* trunc1 = NULL; 350 Node* trunc2 = NULL; 351 const TypeInt* iv_trunc_t = NULL; 352 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { 353 return false; // Funny increment opcode 354 } 355 assert(incr->Opcode() == Op_AddI, "wrong increment code"); 356 357 // Get merge point 358 Node *xphi = incr->in(1); 359 Node *stride = incr->in(2); 360 if (!stride->is_Con()) { // Oops, swap these 361 if (!xphi->is_Con()) // Is the other guy a constant? 362 return false; // Nope, unknown stride, bail out 363 Node *tmp = xphi; // 'incr' is commutative, so ok to swap 364 xphi = stride; 365 stride = tmp; 366 } 367 // Stride must be constant 368 int stride_con = stride->get_int(); 369 if (stride_con == 0) 370 return false; // missed some peephole opt 371 372 if (!xphi->is_Phi()) 373 return false; // Too much math on the trip counter 374 if (phi_incr != NULL && phi_incr != xphi) 375 return false; 376 PhiNode *phi = xphi->as_Phi(); 377 378 // Phi must be of loop header; backedge must wrap to increment 379 if (phi->region() != x) 380 return false; 381 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr || 382 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) { 383 return false; 384 } 385 Node *init_trip = phi->in(LoopNode::EntryControl); 386 387 // If iv trunc type is smaller than int, check for possible wrap. 388 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { 389 assert(trunc1 != NULL, "must have found some truncation"); 390 391 // Get a better type for the phi (filtered thru if's) 392 const TypeInt* phi_ft = filtered_type(phi); 393 394 // Can iv take on a value that will wrap? 395 // 396 // Ensure iv's limit is not within "stride" of the wrap value. 397 // 398 // Example for "short" type 399 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) 400 // If the stride is +10, then the last value of the induction 401 // variable before the increment (phi_ft->_hi) must be 402 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to 403 // ensure no truncation occurs after the increment. 404 405 if (stride_con > 0) { 406 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || 407 iv_trunc_t->_lo > phi_ft->_lo) { 408 return false; // truncation may occur 409 } 410 } else if (stride_con < 0) { 411 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || 412 iv_trunc_t->_hi < phi_ft->_hi) { 413 return false; // truncation may occur 414 } 415 } 416 // No possibility of wrap so truncation can be discarded 417 // Promote iv type to Int 418 } else { 419 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); 420 } 421 422 // If the condition is inverted and we will be rolling 423 // through MININT to MAXINT, then bail out. 424 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice! 425 // Odd stride 426 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 || 427 // Count down loop rolls through MAXINT 428 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 || 429 // Count up loop rolls through MININT 430 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) { 431 return false; // Bail out 432 } 433 434 const TypeInt* init_t = gvn->type(init_trip)->is_int(); 435 const TypeInt* limit_t = gvn->type(limit)->is_int(); 436 437 if (stride_con > 0) { 438 jlong init_p = (jlong)init_t->_lo + stride_con; 439 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi) 440 return false; // cyclic loop or this loop trips only once 441 } else { 442 jlong init_p = (jlong)init_t->_hi + stride_con; 443 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo) 444 return false; // cyclic loop or this loop trips only once 445 } 446 447 if (phi_incr != NULL) { 448 // check if there is a possiblity of IV overflowing after the first increment 449 if (stride_con > 0) { 450 if (init_t->_hi > max_jint - stride_con) { 451 return false; 452 } 453 } else { 454 if (init_t->_lo < min_jint - stride_con) { 455 return false; 456 } 457 } 458 } 459 460 // ================================================= 461 // ---- SUCCESS! Found A Trip-Counted Loop! ----- 462 // 463 assert(x->Opcode() == Op_Loop, "regular loops only"); 464 C->print_method(PHASE_BEFORE_CLOOPS, 3); 465 466 Node *hook = new (C) Node(6); 467 468 if (LoopLimitCheck) { 469 470 // =================================================== 471 // Generate loop limit check to avoid integer overflow 472 // in cases like next (cyclic loops): 473 // 474 // for (i=0; i <= max_jint; i++) {} 475 // for (i=0; i < max_jint; i+=2) {} 476 // 477 // 478 // Limit check predicate depends on the loop test: 479 // 480 // for(;i != limit; i++) --> limit <= (max_jint) 481 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1) 482 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride ) 483 // 484 485 // Check if limit is excluded to do more precise int overflow check. 486 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge); 487 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1)); 488 489 // If compare points directly to the phi we need to adjust 490 // the compare so that it points to the incr. Limit have 491 // to be adjusted to keep trip count the same and the 492 // adjusted limit should be checked for int overflow. 493 if (phi_incr != NULL) { 494 stride_m += stride_con; 495 } 496 497 if (limit->is_Con()) { 498 int limit_con = limit->get_int(); 499 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) || 500 (stride_con < 0 && limit_con < (min_jint - stride_m))) { 501 // Bailout: it could be integer overflow. 502 return false; 503 } 504 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) || 505 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) { 506 // Limit's type may satisfy the condition, for example, 507 // when it is an array length. 508 } else { 509 // Generate loop's limit check. 510 // Loop limit check predicate should be near the loop. 511 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); 512 if (!limit_check_proj) { 513 // The limit check predicate is not generated if this method trapped here before. 514 #ifdef ASSERT 515 if (TraceLoopLimitCheck) { 516 tty->print("missing loop limit check:"); 517 loop->dump_head(); 518 x->dump(1); 519 } 520 #endif 521 return false; 522 } 523 524 IfNode* check_iff = limit_check_proj->in(0)->as_If(); 525 Node* cmp_limit; 526 Node* bol; 527 528 if (stride_con > 0) { 529 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(max_jint - stride_m)); 530 bol = new (C) BoolNode(cmp_limit, BoolTest::le); 531 } else { 532 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(min_jint - stride_m)); 533 bol = new (C) BoolNode(cmp_limit, BoolTest::ge); 534 } 535 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit); 536 bol = _igvn.register_new_node_with_optimizer(bol); 537 set_subtree_ctrl(bol); 538 539 // Replace condition in original predicate but preserve Opaque node 540 // so that previous predicates could be found. 541 assert(check_iff->in(1)->Opcode() == Op_Conv2B && 542 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, ""); 543 Node* opq = check_iff->in(1)->in(1); 544 _igvn.hash_delete(opq); 545 opq->set_req(1, bol); 546 // Update ctrl. 547 set_ctrl(opq, check_iff->in(0)); 548 set_ctrl(check_iff->in(1), check_iff->in(0)); 549 550 #ifndef PRODUCT 551 // report that the loop predication has been actually performed 552 // for this loop 553 if (TraceLoopLimitCheck) { 554 tty->print_cr("Counted Loop Limit Check generated:"); 555 debug_only( bol->dump(2); ) 556 } 557 #endif 558 } 559 560 if (phi_incr != NULL) { 561 // If compare points directly to the phi we need to adjust 562 // the compare so that it points to the incr. Limit have 563 // to be adjusted to keep trip count the same and we 564 // should avoid int overflow. 565 // 566 // i = init; do {} while(i++ < limit); 567 // is converted to 568 // i = init; do {} while(++i < limit+1); 569 // 570 limit = gvn->transform(new (C) AddINode(limit, stride)); 571 } 572 573 // Now we need to canonicalize loop condition. 574 if (bt == BoolTest::ne) { 575 assert(stride_con == 1 || stride_con == -1, "simple increment only"); 576 // 'ne' can be replaced with 'lt' only when init < limit. 577 if (stride_con > 0 && init_t->_hi < limit_t->_lo) 578 bt = BoolTest::lt; 579 // 'ne' can be replaced with 'gt' only when init > limit. 580 if (stride_con < 0 && init_t->_lo > limit_t->_hi) 581 bt = BoolTest::gt; 582 } 583 584 if (incl_limit) { 585 // The limit check guaranties that 'limit <= (max_jint - stride)' so 586 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0. 587 // 588 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1); 589 limit = gvn->transform(new (C) AddINode(limit, one)); 590 if (bt == BoolTest::le) 591 bt = BoolTest::lt; 592 else if (bt == BoolTest::ge) 593 bt = BoolTest::gt; 594 else 595 ShouldNotReachHere(); 596 } 597 set_subtree_ctrl( limit ); 598 599 } else { // LoopLimitCheck 600 601 // If compare points to incr, we are ok. Otherwise the compare 602 // can directly point to the phi; in this case adjust the compare so that 603 // it points to the incr by adjusting the limit. 604 if (cmp->in(1) == phi || cmp->in(2) == phi) 605 limit = gvn->transform(new (C) AddINode(limit,stride)); 606 607 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride. 608 // Final value for iterator should be: trip_count * stride + init_trip. 609 Node *one_p = gvn->intcon( 1); 610 Node *one_m = gvn->intcon(-1); 611 612 Node *trip_count = NULL; 613 switch( bt ) { 614 case BoolTest::eq: 615 ShouldNotReachHere(); 616 case BoolTest::ne: // Ahh, the case we desire 617 if (stride_con == 1) 618 trip_count = gvn->transform(new (C) SubINode(limit,init_trip)); 619 else if (stride_con == -1) 620 trip_count = gvn->transform(new (C) SubINode(init_trip,limit)); 621 else 622 ShouldNotReachHere(); 623 set_subtree_ctrl(trip_count); 624 //_loop.map(trip_count->_idx,loop(limit)); 625 break; 626 case BoolTest::le: // Maybe convert to '<' case 627 limit = gvn->transform(new (C) AddINode(limit,one_p)); 628 set_subtree_ctrl( limit ); 629 hook->init_req(4, limit); 630 631 bt = BoolTest::lt; 632 // Make the new limit be in the same loop nest as the old limit 633 //_loop.map(limit->_idx,limit_loop); 634 // Fall into next case 635 case BoolTest::lt: { // Maybe convert to '!=' case 636 if (stride_con < 0) // Count down loop rolls through MAXINT 637 ShouldNotReachHere(); 638 Node *range = gvn->transform(new (C) SubINode(limit,init_trip)); 639 set_subtree_ctrl( range ); 640 hook->init_req(0, range); 641 642 Node *bias = gvn->transform(new (C) AddINode(range,stride)); 643 set_subtree_ctrl( bias ); 644 hook->init_req(1, bias); 645 646 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_m)); 647 set_subtree_ctrl( bias1 ); 648 hook->init_req(2, bias1); 649 650 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride)); 651 set_subtree_ctrl( trip_count ); 652 hook->init_req(3, trip_count); 653 break; 654 } 655 656 case BoolTest::ge: // Maybe convert to '>' case 657 limit = gvn->transform(new (C) AddINode(limit,one_m)); 658 set_subtree_ctrl( limit ); 659 hook->init_req(4 ,limit); 660 661 bt = BoolTest::gt; 662 // Make the new limit be in the same loop nest as the old limit 663 //_loop.map(limit->_idx,limit_loop); 664 // Fall into next case 665 case BoolTest::gt: { // Maybe convert to '!=' case 666 if (stride_con > 0) // count up loop rolls through MININT 667 ShouldNotReachHere(); 668 Node *range = gvn->transform(new (C) SubINode(limit,init_trip)); 669 set_subtree_ctrl( range ); 670 hook->init_req(0, range); 671 672 Node *bias = gvn->transform(new (C) AddINode(range,stride)); 673 set_subtree_ctrl( bias ); 674 hook->init_req(1, bias); 675 676 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_p)); 677 set_subtree_ctrl( bias1 ); 678 hook->init_req(2, bias1); 679 680 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride)); 681 set_subtree_ctrl( trip_count ); 682 hook->init_req(3, trip_count); 683 break; 684 } 685 } // switch( bt ) 686 687 Node *span = gvn->transform(new (C) MulINode(trip_count,stride)); 688 set_subtree_ctrl( span ); 689 hook->init_req(5, span); 690 691 limit = gvn->transform(new (C) AddINode(span,init_trip)); 692 set_subtree_ctrl( limit ); 693 694 } // LoopLimitCheck 695 696 if (!UseCountedLoopSafepoints) { 697 // Check for SafePoint on backedge and remove 698 Node *sfpt = x->in(LoopNode::LoopBackControl); 699 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { 700 lazy_replace( sfpt, iftrue ); 701 if (loop->_safepts != NULL) { 702 loop->_safepts->yank(sfpt); 703 } 704 loop->_tail = iftrue; 705 } 706 } 707 708 // Build a canonical trip test. 709 // Clone code, as old values may be in use. 710 incr = incr->clone(); 711 incr->set_req(1,phi); 712 incr->set_req(2,stride); 713 incr = _igvn.register_new_node_with_optimizer(incr); 714 set_early_ctrl( incr ); 715 _igvn.hash_delete(phi); 716 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); 717 718 // If phi type is more restrictive than Int, raise to 719 // Int to prevent (almost) infinite recursion in igvn 720 // which can only handle integer types for constants or minint..maxint. 721 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { 722 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); 723 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); 724 nphi = _igvn.register_new_node_with_optimizer(nphi); 725 set_ctrl(nphi, get_ctrl(phi)); 726 _igvn.replace_node(phi, nphi); 727 phi = nphi->as_Phi(); 728 } 729 cmp = cmp->clone(); 730 cmp->set_req(1,incr); 731 cmp->set_req(2,limit); 732 cmp = _igvn.register_new_node_with_optimizer(cmp); 733 set_ctrl(cmp, iff->in(0)); 734 735 test = test->clone()->as_Bool(); 736 (*(BoolTest*)&test->_test)._test = bt; 737 test->set_req(1,cmp); 738 _igvn.register_new_node_with_optimizer(test); 739 set_ctrl(test, iff->in(0)); 740 741 // Replace the old IfNode with a new LoopEndNode 742 Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt )); 743 IfNode *le = lex->as_If(); 744 uint dd = dom_depth(iff); 745 set_idom(le, le->in(0), dd); // Update dominance for loop exit 746 set_loop(le, loop); 747 748 // Get the loop-exit control 749 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); 750 751 // Need to swap loop-exit and loop-back control? 752 if (iftrue_op == Op_IfFalse) { 753 Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le)); 754 Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le)); 755 756 loop->_tail = back_control = ift2; 757 set_loop(ift2, loop); 758 set_loop(iff2, get_loop(iffalse)); 759 760 // Lazy update of 'get_ctrl' mechanism. 761 lazy_replace(iffalse, iff2); 762 lazy_replace(iftrue, ift2); 763 764 // Swap names 765 iffalse = iff2; 766 iftrue = ift2; 767 } else { 768 _igvn.hash_delete(iffalse); 769 _igvn.hash_delete(iftrue); 770 iffalse->set_req_X( 0, le, &_igvn ); 771 iftrue ->set_req_X( 0, le, &_igvn ); 772 } 773 774 set_idom(iftrue, le, dd+1); 775 set_idom(iffalse, le, dd+1); 776 assert(iff->outcnt() == 0, "should be dead now"); 777 lazy_replace( iff, le ); // fix 'get_ctrl' 778 779 // Now setup a new CountedLoopNode to replace the existing LoopNode 780 CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control); 781 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve 782 // The following assert is approximately true, and defines the intention 783 // of can_be_counted_loop. It fails, however, because phase->type 784 // is not yet initialized for this loop and its parts. 785 //assert(l->can_be_counted_loop(this), "sanity"); 786 _igvn.register_new_node_with_optimizer(l); 787 set_loop(l, loop); 788 loop->_head = l; 789 // Fix all data nodes placed at the old loop head. 790 // Uses the lazy-update mechanism of 'get_ctrl'. 