1 /* 2 * Copyright (c) 1998, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciMethodData.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "libadt/vectset.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/connode.hpp" 33 #include "opto/convertnode.hpp" 34 #include "opto/divnode.hpp" 35 #include "opto/idealGraphPrinter.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/mulnode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/superword.hpp" 40 41 //============================================================================= 42 //------------------------------is_loop_iv------------------------------------- 43 // Determine if a node is Counted loop induction variable. 44 // The method is declared in node.hpp. 45 const Node* Node::is_loop_iv() const { 46 if (this->is_Phi() && !this->as_Phi()->is_copy() && 47 this->as_Phi()->region()->is_CountedLoop() && 48 this->as_Phi()->region()->as_CountedLoop()->phi() == this) { 49 return this; 50 } else { 51 return NULL; 52 } 53 } 54 55 //============================================================================= 56 //------------------------------dump_spec-------------------------------------- 57 // Dump special per-node info 58 #ifndef PRODUCT 59 void LoopNode::dump_spec(outputStream *st) const { 60 if (is_inner_loop()) st->print( "inner " ); 61 if (is_partial_peel_loop()) st->print( "partial_peel " ); 62 if (partial_peel_has_failed()) st->print( "partial_peel_failed " ); 63 } 64 #endif 65 66 //------------------------------is_valid_counted_loop------------------------- 67 bool LoopNode::is_valid_counted_loop() const { 68 if (is_CountedLoop()) { 69 CountedLoopNode* l = as_CountedLoop(); 70 CountedLoopEndNode* le = l->loopexit(); 71 if (le != NULL && 72 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) { 73 Node* phi = l->phi(); 74 Node* exit = le->proj_out(0 /* false */); 75 if (exit != NULL && exit->Opcode() == Op_IfFalse && 76 phi != NULL && phi->is_Phi() && 77 phi->in(LoopNode::LoopBackControl) == l->incr() && 78 le->loopnode() == l && le->stride_is_con()) { 79 return true; 80 } 81 } 82 } 83 return false; 84 } 85 86 //------------------------------get_early_ctrl--------------------------------- 87 // Compute earliest legal control 88 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { 89 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); 90 uint i; 91 Node *early; 92 if (n->in(0) && !n->is_expensive()) { 93 early = n->in(0); 94 if (!early->is_CFG()) // Might be a non-CFG multi-def 95 early = get_ctrl(early); // So treat input as a straight data input 96 i = 1; 97 } else { 98 early = get_ctrl(n->in(1)); 99 i = 2; 100 } 101 uint e_d = dom_depth(early); 102 assert( early, "" ); 103 for (; i < n->req(); i++) { 104 Node *cin = get_ctrl(n->in(i)); 105 assert( cin, "" ); 106 // Keep deepest dominator depth 107 uint c_d = dom_depth(cin); 108 if (c_d > e_d) { // Deeper guy? 109 early = cin; // Keep deepest found so far 110 e_d = c_d; 111 } else if (c_d == e_d && // Same depth? 112 early != cin) { // If not equal, must use slower algorithm 113 // If same depth but not equal, one _must_ dominate the other 114 // and we want the deeper (i.e., dominated) guy. 115 Node *n1 = early; 116 Node *n2 = cin; 117 while (1) { 118 n1 = idom(n1); // Walk up until break cycle 119 n2 = idom(n2); 120 if (n1 == cin || // Walked early up to cin 121 dom_depth(n2) < c_d) 122 break; // early is deeper; keep him 123 if (n2 == early || // Walked cin up to early 124 dom_depth(n1) < c_d) { 125 early = cin; // cin is deeper; keep him 126 break; 127 } 128 } 129 e_d = dom_depth(early); // Reset depth register cache 130 } 131 } 132 133 // Return earliest legal location 134 assert(early == find_non_split_ctrl(early), "unexpected early control"); 135 136 if (n->is_expensive() && !_verify_only && !_verify_me) { 137 assert(n->in(0), "should have control input"); 138 early = get_early_ctrl_for_expensive(n, early); 139 } 140 141 return early; 142 } 143 144 //------------------------------get_early_ctrl_for_expensive--------------------------------- 145 // Move node up the dominator tree as high as legal while still beneficial 146 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) { 147 assert(n->in(0) && n->is_expensive(), "expensive node with control input here"); 148 assert(OptimizeExpensiveOps, "optimization off?"); 149 150 Node* ctl = n->in(0); 151 assert(ctl->is_CFG(), "expensive input 0 must be cfg"); 152 uint min_dom_depth = dom_depth(earliest); 153 #ifdef ASSERT 154 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) { 155 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl); 156 assert(false, "Bad graph detected in get_early_ctrl_for_expensive"); 157 } 158 #endif 159 if (dom_depth(ctl) < min_dom_depth) { 160 return earliest; 161 } 162 163 while (1) { 164 Node *next = ctl; 165 // Moving the node out of a loop on the projection of a If 166 // confuses loop predication. So once we hit a Loop in a If branch 167 // that doesn't branch to an UNC, we stop. The code that process 168 // expensive nodes will notice the loop and skip over it to try to 169 // move the node further up. 170 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) { 171 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 172 break; 173 } 174 next = idom(ctl->in(1)->in(0)); 175 } else if (ctl->is_Proj()) { 176 // We only move it up along a projection if the projection is 177 // the single control projection for its parent: same code path, 178 // if it's a If with UNC or fallthrough of a call. 179 Node* parent_ctl = ctl->in(0); 180 if (parent_ctl == NULL) { 181 break; 182 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) { 183 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control(); 184 } else if (parent_ctl->is_If()) { 185 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 186 break; 187 } 188 assert(idom(ctl) == parent_ctl, "strange"); 189 next = idom(parent_ctl); 190 } else if (ctl->is_CatchProj()) { 191 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) { 192 break; 193 } 194 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph"); 195 next = parent_ctl->in(0)->in(0)->in(0); 196 } else { 197 // Check if parent control has a single projection (this 198 // control is the only possible successor of the parent 199 // control). If so, we can try to move the node above the 200 // parent control. 201 int nb_ctl_proj = 0; 202 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) { 203 Node *p = parent_ctl->fast_out(i); 204 if (p->is_Proj() && p->is_CFG()) { 205 nb_ctl_proj++; 206 if (nb_ctl_proj > 1) { 207 break; 208 } 209 } 210 } 211 212 if (nb_ctl_proj > 1) { 213 break; 214 } 215 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node"); 216 assert(idom(ctl) == parent_ctl, "strange"); 217 next = idom(parent_ctl); 218 } 219 } else { 220 next = idom(ctl); 221 } 222 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) { 223 break; 224 } 225 ctl = next; 226 } 227 228 if (ctl != n->in(0)) { 229 _igvn.replace_input_of(n, 0, ctl); 230 _igvn.hash_insert(n); 231 } 232 233 return ctl; 234 } 235 236 237 //------------------------------set_early_ctrl--------------------------------- 238 // Set earliest legal control 239 void PhaseIdealLoop::set_early_ctrl( Node *n ) { 240 Node *early = get_early_ctrl(n); 241 242 // Record earliest legal location 243 set_ctrl(n, early); 244 } 245 246 //------------------------------set_subtree_ctrl------------------------------- 247 // set missing _ctrl entries on new nodes 248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { 249 // Already set? Get out. 250 if( _nodes[n->_idx] ) return; 251 // Recursively set _nodes array to indicate where the Node goes 252 uint i; 253 for( i = 0; i < n->req(); ++i ) { 254 Node *m = n->in(i); 255 if( m && m != C->root() ) 256 set_subtree_ctrl( m ); 257 } 258 259 // Fixup self 260 set_early_ctrl( n ); 261 } 262 263 //------------------------------is_counted_loop-------------------------------- 264 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) { 265 PhaseGVN *gvn = &_igvn; 266 267 // Counted loop head must be a good RegionNode with only 3 not NULL 268 // control input edges: Self, Entry, LoopBack. 269 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) { 270 return false; 271 } 272 Node *init_control = x->in(LoopNode::EntryControl); 273 Node *back_control = x->in(LoopNode::LoopBackControl); 274 if (init_control == NULL || back_control == NULL) // Partially dead 275 return false; 276 // Must also check for TOP when looking for a dead loop 277 if (init_control->is_top() || back_control->is_top()) 278 return false; 279 280 // Allow funny placement of Safepoint 281 if (back_control->Opcode() == Op_SafePoint) 282 back_control = back_control->in(TypeFunc::Control); 283 284 // Controlling test for loop 285 Node *iftrue = back_control; 286 uint iftrue_op = iftrue->Opcode(); 287 if (iftrue_op != Op_IfTrue && 288 iftrue_op != Op_IfFalse) 289 // I have a weird back-control. Probably the loop-exit test is in 290 // the middle of the loop and I am looking at some trailing control-flow 291 // merge point. To fix this I would have to partially peel the loop. 292 return false; // Obscure back-control 293 294 // Get boolean guarding loop-back test 295 Node *iff = iftrue->in(0); 296 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) 297 return false; 298 BoolNode *test = iff->in(1)->as_Bool(); 299 BoolTest::mask bt = test->_test._test; 300 float cl_prob = iff->as_If()->_prob; 301 if (iftrue_op == Op_IfFalse) { 302 bt = BoolTest(bt).negate(); 303 cl_prob = 1.0 - cl_prob; 304 } 305 // Get backedge compare 306 Node *cmp = test->in(1); 307 int cmp_op = cmp->Opcode(); 308 if (cmp_op != Op_CmpI) 309 return false; // Avoid pointer & float compares 310 311 // Find the trip-counter increment & limit. Limit must be loop invariant. 312 Node *incr = cmp->in(1); 313 Node *limit = cmp->in(2); 314 315 // --------- 316 // need 'loop()' test to tell if limit is loop invariant 317 // --------- 318 319 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit? 320 Node *tmp = incr; // Then reverse order into the CmpI 321 incr = limit; 322 limit = tmp; 323 bt = BoolTest(bt).commute(); // And commute the exit test 324 } 325 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant 326 return false; 327 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 328 return false; 329 330 Node* phi_incr = NULL; 331 // Trip-counter increment must be commutative & associative. 332 if (incr->Opcode() == Op_CastII) { 333 incr = incr->in(1); 334 } 335 if (incr->is_Phi()) { 336 if (incr->as_Phi()->region() != x || incr->req() != 3) 337 return false; // Not simple trip counter expression 338 phi_incr = incr; 339 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi 340 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 341 return false; 342 } 343 344 Node* trunc1 = NULL; 345 Node* trunc2 = NULL; 346 const TypeInt* iv_trunc_t = NULL; 347 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { 348 return false; // Funny increment opcode 349 } 350 assert(incr->Opcode() == Op_AddI, "wrong increment code"); 351 352 // Get merge point 353 Node *xphi = incr->in(1); 354 Node *stride = incr->in(2); 355 if (!stride->is_Con()) { // Oops, swap these 356 if (!xphi->is_Con()) // Is the other guy a constant? 357 return false; // Nope, unknown stride, bail out 358 Node *tmp = xphi; // 'incr' is commutative, so ok to swap 359 xphi = stride; 360 stride = tmp; 361 } 362 if (xphi->Opcode() == Op_CastII) { 363 xphi = xphi->in(1); 364 } 365 // Stride must be constant 366 int stride_con = stride->get_int(); 367 if (stride_con == 0) 368 return false; // missed some peephole opt 369 370 if (!xphi->is_Phi()) 371 return false; // Too much math on the trip counter 372 if (phi_incr != NULL && phi_incr != xphi) 373 return false; 374 PhiNode *phi = xphi->as_Phi(); 375 376 // Phi must be of loop header; backedge must wrap to increment 377 if (phi->region() != x) 378 return false; 379 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr || 380 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) { 381 return false; 382 } 383 Node *init_trip = phi->in(LoopNode::EntryControl); 384 385 // If iv trunc type is smaller than int, check for possible wrap. 386 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { 387 assert(trunc1 != NULL, "must have found some truncation"); 388 389 // Get a better type for the phi (filtered thru if's) 390 const TypeInt* phi_ft = filtered_type(phi); 391 392 // Can iv take on a value that will wrap? 393 // 394 // Ensure iv's limit is not within "stride" of the wrap value. 395 // 396 // Example for "short" type 397 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) 398 // If the stride is +10, then the last value of the induction 399 // variable before the increment (phi_ft->_hi) must be 400 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to 401 // ensure no truncation occurs after the increment. 402 403 if (stride_con > 0) { 404 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || 405 iv_trunc_t->_lo > phi_ft->_lo) { 406 return false; // truncation may occur 407 } 408 } else if (stride_con < 0) { 409 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || 410 iv_trunc_t->_hi < phi_ft->_hi) { 411 return false; // truncation may occur 412 } 413 } 414 // No possibility of wrap so truncation can be discarded 415 // Promote iv type to Int 416 } else { 417 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); 418 } 419 420 // If the condition is inverted and we will be rolling 421 // through MININT to MAXINT, then bail out. 422 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice! 423 // Odd stride 424 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 || 425 // Count down loop rolls through MAXINT 426 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 || 427 // Count up loop rolls through MININT 428 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) { 429 return false; // Bail out 430 } 431 432 const TypeInt* init_t = gvn->type(init_trip)->is_int(); 433 const TypeInt* limit_t = gvn->type(limit)->is_int(); 434 435 if (stride_con > 0) { 436 jlong init_p = (jlong)init_t->_lo + stride_con; 437 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi) 438 return false; // cyclic loop or this loop trips only once 439 } else { 440 jlong init_p = (jlong)init_t->_hi + stride_con; 441 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo) 442 return false; // cyclic loop or this loop trips only once 443 } 444 445 if (phi_incr != NULL) { 446 // check if there is a possiblity of IV overflowing after the first increment 447 if (stride_con > 0) { 448 if (init_t->_hi > max_jint - stride_con) { 449 return false; 450 } 451 } else { 452 if (init_t->_lo < min_jint - stride_con) { 453 return false; 454 } 455 } 456 } 457 458 // ================================================= 459 // ---- SUCCESS! Found A Trip-Counted Loop! ----- 460 // 461 assert(x->Opcode() == Op_Loop, "regular loops only"); 462 C->print_method(PHASE_BEFORE_CLOOPS, 3); 463 464 Node *hook = new Node(6); 465 466 if (LoopLimitCheck) { 467 468 // =================================================== 469 // Generate loop limit check to avoid integer overflow 470 // in cases like next (cyclic loops): 471 // 472 // for (i=0; i <= max_jint; i++) {} 473 // for (i=0; i < max_jint; i+=2) {} 474 // 475 // 476 // Limit check predicate depends on the loop test: 477 // 478 // for(;i != limit; i++) --> limit <= (max_jint) 479 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1) 480 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride ) 481 // 482 483 // Check if limit is excluded to do more precise int overflow check. 484 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge); 485 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1)); 486 487 // If compare points directly to the phi we need to adjust 488 // the compare so that it points to the incr. Limit have 489 // to be adjusted to keep trip count the same and the 490 // adjusted limit should be checked for int overflow. 491 if (phi_incr != NULL) { 492 stride_m += stride_con; 493 } 494 495 if (limit->is_Con()) { 496 int limit_con = limit->get_int(); 497 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) || 498 (stride_con < 0 && limit_con < (min_jint - stride_m))) { 499 // Bailout: it could be integer overflow. 500 return false; 501 } 502 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) || 503 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) { 504 // Limit's type may satisfy the condition, for example, 505 // when it is an array length. 506 } else { 507 // Generate loop's limit check. 508 // Loop limit check predicate should be near the loop. 509 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); 510 if (!limit_check_proj) { 511 // The limit check predicate is not generated if this method trapped here before. 512 #ifdef ASSERT 513 if (TraceLoopLimitCheck) { 514 tty->print("missing loop limit check:"); 515 loop->dump_head(); 516 x->dump(1); 517 } 518 #endif 519 return false; 520 } 521 522 IfNode* check_iff = limit_check_proj->in(0)->as_If(); 523 Node* cmp_limit; 524 Node* bol; 525 526 if (stride_con > 0) { 527 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m)); 528 bol = new BoolNode(cmp_limit, BoolTest::le); 529 } else { 530 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m)); 531 bol = new BoolNode(cmp_limit, BoolTest::ge); 532 } 533 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit); 534 bol = _igvn.register_new_node_with_optimizer(bol); 535 set_subtree_ctrl(bol); 536 537 // Replace condition in original predicate but preserve Opaque node 538 // so that previous predicates could be found. 