1 /* 2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciMethodData.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "libadt/vectset.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "memory/resourceArea.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/callnode.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/divnode.hpp" 36 #include "opto/idealGraphPrinter.hpp" 37 #include "opto/loopnode.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/superword.hpp" 41 42 //============================================================================= 43 //------------------------------is_loop_iv------------------------------------- 44 // Determine if a node is Counted loop induction variable. 45 // The method is declared in node.hpp. 46 const Node* Node::is_loop_iv() const { 47 if (this->is_Phi() && !this->as_Phi()->is_copy() && 48 this->as_Phi()->region()->is_CountedLoop() && 49 this->as_Phi()->region()->as_CountedLoop()->phi() == this) { 50 return this; 51 } else { 52 return NULL; 53 } 54 } 55 56 //============================================================================= 57 //------------------------------dump_spec-------------------------------------- 58 // Dump special per-node info 59 #ifndef PRODUCT 60 void LoopNode::dump_spec(outputStream *st) const { 61 if (is_inner_loop()) st->print( "inner " ); 62 if (is_partial_peel_loop()) st->print( "partial_peel " ); 63 if (partial_peel_has_failed()) st->print( "partial_peel_failed " ); 64 } 65 #endif 66 67 //------------------------------is_valid_counted_loop------------------------- 68 bool LoopNode::is_valid_counted_loop() const { 69 if (is_CountedLoop()) { 70 CountedLoopNode* l = as_CountedLoop(); 71 CountedLoopEndNode* le = l->loopexit_or_null(); 72 if (le != NULL && 73 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) { 74 Node* phi = l->phi(); 75 Node* exit = le->proj_out_or_null(0 /* false */); 76 if (exit != NULL && exit->Opcode() == Op_IfFalse && 77 phi != NULL && phi->is_Phi() && 78 phi->in(LoopNode::LoopBackControl) == l->incr() && 79 le->loopnode() == l && le->stride_is_con()) { 80 return true; 81 } 82 } 83 } 84 return false; 85 } 86 87 //------------------------------get_early_ctrl--------------------------------- 88 // Compute earliest legal control 89 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { 90 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); 91 uint i; 92 Node *early; 93 if (n->in(0) && !n->is_expensive()) { 94 early = n->in(0); 95 if (!early->is_CFG()) // Might be a non-CFG multi-def 96 early = get_ctrl(early); // So treat input as a straight data input 97 i = 1; 98 } else { 99 early = get_ctrl(n->in(1)); 100 i = 2; 101 } 102 uint e_d = dom_depth(early); 103 assert( early, "" ); 104 for (; i < n->req(); i++) { 105 Node *cin = get_ctrl(n->in(i)); 106 assert( cin, "" ); 107 // Keep deepest dominator depth 108 uint c_d = dom_depth(cin); 109 if (c_d > e_d) { // Deeper guy? 110 early = cin; // Keep deepest found so far 111 e_d = c_d; 112 } else if (c_d == e_d && // Same depth? 113 early != cin) { // If not equal, must use slower algorithm 114 // If same depth but not equal, one _must_ dominate the other 115 // and we want the deeper (i.e., dominated) guy. 116 Node *n1 = early; 117 Node *n2 = cin; 118 while (1) { 119 n1 = idom(n1); // Walk up until break cycle 120 n2 = idom(n2); 121 if (n1 == cin || // Walked early up to cin 122 dom_depth(n2) < c_d) 123 break; // early is deeper; keep him 124 if (n2 == early || // Walked cin up to early 125 dom_depth(n1) < c_d) { 126 early = cin; // cin is deeper; keep him 127 break; 128 } 129 } 130 e_d = dom_depth(early); // Reset depth register cache 131 } 132 } 133 134 // Return earliest legal location 135 assert(early == find_non_split_ctrl(early), "unexpected early control"); 136 137 if (n->is_expensive() && !_verify_only && !_verify_me) { 138 assert(n->in(0), "should have control input"); 139 early = get_early_ctrl_for_expensive(n, early); 140 } 141 142 return early; 143 } 144 145 //------------------------------get_early_ctrl_for_expensive--------------------------------- 146 // Move node up the dominator tree as high as legal while still beneficial 147 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) { 148 assert(n->in(0) && n->is_expensive(), "expensive node with control input here"); 149 assert(OptimizeExpensiveOps, "optimization off?"); 150 151 Node* ctl = n->in(0); 152 assert(ctl->is_CFG(), "expensive input 0 must be cfg"); 153 uint min_dom_depth = dom_depth(earliest); 154 #ifdef ASSERT 155 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) { 156 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl); 157 assert(false, "Bad graph detected in get_early_ctrl_for_expensive"); 158 } 159 #endif 160 if (dom_depth(ctl) < min_dom_depth) { 161 return earliest; 162 } 163 164 while (1) { 165 Node *next = ctl; 166 // Moving the node out of a loop on the projection of a If 167 // confuses loop predication. So once we hit a Loop in a If branch 168 // that doesn't branch to an UNC, we stop. The code that process 169 // expensive nodes will notice the loop and skip over it to try to 170 // move the node further up. 171 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) { 172 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 173 break; 174 } 175 next = idom(ctl->in(1)->in(0)); 176 } else if (ctl->is_Proj()) { 177 // We only move it up along a projection if the projection is 178 // the single control projection for its parent: same code path, 179 // if it's a If with UNC or fallthrough of a call. 180 Node* parent_ctl = ctl->in(0); 181 if (parent_ctl == NULL) { 182 break; 183 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) { 184 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control(); 185 } else if (parent_ctl->is_If()) { 186 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 187 break; 188 } 189 assert(idom(ctl) == parent_ctl, "strange"); 190 next = idom(parent_ctl); 191 } else if (ctl->is_CatchProj()) { 192 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) { 193 break; 194 } 195 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph"); 196 next = parent_ctl->in(0)->in(0)->in(0); 197 } else { 198 // Check if parent control has a single projection (this 199 // control is the only possible successor of the parent 200 // control). If so, we can try to move the node above the 201 // parent control. 202 int nb_ctl_proj = 0; 203 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) { 204 Node *p = parent_ctl->fast_out(i); 205 if (p->is_Proj() && p->is_CFG()) { 206 nb_ctl_proj++; 207 if (nb_ctl_proj > 1) { 208 break; 209 } 210 } 211 } 212 213 if (nb_ctl_proj > 1) { 214 break; 215 } 216 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node"); 217 assert(idom(ctl) == parent_ctl, "strange"); 218 next = idom(parent_ctl); 219 } 220 } else { 221 next = idom(ctl); 222 } 223 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) { 224 break; 225 } 226 ctl = next; 227 } 228 229 if (ctl != n->in(0)) { 230 _igvn.replace_input_of(n, 0, ctl); 231 _igvn.hash_insert(n); 232 } 233 234 return ctl; 235 } 236 237 238 //------------------------------set_early_ctrl--------------------------------- 239 // Set earliest legal control 240 void PhaseIdealLoop::set_early_ctrl( Node *n ) { 241 Node *early = get_early_ctrl(n); 242 243 // Record earliest legal location 244 set_ctrl(n, early); 245 } 246 247 //------------------------------set_subtree_ctrl------------------------------- 248 // set missing _ctrl entries on new nodes 249 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { 250 // Already set? Get out. 251 if( _nodes[n->_idx] ) return; 252 // Recursively set _nodes array to indicate where the Node goes 253 uint i; 254 for( i = 0; i < n->req(); ++i ) { 255 Node *m = n->in(i); 256 if( m && m != C->root() ) 257 set_subtree_ctrl( m ); 258 } 259 260 // Fixup self 261 set_early_ctrl( n ); 262 } 263 264 // Create a skeleton strip mined outer loop: a Loop head before the 265 // inner strip mined loop, a safepoint and an exit condition guarded 266 // by an opaque node after the inner strip mined loop with a backedge 267 // to the loop head. The inner strip mined loop is left as it is. Only 268 // once loop optimizations are over, do we adjust the inner loop exit 269 // condition to limit its number of iterations, set the outer loop 270 // exit condition and add Phis to the outer loop head. Some loop 271 // optimizations that operate on the inner strip mined loop need to be 272 // aware of the outer strip mined loop: loop unswitching needs to 273 // clone the outer loop as well as the inner, unrolling needs to only 274 // clone the inner loop etc. No optimizations need to change the outer 275 // strip mined loop as it is only a skeleton. 276 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control, 277 IdealLoopTree* loop, float cl_prob, float le_fcnt, 278 Node*& entry_control, Node*& iffalse) { 279 Node* outer_test = _igvn.intcon(0); 280 set_ctrl(outer_test, C->root()); 281 Node *orig = iffalse; 282 iffalse = iffalse->clone(); 283 _igvn.register_new_node_with_optimizer(iffalse); 284 set_idom(iffalse, idom(orig), dom_depth(orig)); 285 286 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt); 287 Node *outer_ift = new IfTrueNode (outer_le); 288 Node* outer_iff = orig; 289 _igvn.replace_input_of(outer_iff, 0, outer_le); 290 291 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift); 292 entry_control = outer_l; 293 294 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift); 295 IdealLoopTree* parent = loop->_parent; 296 IdealLoopTree* sibling = parent->_child; 297 if (sibling == loop) { 298 parent->_child = outer_ilt; 299 } else { 300 while (sibling->_next != loop) { 301 sibling = sibling->_next; 302 } 303 sibling->_next = outer_ilt; 304 } 305 outer_ilt->_next = loop->_next; 306 outer_ilt->_parent = parent; 307 outer_ilt->_child = loop; 308 outer_ilt->_nest = loop->_nest; 309 loop->_parent = outer_ilt; 310 loop->_next = NULL; 311 loop->_nest++; 312 313 set_loop(iffalse, outer_ilt); 314 register_control(outer_le, outer_ilt, iffalse); 315 register_control(outer_ift, outer_ilt, outer_le); 316 set_idom(outer_iff, outer_le, dom_depth(outer_le)); 317 _igvn.register_new_node_with_optimizer(outer_l); 318 set_loop(outer_l, outer_ilt); 319 set_idom(outer_l, init_control, dom_depth(init_control)+1); 320 321 return outer_ilt; 322 } 323 324 //------------------------------is_counted_loop-------------------------------- 325 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) { 326 PhaseGVN *gvn = &_igvn; 327 328 // Counted loop head must be a good RegionNode with only 3 not NULL 329 // control input edges: Self, Entry, LoopBack. 330 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) { 331 return false; 332 } 333 Node *init_control = x->in(LoopNode::EntryControl); 334 Node *back_control = x->in(LoopNode::LoopBackControl); 335 if (init_control == NULL || back_control == NULL) // Partially dead 336 return false; 337 // Must also check for TOP when looking for a dead loop 338 if (init_control->is_top() || back_control->is_top()) 339 return false; 340 341 // Allow funny placement of Safepoint 342 if (back_control->Opcode() == Op_SafePoint) { 343 if (LoopStripMiningIter != 0) { 344 // Leaving the safepoint on the backedge and creating a 345 // CountedLoop will confuse optimizations. We can't move the 346 // safepoint around because its jvm state wouldn't match a new 347 // location. Give up on that loop. 348 return false; 349 } 350 back_control = back_control->in(TypeFunc::Control); 351 } 352 353 // Controlling test for loop 354 Node *iftrue = back_control; 355 uint iftrue_op = iftrue->Opcode(); 356 if (iftrue_op != Op_IfTrue && 357 iftrue_op != Op_IfFalse) 358 // I have a weird back-control. Probably the loop-exit test is in 359 // the middle of the loop and I am looking at some trailing control-flow 360 // merge point. To fix this I would have to partially peel the loop. 361 return false; // Obscure back-control 362 363 // Get boolean guarding loop-back test 364 Node *iff = iftrue->in(0); 365 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) 366 return false; 367 BoolNode *test = iff->in(1)->as_Bool(); 368 BoolTest::mask bt = test->_test._test; 369 float cl_prob = iff->as_If()->_prob; 370 if (iftrue_op == Op_IfFalse) { 371 bt = BoolTest(bt).negate(); 372 cl_prob = 1.0 - cl_prob; 373 } 374 // Get backedge compare 375 Node *cmp = test->in(1); 376 int cmp_op = cmp->Opcode(); 377 if (cmp_op != Op_CmpI) 378 return false; // Avoid pointer & float compares 379 380 // Find the trip-counter increment & limit. Limit must be loop invariant. 381 Node *incr = cmp->in(1); 382 Node *limit = cmp->in(2); 383 384 // --------- 385 // need 'loop()' test to tell if limit is loop invariant 386 // --------- 387 388 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit? 389 Node *tmp = incr; // Then reverse order into the CmpI 390 incr = limit; 391 limit = tmp; 392 bt = BoolTest(bt).commute(); // And commute the exit test 393 } 394 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant 395 return false; 396 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 397 return false; 398 399 Node* phi_incr = NULL; 400 // Trip-counter increment must be commutative & associative. 401 if (incr->Opcode() == Op_CastII) { 402 incr = incr->in(1); 403 } 404 if (incr->is_Phi()) { 405 if (incr->as_Phi()->region() != x || incr->req() != 3) 406 return false; // Not simple trip counter expression 407 phi_incr = incr; 408 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi 409 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 410 return false; 411 } 412 413 Node* trunc1 = NULL; 414 Node* trunc2 = NULL; 415 const TypeInt* iv_trunc_t = NULL; 416 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { 417 return false; // Funny increment opcode 418 } 419 assert(incr->Opcode() == Op_AddI, "wrong increment code"); 420 421 // Get merge point 422 Node *xphi = incr->in(1); 423 Node *stride = incr->in(2); 424 if (!stride->is_Con()) { // Oops, swap these 425 if (!xphi->is_Con()) // Is the other guy a constant? 426 return false; // Nope, unknown stride, bail out 427 Node *tmp = xphi; // 'incr' is commutative, so ok to swap 428 xphi = stride; 429 stride = tmp; 430 } 431 if (xphi->Opcode() == Op_CastII) { 432 xphi = xphi->in(1); 433 } 434 // Stride must be constant 435 int stride_con = stride->get_int(); 436 if (stride_con == 0) 437 return false; // missed some peephole opt 438 439 if (!xphi->is_Phi()) 440 return false; // Too much math on the trip counter 441 if (phi_incr != NULL && phi_incr != xphi) 442 return false; 443 PhiNode *phi = xphi->as_Phi(); 444 445 // Phi must be of loop header; backedge must wrap to increment 446 if (phi->region() != x) 447 return false; 448 if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) || 449 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) { 450 return false; 451 } 452 Node *init_trip = phi->in(LoopNode::EntryControl); 453 454 // If iv trunc type is smaller than int, check for possible wrap. 455 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { 456 assert(trunc1 != NULL, "must have found some truncation"); 457 458 // Get a better type for the phi (filtered thru if's) 459 const TypeInt* phi_ft = filtered_type(phi); 460 461 // Can iv take on a value that will wrap? 462 // 463 // Ensure iv's limit is not within "stride" of the wrap value. 