791 lazy_replace( x, l ); 792 set_idom(l, init_control, dom_depth(x)); 793 794 if (!UseCountedLoopSafepoints) { 795 // Check for immediately preceding SafePoint and remove 796 Node *sfpt2 = le->in(0); 797 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) { 798 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); 799 if (loop->_safepts != NULL) { 800 loop->_safepts->yank(sfpt2); 801 } 802 } 803 } 804 805 // Free up intermediate goo 806 _igvn.remove_dead_node(hook); 807 808 #ifdef ASSERT 809 assert(l->is_valid_counted_loop(), "counted loop shape is messed up"); 810 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" ); 811 #endif 812 #ifndef PRODUCT 813 if (TraceLoopOpts) { 814 tty->print("Counted "); 815 loop->dump_head(); 816 } 817 #endif 818 819 C->print_method(PHASE_AFTER_CLOOPS, 3); 820 821 return true; 822 } 823 824 //----------------------exact_limit------------------------------------------- 825 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) { 826 assert(loop->_head->is_CountedLoop(), ""); 827 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 828 assert(cl->is_valid_counted_loop(), ""); 829 830 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 || 831 cl->limit()->Opcode() == Op_LoopLimit) { 832 // Old code has exact limit (it could be incorrect in case of int overflow). 833 // Loop limit is exact with stride == 1. And loop may already have exact limit. 834 return cl->limit(); 835 } 836 Node *limit = NULL; 837 #ifdef ASSERT 838 BoolTest::mask bt = cl->loopexit()->test_trip(); 839 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 840 #endif 841 if (cl->has_exact_trip_count()) { 842 // Simple case: loop has constant boundaries. 843 // Use jlongs to avoid integer overflow. 844 int stride_con = cl->stride_con(); 845 jlong init_con = cl->init_trip()->get_int(); 846 jlong limit_con = cl->limit()->get_int(); 847 julong trip_cnt = cl->trip_count(); 848 jlong final_con = init_con + trip_cnt*stride_con; 849 int final_int = (int)final_con; 850 // The final value should be in integer range since the loop 851 // is counted and the limit was checked for overflow. 852 assert(final_con == (jlong)final_int, "final value should be integer"); 853 limit = _igvn.intcon(final_int); 854 } else { 855 // Create new LoopLimit node to get exact limit (final iv value). 856 limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride()); 857 register_new_node(limit, cl->in(LoopNode::EntryControl)); 858 } 859 assert(limit != NULL, "sanity"); 860 return limit; 861 } 862 863 //------------------------------Ideal------------------------------------------ 864 // Return a node which is more "ideal" than the current node. 865 // Attempt to convert into a counted-loop. 866 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 867 if (!can_be_counted_loop(phase)) { 868 phase->C->set_major_progress(); 869 } 870 return RegionNode::Ideal(phase, can_reshape); 871 } 872 873 874 //============================================================================= 875 //------------------------------Ideal------------------------------------------ 876 // Return a node which is more "ideal" than the current node. 877 // Attempt to convert into a counted-loop. 878 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 879 return RegionNode::Ideal(phase, can_reshape); 880 } 881 882 //------------------------------dump_spec-------------------------------------- 883 // Dump special per-node info 884 #ifndef PRODUCT 885 void CountedLoopNode::dump_spec(outputStream *st) const { 886 LoopNode::dump_spec(st); 887 if (stride_is_con()) { 888 st->print("stride: %d ",stride_con()); 889 } 890 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx); 891 if (is_main_loop()) st->print("main of N%d", _idx); 892 if (is_post_loop()) st->print("post of N%d", _main_idx); 893 } 894 #endif 895 896 //============================================================================= 897 int CountedLoopEndNode::stride_con() const { 898 return stride()->bottom_type()->is_int()->get_con(); 899 } 900 901 //============================================================================= 902 //------------------------------Value----------------------------------------- 903 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const { 904 const Type* init_t = phase->type(in(Init)); 905 const Type* limit_t = phase->type(in(Limit)); 906 const Type* stride_t = phase->type(in(Stride)); 907 // Either input is TOP ==> the result is TOP 908 if (init_t == Type::TOP) return Type::TOP; 909 if (limit_t == Type::TOP) return Type::TOP; 910 if (stride_t == Type::TOP) return Type::TOP; 911 912 int stride_con = stride_t->is_int()->get_con(); 913 if (stride_con == 1) 914 return NULL; // Identity 915 916 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) { 917 // Use jlongs to avoid integer overflow. 918 jlong init_con = init_t->is_int()->get_con(); 919 jlong limit_con = limit_t->is_int()->get_con(); 920 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 921 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 922 jlong final_con = init_con + stride_con*trip_count; 923 int final_int = (int)final_con; 924 // The final value should be in integer range since the loop 925 // is counted and the limit was checked for overflow. 926 assert(final_con == (jlong)final_int, "final value should be integer"); 927 return TypeInt::make(final_int); 928 } 929 930 return bottom_type(); // TypeInt::INT 931 } 932 933 //------------------------------Ideal------------------------------------------ 934 // Return a node which is more "ideal" than the current node. 935 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) { 936 if (phase->type(in(Init)) == Type::TOP || 937 phase->type(in(Limit)) == Type::TOP || 938 phase->type(in(Stride)) == Type::TOP) 939 return NULL; // Dead 940 941 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 942 if (stride_con == 1) 943 return NULL; // Identity 944 945 if (in(Init)->is_Con() && in(Limit)->is_Con()) 946 return NULL; // Value 947 948 // Delay following optimizations until all loop optimizations 949 // done to keep Ideal graph simple. 950 if (!can_reshape || phase->C->major_progress()) 951 return NULL; 952 953 const TypeInt* init_t = phase->type(in(Init) )->is_int(); 954 const TypeInt* limit_t = phase->type(in(Limit))->is_int(); 955 int stride_p; 956 jlong lim, ini; 957 julong max; 958 if (stride_con > 0) { 959 stride_p = stride_con; 960 lim = limit_t->_hi; 961 ini = init_t->_lo; 962 max = (julong)max_jint; 963 } else { 964 stride_p = -stride_con; 965 lim = init_t->_hi; 966 ini = limit_t->_lo; 967 max = (julong)min_jint; 968 } 969 julong range = lim - ini + stride_p; 970 if (range <= max) { 971 // Convert to integer expression if it is not overflow. 972 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1)); 973 Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init))); 974 Node *bias = phase->transform(new (phase->C) AddINode(range, stride_m)); 975 Node *trip = phase->transform(new (phase->C) DivINode(0, bias, in(Stride))); 976 Node *span = phase->transform(new (phase->C) MulINode(trip, in(Stride))); 977 return new (phase->C) AddINode(span, in(Init)); // exact limit 978 } 979 980 if (is_power_of_2(stride_p) || // divisor is 2^n 981 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node? 982 // Convert to long expression to avoid integer overflow 983 // and let igvn optimizer convert this division. 984 // 985 Node* init = phase->transform( new (phase->C) ConvI2LNode(in(Init))); 986 Node* limit = phase->transform( new (phase->C) ConvI2LNode(in(Limit))); 987 Node* stride = phase->longcon(stride_con); 988 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1)); 989 990 Node *range = phase->transform(new (phase->C) SubLNode(limit, init)); 991 Node *bias = phase->transform(new (phase->C) AddLNode(range, stride_m)); 992 Node *span; 993 if (stride_con > 0 && is_power_of_2(stride_p)) { 994 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride) 995 // and avoid generating rounding for division. Zero trip guard should 996 // guarantee that init < limit but sometimes the guard is missing and 997 // we can get situation when init > limit. Note, for the empty loop 998 // optimization zero trip guard is generated explicitly which leaves 999 // only RCE predicate where exact limit is used and the predicate 1000 // will simply fail forcing recompilation. 1001 Node* neg_stride = phase->longcon(-stride_con); 1002 span = phase->transform(new (phase->C) AndLNode(bias, neg_stride)); 1003 } else { 1004 Node *trip = phase->transform(new (phase->C) DivLNode(0, bias, stride)); 1005 span = phase->transform(new (phase->C) MulLNode(trip, stride)); 1006 } 1007 // Convert back to int 1008 Node *span_int = phase->transform(new (phase->C) ConvL2INode(span)); 1009 return new (phase->C) AddINode(span_int, in(Init)); // exact limit 1010 } 1011 1012 return NULL; // No progress 1013 } 1014 1015 //------------------------------Identity--------------------------------------- 1016 // If stride == 1 return limit node. 1017 Node *LoopLimitNode::Identity( PhaseTransform *phase ) { 1018 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1019 if (stride_con == 1 || stride_con == -1) 1020 return in(Limit); 1021 return this; 1022 } 1023 1024 //============================================================================= 1025 //----------------------match_incr_with_optional_truncation-------------------- 1026 // Match increment with optional truncation: 1027 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 1028 // Return NULL for failure. Success returns the increment node. 1029 Node* CountedLoopNode::match_incr_with_optional_truncation( 1030 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { 1031 // Quick cutouts: 1032 if (expr == NULL || expr->req() != 3) return NULL; 1033 1034 Node *t1 = NULL; 1035 Node *t2 = NULL; 1036 const TypeInt* trunc_t = TypeInt::INT; 1037 Node* n1 = expr; 1038 int n1op = n1->Opcode(); 1039 1040 // Try to strip (n1 & M) or (n1 << N >> N) from n1. 1041 if (n1op == Op_AndI && 1042 n1->in(2)->is_Con() && 1043 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { 1044 // %%% This check should match any mask of 2**K-1. 1045 t1 = n1; 1046 n1 = t1->in(1); 1047 n1op = n1->Opcode(); 1048 trunc_t = TypeInt::CHAR; 1049 } else if (n1op == Op_RShiftI && 1050 n1->in(1) != NULL && 1051 n1->in(1)->Opcode() == Op_LShiftI && 1052 n1->in(2) == n1->in(1)->in(2) && 1053 n1->in(2)->is_Con()) { 1054 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); 1055 // %%% This check should match any shift in [1..31]. 1056 if (shift == 16 || shift == 8) { 1057 t1 = n1; 1058 t2 = t1->in(1); 1059 n1 = t2->in(1); 1060 n1op = n1->Opcode(); 1061 if (shift == 16) { 1062 trunc_t = TypeInt::SHORT; 1063 } else if (shift == 8) { 1064 trunc_t = TypeInt::BYTE; 1065 } 1066 } 1067 } 1068 1069 // If (maybe after stripping) it is an AddI, we won: 1070 if (n1op == Op_AddI) { 1071 *trunc1 = t1; 1072 *trunc2 = t2; 1073 *trunc_type = trunc_t; 1074 return n1; 1075 } 1076 1077 // failed 1078 return NULL; 1079 } 1080 1081 1082 //------------------------------filtered_type-------------------------------- 1083 // Return a type based on condition control flow 1084 // A successful return will be a type that is restricted due 1085 // to a series of dominating if-tests, such as: 1086 // if (i < 10) { 1087 // if (i > 0) { 1088 // here: "i" type is [1..10) 1089 // } 1090 // } 1091 // or a control flow merge 1092 // if (i < 10) { 1093 // do { 1094 // phi( , ) -- at top of loop type is [min_int..10) 1095 // i = ? 1096 // } while ( i < 10) 1097 // 1098 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { 1099 assert(n && n->bottom_type()->is_int(), "must be int"); 1100 const TypeInt* filtered_t = NULL; 1101 if (!n->is_Phi()) { 1102 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); 1103 filtered_t = filtered_type_from_dominators(n, n_ctrl); 1104 1105 } else { 1106 Node* phi = n->as_Phi(); 1107 Node* region = phi->in(0); 1108 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); 1109 if (region && region != C->top()) { 1110 for (uint i = 1; i < phi->req(); i++) { 1111 Node* val = phi->in(i); 1112 Node* use_c = region->in(i); 1113 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); 1114 if (val_t != NULL) { 1115 if (filtered_t == NULL) { 1116 filtered_t = val_t; 1117 } else { 1118 filtered_t = filtered_t->meet(val_t)->is_int(); 1119 } 1120 } 1121 } 1122 } 1123 } 1124 const TypeInt* n_t = _igvn.type(n)->is_int(); 1125 if (filtered_t != NULL) { 1126 n_t = n_t->join(filtered_t)->is_int(); 1127 } 1128 return n_t; 1129 } 1130 1131 1132 //------------------------------filtered_type_from_dominators-------------------------------- 1133 // Return a possibly more restrictive type for val based on condition control flow of dominators 1134 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { 1135 if (val->is_Con()) { 1136 return val->bottom_type()->is_int(); 1137 } 1138 uint if_limit = 10; // Max number of dominating if's visited 1139 const TypeInt* rtn_t = NULL; 1140 1141 if (use_ctrl && use_ctrl != C->top()) { 1142 Node* val_ctrl = get_ctrl(val); 1143 uint val_dom_depth = dom_depth(val_ctrl); 1144 Node* pred = use_ctrl; 1145 uint if_cnt = 0; 1146 while (if_cnt < if_limit) { 1147 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { 1148 if_cnt++; 1149 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); 1150 if (if_t != NULL) { 1151 if (rtn_t == NULL) { 1152 rtn_t = if_t; 1153 } else { 1154 rtn_t = rtn_t->join(if_t)->is_int(); 1155 } 1156 } 1157 } 1158 pred = idom(pred); 1159 if (pred == NULL || pred == C->top()) { 1160 break; 1161 } 1162 // Stop if going beyond definition block of val 1163 if (dom_depth(pred) < val_dom_depth) { 1164 break; 1165 } 1166 } 1167 } 1168 return rtn_t; 1169 } 1170 1171 1172 //------------------------------dump_spec-------------------------------------- 1173 // Dump special per-node info 1174 #ifndef PRODUCT 1175 void CountedLoopEndNode::dump_spec(outputStream *st) const { 1176 if( in(TestValue)->is_Bool() ) { 1177 BoolTest bt( test_trip()); // Added this for g++. 1178 1179 st->print("["); 1180 bt.dump_on(st); 1181 st->print("]"); 1182 } 1183 st->print(" "); 1184 IfNode::dump_spec(st); 1185 } 1186 #endif 1187 1188 //============================================================================= 1189 //------------------------------is_member-------------------------------------- 1190 // Is 'l' a member of 'this'? 1191 int IdealLoopTree::is_member( const IdealLoopTree *l ) const { 1192 while( l->_nest > _nest ) l = l->_parent; 1193 return l == this; 1194 } 1195 1196 //------------------------------set_nest--------------------------------------- 1197 // Set loop tree nesting depth. Accumulate _has_call bits. 1198 int IdealLoopTree::set_nest( uint depth ) { 1199 _nest = depth; 1200 int bits = _has_call; 1201 if( _child ) bits |= _child->set_nest(depth+1); 1202 if( bits ) _has_call = 1; 1203 if( _next ) bits |= _next ->set_nest(depth ); 1204 return bits; 1205 } 1206 1207 //------------------------------split_fall_in---------------------------------- 1208 // Split out multiple fall-in edges from the loop header. Move them to a 1209 // private RegionNode before the loop. This becomes the loop landing pad. 1210 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { 1211 PhaseIterGVN &igvn = phase->_igvn; 1212 uint i; 1213 1214 // Make a new RegionNode to be the landing pad. 1215 Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 ); 1216 phase->set_loop(landing_pad,_parent); 1217 // Gather all the fall-in control paths into the landing pad 1218 uint icnt = fall_in_cnt; 1219 uint oreq = _head->req(); 1220 for( i = oreq-1; i>0; i-- ) 1221 if( !phase->is_member( this, _head->in(i) ) ) 1222 landing_pad->set_req(icnt--,_head->in(i)); 1223 1224 // Peel off PhiNode edges as well 1225 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1226 Node *oj = _head->fast_out(j); 1227 if( oj->is_Phi() ) { 1228 PhiNode* old_phi = oj->as_Phi(); 1229 assert( old_phi->region() == _head, "" ); 1230 igvn.hash_delete(old_phi); // Yank from hash before hacking edges 1231 Node *p = PhiNode::make_blank(landing_pad, old_phi); 1232 uint icnt = fall_in_cnt; 1233 for( i = oreq-1; i>0; i-- ) { 1234 if( !phase->is_member( this, _head->in(i) ) ) { 1235 p->init_req(icnt--, old_phi->in(i)); 1236 // Go ahead and clean out old edges from old phi 1237 old_phi->del_req(i); 1238 } 1239 } 1240 // Search for CSE's here, because ZKM.jar does a lot of 1241 // loop hackery and we need to be a little incremental 1242 // with the CSE to avoid O(N^2) node blow-up. 1243 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE 1244 if( p2 ) { // Found CSE 1245 p->destruct(); // Recover useless new node 1246 p = p2; // Use old node 1247 } else { 1248 igvn.register_new_node_with_optimizer(p, old_phi); 1249 } 1250 // Make old Phi refer to new Phi. 1251 old_phi->add_req(p); 1252 // Check for the special case of making the old phi useless and 1253 // disappear it. In JavaGrande I have a case where this useless 1254 // Phi is the loop limit and prevents recognizing a CountedLoop 1255 // which in turn prevents removing an empty loop. 1256 Node *id_old_phi = old_phi->Identity( &igvn ); 1257 if( id_old_phi != old_phi ) { // Found a simple identity? 