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