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