1258 // Note that I cannot call 'replace_node' here, because 1259 // that will yank the edge from old_phi to the Region and 1260 // I'm mid-iteration over the Region's uses. 1261 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { 1262 Node* use = old_phi->last_out(i); 1263 igvn.rehash_node_delayed(use); 1264 uint uses_found = 0; 1265 for (uint j = 0; j < use->len(); j++) { 1266 if (use->in(j) == old_phi) { 1267 if (j < use->req()) use->set_req (j, id_old_phi); 1268 else use->set_prec(j, id_old_phi); 1269 uses_found++; 1270 } 1271 } 1272 i -= uses_found; // we deleted 1 or more copies of this edge 1273 } 1274 } 1275 igvn._worklist.push(old_phi); 1276 } 1277 } 1278 // Finally clean out the fall-in edges from the RegionNode 1279 for( i = oreq-1; i>0; i-- ) { 1280 if( !phase->is_member( this, _head->in(i) ) ) { 1281 _head->del_req(i); 1282 } 1283 } 1284 // Transform landing pad 1285 igvn.register_new_node_with_optimizer(landing_pad, _head); 1286 // Insert landing pad into the header 1287 _head->add_req(landing_pad); 1288 } 1289 1290 //------------------------------split_outer_loop------------------------------- 1291 // Split out the outermost loop from this shared header. 1292 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { 1293 PhaseIterGVN &igvn = phase->_igvn; 1294 1295 // Find index of outermost loop; it should also be my tail. 1296 uint outer_idx = 1; 1297 while( _head->in(outer_idx) != _tail ) outer_idx++; 1298 1299 // Make a LoopNode for the outermost loop. 1300 Node *ctl = _head->in(LoopNode::EntryControl); 1301 Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) ); 1302 outer = igvn.register_new_node_with_optimizer(outer, _head); 1303 phase->set_created_loop_node(); 1304 1305 // Outermost loop falls into '_head' loop 1306 _head->set_req(LoopNode::EntryControl, outer); 1307 _head->del_req(outer_idx); 1308 // Split all the Phis up between '_head' loop and 'outer' loop. 1309 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1310 Node *out = _head->fast_out(j); 1311 if( out->is_Phi() ) { 1312 PhiNode *old_phi = out->as_Phi(); 1313 assert( old_phi->region() == _head, "" ); 1314 Node *phi = PhiNode::make_blank(outer, old_phi); 1315 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); 1316 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); 1317 phi = igvn.register_new_node_with_optimizer(phi, old_phi); 1318 // Make old Phi point to new Phi on the fall-in path 1319 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi); 1320 old_phi->del_req(outer_idx); 1321 } 1322 } 1323 1324 // Use the new loop head instead of the old shared one 1325 _head = outer; 1326 phase->set_loop(_head, this); 1327 } 1328 1329 //------------------------------fix_parent------------------------------------- 1330 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { 1331 loop->_parent = parent; 1332 if( loop->_child ) fix_parent( loop->_child, loop ); 1333 if( loop->_next ) fix_parent( loop->_next , parent ); 1334 } 1335 1336 //------------------------------estimate_path_freq----------------------------- 1337 static float estimate_path_freq( Node *n ) { 1338 // Try to extract some path frequency info 1339 IfNode *iff; 1340 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests 1341 uint nop = n->Opcode(); 1342 if( nop == Op_SafePoint ) { // Skip any safepoint 1343 n = n->in(0); 1344 continue; 1345 } 1346 if( nop == Op_CatchProj ) { // Get count from a prior call 1347 // Assume call does not always throw exceptions: means the call-site 1348 // count is also the frequency of the fall-through path. 1349 assert( n->is_CatchProj(), "" ); 1350 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) 1351 return 0.0f; // Assume call exception path is rare 1352 Node *call = n->in(0)->in(0)->in(0); 1353 assert( call->is_Call(), "expect a call here" ); 1354 const JVMState *jvms = ((CallNode*)call)->jvms(); 1355 ciMethodData* methodData = jvms->method()->method_data(); 1356 if (!methodData->is_mature()) return 0.0f; // No call-site data 1357 ciProfileData* data = methodData->bci_to_data(jvms->bci()); 1358 if ((data == NULL) || !data->is_CounterData()) { 1359 // no call profile available, try call's control input 1360 n = n->in(0); 1361 continue; 1362 } 1363 return data->as_CounterData()->count()/FreqCountInvocations; 1364 } 1365 // See if there's a gating IF test 1366 Node *n_c = n->in(0); 1367 if( !n_c->is_If() ) break; // No estimate available 1368 iff = n_c->as_If(); 1369 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? 1370 // Compute how much count comes on this path 1371 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; 1372 // Have no count info. Skip dull uncommon-trap like branches. 1373 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || 1374 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) 1375 break; 1376 // Skip through never-taken branch; look for a real loop exit. 1377 n = iff->in(0); 1378 } 1379 return 0.0f; // No estimate available 1380 } 1381 1382 //------------------------------merge_many_backedges--------------------------- 1383 // Merge all the backedges from the shared header into a private Region. 1384 // Feed that region as the one backedge to this loop. 1385 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { 1386 uint i; 1387 1388 // Scan for the top 2 hottest backedges 1389 float hotcnt = 0.0f; 1390 float warmcnt = 0.0f; 1391 uint hot_idx = 0; 1392 // Loop starts at 2 because slot 1 is the fall-in path 1393 for( i = 2; i < _head->req(); i++ ) { 1394 float cnt = estimate_path_freq(_head->in(i)); 1395 if( cnt > hotcnt ) { // Grab hottest path 1396 warmcnt = hotcnt; 1397 hotcnt = cnt; 1398 hot_idx = i; 1399 } else if( cnt > warmcnt ) { // And 2nd hottest path 1400 warmcnt = cnt; 1401 } 1402 } 1403 1404 // See if the hottest backedge is worthy of being an inner loop 1405 // by being much hotter than the next hottest backedge. 1406 if( hotcnt <= 0.0001 || 1407 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge 1408 1409 // Peel out the backedges into a private merge point; peel 1410 // them all except optionally hot_idx. 1411 PhaseIterGVN &igvn = phase->_igvn; 1412 1413 Node *hot_tail = NULL; 1414 // Make a Region for the merge point 1415 Node *r = new (phase->C) RegionNode(1); 1416 for( i = 2; i < _head->req(); i++ ) { 1417 if( i != hot_idx ) 1418 r->add_req( _head->in(i) ); 1419 else hot_tail = _head->in(i); 1420 } 1421 igvn.register_new_node_with_optimizer(r, _head); 1422 // Plug region into end of loop _head, followed by hot_tail 1423 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); 1424 _head->set_req(2, r); 1425 if( hot_idx ) _head->add_req(hot_tail); 1426 1427 // Split all the Phis up between '_head' loop and the Region 'r' 1428 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1429 Node *out = _head->fast_out(j); 1430 if( out->is_Phi() ) { 1431 PhiNode* n = out->as_Phi(); 1432 igvn.hash_delete(n); // Delete from hash before hacking edges 1433 Node *hot_phi = NULL; 1434 Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type()); 1435 // Check all inputs for the ones to peel out 1436 uint j = 1; 1437 for( uint i = 2; i < n->req(); i++ ) { 1438 if( i != hot_idx ) 1439 phi->set_req( j++, n->in(i) ); 1440 else hot_phi = n->in(i); 1441 } 1442 // Register the phi but do not transform until whole place transforms 1443 igvn.register_new_node_with_optimizer(phi, n); 1444 // Add the merge phi to the old Phi 1445 while( n->req() > 3 ) n->del_req( n->req()-1 ); 1446 n->set_req(2, phi); 1447 if( hot_idx ) n->add_req(hot_phi); 1448 } 1449 } 1450 1451 1452 // Insert a new IdealLoopTree inserted below me. Turn it into a clone 1453 // of self loop tree. Turn self into a loop headed by _head and with 1454 // tail being the new merge point. 1455 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); 1456 phase->set_loop(_tail,ilt); // Adjust tail 1457 _tail = r; // Self's tail is new merge point 1458 phase->set_loop(r,this); 1459 ilt->_child = _child; // New guy has my children 1460 _child = ilt; // Self has new guy as only child 1461 ilt->_parent = this; // new guy has self for parent 1462 ilt->_nest = _nest; // Same nesting depth (for now) 1463 1464 // Starting with 'ilt', look for child loop trees using the same shared 1465 // header. Flatten these out; they will no longer be loops in the end. 1466 IdealLoopTree **pilt = &_child; 1467 while( ilt ) { 1468 if( ilt->_head == _head ) { 1469 uint i; 1470 for( i = 2; i < _head->req(); i++ ) 1471 if( _head->in(i) == ilt->_tail ) 1472 break; // Still a loop 1473 if( i == _head->req() ) { // No longer a loop 1474 // Flatten ilt. Hang ilt's "_next" list from the end of 1475 // ilt's '_child' list. Move the ilt's _child up to replace ilt. 1476 IdealLoopTree **cp = &ilt->_child; 1477 while( *cp ) cp = &(*cp)->_next; // Find end of child list 1478 *cp = ilt->_next; // Hang next list at end of child list 1479 *pilt = ilt->_child; // Move child up to replace ilt 1480 ilt->_head = NULL; // Flag as a loop UNIONED into parent 1481 ilt = ilt->_child; // Repeat using new ilt 1482 continue; // do not advance over ilt->_child 1483 } 1484 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); 1485 phase->set_loop(_head,ilt); 1486 } 1487 pilt = &ilt->_child; // Advance to next 1488 ilt = *pilt; 1489 } 1490 1491 if( _child ) fix_parent( _child, this ); 1492 } 1493 1494 //------------------------------beautify_loops--------------------------------- 1495 // Split shared headers and insert loop landing pads. 1496 // Insert a LoopNode to replace the RegionNode. 1497 // Return TRUE if loop tree is structurally changed. 1498 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { 1499 bool result = false; 1500 // Cache parts in locals for easy 1501 PhaseIterGVN &igvn = phase->_igvn; 1502 1503 igvn.hash_delete(_head); // Yank from hash before hacking edges 1504 1505 // Check for multiple fall-in paths. Peel off a landing pad if need be. 1506 int fall_in_cnt = 0; 1507 for( uint i = 1; i < _head->req(); i++ ) 1508 if( !phase->is_member( this, _head->in(i) ) ) 1509 fall_in_cnt++; 1510 assert( fall_in_cnt, "at least 1 fall-in path" ); 1511 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins 1512 split_fall_in( phase, fall_in_cnt ); 1513 1514 // Swap inputs to the _head and all Phis to move the fall-in edge to 1515 // the left. 1516 fall_in_cnt = 1; 1517 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) 1518 fall_in_cnt++; 1519 if( fall_in_cnt > 1 ) { 1520 // Since I am just swapping inputs I do not need to update def-use info 1521 Node *tmp = _head->in(1); 1522 _head->set_req( 1, _head->in(fall_in_cnt) ); 1523 _head->set_req( fall_in_cnt, tmp ); 1524 // Swap also all Phis 1525 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { 1526 Node* phi = _head->fast_out(i); 1527 if( phi->is_Phi() ) { 1528 igvn.hash_delete(phi); // Yank from hash before hacking edges 1529 tmp = phi->in(1); 1530 phi->set_req( 1, phi->in(fall_in_cnt) ); 1531 phi->set_req( fall_in_cnt, tmp ); 1532 } 1533 } 1534 } 1535 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); 1536 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); 1537 1538 // If I am a shared header (multiple backedges), peel off the many 1539 // backedges into a private merge point and use the merge point as 1540 // the one true backedge. 1541 if( _head->req() > 3 ) { 1542 // Merge the many backedges into a single backedge but leave 1543 // the hottest backedge as separate edge for the following peel. 1544 merge_many_backedges( phase ); 1545 result = true; 1546 } 1547 1548 // If I have one hot backedge, peel off myself loop. 1549 // I better be the outermost loop. 1550 if (_head->req() > 3 && !_irreducible) { 1551 split_outer_loop( phase ); 1552 result = true; 1553 1554 } else if (!_head->is_Loop() && !_irreducible) { 1555 // Make a new LoopNode to replace the old loop head 1556 Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) ); 1557 l = igvn.register_new_node_with_optimizer(l, _head); 1558 phase->set_created_loop_node(); 1559 // Go ahead and replace _head 1560 phase->_igvn.replace_node( _head, l ); 1561 _head = l; 1562 phase->set_loop(_head, this); 1563 } 1564 1565 // Now recursively beautify nested loops 1566 if( _child ) result |= _child->beautify_loops( phase ); 1567 if( _next ) result |= _next ->beautify_loops( phase ); 1568 return result; 1569 } 1570 1571 //------------------------------allpaths_check_safepts---------------------------- 1572 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 1573 // encountered. Helper for check_safepts. 1574 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { 1575 assert(stack.size() == 0, "empty stack"); 1576 stack.push(_tail); 1577 visited.Clear(); 1578 visited.set(_tail->_idx); 1579 while (stack.size() > 0) { 1580 Node* n = stack.pop(); 1581 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 1582 // Terminate this path 1583 } else if (n->Opcode() == Op_SafePoint) { 1584 if (_phase->get_loop(n) != this) { 1585 if (_required_safept == NULL) _required_safept = new Node_List(); 1586 _required_safept->push(n); // save the one closest to the tail 1587 } 1588 // Terminate this path 1589 } else { 1590 uint start = n->is_Region() ? 1 : 0; 1591 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; 1592 for (uint i = start; i < end; i++) { 1593 Node* in = n->in(i); 1594 assert(in->is_CFG(), "must be"); 1595 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { 1596 stack.push(in); 1597 } 1598 } 1599 } 1600 } 1601 } 1602 1603 //------------------------------check_safepts---------------------------- 1604 // Given dominators, try to find loops with calls that must always be 1605 // executed (call dominates loop tail). These loops do not need non-call 1606 // safepoints (ncsfpt). 1607 // 1608 // A complication is that a safepoint in a inner loop may be needed 1609 // by an outer loop. In the following, the inner loop sees it has a 1610 // call (block 3) on every path from the head (block 2) to the 1611 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) 1612 // in block 2, _but_ this leaves the outer loop without a safepoint. 1613 // 1614 // entry 0 1615 // | 1616 // v 1617 // outer 1,2 +->1 1618 // | | 1619 // | v 1620 // | 2<---+ ncsfpt in 2 1621 // |_/|\ | 1622 // | v | 1623 // inner 2,3 / 3 | call in 3 1624 // / | | 1625 // v +--+ 1626 // exit 4 1627 // 1628 // 1629 // This method creates a list (_required_safept) of ncsfpt nodes that must 1630 // be protected is created for each loop. When a ncsfpt maybe deleted, it 1631 // is first looked for in the lists for the outer loops of the current loop. 1632 // 1633 // The insights into the problem: 1634 // A) counted loops are okay 1635 // B) innermost loops are okay (only an inner loop can delete 1636 // a ncsfpt needed by an outer loop) 1637 // C) a loop is immune from an inner loop deleting a safepoint 1638 // if the loop has a call on the idom-path 1639 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the 1640 // idom-path that is not in a nested loop 1641 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner 1642 // loop needs to be prevented from deletion by an inner loop 1643 // 1644 // There are two analyses: 1645 // 1) The first, and cheaper one, scans the loop body from 1646 // tail to head following the idom (immediate dominator) 1647 // chain, looking for the cases (C,D,E) above. 1648 // Since inner loops are scanned before outer loops, there is summary 1649 // information about inner loops. Inner loops can be skipped over 1650 // when the tail of an inner loop is encountered. 