464 // 465 // Example for "short" type 466 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) 467 // If the stride is +10, then the last value of the induction 468 // variable before the increment (phi_ft->_hi) must be 469 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to 470 // ensure no truncation occurs after the increment. 471 472 if (stride_con > 0) { 473 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || 474 iv_trunc_t->_lo > phi_ft->_lo) { 475 return false; // truncation may occur 476 } 477 } else if (stride_con < 0) { 478 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || 479 iv_trunc_t->_hi < phi_ft->_hi) { 480 return false; // truncation may occur 481 } 482 } 483 // No possibility of wrap so truncation can be discarded 484 // Promote iv type to Int 485 } else { 486 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); 487 } 488 489 // If the condition is inverted and we will be rolling 490 // through MININT to MAXINT, then bail out. 491 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice! 492 // Odd stride 493 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) || 494 // Count down loop rolls through MAXINT 495 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) || 496 // Count up loop rolls through MININT 497 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) { 498 return false; // Bail out 499 } 500 501 const TypeInt* init_t = gvn->type(init_trip)->is_int(); 502 const TypeInt* limit_t = gvn->type(limit)->is_int(); 503 504 if (stride_con > 0) { 505 jlong init_p = (jlong)init_t->_lo + stride_con; 506 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi) 507 return false; // cyclic loop or this loop trips only once 508 } else { 509 jlong init_p = (jlong)init_t->_hi + stride_con; 510 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo) 511 return false; // cyclic loop or this loop trips only once 512 } 513 514 if (phi_incr != NULL) { 515 // check if there is a possiblity of IV overflowing after the first increment 516 if (stride_con > 0) { 517 if (init_t->_hi > max_jint - stride_con) { 518 return false; 519 } 520 } else { 521 if (init_t->_lo < min_jint - stride_con) { 522 return false; 523 } 524 } 525 } 526 527 // ================================================= 528 // ---- SUCCESS! Found A Trip-Counted Loop! ----- 529 // 530 assert(x->Opcode() == Op_Loop, "regular loops only"); 531 C->print_method(PHASE_BEFORE_CLOOPS, 3); 532 533 Node *hook = new Node(6); 534 535 // =================================================== 536 // Generate loop limit check to avoid integer overflow 537 // in cases like next (cyclic loops): 538 // 539 // for (i=0; i <= max_jint; i++) {} 540 // for (i=0; i < max_jint; i+=2) {} 541 // 542 // 543 // Limit check predicate depends on the loop test: 544 // 545 // for(;i != limit; i++) --> limit <= (max_jint) 546 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1) 547 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride ) 548 // 549 550 // Check if limit is excluded to do more precise int overflow check. 551 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge); 552 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1)); 553 554 // If compare points directly to the phi we need to adjust 555 // the compare so that it points to the incr. Limit have 556 // to be adjusted to keep trip count the same and the 557 // adjusted limit should be checked for int overflow. 558 if (phi_incr != NULL) { 559 stride_m += stride_con; 560 } 561 562 if (limit->is_Con()) { 563 int limit_con = limit->get_int(); 564 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) || 565 (stride_con < 0 && limit_con < (min_jint - stride_m))) { 566 // Bailout: it could be integer overflow. 567 return false; 568 } 569 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) || 570 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) { 571 // Limit's type may satisfy the condition, for example, 572 // when it is an array length. 573 } else { 574 // Generate loop's limit check. 575 // Loop limit check predicate should be near the loop. 576 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); 577 if (!limit_check_proj) { 578 // The limit check predicate is not generated if this method trapped here before. 579 #ifdef ASSERT 580 if (TraceLoopLimitCheck) { 581 tty->print("missing loop limit check:"); 582 loop->dump_head(); 583 x->dump(1); 584 } 585 #endif 586 return false; 587 } 588 589 IfNode* check_iff = limit_check_proj->in(0)->as_If(); 590 Node* cmp_limit; 591 Node* bol; 592 593 if (stride_con > 0) { 594 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m)); 595 bol = new BoolNode(cmp_limit, BoolTest::le); 596 } else { 597 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m)); 598 bol = new BoolNode(cmp_limit, BoolTest::ge); 599 } 600 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit); 601 bol = _igvn.register_new_node_with_optimizer(bol); 602 set_subtree_ctrl(bol); 603 604 // Replace condition in original predicate but preserve Opaque node 605 // so that previous predicates could be found. 606 assert(check_iff->in(1)->Opcode() == Op_Conv2B && 607 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, ""); 608 Node* opq = check_iff->in(1)->in(1); 609 _igvn.replace_input_of(opq, 1, bol); 610 // Update ctrl. 611 set_ctrl(opq, check_iff->in(0)); 612 set_ctrl(check_iff->in(1), check_iff->in(0)); 613 614 #ifndef PRODUCT 615 // report that the loop predication has been actually performed 616 // for this loop 617 if (TraceLoopLimitCheck) { 618 tty->print_cr("Counted Loop Limit Check generated:"); 619 debug_only( bol->dump(2); ) 620 } 621 #endif 622 } 623 624 if (phi_incr != NULL) { 625 // If compare points directly to the phi we need to adjust 626 // the compare so that it points to the incr. Limit have 627 // to be adjusted to keep trip count the same and we 628 // should avoid int overflow. 629 // 630 // i = init; do {} while(i++ < limit); 631 // is converted to 632 // i = init; do {} while(++i < limit+1); 633 // 634 limit = gvn->transform(new AddINode(limit, stride)); 635 } 636 637 // Now we need to canonicalize loop condition. 638 if (bt == BoolTest::ne) { 639 assert(stride_con == 1 || stride_con == -1, "simple increment only"); 640 // 'ne' can be replaced with 'lt' only when init < limit. 641 if (stride_con > 0 && init_t->_hi < limit_t->_lo) 642 bt = BoolTest::lt; 643 // 'ne' can be replaced with 'gt' only when init > limit. 644 if (stride_con < 0 && init_t->_lo > limit_t->_hi) 645 bt = BoolTest::gt; 646 } 647 648 if (incl_limit) { 649 // The limit check guaranties that 'limit <= (max_jint - stride)' so 650 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0. 651 // 652 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1); 653 limit = gvn->transform(new AddINode(limit, one)); 654 if (bt == BoolTest::le) 655 bt = BoolTest::lt; 656 else if (bt == BoolTest::ge) 657 bt = BoolTest::gt; 658 else 659 ShouldNotReachHere(); 660 } 661 set_subtree_ctrl( limit ); 662 663 if (LoopStripMiningIter == 0) { 664 // Check for SafePoint on backedge and remove 665 Node *sfpt = x->in(LoopNode::LoopBackControl); 666 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { 667 lazy_replace( sfpt, iftrue ); 668 if (loop->_safepts != NULL) { 669 loop->_safepts->yank(sfpt); 670 } 671 loop->_tail = iftrue; 672 } 673 } 674 675 // Build a canonical trip test. 676 // Clone code, as old values may be in use. 677 incr = incr->clone(); 678 incr->set_req(1,phi); 679 incr->set_req(2,stride); 680 incr = _igvn.register_new_node_with_optimizer(incr); 681 set_early_ctrl( incr ); 682 _igvn.rehash_node_delayed(phi); 683 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); 684 685 // If phi type is more restrictive than Int, raise to 686 // Int to prevent (almost) infinite recursion in igvn 687 // which can only handle integer types for constants or minint..maxint. 688 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { 689 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); 690 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); 691 nphi = _igvn.register_new_node_with_optimizer(nphi); 692 set_ctrl(nphi, get_ctrl(phi)); 693 _igvn.replace_node(phi, nphi); 694 phi = nphi->as_Phi(); 695 } 696 cmp = cmp->clone(); 697 cmp->set_req(1,incr); 698 cmp->set_req(2,limit); 699 cmp = _igvn.register_new_node_with_optimizer(cmp); 700 set_ctrl(cmp, iff->in(0)); 701 702 test = test->clone()->as_Bool(); 703 (*(BoolTest*)&test->_test)._test = bt; 704 test->set_req(1,cmp); 705 _igvn.register_new_node_with_optimizer(test); 706 set_ctrl(test, iff->in(0)); 707 708 // Replace the old IfNode with a new LoopEndNode 709 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt )); 710 IfNode *le = lex->as_If(); 711 uint dd = dom_depth(iff); 712 set_idom(le, le->in(0), dd); // Update dominance for loop exit 713 set_loop(le, loop); 714 715 // Get the loop-exit control 716 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); 717 718 // Need to swap loop-exit and loop-back control? 719 if (iftrue_op == Op_IfFalse) { 720 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le)); 721 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le)); 722 723 loop->_tail = back_control = ift2; 724 set_loop(ift2, loop); 725 set_loop(iff2, get_loop(iffalse)); 726 727 // Lazy update of 'get_ctrl' mechanism. 728 lazy_replace(iffalse, iff2); 729 lazy_replace(iftrue, ift2); 730 731 // Swap names 732 iffalse = iff2; 733 iftrue = ift2; 734 } else { 735 _igvn.rehash_node_delayed(iffalse); 736 _igvn.rehash_node_delayed(iftrue); 737 iffalse->set_req_X( 0, le, &_igvn ); 738 iftrue ->set_req_X( 0, le, &_igvn ); 739 } 740 741 set_idom(iftrue, le, dd+1); 742 set_idom(iffalse, le, dd+1); 743 assert(iff->outcnt() == 0, "should be dead now"); 744 lazy_replace( iff, le ); // fix 'get_ctrl' 745 746 Node *sfpt2 = le->in(0); 747 748 Node* entry_control = init_control; 749 bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL && 750 sfpt2->Opcode() == Op_SafePoint && !loop->_has_call; 751 IdealLoopTree* outer_ilt = NULL; 752 if (strip_mine_loop) { 753 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop, 754 cl_prob, le->_fcnt, entry_control, 755 iffalse); 756 } 757 758 // Now setup a new CountedLoopNode to replace the existing LoopNode 759 CountedLoopNode *l = new CountedLoopNode(entry_control, back_control); 760 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve 761 // The following assert is approximately true, and defines the intention 762 // of can_be_counted_loop. It fails, however, because phase->type 763 // is not yet initialized for this loop and its parts. 764 //assert(l->can_be_counted_loop(this), "sanity"); 765 _igvn.register_new_node_with_optimizer(l); 766 set_loop(l, loop); 767 loop->_head = l; 768 // Fix all data nodes placed at the old loop head. 769 // Uses the lazy-update mechanism of 'get_ctrl'. 770 lazy_replace( x, l ); 771 set_idom(l, entry_control, dom_depth(entry_control) + 1); 772 773 if (LoopStripMiningIter == 0 || strip_mine_loop) { 774 // Check for immediately preceding SafePoint and remove 775 if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) { 776 if (strip_mine_loop) { 777 Node* outer_le = outer_ilt->_tail->in(0); 778 Node* sfpt = sfpt2->clone(); 779 sfpt->set_req(0, iffalse); 780 outer_le->set_req(0, sfpt); 781 register_control(sfpt, outer_ilt, iffalse); 782 set_idom(outer_le, sfpt, dom_depth(sfpt)); 783 } 784 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); 785 if (loop->_safepts != NULL) { 786 loop->_safepts->yank(sfpt2); 787 } 788 } 789 } 790 791 // Free up intermediate goo 792 _igvn.remove_dead_node(hook); 793 794 #ifdef ASSERT 795 assert(l->is_valid_counted_loop(), "counted loop shape is messed up"); 796 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" ); 797 #endif 798 #ifndef PRODUCT 799 if (TraceLoopOpts) { 800 tty->print("Counted "); 801 loop->dump_head(); 802 } 803 #endif 804 805 C->print_method(PHASE_AFTER_CLOOPS, 3); 806 807 // Capture bounds of the loop in the induction variable Phi before 808 // subsequent transformation (iteration splitting) obscures the 809 // bounds 810 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn)); 811 812 if (strip_mine_loop) { 813 l->mark_strip_mined(); 814 l->verify_strip_mined(1); 815 outer_ilt->_head->as_Loop()->verify_strip_mined(1); 816 loop = outer_ilt; 817 } 818 819 return true; 820 } 821 822 //----------------------exact_limit------------------------------------------- 823 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) { 824 assert(loop->_head->is_CountedLoop(), ""); 825 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 826 assert(cl->is_valid_counted_loop(), ""); 827 828 if (ABS(cl->stride_con()) == 1 || 829 cl->limit()->Opcode() == Op_LoopLimit) { 830 // Old code has exact limit (it could be incorrect in case of int overflow). 831 // Loop limit is exact with stride == 1. And loop may already have exact limit. 832 return cl->limit(); 833 } 834 Node *limit = NULL; 835 #ifdef ASSERT 836 BoolTest::mask bt = cl->loopexit()->test_trip(); 837 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 838 #endif 839 if (cl->has_exact_trip_count()) { 840 // Simple case: loop has constant boundaries. 841 // Use jlongs to avoid integer overflow. 842 int stride_con = cl->stride_con(); 843 jlong init_con = cl->init_trip()->get_int(); 844 jlong limit_con = cl->limit()->get_int(); 845 julong trip_cnt = cl->trip_count(); 846 jlong final_con = init_con + trip_cnt*stride_con; 847 int final_int = (int)final_con; 848 // The final value should be in integer range since the loop 849 // is counted and the limit was checked for overflow. 850 assert(final_con == (jlong)final_int, "final value should be integer"); 851 limit = _igvn.intcon(final_int); 852 } else { 853 // Create new LoopLimit node to get exact limit (final iv value). 