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