1651 // 1652 // 2) The second, invoked if the first fails to find a call or ncsfpt on 1653 // the idom path (which is rare), scans all predecessor control paths 1654 // from the tail to the head, terminating a path when a call or sfpt 1655 // is encountered, to find the ncsfpt's that are closest to the tail. 1656 // 1657 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { 1658 // Bottom up traversal 1659 IdealLoopTree* ch = _child; 1660 if (_child) _child->check_safepts(visited, stack); 1661 if (_next) _next ->check_safepts(visited, stack); 1662 1663 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { 1664 bool has_call = false; // call on dom-path 1665 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth 1666 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth 1667 // Scan the dom-path nodes from tail to head 1668 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { 1669 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 1670 has_call = true; 1671 _has_sfpt = 1; // Then no need for a safept! 1672 break; 1673 } else if (n->Opcode() == Op_SafePoint) { 1674 if (_phase->get_loop(n) == this) { 1675 has_local_ncsfpt = true; 1676 break; 1677 } 1678 if (nonlocal_ncsfpt == NULL) { 1679 nonlocal_ncsfpt = n; // save the one closest to the tail 1680 } 1681 } else { 1682 IdealLoopTree* nlpt = _phase->get_loop(n); 1683 if (this != nlpt) { 1684 // If at an inner loop tail, see if the inner loop has already 1685 // recorded seeing a call on the dom-path (and stop.) If not, 1686 // jump to the head of the inner loop. 1687 assert(is_member(nlpt), "nested loop"); 1688 Node* tail = nlpt->_tail; 1689 if (tail->in(0)->is_If()) tail = tail->in(0); 1690 if (n == tail) { 1691 // If inner loop has call on dom-path, so does outer loop 1692 if (nlpt->_has_sfpt) { 1693 has_call = true; 1694 _has_sfpt = 1; 1695 break; 1696 } 1697 // Skip to head of inner loop 1698 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); 1699 n = nlpt->_head; 1700 } 1701 } 1702 } 1703 } 1704 // Record safept's that this loop needs preserved when an 1705 // inner loop attempts to delete it's safepoints. 1706 if (_child != NULL && !has_call && !has_local_ncsfpt) { 1707 if (nonlocal_ncsfpt != NULL) { 1708 if (_required_safept == NULL) _required_safept = new Node_List(); 1709 _required_safept->push(nonlocal_ncsfpt); 1710 } else { 1711 // Failed to find a suitable safept on the dom-path. Now use 1712 // an all paths walk from tail to head, looking for safepoints to preserve. 1713 allpaths_check_safepts(visited, stack); 1714 } 1715 } 1716 } 1717 } 1718 1719 //---------------------------is_deleteable_safept---------------------------- 1720 // Is safept not required by an outer loop? 1721 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { 1722 assert(sfpt->Opcode() == Op_SafePoint, ""); 1723 IdealLoopTree* lp = get_loop(sfpt)->_parent; 1724 while (lp != NULL) { 1725 Node_List* sfpts = lp->_required_safept; 1726 if (sfpts != NULL) { 1727 for (uint i = 0; i < sfpts->size(); i++) { 1728 if (sfpt == sfpts->at(i)) 1729 return false; 1730 } 1731 } 1732 lp = lp->_parent; 1733 } 1734 return true; 1735 } 1736 1737 //---------------------------replace_parallel_iv------------------------------- 1738 // Replace parallel induction variable (parallel to trip counter) 1739 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) { 1740 assert(loop->_head->is_CountedLoop(), ""); 1741 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1742 if (!cl->is_valid_counted_loop()) 1743 return; // skip malformed counted loop 1744 Node *incr = cl->incr(); 1745 if (incr == NULL) 1746 return; // Dead loop? 1747 Node *init = cl->init_trip(); 1748 Node *phi = cl->phi(); 1749 int stride_con = cl->stride_con(); 1750 1751 // Visit all children, looking for Phis 1752 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { 1753 Node *out = cl->out(i); 1754 // Look for other phis (secondary IVs). Skip dead ones 1755 if (!out->is_Phi() || out == phi || !has_node(out)) 1756 continue; 1757 PhiNode* phi2 = out->as_Phi(); 1758 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); 1759 // Look for induction variables of the form: X += constant 1760 if (phi2->region() != loop->_head || 1761 incr2->req() != 3 || 1762 incr2->in(1) != phi2 || 1763 incr2 == incr || 1764 incr2->Opcode() != Op_AddI || 1765 !incr2->in(2)->is_Con()) 1766 continue; 1767 1768 // Check for parallel induction variable (parallel to trip counter) 1769 // via an affine function. In particular, count-down loops with 1770 // count-up array indices are common. We only RCE references off 1771 // the trip-counter, so we need to convert all these to trip-counter 1772 // expressions. 1773 Node *init2 = phi2->in( LoopNode::EntryControl ); 1774 int stride_con2 = incr2->in(2)->get_int(); 1775 1776 // The ratio of the two strides cannot be represented as an int 1777 // if stride_con2 is min_int and stride_con is -1. 1778 if (stride_con2 == min_jint && stride_con == -1) { 1779 continue; 1780 } 1781 1782 // The general case here gets a little tricky. We want to find the 1783 // GCD of all possible parallel IV's and make a new IV using this 1784 // GCD for the loop. Then all possible IVs are simple multiples of 1785 // the GCD. In practice, this will cover very few extra loops. 1786 // Instead we require 'stride_con2' to be a multiple of 'stride_con', 1787 // where +/-1 is the common case, but other integer multiples are 1788 // also easy to handle. 1789 int ratio_con = stride_con2/stride_con; 1790 1791 if ((ratio_con * stride_con) == stride_con2) { // Check for exact 1792 #ifndef PRODUCT 1793 if (TraceLoopOpts) { 1794 tty->print("Parallel IV: %d ", phi2->_idx); 1795 loop->dump_head(); 1796 } 1797 #endif 1798 // Convert to using the trip counter. The parallel induction 1799 // variable differs from the trip counter by a loop-invariant 1800 // amount, the difference between their respective initial values. 1801 // It is scaled by the 'ratio_con'. 1802 Node* ratio = _igvn.intcon(ratio_con); 1803 set_ctrl(ratio, C->root()); 1804 Node* ratio_init = new (C) MulINode(init, ratio); 1805 _igvn.register_new_node_with_optimizer(ratio_init, init); 1806 set_early_ctrl(ratio_init); 1807 Node* diff = new (C) SubINode(init2, ratio_init); 1808 _igvn.register_new_node_with_optimizer(diff, init2); 1809 set_early_ctrl(diff); 1810 Node* ratio_idx = new (C) MulINode(phi, ratio); 1811 _igvn.register_new_node_with_optimizer(ratio_idx, phi); 1812 set_ctrl(ratio_idx, cl); 1813 Node* add = new (C) AddINode(ratio_idx, diff); 1814 _igvn.register_new_node_with_optimizer(add); 1815 set_ctrl(add, cl); 1816 _igvn.replace_node( phi2, add ); 1817 // Sometimes an induction variable is unused 1818 if (add->outcnt() == 0) { 1819 _igvn.remove_dead_node(add); 1820 } 1821 --i; // deleted this phi; rescan starting with next position 1822 continue; 1823 } 1824 } 1825 } 1826 1827 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) { 1828 Node* keep = NULL; 1829 if (keep_one) { 1830 // Look for a safepoint on the idom-path. 1831 for (Node* i = tail(); i != _head; i = phase->idom(i)) { 1832 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) { 1833 keep = i; 1834 break; // Found one 1835 } 1836 } 1837 } 1838 1839 // Don't remove any safepoints if it is requested to keep a single safepoint and 1840 // no safepoint was found on idom-path. It is not safe to remove any safepoint 1841 // in this case since there's no safepoint dominating all paths in the loop body. 1842 bool prune = !keep_one || keep != NULL; 1843 1844 // Delete other safepoints in this loop. 1845 Node_List* sfpts = _safepts; 1846 if (prune && sfpts != NULL) { 1847 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint"); 1848 for (uint i = 0; i < sfpts->size(); i++) { 1849 Node* n = sfpts->at(i); 1850 assert(phase->get_loop(n) == this, ""); 1851 if (n != keep && phase->is_deleteable_safept(n)) { 1852 phase->lazy_replace(n, n->in(TypeFunc::Control)); 1853 } 1854 } 1855 } 1856 } 1857 1858 //------------------------------counted_loop----------------------------------- 1859 // Convert to counted loops where possible 1860 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { 1861 1862 // For grins, set the inner-loop flag here 1863 if (!_child) { 1864 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop(); 1865 } 1866 1867 if (_head->is_CountedLoop() || 1868 phase->is_counted_loop(_head, this)) { 1869 1870 if (!UseCountedLoopSafepoints) { 1871 // Indicate we do not need a safepoint here 1872 _has_sfpt = 1; 1873 } 1874 1875 // Remove safepoints 1876 bool keep_one_sfpt = !(_has_call || _has_sfpt); 1877 remove_safepoints(phase, keep_one_sfpt); 1878 1879 // Look for induction variables 1880 phase->replace_parallel_iv(this); 1881 1882 } else if (_parent != NULL && !_irreducible) { 1883 // Not a counted loop. Keep one safepoint. 1884 bool keep_one_sfpt = true; 1885 remove_safepoints(phase, keep_one_sfpt); 1886 } 1887 1888 // Recursively 1889 if (_child) _child->counted_loop( phase ); 1890 if (_next) _next ->counted_loop( phase ); 1891 } 1892 1893 #ifndef PRODUCT 1894 //------------------------------dump_head-------------------------------------- 1895 // Dump 1 liner for loop header info 1896 void IdealLoopTree::dump_head( ) const { 1897 for (uint i=0; i<_nest; i++) 1898 tty->print(" "); 1899 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); 1900 if (_irreducible) tty->print(" IRREDUCIBLE"); 1901 Node* entry = _head->in(LoopNode::EntryControl); 1902 if (LoopLimitCheck) { 1903 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 1904 if (predicate != NULL ) { 1905 tty->print(" limit_check"); 1906 entry = entry->in(0)->in(0); 1907 } 1908 } 1909 if (UseLoopPredicate) { 1910 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 1911 if (entry != NULL) { 1912 tty->print(" predicated"); 1913 } 1914 } 1915 if (_head->is_CountedLoop()) { 1916 CountedLoopNode *cl = _head->as_CountedLoop(); 1917 tty->print(" counted"); 1918 1919 Node* init_n = cl->init_trip(); 1920 if (init_n != NULL && init_n->is_Con()) 1921 tty->print(" [%d,", cl->init_trip()->get_int()); 1922 else 1923 tty->print(" [int,"); 1924 Node* limit_n = cl->limit(); 1925 if (limit_n != NULL && limit_n->is_Con()) 1926 tty->print("%d),", cl->limit()->get_int()); 1927 else 1928 tty->print("int),"); 1929 int stride_con = cl->stride_con(); 1930 if (stride_con > 0) tty->print("+"); 1931 tty->print("%d", stride_con); 1932 1933 tty->print(" (%d iters) ", (int)cl->profile_trip_cnt()); 1934 1935 if (cl->is_pre_loop ()) tty->print(" pre" ); 1936 if (cl->is_main_loop()) tty->print(" main"); 1937 if (cl->is_post_loop()) tty->print(" post"); 1938 } 1939 if (_has_call) tty->print(" has_call"); 1940 if (_has_sfpt) tty->print(" has_sfpt"); 1941 if (_rce_candidate) tty->print(" rce"); 1942 if (_safepts != NULL && _safepts->size() > 0) { 1943 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }"); 1944 } 1945 if (_required_safept != NULL && _required_safept->size() > 0) { 1946 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }"); 1947 } 1948 tty->cr(); 1949 } 1950 1951 //------------------------------dump------------------------------------------- 1952 // Dump loops by loop tree 1953 void IdealLoopTree::dump( ) const { 1954 dump_head(); 1955 if (_child) _child->dump(); 1956 if (_next) _next ->dump(); 1957 } 1958 1959 #endif 1960 1961 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { 1962 if (loop == root) { 1963 if (loop->_child != NULL) { 1964 log->begin_head("loop_tree"); 1965 log->end_head(); 1966 if( loop->_child ) log_loop_tree(root, loop->_child, log); 1967 log->tail("loop_tree"); 1968 assert(loop->_next == NULL, "what?"); 1969 } 1970 } else { 1971 Node* head = loop->_head; 1972 log->begin_head("loop"); 1973 log->print(" idx='%d' ", head->_idx); 1974 if (loop->_irreducible) log->print("irreducible='1' "); 1975 if (head->is_Loop()) { 1976 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); 1977 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); 1978 } 1979 if (head->is_CountedLoop()) { 1980 CountedLoopNode* cl = head->as_CountedLoop(); 1981 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); 1982 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); 1983 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); 1984 } 1985 log->end_head(); 1986 if( loop->_child ) log_loop_tree(root, loop->_child, log); 1987 log->tail("loop"); 1988 if( loop->_next ) log_loop_tree(root, loop->_next, log); 1989 } 1990 } 1991 1992 //---------------------collect_potentially_useful_predicates----------------------- 1993 // Helper function to collect potentially useful predicates to prevent them from 1994 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates 1995 void PhaseIdealLoop::collect_potentially_useful_predicates( 1996 IdealLoopTree * loop, Unique_Node_List &useful_predicates) { 1997 if (loop->_child) { // child 1998 collect_potentially_useful_predicates(loop->_child, useful_predicates); 1999 } 2000 2001 // self (only loops that we can apply loop predication may use their predicates) 2002 if (loop->_head->is_Loop() && 2003 !loop->_irreducible && 2004 !loop->tail()->is_top()) { 2005 LoopNode* lpn = loop->_head->as_Loop(); 2006 Node* entry = lpn->in(LoopNode::EntryControl); 2007 Node* predicate_proj = find_predicate(entry); // loop_limit_check first 2008 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication 2009 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be"); 2010 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2011 entry = entry->in(0)->in(0); 2012 } 2013 predicate_proj = find_predicate(entry); // Predicate 2014 if (predicate_proj != NULL ) { 2015 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2016 } 2017 } 2018 2019 if (loop->_next) { // sibling 2020 collect_potentially_useful_predicates(loop->_next, useful_predicates); 2021 } 2022 } 2023 2024 //------------------------eliminate_useless_predicates----------------------------- 2025 // Eliminate all inserted predicates if they could not be used by loop predication. 2026 // Note: it will also eliminates loop limits check predicate since it also uses 2027 // Opaque1 node (see Parse::add_predicate()). 2028 void PhaseIdealLoop::eliminate_useless_predicates() { 2029 if (C->predicate_count() == 0) 2030 return; // no predicate left 2031 2032 Unique_Node_List useful_predicates; // to store useful predicates 2033 if (C->has_loops()) { 2034 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); 2035 } 2036 2037 for (int i = C->predicate_count(); i > 0; i--) { 2038 Node * n = C->predicate_opaque1_node(i-1); 2039 assert(n->Opcode() == Op_Opaque1, "must be"); 2040 if (!useful_predicates.member(n)) { // not in the useful list 2041 _igvn.replace_node(n, n->in(1)); 2042 } 2043 } 2044 } 2045 2046 //------------------------process_expensive_nodes----------------------------- 2047 // Expensive nodes have their control input set to prevent the GVN 2048 // from commoning them and as a result forcing the resulting node to 2049 // be in a more frequent path. Use CFG information here, to change the 2050 // control inputs so that some expensive nodes can be commoned while 2051 // not executed more frequently. 2052 bool PhaseIdealLoop::process_expensive_nodes() { 2053 assert(OptimizeExpensiveOps, "optimization off?"); 2054 2055 // Sort nodes to bring similar nodes together 2056 C->sort_expensive_nodes(); 2057 2058 bool progress = false; 2059 2060 for (int i = 0; i < C->expensive_count(); ) { 2061 Node* n = C->expensive_node(i); 2062 int start = i; 2063 // Find nodes similar to n 2064 i++; 2065 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++); 2066 int end = i; 2067 // And compare them two by two 2068 for (int j = start; j < end; j++) { 2069 Node* n1 = C->expensive_node(j); 2070 if (is_node_unreachable(n1)) { 2071 continue; 2072 } 2073 for (int k = j+1; k < end; k++) { 2074 Node* n2 = C->expensive_node(k); 2075 if (is_node_unreachable(n2)) { 2076 continue; 2077 } 2078 2079 assert(n1 != n2, "should be pair of nodes"); 2080 2081 Node* c1 = n1->in(0); 2082 Node* c2 = n2->in(0); 2083 2084 Node* parent_c1 = c1; 2085 Node* parent_c2 = c2; 2086 2087 // The call to get_early_ctrl_for_expensive() moves the 2088 // expensive nodes up but stops at loops that are in a if 2089 // branch. See whether we can exit the loop and move above the 2090 // If. 2091 if (c1->is_Loop()) { 2092 parent_c1 = c1->in(1); 2093 } 2094 if (c2->is_Loop()) { 2095 parent_c2 = c2->in(1); 2096 } 2097 2098 if (parent_c1 == parent_c2) { 2099 _igvn._worklist.push(n1); 2100 _igvn._worklist.push(n2); 2101 continue; 2102 } 2103 2104 // Look for identical expensive node up the dominator chain. 