854 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride()); 855 register_new_node(limit, cl->in(LoopNode::EntryControl)); 856 } 857 assert(limit != NULL, "sanity"); 858 return limit; 859 } 860 861 //------------------------------Ideal------------------------------------------ 862 // Return a node which is more "ideal" than the current node. 863 // Attempt to convert into a counted-loop. 864 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 865 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) { 866 phase->C->set_major_progress(); 867 } 868 return RegionNode::Ideal(phase, can_reshape); 869 } 870 871 void LoopNode::verify_strip_mined(int expect_skeleton) const { 872 #ifdef ASSERT 873 const OuterStripMinedLoopNode* outer = NULL; 874 const CountedLoopNode* inner = NULL; 875 if (is_strip_mined()) { 876 assert(is_CountedLoop(), "no Loop should be marked strip mined"); 877 inner = as_CountedLoop(); 878 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop(); 879 } else if (is_OuterStripMinedLoop()) { 880 outer = this->as_OuterStripMinedLoop(); 881 inner = outer->unique_ctrl_out()->as_CountedLoop(); 882 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined"); 883 } 884 if (inner != NULL || outer != NULL) { 885 assert(inner != NULL && outer != NULL, "missing loop in strip mined nest"); 886 Node* outer_tail = outer->in(LoopNode::LoopBackControl); 887 Node* outer_le = outer_tail->in(0); 888 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If"); 889 Node* sfpt = outer_le->in(0); 890 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?"); 891 Node* inner_out = sfpt->in(0); 892 if (inner_out->outcnt() != 1) { 893 ResourceMark rm; 894 Unique_Node_List wq; 895 896 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) { 897 Node* u = inner_out->fast_out(i); 898 if (u == sfpt) { 899 continue; 900 } 901 wq.clear(); 902 wq.push(u); 903 bool found_sfpt = false; 904 for (uint next = 0; next < wq.size() && !found_sfpt; next++) { 905 Node *n = wq.at(next); 906 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) { 907 Node* u = n->fast_out(i); 908 if (u == sfpt) { 909 found_sfpt = true; 910 } 911 if (!u->is_CFG()) { 912 wq.push(u); 913 } 914 } 915 } 916 assert(found_sfpt, "no node in loop that's not input to safepoint"); 917 } 918 } 919 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd(); 920 assert(cle == inner->loopexit_or_null(), "mismatch"); 921 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0; 922 if (has_skeleton) { 923 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node"); 924 assert(outer->outcnt() == 2, "only phis"); 925 } else { 926 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?"); 927 uint phis = 0; 928 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) { 929 Node* u = inner->fast_out(i); 930 if (u->is_Phi()) { 931 phis++; 932 } 933 } 934 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) { 935 Node* u = outer->fast_out(i); 936 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop"); 937 } 938 uint stores = 0; 939 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) { 940 Node* u = inner_out->fast_out(i); 941 if (u->is_Store()) { 942 stores++; 943 } 944 } 945 assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis"); 946 } 947 assert(sfpt->outcnt() == 1, "no data node"); 948 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node"); 949 } 950 #endif 951 } 952 953 //============================================================================= 954 //------------------------------Ideal------------------------------------------ 955 // Return a node which is more "ideal" than the current node. 956 // Attempt to convert into a counted-loop. 957 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 958 return RegionNode::Ideal(phase, can_reshape); 959 } 960 961 //------------------------------dump_spec-------------------------------------- 962 // Dump special per-node info 963 #ifndef PRODUCT 964 void CountedLoopNode::dump_spec(outputStream *st) const { 965 LoopNode::dump_spec(st); 966 if (stride_is_con()) { 967 st->print("stride: %d ",stride_con()); 968 } 969 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx); 970 if (is_main_loop()) st->print("main of N%d", _idx); 971 if (is_post_loop()) st->print("post of N%d", _main_idx); 972 if (is_strip_mined()) st->print(" strip mined"); 973 } 974 #endif 975 976 //============================================================================= 977 int CountedLoopEndNode::stride_con() const { 978 return stride()->bottom_type()->is_int()->get_con(); 979 } 980 981 //============================================================================= 982 //------------------------------Value----------------------------------------- 983 const Type* LoopLimitNode::Value(PhaseGVN* phase) const { 984 const Type* init_t = phase->type(in(Init)); 985 const Type* limit_t = phase->type(in(Limit)); 986 const Type* stride_t = phase->type(in(Stride)); 987 // Either input is TOP ==> the result is TOP 988 if (init_t == Type::TOP) return Type::TOP; 989 if (limit_t == Type::TOP) return Type::TOP; 990 if (stride_t == Type::TOP) return Type::TOP; 991 992 int stride_con = stride_t->is_int()->get_con(); 993 if (stride_con == 1) 994 return NULL; // Identity 995 996 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) { 997 // Use jlongs to avoid integer overflow. 998 jlong init_con = init_t->is_int()->get_con(); 999 jlong limit_con = limit_t->is_int()->get_con(); 1000 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 1001 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 1002 jlong final_con = init_con + stride_con*trip_count; 1003 int final_int = (int)final_con; 1004 // The final value should be in integer range since the loop 1005 // is counted and the limit was checked for overflow. 1006 assert(final_con == (jlong)final_int, "final value should be integer"); 1007 return TypeInt::make(final_int); 1008 } 1009 1010 return bottom_type(); // TypeInt::INT 1011 } 1012 1013 //------------------------------Ideal------------------------------------------ 1014 // Return a node which is more "ideal" than the current node. 1015 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1016 if (phase->type(in(Init)) == Type::TOP || 1017 phase->type(in(Limit)) == Type::TOP || 1018 phase->type(in(Stride)) == Type::TOP) 1019 return NULL; // Dead 1020 1021 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1022 if (stride_con == 1) 1023 return NULL; // Identity 1024 1025 if (in(Init)->is_Con() && in(Limit)->is_Con()) 1026 return NULL; // Value 1027 1028 // Delay following optimizations until all loop optimizations 1029 // done to keep Ideal graph simple. 1030 if (!can_reshape || phase->C->major_progress()) 1031 return NULL; 1032 1033 const TypeInt* init_t = phase->type(in(Init) )->is_int(); 1034 const TypeInt* limit_t = phase->type(in(Limit))->is_int(); 1035 int stride_p; 1036 jlong lim, ini; 1037 julong max; 1038 if (stride_con > 0) { 1039 stride_p = stride_con; 1040 lim = limit_t->_hi; 1041 ini = init_t->_lo; 1042 max = (julong)max_jint; 1043 } else { 1044 stride_p = -stride_con; 1045 lim = init_t->_hi; 1046 ini = limit_t->_lo; 1047 max = (julong)min_jint; 1048 } 1049 julong range = lim - ini + stride_p; 1050 if (range <= max) { 1051 // Convert to integer expression if it is not overflow. 1052 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1)); 1053 Node *range = phase->transform(new SubINode(in(Limit), in(Init))); 1054 Node *bias = phase->transform(new AddINode(range, stride_m)); 1055 Node *trip = phase->transform(new DivINode(0, bias, in(Stride))); 1056 Node *span = phase->transform(new MulINode(trip, in(Stride))); 1057 return new AddINode(span, in(Init)); // exact limit 1058 } 1059 1060 if (is_power_of_2(stride_p) || // divisor is 2^n 1061 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node? 1062 // Convert to long expression to avoid integer overflow 1063 // and let igvn optimizer convert this division. 1064 // 1065 Node* init = phase->transform( new ConvI2LNode(in(Init))); 1066 Node* limit = phase->transform( new ConvI2LNode(in(Limit))); 1067 Node* stride = phase->longcon(stride_con); 1068 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1)); 1069 1070 Node *range = phase->transform(new SubLNode(limit, init)); 1071 Node *bias = phase->transform(new AddLNode(range, stride_m)); 1072 Node *span; 1073 if (stride_con > 0 && is_power_of_2(stride_p)) { 1074 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride) 1075 // and avoid generating rounding for division. Zero trip guard should 1076 // guarantee that init < limit but sometimes the guard is missing and 1077 // we can get situation when init > limit. Note, for the empty loop 1078 // optimization zero trip guard is generated explicitly which leaves 1079 // only RCE predicate where exact limit is used and the predicate 1080 // will simply fail forcing recompilation. 1081 Node* neg_stride = phase->longcon(-stride_con); 1082 span = phase->transform(new AndLNode(bias, neg_stride)); 1083 } else { 1084 Node *trip = phase->transform(new DivLNode(0, bias, stride)); 1085 span = phase->transform(new MulLNode(trip, stride)); 1086 } 1087 // Convert back to int 1088 Node *span_int = phase->transform(new ConvL2INode(span)); 1089 return new AddINode(span_int, in(Init)); // exact limit 1090 } 1091 1092 return NULL; // No progress 1093 } 1094 1095 //------------------------------Identity--------------------------------------- 1096 // If stride == 1 return limit node. 1097 Node* LoopLimitNode::Identity(PhaseGVN* phase) { 1098 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1099 if (stride_con == 1 || stride_con == -1) 1100 return in(Limit); 1101 return this; 1102 } 1103 1104 //============================================================================= 1105 //----------------------match_incr_with_optional_truncation-------------------- 1106 // Match increment with optional truncation: 1107 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 1108 // Return NULL for failure. Success returns the increment node. 1109 Node* CountedLoopNode::match_incr_with_optional_truncation( 1110 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { 1111 // Quick cutouts: 1112 if (expr == NULL || expr->req() != 3) return NULL; 1113 1114 Node *t1 = NULL; 1115 Node *t2 = NULL; 1116 const TypeInt* trunc_t = TypeInt::INT; 1117 Node* n1 = expr; 1118 int n1op = n1->Opcode(); 1119 1120 // Try to strip (n1 & M) or (n1 << N >> N) from n1. 1121 if (n1op == Op_AndI && 1122 n1->in(2)->is_Con() && 1123 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { 1124 // %%% This check should match any mask of 2**K-1. 1125 t1 = n1; 1126 n1 = t1->in(1); 1127 n1op = n1->Opcode(); 1128 trunc_t = TypeInt::CHAR; 1129 } else if (n1op == Op_RShiftI && 1130 n1->in(1) != NULL && 1131 n1->in(1)->Opcode() == Op_LShiftI && 1132 n1->in(2) == n1->in(1)->in(2) && 1133 n1->in(2)->is_Con()) { 1134 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); 1135 // %%% This check should match any shift in [1..31]. 1136 if (shift == 16 || shift == 8) { 1137 t1 = n1; 1138 t2 = t1->in(1); 1139 n1 = t2->in(1); 1140 n1op = n1->Opcode(); 1141 if (shift == 16) { 1142 trunc_t = TypeInt::SHORT; 1143 } else if (shift == 8) { 1144 trunc_t = TypeInt::BYTE; 1145 } 1146 } 1147 } 1148 1149 // If (maybe after stripping) it is an AddI, we won: 1150 if (n1op == Op_AddI) { 1151 *trunc1 = t1; 1152 *trunc2 = t2; 1153 *trunc_type = trunc_t; 1154 return n1; 1155 } 1156 1157 // failed 1158 return NULL; 1159 } 1160 1161 LoopNode* CountedLoopNode::skip_strip_mined(int expect_opaq) { 1162 if (is_strip_mined()) { 1163 verify_strip_mined(expect_opaq); 1164 return in(EntryControl)->as_Loop(); 1165 } 1166 return this; 1167 } 1168 1169 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const { 1170 assert(is_strip_mined(), "not a strip mined loop"); 1171 Node* c = in(EntryControl); 1172 if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) { 1173 return NULL; 1174 } 1175 return c->as_OuterStripMinedLoop(); 1176 } 1177 1178 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const { 1179 Node* c = in(LoopBackControl); 1180 if (c == NULL || c->is_top()) { 1181 return NULL; 1182 } 1183 return c->as_IfTrue(); 1184 } 1185 1186 IfTrueNode* CountedLoopNode::outer_loop_tail() const { 1187 LoopNode* l = outer_loop(); 1188 if (l == NULL) { 1189 return NULL; 1190 } 1191 return l->outer_loop_tail(); 1192 } 1193 1194 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const { 1195 IfTrueNode* proj = outer_loop_tail(); 1196 if (proj == NULL) { 1197 return NULL; 1198 } 1199 Node* c = proj->in(0); 1200 if (c == NULL || c->is_top() || c->outcnt() != 2) { 1201 return NULL; 1202 } 1203 return c->as_OuterStripMinedLoopEnd(); 1204 } 1205 1206 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const { 1207 LoopNode* l = outer_loop(); 1208 if (l == NULL) { 1209 return NULL; 1210 } 1211 return l->outer_loop_end(); 1212 } 1213 1214 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const { 1215 IfNode* le = outer_loop_end(); 1216 if (le == NULL) { 1217 return NULL; 1218 } 1219 Node* c = le->proj_out_or_null(false); 1220 if (c == NULL) { 1221 return NULL; 1222 } 1223 return c->as_IfFalse(); 1224 } 1225 1226 IfFalseNode* CountedLoopNode::outer_loop_exit() const { 1227 LoopNode* l = outer_loop(); 1228 if (l == NULL) { 1229 return NULL; 1230 } 1231 return l->outer_loop_exit(); 1232 } 1233 1234 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const { 1235 IfNode* le = outer_loop_end(); 1236 if (le == NULL) { 1237 return NULL; 1238 } 1239 Node* c = le->in(0); 1240 if (c == NULL || c->is_top()) { 1241 return NULL; 1242 } 1243 assert(c->Opcode() == Op_SafePoint, "broken outer loop"); 1244 return c->as_SafePoint(); 1245 } 1246 1247 SafePointNode* CountedLoopNode::outer_safepoint() const { 1248 LoopNode* l = outer_loop(); 1249 if (l == NULL) { 1250 return NULL; 1251 } 1252 return l->outer_safepoint(); 1253 } 1254 1255 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) { 1256 // Look for the outer & inner strip mined loop, reduce number of 1257 // iterations of the inner loop, set exit condition of outer loop, 1258 // construct required phi nodes for outer loop. 1259 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop(); 1260 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined"); 1261 Node* inner_iv_phi = inner_cl->phi(); 1262 if (inner_iv_phi == NULL) { 1263 IfNode* outer_le = outer_loop_end(); 1264 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1265 igvn->replace_node(outer_le, iff); 1266 inner_cl->clear_strip_mined(); 1267 return; 1268 } 1269 CountedLoopEndNode* inner_cle = inner_cl->loopexit(); 1270 1271 int stride = inner_cl->stride_con(); 1272 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride); 1273 int scaled_iters = (int)scaled_iters_long; 1274 int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride); 1275 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int(); 1276 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo; 1277 assert(iter_estimate > 0, "broken"); 1278 if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) { 1279 // Remove outer loop and safepoint (too few iterations) 1280 Node* outer_sfpt = outer_safepoint(); 1281 Node* outer_out = outer_loop_exit(); 1282 igvn->replace_node(outer_out, outer_sfpt->in(0)); 1283 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top()); 1284 inner_cl->clear_strip_mined(); 1285 return; 1286 } 1287 if (iter_estimate <= scaled_iters_long) { 1288 // We would only go through one iteration of 1289 // the outer loop: drop the outer loop but 1290 // keep the safepoint so we don't run for 1291 // too long without a safepoint 1292 IfNode* outer_le = outer_loop_end(); 1293 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1294 igvn->replace_node(outer_le, iff); 1295 inner_cl->clear_strip_mined(); 1296 return; 1297 } 1298 1299 Node* cle_tail = inner_cle->proj_out(true); 1300 ResourceMark rm; 1301 Node_List old_new; 1302 if (cle_tail->outcnt() > 1) { 1303 // Look for nodes on backedge of inner loop and clone them 1304 Unique_Node_List backedge_nodes; 1305 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) { 1306 Node* u = cle_tail->fast_out(i); 1307 if (u != inner_cl) { 1308 assert(!u->is_CFG(), "control flow on the backedge?"); 1309 backedge_nodes.push(u); 1310 } 1311 } 1312 uint last = igvn->C->unique(); 1313 for (uint next = 0; next < backedge_nodes.size(); next++) { 1314 Node* n = backedge_nodes.at(next); 1315 old_new.map(n->_idx, n->clone()); 1316 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1317 Node* u = n->fast_out(i); 1318 assert(!u->is_CFG(), "broken"); 1319 if (u->_idx >= last) { 1320 continue; 1321 } 1322 if (!u->is_Phi()) { 1323 backedge_nodes.push(u); 1324 } else { 1325 assert(u->in(0) == inner_cl, "strange phi on the backedge"); 1326 } 1327 } 1328 } 1329 // Put the clones