539 assert(check_iff->in(1)->Opcode() == Op_Conv2B && 540 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, ""); 541 Node* opq = check_iff->in(1)->in(1); 542 _igvn.replace_input_of(opq, 1, bol); 543 // Update ctrl. 544 set_ctrl(opq, check_iff->in(0)); 545 set_ctrl(check_iff->in(1), check_iff->in(0)); 546 547 #ifndef PRODUCT 548 // report that the loop predication has been actually performed 549 // for this loop 550 if (TraceLoopLimitCheck) { 551 tty->print_cr("Counted Loop Limit Check generated:"); 552 debug_only( bol->dump(2); ) 553 } 554 #endif 555 } 556 557 if (phi_incr != NULL) { 558 // If compare points directly to the phi we need to adjust 559 // the compare so that it points to the incr. Limit have 560 // to be adjusted to keep trip count the same and we 561 // should avoid int overflow. 562 // 563 // i = init; do {} while(i++ < limit); 564 // is converted to 565 // i = init; do {} while(++i < limit+1); 566 // 567 limit = gvn->transform(new AddINode(limit, stride)); 568 } 569 570 // Now we need to canonicalize loop condition. 571 if (bt == BoolTest::ne) { 572 assert(stride_con == 1 || stride_con == -1, "simple increment only"); 573 // 'ne' can be replaced with 'lt' only when init < limit. 574 if (stride_con > 0 && init_t->_hi < limit_t->_lo) 575 bt = BoolTest::lt; 576 // 'ne' can be replaced with 'gt' only when init > limit. 577 if (stride_con < 0 && init_t->_lo > limit_t->_hi) 578 bt = BoolTest::gt; 579 } 580 581 if (incl_limit) { 582 // The limit check guaranties that 'limit <= (max_jint - stride)' so 583 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0. 584 // 585 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1); 586 limit = gvn->transform(new AddINode(limit, one)); 587 if (bt == BoolTest::le) 588 bt = BoolTest::lt; 589 else if (bt == BoolTest::ge) 590 bt = BoolTest::gt; 591 else 592 ShouldNotReachHere(); 593 } 594 set_subtree_ctrl( limit ); 595 596 } else { // LoopLimitCheck 597 598 // If compare points to incr, we are ok. Otherwise the compare 599 // can directly point to the phi; in this case adjust the compare so that 600 // it points to the incr by adjusting the limit. 601 if (cmp->in(1) == phi || cmp->in(2) == phi) 602 limit = gvn->transform(new AddINode(limit,stride)); 603 604 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride. 605 // Final value for iterator should be: trip_count * stride + init_trip. 606 Node *one_p = gvn->intcon( 1); 607 Node *one_m = gvn->intcon(-1); 608 609 Node *trip_count = NULL; 610 switch( bt ) { 611 case BoolTest::eq: 612 ShouldNotReachHere(); 613 case BoolTest::ne: // Ahh, the case we desire 614 if (stride_con == 1) 615 trip_count = gvn->transform(new SubINode(limit,init_trip)); 616 else if (stride_con == -1) 617 trip_count = gvn->transform(new SubINode(init_trip,limit)); 618 else 619 ShouldNotReachHere(); 620 set_subtree_ctrl(trip_count); 621 //_loop.map(trip_count->_idx,loop(limit)); 622 break; 623 case BoolTest::le: // Maybe convert to '<' case 624 limit = gvn->transform(new AddINode(limit,one_p)); 625 set_subtree_ctrl( limit ); 626 hook->init_req(4, limit); 627 628 bt = BoolTest::lt; 629 // Make the new limit be in the same loop nest as the old limit 630 //_loop.map(limit->_idx,limit_loop); 631 // Fall into next case 632 case BoolTest::lt: { // Maybe convert to '!=' case 633 if (stride_con < 0) // Count down loop rolls through MAXINT 634 ShouldNotReachHere(); 635 Node *range = gvn->transform(new SubINode(limit,init_trip)); 636 set_subtree_ctrl( range ); 637 hook->init_req(0, range); 638 639 Node *bias = gvn->transform(new AddINode(range,stride)); 640 set_subtree_ctrl( bias ); 641 hook->init_req(1, bias); 642 643 Node *bias1 = gvn->transform(new AddINode(bias,one_m)); 644 set_subtree_ctrl( bias1 ); 645 hook->init_req(2, bias1); 646 647 trip_count = gvn->transform(new DivINode(0,bias1,stride)); 648 set_subtree_ctrl( trip_count ); 649 hook->init_req(3, trip_count); 650 break; 651 } 652 653 case BoolTest::ge: // Maybe convert to '>' case 654 limit = gvn->transform(new AddINode(limit,one_m)); 655 set_subtree_ctrl( limit ); 656 hook->init_req(4 ,limit); 657 658 bt = BoolTest::gt; 659 // Make the new limit be in the same loop nest as the old limit 660 //_loop.map(limit->_idx,limit_loop); 661 // Fall into next case 662 case BoolTest::gt: { // Maybe convert to '!=' case 663 if (stride_con > 0) // count up loop rolls through MININT 664 ShouldNotReachHere(); 665 Node *range = gvn->transform(new SubINode(limit,init_trip)); 666 set_subtree_ctrl( range ); 667 hook->init_req(0, range); 668 669 Node *bias = gvn->transform(new AddINode(range,stride)); 670 set_subtree_ctrl( bias ); 671 hook->init_req(1, bias); 672 673 Node *bias1 = gvn->transform(new AddINode(bias,one_p)); 674 set_subtree_ctrl( bias1 ); 675 hook->init_req(2, bias1); 676 677 trip_count = gvn->transform(new DivINode(0,bias1,stride)); 678 set_subtree_ctrl( trip_count ); 679 hook->init_req(3, trip_count); 680 break; 681 } 682 } // switch( bt ) 683 684 Node *span = gvn->transform(new MulINode(trip_count,stride)); 685 set_subtree_ctrl( span ); 686 hook->init_req(5, span); 687 688 limit = gvn->transform(new AddINode(span,init_trip)); 689 set_subtree_ctrl( limit ); 690 691 } // LoopLimitCheck 692 693 if (!UseCountedLoopSafepoints) { 694 // Check for SafePoint on backedge and remove 695 Node *sfpt = x->in(LoopNode::LoopBackControl); 696 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { 697 lazy_replace( sfpt, iftrue ); 698 if (loop->_safepts != NULL) { 699 loop->_safepts->yank(sfpt); 700 } 701 loop->_tail = iftrue; 702 } 703 } 704 705 // Build a canonical trip test. 706 // Clone code, as old values may be in use. 707 incr = incr->clone(); 708 incr->set_req(1,phi); 709 incr->set_req(2,stride); 710 incr = _igvn.register_new_node_with_optimizer(incr); 711 set_early_ctrl( incr ); 712 _igvn.rehash_node_delayed(phi); 713 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); 714 715 // If phi type is more restrictive than Int, raise to 716 // Int to prevent (almost) infinite recursion in igvn 717 // which can only handle integer types for constants or minint..maxint. 718 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { 719 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); 720 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); 721 nphi = _igvn.register_new_node_with_optimizer(nphi); 722 set_ctrl(nphi, get_ctrl(phi)); 723 _igvn.replace_node(phi, nphi); 724 phi = nphi->as_Phi(); 725 } 726 cmp = cmp->clone(); 727 cmp->set_req(1,incr); 728 cmp->set_req(2,limit); 729 cmp = _igvn.register_new_node_with_optimizer(cmp); 730 set_ctrl(cmp, iff->in(0)); 731 732 test = test->clone()->as_Bool(); 733 (*(BoolTest*)&test->_test)._test = bt; 734 test->set_req(1,cmp); 735 _igvn.register_new_node_with_optimizer(test); 736 set_ctrl(test, iff->in(0)); 737 738 // Replace the old IfNode with a new LoopEndNode 739 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt )); 740 IfNode *le = lex->as_If(); 741 uint dd = dom_depth(iff); 742 set_idom(le, le->in(0), dd); // Update dominance for loop exit 743 set_loop(le, loop); 744 745 // Get the loop-exit control 746 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); 747 748 // Need to swap loop-exit and loop-back control? 749 if (iftrue_op == Op_IfFalse) { 750 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le)); 751 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le)); 752 753 loop->_tail = back_control = ift2; 754 set_loop(ift2, loop); 755 set_loop(iff2, get_loop(iffalse)); 756 757 // Lazy update of 'get_ctrl' mechanism. 758 lazy_replace_proj( iffalse, iff2 ); 759 lazy_replace_proj( iftrue, ift2 ); 760 761 // Swap names 762 iffalse = iff2; 763 iftrue = ift2; 764 } else { 765 _igvn.rehash_node_delayed(iffalse); 766 _igvn.rehash_node_delayed(iftrue); 767 iffalse->set_req_X( 0, le, &_igvn ); 768 iftrue ->set_req_X( 0, le, &_igvn ); 769 } 770 771 set_idom(iftrue, le, dd+1); 772 set_idom(iffalse, le, dd+1); 773 assert(iff->outcnt() == 0, "should be dead now"); 774 lazy_replace( iff, le ); // fix 'get_ctrl' 775 776 // Now setup a new CountedLoopNode to replace the existing LoopNode 777 CountedLoopNode *l = new CountedLoopNode(init_control, back_control); 778 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve 779 // The following assert is approximately true, and defines the intention 780 // of can_be_counted_loop. It fails, however, because phase->type 781 // is not yet initialized for this loop and its parts. 782 //assert(l->can_be_counted_loop(this), "sanity"); 783 _igvn.register_new_node_with_optimizer(l); 784 set_loop(l, loop); 785 loop->_head = l; 786 // Fix all data nodes placed at the old loop head. 787 // Uses the lazy-update mechanism of 'get_ctrl'. 788 lazy_replace( x, l ); 789 set_idom(l, init_control, dom_depth(x)); 790 791 if (!UseCountedLoopSafepoints) { 792 // Check for immediately preceding SafePoint and remove 793 Node *sfpt2 = le->in(0); 794 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) { 795 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); 796 if (loop->_safepts != NULL) { 797 loop->_safepts->yank(sfpt2); 798 } 799 } 800 } 801 802 // Free up intermediate goo 803 _igvn.remove_dead_node(hook); 804 805 #ifdef ASSERT 806 assert(l->is_valid_counted_loop(), "counted loop shape is messed up"); 807 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" ); 808 #endif 809 #ifndef PRODUCT 810 if (TraceLoopOpts) { 811 tty->print("Counted "); 812 loop->dump_head(); 813 } 814 #endif 815 816 C->print_method(PHASE_AFTER_CLOOPS, 3); 817 818 // Capture bounds of the loop in the induction variable Phi before 819 // subsequent transformation (iteration splitting) obscures the 820 // bounds 821 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn)); 822 823 return true; 824 } 825 826 //----------------------exact_limit------------------------------------------- 827 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) { 828 assert(loop->_head->is_CountedLoop(), ""); 829 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 830 assert(cl->is_valid_counted_loop(), ""); 831 832 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 || 833 cl->limit()->Opcode() == Op_LoopLimit) { 834 // Old code has exact limit (it could be incorrect in case of int overflow). 835 // Loop limit is exact with stride == 1. And loop may already have exact limit. 836 return cl->limit(); 837 } 838 Node *limit = NULL; 839 #ifdef ASSERT 840 BoolTest::mask bt = cl->loopexit()->test_trip(); 841 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 842 #endif 843 if (cl->has_exact_trip_count()) { 844 // Simple case: loop has constant boundaries. 845 // Use jlongs to avoid integer overflow. 846 int stride_con = cl->stride_con(); 847 jlong init_con = cl->init_trip()->get_int(); 848 jlong limit_con = cl->limit()->get_int(); 849 julong trip_cnt = cl->trip_count(); 850 jlong final_con = init_con + trip_cnt*stride_con; 851 int final_int = (int)final_con; 852 // The final value should be in integer range since the loop 853 // is counted and the limit was checked for overflow. 854 assert(final_con == (jlong)final_int, "final value should be integer"); 855 limit = _igvn.intcon(final_int); 856 } else { 857 // Create new LoopLimit node to get exact limit (final iv value). 858 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride()); 859 register_new_node(limit, cl->in(LoopNode::EntryControl)); 860 } 861 assert(limit != NULL, "sanity"); 862 return limit; 863 } 864 865 //------------------------------Ideal------------------------------------------ 866 // Return a node which is more "ideal" than the current node. 867 // Attempt to convert into a counted-loop. 868 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 869 if (!can_be_counted_loop(phase)) { 870 phase->C->set_major_progress(); 871 } 872 return RegionNode::Ideal(phase, can_reshape); 873 } 874 875 876 //============================================================================= 877 //------------------------------Ideal------------------------------------------ 878 // Return a node which is more "ideal" than the current node. 879 // Attempt to convert into a counted-loop. 880 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 881 return RegionNode::Ideal(phase, can_reshape); 882 } 883 884 //------------------------------dump_spec-------------------------------------- 885 // Dump special per-node info 886 #ifndef PRODUCT 887 void CountedLoopNode::dump_spec(outputStream *st) const { 888 LoopNode::dump_spec(st); 889 if (stride_is_con()) { 890 st->print("stride: %d ",stride_con()); 891 } 892 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx); 893 if (is_main_loop()) st->print("main of N%d", _idx); 894 if (is_post_loop()) st->print("post of N%d", _main_idx); 895 } 896 #endif 897 898 //============================================================================= 899 int CountedLoopEndNode::stride_con() const { 900 return stride()->bottom_type()->is_int()->get_con(); 901 } 902 903 //============================================================================= 904 //------------------------------Value----------------------------------------- 905 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const { 906 const Type* init_t = phase->type(in(Init)); 907 const Type* limit_t = phase->type(in(Limit)); 908 const Type* stride_t = phase->type(in(Stride)); 909 // Either input is TOP ==> the result is TOP 910 if (init_t == Type::TOP) return Type::TOP; 911 if (limit_t == Type::TOP) return Type::TOP; 912 if (stride_t == Type::TOP) return Type::TOP; 913 914 int stride_con = stride_t->is_int()->get_con(); 915 if (stride_con == 1) 916 return NULL; // Identity 917 918 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) { 919 // Use jlongs to avoid integer overflow. 920 jlong init_con = init_t->is_int()->get_con(); 921 jlong limit_con = limit_t->is_int()->get_con(); 922 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 923 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 924 jlong final_con = init_con + stride_con*trip_count; 925 int final_int = (int)final_con; 926 // The final value should be in integer range since the loop 927 // is counted and the limit was checked for overflow. 928 assert(final_con == (jlong)final_int, "final value should be integer"); 929 return TypeInt::make(final_int); 930 } 931 932 return bottom_type(); // TypeInt::INT 933 } 934 935 //------------------------------Ideal------------------------------------------ 936 // Return a node which is more "ideal" than the current node. 937 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) { 938 if (phase->type(in(Init)) == Type::TOP || 939 phase->type(in(Limit)) == Type::TOP || 940 phase->type(in(Stride)) == Type::TOP) 941 return NULL; // Dead 942 943 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 944 if (stride_con == 1) 945 return NULL; // Identity 946 947 if (in(Init)->is_Con() && in(Limit)->is_Con()) 948 return NULL; // Value 949 950 // Delay following optimizations until all loop optimizations 951 // done to keep Ideal graph simple. 952 if (!can_reshape || phase->C->major_progress()) 953 return NULL; 954 955 const TypeInt* init_t = phase->type(in(Init) )->is_int(); 956 const TypeInt* limit_t = phase->type(in(Limit))->is_int(); 957 int stride_p; 958 jlong lim, ini; 959 julong max; 960 if (stride_con > 0) { 961 stride_p = stride_con; 962 lim = limit_t->_hi; 963 ini = init_t->_lo; 964 max = (julong)max_jint; 965 } else { 966 stride_p = -stride_con; 967 lim = init_t->_hi; 968 ini = limit_t->_lo; 969 max = (julong)min_jint; 970 } 971 julong range = lim - ini + stride_p; 972 if (range <= max) { 973 // Convert to integer expression if it is not overflow. 974 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1)); 975 Node *range = phase->transform(new SubINode(in(Limit), in(Init))); 976 Node *bias = phase->transform(new AddINode(range, stride_m)); 977 Node *trip = phase->transform(new DivINode(0, bias, in(Stride))); 978 Node *span = phase->transform(new MulINode(trip, in(Stride))); 979 return new AddINode(span, in(Init)); // exact limit 980 } 981 982 if (is_power_of_2(stride_p) || // divisor is 2^n 983 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node? 984 // Convert to long expression to avoid integer overflow 985 // and let igvn optimizer convert this division. 986 // 987 Node* init = phase->transform( new ConvI2LNode(in(Init))); 988 Node* limit = phase->transform( new ConvI2LNode(in(Limit))); 989 Node* stride = phase->longcon(stride_con); 990 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1)); 991 992 Node *range = phase->transform(new SubLNode(limit, init)); 993 Node *bias = phase->transform(new AddLNode(range, stride_m)); 994 Node *span; 995 if (stride_con > 0 && is_power_of_2(stride_p)) { 996 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride) 997 // and avoid generating rounding for division. Zero trip guard should 998 // guarantee that init < limit but sometimes the guard is missing and 999 // we can get situation when init > limit. Note, for the empty loop 1000 // optimization zero trip guard is generated explicitly which leaves 1001 // only RCE predicate where exact limit is used and the predicate 1002 // will simply fail forcing recompilation. 