2105 if (is_dominator(c1, c2)) { 2106 c2 = c1; 2107 } else if (is_dominator(c2, c1)) { 2108 c1 = c2; 2109 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() && 2110 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) { 2111 // Both branches have the same expensive node so move it up 2112 // before the if. 2113 c1 = c2 = idom(parent_c1->in(0)); 2114 } 2115 // Do the actual moves 2116 if (n1->in(0) != c1) { 2117 _igvn.hash_delete(n1); 2118 n1->set_req(0, c1); 2119 _igvn.hash_insert(n1); 2120 _igvn._worklist.push(n1); 2121 progress = true; 2122 } 2123 if (n2->in(0) != c2) { 2124 _igvn.hash_delete(n2); 2125 n2->set_req(0, c2); 2126 _igvn.hash_insert(n2); 2127 _igvn._worklist.push(n2); 2128 progress = true; 2129 } 2130 } 2131 } 2132 } 2133 2134 return progress; 2135 } 2136 2137 2138 //============================================================================= 2139 //----------------------------build_and_optimize------------------------------- 2140 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to 2141 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. 2142 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) { 2143 ResourceMark rm; 2144 2145 int old_progress = C->major_progress(); 2146 uint orig_worklist_size = _igvn._worklist.size(); 2147 2148 // Reset major-progress flag for the driver's heuristics 2149 C->clear_major_progress(); 2150 2151 #ifndef PRODUCT 2152 // Capture for later assert 2153 uint unique = C->unique(); 2154 _loop_invokes++; 2155 _loop_work += unique; 2156 #endif 2157 2158 // True if the method has at least 1 irreducible loop 2159 _has_irreducible_loops = false; 2160 2161 _created_loop_node = false; 2162 2163 Arena *a = Thread::current()->resource_area(); 2164 VectorSet visited(a); 2165 // Pre-grow the mapping from Nodes to IdealLoopTrees. 2166 _nodes.map(C->unique(), NULL); 2167 memset(_nodes.adr(), 0, wordSize * C->unique()); 2168 2169 // Pre-build the top-level outermost loop tree entry 2170 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); 2171 // Do not need a safepoint at the top level 2172 _ltree_root->_has_sfpt = 1; 2173 2174 // Initialize Dominators. 2175 // Checked in clone_loop_predicate() during beautify_loops(). 2176 _idom_size = 0; 2177 _idom = NULL; 2178 _dom_depth = NULL; 2179 _dom_stk = NULL; 2180 2181 // Empty pre-order array 2182 allocate_preorders(); 2183 2184 // Build a loop tree on the fly. Build a mapping from CFG nodes to 2185 // IdealLoopTree entries. Data nodes are NOT walked. 2186 build_loop_tree(); 2187 // Check for bailout, and return 2188 if (C->failing()) { 2189 return; 2190 } 2191 2192 // No loops after all 2193 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); 2194 2195 // There should always be an outer loop containing the Root and Return nodes. 2196 // If not, we have a degenerate empty program. Bail out in this case. 2197 if (!has_node(C->root())) { 2198 if (!_verify_only) { 2199 C->clear_major_progress(); 2200 C->record_method_not_compilable("empty program detected during loop optimization"); 2201 } 2202 return; 2203 } 2204 2205 // Nothing to do, so get out 2206 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only; 2207 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn); 2208 if (stop_early && !do_expensive_nodes) { 2209 _igvn.optimize(); // Cleanup NeverBranches 2210 return; 2211 } 2212 2213 // Set loop nesting depth 2214 _ltree_root->set_nest( 0 ); 2215 2216 // Split shared headers and insert loop landing pads. 2217 // Do not bother doing this on the Root loop of course. 2218 if( !_verify_me && !_verify_only && _ltree_root->_child ) { 2219 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3); 2220 if( _ltree_root->_child->beautify_loops( this ) ) { 2221 // Re-build loop tree! 2222 _ltree_root->_child = NULL; 2223 _nodes.clear(); 2224 reallocate_preorders(); 2225 build_loop_tree(); 2226 // Check for bailout, and return 2227 if (C->failing()) { 2228 return; 2229 } 2230 // Reset loop nesting depth 2231 _ltree_root->set_nest( 0 ); 2232 2233 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3); 2234 } 2235 } 2236 2237 // Build Dominators for elision of NULL checks & loop finding. 2238 // Since nodes do not have a slot for immediate dominator, make 2239 // a persistent side array for that info indexed on node->_idx. 2240 _idom_size = C->unique(); 2241 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); 2242 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); 2243 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth 2244 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); 2245 2246 Dominators(); 2247 2248 if (!_verify_only) { 2249 // As a side effect, Dominators removed any unreachable CFG paths 2250 // into RegionNodes. It doesn't do this test against Root, so 2251 // we do it here. 2252 for( uint i = 1; i < C->root()->req(); i++ ) { 2253 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? 2254 _igvn.delete_input_of(C->root(), i); 2255 i--; // Rerun same iteration on compressed edges 2256 } 2257 } 2258 2259 // Given dominators, try to find inner loops with calls that must 2260 // always be executed (call dominates loop tail). These loops do 2261 // not need a separate safepoint. 2262 Node_List cisstack(a); 2263 _ltree_root->check_safepts(visited, cisstack); 2264 } 2265 2266 // Walk the DATA nodes and place into loops. Find earliest control 2267 // node. For CFG nodes, the _nodes array starts out and remains 2268 // holding the associated IdealLoopTree pointer. For DATA nodes, the 2269 // _nodes array holds the earliest legal controlling CFG node. 2270 2271 // Allocate stack with enough space to avoid frequent realloc 2272 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats 2273 Node_Stack nstack( a, stack_size ); 2274 2275 visited.Clear(); 2276 Node_List worklist(a); 2277 // Don't need C->root() on worklist since 2278 // it will be processed among C->top() inputs 2279 worklist.push( C->top() ); 2280 visited.set( C->top()->_idx ); // Set C->top() as visited now 2281 build_loop_early( visited, worklist, nstack ); 2282 2283 // Given early legal placement, try finding counted loops. This placement 2284 // is good enough to discover most loop invariants. 2285 if( !_verify_me && !_verify_only ) 2286 _ltree_root->counted_loop( this ); 2287 2288 // Find latest loop placement. Find ideal loop placement. 2289 visited.Clear(); 2290 init_dom_lca_tags(); 2291 // Need C->root() on worklist when processing outs 2292 worklist.push( C->root() ); 2293 NOT_PRODUCT( C->verify_graph_edges(); ) 2294 worklist.push( C->top() ); 2295 build_loop_late( visited, worklist, nstack ); 2296 2297 if (_verify_only) { 2298 // restore major progress flag 2299 for (int i = 0; i < old_progress; i++) 2300 C->set_major_progress(); 2301 assert(C->unique() == unique, "verification mode made Nodes? ? ?"); 2302 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything"); 2303 return; 2304 } 2305 2306 // clear out the dead code after build_loop_late 2307 while (_deadlist.size()) { 2308 _igvn.remove_globally_dead_node(_deadlist.pop()); 2309 } 2310 2311 if (stop_early) { 2312 assert(do_expensive_nodes, "why are we here?"); 2313 if (process_expensive_nodes()) { 2314 // If we made some progress when processing expensive nodes then 2315 // the IGVN may modify the graph in a way that will allow us to 2316 // make some more progress: we need to try processing expensive 2317 // nodes again. 2318 C->set_major_progress(); 2319 } 2320 _igvn.optimize(); 2321 return; 2322 } 2323 2324 // Some parser-inserted loop predicates could never be used by loop 2325 // predication or they were moved away from loop during some optimizations. 2326 // For example, peeling. Eliminate them before next loop optimizations. 2327 if (UseLoopPredicate || LoopLimitCheck) { 2328 eliminate_useless_predicates(); 2329 } 2330 2331 #ifndef PRODUCT 2332 C->verify_graph_edges(); 2333 if (_verify_me) { // Nested verify pass? 2334 // Check to see if the verify mode is broken 2335 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); 2336 return; 2337 } 2338 if(VerifyLoopOptimizations) verify(); 2339 if(TraceLoopOpts && C->has_loops()) { 2340 _ltree_root->dump(); 2341 } 2342 #endif 2343 2344 if (skip_loop_opts) { 2345 // restore major progress flag 2346 for (int i = 0; i < old_progress; i++) { 2347 C->set_major_progress(); 2348 } 2349 2350 if (UseShenandoahGC && !C->major_progress()) { 2351 shenandoah_pin_and_expand_barriers(); 2352 } 2353 2354 // Cleanup any modified bits 2355 _igvn.optimize(); 2356 2357 if (C->log() != NULL) { 2358 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2359 } 2360 return; 2361 } 2362 2363 if (ReassociateInvariants) { 2364 // Reassociate invariants and prep for split_thru_phi 2365 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2366 IdealLoopTree* lpt = iter.current(); 2367 if (!lpt->is_counted() || !lpt->is_inner()) continue; 2368 2369 lpt->reassociate_invariants(this); 2370 2371 // Because RCE opportunities can be masked by split_thru_phi, 2372 // look for RCE candidates and inhibit split_thru_phi 2373 // on just their loop-phi's for this pass of loop opts 2374 if (SplitIfBlocks && do_split_ifs) { 2375 if (lpt->policy_range_check(this)) { 2376 lpt->_rce_candidate = 1; // = true 2377 } 2378 } 2379 } 2380 } 2381 2382 // Check for aggressive application of split-if and other transforms 2383 // that require basic-block info (like cloning through Phi's) 2384 if( SplitIfBlocks && do_split_ifs ) { 2385 visited.Clear(); 2386 split_if_with_blocks( visited, nstack ); 2387 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); 2388 } 2389 2390 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) { 2391 C->set_major_progress(); 2392 } 2393 2394 // Perform loop predication before iteration splitting 2395 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) { 2396 _ltree_root->_child->loop_predication(this); 2397 } 2398 2399 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { 2400 if (do_intrinsify_fill()) { 2401 C->set_major_progress(); 2402 } 2403 } 2404 2405 // Perform iteration-splitting on inner loops. Split iterations to avoid 2406 // range checks or one-shot null checks. 2407 2408 // If split-if's didn't hack the graph too bad (no CFG changes) 2409 // then do loop opts. 2410 if (C->has_loops() && !C->major_progress()) { 2411 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 2412 _ltree_root->_child->iteration_split( this, worklist ); 2413 // No verify after peeling! GCM has hoisted code out of the loop. 2414 // After peeling, the hoisted code could sink inside the peeled area. 2415 // The peeling code does not try to recompute the best location for 2416 // all the code before the peeled area, so the verify pass will always 2417 // complain about it. 2418 } 2419 // Do verify graph edges in any case 2420 NOT_PRODUCT( C->verify_graph_edges(); ); 2421 2422 if (!do_split_ifs) { 2423 // We saw major progress in Split-If to get here. We forced a 2424 // pass with unrolling and not split-if, however more split-if's 2425 // might make progress. If the unrolling didn't make progress 2426 // then the major-progress flag got cleared and we won't try 2427 // another round of Split-If. In particular the ever-common 2428 // instance-of/check-cast pattern requires at least 2 rounds of 2429 // Split-If to clear out. 2430 C->set_major_progress(); 2431 } 2432 2433 // Repeat loop optimizations if new loops were seen 2434 if (created_loop_node()) { 2435 C->set_major_progress(); 2436 } 2437 2438 // Keep loop predicates and perform optimizations with them 2439 // until no more loop optimizations could be done. 2440 // After that switch predicates off and do more loop optimizations. 2441 if (!C->major_progress() && (C->predicate_count() > 0)) { 2442 C->cleanup_loop_predicates(_igvn); 2443 #ifndef PRODUCT 2444 if (TraceLoopOpts) { 2445 tty->print_cr("PredicatesOff"); 2446 } 2447 #endif 2448 C->set_major_progress(); 2449 } 2450 2451 // Convert scalar to superword operations at the end of all loop opts. 2452 if (UseSuperWord && C->has_loops() && !C->major_progress()) { 2453 // SuperWord transform 2454 SuperWord sw(this); 2455 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2456 IdealLoopTree* lpt = iter.current(); 2457 if (lpt->is_counted()) { 2458 sw.transform_loop(lpt); 2459 } 2460 } 2461 } 2462 2463 // Cleanup any modified bits 2464 _igvn.optimize(); 2465 2466 // disable assert until issue with split_flow_path is resolved (6742111) 2467 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), 2468 // "shouldn't introduce irreducible loops"); 2469 2470 if (C->log() != NULL) { 2471 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2472 } 2473 } 2474 2475 #ifndef PRODUCT 2476 //------------------------------print_statistics------------------------------- 2477 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes 2478 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique 2479 void PhaseIdealLoop::print_statistics() { 2480 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); 2481 } 2482 2483 //------------------------------verify----------------------------------------- 2484 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. 2485 static int fail; // debug only, so its multi-thread dont care 2486 void PhaseIdealLoop::verify() const { 2487 int old_progress = C->major_progress(); 2488 ResourceMark rm; 2489 PhaseIdealLoop loop_verify( _igvn, this ); 2490 VectorSet visited(Thread::current()->resource_area()); 2491 2492 fail = 0; 2493 verify_compare( C->root(), &loop_verify, visited ); 2494 assert( fail == 0, "verify loops failed" ); 2495 // Verify loop structure is the same 2496 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); 2497 // Reset major-progress. It was cleared by creating a verify version of 2498 // PhaseIdealLoop. 2499 for( int i=0; i<old_progress; i++ ) 2500 C->set_major_progress(); 2501 } 2502 2503 //------------------------------verify_compare--------------------------------- 2504 // Make sure me and the given PhaseIdealLoop agree on key data structures 2505 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { 2506 if( !n ) return; 2507 if( visited.test_set( n->_idx ) ) return; 2508 if( !_nodes[n->_idx] ) { // Unreachable 2509 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); 2510 return; 2511 } 2512 2513 uint i; 2514 for( i = 0; i < n->req(); i++ ) 2515 verify_compare( n->in(i), loop_verify, visited ); 2516 2517 // Check the '_nodes' block/loop structure 2518 i = n->_idx; 2519 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl 2520 if( _nodes[i] != loop_verify->_nodes[i] && 2521 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { 2522 tty->print("Mismatched control setting for: "); 2523 n->dump(); 2524 if( fail++ > 10 ) return; 2525 Node *c = get_ctrl_no_update(n); 2526 tty->print("We have it as: "); 2527 if( c->in(0) ) c->dump(); 2528 else tty->print_cr("N%d",c->_idx); 2529 tty->print("Verify thinks: "); 2530 if( loop_verify->has_ctrl(n) ) 2531 loop_verify->get_ctrl_no_update(n)->dump(); 2532 else 2533 loop_verify->get_loop_idx(n)->dump(); 2534 tty->cr(); 2535 } 2536 } else { // We have a loop 2537 IdealLoopTree *us = get_loop_idx(n); 2538 if( loop_verify->has_ctrl(n) ) { 2539 tty->print("Mismatched loop setting for: "); 2540 n->dump(); 2541 if( fail++ > 10 ) return; 2542 tty->print("We have it as: "); 2543 us->dump(); 2544 tty->print("Verify thinks: "); 2545 loop_verify->get_ctrl_no_update(n)->dump(); 2546 tty->cr(); 2547 } else if (!C->major_progress()) { 2548 // Loop selection can be messed up if we did a major progress 2549 // operation, like split-if. Do not verify in that case. 2550 IdealLoopTree *them = loop_verify->get_loop_idx(n); 2551 if( us->_head != them->_head || us->_tail != them->_tail ) { 2552 tty->print("Unequals loops for: "); 2553 n->dump(); 2554 if( fail++ > 10 ) return; 2555 tty->print("We have it as: "); 2556 us->dump(); 2557 tty->print("Verify thinks: "); 2558 them->dump(); 2559 tty->cr(); 2560 } 2561 } 2562 } 2563 2564 // Check for immediate dominators being equal 2565 if( i >= _idom_size ) { 2566 if( !n->is_CFG() ) return; 2567 tty->print("CFG Node with no idom: "); 2568 n->dump(); 2569 return; 2570 } 2571 if( !n->is_CFG() ) return; 2572 if( n == C->root() ) return; // No IDOM here 2573 2574 assert(n->_idx == i, "sanity"); 2575 Node *id = idom_no_update(n); 2576 if( id != loop_verify->idom_no_update(n) ) { 2577 tty->print("Unequals idoms for: "); 2578 n->dump(); 2579 if( fail++ > 10 ) return; 2580 tty->print("We have it as: "); 2581 id->dump(); 2582 tty->print("Verify thinks: "); 2583 loop_verify->idom_no_update(n)->dump(); 2584 tty->cr(); 2585 } 2586 2587 } 2588 2589 //------------------------------verify_tree------------------------------------ 2590 // Verify that tree structures match. Because the CFG can change, siblings 2591 // within the loop tree can be reordered. We attempt to deal with that by 2592 // reordering the verify's loop tree if possible. 