on the outer loop backedge 1330 Node* le_tail = outer_loop_tail(); 1331 for (uint next = 0; next < backedge_nodes.size(); next++) { 1332 Node *n = old_new[backedge_nodes.at(next)->_idx]; 1333 for (uint i = 1; i < n->req(); i++) { 1334 if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) { 1335 n->set_req(i, old_new[n->in(i)->_idx]); 1336 } 1337 } 1338 if (n->in(0) != NULL && n->in(0) == cle_tail) { 1339 n->set_req(0, le_tail); 1340 } 1341 igvn->register_new_node_with_optimizer(n); 1342 } 1343 } 1344 1345 Node* iv_phi = NULL; 1346 // Make a clone of each phi in the inner loop 1347 // for the outer loop 1348 for (uint i = 0; i < inner_cl->outcnt(); i++) { 1349 Node* u = inner_cl->raw_out(i); 1350 if (u->is_Phi()) { 1351 assert(u->in(0) == inner_cl, "inconsistent"); 1352 Node* phi = u->clone(); 1353 phi->set_req(0, this); 1354 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx]; 1355 if (be != NULL) { 1356 phi->set_req(LoopNode::LoopBackControl, be); 1357 } 1358 phi = igvn->transform(phi); 1359 igvn->replace_input_of(u, LoopNode::EntryControl, phi); 1360 if (u == inner_iv_phi) { 1361 iv_phi = phi; 1362 } 1363 } 1364 } 1365 Node* cle_out = inner_cle->proj_out(false); 1366 if (cle_out->outcnt() > 1) { 1367 // Look for chains of stores that were sunk 1368 // out of the inner loop and are in the outer loop 1369 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) { 1370 Node* u = cle_out->fast_out(i); 1371 if (u->is_Store()) { 1372 Node* first = u; 1373 for(;;) { 1374 Node* next = first->in(MemNode::Memory); 1375 if (!next->is_Store() || next->in(0) != cle_out) { 1376 break; 1377 } 1378 first = next; 1379 } 1380 Node* last = u; 1381 for(;;) { 1382 Node* next = NULL; 1383 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) { 1384 Node* uu = last->fast_out(j); 1385 if (uu->is_Store() && uu->in(0) == cle_out) { 1386 assert(next == NULL, "only one in the outer loop"); 1387 next = uu; 1388 } 1389 } 1390 if (next == NULL) { 1391 break; 1392 } 1393 last = next; 1394 } 1395 Node* phi = NULL; 1396 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1397 Node* uu = fast_out(j); 1398 if (uu->is_Phi()) { 1399 Node* be = uu->in(LoopNode::LoopBackControl); 1400 if (be->is_Store() && old_new[be->_idx] != NULL) { 1401 assert(false, "store on the backedge + sunk stores: unsupported"); 1402 // drop outer loop 1403 IfNode* outer_le = outer_loop_end(); 1404 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1405 igvn->replace_node(outer_le, iff); 1406 inner_cl->clear_strip_mined(); 1407 return; 1408 } 1409 if (be == last || be == first->in(MemNode::Memory)) { 1410 assert(phi == NULL, "only one phi"); 1411 phi = uu; 1412 } 1413 } 1414 } 1415 #ifdef ASSERT 1416 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1417 Node* uu = fast_out(j); 1418 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) { 1419 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) { 1420 assert(phi == uu, "what's that phi?"); 1421 } else if (uu->adr_type() == TypePtr::BOTTOM) { 1422 Node* n = uu->in(LoopNode::LoopBackControl); 1423 uint limit = igvn->C->live_nodes(); 1424 uint i = 0; 1425 while (n != uu) { 1426 i++; 1427 assert(i < limit, "infinite loop"); 1428 if (n->is_Proj()) { 1429 n = n->in(0); 1430 } else if (n->is_SafePoint() || n->is_MemBar()) { 1431 n = n->in(TypeFunc::Memory); 1432 } else if (n->is_Phi()) { 1433 n = n->in(1); 1434 } else if (n->is_MergeMem()) { 1435 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type())); 1436 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) { 1437 n = n->in(MemNode::Memory); 1438 } else { 1439 n->dump(); 1440 ShouldNotReachHere(); 1441 } 1442 } 1443 } 1444 } 1445 } 1446 #endif 1447 if (phi == NULL) { 1448 // If the an entire chains was sunk, the 1449 // inner loop has no phi for that memory 1450 // slice, create one for the outer loop 1451 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY, 1452 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))); 1453 phi->set_req(LoopNode::LoopBackControl, last); 1454 phi = igvn->transform(phi); 1455 igvn->replace_input_of(first, MemNode::Memory, phi); 1456 } else { 1457 // Or fix the outer loop fix to include 1458 // that chain of stores. 1459 Node* be = phi->in(LoopNode::LoopBackControl); 1460 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported"); 1461 if (be == first->in(MemNode::Memory)) { 1462 if (be == phi->in(LoopNode::LoopBackControl)) { 1463 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last); 1464 } else { 1465 igvn->replace_input_of(be, MemNode::Memory, last); 1466 } 1467 } else { 1468 #ifdef ASSERT 1469 if (be == phi->in(LoopNode::LoopBackControl)) { 1470 assert(phi->in(LoopNode::LoopBackControl) == last, ""); 1471 } else { 1472 assert(be->in(MemNode::Memory) == last, ""); 1473 } 1474 #endif 1475 } 1476 } 1477 } 1478 } 1479 } 1480 1481 if (iv_phi != NULL) { 1482 // Now adjust the inner loop's exit condition 1483 Node* limit = inner_cl->limit(); 1484 Node* sub = NULL; 1485 if (stride > 0) { 1486 sub = igvn->transform(new SubINode(limit, iv_phi)); 1487 } else { 1488 sub = igvn->transform(new SubINode(iv_phi, limit)); 1489 } 1490 Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters))); 1491 Node* new_limit = NULL; 1492 if (stride > 0) { 1493 new_limit = igvn->transform(new AddINode(min, iv_phi)); 1494 } else { 1495 new_limit = igvn->transform(new SubINode(iv_phi, min)); 1496 } 1497 Node* cmp = inner_cle->cmp_node()->clone(); 1498 igvn->replace_input_of(cmp, 2, new_limit); 1499 Node* bol = inner_cle->in(CountedLoopEndNode::TestValue)->clone(); 1500 cmp->set_req(2, limit); 1501 bol->set_req(1, igvn->transform(cmp)); 1502 igvn->replace_input_of(outer_loop_end(), 1, igvn->transform(bol)); 1503 } else { 1504 assert(false, "should be able to adjust outer loop"); 1505 IfNode* outer_le = outer_loop_end(); 1506 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1507 igvn->replace_node(outer_le, iff); 1508 inner_cl->clear_strip_mined(); 1509 } 1510 } 1511 1512 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const { 1513 if (!in(0)) return Type::TOP; 1514 if (phase->type(in(0)) == Type::TOP) 1515 return Type::TOP; 1516 1517 return TypeTuple::IFBOTH; 1518 } 1519 1520 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1521 if (remove_dead_region(phase, can_reshape)) return this; 1522 1523 return NULL; 1524 } 1525 1526 //------------------------------filtered_type-------------------------------- 1527 // Return a type based on condition control flow 1528 // A successful return will be a type that is restricted due 1529 // to a series of dominating if-tests, such as: 1530 // if (i < 10) { 1531 // if (i > 0) { 1532 // here: "i" type is [1..10) 1533 // } 1534 // } 1535 // or a control flow merge 1536 // if (i < 10) { 1537 // do { 1538 // phi( , ) -- at top of loop type is [min_int..10) 1539 // i = ? 1540 // } while ( i < 10) 1541 // 1542 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { 1543 assert(n && n->bottom_type()->is_int(), "must be int"); 1544 const TypeInt* filtered_t = NULL; 1545 if (!n->is_Phi()) { 1546 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); 1547 filtered_t = filtered_type_from_dominators(n, n_ctrl); 1548 1549 } else { 1550 Node* phi = n->as_Phi(); 1551 Node* region = phi->in(0); 1552 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); 1553 if (region && region != C->top()) { 1554 for (uint i = 1; i < phi->req(); i++) { 1555 Node* val = phi->in(i); 1556 Node* use_c = region->in(i); 1557 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); 1558 if (val_t != NULL) { 1559 if (filtered_t == NULL) { 1560 filtered_t = val_t; 1561 } else { 1562 filtered_t = filtered_t->meet(val_t)->is_int(); 1563 } 1564 } 1565 } 1566 } 1567 } 1568 const TypeInt* n_t = _igvn.type(n)->is_int(); 1569 if (filtered_t != NULL) { 1570 n_t = n_t->join(filtered_t)->is_int(); 1571 } 1572 return n_t; 1573 } 1574 1575 1576 //------------------------------filtered_type_from_dominators-------------------------------- 1577 // Return a possibly more restrictive type for val based on condition control flow of dominators 1578 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { 1579 if (val->is_Con()) { 1580 return val->bottom_type()->is_int(); 1581 } 1582 uint if_limit = 10; // Max number of dominating if's visited 1583 const TypeInt* rtn_t = NULL; 1584 1585 if (use_ctrl && use_ctrl != C->top()) { 1586 Node* val_ctrl = get_ctrl(val); 1587 uint val_dom_depth = dom_depth(val_ctrl); 1588 Node* pred = use_ctrl; 1589 uint if_cnt = 0; 1590 while (if_cnt < if_limit) { 1591 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { 1592 if_cnt++; 1593 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); 1594 if (if_t != NULL) { 1595 if (rtn_t == NULL) { 1596 rtn_t = if_t; 1597 } else { 1598 rtn_t = rtn_t->join(if_t)->is_int(); 1599 } 1600 } 1601 } 1602 pred = idom(pred); 1603 if (pred == NULL || pred == C->top()) { 1604 break; 1605 } 1606 // Stop if going beyond definition block of val 1607 if (dom_depth(pred) < val_dom_depth) { 1608 break; 1609 } 1610 } 1611 } 1612 return rtn_t; 1613 } 1614 1615 1616 //------------------------------dump_spec-------------------------------------- 1617 // Dump special per-node info 1618 #ifndef PRODUCT 1619 void CountedLoopEndNode::dump_spec(outputStream *st) const { 1620 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) { 1621 BoolTest bt( test_trip()); // Added this for g++. 1622 1623 st->print("["); 1624 bt.dump_on(st); 1625 st->print("]"); 1626 } 1627 st->print(" "); 1628 IfNode::dump_spec(st); 1629 } 1630 #endif 1631 1632 //============================================================================= 1633 //------------------------------is_member-------------------------------------- 1634 // Is 'l' a member of 'this'? 1635 bool IdealLoopTree::is_member(const IdealLoopTree *l) const { 1636 while( l->_nest > _nest ) l = l->_parent; 1637 return l == this; 1638 } 1639 1640 //------------------------------set_nest--------------------------------------- 1641 // Set loop tree nesting depth. Accumulate _has_call bits. 1642 int IdealLoopTree::set_nest( uint depth ) { 1643 _nest = depth; 1644 int bits = _has_call; 1645 if( _child ) bits |= _child->set_nest(depth+1); 1646 if( bits ) _has_call = 1; 1647 if( _next ) bits |= _next ->set_nest(depth ); 1648 return bits; 1649 } 1650 1651 //------------------------------split_fall_in---------------------------------- 1652 // Split out multiple fall-in edges from the loop header. Move them to a 1653 // private RegionNode before the loop. This becomes the loop landing pad. 1654 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { 1655 PhaseIterGVN &igvn = phase->_igvn; 1656 uint i; 1657 1658 // Make a new RegionNode to be the landing pad. 1659 Node *landing_pad = new RegionNode( fall_in_cnt+1 ); 1660 phase->set_loop(landing_pad,_parent); 1661 // Gather all the fall-in control paths into the landing pad 1662 uint icnt = fall_in_cnt; 1663 uint oreq = _head->req(); 1664 for( i = oreq-1; i>0; i-- ) 1665 if( !phase->is_member( this, _head->in(i) ) ) 1666 landing_pad->set_req(icnt--,_head->in(i)); 1667 1668 // Peel off PhiNode edges as well 1669 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1670 Node *oj = _head->fast_out(j); 1671 if( oj->is_Phi() ) { 1672 PhiNode* old_phi = oj->as_Phi(); 1673 assert( old_phi->region() == _head, "" ); 1674 igvn.hash_delete(old_phi); // Yank from hash before hacking edges 1675 Node *p = PhiNode::make_blank(landing_pad, old_phi); 1676 uint icnt = fall_in_cnt; 1677 for( i = oreq-1; i>0; i-- ) { 1678 if( !phase->is_member( this, _head->in(i) ) ) { 1679 p->init_req(icnt--, old_phi->in(i)); 1680 // Go ahead and clean out old edges from old phi 1681 old_phi->del_req(i); 1682 } 1683 } 1684 // Search for CSE's here, because ZKM.jar does a lot of 1685 // loop hackery and we need to be a little incremental 1686 // with the CSE to avoid O(N^2) node blow-up. 1687 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE 1688 if( p2 ) { // Found CSE 1689 p->destruct(); // Recover useless new node 1690 p = p2; // Use old node 1691 } else { 1692 igvn.register_new_node_with_optimizer(p, old_phi); 1693 } 1694 // Make old Phi refer to new Phi. 1695 old_phi->add_req(p); 1696 // Check for the special case of making the old phi useless and 1697 // disappear it. In JavaGrande I have a case where this useless 1698 // Phi is the loop limit and prevents recognizing a CountedLoop 1699 // which in turn prevents removing an empty loop. 1700 Node *id_old_phi = old_phi->Identity( &igvn ); 1701 if( id_old_phi != old_phi ) { // Found a simple identity? 1702 // Note that I cannot call 'replace_node' here, because 1703 // that will yank the edge from old_phi to the Region and 1704 // I'm mid-iteration over the Region's uses. 1705 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { 1706 Node* use = old_phi->last_out(i); 1707 igvn.rehash_node_delayed(use); 1708 uint uses_found = 0; 1709 for (uint j = 0; j < use->len(); j++) { 1710 if (use->in(j) == old_phi) { 1711 if (j < use->req()) use->set_req (j, id_old_phi); 1712 else use->set_prec(j, id_old_phi); 1713 uses_found++; 1714 } 1715 } 1716 i -= uses_found; // we deleted 1 or more copies of this edge 1717 } 1718 } 1719 igvn._worklist.push(old_phi); 1720 } 1721 } 1722 // Finally clean out the fall-in edges from the RegionNode 1723 for( i = oreq-1; i>0; i-- ) { 1724 if( !phase->is_member( this, _head->in(i) ) ) { 1725 _head->del_req(i); 1726 } 1727 } 1728 igvn.rehash_node_delayed(_head); 1729 // Transform landing pad 1730 igvn.register_new_node_with_optimizer(landing_pad, _head); 1731 // Insert landing pad into the header 1732 _head->add_req(landing_pad); 1733 } 1734 1735 //------------------------------split_outer_loop------------------------------- 1736 // Split out the outermost loop from this shared header. 