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