2593 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { 2594 assert( _parent == parent, "Badly formed loop tree" ); 2595 2596 // Siblings not in same order? Attempt to re-order. 2597 if( _head != loop->_head ) { 2598 // Find _next pointer to update 2599 IdealLoopTree **pp = &loop->_parent->_child; 2600 while( *pp != loop ) 2601 pp = &((*pp)->_next); 2602 // Find proper sibling to be next 2603 IdealLoopTree **nn = &loop->_next; 2604 while( (*nn) && (*nn)->_head != _head ) 2605 nn = &((*nn)->_next); 2606 2607 // Check for no match. 2608 if( !(*nn) ) { 2609 // Annoyingly, irreducible loops can pick different headers 2610 // after a major_progress operation, so the rest of the loop 2611 // tree cannot be matched. 2612 if (_irreducible && Compile::current()->major_progress()) return; 2613 assert( 0, "failed to match loop tree" ); 2614 } 2615 2616 // Move (*nn) to (*pp) 2617 IdealLoopTree *hit = *nn; 2618 *nn = hit->_next; 2619 hit->_next = loop; 2620 *pp = loop; 2621 loop = hit; 2622 // Now try again to verify 2623 } 2624 2625 assert( _head == loop->_head , "mismatched loop head" ); 2626 Node *tail = _tail; // Inline a non-updating version of 2627 while( !tail->in(0) ) // the 'tail()' call. 2628 tail = tail->in(1); 2629 assert( tail == loop->_tail, "mismatched loop tail" ); 2630 2631 // Counted loops that are guarded should be able to find their guards 2632 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { 2633 CountedLoopNode *cl = _head->as_CountedLoop(); 2634 Node *init = cl->init_trip(); 2635 Node *ctrl = cl->in(LoopNode::EntryControl); 2636 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 2637 Node *iff = ctrl->in(0); 2638 assert( iff->Opcode() == Op_If, "" ); 2639 Node *bol = iff->in(1); 2640 assert( bol->Opcode() == Op_Bool, "" ); 2641 Node *cmp = bol->in(1); 2642 assert( cmp->Opcode() == Op_CmpI, "" ); 2643 Node *add = cmp->in(1); 2644 Node *opaq; 2645 if( add->Opcode() == Op_Opaque1 ) { 2646 opaq = add; 2647 } else { 2648 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); 2649 assert( add == init, "" ); 2650 opaq = cmp->in(2); 2651 } 2652 assert( opaq->Opcode() == Op_Opaque1, "" ); 2653 2654 } 2655 2656 if (_child != NULL) _child->verify_tree(loop->_child, this); 2657 if (_next != NULL) _next ->verify_tree(loop->_next, parent); 2658 // Innermost loops need to verify loop bodies, 2659 // but only if no 'major_progress' 2660 int fail = 0; 2661 if (!Compile::current()->major_progress() && _child == NULL) { 2662 for( uint i = 0; i < _body.size(); i++ ) { 2663 Node *n = _body.at(i); 2664 if (n->outcnt() == 0) continue; // Ignore dead 2665 uint j; 2666 for( j = 0; j < loop->_body.size(); j++ ) 2667 if( loop->_body.at(j) == n ) 2668 break; 2669 if( j == loop->_body.size() ) { // Not found in loop body 2670 // Last ditch effort to avoid assertion: Its possible that we 2671 // have some users (so outcnt not zero) but are still dead. 2672 // Try to find from root. 2673 if (Compile::current()->root()->find(n->_idx)) { 2674 fail++; 2675 tty->print("We have that verify does not: "); 2676 n->dump(); 2677 } 2678 } 2679 } 2680 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { 2681 Node *n = loop->_body.at(i2); 2682 if (n->outcnt() == 0) continue; // Ignore dead 2683 uint j; 2684 for( j = 0; j < _body.size(); j++ ) 2685 if( _body.at(j) == n ) 2686 break; 2687 if( j == _body.size() ) { // Not found in loop body 2688 // Last ditch effort to avoid assertion: Its possible that we 2689 // have some users (so outcnt not zero) but are still dead. 2690 // Try to find from root. 2691 if (Compile::current()->root()->find(n->_idx)) { 2692 fail++; 2693 tty->print("Verify has that we do not: "); 2694 n->dump(); 2695 } 2696 } 2697 } 2698 assert( !fail, "loop body mismatch" ); 2699 } 2700 } 2701 2702 #endif 2703 2704 //------------------------------set_idom--------------------------------------- 2705 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { 2706 uint idx = d->_idx; 2707 if (idx >= _idom_size) { 2708 uint newsize = _idom_size<<1; 2709 while( idx >= newsize ) { 2710 newsize <<= 1; 2711 } 2712 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); 2713 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); 2714 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); 2715 _idom_size = newsize; 2716 } 2717 _idom[idx] = n; 2718 _dom_depth[idx] = dom_depth; 2719 } 2720 2721 //------------------------------recompute_dom_depth--------------------------------------- 2722 // The dominator tree is constructed with only parent pointers. 2723 // This recomputes the depth in the tree by first tagging all 2724 // nodes as "no depth yet" marker. The next pass then runs up 2725 // the dom tree from each node marked "no depth yet", and computes 2726 // the depth on the way back down. 2727 void PhaseIdealLoop::recompute_dom_depth() { 2728 uint no_depth_marker = C->unique(); 2729 uint i; 2730 // Initialize depth to "no depth yet" 2731 for (i = 0; i < _idom_size; i++) { 2732 if (_dom_depth[i] > 0 && _idom[i] != NULL) { 2733 _dom_depth[i] = no_depth_marker; 2734 } 2735 } 2736 if (_dom_stk == NULL) { 2737 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size. 2738 if (init_size < 10) init_size = 10; 2739 _dom_stk = new GrowableArray<uint>(init_size); 2740 } 2741 // Compute new depth for each node. 2742 for (i = 0; i < _idom_size; i++) { 2743 uint j = i; 2744 // Run up the dom tree to find a node with a depth 2745 while (_dom_depth[j] == no_depth_marker) { 2746 _dom_stk->push(j); 2747 j = _idom[j]->_idx; 2748 } 2749 // Compute the depth on the way back down this tree branch 2750 uint dd = _dom_depth[j] + 1; 2751 while (_dom_stk->length() > 0) { 2752 uint j = _dom_stk->pop(); 2753 _dom_depth[j] = dd; 2754 dd++; 2755 } 2756 } 2757 } 2758 2759 //------------------------------sort------------------------------------------- 2760 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the 2761 // loop tree, not the root. 2762 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { 2763 if( !innermost ) return loop; // New innermost loop 2764 2765 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number 2766 assert( loop_preorder, "not yet post-walked loop" ); 2767 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer 2768 IdealLoopTree *l = *pp; // Do I go before or after 'l'? 2769 2770 // Insert at start of list 2771 while( l ) { // Insertion sort based on pre-order 2772 if( l == loop ) return innermost; // Already on list! 2773 int l_preorder = get_preorder(l->_head); // Cache pre-order number 2774 assert( l_preorder, "not yet post-walked l" ); 2775 // Check header pre-order number to figure proper nesting 2776 if( loop_preorder > l_preorder ) 2777 break; // End of insertion 2778 // If headers tie (e.g., shared headers) check tail pre-order numbers. 2779 // Since I split shared headers, you'd think this could not happen. 2780 // BUT: I must first do the preorder numbering before I can discover I 2781 // have shared headers, so the split headers all get the same preorder 2782 // number as the RegionNode they split from. 2783 if( loop_preorder == l_preorder && 2784 get_preorder(loop->_tail) < get_preorder(l->_tail) ) 2785 break; // Also check for shared headers (same pre#) 2786 pp = &l->_parent; // Chain up list 2787 l = *pp; 2788 } 2789 // Link into list 2790 // Point predecessor to me 2791 *pp = loop; 2792 // Point me to successor 2793 IdealLoopTree *p = loop->_parent; 2794 loop->_parent = l; // Point me to successor 2795 if( p ) sort( p, innermost ); // Insert my parents into list as well 2796 return innermost; 2797 } 2798 2799 //------------------------------build_loop_tree-------------------------------- 2800 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit 2801 // bits. The _nodes[] array is mapped by Node index and holds a NULL for 2802 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the 2803 // tightest enclosing IdealLoopTree for post-walked. 2804 // 2805 // During my forward walk I do a short 1-layer lookahead to see if I can find 2806 // a loop backedge with that doesn't have any work on the backedge. This 2807 // helps me construct nested loops with shared headers better. 2808 // 2809 // Once I've done the forward recursion, I do the post-work. For each child 2810 // I check to see if there is a backedge. Backedges define a loop! I 2811 // insert an IdealLoopTree at the target of the backedge. 2812 // 2813 // During the post-work I also check to see if I have several children 2814 // belonging to different loops. If so, then this Node is a decision point 2815 // where control flow can choose to change loop nests. It is at this 2816 // decision point where I can figure out how loops are nested. At this 2817 // time I can properly order the different loop nests from my children. 2818 // Note that there may not be any backedges at the decision point! 2819 // 2820 // Since the decision point can be far removed from the backedges, I can't 2821 // order my loops at the time I discover them. Thus at the decision point 2822 // I need to inspect loop header pre-order numbers to properly nest my 2823 // loops. This means I need to sort my childrens' loops by pre-order. 2824 // The sort is of size number-of-control-children, which generally limits 2825 // it to size 2 (i.e., I just choose between my 2 target loops). 2826 void PhaseIdealLoop::build_loop_tree() { 2827 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 2828 GrowableArray <Node *> bltstack(C->live_nodes() >> 1); 2829 Node *n = C->root(); 2830 bltstack.push(n); 2831 int pre_order = 1; 2832 int stack_size; 2833 2834 while ( ( stack_size = bltstack.length() ) != 0 ) { 2835 n = bltstack.top(); // Leave node on stack 2836 if ( !is_visited(n) ) { 2837 // ---- Pre-pass Work ---- 2838 // Pre-walked but not post-walked nodes need a pre_order number. 2839 2840 set_preorder_visited( n, pre_order ); // set as visited 2841 2842 // ---- Scan over children ---- 2843 // Scan first over control projections that lead to loop headers. 2844 // This helps us find inner-to-outer loops with shared headers better. 2845 2846 // Scan children's children for loop headers. 2847 for ( int i = n->outcnt() - 1; i >= 0; --i ) { 2848 Node* m = n->raw_out(i); // Child 2849 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children 2850 // Scan over children's children to find loop 2851 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 2852 Node* l = m->fast_out(j); 2853 if( is_visited(l) && // Been visited? 2854 !is_postvisited(l) && // But not post-visited 2855 get_preorder(l) < pre_order ) { // And smaller pre-order 2856 // Found! Scan the DFS down this path before doing other paths 2857 bltstack.push(m); 2858 break; 2859 } 2860 } 2861 } 2862 } 2863 pre_order++; 2864 } 2865 else if ( !is_postvisited(n) ) { 2866 // Note: build_loop_tree_impl() adds out edges on rare occasions, 2867 // such as com.sun.rsasign.am::a. 2868 // For non-recursive version, first, process current children. 2869 // On next iteration, check if additional children were added. 2870 for ( int k = n->outcnt() - 1; k >= 0; --k ) { 2871 Node* u = n->raw_out(k); 2872 if ( u->is_CFG() && !is_visited(u) ) { 2873 bltstack.push(u); 2874 } 2875 } 2876 if ( bltstack.length() == stack_size ) { 2877 // There were no additional children, post visit node now 2878 (void)bltstack.pop(); // Remove node from stack 2879 pre_order = build_loop_tree_impl( n, pre_order ); 2880 // Check for bailout 2881 if (C->failing()) { 2882 return; 2883 } 2884 // Check to grow _preorders[] array for the case when 2885 // build_loop_tree_impl() adds new nodes. 2886 check_grow_preorders(); 2887 } 2888 } 2889 else { 2890 (void)bltstack.pop(); // Remove post-visited node from stack 2891 } 2892 } 2893 } 2894 2895 //------------------------------build_loop_tree_impl--------------------------- 2896 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { 2897 // ---- Post-pass Work ---- 2898 // Pre-walked but not post-walked nodes need a pre_order number. 2899 2900 // Tightest enclosing loop for this Node 2901 IdealLoopTree *innermost = NULL; 2902 2903 // For all children, see if any edge is a backedge. If so, make a loop 2904 // for it. Then find the tightest enclosing loop for the self Node. 2905 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2906 Node* m = n->fast_out(i); // Child 2907 if( n == m ) continue; // Ignore control self-cycles 2908 if( !m->is_CFG() ) continue;// Ignore non-CFG edges 2909 2910 IdealLoopTree *l; // Child's loop 2911 if( !is_postvisited(m) ) { // Child visited but not post-visited? 2912 // Found a backedge 2913 assert( get_preorder(m) < pre_order, "should be backedge" ); 2914 // Check for the RootNode, which is already a LoopNode and is allowed 2915 // to have multiple "backedges". 2916 if( m == C->root()) { // Found the root? 2917 l = _ltree_root; // Root is the outermost LoopNode 2918 } else { // Else found a nested loop 2919 // Insert a LoopNode to mark this loop. 2920 l = new IdealLoopTree(this, m, n); 2921 } // End of Else found a nested loop 2922 if( !has_loop(m) ) // If 'm' does not already have a loop set 2923 set_loop(m, l); // Set loop header to loop now 2924 2925 } else { // Else not a nested loop 2926 if( !_nodes[m->_idx] ) continue; // Dead code has no loop 2927 l = get_loop(m); // Get previously determined loop 2928 // If successor is header of a loop (nest), move up-loop till it 2929 // is a member of some outer enclosing loop. Since there are no 2930 // shared headers (I've split them already) I only need to go up 2931 // at most 1 level. 2932 while( l && l->_head == m ) // Successor heads loop? 2933 l = l->_parent; // Move up 1 for me 2934 // If this loop is not properly parented, then this loop 2935 // has no exit path out, i.e. its an infinite loop. 2936 if( !l ) { 2937 // Make loop "reachable" from root so the CFG is reachable. Basically 2938 // insert a bogus loop exit that is never taken. 'm', the loop head, 2939 // points to 'n', one (of possibly many) fall-in paths. There may be 2940 // many backedges as well. 2941 2942 // Here I set the loop to be the root loop. I could have, after 2943 // inserting a bogus loop exit, restarted the recursion and found my 2944 // new loop exit. This would make the infinite loop a first-class 2945 // loop and it would then get properly optimized. What's the use of 2946 // optimizing an infinite loop? 2947 l = _ltree_root; // Oops, found infinite loop 2948 2949 if (!