1737 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { 1738 PhaseIterGVN &igvn = phase->_igvn; 1739 1740 // Find index of outermost loop; it should also be my tail. 1741 uint outer_idx = 1; 1742 while( _head->in(outer_idx) != _tail ) outer_idx++; 1743 1744 // Make a LoopNode for the outermost loop. 1745 Node *ctl = _head->in(LoopNode::EntryControl); 1746 Node *outer = new LoopNode( ctl, _head->in(outer_idx) ); 1747 outer = igvn.register_new_node_with_optimizer(outer, _head); 1748 phase->set_created_loop_node(); 1749 1750 // Outermost loop falls into '_head' loop 1751 _head->set_req(LoopNode::EntryControl, outer); 1752 _head->del_req(outer_idx); 1753 // Split all the Phis up between '_head' loop and 'outer' loop. 1754 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1755 Node *out = _head->fast_out(j); 1756 if( out->is_Phi() ) { 1757 PhiNode *old_phi = out->as_Phi(); 1758 assert( old_phi->region() == _head, "" ); 1759 Node *phi = PhiNode::make_blank(outer, old_phi); 1760 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); 1761 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); 1762 phi = igvn.register_new_node_with_optimizer(phi, old_phi); 1763 // Make old Phi point to new Phi on the fall-in path 1764 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi); 1765 old_phi->del_req(outer_idx); 1766 } 1767 } 1768 1769 // Use the new loop head instead of the old shared one 1770 _head = outer; 1771 phase->set_loop(_head, this); 1772 } 1773 1774 //------------------------------fix_parent------------------------------------- 1775 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { 1776 loop->_parent = parent; 1777 if( loop->_child ) fix_parent( loop->_child, loop ); 1778 if( loop->_next ) fix_parent( loop->_next , parent ); 1779 } 1780 1781 //------------------------------estimate_path_freq----------------------------- 1782 static float estimate_path_freq( Node *n ) { 1783 // Try to extract some path frequency info 1784 IfNode *iff; 1785 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests 1786 uint nop = n->Opcode(); 1787 if( nop == Op_SafePoint ) { // Skip any safepoint 1788 n = n->in(0); 1789 continue; 1790 } 1791 if( nop == Op_CatchProj ) { // Get count from a prior call 1792 // Assume call does not always throw exceptions: means the call-site 1793 // count is also the frequency of the fall-through path. 1794 assert( n->is_CatchProj(), "" ); 1795 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) 1796 return 0.0f; // Assume call exception path is rare 1797 Node *call = n->in(0)->in(0)->in(0); 1798 assert( call->is_Call(), "expect a call here" ); 1799 const JVMState *jvms = ((CallNode*)call)->jvms(); 1800 ciMethodData* methodData = jvms->method()->method_data(); 1801 if (!methodData->is_mature()) return 0.0f; // No call-site data 1802 ciProfileData* data = methodData->bci_to_data(jvms->bci()); 1803 if ((data == NULL) || !data->is_CounterData()) { 1804 // no call profile available, try call's control input 1805 n = n->in(0); 1806 continue; 1807 } 1808 return data->as_CounterData()->count()/FreqCountInvocations; 1809 } 1810 // See if there's a gating IF test 1811 Node *n_c = n->in(0); 1812 if( !n_c->is_If() ) break; // No estimate available 1813 iff = n_c->as_If(); 1814 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? 1815 // Compute how much count comes on this path 1816 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; 1817 // Have no count info. Skip dull uncommon-trap like branches. 1818 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || 1819 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) 1820 break; 1821 // Skip through never-taken branch; look for a real loop exit. 1822 n = iff->in(0); 1823 } 1824 return 0.0f; // No estimate available 1825 } 1826 1827 //------------------------------merge_many_backedges--------------------------- 1828 // Merge all the backedges from the shared header into a private Region. 1829 // Feed that region as the one backedge to this loop. 1830 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { 1831 uint i; 1832 1833 // Scan for the top 2 hottest backedges 1834 float hotcnt = 0.0f; 1835 float warmcnt = 0.0f; 1836 uint hot_idx = 0; 1837 // Loop starts at 2 because slot 1 is the fall-in path 1838 for( i = 2; i < _head->req(); i++ ) { 1839 float cnt = estimate_path_freq(_head->in(i)); 1840 if( cnt > hotcnt ) { // Grab hottest path 1841 warmcnt = hotcnt; 1842 hotcnt = cnt; 1843 hot_idx = i; 1844 } else if( cnt > warmcnt ) { // And 2nd hottest path 1845 warmcnt = cnt; 1846 } 1847 } 1848 1849 // See if the hottest backedge is worthy of being an inner loop 1850 // by being much hotter than the next hottest backedge. 1851 if( hotcnt <= 0.0001 || 1852 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge 1853 1854 // Peel out the backedges into a private merge point; peel 1855 // them all except optionally hot_idx. 1856 PhaseIterGVN &igvn = phase->_igvn; 1857 1858 Node *hot_tail = NULL; 1859 // Make a Region for the merge point 1860 Node *r = new RegionNode(1); 1861 for( i = 2; i < _head->req(); i++ ) { 1862 if( i != hot_idx ) 1863 r->add_req( _head->in(i) ); 1864 else hot_tail = _head->in(i); 1865 } 1866 igvn.register_new_node_with_optimizer(r, _head); 1867 // Plug region into end of loop _head, followed by hot_tail 1868 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); 1869 igvn.replace_input_of(_head, 2, r); 1870 if( hot_idx ) _head->add_req(hot_tail); 1871 1872 // Split all the Phis up between '_head' loop and the Region 'r' 1873 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1874 Node *out = _head->fast_out(j); 1875 if( out->is_Phi() ) { 1876 PhiNode* n = out->as_Phi(); 1877 igvn.hash_delete(n); // Delete from hash before hacking edges 1878 Node *hot_phi = NULL; 1879 Node *phi = new PhiNode(r, n->type(), n->adr_type()); 1880 // Check all inputs for the ones to peel out 1881 uint j = 1; 1882 for( uint i = 2; i < n->req(); i++ ) { 1883 if( i != hot_idx ) 1884 phi->set_req( j++, n->in(i) ); 1885 else hot_phi = n->in(i); 1886 } 1887 // Register the phi but do not transform until whole place transforms 1888 igvn.register_new_node_with_optimizer(phi, n); 1889 // Add the merge phi to the old Phi 1890 while( n->req() > 3 ) n->del_req( n->req()-1 ); 1891 igvn.replace_input_of(n, 2, phi); 1892 if( hot_idx ) n->add_req(hot_phi); 1893 } 1894 } 1895 1896 1897 // Insert a new IdealLoopTree inserted below me. Turn it into a clone 1898 // of self loop tree. Turn self into a loop headed by _head and with 1899 // tail being the new merge point. 1900 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); 1901 phase->set_loop(_tail,ilt); // Adjust tail 1902 _tail = r; // Self's tail is new merge point 1903 phase->set_loop(r,this); 1904 ilt->_child = _child; // New guy has my children 1905 _child = ilt; // Self has new guy as only child 1906 ilt->_parent = this; // new guy has self for parent 1907 ilt->_nest = _nest; // Same nesting depth (for now) 1908 1909 // Starting with 'ilt', look for child loop trees using the same shared 1910 // header. Flatten these out; they will no longer be loops in the end. 1911 IdealLoopTree **pilt = &_child; 1912 while( ilt ) { 1913 if( ilt->_head == _head ) { 1914 uint i; 1915 for( i = 2; i < _head->req(); i++ ) 1916 if( _head->in(i) == ilt->_tail ) 1917 break; // Still a loop 1918 if( i == _head->req() ) { // No longer a loop 1919 // Flatten ilt. Hang ilt's "_next" list from the end of 1920 // ilt's '_child' list. Move the ilt's _child up to replace ilt. 1921 IdealLoopTree **cp = &ilt->_child; 1922 while( *cp ) cp = &(*cp)->_next; // Find end of child list 1923 *cp = ilt->_next; // Hang next list at end of child list 1924 *pilt = ilt->_child; // Move child up to replace ilt 1925 ilt->_head = NULL; // Flag as a loop UNIONED into parent 1926 ilt = ilt->_child; // Repeat using new ilt 1927 continue; // do not advance over ilt->_child 1928 } 1929 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); 1930 phase->set_loop(_head,ilt); 1931 } 1932 pilt = &ilt->_child; // Advance to next 1933 ilt = *pilt; 1934 } 1935 1936 if( _child ) fix_parent( _child, this ); 1937 } 1938 1939 //------------------------------beautify_loops--------------------------------- 1940 // Split shared headers and insert loop landing pads. 1941 // Insert a LoopNode to replace the RegionNode. 1942 // Return TRUE if loop tree is structurally changed. 1943 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { 1944 bool result = false; 1945 // Cache parts in locals for easy 1946 PhaseIterGVN &igvn = phase->_igvn; 1947 1948 igvn.hash_delete(_head); // Yank from hash before hacking edges 1949 1950 // Check for multiple fall-in paths. Peel off a landing pad if need be. 1951 int fall_in_cnt = 0; 1952 for( uint i = 1; i < _head->req(); i++ ) 1953 if( !phase->is_member( this, _head->in(i) ) ) 1954 fall_in_cnt++; 1955 assert( fall_in_cnt, "at least 1 fall-in path" ); 1956 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins 1957 split_fall_in( phase, fall_in_cnt ); 1958 1959 // Swap inputs to the _head and all Phis to move the fall-in edge to 1960 // the left. 1961 fall_in_cnt = 1; 1962 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) 1963 fall_in_cnt++; 1964 if( fall_in_cnt > 1 ) { 1965 // Since I am just swapping inputs I do not need to update def-use info 1966 Node *tmp = _head->in(1); 1967 igvn.rehash_node_delayed(_head); 1968 _head->set_req( 1, _head->in(fall_in_cnt) ); 1969 _head->set_req( fall_in_cnt, tmp ); 1970 // Swap also all Phis 1971 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { 1972 Node* phi = _head->fast_out(i); 1973 if( phi->is_Phi() ) { 1974 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges 1975 tmp = phi->in(1); 1976 phi->set_req( 1, phi->in(fall_in_cnt) ); 1977 phi->set_req( fall_in_cnt, tmp ); 1978 } 1979 } 1980 } 1981 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); 1982 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); 1983 1984 // If I am a shared header (multiple backedges), peel off the many 1985 // backedges into a private merge point and use the merge point as 1986 // the one true backedge. 1987 if( _head->req() > 3 ) { 1988 // Merge the many backedges into a single backedge but leave 1989 // the hottest backedge as separate edge for the following peel. 1990 merge_many_backedges( phase ); 1991 result = true; 1992 } 1993 1994 // If I have one hot backedge, peel off myself loop. 1995 // I better be the outermost loop. 1996 if (_head->req() > 3 && !_irreducible) { 1997 split_outer_loop( phase ); 1998 result = true; 1999 2000 } else if (!_head->is_Loop() && !_irreducible) { 2001 // Make a new LoopNode to replace the old loop head 2002 Node *l = new LoopNode( _head->in(1), _head->in(2) ); 2003 l = igvn.register_new_node_with_optimizer(l, _head); 2004 phase->set_created_loop_node(); 2005 // Go ahead and replace _head 2006 phase->_igvn.replace_node( _head, l ); 2007 _head = l; 2008 phase->set_loop(_head, this); 2009 } 2010 2011 // Now recursively beautify nested loops 2012 if( _child ) result |= _child->beautify_loops( phase ); 2013 if( _next ) result |= _next ->beautify_loops( phase ); 2014 return result; 2015 } 2016 2017 //------------------------------allpaths_check_safepts---------------------------- 2018 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 2019 // encountered. Helper for check_safepts. 2020 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { 2021 assert(stack.size() == 0, "empty stack"); 2022 stack.push(_tail); 2023 visited.Clear(); 2024 visited.set(_tail->_idx); 2025 while (stack.size() > 0) { 2026 Node* n = stack.pop(); 2027 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2028 // Terminate this path 2029 } else if (n->Opcode() == Op_SafePoint) { 2030 if (_phase->get_loop(n) != this) { 2031 if (_required_safept == NULL) _required_safept = new Node_List(); 2032 _required_safept->push(n); // save the one closest to the tail 2033 } 2034 // Terminate this path 2035 } else { 2036 uint start = n->is_Region() ? 1 : 0; 2037 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; 2038 for (uint i = start; i < end; i++) { 2039 Node* in = n->in(i); 2040 assert(in->is_CFG(), "must be"); 2041 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { 2042 stack.push(in); 2043 } 2044 } 2045 } 2046 } 2047 } 2048 2049 //------------------------------check_safepts---------------------------- 2050 // Given dominators, try to find loops with calls that must always be 2051 // executed (call dominates loop tail). These loops do not need non-call 2052 // safepoints (ncsfpt). 2053 // 2054 // A complication is that a safepoint in a inner loop may be needed 2055 // by an outer loop. In the following, the inner loop sees it has a 2056 // call (block 3) on every path from the head (block 2) to the 2057 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) 2058 // in block 2, _but_ this leaves the outer loop without a safepoint. 2059 // 2060 // entry 0 2061 // | 2062 // v 2063 // outer 1,2 +->1 2064 // | | 2065 // | v 2066 // | 2<---+ ncsfpt in 2 2067 // |_/|\ | 2068 // | v | 2069 // inner 2,3 / 3 | call in 3 2070 // / | | 2071 // v +--+ 2072 // exit 4 2073 // 2074 // 2075 // This method creates a list (_required_safept) of ncsfpt nodes that must 2076 // be protected is created for each loop. When a ncsfpt maybe deleted, it 2077 // is first looked for in the lists for the outer loops of the current loop. 2078 // 2079 // The insights into the problem: 2080 // A) counted loops are okay 2081 // B) innermost loops are okay (only an inner loop can delete 2082 // a ncsfpt needed by an outer loop) 2083 // C) a loop is immune from an inner loop deleting a safepoint 2084 // if the loop has a call on the idom-path 2085 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the 2086 // idom-path that is not in a nested loop 2087 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner 2088 // loop needs to be prevented from deletion by an inner loop 2089 // 2090 // There are two analyses: 2091 // 1) The first, and cheaper one, scans the loop body from 2092 // tail to head following the idom (immediate dominator) 2093 // chain, looking for the cases (C,D,E) above. 2094 // Since inner loops are scanned before outer loops, there is summary 2095 // information about inner loops. Inner loops can be skipped over 2096 // when the tail of an inner loop is encountered. 2097 // 2098 // 2) The second, invoked if the first fails to find a call or ncsfpt on 2099 // the idom path (which is rare), scans all predecessor control paths 2100 // from the tail to the head, terminating a path when a call or sfpt 2101 // is encountered, to find the ncsfpt's that are closest to the tail. 2102 // 2103 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { 2104 // Bottom up traversal 2105 IdealLoopTree* ch = _child; 2106 if (_child) _child->check_safepts(visited, stack); 2107 if (_next) _next ->check_safepts(visited, stack); 2108 2109 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { 2110 bool has_call = false; // call on dom-path 2111 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth 2112 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth 2113 // Scan the dom-path nodes from tail to head 2114 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { 2115 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2116 has_call = true; 2117 _has_sfpt = 1; // Then no need for a safept! 2118 break; 2119 } else if (n->Opcode() == Op_SafePoint) { 2120 if (_phase->get_loop(n) == this) { 2121 has_local_ncsfpt = true; 2122 break; 2123 } 2124 if (nonlocal_ncsfpt == NULL) { 2125 nonlocal_ncsfpt = n; // save the one closest to the tail 2126 } 2127 } else { 2128 IdealLoopTree* nlpt = _phase->get_loop(n); 2129 if (this != nlpt) { 2130 // If at an inner loop tail, see if the inner loop has already 2131 // recorded seeing a call on the dom-path (and stop.) If not, 2132 // jump to the head of the inner loop. 2133 assert(is_member(nlpt), "nested loop"); 2134 Node* tail = nlpt->_tail; 2135 if (tail->in(0)->is_If()) tail = tail->in(0); 2136 if (n == tail) { 2137 // If inner loop has call on dom-path, so does outer loop 2138 if (nlpt->_has_sfpt) { 2139 has_call = true; 2140 _has_sfpt = 1; 2141 break; 2142 } 2143 // Skip to head of inner loop 2144 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); 2145 n = nlpt->_head; 2146 } 2147 } 2148 } 2149 } 2150 // Record safept's that this loop needs preserved when an 2151 // inner loop attempts to delete it's safepoints. 2152 if (_child != NULL && !has_call && !has_local_ncsfpt) { 2153 if (nonlocal_ncsfpt != NULL) { 2154 if (_required_safept == NULL) _required_safept = new Node_List(); 2155 _required_safept->push(nonlocal_ncsfpt); 2156 } else { 2157 // Failed to find a suitable safept on the dom-path. Now use 2158 // an all paths walk from tail to head, looking for safepoints to preserve. 