1003 Node* neg_stride = phase->longcon(-stride_con); 1004 span = phase->transform(new AndLNode(bias, neg_stride)); 1005 } else { 1006 Node *trip = phase->transform(new DivLNode(0, bias, stride)); 1007 span = phase->transform(new MulLNode(trip, stride)); 1008 } 1009 // Convert back to int 1010 Node *span_int = phase->transform(new ConvL2INode(span)); 1011 return new AddINode(span_int, in(Init)); // exact limit 1012 } 1013 1014 return NULL; // No progress 1015 } 1016 1017 //------------------------------Identity--------------------------------------- 1018 // If stride == 1 return limit node. 1019 Node *LoopLimitNode::Identity( PhaseTransform *phase ) { 1020 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1021 if (stride_con == 1 || stride_con == -1) 1022 return in(Limit); 1023 return this; 1024 } 1025 1026 //============================================================================= 1027 //----------------------match_incr_with_optional_truncation-------------------- 1028 // Match increment with optional truncation: 1029 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 1030 // Return NULL for failure. Success returns the increment node. 1031 Node* CountedLoopNode::match_incr_with_optional_truncation( 1032 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { 1033 // Quick cutouts: 1034 if (expr == NULL || expr->req() != 3) return NULL; 1035 1036 Node *t1 = NULL; 1037 Node *t2 = NULL; 1038 const TypeInt* trunc_t = TypeInt::INT; 1039 Node* n1 = expr; 1040 int n1op = n1->Opcode(); 1041 1042 // Try to strip (n1 & M) or (n1 << N >> N) from n1. 1043 if (n1op == Op_AndI && 1044 n1->in(2)->is_Con() && 1045 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { 1046 // %%% This check should match any mask of 2**K-1. 1047 t1 = n1; 1048 n1 = t1->in(1); 1049 n1op = n1->Opcode(); 1050 trunc_t = TypeInt::CHAR; 1051 } else if (n1op == Op_RShiftI && 1052 n1->in(1) != NULL && 1053 n1->in(1)->Opcode() == Op_LShiftI && 1054 n1->in(2) == n1->in(1)->in(2) && 1055 n1->in(2)->is_Con()) { 1056 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); 1057 // %%% This check should match any shift in [1..31]. 1058 if (shift == 16 || shift == 8) { 1059 t1 = n1; 1060 t2 = t1->in(1); 1061 n1 = t2->in(1); 1062 n1op = n1->Opcode(); 1063 if (shift == 16) { 1064 trunc_t = TypeInt::SHORT; 1065 } else if (shift == 8) { 1066 trunc_t = TypeInt::BYTE; 1067 } 1068 } 1069 } 1070 1071 // If (maybe after stripping) it is an AddI, we won: 1072 if (n1op == Op_AddI) { 1073 *trunc1 = t1; 1074 *trunc2 = t2; 1075 *trunc_type = trunc_t; 1076 return n1; 1077 } 1078 1079 // failed 1080 return NULL; 1081 } 1082 1083 1084 //------------------------------filtered_type-------------------------------- 1085 // Return a type based on condition control flow 1086 // A successful return will be a type that is restricted due 1087 // to a series of dominating if-tests, such as: 1088 // if (i < 10) { 1089 // if (i > 0) { 1090 // here: "i" type is [1..10) 1091 // } 1092 // } 1093 // or a control flow merge 1094 // if (i < 10) { 1095 // do { 1096 // phi( , ) -- at top of loop type is [min_int..10) 1097 // i = ? 1098 // } while ( i < 10) 1099 // 1100 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { 1101 assert(n && n->bottom_type()->is_int(), "must be int"); 1102 const TypeInt* filtered_t = NULL; 1103 if (!n->is_Phi()) { 1104 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); 1105 filtered_t = filtered_type_from_dominators(n, n_ctrl); 1106 1107 } else { 1108 Node* phi = n->as_Phi(); 1109 Node* region = phi->in(0); 1110 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); 1111 if (region && region != C->top()) { 1112 for (uint i = 1; i < phi->req(); i++) { 1113 Node* val = phi->in(i); 1114 Node* use_c = region->in(i); 1115 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); 1116 if (val_t != NULL) { 1117 if (filtered_t == NULL) { 1118 filtered_t = val_t; 1119 } else { 1120 filtered_t = filtered_t->meet(val_t)->is_int(); 1121 } 1122 } 1123 } 1124 } 1125 } 1126 const TypeInt* n_t = _igvn.type(n)->is_int(); 1127 if (filtered_t != NULL) { 1128 n_t = n_t->join(filtered_t)->is_int(); 1129 } 1130 return n_t; 1131 } 1132 1133 1134 //------------------------------filtered_type_from_dominators-------------------------------- 1135 // Return a possibly more restrictive type for val based on condition control flow of dominators 1136 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { 1137 if (val->is_Con()) { 1138 return val->bottom_type()->is_int(); 1139 } 1140 uint if_limit = 10; // Max number of dominating if's visited 1141 const TypeInt* rtn_t = NULL; 1142 1143 if (use_ctrl && use_ctrl != C->top()) { 1144 Node* val_ctrl = get_ctrl(val); 1145 uint val_dom_depth = dom_depth(val_ctrl); 1146 Node* pred = use_ctrl; 1147 uint if_cnt = 0; 1148 while (if_cnt < if_limit) { 1149 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { 1150 if_cnt++; 1151 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); 1152 if (if_t != NULL) { 1153 if (rtn_t == NULL) { 1154 rtn_t = if_t; 1155 } else { 1156 rtn_t = rtn_t->join(if_t)->is_int(); 1157 } 1158 } 1159 } 1160 pred = idom(pred); 1161 if (pred == NULL || pred == C->top()) { 1162 break; 1163 } 1164 // Stop if going beyond definition block of val 1165 if (dom_depth(pred) < val_dom_depth) { 1166 break; 1167 } 1168 } 1169 } 1170 return rtn_t; 1171 } 1172 1173 1174 //------------------------------dump_spec-------------------------------------- 1175 // Dump special per-node info 1176 #ifndef PRODUCT 1177 void CountedLoopEndNode::dump_spec(outputStream *st) const { 1178 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) { 1179 BoolTest bt( test_trip()); // Added this for g++. 1180 1181 st->print("["); 1182 bt.dump_on(st); 1183 st->print("]"); 1184 } 1185 st->print(" "); 1186 IfNode::dump_spec(st); 1187 } 1188 #endif 1189 1190 //============================================================================= 1191 //------------------------------is_member-------------------------------------- 1192 // Is 'l' a member of 'this'? 1193 bool IdealLoopTree::is_member(const IdealLoopTree *l) const { 1194 while( l->_nest > _nest ) l = l->_parent; 1195 return l == this; 1196 } 1197 1198 //------------------------------set_nest--------------------------------------- 1199 // Set loop tree nesting depth. Accumulate _has_call bits. 1200 int IdealLoopTree::set_nest( uint depth ) { 1201 _nest = depth; 1202 int bits = _has_call; 1203 if( _child ) bits |= _child->set_nest(depth+1); 1204 if( bits ) _has_call = 1; 1205 if( _next ) bits |= _next ->set_nest(depth ); 1206 return bits; 1207 } 1208 1209 //------------------------------split_fall_in---------------------------------- 1210 // Split out multiple fall-in edges from the loop header. Move them to a 1211 // private RegionNode before the loop. This becomes the loop landing pad. 1212 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { 1213 PhaseIterGVN &igvn = phase->_igvn; 1214 uint i; 1215 1216 // Make a new RegionNode to be the landing pad. 1217 Node *landing_pad = new RegionNode( fall_in_cnt+1 ); 1218 phase->set_loop(landing_pad,_parent); 1219 // Gather all the fall-in control paths into the landing pad 1220 uint icnt = fall_in_cnt; 1221 uint oreq = _head->req(); 1222 for( i = oreq-1; i>0; i-- ) 1223 if( !phase->is_member( this, _head->in(i) ) ) 1224 landing_pad->set_req(icnt--,_head->in(i)); 1225 1226 // Peel off PhiNode edges as well 1227 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1228 Node *oj = _head->fast_out(j); 1229 if( oj->is_Phi() ) { 1230 PhiNode* old_phi = oj->as_Phi(); 1231 assert( old_phi->region() == _head, "" ); 1232 igvn.hash_delete(old_phi); // Yank from hash before hacking edges 1233 Node *p = PhiNode::make_blank(landing_pad, old_phi); 1234 uint icnt = fall_in_cnt; 1235 for( i = oreq-1; i>0; i-- ) { 1236 if( !phase->is_member( this, _head->in(i) ) ) { 1237 p->init_req(icnt--, old_phi->in(i)); 1238 // Go ahead and clean out old edges from old phi 1239 old_phi->del_req(i); 1240 } 1241 } 1242 // Search for CSE's here, because ZKM.jar does a lot of 1243 // loop hackery and we need to be a little incremental 1244 // with the CSE to avoid O(N^2) node blow-up. 1245 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE 1246 if( p2 ) { // Found CSE 1247 p->destruct(); // Recover useless new node 1248 p = p2; // Use old node 1249 } else { 1250 igvn.register_new_node_with_optimizer(p, old_phi); 1251 } 1252 // Make old Phi refer to new Phi. 1253 old_phi->add_req(p); 1254 // Check for the special case of making the old phi useless and 1255 // disappear it. In JavaGrande I have a case where this useless 1256 // Phi is the loop limit and prevents recognizing a CountedLoop 1257 // which in turn prevents removing an empty loop. 1258 Node *id_old_phi = old_phi->Identity( &igvn ); 1259 if( id_old_phi != old_phi ) { // Found a simple identity? 1260 // Note that I cannot call 'replace_node' here, because 1261 // that will yank the edge from old_phi to the Region and 1262 // I'm mid-iteration over the Region's uses. 1263 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { 1264 Node* use = old_phi->last_out(i); 1265 igvn.rehash_node_delayed(use); 1266 uint uses_found = 0; 1267 for (uint j = 0; j < use->len(); j++) { 1268 if (use->in(j) == old_phi) { 1269 if (j < use->req()) use->set_req (j, id_old_phi); 1270 else use->set_prec(j, id_old_phi); 1271 uses_found++; 1272 } 1273 } 1274 i -= uses_found; // we deleted 1 or more copies of this edge 1275 } 1276 } 1277 igvn._worklist.push(old_phi); 1278 } 1279 } 1280 // Finally clean out the fall-in edges from the RegionNode 1281 for( i = oreq-1; i>0; i-- ) { 1282 if( !phase->is_member( this, _head->in(i) ) ) { 1283 _head->del_req(i); 1284 } 1285 } 1286 igvn.rehash_node_delayed(_head); 1287 // Transform landing pad 1288 igvn.register_new_node_with_optimizer(landing_pad, _head); 1289 // Insert landing pad into the header 1290 _head->add_req(landing_pad); 1291 } 1292 1293 //------------------------------split_outer_loop------------------------------- 1294 // Split out the outermost loop from this shared header. 1295 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { 1296 PhaseIterGVN &igvn = phase->_igvn; 1297 1298 // Find index of outermost loop; it should also be my tail. 1299 uint outer_idx = 1; 1300 while( _head->in(outer_idx) != _tail ) outer_idx++; 1301 1302 // Make a LoopNode for the outermost loop. 1303 Node *ctl = _head->in(LoopNode::EntryControl); 1304 Node *outer = new LoopNode( ctl, _head->in(outer_idx) ); 1305 outer = igvn.register_new_node_with_optimizer(outer, _head); 1306 phase->set_created_loop_node(); 1307 1308 // Outermost loop falls into '_head' loop 1309 _head->set_req(LoopNode::EntryControl, outer); 1310 _head->del_req(outer_idx); 1311 // Split all the Phis up between '_head' loop and 'outer' loop. 1312 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1313 Node *out = _head->fast_out(j); 1314 if( out->is_Phi() ) { 1315 PhiNode *old_phi = out->as_Phi(); 1316 assert( old_phi->region() == _head, "" ); 1317 Node *phi = PhiNode::make_blank(outer, old_phi); 1318 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); 1319 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); 1320 phi = igvn.register_new_node_with_optimizer(phi, old_phi); 1321 // Make old Phi point to new Phi on the fall-in path 1322 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi); 1323 old_phi->del_req(outer_idx); 1324 } 1325 } 1326 1327 // Use the new loop head instead of the old shared one 1328 _head = outer; 1329 phase->set_loop(_head, this); 1330 } 1331 1332 //------------------------------fix_parent------------------------------------- 1333 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { 1334 loop->_parent = parent; 1335 if( loop->_child ) fix_parent( loop->_child, loop ); 1336 if( loop->_next ) fix_parent( loop->_next , parent ); 1337 } 1338 1339 //------------------------------estimate_path_freq----------------------------- 1340 static float estimate_path_freq( Node *n ) { 1341 // Try to extract some path frequency info 1342 IfNode *iff; 1343 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests 1344 uint nop = n->Opcode(); 1345 if( nop == Op_SafePoint ) { // Skip any safepoint 1346 n = n->in(0); 1347 continue; 1348 } 1349 if( nop == Op_CatchProj ) { // Get count from a prior call 1350 // Assume call does not always throw exceptions: means the call-site 1351 // count is also the frequency of the fall-through path. 1352 assert( n->is_CatchProj(), "" ); 1353 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) 1354 return 0.0f; // Assume call exception path is rare 1355 Node *call = n->in(0)->in(0)->in(0); 1356 assert( call->is_Call(), "expect a call here" ); 1357 const JVMState *jvms = ((CallNode*)call)->jvms(); 1358 ciMethodData* methodData = jvms->method()->method_data(); 1359 if (!methodData->is_mature()) return 0.0f; // No call-site data 1360 ciProfileData* data = methodData->bci_to_data(jvms->bci()); 1361 if ((data == NULL) || !data->is_CounterData()) { 1362 // no call profile available, try call's control input 1363 n = n->in(0); 1364 continue; 1365 } 1366 return data->as_CounterData()->count()/FreqCountInvocations; 1367 } 1368 // See if there's a gating IF test 1369 Node *n_c = n->in(0); 1370 if( !n_c->is_If() ) break; // No estimate available 1371 iff = n_c->as_If(); 1372 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? 1373 // Compute how much count comes on this path 1374 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; 1375 // Have no count info. Skip dull uncommon-trap like branches. 1376 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || 1377 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) 1378 break; 1379 // Skip through never-taken branch; look for a real loop exit. 1380 n = iff->in(0); 1381 } 1382 return 0.0f; // No estimate available 1383 } 1384 1385 //------------------------------merge_many_backedges--------------------------- 1386 // Merge all the backedges from the shared header into a private Region. 1387 // Feed that region as the one backedge to this loop. 1388 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { 1389 uint i; 1390 1391 // Scan for the top 2 hottest backedges 1392 float hotcnt = 0.0f; 1393 float warmcnt = 0.0f; 1394 uint hot_idx = 0; 1395 // Loop starts at 2 because slot 1 is the fall-in path 1396 for( i = 2; i < _head->req(); i++ ) { 1397 float cnt = estimate_path_freq(_head->in(i)); 1398 if( cnt > hotcnt ) { // Grab hottest path 1399 warmcnt = hotcnt; 1400 hotcnt = cnt; 1401 hot_idx = i; 1402 } else if( cnt > warmcnt ) { // And 2nd hottest path 1403 warmcnt = cnt; 1404 } 1405 } 1406 1407 // See if the hottest backedge is worthy of being an inner loop 1408 // by being much hotter than the next hottest backedge. 1409 if( hotcnt <= 0.0001 || 1410 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge 1411 1412 // Peel out the backedges into a private merge point; peel 1413 // them all except optionally hot_idx. 1414 PhaseIterGVN &igvn = phase->_igvn; 1415 1416 Node *hot_tail = NULL; 1417 // Make a Region for the merge point 1418 Node *r = new RegionNode(1); 1419 for( i = 2; i < _head->req(); i++ ) { 1420 if( i != hot_idx ) 1421 r->add_req( _head->in(i) ); 1422 else hot_tail = _head->in(i); 1423 } 1424 igvn.register_new_node_with_optimizer(r, _head); 1425 // Plug region into end of loop _head, followed by hot_tail 1426 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); 1427 igvn.replace_input_of(_head, 2, r); 1428 if( hot_idx ) _head->add_req(hot_tail); 1429 1430 // Split all the Phis up between '_head' loop and the Region 'r' 1431 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1432 Node *out = _head->fast_out(j); 1433 if( out->is_Phi() ) { 1434 PhiNode* n = out->as_Phi(); 1435 igvn.hash_delete(n); // Delete from hash before hacking edges 1436 Node *hot_phi = NULL; 1437 Node *phi = new PhiNode(r, n->type(), n->adr_type()); 1438 // Check all inputs for the ones to peel out 1439 uint j = 1; 1440 for( uint i = 2; i < n->req(); i++ ) { 1441 if( i != hot_idx ) 1442 phi->set_req( j++, n->in(i) ); 1443 else hot_phi = n->in(i); 1444 } 1445 // Register the phi but do not transform until whole place transforms 1446 igvn.register_new_node_with_optimizer(phi, n); 1447 // Add the merge phi to the old Phi 1448 while( n->req() > 3 ) n->del_req( n->req()-1 ); 1449 igvn.replace_input_of(n, 2, phi); 1450 if( hot_idx ) n->add_req(hot_phi); 1451 } 1452 } 1453 1454 1455 // Insert a new IdealLoopTree inserted below me. Turn it into a clone 1456 // of self loop tree. Turn self into a loop headed by _head and with 1457 // tail being the new merge point. 