_verify_only) { 2950 // Insert the NeverBranch between 'm' and it's control user. 2951 NeverBranchNode *iff = new (C) NeverBranchNode( m ); 2952 _igvn.register_new_node_with_optimizer(iff); 2953 set_loop(iff, l); 2954 Node *if_t = new (C) CProjNode( iff, 0 ); 2955 _igvn.register_new_node_with_optimizer(if_t); 2956 set_loop(if_t, l); 2957 2958 Node* cfg = NULL; // Find the One True Control User of m 2959 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 2960 Node* x = m->fast_out(j); 2961 if (x->is_CFG() && x != m && x != iff) 2962 { cfg = x; break; } 2963 } 2964 assert(cfg != NULL, "must find the control user of m"); 2965 uint k = 0; // Probably cfg->in(0) 2966 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region 2967 cfg->set_req( k, if_t ); // Now point to NeverBranch 2968 2969 // Now create the never-taken loop exit 2970 Node *if_f = new (C) CProjNode( iff, 1 ); 2971 _igvn.register_new_node_with_optimizer(if_f); 2972 set_loop(if_f, l); 2973 // Find frame ptr for Halt. Relies on the optimizer 2974 // V-N'ing. Easier and quicker than searching through 2975 // the program structure. 2976 Node *frame = new (C) ParmNode( C->start(), TypeFunc::FramePtr ); 2977 _igvn.register_new_node_with_optimizer(frame); 2978 // Halt & Catch Fire 2979 Node *halt = new (C) HaltNode( if_f, frame ); 2980 _igvn.register_new_node_with_optimizer(halt); 2981 set_loop(halt, l); 2982 C->root()->add_req(halt); 2983 } 2984 set_loop(C->root(), _ltree_root); 2985 } 2986 } 2987 // Weeny check for irreducible. This child was already visited (this 2988 // IS the post-work phase). Is this child's loop header post-visited 2989 // as well? If so, then I found another entry into the loop. 2990 if (!_verify_only) { 2991 while( is_postvisited(l->_head) ) { 2992 // found irreducible 2993 l->_irreducible = 1; // = true 2994 l = l->_parent; 2995 _has_irreducible_loops = true; 2996 // Check for bad CFG here to prevent crash, and bailout of compile 2997 if (l == NULL) { 2998 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); 2999 return pre_order; 3000 } 3001 } 3002 C->set_has_irreducible_loop(_has_irreducible_loops); 3003 } 3004 3005 // This Node might be a decision point for loops. It is only if 3006 // it's children belong to several different loops. The sort call 3007 // does a trivial amount of work if there is only 1 child or all 3008 // children belong to the same loop. If however, the children 3009 // belong to different loops, the sort call will properly set the 3010 // _parent pointers to show how the loops nest. 3011 // 3012 // In any case, it returns the tightest enclosing loop. 3013 innermost = sort( l, innermost ); 3014 } 3015 3016 // Def-use info will have some dead stuff; dead stuff will have no 3017 // loop decided on. 3018 3019 // Am I a loop header? If so fix up my parent's child and next ptrs. 3020 if( innermost && innermost->_head == n ) { 3021 assert( get_loop(n) == innermost, "" ); 3022 IdealLoopTree *p = innermost->_parent; 3023 IdealLoopTree *l = innermost; 3024 while( p && l->_head == n ) { 3025 l->_next = p->_child; // Put self on parents 'next child' 3026 p->_child = l; // Make self as first child of parent 3027 l = p; // Now walk up the parent chain 3028 p = l->_parent; 3029 } 3030 } else { 3031 // Note that it is possible for a LoopNode to reach here, if the 3032 // backedge has been made unreachable (hence the LoopNode no longer 3033 // denotes a Loop, and will eventually be removed). 3034 3035 // Record tightest enclosing loop for self. Mark as post-visited. 3036 set_loop(n, innermost); 3037 // Also record has_call flag early on 3038 if( innermost ) { 3039 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { 3040 // Do not count uncommon calls 3041 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { 3042 Node *iff = n->in(0)->in(0); 3043 // No any calls for vectorized loops. 3044 if( UseSuperWord || !iff->is_If() || 3045 (n->in(0)->Opcode() == Op_IfFalse && 3046 (1.0 - iff->as_If()->_prob) >= 0.01) || 3047 (iff->as_If()->_prob >= 0.01) ) 3048 innermost->_has_call = 1; 3049 } 3050 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { 3051 // Disable loop optimizations if the loop has a scalar replaceable 3052 // allocation. This disabling may cause a potential performance lost 3053 // if the allocation is not eliminated for some reason. 3054 innermost->_allow_optimizations = false; 3055 innermost->_has_call = 1; // = true 3056 } else if (n->Opcode() == Op_SafePoint) { 3057 // Record all safepoints in this loop. 3058 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List(); 3059 innermost->_safepts->push(n); 3060 } 3061 } 3062 } 3063 3064 // Flag as post-visited now 3065 set_postvisited(n); 3066 return pre_order; 3067 } 3068 3069 3070 //------------------------------build_loop_early------------------------------- 3071 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3072 // First pass computes the earliest controlling node possible. This is the 3073 // controlling input with the deepest dominating depth. 3074 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3075 while (worklist.size() != 0) { 3076 // Use local variables nstack_top_n & nstack_top_i to cache values 3077 // on nstack's top. 3078 Node *nstack_top_n = worklist.pop(); 3079 uint nstack_top_i = 0; 3080 //while_nstack_nonempty: 3081 while (true) { 3082 // Get parent node and next input's index from stack's top. 3083 Node *n = nstack_top_n; 3084 uint i = nstack_top_i; 3085 uint cnt = n->req(); // Count of inputs 3086 if (i == 0) { // Pre-process the node. 3087 if( has_node(n) && // Have either loop or control already? 3088 !has_ctrl(n) ) { // Have loop picked out already? 3089 // During "merge_many_backedges" we fold up several nested loops 3090 // into a single loop. This makes the members of the original 3091 // loop bodies pointing to dead loops; they need to move up 3092 // to the new UNION'd larger loop. I set the _head field of these 3093 // dead loops to NULL and the _parent field points to the owning 3094 // loop. Shades of UNION-FIND algorithm. 3095 IdealLoopTree *ilt; 3096 while( !(ilt = get_loop(n))->_head ) { 3097 // Normally I would use a set_loop here. But in this one special 3098 // case, it is legal (and expected) to change what loop a Node 3099 // belongs to. 3100 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); 3101 } 3102 // Remove safepoints ONLY if I've already seen I don't need one. 3103 // (the old code here would yank a 2nd safepoint after seeing a 3104 // first one, even though the 1st did not dominate in the loop body 3105 // and thus could be avoided indefinitely) 3106 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && 3107 is_deleteable_safept(n)) { 3108 Node *in = n->in(TypeFunc::Control); 3109 lazy_replace(n,in); // Pull safepoint now 3110 if (ilt->_safepts != NULL) { 3111 ilt->_safepts->yank(n); 3112 } 3113 // Carry on with the recursion "as if" we are walking 3114 // only the control input 3115 if( !visited.test_set( in->_idx ) ) { 3116 worklist.push(in); // Visit this guy later, using worklist 3117 } 3118 // Get next node from nstack: 3119 // - skip n's inputs processing by setting i > cnt; 3120 // - we also will not call set_early_ctrl(n) since 3121 // has_node(n) == true (see the condition above). 3122 i = cnt + 1; 3123 } 3124 } 3125 } // if (i == 0) 3126 3127 // Visit all inputs 3128 bool done = true; // Assume all n's inputs will be processed 3129 while (i < cnt) { 3130 Node *in = n->in(i); 3131 ++i; 3132 if (in == NULL) continue; 3133 if (in->pinned() && !in->is_CFG()) 3134 set_ctrl(in, in->in(0)); 3135 int is_visited = visited.test_set( in->_idx ); 3136 if (!has_node(in)) { // No controlling input yet? 3137 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); 3138 assert( !is_visited, "visit only once" ); 3139 nstack.push(n, i); // Save parent node and next input's index. 3140 nstack_top_n = in; // Process current input now. 3141 nstack_top_i = 0; 3142 done = false; // Not all n's inputs processed. 3143 break; // continue while_nstack_nonempty; 3144 } else if (!is_visited) { 3145 // This guy has a location picked out for him, but has not yet 3146 // been visited. Happens to all CFG nodes, for instance. 3147 // Visit him using the worklist instead of recursion, to break 3148 // cycles. Since he has a location already we do not need to 3149 // find his location before proceeding with the current Node. 3150 worklist.push(in); // Visit this guy later, using worklist 3151 } 3152 } 3153 if (done) { 3154 // All of n's inputs have been processed, complete post-processing. 3155 3156 // Compute earliest point this Node can go. 3157 // CFG, Phi, pinned nodes already know their controlling input. 3158 if (!has_node(n)) { 3159 // Record earliest legal location 3160 set_early_ctrl( n ); 3161 } 3162 if (nstack.is_empty()) { 3163 // Finished all nodes on stack. 3164 // Process next node on the worklist. 3165 break; 3166 } 3167 // Get saved parent node and next input's index. 3168 nstack_top_n = nstack.node(); 3169 nstack_top_i = nstack.index(); 3170 nstack.pop(); 3171 } 3172 } // while (true) 3173 } 3174 } 3175 3176 //------------------------------dom_lca_internal-------------------------------- 3177 // Pair-wise LCA 3178 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { 3179 if( !n1 ) return n2; // Handle NULL original LCA 3180 assert( n1->is_CFG(), "" ); 3181 assert( n2->is_CFG(), "" ); 3182 // find LCA of all uses 3183 uint d1 = dom_depth(n1); 3184 uint d2 = dom_depth(n2); 3185 while (n1 != n2) { 3186 if (d1 > d2) { 3187 n1 = idom(n1); 3188 d1 = dom_depth(n1); 3189 } else if (d1 < d2) { 3190 n2 = idom(n2); 3191 d2 = dom_depth(n2); 3192 } else { 3193 // Here d1 == d2. Due to edits of the dominator-tree, sections 3194 // of the tree might have the same depth. These sections have 3195 // to be searched more carefully. 3196 3197 // Scan up all the n1's with equal depth, looking for n2. 3198 Node *t1 = idom(n1); 3199 while (dom_depth(t1) == d1) { 3200 if (t1 == n2) return n2; 3201 t1 = idom(t1); 3202 } 3203 // Scan up all the n2's with equal depth, looking for n1. 3204 Node *t2 = idom(n2); 3205 while (dom_depth(t2) == d2) { 3206 if (t2 == n1) return n1; 3207 t2 = idom(t2); 3208 } 3209 // Move up to a new dominator-depth value as well as up the dom-tree. 3210 n1 = t1; 3211 n2 = t2; 3212 d1 = dom_depth(n1); 3213 d2 = dom_depth(n2); 3214 } 3215 } 3216 return n1; 3217 } 3218 3219 //------------------------------compute_idom----------------------------------- 3220 // Locally compute IDOM using dom_lca call. Correct only if the incoming 3221 // IDOMs are correct. 3222 Node *PhaseIdealLoop::compute_idom( Node *region ) const { 3223 assert( region->is_Region(), "" ); 3224 Node *LCA = NULL; 3225 for( uint i = 1; i < region->req(); i++ ) { 3226 if( region->in(i) != C->top() ) 3227 LCA = dom_lca( LCA, region->in(i) ); 3228 } 3229 return LCA; 3230 } 3231 3232 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { 3233 bool had_error = false; 3234 #ifdef ASSERT 3235 if (early != C->root()) { 3236 // Make sure that there's a dominance path from LCA to early 3237 Node* d = LCA; 3238 while (d != early) { 3239 if (d == C->root()) { 3240 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA); 3241 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx); 3242 had_error = true; 3243 break; 3244 } 3245 d = idom(d); 3246 } 3247 } 3248 #endif 3249 return had_error; 3250 } 3251 3252 3253 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { 3254 // Compute LCA over list of uses 3255 bool had_error = false; 3256 Node *LCA = NULL; 3257 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { 3258 Node* c = n->fast_out(i); 3259 if (_nodes[c->_idx] == NULL) 3260 continue; // Skip the occasional dead node 3261 if( c->is_Phi() ) { // For Phis, we must land above on the path 3262 for( uint j=1; j<c->req(); j++ ) {// For all inputs 3263 if( c->in(j) == n ) { // Found matching input? 3264 Node *use = c->in(0)->in(j); 3265 if (_verify_only && use->is_top()) continue; 3266 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3267 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3268 } 3269 } 3270 } else { 3271 // For CFG data-users, use is in the block just prior 3272 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); 3273 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3274 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3275 } 3276 } 3277 assert(!had_error, "bad dominance"); 3278 return LCA; 3279 } 3280 3281 //------------------------------get_late_ctrl---------------------------------- 3282 // Compute latest legal control. 3283 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { 3284 assert(early != NULL, "early control should not be NULL"); 3285 3286 Node* LCA = compute_lca_of_uses(n, early); 3287 #ifdef ASSERT 3288 if (LCA == C->root() && LCA != early) { 3289 // def doesn't dominate uses so print some useful debugging output 3290 compute_lca_of_uses(n, early, true); 3291 } 3292 #endif 3293 3294 // if this is a load, check for anti-dependent stores 3295 // We use a conservative algorithm to identify potential interfering 3296 // instructions and for rescheduling the load. The users of the memory 3297 // input of this load are examined. Any use which is not a load and is 3298 // dominated by early is considered a potentially interfering store. 3299 // This can produce false positives. 3300 if (n->is_Load() && LCA != early) { 3301 Node_List worklist; 3302 3303 Node *mem = n->in(MemNode::Memory); 3304 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 3305 Node* s = mem->fast_out(i); 3306 worklist.push(s); 3307 } 3308 while(worklist.size() != 0 && LCA != early) { 3309 Node* s = worklist.pop(); 3310 if (s->is_Load() || 3311 (UseShenandoahGC && 3312 (s->is_ShenandoahBarrier() || s->Opcode() == Op_SafePoint || 3313 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)))) { 3314 continue; 3315 } else if (s->is_MergeMem()) { 3316 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 3317 Node* s1 = s->fast_out(i); 3318 worklist.push(s1); 3319 } 3320 } else { 3321 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); 3322 assert(sctrl != NULL || s->outcnt() == 0 || s->is_ShenandoahBarrier(), "must have control"); 3323 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { 3324 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); 3325 } 3326 } 3327 } 3328 } 3329 3330 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); 3331 return LCA; 3332 } 3333 3334 // true if CFG node d dominates CFG node n 3335 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { 3336 if (d == n) 3337 return true; 3338 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 3339 uint dd = dom_depth(d); 3340 while (dom_depth(n) >= dd) { 3341 if (n == d) 3342 return true; 3343 n = idom(n); 3344 } 3345 return false; 3346 } 3347 3348 //------------------------------dom_lca_for_get_late_ctrl_internal------------- 3349 // Pair-wise LCA with tags. 3350 // Tag each index with the node 'tag' currently being processed 3351 // before advancing up the dominator chain using idom(). 3352 // Later calls that find a match to 'tag' know that this path has already 3353 // been considered in the current LCA (which is input 'n1' by convention). 3354 // Since get_late_ctrl() is only called once for each node, the tag array 3355 // does not need to be cleared between calls to get_late_ctrl(). 3356 // Algorithm trades a larger constant factor for better asymptotic behavior 3357 // 3358 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { 3359 uint d1 = dom_depth(n1); 3360 uint d2 = dom_depth(n2); 3361 3362 do { 3363 if (d1 > d2) { 3364 // current lca is deeper than n2 3365 _dom_lca_tags.map(n1->_idx, tag); 3366 n1 = idom(n1); 3367 d1 = dom_depth(n1); 3368 } else if (d1 < d2) { 3369 // n2 is deeper than current lca 3370 Node *memo = _dom_lca_tags[n2->_idx]; 3371 if( memo == tag ) { 3372 return n1; // Return the current LCA 3373 } 3374 _dom_lca_tags.map(n2->_idx, tag); 3375 n2 = idom(n2); 3376 d2 = dom_depth(n2); 3377 } else { 3378 // Here d1 == d2. Due to edits of the dominator-tree, sections 3379 // of the tree might have the same depth. These sections have 3380 // to be searched more carefully. 3381 3382 // Scan up all the n1's with equal depth, looking for n2. 3383 _dom_lca_tags.map(n1->_idx, tag); 3384 Node *t1 = idom(n1); 3385 while (dom_depth(t1) == d1) { 3386 if (t1 == n2) return n2; 3387 _dom_lca_tags.map(t1->_idx, tag); 3388 t1 = idom(t1); 3389 } 3390 // Scan up all the n2's with equal depth, looking for n1. 