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