1458 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); 1459 phase->set_loop(_tail,ilt); // Adjust tail 1460 _tail = r; // Self's tail is new merge point 1461 phase->set_loop(r,this); 1462 ilt->_child = _child; // New guy has my children 1463 _child = ilt; // Self has new guy as only child 1464 ilt->_parent = this; // new guy has self for parent 1465 ilt->_nest = _nest; // Same nesting depth (for now) 1466 1467 // Starting with 'ilt', look for child loop trees using the same shared 1468 // header. Flatten these out; they will no longer be loops in the end. 1469 IdealLoopTree **pilt = &_child; 1470 while( ilt ) { 1471 if( ilt->_head == _head ) { 1472 uint i; 1473 for( i = 2; i < _head->req(); i++ ) 1474 if( _head->in(i) == ilt->_tail ) 1475 break; // Still a loop 1476 if( i == _head->req() ) { // No longer a loop 1477 // Flatten ilt. Hang ilt's "_next" list from the end of 1478 // ilt's '_child' list. Move the ilt's _child up to replace ilt. 1479 IdealLoopTree **cp = &ilt->_child; 1480 while( *cp ) cp = &(*cp)->_next; // Find end of child list 1481 *cp = ilt->_next; // Hang next list at end of child list 1482 *pilt = ilt->_child; // Move child up to replace ilt 1483 ilt->_head = NULL; // Flag as a loop UNIONED into parent 1484 ilt = ilt->_child; // Repeat using new ilt 1485 continue; // do not advance over ilt->_child 1486 } 1487 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); 1488 phase->set_loop(_head,ilt); 1489 } 1490 pilt = &ilt->_child; // Advance to next 1491 ilt = *pilt; 1492 } 1493 1494 if( _child ) fix_parent( _child, this ); 1495 } 1496 1497 //------------------------------beautify_loops--------------------------------- 1498 // Split shared headers and insert loop landing pads. 1499 // Insert a LoopNode to replace the RegionNode. 1500 // Return TRUE if loop tree is structurally changed. 1501 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { 1502 bool result = false; 1503 // Cache parts in locals for easy 1504 PhaseIterGVN &igvn = phase->_igvn; 1505 1506 igvn.hash_delete(_head); // Yank from hash before hacking edges 1507 1508 // Check for multiple fall-in paths. Peel off a landing pad if need be. 1509 int fall_in_cnt = 0; 1510 for( uint i = 1; i < _head->req(); i++ ) 1511 if( !phase->is_member( this, _head->in(i) ) ) 1512 fall_in_cnt++; 1513 assert( fall_in_cnt, "at least 1 fall-in path" ); 1514 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins 1515 split_fall_in( phase, fall_in_cnt ); 1516 1517 // Swap inputs to the _head and all Phis to move the fall-in edge to 1518 // the left. 1519 fall_in_cnt = 1; 1520 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) 1521 fall_in_cnt++; 1522 if( fall_in_cnt > 1 ) { 1523 // Since I am just swapping inputs I do not need to update def-use info 1524 Node *tmp = _head->in(1); 1525 igvn.rehash_node_delayed(_head); 1526 _head->set_req( 1, _head->in(fall_in_cnt) ); 1527 _head->set_req( fall_in_cnt, tmp ); 1528 // Swap also all Phis 1529 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { 1530 Node* phi = _head->fast_out(i); 1531 if( phi->is_Phi() ) { 1532 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges 1533 tmp = phi->in(1); 1534 phi->set_req( 1, phi->in(fall_in_cnt) ); 1535 phi->set_req( fall_in_cnt, tmp ); 1536 } 1537 } 1538 } 1539 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); 1540 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); 1541 1542 // If I am a shared header (multiple backedges), peel off the many 1543 // backedges into a private merge point and use the merge point as 1544 // the one true backedge. 1545 if( _head->req() > 3 ) { 1546 // Merge the many backedges into a single backedge but leave 1547 // the hottest backedge as separate edge for the following peel. 1548 merge_many_backedges( phase ); 1549 result = true; 1550 } 1551 1552 // If I have one hot backedge, peel off myself loop. 1553 // I better be the outermost loop. 1554 if (_head->req() > 3 && !_irreducible) { 1555 split_outer_loop( phase ); 1556 result = true; 1557 1558 } else if (!_head->is_Loop() && !_irreducible) { 1559 // Make a new LoopNode to replace the old loop head 1560 Node *l = new LoopNode( _head->in(1), _head->in(2) ); 1561 l = igvn.register_new_node_with_optimizer(l, _head); 1562 phase->set_created_loop_node(); 1563 // Go ahead and replace _head 1564 phase->_igvn.replace_node( _head, l ); 1565 _head = l; 1566 phase->set_loop(_head, this); 1567 } 1568 1569 // Now recursively beautify nested loops 1570 if( _child ) result |= _child->beautify_loops( phase ); 1571 if( _next ) result |= _next ->beautify_loops( phase ); 1572 return result; 1573 } 1574 1575 //------------------------------allpaths_check_safepts---------------------------- 1576 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 1577 // encountered. Helper for check_safepts. 1578 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { 1579 assert(stack.size() == 0, "empty stack"); 1580 stack.push(_tail); 1581 visited.Clear(); 1582 visited.set(_tail->_idx); 1583 while (stack.size() > 0) { 1584 Node* n = stack.pop(); 1585 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 1586 // Terminate this path 1587 } else if (n->Opcode() == Op_SafePoint) { 1588 if (_phase->get_loop(n) != this) { 1589 if (_required_safept == NULL) _required_safept = new Node_List(); 1590 _required_safept->push(n); // save the one closest to the tail 1591 } 1592 // Terminate this path 1593 } else { 1594 uint start = n->is_Region() ? 1 : 0; 1595 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; 1596 for (uint i = start; i < end; i++) { 1597 Node* in = n->in(i); 1598 assert(in->is_CFG(), "must be"); 1599 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { 1600 stack.push(in); 1601 } 1602 } 1603 } 1604 } 1605 } 1606 1607 //------------------------------check_safepts---------------------------- 1608 // Given dominators, try to find loops with calls that must always be 1609 // executed (call dominates loop tail). These loops do not need non-call 1610 // safepoints (ncsfpt). 1611 // 1612 // A complication is that a safepoint in a inner loop may be needed 1613 // by an outer loop. In the following, the inner loop sees it has a 1614 // call (block 3) on every path from the head (block 2) to the 1615 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) 1616 // in block 2, _but_ this leaves the outer loop without a safepoint. 1617 // 1618 // entry 0 1619 // | 1620 // v 1621 // outer 1,2 +->1 1622 // | | 1623 // | v 1624 // | 2<---+ ncsfpt in 2 1625 // |_/|\ | 1626 // | v | 1627 // inner 2,3 / 3 | call in 3 1628 // / | | 1629 // v +--+ 1630 // exit 4 1631 // 1632 // 1633 // This method creates a list (_required_safept) of ncsfpt nodes that must 1634 // be protected is created for each loop. When a ncsfpt maybe deleted, it 1635 // is first looked for in the lists for the outer loops of the current loop. 1636 // 1637 // The insights into the problem: 1638 // A) counted loops are okay 1639 // B) innermost loops are okay (only an inner loop can delete 1640 // a ncsfpt needed by an outer loop) 1641 // C) a loop is immune from an inner loop deleting a safepoint 1642 // if the loop has a call on the idom-path 1643 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the 1644 // idom-path that is not in a nested loop 1645 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner 1646 // loop needs to be prevented from deletion by an inner loop 1647 // 1648 // There are two analyses: 1649 // 1) The first, and cheaper one, scans the loop body from 1650 // tail to head following the idom (immediate dominator) 1651 // chain, looking for the cases (C,D,E) above. 1652 // Since inner loops are scanned before outer loops, there is summary 1653 // information about inner loops. Inner loops can be skipped over 1654 // when the tail of an inner loop is encountered. 1655 // 1656 // 2) The second, invoked if the first fails to find a call or ncsfpt on 1657 // the idom path (which is rare), scans all predecessor control paths 1658 // from the tail to the head, terminating a path when a call or sfpt 1659 // is encountered, to find the ncsfpt's that are closest to the tail. 1660 // 1661 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { 1662 // Bottom up traversal 1663 IdealLoopTree* ch = _child; 1664 if (_child) _child->check_safepts(visited, stack); 1665 if (_next) _next ->check_safepts(visited, stack); 1666 1667 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { 1668 bool has_call = false; // call on dom-path 1669 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth 1670 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth 1671 // Scan the dom-path nodes from tail to head 1672 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { 1673 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 1674 has_call = true; 1675 _has_sfpt = 1; // Then no need for a safept! 1676 break; 1677 } else if (n->Opcode() == Op_SafePoint) { 1678 if (_phase->get_loop(n) == this) { 1679 has_local_ncsfpt = true; 1680 break; 1681 } 1682 if (nonlocal_ncsfpt == NULL) { 1683 nonlocal_ncsfpt = n; // save the one closest to the tail 1684 } 1685 } else { 1686 IdealLoopTree* nlpt = _phase->get_loop(n); 1687 if (this != nlpt) { 1688 // If at an inner loop tail, see if the inner loop has already 1689 // recorded seeing a call on the dom-path (and stop.) If not, 1690 // jump to the head of the inner loop. 1691 assert(is_member(nlpt), "nested loop"); 1692 Node* tail = nlpt->_tail; 1693 if (tail->in(0)->is_If()) tail = tail->in(0); 1694 if (n == tail) { 1695 // If inner loop has call on dom-path, so does outer loop 1696 if (nlpt->_has_sfpt) { 1697 has_call = true; 1698 _has_sfpt = 1; 1699 break; 1700 } 1701 // Skip to head of inner loop 1702 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); 1703 n = nlpt->_head; 1704 } 1705 } 1706 } 1707 } 1708 // Record safept's that this loop needs preserved when an 1709 // inner loop attempts to delete it's safepoints. 1710 if (_child != NULL && !has_call && !has_local_ncsfpt) { 1711 if (nonlocal_ncsfpt != NULL) { 1712 if (_required_safept == NULL) _required_safept = new Node_List(); 1713 _required_safept->push(nonlocal_ncsfpt); 1714 } else { 1715 // Failed to find a suitable safept on the dom-path. Now use 1716 // an all paths walk from tail to head, looking for safepoints to preserve. 1717 allpaths_check_safepts(visited, stack); 1718 } 1719 } 1720 } 1721 } 1722 1723 //---------------------------is_deleteable_safept---------------------------- 1724 // Is safept not required by an outer loop? 1725 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { 1726 assert(sfpt->Opcode() == Op_SafePoint, ""); 1727 IdealLoopTree* lp = get_loop(sfpt)->_parent; 1728 while (lp != NULL) { 1729 Node_List* sfpts = lp->_required_safept; 1730 if (sfpts != NULL) { 1731 for (uint i = 0; i < sfpts->size(); i++) { 1732 if (sfpt == sfpts->at(i)) 1733 return false; 1734 } 1735 } 1736 lp = lp->_parent; 1737 } 1738 return true; 1739 } 1740 1741 //---------------------------replace_parallel_iv------------------------------- 1742 // Replace parallel induction variable (parallel to trip counter) 1743 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) { 1744 assert(loop->_head->is_CountedLoop(), ""); 1745 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1746 if (!cl->is_valid_counted_loop()) 1747 return; // skip malformed counted loop 1748 Node *incr = cl->incr(); 1749 if (incr == NULL) 1750 return; // Dead loop? 1751 Node *init = cl->init_trip(); 1752 Node *phi = cl->phi(); 1753 int stride_con = cl->stride_con(); 1754 1755 // Visit all children, looking for Phis 1756 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { 1757 Node *out = cl->out(i); 1758 // Look for other phis (secondary IVs). Skip dead ones 1759 if (!out->is_Phi() || out == phi || !has_node(out)) 1760 continue; 1761 PhiNode* phi2 = out->as_Phi(); 1762 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); 1763 // Look for induction variables of the form: X += constant 1764 if (phi2->region() != loop->_head || 1765 incr2->req() != 3 || 1766 incr2->in(1) != phi2 || 1767 incr2 == incr || 1768 incr2->Opcode() != Op_AddI || 1769 !incr2->in(2)->is_Con()) 1770 continue; 1771 1772 // Check for parallel induction variable (parallel to trip counter) 1773 // via an affine function. In particular, count-down loops with 1774 // count-up array indices are common. We only RCE references off 1775 // the trip-counter, so we need to convert all these to trip-counter 1776 // expressions. 1777 Node *init2 = phi2->in( LoopNode::EntryControl ); 1778 int stride_con2 = incr2->in(2)->get_int(); 1779 1780 // The general case here gets a little tricky. We want to find the 1781 // GCD of all possible parallel IV's and make a new IV using this 1782 // GCD for the loop. Then all possible IVs are simple multiples of 1783 // the GCD. In practice, this will cover very few extra loops. 1784 // Instead we require 'stride_con2' to be a multiple of 'stride_con', 1785 // where +/-1 is the common case, but other integer multiples are 1786 // also easy to handle. 1787 int ratio_con = stride_con2/stride_con; 1788 1789 if ((ratio_con * stride_con) == stride_con2) { // Check for exact 1790 #ifndef PRODUCT 1791 if (TraceLoopOpts) { 1792 tty->print("Parallel IV: %d ", phi2->_idx); 1793 loop->dump_head(); 1794 } 1795 #endif 1796 // Convert to using the trip counter. The parallel induction 1797 // variable differs from the trip counter by a loop-invariant 1798 // amount, the difference between their respective initial values. 1799 // It is scaled by the 'ratio_con'. 1800 Node* ratio = _igvn.intcon(ratio_con); 1801 set_ctrl(ratio, C->root()); 1802 Node* ratio_init = new MulINode(init, ratio); 1803 _igvn.register_new_node_with_optimizer(ratio_init, init); 1804 set_early_ctrl(ratio_init); 1805 Node* diff = new SubINode(init2, ratio_init); 1806 _igvn.register_new_node_with_optimizer(diff, init2); 1807 set_early_ctrl(diff); 1808 Node* ratio_idx = new MulINode(phi, ratio); 1809 _igvn.register_new_node_with_optimizer(ratio_idx, phi); 1810 set_ctrl(ratio_idx, cl); 1811 Node* add = new AddINode(ratio_idx, diff); 1812 _igvn.register_new_node_with_optimizer(add); 1813 set_ctrl(add, cl); 1814 _igvn.replace_node( phi2, add ); 1815 // Sometimes an induction variable is unused 1816 if (add->outcnt() == 0) { 1817 _igvn.remove_dead_node(add); 1818 } 1819 --i; // deleted this phi; rescan starting with next position 1820 continue; 1821 } 1822 } 1823 } 1824 1825 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) { 1826 Node* keep = NULL; 1827 if (keep_one) { 1828 // Look for a safepoint on the idom-path. 1829 for (Node* i = tail(); i != _head; i = phase->idom(i)) { 1830 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) { 1831 keep = i; 1832 break; // Found one 1833 } 1834 } 1835 } 1836 1837 // Don't remove any safepoints if it is requested to keep a single safepoint and 1838 // no safepoint was found on idom-path. It is not safe to remove any safepoint 1839 // in this case since there's no safepoint dominating all paths in the loop body. 1840 bool prune = !keep_one || keep != NULL; 1841 1842 // Delete other safepoints in this loop. 1843 Node_List* sfpts = _safepts; 1844 if (prune && sfpts != NULL) { 1845 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint"); 1846 for (uint i = 0; i < sfpts->size(); i++) { 1847 Node* n = sfpts->at(i); 1848 assert(phase->get_loop(n) == this, ""); 1849 if (n != keep && phase->is_deleteable_safept(n)) { 1850 phase->lazy_replace(n, n->in(TypeFunc::Control)); 1851 } 1852 } 1853 } 1854 } 1855 1856 //------------------------------counted_loop----------------------------------- 1857 // Convert to counted loops where possible 1858 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { 1859 1860 // For grins, set the inner-loop flag here 1861 if (!_child) { 1862 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop(); 1863 } 1864 1865 if (_head->is_CountedLoop() || 1866 phase->is_counted_loop(_head, this)) { 1867 1868 if (!UseCountedLoopSafepoints) { 1869 // Indicate we do not need a safepoint here 1870 _has_sfpt = 1; 1871 } 1872 1873 // Remove safepoints 1874 bool keep_one_sfpt = !(_has_call || _has_sfpt); 1875 remove_safepoints(phase, keep_one_sfpt); 1876 1877 // Look for induction variables 1878 phase->replace_parallel_iv(this); 1879 1880 } else if (_parent != NULL && !_irreducible) { 1881 // Not a counted loop. Keep one safepoint. 