2159 allpaths_check_safepts(visited, stack); 2160 } 2161 } 2162 } 2163 } 2164 2165 //---------------------------is_deleteable_safept---------------------------- 2166 // Is safept not required by an outer loop? 2167 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { 2168 assert(sfpt->Opcode() == Op_SafePoint, ""); 2169 IdealLoopTree* lp = get_loop(sfpt)->_parent; 2170 while (lp != NULL) { 2171 Node_List* sfpts = lp->_required_safept; 2172 if (sfpts != NULL) { 2173 for (uint i = 0; i < sfpts->size(); i++) { 2174 if (sfpt == sfpts->at(i)) 2175 return false; 2176 } 2177 } 2178 lp = lp->_parent; 2179 } 2180 return true; 2181 } 2182 2183 //---------------------------replace_parallel_iv------------------------------- 2184 // Replace parallel induction variable (parallel to trip counter) 2185 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) { 2186 assert(loop->_head->is_CountedLoop(), ""); 2187 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2188 if (!cl->is_valid_counted_loop()) 2189 return; // skip malformed counted loop 2190 Node *incr = cl->incr(); 2191 if (incr == NULL) 2192 return; // Dead loop? 2193 Node *init = cl->init_trip(); 2194 Node *phi = cl->phi(); 2195 int stride_con = cl->stride_con(); 2196 2197 // Visit all children, looking for Phis 2198 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { 2199 Node *out = cl->out(i); 2200 // Look for other phis (secondary IVs). Skip dead ones 2201 if (!out->is_Phi() || out == phi || !has_node(out)) 2202 continue; 2203 PhiNode* phi2 = out->as_Phi(); 2204 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); 2205 // Look for induction variables of the form: X += constant 2206 if (phi2->region() != loop->_head || 2207 incr2->req() != 3 || 2208 incr2->in(1) != phi2 || 2209 incr2 == incr || 2210 incr2->Opcode() != Op_AddI || 2211 !incr2->in(2)->is_Con()) 2212 continue; 2213 2214 // Check for parallel induction variable (parallel to trip counter) 2215 // via an affine function. In particular, count-down loops with 2216 // count-up array indices are common. We only RCE references off 2217 // the trip-counter, so we need to convert all these to trip-counter 2218 // expressions. 2219 Node *init2 = phi2->in( LoopNode::EntryControl ); 2220 int stride_con2 = incr2->in(2)->get_int(); 2221 2222 // The ratio of the two strides cannot be represented as an int 2223 // if stride_con2 is min_int and stride_con is -1. 2224 if (stride_con2 == min_jint && stride_con == -1) { 2225 continue; 2226 } 2227 2228 // The general case here gets a little tricky. We want to find the 2229 // GCD of all possible parallel IV's and make a new IV using this 2230 // GCD for the loop. Then all possible IVs are simple multiples of 2231 // the GCD. In practice, this will cover very few extra loops. 2232 // Instead we require 'stride_con2' to be a multiple of 'stride_con', 2233 // where +/-1 is the common case, but other integer multiples are 2234 // also easy to handle. 2235 int ratio_con = stride_con2/stride_con; 2236 2237 if ((ratio_con * stride_con) == stride_con2) { // Check for exact 2238 #ifndef PRODUCT 2239 if (TraceLoopOpts) { 2240 tty->print("Parallel IV: %d ", phi2->_idx); 2241 loop->dump_head(); 2242 } 2243 #endif 2244 // Convert to using the trip counter. The parallel induction 2245 // variable differs from the trip counter by a loop-invariant 2246 // amount, the difference between their respective initial values. 2247 // It is scaled by the 'ratio_con'. 2248 Node* ratio = _igvn.intcon(ratio_con); 2249 set_ctrl(ratio, C->root()); 2250 Node* ratio_init = new MulINode(init, ratio); 2251 _igvn.register_new_node_with_optimizer(ratio_init, init); 2252 set_early_ctrl(ratio_init); 2253 Node* diff = new SubINode(init2, ratio_init); 2254 _igvn.register_new_node_with_optimizer(diff, init2); 2255 set_early_ctrl(diff); 2256 Node* ratio_idx = new MulINode(phi, ratio); 2257 _igvn.register_new_node_with_optimizer(ratio_idx, phi); 2258 set_ctrl(ratio_idx, cl); 2259 Node* add = new AddINode(ratio_idx, diff); 2260 _igvn.register_new_node_with_optimizer(add); 2261 set_ctrl(add, cl); 2262 _igvn.replace_node( phi2, add ); 2263 // Sometimes an induction variable is unused 2264 if (add->outcnt() == 0) { 2265 _igvn.remove_dead_node(add); 2266 } 2267 --i; // deleted this phi; rescan starting with next position 2268 continue; 2269 } 2270 } 2271 } 2272 2273 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) { 2274 Node* keep = NULL; 2275 if (keep_one) { 2276 // Look for a safepoint on the idom-path. 2277 for (Node* i = tail(); i != _head; i = phase->idom(i)) { 2278 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) { 2279 keep = i; 2280 break; // Found one 2281 } 2282 } 2283 } 2284 2285 // Don't remove any safepoints if it is requested to keep a single safepoint and 2286 // no safepoint was found on idom-path. It is not safe to remove any safepoint 2287 // in this case since there's no safepoint dominating all paths in the loop body. 2288 bool prune = !keep_one || keep != NULL; 2289 2290 // Delete other safepoints in this loop. 2291 Node_List* sfpts = _safepts; 2292 if (prune && sfpts != NULL) { 2293 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint"); 2294 for (uint i = 0; i < sfpts->size(); i++) { 2295 Node* n = sfpts->at(i); 2296 assert(phase->get_loop(n) == this, ""); 2297 if (n != keep && phase->is_deleteable_safept(n)) { 2298 phase->lazy_replace(n, n->in(TypeFunc::Control)); 2299 } 2300 } 2301 } 2302 } 2303 2304 //------------------------------counted_loop----------------------------------- 2305 // Convert to counted loops where possible 2306 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { 2307 2308 // For grins, set the inner-loop flag here 2309 if (!_child) { 2310 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop(); 2311 } 2312 2313 IdealLoopTree* loop = this; 2314 if (_head->is_CountedLoop() || 2315 phase->is_counted_loop(_head, loop)) { 2316 2317 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) { 2318 // Indicate we do not need a safepoint here 2319 _has_sfpt = 1; 2320 } 2321 2322 // Remove safepoints 2323 bool keep_one_sfpt = !(_has_call || _has_sfpt); 2324 remove_safepoints(phase, keep_one_sfpt); 2325 2326 // Look for induction variables 2327 phase->replace_parallel_iv(this); 2328 2329 } else if (_parent != NULL && !_irreducible) { 2330 // Not a counted loop. Keep one safepoint. 2331 bool keep_one_sfpt = true; 2332 remove_safepoints(phase, keep_one_sfpt); 2333 } 2334 2335 // Recursively 2336 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?"); 2337 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops"); 2338 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase); 2339 if (loop->_next) loop->_next ->counted_loop(phase); 2340 } 2341 2342 #ifndef PRODUCT 2343 //------------------------------dump_head-------------------------------------- 2344 // Dump 1 liner for loop header info 2345 void IdealLoopTree::dump_head( ) const { 2346 for (uint i=0; i<_nest; i++) 2347 tty->print(" "); 2348 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); 2349 if (_irreducible) tty->print(" IRREDUCIBLE"); 2350 Node* entry = _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl); 2351 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 2352 if (predicate != NULL ) { 2353 tty->print(" limit_check"); 2354 entry = entry->in(0)->in(0); 2355 } 2356 if (UseLoopPredicate) { 2357 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 2358 if (entry != NULL) { 2359 tty->print(" predicated"); 2360 } 2361 } 2362 if (_head->is_CountedLoop()) { 2363 CountedLoopNode *cl = _head->as_CountedLoop(); 2364 tty->print(" counted"); 2365 2366 Node* init_n = cl->init_trip(); 2367 if (init_n != NULL && init_n->is_Con()) 2368 tty->print(" [%d,", cl->init_trip()->get_int()); 2369 else 2370 tty->print(" [int,"); 2371 Node* limit_n = cl->limit(); 2372 if (limit_n != NULL && limit_n->is_Con()) 2373 tty->print("%d),", cl->limit()->get_int()); 2374 else 2375 tty->print("int),"); 2376 int stride_con = cl->stride_con(); 2377 if (stride_con > 0) tty->print("+"); 2378 tty->print("%d", stride_con); 2379 2380 tty->print(" (%0.f iters) ", cl->profile_trip_cnt()); 2381 2382 if (cl->is_pre_loop ()) tty->print(" pre" ); 2383 if (cl->is_main_loop()) tty->print(" main"); 2384 if (cl->is_post_loop()) tty->print(" post"); 2385 if (cl->is_vectorized_loop()) tty->print(" vector"); 2386 if (cl->range_checks_present()) tty->print(" rc "); 2387 if (cl->is_multiversioned()) tty->print(" multi "); 2388 } 2389 if (_has_call) tty->print(" has_call"); 2390 if (_has_sfpt) tty->print(" has_sfpt"); 2391 if (_rce_candidate) tty->print(" rce"); 2392 if (_safepts != NULL && _safepts->size() > 0) { 2393 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }"); 2394 } 2395 if (_required_safept != NULL && _required_safept->size() > 0) { 2396 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }"); 2397 } 2398 if (Verbose) { 2399 tty->print(" body={"); _body.dump_simple(); tty->print(" }"); 2400 } 2401 if (_head->as_Loop()->is_strip_mined()) { 2402 tty->print(" strip_mined"); 2403 } 2404 tty->cr(); 2405 } 2406 2407 //------------------------------dump------------------------------------------- 2408 // Dump loops by loop tree 2409 void IdealLoopTree::dump( ) const { 2410 dump_head(); 2411 if (_child) _child->dump(); 2412 if (_next) _next ->dump(); 2413 } 2414 2415 #endif 2416 2417 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { 2418 if (loop == root) { 2419 if (loop->_child != NULL) { 2420 log->begin_head("loop_tree"); 2421 log->end_head(); 2422 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2423 log->tail("loop_tree"); 2424 assert(loop->_next == NULL, "what?"); 2425 } 2426 } else { 2427 Node* head = loop->_head; 2428 log->begin_head("loop"); 2429 log->print(" idx='%d' ", head->_idx); 2430 if (loop->_irreducible) log->print("irreducible='1' "); 2431 if (head->is_Loop()) { 2432 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); 2433 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); 2434 } 2435 if (head->is_CountedLoop()) { 2436 CountedLoopNode* cl = head->as_CountedLoop(); 2437 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); 2438 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); 2439 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); 2440 } 2441 log->end_head(); 2442 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2443 log->tail("loop"); 2444 if( loop->_next ) log_loop_tree(root, loop->_next, log); 2445 } 2446 } 2447 2448 //---------------------collect_potentially_useful_predicates----------------------- 2449 // Helper function to collect potentially useful predicates to prevent them from 2450 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates 2451 void PhaseIdealLoop::collect_potentially_useful_predicates( 2452 IdealLoopTree * loop, Unique_Node_List &useful_predicates) { 2453 if (loop->_child) { // child 2454 collect_potentially_useful_predicates(loop->_child, useful_predicates); 2455 } 2456 2457 // self (only loops that we can apply loop predication may use their predicates) 2458 if (loop->_head->is_Loop() && 2459 !loop->_irreducible && 2460 !loop->tail()->is_top()) { 2461 LoopNode* lpn = loop->_head->as_Loop(); 2462 Node* entry = lpn->in(LoopNode::EntryControl); 2463 Node* predicate_proj = find_predicate(entry); // loop_limit_check first 2464 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication 2465 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be"); 2466 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2467 entry = entry->in(0)->in(0); 2468 } 2469 predicate_proj = find_predicate(entry); // Predicate 2470 if (predicate_proj != NULL ) { 2471 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2472 } 2473 } 2474 2475 if (loop->_next) { // sibling 2476 collect_potentially_useful_predicates(loop->_next, useful_predicates); 2477 } 2478 } 2479 2480 //------------------------eliminate_useless_predicates----------------------------- 2481 // Eliminate all inserted predicates if they could not be used by loop predication. 2482 // Note: it will also eliminates loop limits check predicate since it also uses 2483 // Opaque1 node (see Parse::add_predicate()). 2484 void PhaseIdealLoop::eliminate_useless_predicates() { 2485 if (C->predicate_count() == 0) 2486 return; // no predicate left 2487 2488 Unique_Node_List useful_predicates; // to store useful predicates 2489 if (C->has_loops()) { 2490 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); 2491 } 2492 2493 for (int i = C->predicate_count(); i > 0; i--) { 2494 Node * n = C->predicate_opaque1_node(i-1); 2495 assert(n->Opcode() == Op_Opaque1, "must be"); 2496 if (!useful_predicates.member(n)) { // not in the useful list 2497 _igvn.replace_node(n, n->in(1)); 2498 } 2499 } 2500 } 2501 2502 //------------------------process_expensive_nodes----------------------------- 2503 // Expensive nodes have their control input set to prevent the GVN 2504 // from commoning them and as a result forcing the resulting node to 2505 // be in a more frequent path. Use CFG information here, to change the 2506 // control inputs so that some expensive nodes can be commoned while 2507 // not executed more frequently. 2508 bool PhaseIdealLoop::process_expensive_nodes() { 2509 assert(OptimizeExpensiveOps, "optimization off?"); 2510 2511 // Sort nodes to bring similar nodes together 2512 C->sort_expensive_nodes(); 2513 2514 bool progress = false; 2515 2516 for (int i = 0; i < C->expensive_count(); ) { 2517 Node* n = C->expensive_node(i); 2518 int start = i; 2519 // Find nodes similar to n 2520 i++; 2521 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++); 2522 int end = i; 2523 // And compare them two by two 2524 for (int j = start; j < end; j++) { 2525 Node* n1 = C->expensive_node(j); 2526 if (is_node_unreachable(n1)) { 2527 continue; 2528 } 2529 for (int k = j+1; k < end; k++) { 2530 Node* n2 = C->expensive_node(k); 2531 if (is_node_unreachable(n2)) { 2532 continue; 2533 } 2534 2535 assert(n1 != n2, "should be pair of nodes"); 2536 2537 Node* c1 = n1->in(0); 2538 Node* c2 = n2->in(0); 2539 2540 Node* parent_c1 = c1; 2541 Node* parent_c2 = c2; 2542 2543 // The call to get_early_ctrl_for_expensive() moves the 2544 // expensive nodes up but stops at loops that are in a if 2545 // branch. See whether we can exit the loop and move above the 2546 // If. 2547 if (c1->is_Loop()) { 2548 parent_c1 = c1->in(1); 2549 } 2550 if (c2->is_Loop()) { 2551 parent_c2 = c2->in(1); 2552 } 2553 2554 if (parent_c1 == parent_c2) { 2555 _igvn._worklist.push(n1); 2556 _igvn._worklist.push(n2); 2557 continue; 2558 } 2559 2560 // Look for identical expensive node up the dominator chain. 2561 if (is_dominator(c1, c2)) { 2562 c2 = c1; 2563 } else if (is_dominator(c2, c1)) { 2564 c1 = c2; 2565 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() && 2566 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) { 2567 // Both branches have the same expensive node so move it up 2568 // before the if. 2569 c1 = c2 = idom(parent_c1->in(0)); 2570 } 2571 // Do the actual moves 2572 if (n1->in(0) != c1) { 2573 _igvn.hash_delete(n1); 2574 n1->set_req(0, c1); 2575 _igvn.hash_insert(n1); 2576 _igvn._worklist.push(n1); 2577 progress = true; 2578 } 2579 if (n2->in(0) != c2) { 2580 _igvn.hash_delete(n2); 2581 n2->set_req(0, c2); 2582 _igvn.hash_insert(n2); 2583 _igvn._worklist.push(n2); 2584 progress = true; 2585 } 2586 } 2587 } 2588 } 2589 2590 return progress; 2591 } 2592 2593 2594 //============================================================================= 2595 //----------------------------build_and_optimize------------------------------- 2596 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to 2597 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. 