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