1882 bool keep_one_sfpt = true; 1883 remove_safepoints(phase, keep_one_sfpt); 1884 } 1885 1886 // Recursively 1887 if (_child) _child->counted_loop( phase ); 1888 if (_next) _next ->counted_loop( phase ); 1889 } 1890 1891 #ifndef PRODUCT 1892 //------------------------------dump_head-------------------------------------- 1893 // Dump 1 liner for loop header info 1894 void IdealLoopTree::dump_head( ) const { 1895 for (uint i=0; i<_nest; i++) 1896 tty->print(" "); 1897 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); 1898 if (_irreducible) tty->print(" IRREDUCIBLE"); 1899 Node* entry = _head->in(LoopNode::EntryControl); 1900 if (LoopLimitCheck) { 1901 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 1902 if (predicate != NULL ) { 1903 tty->print(" limit_check"); 1904 entry = entry->in(0)->in(0); 1905 } 1906 } 1907 if (UseLoopPredicate) { 1908 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 1909 if (entry != NULL) { 1910 tty->print(" predicated"); 1911 } 1912 } 1913 if (_head->is_CountedLoop()) { 1914 CountedLoopNode *cl = _head->as_CountedLoop(); 1915 tty->print(" counted"); 1916 1917 Node* init_n = cl->init_trip(); 1918 if (init_n != NULL && init_n->is_Con()) 1919 tty->print(" [%d,", cl->init_trip()->get_int()); 1920 else 1921 tty->print(" [int,"); 1922 Node* limit_n = cl->limit(); 1923 if (limit_n != NULL && limit_n->is_Con()) 1924 tty->print("%d),", cl->limit()->get_int()); 1925 else 1926 tty->print("int),"); 1927 int stride_con = cl->stride_con(); 1928 if (stride_con > 0) tty->print("+"); 1929 tty->print("%d", stride_con); 1930 1931 tty->print(" (%0.f iters) ", cl->profile_trip_cnt()); 1932 1933 if (cl->is_pre_loop ()) tty->print(" pre" ); 1934 if (cl->is_main_loop()) tty->print(" main"); 1935 if (cl->is_post_loop()) tty->print(" post"); 1936 } 1937 if (_has_call) tty->print(" has_call"); 1938 if (_has_sfpt) tty->print(" has_sfpt"); 1939 if (_rce_candidate) tty->print(" rce"); 1940 if (_safepts != NULL && _safepts->size() > 0) { 1941 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }"); 1942 } 1943 if (_required_safept != NULL && _required_safept->size() > 0) { 1944 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }"); 1945 } 1946 tty->cr(); 1947 } 1948 1949 //------------------------------dump------------------------------------------- 1950 // Dump loops by loop tree 1951 void IdealLoopTree::dump( ) const { 1952 dump_head(); 1953 if (_child) _child->dump(); 1954 if (_next) _next ->dump(); 1955 } 1956 1957 #endif 1958 1959 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { 1960 if (loop == root) { 1961 if (loop->_child != NULL) { 1962 log->begin_head("loop_tree"); 1963 log->end_head(); 1964 if( loop->_child ) log_loop_tree(root, loop->_child, log); 1965 log->tail("loop_tree"); 1966 assert(loop->_next == NULL, "what?"); 1967 } 1968 } else { 1969 Node* head = loop->_head; 1970 log->begin_head("loop"); 1971 log->print(" idx='%d' ", head->_idx); 1972 if (loop->_irreducible) log->print("irreducible='1' "); 1973 if (head->is_Loop()) { 1974 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); 1975 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); 1976 } 1977 if (head->is_CountedLoop()) { 1978 CountedLoopNode* cl = head->as_CountedLoop(); 1979 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); 1980 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); 1981 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); 1982 } 1983 log->end_head(); 1984 if( loop->_child ) log_loop_tree(root, loop->_child, log); 1985 log->tail("loop"); 1986 if( loop->_next ) log_loop_tree(root, loop->_next, log); 1987 } 1988 } 1989 1990 //---------------------collect_potentially_useful_predicates----------------------- 1991 // Helper function to collect potentially useful predicates to prevent them from 1992 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates 1993 void PhaseIdealLoop::collect_potentially_useful_predicates( 1994 IdealLoopTree * loop, Unique_Node_List &useful_predicates) { 1995 if (loop->_child) { // child 1996 collect_potentially_useful_predicates(loop->_child, useful_predicates); 1997 } 1998 1999 // self (only loops that we can apply loop predication may use their predicates) 2000 if (loop->_head->is_Loop() && 2001 !loop->_irreducible && 2002 !loop->tail()->is_top()) { 2003 LoopNode* lpn = loop->_head->as_Loop(); 2004 Node* entry = lpn->in(LoopNode::EntryControl); 2005 Node* predicate_proj = find_predicate(entry); // loop_limit_check first 2006 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication 2007 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be"); 2008 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2009 entry = entry->in(0)->in(0); 2010 } 2011 predicate_proj = find_predicate(entry); // Predicate 2012 if (predicate_proj != NULL ) { 2013 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2014 } 2015 } 2016 2017 if (loop->_next) { // sibling 2018 collect_potentially_useful_predicates(loop->_next, useful_predicates); 2019 } 2020 } 2021 2022 //------------------------eliminate_useless_predicates----------------------------- 2023 // Eliminate all inserted predicates if they could not be used by loop predication. 2024 // Note: it will also eliminates loop limits check predicate since it also uses 2025 // Opaque1 node (see Parse::add_predicate()). 2026 void PhaseIdealLoop::eliminate_useless_predicates() { 2027 if (C->predicate_count() == 0) 2028 return; // no predicate left 2029 2030 Unique_Node_List useful_predicates; // to store useful predicates 2031 if (C->has_loops()) { 2032 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); 2033 } 2034 2035 for (int i = C->predicate_count(); i > 0; i--) { 2036 Node * n = C->predicate_opaque1_node(i-1); 2037 assert(n->Opcode() == Op_Opaque1, "must be"); 2038 if (!useful_predicates.member(n)) { // not in the useful list 2039 _igvn.replace_node(n, n->in(1)); 2040 } 2041 } 2042 } 2043 2044 //------------------------process_expensive_nodes----------------------------- 2045 // Expensive nodes have their control input set to prevent the GVN 2046 // from commoning them and as a result forcing the resulting node to 2047 // be in a more frequent path. Use CFG information here, to change the 2048 // control inputs so that some expensive nodes can be commoned while 2049 // not executed more frequently. 2050 bool PhaseIdealLoop::process_expensive_nodes() { 2051 assert(OptimizeExpensiveOps, "optimization off?"); 2052 2053 // Sort nodes to bring similar nodes together 2054 C->sort_expensive_nodes(); 2055 2056 bool progress = false; 2057 2058 for (int i = 0; i < C->expensive_count(); ) { 2059 Node* n = C->expensive_node(i); 2060 int start = i; 2061 // Find nodes similar to n 2062 i++; 2063 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++); 2064 int end = i; 2065 // And compare them two by two 2066 for (int j = start; j < end; j++) { 2067 Node* n1 = C->expensive_node(j); 2068 if (is_node_unreachable(n1)) { 2069 continue; 2070 } 2071 for (int k = j+1; k < end; k++) { 2072 Node* n2 = C->expensive_node(k); 2073 if (is_node_unreachable(n2)) { 2074 continue; 2075 } 2076 2077 assert(n1 != n2, "should be pair of nodes"); 2078 2079 Node* c1 = n1->in(0); 2080 Node* c2 = n2->in(0); 2081 2082 Node* parent_c1 = c1; 2083 Node* parent_c2 = c2; 2084 2085 // The call to get_early_ctrl_for_expensive() moves the 2086 // expensive nodes up but stops at loops that are in a if 2087 // branch. See whether we can exit the loop and move above the 2088 // If. 2089 if (c1->is_Loop()) { 2090 parent_c1 = c1->in(1); 2091 } 2092 if (c2->is_Loop()) { 2093 parent_c2 = c2->in(1); 2094 } 2095 2096 if (parent_c1 == parent_c2) { 2097 _igvn._worklist.push(n1); 2098 _igvn._worklist.push(n2); 2099 continue; 2100 } 2101 2102 // Look for identical expensive node up the dominator chain. 2103 if (is_dominator(c1, c2)) { 2104 c2 = c1; 2105 } else if (is_dominator(c2, c1)) { 2106 c1 = c2; 2107 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() && 2108 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) { 2109 // Both branches have the same expensive node so move it up 2110 // before the if. 2111 c1 = c2 = idom(parent_c1->in(0)); 2112 } 2113 // Do the actual moves 2114 if (n1->in(0) != c1) { 2115 _igvn.hash_delete(n1); 2116 n1->set_req(0, c1); 2117 _igvn.hash_insert(n1); 2118 _igvn._worklist.push(n1); 2119 progress = true; 2120 } 2121 if (n2->in(0) != c2) { 2122 _igvn.hash_delete(n2); 2123 n2->set_req(0, c2); 2124 _igvn.hash_insert(n2); 2125 _igvn._worklist.push(n2); 2126 progress = true; 2127 } 2128 } 2129 } 2130 } 2131 2132 return progress; 2133 } 2134 2135 2136 //============================================================================= 2137 //----------------------------build_and_optimize------------------------------- 2138 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to 2139 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. 2140 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) { 2141 ResourceMark rm; 2142 2143 int old_progress = C->major_progress(); 2144 uint orig_worklist_size = _igvn._worklist.size(); 2145 2146 // Reset major-progress flag for the driver's heuristics 2147 C->clear_major_progress(); 2148 2149 #ifndef PRODUCT 2150 // Capture for later assert 2151 uint unique = C->unique(); 2152 _loop_invokes++; 2153 _loop_work += unique; 2154 #endif 2155 2156 // True if the method has at least 1 irreducible loop 2157 _has_irreducible_loops = false; 2158 2159 _created_loop_node = false; 2160 2161 Arena *a = Thread::current()->resource_area(); 2162 VectorSet visited(a); 2163 // Pre-grow the mapping from Nodes to IdealLoopTrees. 2164 _nodes.map(C->unique(), NULL); 2165 memset(_nodes.adr(), 0, wordSize * C->unique()); 2166 2167 // Pre-build the top-level outermost loop tree entry 2168 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); 2169 // Do not need a safepoint at the top level 2170 _ltree_root->_has_sfpt = 1; 2171 2172 // Initialize Dominators. 2173 // Checked in clone_loop_predicate() during beautify_loops(). 2174 _idom_size = 0; 2175 _idom = NULL; 2176 _dom_depth = NULL; 2177 _dom_stk = NULL; 2178 2179 // Empty pre-order array 2180 allocate_preorders(); 2181 2182 // Build a loop tree on the fly. Build a mapping from CFG nodes to 2183 // IdealLoopTree entries. Data nodes are NOT walked. 2184 build_loop_tree(); 2185 // Check for bailout, and return 2186 if (C->failing()) { 2187 return; 2188 } 2189 2190 // No loops after all 2191 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); 2192 2193 // There should always be an outer loop containing the Root and Return nodes. 2194 // If not, we have a degenerate empty program. Bail out in this case. 2195 if (!has_node(C->root())) { 2196 if (!_verify_only) { 2197 C->clear_major_progress(); 2198 C->record_method_not_compilable("empty program detected during loop optimization"); 2199 } 2200 return; 2201 } 2202 2203 // Nothing to do, so get out 2204 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only; 2205 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn); 2206 if (stop_early && !do_expensive_nodes) { 2207 _igvn.optimize(); // Cleanup NeverBranches 2208 return; 2209 } 2210 2211 // Set loop nesting depth 2212 _ltree_root->set_nest( 0 ); 2213 2214 // Split shared headers and insert loop landing pads. 2215 // Do not bother doing this on the Root loop of course. 2216 if( !_verify_me && !_verify_only && _ltree_root->_child ) { 2217 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3); 2218 if( _ltree_root->_child->beautify_loops( this ) ) { 2219 // Re-build loop tree! 2220 _ltree_root->_child = NULL; 2221 _nodes.clear(); 2222 reallocate_preorders(); 2223 build_loop_tree(); 2224 // Check for bailout, and return 2225 if (C->failing()) { 2226 return; 2227 } 2228 // Reset loop nesting depth 2229 _ltree_root->set_nest( 0 ); 2230 2231 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3); 2232 } 2233 } 2234 2235 // Build Dominators for elision of NULL checks & loop finding. 2236 // Since nodes do not have a slot for immediate dominator, make 2237 // a persistent side array for that info indexed on node->_idx. 2238 _idom_size = C->unique(); 2239 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); 2240 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); 2241 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth 2242 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); 2243 2244 Dominators(); 2245 2246 if (!_verify_only) { 2247 // As a side effect, Dominators removed any unreachable CFG paths 2248 // into RegionNodes. It doesn't do this test against Root, so 2249 // we do it here. 2250 for( uint i = 1; i < C->root()->req(); i++ ) { 2251 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? 2252 _igvn.delete_input_of(C->root(), i); 2253 i--; // Rerun same iteration on compressed edges 2254 } 2255 } 2256 2257 // Given dominators, try to find inner loops with calls that must 2258 // always be executed (call dominates loop tail). These loops do 2259 // not need a separate safepoint. 2260 Node_List cisstack(a); 2261 _ltree_root->check_safepts(visited, cisstack); 2262 } 2263 2264 // Walk the DATA nodes and place into loops. Find earliest control 2265 // node. For CFG nodes, the _nodes array starts out and remains 2266 // holding the associated IdealLoopTree pointer. For DATA nodes, the 2267 // _nodes array holds the earliest legal controlling CFG node. 2268 2269 // Allocate stack with enough space to avoid frequent realloc 2270 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats 2271 Node_Stack nstack( a, stack_size ); 2272 2273 visited.Clear(); 2274 Node_List worklist(a); 2275 // Don't need C->root() on worklist since 2276 // it will be processed among C->top() inputs 2277 worklist.push( C->top() ); 2278 visited.set( C->top()->_idx ); // Set C->top() as visited now 2279 build_loop_early( visited, worklist, nstack ); 2280 2281 // Given early legal placement, try finding counted loops. This placement 2282 // is good enough to discover most loop invariants. 2283 if( !_verify_me && !_verify_only ) 2284 _ltree_root->counted_loop( this ); 2285 2286 // Find latest loop placement. Find ideal loop placement. 2287 visited.Clear(); 2288 init_dom_lca_tags(); 2289 // Need C->root() on worklist when processing outs 2290 worklist.push( C->root() ); 2291 NOT_PRODUCT( C->verify_graph_edges(); ) 2292 worklist.push( C->top() ); 2293 build_loop_late( visited, worklist, nstack ); 2294 2295 if (_verify_only) { 2296 // restore major progress flag 2297 for (int i = 0; i < old_progress; i++) 2298 C->set_major_progress(); 2299 assert(C->unique() == unique, "verification mode made Nodes? ? ?"); 2300 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything"); 2301 return; 2302 } 2303 2304 // clear out the dead code after build_loop_late 2305 while (_deadlist.size()) { 2306 _igvn.remove_globally_dead_node(_deadlist.pop()); 2307 } 2308 2309 if (stop_early) { 2310 assert(do_expensive_nodes, "why are we here?"); 2311 if (process_expensive_nodes()) { 2312 // If we made some progress when processing expensive nodes then 2313 // the IGVN may modify the graph in a way that will allow us to 2314 // make some more progress: we need to try processing expensive 2315 // nodes again. 2316 C->set_major_progress(); 2317 } 2318 _igvn.optimize(); 2319 return; 2320 } 2321 2322 // Some parser-inserted loop predicates could never be used by loop 2323 // predication or they were moved away from loop during some optimizations. 2324 // For example, peeling. Eliminate them before next loop optimizations. 2325 if (UseLoopPredicate || LoopLimitCheck) { 2326 eliminate_useless_predicates(); 2327 } 2328 2329 #ifndef PRODUCT 2330 C->verify_graph_edges(); 2331 if (_verify_me) { // Nested verify pass? 2332 // Check to see if the verify mode is broken 2333 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); 2334 return; 2335 } 2336 if(VerifyLoopOptimizations) verify(); 2337 if(TraceLoopOpts && C->has_loops()) { 2338 _ltree_root->dump(); 2339 } 2340 #endif 2341 2342 if (skip_loop_opts) { 2343 // Cleanup any modified bits 2344 _igvn.optimize(); 2345 2346 if (C->log() != NULL) { 2347 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2348 } 2349 return; 2350 } 2351 2352 if (ReassociateInvariants) { 2353 // Reassociate invariants and prep for split_thru_phi 2354 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2355 IdealLoopTree* lpt = iter.current(); 2356 bool is_counted = lpt->is_counted(); 2357 if (!is_counted || !lpt->is_inner()) continue; 2358 2359 // check for vectorized loops, any reassociation of invariants was already done 2360 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue; 2361 2362 lpt->reassociate_invariants(this); 2363 2364 // Because RCE opportunities can be masked by split_thru_phi, 2365 // look for RCE candidates and inhibit split_thru_phi 2366 // on just their loop-phi's for this pass of loop opts 2367 if (SplitIfBlocks && do_split_ifs) { 2368 if (lpt->policy_range_check(this)) { 2369 lpt->_rce_candidate = 1; // = true 2370 } 2371 } 2372 } 2373 } 2374 2375 // Check for aggressive application of split-if and other transforms 2376 // that require basic-block info (like cloning through Phi's) 2377 if( SplitIfBlocks && do_split_ifs ) { 2378 visited.Clear(); 2379 split_if_with_blocks( visited, nstack ); 2380 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); 2381 } 2382 2383 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) { 2384 C->set_major_progress(); 2385 } 2386 2387 // Perform loop predication before iteration splitting 2388 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) { 2389 _ltree_root->_child->loop_predication(this); 2390 } 2391 2392 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { 2393 if (do_intrinsify_fill()) { 2394 C->set_major_progress(); 2395 } 2396 } 2397 2398 // Perform iteration-splitting on inner loops. Split iterations to avoid 2399 // range checks or one-shot null checks. 