3391 _dom_lca_tags.map(n2->_idx, tag); 3392 Node *t2 = idom(n2); 3393 while (dom_depth(t2) == d2) { 3394 if (t2 == n1) return n1; 3395 _dom_lca_tags.map(t2->_idx, tag); 3396 t2 = idom(t2); 3397 } 3398 // Move up to a new dominator-depth value as well as up the dom-tree. 3399 n1 = t1; 3400 n2 = t2; 3401 d1 = dom_depth(n1); 3402 d2 = dom_depth(n2); 3403 } 3404 } while (n1 != n2); 3405 return n1; 3406 } 3407 3408 //------------------------------init_dom_lca_tags------------------------------ 3409 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3410 // Intended use does not involve any growth for the array, so it could 3411 // be of fixed size. 3412 void PhaseIdealLoop::init_dom_lca_tags() { 3413 uint limit = C->unique() + 1; 3414 _dom_lca_tags.map( limit, NULL ); 3415 #ifdef ASSERT 3416 for( uint i = 0; i < limit; ++i ) { 3417 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3418 } 3419 #endif // ASSERT 3420 } 3421 3422 //------------------------------clear_dom_lca_tags------------------------------ 3423 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3424 // Intended use does not involve any growth for the array, so it could 3425 // be of fixed size. 3426 void PhaseIdealLoop::clear_dom_lca_tags() { 3427 uint limit = C->unique() + 1; 3428 _dom_lca_tags.map( limit, NULL ); 3429 _dom_lca_tags.clear(); 3430 #ifdef ASSERT 3431 for( uint i = 0; i < limit; ++i ) { 3432 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3433 } 3434 #endif // ASSERT 3435 } 3436 3437 //------------------------------build_loop_late-------------------------------- 3438 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3439 // Second pass finds latest legal placement, and ideal loop placement. 3440 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3441 while (worklist.size() != 0) { 3442 Node *n = worklist.pop(); 3443 // Only visit once 3444 if (visited.test_set(n->_idx)) continue; 3445 uint cnt = n->outcnt(); 3446 uint i = 0; 3447 while (true) { 3448 assert( _nodes[n->_idx], "no dead nodes" ); 3449 // Visit all children 3450 if (i < cnt) { 3451 Node* use = n->raw_out(i); 3452 ++i; 3453 // Check for dead uses. Aggressively prune such junk. It might be 3454 // dead in the global sense, but still have local uses so I cannot 3455 // easily call 'remove_dead_node'. 3456 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? 3457 // Due to cycles, we might not hit the same fixed point in the verify 3458 // pass as we do in the regular pass. Instead, visit such phis as 3459 // simple uses of the loop head. 3460 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { 3461 if( !visited.test(use->_idx) ) 3462 worklist.push(use); 3463 } else if( !visited.test_set(use->_idx) ) { 3464 nstack.push(n, i); // Save parent and next use's index. 3465 n = use; // Process all children of current use. 3466 cnt = use->outcnt(); 3467 i = 0; 3468 } 3469 } else { 3470 // Do not visit around the backedge of loops via data edges. 3471 // push dead code onto a worklist 3472 _deadlist.push(use); 3473 } 3474 } else { 3475 // All of n's children have been processed, complete post-processing. 3476 build_loop_late_post(n); 3477 if (nstack.is_empty()) { 3478 // Finished all nodes on stack. 3479 // Process next node on the worklist. 3480 break; 3481 } 3482 // Get saved parent node and next use's index. Visit the rest of uses. 3483 n = nstack.node(); 3484 cnt = n->outcnt(); 3485 i = nstack.index(); 3486 nstack.pop(); 3487 } 3488 } 3489 } 3490 } 3491 3492 //------------------------------build_loop_late_post--------------------------- 3493 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3494 // Second pass finds latest legal placement, and ideal loop placement. 3495 void PhaseIdealLoop::build_loop_late_post( Node *n ) { 3496 3497 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) { 3498 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. 3499 } 3500 3501 #ifdef ASSERT 3502 if (_verify_only && !n->is_CFG()) { 3503 // Check def-use domination. 3504 compute_lca_of_uses(n, get_ctrl(n), true /* verify */); 3505 } 3506 #endif 3507 3508 // CFG and pinned nodes already handled 3509 if( n->in(0) ) { 3510 if( n->in(0)->is_top() ) return; // Dead? 3511 3512 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads 3513 // _must_ be pinned (they have to observe their control edge of course). 3514 // Unlike Stores (which modify an unallocable resource, the memory 3515 // state), Mods/Loads can float around. So free them up. 3516 bool pinned = true; 3517 switch( n->Opcode() ) { 3518 case Op_DivI: 3519 case Op_DivF: 3520 case Op_DivD: 3521 case Op_ModI: 3522 case Op_ModF: 3523 case Op_ModD: 3524 case Op_LoadB: // Same with Loads; they can sink 3525 case Op_LoadUB: // during loop optimizations. 3526 case Op_LoadUS: 3527 case Op_LoadD: 3528 case Op_LoadF: 3529 case Op_LoadI: 3530 case Op_LoadKlass: 3531 case Op_LoadNKlass: 3532 case Op_LoadL: 3533 case Op_LoadS: 3534 case Op_LoadP: 3535 case Op_LoadN: 3536 case Op_LoadRange: 3537 case Op_LoadD_unaligned: 3538 case Op_LoadL_unaligned: 3539 case Op_StrComp: // Does a bunch of load-like effects 3540 case Op_StrEquals: 3541 case Op_StrIndexOf: 3542 case Op_AryEq: 3543 case Op_ShenandoahReadBarrier: 3544 case Op_ShenandoahWriteBarrier: 3545 case Op_ShenandoahWBMemProj: 3546 pinned = false; 3547 } 3548 if( pinned ) { 3549 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); 3550 if( !chosen_loop->_child ) // Inner loop? 3551 chosen_loop->_body.push(n); // Collect inner loops 3552 return; 3553 } 3554 } else { // No slot zero 3555 if( n->is_CFG() ) { // CFG with no slot 0 is dead 3556 _nodes.map(n->_idx,0); // No block setting, it's globally dead 3557 return; 3558 } 3559 assert(!n->is_CFG() || n->outcnt() == 0, ""); 3560 } 3561 3562 // Do I have a "safe range" I can select over? 3563 Node *early = get_ctrl(n);// Early location already computed 3564 3565 // Compute latest point this Node can go 3566 Node *LCA = get_late_ctrl( n, early ); 3567 // LCA is NULL due to uses being dead 3568 if( LCA == NULL ) { 3569 #ifdef ASSERT 3570 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { 3571 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); 3572 } 3573 #endif 3574 _nodes.map(n->_idx, 0); // This node is useless 3575 _deadlist.push(n); 3576 return; 3577 } 3578 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); 3579 3580 Node *legal = LCA; // Walk 'legal' up the IDOM chain 3581 Node *least = legal; // Best legal position so far 3582 while( early != legal ) { // While not at earliest legal 3583 #ifdef ASSERT 3584 if (legal->is_Start() && !early->is_Root()) { 3585 // Bad graph. Print idom path and fail. 3586 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA); 3587 assert(false, "Bad graph detected in build_loop_late"); 3588 } 3589 #endif 3590 // Find least loop nesting depth 3591 legal = idom(legal); // Bump up the IDOM tree 3592 // Check for lower nesting depth 3593 if( get_loop(legal)->_nest < get_loop(least)->_nest ) 3594 least = legal; 3595 } 3596 assert(early == legal || legal != C->root(), "bad dominance of inputs"); 3597 3598 // Try not to place code on a loop entry projection 3599 // which can inhibit range check elimination. 3600 if (least != early) { 3601 Node* ctrl_out = least->unique_ctrl_out(); 3602 if (UseShenandoahGC && 3603 ctrl_out && ctrl_out->is_Loop() && 3604 least == ctrl_out->in(LoopNode::EntryControl)) { 3605 Node* new_ctrl = least; 3606 if (find_predicate_insertion_point(new_ctrl, Deoptimization::Reason_loop_limit_check) != NULL) { 3607 new_ctrl = new_ctrl->in(0)->in(0); 3608 assert(is_dominator(early, new_ctrl), "least != early so we can move up the dominator tree"); 3609 } 3610 if (find_predicate_insertion_point(new_ctrl, Deoptimization::Reason_predicate) != NULL) { 3611 Node* c = new_ctrl->in(0)->in(0); 3612 assert(is_dominator(early, c), "least != early so we can move up the dominator tree"); 3613 new_ctrl = c; 3614 } 3615 if (new_ctrl != least) { 3616 least = new_ctrl; 3617 } else if (ctrl_out->is_CountedLoop()) { 3618 Node* least_dom = idom(least); 3619 if (get_loop(least_dom)->is_member(get_loop(least))) { 3620 least = least_dom; 3621 } 3622 } 3623 } else if (ctrl_out && ctrl_out->is_CountedLoop() && 3624 least == ctrl_out->in(LoopNode::EntryControl)) { 3625 Node* least_dom = idom(least); 3626 if (get_loop(least_dom)->is_member(get_loop(least))) { 3627 least = least_dom; 3628 } 3629 } 3630 } 3631 3632 #ifdef ASSERT 3633 // If verifying, verify that 'verify_me' has a legal location 3634 // and choose it as our location. 3635 if( _verify_me ) { 3636 Node *v_ctrl = _verify_me->get_ctrl_no_update(n); 3637 Node *legal = LCA; 3638 while( early != legal ) { // While not at earliest legal 3639 if( legal == v_ctrl ) break; // Check for prior good location 3640 legal = idom(legal) ;// Bump up the IDOM tree 3641 } 3642 // Check for prior good location 3643 if( legal == v_ctrl ) least = legal; // Keep prior if found 3644 } 3645 #endif 3646 3647 // Assign discovered "here or above" point 3648 least = find_non_split_ctrl(least); 3649 set_ctrl(n, least); 3650 3651 // Collect inner loop bodies 3652 IdealLoopTree *chosen_loop = get_loop(least); 3653 if( !chosen_loop->_child ) // Inner loop? 3654 chosen_loop->_body.push(n);// Collect inner loops 3655 3656 if (n->Opcode() == Op_ShenandoahWriteBarrier) { 3657 // The write barrier and its memory proj must have the same 3658 // control otherwise some loop opts could put nodes (Phis) between 3659 // them 3660 Node* proj = n->find_out_with(Op_ShenandoahWBMemProj); 3661 if (proj != NULL) { 3662 set_ctrl_and_loop(proj, least); 3663 } 3664 } 3665 } 3666 3667 #ifdef ASSERT 3668 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) { 3669 tty->print_cr("%s", msg); 3670 tty->print("n: "); n->dump(); 3671 tty->print("early(n): "); early->dump(); 3672 if (n->in(0) != NULL && !n->in(0)->is_top() && 3673 n->in(0) != early && !n->in(0)->is_Root()) { 3674 tty->print("n->in(0): "); n->in(0)->dump(); 3675 } 3676 for (uint i = 1; i < n->req(); i++) { 3677 Node* in1 = n->in(i); 3678 if (in1 != NULL && in1 != n && !in1->is_top()) { 3679 tty->print("n->in(%d): ", i); in1->dump(); 3680 Node* in1_early = get_ctrl(in1); 3681 tty->print("early(n->in(%d)): ", i); in1_early->dump(); 3682 if (in1->in(0) != NULL && !in1->in(0)->is_top() && 3683 in1->in(0) != in1_early && !in1->in(0)->is_Root()) { 3684 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump(); 3685 } 3686 for (uint j = 1; j < in1->req(); j++) { 3687 Node* in2 = in1->in(j); 3688 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) { 3689 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump(); 3690 Node* in2_early = get_ctrl(in2); 3691 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); 3692 if (in2->in(0) != NULL && !in2->in(0)->is_top() && 3693 in2->in(0) != in2_early && !in2->in(0)->is_Root()) { 3694 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump(); 3695 } 3696 } 3697 } 3698 } 3699 } 3700 tty->cr(); 3701 tty->print("LCA(n): "); LCA->dump(); 3702 for (uint i = 0; i < n->outcnt(); i++) { 3703 Node* u1 = n->raw_out(i); 3704 if (u1 == n) 3705 continue; 3706 tty->print("n->out(%d): ", i); u1->dump(); 3707 if (u1->is_CFG()) { 3708 for (uint j = 0; j < u1->outcnt(); j++) { 3709 Node* u2 = u1->raw_out(j); 3710 if (u2 != u1 && u2 != n && u2->is_CFG()) { 3711 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 3712 } 3713 } 3714 } else { 3715 Node* u1_later = get_ctrl(u1); 3716 tty->print("later(n->out(%d)): ", i); u1_later->dump(); 3717 if (u1->in(0) != NULL && !u1->in(0)->is_top() && 3718 u1->in(0) != u1_later && !u1->in(0)->is_Root()) { 3719 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump(); 3720 } 3721 for (uint j = 0; j < u1->outcnt(); j++) { 3722 Node* u2 = u1->raw_out(j); 3723 if (u2 == n || u2 == u1) 3724 continue; 3725 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 3726 if (!u2->is_CFG()) { 3727 Node* u2_later = get_ctrl(u2); 3728 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); 3729 if (u2->in(0) != NULL && !u2->in(0)->is_top() && 3730 u2->in(0) != u2_later && !u2->in(0)->is_Root()) { 3731 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump(); 3732 } 3733 } 3734 } 3735 } 3736 } 3737 tty->cr(); 3738 int ct = 0; 3739 Node *dbg_legal = LCA; 3740 while(!dbg_legal->is_Start() && ct < 100) { 3741 tty->print("idom[%d] ",ct); dbg_legal->dump(); 3742 ct++; 3743 dbg_legal = idom(dbg_legal); 3744 } 3745 tty->cr(); 3746 } 3747 #endif 3748 3749 #ifndef PRODUCT 3750 //------------------------------dump------------------------------------------- 3751 void PhaseIdealLoop::dump( ) const { 3752 ResourceMark rm; 3753 Arena* arena = Thread::current()->resource_area(); 3754 Node_Stack stack(arena, C->live_nodes() >> 2); 3755 Node_List rpo_list; 3756 VectorSet visited(arena); 3757 visited.set(C->top()->_idx); 3758 rpo( C->root(), stack, visited, rpo_list ); 3759 // Dump root loop indexed by last element in PO order 3760 dump( _ltree_root, rpo_list.size(), rpo_list ); 3761 } 3762 3763 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { 3764 loop->dump_head(); 3765 3766 // Now scan for CFG nodes in the same loop 3767 for( uint j=idx; j > 0; j-- ) { 3768 Node *n = rpo_list[j-1]; 3769 if( !_nodes[n->_idx] ) // Skip dead nodes 3770 continue; 3771 if( get_loop(n) != loop ) { // Wrong loop nest 3772 if( get_loop(n)->_head == n && // Found nested loop? 3773 get_loop(n)->_parent == loop ) 3774 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly 3775 continue; 3776 } 3777 3778 // Dump controlling node 3779 for( uint x = 0; x < loop->_nest; x++ ) 3780 tty->print(" "); 3781 tty->print("C"); 3782 if( n == C->root() ) { 3783 n->dump(); 3784 } else { 3785 Node* cached_idom = idom_no_update(n); 3786 Node *computed_idom = n->in(0); 3787 if( n->is_Region() ) { 3788 computed_idom = compute_idom(n); 3789 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or 3790 // any MultiBranch ctrl node), so apply a similar transform to 3791 // the cached idom returned from idom_no_update. 3792 cached_idom = find_non_split_ctrl(cached_idom); 3793 } 3794 tty->print(" ID:%d",computed_idom->_idx); 3795 n->dump(); 3796 if( cached_idom != computed_idom ) { 3797 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", 3798 computed_idom->_idx, cached_idom->_idx); 3799 } 3800 } 3801 // Dump nodes it controls 3802 for( uint k = 0; k < _nodes.Size(); k++ ) { 3803 // (k < C->unique() && get_ctrl(find(k)) == n) 3804 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { 3805 Node *m = C->root()->find(k); 3806 if( m && m->outcnt() > 0 ) { 3807 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { 3808 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", 3809 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); 3810 } 3811 for( uint j = 0; j < loop->_nest; j++ ) 3812 tty->print(" "); 3813 tty->print(" "); 3814 m->dump(); 3815 } 3816 } 3817 } 3818 } 3819 } 3820 #endif 3821 3822 // Collect a R-P-O for the whole CFG. 3823 // Result list is in post-order (scan backwards for RPO) 3824 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { 3825 stk.push(start, 0); 3826 visited.set(start->_idx); 3827 3828 while (stk.is_nonempty()) { 3829 Node* m = stk.node(); 3830 uint idx = stk.index(); 3831 if (idx < m->outcnt()) { 3832 stk.set_index(idx + 1); 3833 Node* n = m->raw_out(idx); 3834 if (n->is_CFG() && !visited.test_set(n->_idx)) { 3835 stk.push(n, 0); 3836 } 3837 } else { 3838 rpo_list.push(m); 3839 stk.pop(); 3840 } 3841 } 3842 } 3843 3844 3845 //============================================================================= 3846 //------------------------------LoopTreeIterator----------------------------------- 3847 3848 // Advance to next loop tree using a preorder, left-to-right traversal. 3849 void LoopTreeIterator::next() { 3850 assert(!done(), "must not be done."); 3851 if (_curnt->_child != NULL) { 3852 _curnt = _curnt->_child; 3853 } else if (_curnt->_next != NULL) { 3854 _curnt = _curnt->_next; 3855 } else { 3856 while (_curnt != _root && _curnt->_next == NULL) { 3857 _curnt = _curnt->_parent; 3858 } 3859 if (_curnt == _root) { 3860 _curnt = NULL; 3861 assert(done(), "must be done."); 3862 } else { 3863 assert(_curnt->_next != NULL, "must be more to do"); 3864 _curnt = _curnt->_next; 3865 } 3866 } 3867 }