2598 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) { 2599 ResourceMark rm; 2600 2601 int old_progress = C->major_progress(); 2602 uint orig_worklist_size = _igvn._worklist.size(); 2603 2604 // Reset major-progress flag for the driver's heuristics 2605 C->clear_major_progress(); 2606 2607 #ifndef PRODUCT 2608 // Capture for later assert 2609 uint unique = C->unique(); 2610 _loop_invokes++; 2611 _loop_work += unique; 2612 #endif 2613 2614 // True if the method has at least 1 irreducible loop 2615 _has_irreducible_loops = false; 2616 2617 _created_loop_node = false; 2618 2619 Arena *a = Thread::current()->resource_area(); 2620 VectorSet visited(a); 2621 // Pre-grow the mapping from Nodes to IdealLoopTrees. 2622 _nodes.map(C->unique(), NULL); 2623 memset(_nodes.adr(), 0, wordSize * C->unique()); 2624 2625 // Pre-build the top-level outermost loop tree entry 2626 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); 2627 // Do not need a safepoint at the top level 2628 _ltree_root->_has_sfpt = 1; 2629 2630 // Initialize Dominators. 2631 // Checked in clone_loop_predicate() during beautify_loops(). 2632 _idom_size = 0; 2633 _idom = NULL; 2634 _dom_depth = NULL; 2635 _dom_stk = NULL; 2636 2637 // Empty pre-order array 2638 allocate_preorders(); 2639 2640 // Build a loop tree on the fly. Build a mapping from CFG nodes to 2641 // IdealLoopTree entries. Data nodes are NOT walked. 2642 build_loop_tree(); 2643 // Check for bailout, and return 2644 if (C->failing()) { 2645 return; 2646 } 2647 2648 // No loops after all 2649 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); 2650 2651 // There should always be an outer loop containing the Root and Return nodes. 2652 // If not, we have a degenerate empty program. Bail out in this case. 2653 if (!has_node(C->root())) { 2654 if (!_verify_only) { 2655 C->clear_major_progress(); 2656 C->record_method_not_compilable("empty program detected during loop optimization"); 2657 } 2658 return; 2659 } 2660 2661 // Nothing to do, so get out 2662 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only; 2663 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn); 2664 if (stop_early && !do_expensive_nodes) { 2665 _igvn.optimize(); // Cleanup NeverBranches 2666 return; 2667 } 2668 2669 // Set loop nesting depth 2670 _ltree_root->set_nest( 0 ); 2671 2672 // Split shared headers and insert loop landing pads. 2673 // Do not bother doing this on the Root loop of course. 2674 if( !_verify_me && !_verify_only && _ltree_root->_child ) { 2675 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3); 2676 if( _ltree_root->_child->beautify_loops( this ) ) { 2677 // Re-build loop tree! 2678 _ltree_root->_child = NULL; 2679 _nodes.clear(); 2680 reallocate_preorders(); 2681 build_loop_tree(); 2682 // Check for bailout, and return 2683 if (C->failing()) { 2684 return; 2685 } 2686 // Reset loop nesting depth 2687 _ltree_root->set_nest( 0 ); 2688 2689 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3); 2690 } 2691 } 2692 2693 // Build Dominators for elision of NULL checks & loop finding. 2694 // Since nodes do not have a slot for immediate dominator, make 2695 // a persistent side array for that info indexed on node->_idx. 2696 _idom_size = C->unique(); 2697 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); 2698 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); 2699 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth 2700 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); 2701 2702 Dominators(); 2703 2704 if (!_verify_only) { 2705 // As a side effect, Dominators removed any unreachable CFG paths 2706 // into RegionNodes. It doesn't do this test against Root, so 2707 // we do it here. 2708 for( uint i = 1; i < C->root()->req(); i++ ) { 2709 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? 2710 _igvn.delete_input_of(C->root(), i); 2711 i--; // Rerun same iteration on compressed edges 2712 } 2713 } 2714 2715 // Given dominators, try to find inner loops with calls that must 2716 // always be executed (call dominates loop tail). These loops do 2717 // not need a separate safepoint. 2718 Node_List cisstack(a); 2719 _ltree_root->check_safepts(visited, cisstack); 2720 } 2721 2722 // Walk the DATA nodes and place into loops. Find earliest control 2723 // node. For CFG nodes, the _nodes array starts out and remains 2724 // holding the associated IdealLoopTree pointer. For DATA nodes, the 2725 // _nodes array holds the earliest legal controlling CFG node. 2726 2727 // Allocate stack with enough space to avoid frequent realloc 2728 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats 2729 Node_Stack nstack( a, stack_size ); 2730 2731 visited.Clear(); 2732 Node_List worklist(a); 2733 // Don't need C->root() on worklist since 2734 // it will be processed among C->top() inputs 2735 worklist.push( C->top() ); 2736 visited.set( C->top()->_idx ); // Set C->top() as visited now 2737 build_loop_early( visited, worklist, nstack ); 2738 2739 // Given early legal placement, try finding counted loops. This placement 2740 // is good enough to discover most loop invariants. 2741 if( !_verify_me && !_verify_only ) 2742 _ltree_root->counted_loop( this ); 2743 2744 // Find latest loop placement. Find ideal loop placement. 2745 visited.Clear(); 2746 init_dom_lca_tags(); 2747 // Need C->root() on worklist when processing outs 2748 worklist.push( C->root() ); 2749 NOT_PRODUCT( C->verify_graph_edges(); ) 2750 worklist.push( C->top() ); 2751 build_loop_late( visited, worklist, nstack ); 2752 2753 if (_verify_only) { 2754 // restore major progress flag 2755 for (int i = 0; i < old_progress; i++) 2756 C->set_major_progress(); 2757 assert(C->unique() == unique, "verification mode made Nodes? ? ?"); 2758 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything"); 2759 return; 2760 } 2761 2762 // clear out the dead code after build_loop_late 2763 while (_deadlist.size()) { 2764 _igvn.remove_globally_dead_node(_deadlist.pop()); 2765 } 2766 2767 if (stop_early) { 2768 assert(do_expensive_nodes, "why are we here?"); 2769 if (process_expensive_nodes()) { 2770 // If we made some progress when processing expensive nodes then 2771 // the IGVN may modify the graph in a way that will allow us to 2772 // make some more progress: we need to try processing expensive 2773 // nodes again. 2774 C->set_major_progress(); 2775 } 2776 _igvn.optimize(); 2777 return; 2778 } 2779 2780 // Some parser-inserted loop predicates could never be used by loop 2781 // predication or they were moved away from loop during some optimizations. 2782 // For example, peeling. Eliminate them before next loop optimizations. 2783 eliminate_useless_predicates(); 2784 2785 #ifndef PRODUCT 2786 C->verify_graph_edges(); 2787 if (_verify_me) { // Nested verify pass? 2788 // Check to see if the verify mode is broken 2789 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); 2790 return; 2791 } 2792 if(VerifyLoopOptimizations) verify(); 2793 if(TraceLoopOpts && C->has_loops()) { 2794 _ltree_root->dump(); 2795 } 2796 #endif 2797 2798 if (skip_loop_opts) { 2799 // restore major progress flag 2800 for (int i = 0; i < old_progress; i++) { 2801 C->set_major_progress(); 2802 } 2803 2804 // Cleanup any modified bits 2805 _igvn.optimize(); 2806 2807 if (C->log() != NULL) { 2808 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2809 } 2810 return; 2811 } 2812 2813 if (ReassociateInvariants) { 2814 // Reassociate invariants and prep for split_thru_phi 2815 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2816 IdealLoopTree* lpt = iter.current(); 2817 bool is_counted = lpt->is_counted(); 2818 if (!is_counted || !lpt->is_inner()) continue; 2819 2820 // check for vectorized loops, any reassociation of invariants was already done 2821 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue; 2822 2823 lpt->reassociate_invariants(this); 2824 2825 // Because RCE opportunities can be masked by split_thru_phi, 2826 // look for RCE candidates and inhibit split_thru_phi 2827 // on just their loop-phi's for this pass of loop opts 2828 if (SplitIfBlocks && do_split_ifs) { 2829 if (lpt->policy_range_check(this)) { 2830 lpt->_rce_candidate = 1; // = true 2831 } 2832 } 2833 } 2834 } 2835 2836 // Check for aggressive application of split-if and other transforms 2837 // that require basic-block info (like cloning through Phi's) 2838 if( SplitIfBlocks && do_split_ifs ) { 2839 visited.Clear(); 2840 split_if_with_blocks( visited, nstack ); 2841 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); 2842 } 2843 2844 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) { 2845 C->set_major_progress(); 2846 } 2847 2848 // Perform loop predication before iteration splitting 2849 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) { 2850 _ltree_root->_child->loop_predication(this); 2851 } 2852 2853 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { 2854 if (do_intrinsify_fill()) { 2855 C->set_major_progress(); 2856 } 2857 } 2858 2859 // Perform iteration-splitting on inner loops. Split iterations to avoid 2860 // range checks or one-shot null checks. 2861 2862 // If split-if's didn't hack the graph too bad (no CFG changes) 2863 // then do loop opts. 2864 if (C->has_loops() && !C->major_progress()) { 2865 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 2866 _ltree_root->_child->iteration_split( this, worklist ); 2867 // No verify after peeling! GCM has hoisted code out of the loop. 2868 // After peeling, the hoisted code could sink inside the peeled area. 2869 // The peeling code does not try to recompute the best location for 2870 // all the code before the peeled area, so the verify pass will always 2871 // complain about it. 2872 } 2873 // Do verify graph edges in any case 2874 NOT_PRODUCT( C->verify_graph_edges(); ); 2875 2876 if (!do_split_ifs) { 2877 // We saw major progress in Split-If to get here. We forced a 2878 // pass with unrolling and not split-if, however more split-if's 2879 // might make progress. If the unrolling didn't make progress 2880 // then the major-progress flag got cleared and we won't try 2881 // another round of Split-If. In particular the ever-common 2882 // instance-of/check-cast pattern requires at least 2 rounds of 2883 // Split-If to clear out. 2884 C->set_major_progress(); 2885 } 2886 2887 // Repeat loop optimizations if new loops were seen 2888 if (created_loop_node()) { 2889 C->set_major_progress(); 2890 } 2891 2892 // Keep loop predicates and perform optimizations with them 2893 // until no more loop optimizations could be done. 2894 // After that switch predicates off and do more loop optimizations. 2895 if (!C->major_progress() && (C->predicate_count() > 0)) { 2896 C->cleanup_loop_predicates(_igvn); 2897 if (TraceLoopOpts) { 2898 tty->print_cr("PredicatesOff"); 2899 } 2900 C->set_major_progress(); 2901 } 2902 2903 // Convert scalar to superword operations at the end of all loop opts. 2904 if (UseSuperWord && C->has_loops() && !C->major_progress()) { 2905 // SuperWord transform 2906 SuperWord sw(this); 2907 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2908 IdealLoopTree* lpt = iter.current(); 2909 if (lpt->is_counted()) { 2910 CountedLoopNode *cl = lpt->_head->as_CountedLoop(); 2911 2912 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) { 2913 // Check that the rce'd post loop is encountered first, multiversion after all 2914 // major main loop optimization are concluded 2915 if (!C->major_progress()) { 2916 IdealLoopTree *lpt_next = lpt->_next; 2917 if (lpt_next && lpt_next->is_counted()) { 2918 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop(); 2919 has_range_checks(lpt_next); 2920 if (cl->is_post_loop() && cl->range_checks_present()) { 2921 if (!cl->is_multiversioned()) { 2922 if (multi_version_post_loops(lpt, lpt_next) == false) { 2923 // Cause the rce loop to be optimized away if we fail 2924 cl->mark_is_multiversioned(); 2925 cl->set_slp_max_unroll(0); 2926 poison_rce_post_loop(lpt); 2927 } 2928 } 2929 } 2930 } 2931 sw.transform_loop(lpt, true); 2932 } 2933 } else if (cl->is_main_loop()) { 2934 sw.transform_loop(lpt, true); 2935 } 2936 } 2937 } 2938 } 2939 2940 // Cleanup any modified bits 2941 _igvn.optimize(); 2942 2943 // disable assert until issue with split_flow_path is resolved (6742111) 2944 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), 2945 // "shouldn't introduce irreducible loops"); 2946 2947 if (C->log() != NULL) { 2948 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2949 } 2950 } 2951 2952 #ifndef PRODUCT 2953 //------------------------------print_statistics------------------------------- 2954 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes 2955 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique 2956 void PhaseIdealLoop::print_statistics() { 2957 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); 2958 } 2959 2960 //------------------------------verify----------------------------------------- 2961 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. 2962 static int fail; // debug only, so its multi-thread dont care 2963 void PhaseIdealLoop::verify() const { 2964 int old_progress = C->major_progress(); 2965 ResourceMark rm; 2966 PhaseIdealLoop loop_verify( _igvn, this ); 2967 VectorSet visited(Thread::current()->resource_area()); 2968 2969 fail = 0; 2970 verify_compare( C->root(), &loop_verify, visited ); 2971 assert( fail == 0, "verify loops failed" ); 2972 // Verify loop structure is the same 2973 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); 2974 // Reset major-progress. It was cleared by creating a verify version of 2975 // PhaseIdealLoop. 2976 for( int i=0; i<old_progress; i++ ) 2977 C->set_major_progress(); 2978 } 2979 2980 //------------------------------verify_compare--------------------------------- 2981 // Make sure me and the given PhaseIdealLoop agree on key data structures 2982 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { 2983 if( !n ) return; 2984 if( visited.test_set( n->_idx ) ) return; 2985 if( !