2400 2401 // If split-if's didn't hack the graph too bad (no CFG changes) 2402 // then do loop opts. 2403 if (C->has_loops() && !C->major_progress()) { 2404 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 2405 _ltree_root->_child->iteration_split( this, worklist ); 2406 // No verify after peeling! GCM has hoisted code out of the loop. 2407 // After peeling, the hoisted code could sink inside the peeled area. 2408 // The peeling code does not try to recompute the best location for 2409 // all the code before the peeled area, so the verify pass will always 2410 // complain about it. 2411 } 2412 // Do verify graph edges in any case 2413 NOT_PRODUCT( C->verify_graph_edges(); ); 2414 2415 if (!do_split_ifs) { 2416 // We saw major progress in Split-If to get here. We forced a 2417 // pass with unrolling and not split-if, however more split-if's 2418 // might make progress. If the unrolling didn't make progress 2419 // then the major-progress flag got cleared and we won't try 2420 // another round of Split-If. In particular the ever-common 2421 // instance-of/check-cast pattern requires at least 2 rounds of 2422 // Split-If to clear out. 2423 C->set_major_progress(); 2424 } 2425 2426 // Repeat loop optimizations if new loops were seen 2427 if (created_loop_node()) { 2428 C->set_major_progress(); 2429 } 2430 2431 // Keep loop predicates and perform optimizations with them 2432 // until no more loop optimizations could be done. 2433 // After that switch predicates off and do more loop optimizations. 2434 if (!C->major_progress() && (C->predicate_count() > 0)) { 2435 C->cleanup_loop_predicates(_igvn); 2436 if (TraceLoopOpts) { 2437 tty->print_cr("PredicatesOff"); 2438 } 2439 C->set_major_progress(); 2440 } 2441 2442 // Convert scalar to superword operations at the end of all loop opts. 2443 if (UseSuperWord && C->has_loops() && !C->major_progress()) { 2444 // SuperWord transform 2445 SuperWord sw(this); 2446 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2447 IdealLoopTree* lpt = iter.current(); 2448 if (lpt->is_counted()) { 2449 sw.transform_loop(lpt, true); 2450 } 2451 } 2452 } 2453 2454 // Cleanup any modified bits 2455 _igvn.optimize(); 2456 2457 // disable assert until issue with split_flow_path is resolved (6742111) 2458 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), 2459 // "shouldn't introduce irreducible loops"); 2460 2461 if (C->log() != NULL) { 2462 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2463 } 2464 } 2465 2466 #ifndef PRODUCT 2467 //------------------------------print_statistics------------------------------- 2468 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes 2469 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique 2470 void PhaseIdealLoop::print_statistics() { 2471 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); 2472 } 2473 2474 //------------------------------verify----------------------------------------- 2475 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. 2476 static int fail; // debug only, so its multi-thread dont care 2477 void PhaseIdealLoop::verify() const { 2478 int old_progress = C->major_progress(); 2479 ResourceMark rm; 2480 PhaseIdealLoop loop_verify( _igvn, this ); 2481 VectorSet visited(Thread::current()->resource_area()); 2482 2483 fail = 0; 2484 verify_compare( C->root(), &loop_verify, visited ); 2485 assert( fail == 0, "verify loops failed" ); 2486 // Verify loop structure is the same 2487 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); 2488 // Reset major-progress. It was cleared by creating a verify version of 2489 // PhaseIdealLoop. 2490 for( int i=0; i<old_progress; i++ ) 2491 C->set_major_progress(); 2492 } 2493 2494 //------------------------------verify_compare--------------------------------- 2495 // Make sure me and the given PhaseIdealLoop agree on key data structures 2496 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { 2497 if( !n ) return; 2498 if( visited.test_set( n->_idx ) ) return; 2499 if( !_nodes[n->_idx] ) { // Unreachable 2500 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); 2501 return; 2502 } 2503 2504 uint i; 2505 for( i = 0; i < n->req(); i++ ) 2506 verify_compare( n->in(i), loop_verify, visited ); 2507 2508 // Check the '_nodes' block/loop structure 2509 i = n->_idx; 2510 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl 2511 if( _nodes[i] != loop_verify->_nodes[i] && 2512 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { 2513 tty->print("Mismatched control setting for: "); 2514 n->dump(); 2515 if( fail++ > 10 ) return; 2516 Node *c = get_ctrl_no_update(n); 2517 tty->print("We have it as: "); 2518 if( c->in(0) ) c->dump(); 2519 else tty->print_cr("N%d",c->_idx); 2520 tty->print("Verify thinks: "); 2521 if( loop_verify->has_ctrl(n) ) 2522 loop_verify->get_ctrl_no_update(n)->dump(); 2523 else 2524 loop_verify->get_loop_idx(n)->dump(); 2525 tty->cr(); 2526 } 2527 } else { // We have a loop 2528 IdealLoopTree *us = get_loop_idx(n); 2529 if( loop_verify->has_ctrl(n) ) { 2530 tty->print("Mismatched loop setting for: "); 2531 n->dump(); 2532 if( fail++ > 10 ) return; 2533 tty->print("We have it as: "); 2534 us->dump(); 2535 tty->print("Verify thinks: "); 2536 loop_verify->get_ctrl_no_update(n)->dump(); 2537 tty->cr(); 2538 } else if (!C->major_progress()) { 2539 // Loop selection can be messed up if we did a major progress 2540 // operation, like split-if. Do not verify in that case. 2541 IdealLoopTree *them = loop_verify->get_loop_idx(n); 2542 if( us->_head != them->_head || us->_tail != them->_tail ) { 2543 tty->print("Unequals loops for: "); 2544 n->dump(); 2545 if( fail++ > 10 ) return; 2546 tty->print("We have it as: "); 2547 us->dump(); 2548 tty->print("Verify thinks: "); 2549 them->dump(); 2550 tty->cr(); 2551 } 2552 } 2553 } 2554 2555 // Check for immediate dominators being equal 2556 if( i >= _idom_size ) { 2557 if( !n->is_CFG() ) return; 2558 tty->print("CFG Node with no idom: "); 2559 n->dump(); 2560 return; 2561 } 2562 if( !n->is_CFG() ) return; 2563 if( n == C->root() ) return; // No IDOM here 2564 2565 assert(n->_idx == i, "sanity"); 2566 Node *id = idom_no_update(n); 2567 if( id != loop_verify->idom_no_update(n) ) { 2568 tty->print("Unequals idoms for: "); 2569 n->dump(); 2570 if( fail++ > 10 ) return; 2571 tty->print("We have it as: "); 2572 id->dump(); 2573 tty->print("Verify thinks: "); 2574 loop_verify->idom_no_update(n)->dump(); 2575 tty->cr(); 2576 } 2577 2578 } 2579 2580 //------------------------------verify_tree------------------------------------ 2581 // Verify that tree structures match. Because the CFG can change, siblings 2582 // within the loop tree can be reordered. We attempt to deal with that by 2583 // reordering the verify's loop tree if possible. 2584 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { 2585 assert( _parent == parent, "Badly formed loop tree" ); 2586 2587 // Siblings not in same order? Attempt to re-order. 2588 if( _head != loop->_head ) { 2589 // Find _next pointer to update 2590 IdealLoopTree **pp = &loop->_parent->_child; 2591 while( *pp != loop ) 2592 pp = &((*pp)->_next); 2593 // Find proper sibling to be next 2594 IdealLoopTree **nn = &loop->_next; 2595 while( (*nn) && (*nn)->_head != _head ) 2596 nn = &((*nn)->_next); 2597 2598 // Check for no match. 2599 if( !(*nn) ) { 2600 // Annoyingly, irreducible loops can pick different headers 2601 // after a major_progress operation, so the rest of the loop 2602 // tree cannot be matched. 2603 if (_irreducible && Compile::current()->major_progress()) return; 2604 assert( 0, "failed to match loop tree" ); 2605 } 2606 2607 // Move (*nn) to (*pp) 2608 IdealLoopTree *hit = *nn; 2609 *nn = hit->_next; 2610 hit->_next = loop; 2611 *pp = loop; 2612 loop = hit; 2613 // Now try again to verify 2614 } 2615 2616 assert( _head == loop->_head , "mismatched loop head" ); 2617 Node *tail = _tail; // Inline a non-updating version of 2618 while( !tail->in(0) ) // the 'tail()' call. 2619 tail = tail->in(1); 2620 assert( tail == loop->_tail, "mismatched loop tail" ); 2621 2622 // Counted loops that are guarded should be able to find their guards 2623 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { 2624 CountedLoopNode *cl = _head->as_CountedLoop(); 2625 Node *init = cl->init_trip(); 2626 Node *ctrl = cl->in(LoopNode::EntryControl); 2627 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 2628 Node *iff = ctrl->in(0); 2629 assert( iff->Opcode() == Op_If, "" ); 2630 Node *bol = iff->in(1); 2631 assert( bol->Opcode() == Op_Bool, "" ); 2632 Node *cmp = bol->in(1); 2633 assert( cmp->Opcode() == Op_CmpI, "" ); 2634 Node *add = cmp->in(1); 2635 Node *opaq; 2636 if( add->Opcode() == Op_Opaque1 ) { 2637 opaq = add; 2638 } else { 2639 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); 2640 assert( add == init, "" ); 2641 opaq = cmp->in(2); 2642 } 2643 assert( opaq->Opcode() == Op_Opaque1, "" ); 2644 2645 } 2646 2647 if (_child != NULL) _child->verify_tree(loop->_child, this); 2648 if (_next != NULL) _next ->verify_tree(loop->_next, parent); 2649 // Innermost loops need to verify loop bodies, 2650 // but only if no 'major_progress' 2651 int fail = 0; 2652 if (!Compile::current()->major_progress() && _child == NULL) { 2653 for( uint i = 0; i < _body.size(); i++ ) { 2654 Node *n = _body.at(i); 2655 if (n->outcnt() == 0) continue; // Ignore dead 2656 uint j; 2657 for( j = 0; j < loop->_body.size(); j++ ) 2658 if( loop->_body.at(j) == n ) 2659 break; 2660 if( j == loop->_body.size() ) { // Not found in loop body 2661 // Last ditch effort to avoid assertion: Its possible that we 2662 // have some users (so outcnt not zero) but are still dead. 2663 // Try to find from root. 2664 if (Compile::current()->root()->find(n->_idx)) { 2665 fail++; 2666 tty->print("We have that verify does not: "); 2667 n->dump(); 2668 } 2669 } 2670 } 2671 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { 2672 Node *n = loop->_body.at(i2); 2673 if (n->outcnt() == 0) continue; // Ignore dead 2674 uint j; 2675 for( j = 0; j < _body.size(); j++ ) 2676 if( _body.at(j) == n ) 2677 break; 2678 if( j == _body.size() ) { // Not found in loop body 2679 // Last ditch effort to avoid assertion: Its possible that we 2680 // have some users (so outcnt not zero) but are still dead. 2681 // Try to find from root. 2682 if (Compile::current()->root()->find(n->_idx)) { 2683 fail++; 2684 tty->print("Verify has that we do not: "); 2685 n->dump(); 2686 } 2687 } 2688 } 2689 assert( !fail, "loop body mismatch" ); 2690 } 2691 } 2692 2693 #endif 2694 2695 //------------------------------set_idom--------------------------------------- 2696 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { 2697 uint idx = d->_idx; 2698 if (idx >= _idom_size) { 2699 uint newsize = _idom_size<<1; 2700 while( idx >= newsize ) { 2701 newsize <<= 1; 2702 } 2703 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); 2704 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); 2705 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); 2706 _idom_size = newsize; 2707 } 2708 _idom[idx] = n; 2709 _dom_depth[idx] = dom_depth; 2710 } 2711 2712 //------------------------------recompute_dom_depth--------------------------------------- 2713 // The dominator tree is constructed with only parent pointers. 2714 // This recomputes the depth in the tree by first tagging all 2715 // nodes as "no depth yet" marker. The next pass then runs up 2716 // the dom tree from each node marked "no depth yet", and computes 2717 // the depth on the way back down. 2718 void PhaseIdealLoop::recompute_dom_depth() { 2719 uint no_depth_marker = C->unique(); 2720 uint i; 2721 // Initialize depth to "no depth yet" 2722 for (i = 0; i < _idom_size; i++) { 2723 if (_dom_depth[i] > 0 && _idom[i] != NULL) { 2724 _dom_depth[i] = no_depth_marker; 2725 } 2726 } 2727 if (_dom_stk == NULL) { 2728 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size. 2729 if (init_size < 10) init_size = 10; 2730 _dom_stk = new GrowableArray<uint>(init_size); 2731 } 2732 // Compute new depth for each node. 2733 for (i = 0; i < _idom_size; i++) { 2734 uint j = i; 2735 // Run up the dom tree to find a node with a depth 2736 while (_dom_depth[j] == no_depth_marker) { 2737 _dom_stk->push(j); 2738 j = _idom[j]->_idx; 2739 } 2740 // Compute the depth on the way back down this tree branch 2741 uint dd = _dom_depth[j] + 1; 2742 while (_dom_stk->length() > 0) { 2743 uint j = _dom_stk->pop(); 2744 _dom_depth[j] = dd; 2745 dd++; 2746 } 2747 } 2748 } 2749 2750 //------------------------------sort------------------------------------------- 2751 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the 2752 // loop tree, not the root. 2753 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { 2754 if( !innermost ) return loop; // New innermost loop 2755 2756 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number 2757 assert( loop_preorder, "not yet post-walked loop" ); 2758 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer 2759 IdealLoopTree *l = *pp; // Do I go before or after 'l'? 2760 2761 // Insert at start of list 2762 while( l ) { // Insertion sort based on pre-order 2763 if( l == loop ) return innermost; // Already on list! 2764 int l_preorder = get_preorder(l->_head); // Cache pre-order number 2765 assert( l_preorder, "not yet post-walked l" ); 2766 // Check header pre-order number to figure proper nesting 2767 if( loop_preorder > l_preorder ) 2768 break; // End of insertion 2769 // If headers tie (e.g., shared headers) check tail pre-order numbers. 2770 // Since I split shared headers, you'd think this could not happen. 2771 // BUT: I must first do the preorder numbering before I can discover I 2772 // have shared headers, so the split headers all get the same preorder 2773 // number as the RegionNode they split from. 2774 if( loop_preorder == l_preorder && 2775 get_preorder(loop->_tail) < get_preorder(l->_tail) ) 2776 break; // Also check for shared headers (same pre#) 2777 pp = &l->_parent; // Chain up list 2778 l = *pp; 2779 } 2780 // Link into list 2781 // Point predecessor to me 2782 *pp = loop; 2783 // Point me to successor 2784 IdealLoopTree *p = loop->_parent; 2785 loop->_parent = l; // Point me to successor 2786 if( p ) sort( p, innermost ); // Insert my parents into list as well 2787 return innermost; 2788 } 2789 2790 //------------------------------build_loop_tree-------------------------------- 2791 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit 2792 // bits. The _nodes[] array is mapped by Node index and holds a NULL for 2793 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the 2794 // tightest enclosing IdealLoopTree for post-walked. 2795 // 2796 // During my forward walk I do a short 1-layer lookahead to see if I can find 2797 // a loop backedge with that doesn't have any work on the backedge. This 2798 // helps me construct nested loops with shared headers better. 2799 // 2800 // Once I've done the forward recursion, I do the post-work. For each child 2801 // I check to see if there is a backedge. Backedges define a loop! I 2802 // insert an IdealLoopTree at the target of the backedge. 2803 // 2804 // During the post-work I also check to see if I have several children 2805 // belonging to different loops. If so, then this Node is a decision point 2806 // where control flow can choose to change loop nests. It is at this 2807 // decision point where I can figure out how loops are nested. At this 2808 // time I can properly order the different loop nests from my children. 2809 // Note that there may not be any backedges at the decision point! 2810 // 2811 // Since the decision point can be far removed from the backedges, I can't 2812 // order my loops at the time I discover them. Thus at the decision point 2813 // I need to inspect loop header pre-order numbers to properly nest my 2814 // loops. This means I need to sort my childrens' loops by pre-order. 2815 // The sort is of size number-of-control-children, which generally limits 2816 // it to size 2 (i.e., I just choose between my 2 target loops). 2817 void PhaseIdealLoop::build_loop_tree() { 2818 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 2819 GrowableArray <Node *> bltstack(C->live_nodes() >> 1); 2820 Node *n = C->root(); 2821 bltstack.push(n); 2822 int pre_order = 1; 2823 int stack_size; 2824 2825 while ( ( stack_size = bltstack.length() ) != 0 ) { 2826 n = bltstack.top(); // Leave node on stack 2827 if ( !is_visited(n) ) { 2828 // ---- Pre-pass Work ---- 2829 // Pre-walked but not post-walked nodes need a pre_order number. 2830 2831 set_preorder_visited( n, pre_order ); // set as visited 2832 2833 // ---- Scan over children ---- 2834 // Scan first over control projections that lead to loop headers. 2835 // This helps us find inner-to-outer loops with shared headers better. 2836 2837 // Scan children's children for loop headers. 2838 for ( int i = n->outcnt() - 1; i >= 0; --i ) { 2839 Node* m = n->raw_out(i); // Child 2840 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children 2841 // Scan over children's children to find loop 2842 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 2843 Node* l = m->fast_out(j); 2844 if( is_visited(l) && // Been visited? 2845 !is_postvisited(l) && // But not post-visited 2846 get_preorder(l) < pre_order ) { // And smaller pre-order 2847 // Found! Scan the DFS down this path before doing other paths 2848 bltstack.push(m); 2849 break; 2850 } 2851 } 2852 } 2853 } 2854 pre_order++; 2855 } 2856 else if ( !is_postvisited(n) ) { 2857 // Note: build_loop_tree_impl() adds out edges on rare occasions, 2858 // such as com.sun.rsasign.am::a. 2859 // For non-recursive version, first, process current children. 2860 // On next iteration, check if additional children were added. 2861 for ( int k = n->outcnt() - 1; k >= 0; --k ) { 2862 Node* u = n->raw_out(k); 2863 if ( u->is_CFG() && !is_visited(u) ) { 2864 bltstack.push(u); 2865 } 2866 } 2867 if ( bltstack.length() == stack_size ) { 2868 // There were no additional children, post visit node now 2869 (void)bltstack.pop(); // Remove node from stack 2870 pre_order = build_loop_tree_impl( n, pre_order ); 2871 // Check for bailout 2872 if (C->failing()) { 2873 return; 2874 } 2875 // Check to grow _preorders[] array for the case when 2876 // build_loop_tree_impl() adds new nodes. 2877 check_grow_preorders(); 2878 } 2879 } 2880 else { 2881 (void)bltstack.pop(); // Remove post-visited node from stack 2882 } 2883 } 2884 } 2885 2886 //------------------------------build_loop_tree_impl--------------------------- 2887 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { 2888 // ---- Post-pass Work ---- 2889 // Pre-walked but not post-walked nodes need a pre_order number. 2890 2891 // Tightest enclosing loop for this Node 2892 IdealLoopTree *innermost = NULL; 2893 2894 // For all children, see if any edge is a backedge. If so, make a loop 2895 // for it. Then find the tightest enclosing loop for the self Node. 2896 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2897 Node* m = n->fast_out(i); // Child 2898 if( n == m ) continue; // Ignore control self-cycles 2899 if( !m->is_CFG() ) continue;// Ignore non-CFG edges 2900 2901 IdealLoopTree *l; // Child's loop 2902 if( !is_postvisited(m) ) { // Child visited but not post-visited? 2903 // Found a backedge 2904 assert( get_preorder(m) < pre_order, "should be backedge" ); 2905 // Check for the RootNode, which is already a LoopNode and is allowed 2906 // to have multiple "backedges". 2907 if( m == C->root()) { // Found the root? 2908 l = _ltree_root; // Root is the outermost LoopNode 2909 } else { // Else found a nested loop 2910 // Insert a LoopNode to mark this loop. 2911 l = new IdealLoopTree(this, m, n); 2912 } // End of Else found a nested loop 2913 if( !has_loop(m) ) // If 'm' does not already have a loop set 2914 set_loop(m, l); // Set loop header to loop now 2915 2916 } else { // Else not a nested loop 2917 if( !_nodes[m->_idx] ) continue; // Dead code has no loop 2918 l = get_loop(m); // Get previously determined loop 2919 // If successor is header of a loop (nest), move up-loop till it 2920 // is a member of some outer enclosing loop. Since there are no 2921 // shared headers (I've split them already) I only need to go up 2922 // at most 1 level. 2923 while( l && l->_head == m ) // Successor heads loop? 2924 l = l->_parent; // Move up 1 for me 2925 // If this loop is not properly parented, then this loop 2926 // has no exit path out, i.e. its an infinite loop. 2927 if( !l ) { 2928 // Make loop "reachable" from root so the CFG is reachable. Basically 2929 // insert a bogus loop exit that is never taken. 'm', the loop head, 2930 // points to 'n', one (of possibly many) fall-in paths. There may be 2931 // many backedges as well. 2932 2933 // Here I set the loop to be the root loop. I could have, after 2934 // inserting a bogus loop exit, restarted the recursion and found my 2935 // new loop exit. This would make the infinite loop a first-class 2936 // loop and it would then get properly optimized. What's the use of 2937 // optimizing an infinite loop? 2938 l = _ltree_root; // Oops, found infinite loop 2939 2940 if (!_verify_only) { 2941 // Insert the NeverBranch between 'm' and it's control user. 2942 NeverBranchNode *iff = new NeverBranchNode( m ); 2943 _igvn.register_new_node_with_optimizer(iff); 2944 set_loop(iff, l); 2945 Node *if_t = new CProjNode( iff, 0 ); 2946 _igvn.register_new_node_with_optimizer(if_t); 2947 set_loop(if_t, l); 2948 2949 Node* cfg = NULL; // Find the One True Control User of m 2950 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 2951 Node* x = m->fast_out(j); 2952 if (x->is_CFG() && x != m && x != iff) 2953 { cfg = x; break; } 2954 } 2955 assert(cfg != NULL, "must find the control user of m"); 2956 uint k = 0; // Probably cfg->in(0) 2957 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region 2958 cfg->set_req( k, if_t ); // Now point to NeverBranch 2959 _igvn._worklist.push(cfg); 2960 2961 // Now create the never-taken loop exit 2962 Node *if_f = new CProjNode( iff, 1 ); 2963 _igvn.register_new_node_with_optimizer(if_f); 2964 set_loop(if_f, l); 2965 // Find frame ptr for Halt. Relies on the optimizer 2966 // V-N'ing. Easier and quicker than searching through 2967 // the program structure. 2968 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr ); 2969 _igvn.register_new_node_with_optimizer(frame); 2970 // Halt & Catch Fire 2971 Node *halt = new HaltNode( if_f, frame ); 2972 _igvn.register_new_node_with_optimizer(halt); 2973 set_loop(halt, l); 2974 C->root()->add_req(halt); 2975 } 2976 set_loop(C->root(), _ltree_root); 2977 } 2978 } 2979 // Weeny check for irreducible. This child was already visited (this 2980 // IS the post-work phase). Is this child's loop header post-visited 2981 // as well? If so, then I found another entry into the loop. 2982 if (!_verify_only) { 2983 while( is_postvisited(l->_head) ) { 2984 // found irreducible 2985 l->_irreducible = 1; // = true 2986 l = l->_parent; 2987 _has_irreducible_loops = true; 2988 // Check for bad CFG here to prevent crash, and bailout of compile 2989 if (l == NULL) { 2990 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); 2991 return pre_order; 2992 } 2993 } 2994 C->set_has_irreducible_loop(_has_irreducible_loops); 2995 } 2996 2997 // This Node might be a decision point for loops. It is only if 2998 // it's children belong to several different loops. The sort call 2999 // does a trivial amount of work if there is only 1 child or all 3000 // children belong to the same loop. If however, the children 3001 // belong to different loops, the sort call will properly set the 3002 // _parent pointers to show how the loops nest. 3003 // 3004 // In any case, it returns the tightest enclosing loop. 3005 innermost = sort( l, innermost ); 3006 } 3007 3008 // Def-use info will have some dead stuff; dead stuff will have no 3009 // loop decided on. 3010 3011 // Am I a loop header? If so fix up my parent's child and next ptrs. 3012 if( innermost && innermost->_head == n ) { 3013 assert( get_loop(n) == innermost, "" ); 3014 IdealLoopTree *p = innermost->_parent; 3015 IdealLoopTree *l = innermost; 3016 while( p && l->_head == n ) { 3017 l->_next = p->_child; // Put self on parents 'next child' 3018 p->_child = l; // Make self as first child of parent 3019 l = p; // Now walk up the parent chain 3020 p = l->_parent; 3021 } 3022 } else { 3023 // Note that it is possible for a LoopNode to reach here, if the 3024 // backedge has been made unreachable (hence the LoopNode no longer 3025 // denotes a Loop, and will eventually be removed). 3026 3027 // Record tightest enclosing loop for self. Mark as post-visited. 3028 set_loop(n, innermost); 3029 // Also record has_call flag early on 3030 if( innermost ) { 3031 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { 3032 // Do not count uncommon calls 3033 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { 3034 Node *iff = n->in(0)->in(0); 3035 // No any calls for vectorized loops. 3036 if( UseSuperWord || !iff->is_If() || 3037 (n->in(0)->Opcode() == Op_IfFalse && 3038 (1.0 - iff->as_If()->_prob) >= 0.01) || 3039 (iff->as_If()->_prob >= 0.01) ) 3040 innermost->_has_call = 1; 3041 } 3042 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { 3043 // Disable loop optimizations if the loop has a scalar replaceable 3044 // allocation. This disabling may cause a potential performance lost 3045 // if the allocation is not eliminated for some reason. 3046 innermost->_allow_optimizations = false; 3047 innermost->_has_call = 1; // = true 3048 } else if (n->Opcode() == Op_SafePoint) { 3049 // Record all safepoints in this loop. 3050 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List(); 3051 innermost->_safepts->push(n); 3052 } 3053 } 3054 } 3055 3056 // Flag as post-visited now 3057 set_postvisited(n); 3058 return pre_order; 3059 } 3060 3061 3062 //------------------------------build_loop_early------------------------------- 3063 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3064 // First pass computes the earliest controlling node possible. This is the 3065 // controlling input with the deepest dominating depth. 3066 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3067 while (worklist.size() != 0) { 3068 // Use local variables nstack_top_n & nstack_top_i to cache values 3069 // on nstack's top. 3070 Node *nstack_top_n = worklist.pop(); 3071 uint nstack_top_i = 0; 3072 //while_nstack_nonempty: 3073 while (true) { 3074 // Get parent node and next input's index from stack's top. 3075 Node *n = nstack_top_n; 3076 uint i = nstack_top_i; 3077 uint cnt = n->req(); // Count of inputs 3078 if (i == 0) { // Pre-process the node. 3079 if( has_node(n) && // Have either loop or control already? 3080 !has_ctrl(n) ) { // Have loop picked out already? 3081 // During "merge_many_backedges" we fold up several nested loops 3082 // into a single loop. This makes the members of the original 3083 // loop bodies pointing to dead loops; they need to move up 3084 // to the new UNION'd larger loop. I set the _head field of these 3085 // dead loops to NULL and the _parent field points to the owning 3086 // loop. Shades of UNION-FIND algorithm. 3087 IdealLoopTree *ilt; 3088 while( !(ilt = get_loop(n))->_head ) { 3089 // Normally I would use a set_loop here. But in this one special 3090 // case, it is legal (and expected) to change what loop a Node 3091 // belongs to. 3092 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); 3093 } 3094 // Remove safepoints ONLY if I've already seen I don't need one. 3095 // (the old code here would yank a 2nd safepoint after seeing a 3096 // first one, even though the 1st did not dominate in the loop body 3097 // and thus could be avoided indefinitely) 3098 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && 3099 is_deleteable_safept(n)) { 3100 Node *in = n->in(TypeFunc::Control); 3101 lazy_replace(n,in); // Pull safepoint now 3102 if (ilt->_safepts != NULL) { 3103 ilt->_safepts->yank(n); 3104 } 3105 // Carry on with the recursion "as if" we are walking 3106 // only the control input 3107 if( !visited.test_set( in->_idx ) ) { 3108 worklist.push(in); // Visit this guy later, using worklist 3109 } 3110 // Get next node from nstack: 3111 // - skip n's inputs processing by setting i > cnt; 3112 // - we also will not call set_early_ctrl(n) since 3113 // has_node(n) == true (see the condition above). 3114 i = cnt + 1; 3115 } 3116 } 3117 } // if (i == 0) 3118 3119 // Visit all inputs 3120 bool done = true; // Assume all n's inputs will be processed 3121 while (i < cnt) { 3122 Node *in = n->in(i); 3123 ++i; 3124 if (in == NULL) continue; 3125 if (in->pinned() && !in->is_CFG()) 3126 set_ctrl(in, in->in(0)); 3127 int is_visited = visited.test_set( in->_idx ); 3128 if (!has_node(in)) { // No controlling input yet? 3129 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); 3130 assert( !is_visited, "visit only once" ); 3131 nstack.push(n, i); // Save parent node and next input's index. 3132 nstack_top_n = in; // Process current input now. 3133 nstack_top_i = 0; 3134 done = false; // Not all n's inputs processed. 3135 break; // continue while_nstack_nonempty; 3136 } else if (!is_visited) { 3137 // This guy has a location picked out for him, but has not yet 3138 // been visited. Happens to all CFG nodes, for instance. 3139 // Visit him using the worklist instead of recursion, to break 3140 // cycles. Since he has a location already we do not need to 3141 // find his location before proceeding with the current Node. 3142 worklist.push(in); // Visit this guy later, using worklist 3143 } 3144 } 3145 if (done) { 3146 // All of n's inputs have been processed, complete post-processing. 3147 3148 // Compute earliest point this Node can go. 3149 // CFG, Phi, pinned nodes already know their controlling input. 3150 if (!has_node(n)) { 3151 // Record earliest legal location 3152 set_early_ctrl( n ); 3153 } 3154 if (nstack.is_empty()) { 3155 // Finished all nodes on stack. 3156 // Process next node on the worklist. 3157 break; 3158 } 3159 // Get saved parent node and next input's index. 3160 nstack_top_n = nstack.node(); 3161 nstack_top_i = nstack.index(); 3162 nstack.pop(); 3163 } 3164 } // while (true) 3165 } 3166 } 3167 3168 //------------------------------dom_lca_internal-------------------------------- 3169 // Pair-wise LCA 3170 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { 3171 if( !n1 ) return n2; // Handle NULL original LCA 3172 assert( n1->is_CFG(), "" ); 3173 assert( n2->is_CFG(), "" ); 3174 // find LCA of all uses 3175 uint d1 = dom_depth(n1); 3176 uint d2 = dom_depth(n2); 3177 while (n1 != n2) { 3178 if (d1 > d2) { 3179 n1 = idom(n1); 3180 d1 = dom_depth(n1); 3181 } else if (d1 < d2) { 3182 n2 = idom(n2); 3183 d2 = dom_depth(n2); 3184 } else { 3185 // Here d1 == d2. Due to edits of the dominator-tree, sections 3186 // of the tree might have the same depth. These sections have 3187 // to be searched more carefully. 3188 3189 // Scan up all the n1's with equal depth, looking for n2. 3190 Node *t1 = idom(n1); 3191 while (dom_depth(t1) == d1) { 3192 if (t1 == n2) return n2; 3193 t1 = idom(t1); 3194 } 3195 // Scan up all the n2's with equal depth, looking for n1. 3196 Node *t2 = idom(n2); 3197 while (dom_depth(t2) == d2) { 3198 if (t2 == n1) return n1; 3199 t2 = idom(t2); 3200 } 3201 // Move up to a new dominator-depth value as well as up the dom-tree. 3202 n1 = t1; 3203 n2 = t2; 3204 d1 = dom_depth(n1); 3205 d2 = dom_depth(n2); 3206 } 3207 } 3208 return n1; 3209 } 3210 3211 //------------------------------compute_idom----------------------------------- 3212 // Locally compute IDOM using dom_lca call. Correct only if the incoming 3213 // IDOMs are correct. 3214 Node *PhaseIdealLoop::compute_idom( Node *region ) const { 3215 assert( region->is_Region(), "" ); 3216 Node *LCA = NULL; 3217 for( uint i = 1; i < region->req(); i++ ) { 3218 if( region->in(i) != C->top() ) 3219 LCA = dom_lca( LCA, region->in(i) ); 3220 } 3221 return LCA; 3222 } 3223 3224 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { 3225 bool had_error = false; 3226 #ifdef ASSERT 3227 if (early != C->root()) { 3228 // Make sure that there's a dominance path from LCA to early 3229 Node* d = LCA; 3230 while (d != early) { 3231 if (d == C->root()) { 3232 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA); 3233 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx); 3234 had_error = true; 3235 break; 3236 } 3237 d = idom(d); 3238 } 3239 } 3240 #endif 3241 return had_error; 3242 } 3243 3244 3245 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { 3246 // Compute LCA over list of uses 3247 bool had_error = false; 3248 Node *LCA = NULL; 3249 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { 3250 Node* c = n->fast_out(i); 3251 if (_nodes[c->_idx] == NULL) 3252 continue; // Skip the occasional dead node 3253 if( c->is_Phi() ) { // For Phis, we must land above on the path 3254 for( uint j=1; j<c->req(); j++ ) {// For all inputs 3255 if( c->in(j) == n ) { // Found matching input? 3256 Node *use = c->in(0)->in(j); 3257 if (_verify_only && use->is_top()) continue; 3258 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3259 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3260 } 3261 } 3262 } else { 3263 // For CFG data-users, use is in the block just prior 3264 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); 3265 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3266 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3267 } 3268 } 3269 assert(!had_error, "bad dominance"); 3270 return LCA; 3271 } 3272 3273 //------------------------------get_late_ctrl---------------------------------- 3274 // Compute latest legal control. 3275 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { 3276 assert(early != NULL, "early control should not be NULL"); 3277 3278 Node* LCA = compute_lca_of_uses(n, early); 3279 #ifdef ASSERT 3280 if (LCA == C->root() && LCA != early) { 3281 // def doesn't dominate uses so print some useful debugging output 3282 compute_lca_of_uses(n, early, true); 3283 } 3284 #endif 3285 3286 // if this is a load, check for anti-dependent stores 3287 // We use a conservative algorithm to identify potential interfering 3288 // instructions and for rescheduling the load. The users of the memory 3289 // input of this load are examined. Any use which is not a load and is 3290 // dominated by early is considered a potentially interfering store. 3291 // This can produce false positives. 3292 if (n->is_Load() && LCA != early) { 3293 Node_List worklist; 3294 3295 Node *mem = n->in(MemNode::Memory); 3296 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 3297 Node* s = mem->fast_out(i); 3298 worklist.push(s); 3299 } 3300 while(worklist.size() != 0 && LCA != early) { 3301 Node* s = worklist.pop(); 3302 if (s->is_Load()) { 3303 continue; 3304 } else if (s->is_MergeMem()) { 3305 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 3306 Node* s1 = s->fast_out(i); 3307 worklist.push(s1); 3308 } 3309 } else { 3310 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); 3311 assert(sctrl != NULL || s->outcnt() == 0, "must have control"); 3312 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { 3313 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); 3314 } 3315 } 3316 } 3317 } 3318 3319 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); 3320 return LCA; 3321 } 3322 3323 // true if CFG node d dominates CFG node n 3324 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { 3325 if (d == n) 3326 return true; 3327 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 3328 uint dd = dom_depth(d); 3329 while (dom_depth(n) >= dd) { 3330 if (n == d) 3331 return true; 3332 n = idom(n); 3333 } 3334 return false; 3335 } 3336 3337 //------------------------------dom_lca_for_get_late_ctrl_internal------------- 3338 // Pair-wise LCA with tags. 3339 // Tag each index with the node 'tag' currently being processed 3340 // before advancing up the dominator chain using idom(). 3341 // Later calls that find a match to 'tag' know that this path has already 3342 // been considered in the current LCA (which is input 'n1' by convention). 3343 // Since get_late_ctrl() is only called once for each node, the tag array 3344 // does not need to be cleared between calls to get_late_ctrl(). 3345 // Algorithm trades a larger constant factor for better asymptotic behavior 3346 // 3347 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { 3348 uint d1 = dom_depth(n1); 3349 uint d2 = dom_depth(n2); 3350 3351 do { 3352 if (d1 > d2) { 3353 // current lca is deeper than n2 3354 _dom_lca_tags.map(n1->_idx, tag); 3355 n1 = idom(n1); 3356 d1 = dom_depth(n1); 3357 } else if (d1 < d2) { 3358 // n2 is deeper than current lca 3359 Node *memo = _dom_lca_tags[n2->_idx]; 3360 if( memo == tag ) { 3361 return n1; // Return the current LCA 3362 } 3363 _dom_lca_tags.map(n2->_idx, tag); 3364 n2 = idom(n2); 3365 d2 = dom_depth(n2); 3366 } else { 3367 // Here d1 == d2. Due to edits of the dominator-tree, sections 3368 // of the tree might have the same depth. These sections have 3369 // to be searched more carefully. 3370 3371 // Scan up all the n1's with equal depth, looking for n2. 3372 _dom_lca_tags.map(n1->_idx, tag); 3373 Node *t1 = idom(n1); 3374 while (dom_depth(t1) == d1) { 3375 if (t1 == n2) return n2; 3376 _dom_lca_tags.map(t1->_idx, tag); 3377 t1 = idom(t1); 3378 } 3379 // Scan up all the n2's with equal depth, looking for n1. 3380 _dom_lca_tags.map(n2->_idx, tag); 3381 Node *t2 = idom(n2); 3382 while (dom_depth(t2) == d2) { 3383 if (t2 == n1) return n1; 3384 _dom_lca_tags.map(t2->_idx, tag); 3385 t2 = idom(t2); 3386 } 3387 // Move up to a new dominator-depth value as well as up the dom-tree. 3388 n1 = t1; 3389 n2 = t2; 3390 d1 = dom_depth(n1); 3391 d2 = dom_depth(n2); 3392 } 3393 } while (n1 != n2); 3394 return n1; 3395 } 3396 3397 //------------------------------init_dom_lca_tags------------------------------ 3398 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3399 // Intended use does not involve any growth for the array, so it could 3400 // be of fixed size. 3401 void PhaseIdealLoop::init_dom_lca_tags() { 3402 uint limit = C->unique() + 1; 3403 _dom_lca_tags.map( limit, NULL ); 3404 #ifdef ASSERT 3405 for( uint i = 0; i < limit; ++i ) { 3406 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3407 } 3408 #endif // ASSERT 3409 } 3410 3411 //------------------------------clear_dom_lca_tags------------------------------ 3412 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3413 // Intended use does not involve any growth for the array, so it could 3414 // be of fixed size. 3415 void PhaseIdealLoop::clear_dom_lca_tags() { 3416 uint limit = C->unique() + 1; 3417 _dom_lca_tags.map( limit, NULL ); 3418 _dom_lca_tags.clear(); 3419 #ifdef ASSERT 3420 for( uint i = 0; i < limit; ++i ) { 3421 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3422 } 3423 #endif // ASSERT 3424 } 3425 3426 //------------------------------build_loop_late-------------------------------- 3427 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3428 // Second pass finds latest legal placement, and ideal loop placement. 3429 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3430 while (worklist.size() != 0) { 3431 Node *n = worklist.pop(); 3432 // Only visit once 3433 if (visited.test_set(n->_idx)) continue; 3434 uint cnt = n->outcnt(); 3435 uint i = 0; 3436 while (true) { 3437 assert( _nodes[n->_idx], "no dead nodes" ); 3438 // Visit all children 3439 if (i < cnt) { 3440 Node* use = n->raw_out(i); 3441 ++i; 3442 // Check for dead uses. Aggressively prune such junk. It might be 3443 // dead in the global sense, but still have local uses so I cannot 3444 // easily call 'remove_dead_node'. 3445 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? 3446 // Due to cycles, we might not hit the same fixed point in the verify 3447 // pass as we do in the regular pass. Instead, visit such phis as 3448 // simple uses of the loop head. 3449 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { 3450 if( !visited.test(use->_idx) ) 3451 worklist.push(use); 3452 } else if( !visited.test_set(use->_idx) ) { 3453 nstack.push(n, i); // Save parent and next use's index. 3454 n = use; // Process all children of current use. 3455 cnt = use->outcnt(); 3456 i = 0; 3457 } 3458 } else { 3459 // Do not visit around the backedge of loops via data edges. 3460 // push dead code onto a worklist 3461 _deadlist.push(use); 3462 } 3463 } else { 3464 // All of n's children have been processed, complete post-processing. 3465 build_loop_late_post(n); 3466 if (nstack.is_empty()) { 3467 // Finished all nodes on stack. 3468 // Process next node on the worklist. 3469 break; 3470 } 3471 // Get saved parent node and next use's index. Visit the rest of uses. 3472 n = nstack.node(); 3473 cnt = n->outcnt(); 3474 i = nstack.index(); 3475 nstack.pop(); 3476 } 3477 } 3478 } 3479 } 3480 3481 //------------------------------build_loop_late_post--------------------------- 3482 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3483 // Second pass finds latest legal placement, and ideal loop placement. 3484 void PhaseIdealLoop::build_loop_late_post( Node *n ) { 3485 3486 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) { 3487 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. 3488 } 3489 3490 #ifdef ASSERT 3491 if (_verify_only && !n->is_CFG()) { 3492 // Check def-use domination. 3493 compute_lca_of_uses(n, get_ctrl(n), true /* verify */); 3494 } 3495 #endif 3496 3497 // CFG and pinned nodes already handled 3498 if( n->in(0) ) { 3499 if( n->in(0)->is_top() ) return; // Dead? 3500 3501 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads 3502 // _must_ be pinned (they have to observe their control edge of course). 3503 // Unlike Stores (which modify an unallocable resource, the memory 3504 // state), Mods/Loads can float around. So free them up. 3505 bool pinned = true; 3506 switch( n->Opcode() ) { 3507 case Op_DivI: 3508 case Op_DivF: 3509 case Op_DivD: 3510 case Op_ModI: 3511 case Op_ModF: 3512 case Op_ModD: 3513 case Op_LoadB: // Same with Loads; they can sink 3514 case Op_LoadUB: // during loop optimizations. 3515 case Op_LoadUS: 3516 case Op_LoadD: 3517 case Op_LoadF: 3518 case Op_LoadI: 3519 case Op_LoadKlass: 3520 case Op_LoadNKlass: 3521 case Op_LoadL: 3522 case Op_LoadS: 3523 case Op_LoadP: 3524 case Op_LoadN: 3525 case Op_LoadRange: 3526 case Op_LoadD_unaligned: 3527 case Op_LoadL_unaligned: 3528 case Op_StrComp: // Does a bunch of load-like effects 3529 case Op_StrEquals: 3530 case Op_StrIndexOf: 3531 case Op_StrIndexOfChar: 3532 case Op_AryEq: 3533 case Op_HasNegatives: 3534 pinned = false; 3535 } 3536 if( pinned ) { 3537 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); 3538 if( !chosen_loop->_child ) // Inner loop? 3539 chosen_loop->_body.push(n); // Collect inner loops 3540 return; 3541 } 3542 } else { // No slot zero 3543 if( n->is_CFG() ) { // CFG with no slot 0 is dead 3544 _nodes.map(n->_idx,0); // No block setting, it's globally dead 3545 return; 3546 } 3547 assert(!n->is_CFG() || n->outcnt() == 0, ""); 3548 } 3549 3550 // Do I have a "safe range" I can select over? 3551 Node *early = get_ctrl(n);// Early location already computed 3552 3553 // Compute latest point this Node can go 3554 Node *LCA = get_late_ctrl( n, early ); 3555 // LCA is NULL due to uses being dead 3556 if( LCA == NULL ) { 3557 #ifdef ASSERT 3558 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { 3559 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); 3560 } 3561 #endif 3562 _nodes.map(n->_idx, 0); // This node is useless 3563 _deadlist.push(n); 3564 return; 3565 } 3566 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); 3567 3568 Node *legal = LCA; // Walk 'legal' up the IDOM chain 3569 Node *least = legal; // Best legal position so far 3570 while( early != legal ) { // While not at earliest legal 3571 #ifdef ASSERT 3572 if (legal->is_Start() && !early->is_Root()) { 3573 // Bad graph. Print idom path and fail. 3574 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA); 3575 assert(false, "Bad graph detected in build_loop_late"); 3576 } 3577 #endif 3578 // Find least loop nesting depth 3579 legal = idom(legal); // Bump up the IDOM tree 3580 // Check for lower nesting depth 3581 if( get_loop(legal)->_nest < get_loop(least)->_nest ) 3582 least = legal; 3583 } 3584 assert(early == legal || legal != C->root(), "bad dominance of inputs"); 3585 3586 // Try not to place code on a loop entry projection 3587 // which can inhibit range check elimination. 3588 if (least != early) { 3589 Node* ctrl_out = least->unique_ctrl_out(); 3590 if (ctrl_out && ctrl_out->is_CountedLoop() && 3591 least == ctrl_out->in(LoopNode::EntryControl)) { 3592 Node* least_dom = idom(least); 3593 if (get_loop(least_dom)->is_member(get_loop(least))) { 3594 least = least_dom; 3595 } 3596 } 3597 } 3598 3599 #ifdef ASSERT 3600 // If verifying, verify that 'verify_me' has a legal location 3601 // and choose it as our location. 3602 if( _verify_me ) { 3603 Node *v_ctrl = _verify_me->get_ctrl_no_update(n); 3604 Node *legal = LCA; 3605 while( early != legal ) { // While not at earliest legal 3606 if( legal == v_ctrl ) break; // Check for prior good location 3607 legal = idom(legal) ;// Bump up the IDOM tree 3608 } 3609 // Check for prior good location 3610 if( legal == v_ctrl ) least = legal; // Keep prior if found 3611 } 3612 #endif 3613 3614 // Assign discovered "here or above" point 3615 least = find_non_split_ctrl(least); 3616 set_ctrl(n, least); 3617 3618 // Collect inner loop bodies 3619 IdealLoopTree *chosen_loop = get_loop(least); 3620 if( !chosen_loop->_child ) // Inner loop? 3621 chosen_loop->_body.push(n);// Collect inner loops 3622 } 3623 3624 #ifdef ASSERT 3625 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) { 3626 tty->print_cr("%s", msg); 3627 tty->print("n: "); n->dump(); 3628 tty->print("early(n): "); early->dump(); 3629 if (n->in(0) != NULL && !n->in(0)->is_top() && 3630 n->in(0) != early && !n->in(0)->is_Root()) { 3631 tty->print("n->in(0): "); n->in(0)->dump(); 3632 } 3633 for (uint i = 1; i < n->req(); i++) { 3634 Node* in1 = n->in(i); 3635 if (in1 != NULL && in1 != n && !in1->is_top()) { 3636 tty->print("n->in(%d): ", i); in1->dump(); 3637 Node* in1_early = get_ctrl(in1); 3638 tty->print("early(n->in(%d)): ", i); in1_early->dump(); 3639 if (in1->in(0) != NULL && !in1->in(0)->is_top() && 3640 in1->in(0) != in1_early && !in1->in(0)->is_Root()) { 3641 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump(); 3642 } 3643 for (uint j = 1; j < in1->req(); j++) { 3644 Node* in2 = in1->in(j); 3645 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) { 3646 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump(); 3647 Node* in2_early = get_ctrl(in2); 3648 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); 3649 if (in2->in(0) != NULL && !in2->in(0)->is_top() && 3650 in2->in(0) != in2_early && !in2->in(0)->is_Root()) { 3651 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump(); 3652 } 3653 } 3654 } 3655 } 3656 } 3657 tty->cr(); 3658 tty->print("LCA(n): "); LCA->dump(); 3659 for (uint i = 0; i < n->outcnt(); i++) { 3660 Node* u1 = n->raw_out(i); 3661 if (u1 == n) 3662 continue; 3663 tty->print("n->out(%d): ", i); u1->dump(); 3664 if (u1->is_CFG()) { 3665 for (uint j = 0; j < u1->outcnt(); j++) { 3666 Node* u2 = u1->raw_out(j); 3667 if (u2 != u1 && u2 != n && u2->is_CFG()) { 3668 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 3669 } 3670 } 3671 } else { 3672 Node* u1_later = get_ctrl(u1); 3673 tty->print("later(n->out(%d)): ", i); u1_later->dump(); 3674 if (u1->in(0) != NULL && !u1->in(0)->is_top() && 3675 u1->in(0) != u1_later && !u1->in(0)->is_Root()) { 3676 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump(); 3677 } 3678 for (uint j = 0; j < u1->outcnt(); j++) { 3679 Node* u2 = u1->raw_out(j); 3680 if (u2 == n || u2 == u1) 3681 continue; 3682 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 3683 if (!u2->is_CFG()) { 3684 Node* u2_later = get_ctrl(u2); 3685 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); 3686 if (u2->in(0) != NULL && !u2->in(0)->is_top() && 3687 u2->in(0) != u2_later && !u2->in(0)->is_Root()) { 3688 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump(); 3689 } 3690 } 3691 } 3692 } 3693 } 3694 tty->cr(); 3695 int ct = 0; 3696 Node *dbg_legal = LCA; 3697 while(!dbg_legal->is_Start() && ct < 100) { 3698 tty->print("idom[%d] ",ct); dbg_legal->dump(); 3699 ct++; 3700 dbg_legal = idom(dbg_legal); 3701 } 3702 tty->cr(); 3703 } 3704 #endif 3705 3706 #ifndef PRODUCT 3707 //------------------------------dump------------------------------------------- 3708 void PhaseIdealLoop::dump( ) const { 3709 ResourceMark rm; 3710 Arena* arena = Thread::current()->resource_area(); 3711 Node_Stack stack(arena, C->live_nodes() >> 2); 3712 Node_List rpo_list; 3713 VectorSet visited(arena); 3714 visited.set(C->top()->_idx); 3715 rpo( C->root(), stack, visited, rpo_list ); 3716 // Dump root loop indexed by last element in PO order 3717 dump( _ltree_root, rpo_list.size(), rpo_list ); 3718 } 3719 3720 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { 3721 loop->dump_head(); 3722 3723 // Now scan for CFG nodes in the same loop 3724 for( uint j=idx; j > 0; j-- ) { 3725 Node *n = rpo_list[j-1]; 3726 if( !_nodes[n->_idx] ) // Skip dead nodes 3727 continue; 3728 if( get_loop(n) != loop ) { // Wrong loop nest 3729 if( get_loop(n)->_head == n && // Found nested loop? 3730 get_loop(n)->_parent == loop ) 3731 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly 3732 continue; 3733 } 3734 3735 // Dump controlling node 3736 for( uint x = 0; x < loop->_nest; x++ ) 3737 tty->print(" "); 3738 tty->print("C"); 3739 if( n == C->root() ) { 3740 n->dump(); 3741 } else { 3742 Node* cached_idom = idom_no_update(n); 3743 Node *computed_idom = n->in(0); 3744 if( n->is_Region() ) { 3745 computed_idom = compute_idom(n); 3746 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or 3747 // any MultiBranch ctrl node), so apply a similar transform to 3748 // the cached idom returned from idom_no_update. 3749 cached_idom = find_non_split_ctrl(cached_idom); 3750 } 3751 tty->print(" ID:%d",computed_idom->_idx); 3752 n->dump(); 3753 if( cached_idom != computed_idom ) { 3754 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", 3755 computed_idom->_idx, cached_idom->_idx); 3756 } 3757 } 3758 // Dump nodes it controls 3759 for( uint k = 0; k < _nodes.Size(); k++ ) { 3760 // (k < C->unique() && get_ctrl(find(k)) == n) 3761 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { 3762 Node *m = C->root()->find(k); 3763 if( m && m->outcnt() > 0 ) { 3764 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { 3765 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", 3766 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); 3767 } 3768 for( uint j = 0; j < loop->_nest; j++ ) 3769 tty->print(" "); 3770 tty->print(" "); 3771 m->dump(); 3772 } 3773 } 3774 } 3775 } 3776 } 3777 3778 // Collect a R-P-O for the whole CFG. 3779 // Result list is in post-order (scan backwards for RPO) 3780 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { 3781 stk.push(start, 0); 3782 visited.set(start->_idx); 3783 3784 while (stk.is_nonempty()) { 3785 Node* m = stk.node(); 3786 uint idx = stk.index(); 3787 if (idx < m->outcnt()) { 3788 stk.set_index(idx + 1); 3789 Node* n = m->raw_out(idx); 3790 if (n->is_CFG() && !visited.test_set(n->_idx)) { 3791 stk.push(n, 0); 3792 } 3793 } else { 3794 rpo_list.push(m); 3795 stk.pop(); 3796 } 3797 } 3798 } 3799 #endif 3800 3801 3802 //============================================================================= 3803 //------------------------------LoopTreeIterator----------------------------------- 3804 3805 // Advance to next loop tree using a preorder, left-to-right traversal. 3806 void LoopTreeIterator::next() { 3807 assert(!done(), "must not be done."); 3808 if (_curnt->_child != NULL) { 3809 _curnt = _curnt->_child; 3810 } else if (_curnt->_next != NULL) { 3811 _curnt = _curnt->_next; 3812 } else { 3813 while (_curnt != _root && _curnt->_next == NULL) { 3814 _curnt = _curnt->_parent; 3815 } 3816 if (_curnt == _root) { 3817 _curnt = NULL; 3818 assert(done(), "must be done."); 3819 } else { 3820 assert(_curnt->_next != NULL, "must be more to do"); 3821 _curnt = _curnt->_next; 3822 } 3823 } 3824 }