_nodes[n->_idx] ) { // Unreachable 2986 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); 2987 return; 2988 } 2989 2990 uint i; 2991 for( i = 0; i < n->req(); i++ ) 2992 verify_compare( n->in(i), loop_verify, visited ); 2993 2994 // Check the '_nodes' block/loop structure 2995 i = n->_idx; 2996 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl 2997 if( _nodes[i] != loop_verify->_nodes[i] && 2998 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { 2999 tty->print("Mismatched control setting for: "); 3000 n->dump(); 3001 if( fail++ > 10 ) return; 3002 Node *c = get_ctrl_no_update(n); 3003 tty->print("We have it as: "); 3004 if( c->in(0) ) c->dump(); 3005 else tty->print_cr("N%d",c->_idx); 3006 tty->print("Verify thinks: "); 3007 if( loop_verify->has_ctrl(n) ) 3008 loop_verify->get_ctrl_no_update(n)->dump(); 3009 else 3010 loop_verify->get_loop_idx(n)->dump(); 3011 tty->cr(); 3012 } 3013 } else { // We have a loop 3014 IdealLoopTree *us = get_loop_idx(n); 3015 if( loop_verify->has_ctrl(n) ) { 3016 tty->print("Mismatched loop setting for: "); 3017 n->dump(); 3018 if( fail++ > 10 ) return; 3019 tty->print("We have it as: "); 3020 us->dump(); 3021 tty->print("Verify thinks: "); 3022 loop_verify->get_ctrl_no_update(n)->dump(); 3023 tty->cr(); 3024 } else if (!C->major_progress()) { 3025 // Loop selection can be messed up if we did a major progress 3026 // operation, like split-if. Do not verify in that case. 3027 IdealLoopTree *them = loop_verify->get_loop_idx(n); 3028 if( us->_head != them->_head || us->_tail != them->_tail ) { 3029 tty->print("Unequals loops for: "); 3030 n->dump(); 3031 if( fail++ > 10 ) return; 3032 tty->print("We have it as: "); 3033 us->dump(); 3034 tty->print("Verify thinks: "); 3035 them->dump(); 3036 tty->cr(); 3037 } 3038 } 3039 } 3040 3041 // Check for immediate dominators being equal 3042 if( i >= _idom_size ) { 3043 if( !n->is_CFG() ) return; 3044 tty->print("CFG Node with no idom: "); 3045 n->dump(); 3046 return; 3047 } 3048 if( !n->is_CFG() ) return; 3049 if( n == C->root() ) return; // No IDOM here 3050 3051 assert(n->_idx == i, "sanity"); 3052 Node *id = idom_no_update(n); 3053 if( id != loop_verify->idom_no_update(n) ) { 3054 tty->print("Unequals idoms for: "); 3055 n->dump(); 3056 if( fail++ > 10 ) return; 3057 tty->print("We have it as: "); 3058 id->dump(); 3059 tty->print("Verify thinks: "); 3060 loop_verify->idom_no_update(n)->dump(); 3061 tty->cr(); 3062 } 3063 3064 } 3065 3066 //------------------------------verify_tree------------------------------------ 3067 // Verify that tree structures match. Because the CFG can change, siblings 3068 // within the loop tree can be reordered. We attempt to deal with that by 3069 // reordering the verify's loop tree if possible. 3070 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { 3071 assert( _parent == parent, "Badly formed loop tree" ); 3072 3073 // Siblings not in same order? Attempt to re-order. 3074 if( _head != loop->_head ) { 3075 // Find _next pointer to update 3076 IdealLoopTree **pp = &loop->_parent->_child; 3077 while( *pp != loop ) 3078 pp = &((*pp)->_next); 3079 // Find proper sibling to be next 3080 IdealLoopTree **nn = &loop->_next; 3081 while( (*nn) && (*nn)->_head != _head ) 3082 nn = &((*nn)->_next); 3083 3084 // Check for no match. 3085 if( !(*nn) ) { 3086 // Annoyingly, irreducible loops can pick different headers 3087 // after a major_progress operation, so the rest of the loop 3088 // tree cannot be matched. 3089 if (_irreducible && Compile::current()->major_progress()) return; 3090 assert( 0, "failed to match loop tree" ); 3091 } 3092 3093 // Move (*nn) to (*pp) 3094 IdealLoopTree *hit = *nn; 3095 *nn = hit->_next; 3096 hit->_next = loop; 3097 *pp = loop; 3098 loop = hit; 3099 // Now try again to verify 3100 } 3101 3102 assert( _head == loop->_head , "mismatched loop head" ); 3103 Node *tail = _tail; // Inline a non-updating version of 3104 while( !tail->in(0) ) // the 'tail()' call. 3105 tail = tail->in(1); 3106 assert( tail == loop->_tail, "mismatched loop tail" ); 3107 3108 // Counted loops that are guarded should be able to find their guards 3109 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { 3110 CountedLoopNode *cl = _head->as_CountedLoop(); 3111 Node *init = cl->init_trip(); 3112 Node *ctrl = cl->in(LoopNode::EntryControl); 3113 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 3114 Node *iff = ctrl->in(0); 3115 assert( iff->Opcode() == Op_If, "" ); 3116 Node *bol = iff->in(1); 3117 assert( bol->Opcode() == Op_Bool, "" ); 3118 Node *cmp = bol->in(1); 3119 assert( cmp->Opcode() == Op_CmpI, "" ); 3120 Node *add = cmp->in(1); 3121 Node *opaq; 3122 if( add->Opcode() == Op_Opaque1 ) { 3123 opaq = add; 3124 } else { 3125 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); 3126 assert( add == init, "" ); 3127 opaq = cmp->in(2); 3128 } 3129 assert( opaq->Opcode() == Op_Opaque1, "" ); 3130 3131 } 3132 3133 if (_child != NULL) _child->verify_tree(loop->_child, this); 3134 if (_next != NULL) _next ->verify_tree(loop->_next, parent); 3135 // Innermost loops need to verify loop bodies, 3136 // but only if no 'major_progress' 3137 int fail = 0; 3138 if (!Compile::current()->major_progress() && _child == NULL) { 3139 for( uint i = 0; i < _body.size(); i++ ) { 3140 Node *n = _body.at(i); 3141 if (n->outcnt() == 0) continue; // Ignore dead 3142 uint j; 3143 for( j = 0; j < loop->_body.size(); j++ ) 3144 if( loop->_body.at(j) == n ) 3145 break; 3146 if( j == loop->_body.size() ) { // Not found in loop body 3147 // Last ditch effort to avoid assertion: Its possible that we 3148 // have some users (so outcnt not zero) but are still dead. 3149 // Try to find from root. 3150 if (Compile::current()->root()->find(n->_idx)) { 3151 fail++; 3152 tty->print("We have that verify does not: "); 3153 n->dump(); 3154 } 3155 } 3156 } 3157 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { 3158 Node *n = loop->_body.at(i2); 3159 if (n->outcnt() == 0) continue; // Ignore dead 3160 uint j; 3161 for( j = 0; j < _body.size(); j++ ) 3162 if( _body.at(j) == n ) 3163 break; 3164 if( j == _body.size() ) { // Not found in loop body 3165 // Last ditch effort to avoid assertion: Its possible that we 3166 // have some users (so outcnt not zero) but are still dead. 3167 // Try to find from root. 3168 if (Compile::current()->root()->find(n->_idx)) { 3169 fail++; 3170 tty->print("Verify has that we do not: "); 3171 n->dump(); 3172 } 3173 } 3174 } 3175 assert( !fail, "loop body mismatch" ); 3176 } 3177 } 3178 3179 #endif 3180 3181 //------------------------------set_idom--------------------------------------- 3182 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { 3183 uint idx = d->_idx; 3184 if (idx >= _idom_size) { 3185 uint newsize = _idom_size<<1; 3186 while( idx >= newsize ) { 3187 newsize <<= 1; 3188 } 3189 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); 3190 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); 3191 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); 3192 _idom_size = newsize; 3193 } 3194 _idom[idx] = n; 3195 _dom_depth[idx] = dom_depth; 3196 } 3197 3198 //------------------------------recompute_dom_depth--------------------------------------- 3199 // The dominator tree is constructed with only parent pointers. 3200 // This recomputes the depth in the tree by first tagging all 3201 // nodes as "no depth yet" marker. The next pass then runs up 3202 // the dom tree from each node marked "no depth yet", and computes 3203 // the depth on the way back down. 3204 void PhaseIdealLoop::recompute_dom_depth() { 3205 uint no_depth_marker = C->unique(); 3206 uint i; 3207 // Initialize depth to "no depth yet" 3208 for (i = 0; i < _idom_size; i++) { 3209 if (_dom_depth[i] > 0 && _idom[i] != NULL) { 3210 _dom_depth[i] = no_depth_marker; 3211 } 3212 } 3213 if (_dom_stk == NULL) { 3214 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size. 3215 if (init_size < 10) init_size = 10; 3216 _dom_stk = new GrowableArray<uint>(init_size); 3217 } 3218 // Compute new depth for each node. 3219 for (i = 0; i < _idom_size; i++) { 3220 uint j = i; 3221 // Run up the dom tree to find a node with a depth 3222 while (_dom_depth[j] == no_depth_marker) { 3223 _dom_stk->push(j); 3224 j = _idom[j]->_idx; 3225 } 3226 // Compute the depth on the way back down this tree branch 3227 uint dd = _dom_depth[j] + 1; 3228 while (_dom_stk->length() > 0) { 3229 uint j = _dom_stk->pop(); 3230 _dom_depth[j] = dd; 3231 dd++; 3232 } 3233 } 3234 } 3235 3236 //------------------------------sort------------------------------------------- 3237 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the 3238 // loop tree, not the root. 3239 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { 3240 if( !innermost ) return loop; // New innermost loop 3241 3242 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number 3243 assert( loop_preorder, "not yet post-walked loop" ); 3244 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer 3245 IdealLoopTree *l = *pp; // Do I go before or after 'l'? 3246 3247 // Insert at start of list 3248 while( l ) { // Insertion sort based on pre-order 3249 if( l == loop ) return innermost; // Already on list! 3250 int l_preorder = get_preorder(l->_head); // Cache pre-order number 3251 assert( l_preorder, "not yet post-walked l" ); 3252 // Check header pre-order number to figure proper nesting 3253 if( loop_preorder > l_preorder ) 3254 break; // End of insertion 3255 // If headers tie (e.g., shared headers) check tail pre-order numbers. 3256 // Since I split shared headers, you'd think this could not happen. 3257 // BUT: I must first do the preorder numbering before I can discover I 3258 // have shared headers, so the split headers all get the same preorder 3259 // number as the RegionNode they split from. 3260 if( loop_preorder == l_preorder && 3261 get_preorder(loop->_tail) < get_preorder(l->_tail) ) 3262 break; // Also check for shared headers (same pre#) 3263 pp = &l->_parent; // Chain up list 3264 l = *pp; 3265 } 3266 // Link into list 3267 // Point predecessor to me 3268 *pp = loop; 3269 // Point me to successor 3270 IdealLoopTree *p = loop->_parent; 3271 loop->_parent = l; // Point me to successor 3272 if( p ) sort( p, innermost ); // Insert my parents into list as well 3273 return innermost; 3274 } 3275 3276 //------------------------------build_loop_tree-------------------------------- 3277 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit 3278 // bits. The _nodes[] array is mapped by Node index and holds a NULL for 3279 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the 3280 // tightest enclosing IdealLoopTree for post-walked. 3281 // 3282 // During my forward walk I do a short 1-layer lookahead to see if I can find 3283 // a loop backedge with that doesn't have any work on the backedge. This 3284 // helps me construct nested loops with shared headers better. 3285 // 3286 // Once I've done the forward recursion, I do the post-work. For each child 3287 // I check to see if there is a backedge. Backedges define a loop! I 3288 // insert an IdealLoopTree at the target of the backedge. 3289 // 3290 // During the post-work I also check to see if I have several children 3291 // belonging to different loops. If so, then this Node is a decision point 3292 // where control flow can choose to change loop nests. It is at this 3293 // decision point where I can figure out how loops are nested. At this 3294 // time I can properly order the different loop nests from my children. 3295 // Note that there may not be any backedges at the decision point! 3296 // 3297 // Since the decision point can be far removed from the backedges, I can't 3298 // order my loops at the time I discover them. Thus at the decision point 3299 // I need to inspect loop header pre-order numbers to properly nest my 3300 // loops. This means I need to sort my childrens' loops by pre-order. 3301 // The sort is of size number-of-control-children, which generally limits 3302 // it to size 2 (i.e., I just choose between my 2 target loops). 3303 void PhaseIdealLoop::build_loop_tree() { 3304 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3305 GrowableArray <Node *> bltstack(C->live_nodes() >> 1); 3306 Node *n = C->root(); 3307 bltstack.push(n); 3308 int pre_order = 1; 3309 int stack_size; 3310 3311 while ( ( stack_size = bltstack.length() ) != 0 ) { 3312 n = bltstack.top(); // Leave node on stack 3313 if ( !is_visited(n) ) { 3314 // ---- Pre-pass Work ---- 3315 // Pre-walked but not post-walked nodes need a pre_order number. 3316 3317 set_preorder_visited( n, pre_order ); // set as visited 3318 3319 // ---- Scan over children ---- 3320 // Scan first over control projections that lead to loop headers. 3321 // This helps us find inner-to-outer loops with shared headers better. 3322 3323 // Scan children's children for loop headers. 3324 for ( int i = n->outcnt() - 1; i >= 0; --i ) { 3325 Node* m = n->raw_out(i); // Child 3326 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children 3327 // Scan over children's children to find loop 3328 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3329 Node* l = m->fast_out(j); 3330 if( is_visited(l) && // Been visited? 3331 !is_postvisited(l) && // But not post-visited 3332 get_preorder(l) < pre_order ) { // And smaller pre-order 3333 // Found! Scan the DFS down this path before doing other paths 3334 bltstack.push(m); 3335 break; 3336 } 3337 } 3338 } 3339 } 3340 pre_order++; 3341 } 3342 else if ( !is_postvisited(n) ) { 3343 // Note: build_loop_tree_impl() adds out edges on rare occasions, 3344 // such as com.sun.rsasign.am::a. 3345 // For non-recursive version, first, process current children. 3346 // On next iteration, check if additional children were added. 3347 for ( int k = n->outcnt() - 1; k >= 0; --k ) { 3348 Node* u = n->raw_out(k); 3349 if ( u->is_CFG() && !is_visited(u) ) { 3350 bltstack.push(u); 3351 } 3352 } 3353 if ( bltstack.length() == stack_size ) { 3354 // There were no additional children, post visit node now 3355 (void)bltstack.pop(); // Remove node from stack 3356 pre_order = build_loop_tree_impl( n, pre_order ); 3357 // Check for bailout 3358 if (C->failing()) { 3359 return; 3360 } 3361 // Check to grow _preorders[] array for the case when 3362 // build_loop_tree_impl() adds new nodes. 3363 check_grow_preorders(); 3364 } 3365 } 3366 else { 3367 (void)bltstack.pop(); // Remove post-visited node from stack 3368 } 3369 } 3370 } 3371 3372 //------------------------------build_loop_tree_impl--------------------------- 3373 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { 3374 // ---- Post-pass Work ---- 3375 // Pre-walked but not post-walked nodes need a pre_order number. 3376 3377 // Tightest enclosing loop for this Node 3378 IdealLoopTree *innermost = NULL; 3379 3380 // For all children, see if any edge is a backedge. If so, make a loop 3381 // for it. Then find the tightest enclosing loop for the self Node. 3382 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3383 Node* m = n->fast_out(i); // Child 3384 if( n == m ) continue; // Ignore control self-cycles 3385 if( !m->is_CFG() ) continue;// Ignore non-CFG edges 3386 3387 IdealLoopTree *l; // Child's loop 3388 if( !is_postvisited(m) ) { // Child visited but not post-visited? 3389 // Found a backedge 3390 assert( get_preorder(m) < pre_order, "should be backedge" ); 3391 // Check for the RootNode, which is already a LoopNode and is allowed 3392 // to have multiple "backedges". 3393 if( m == C->root()) { // Found the root? 3394 l = _ltree_root; // Root is the outermost LoopNode 3395 } else { // Else found a nested loop 3396 // Insert a LoopNode to mark this loop. 3397 l = new IdealLoopTree(this, m, n); 3398 } // End of Else found a nested loop 3399 if( !has_loop(m) ) // If 'm' does not already have a loop set 3400 set_loop(m, l); // Set loop header to loop now 3401 3402 } else { // Else not a nested loop 3403 if( !_nodes[m->_idx] ) continue; // Dead code has no loop 3404 l = get_loop(m); // Get previously determined loop 3405 // If successor is header of a loop (nest), move up-loop till it 3406 // is a member of some outer enclosing loop. Since there are no 3407 // shared headers (I've split them already) I only need to go up 3408 // at most 1 level. 3409 while( l && l->_head == m ) // Successor heads loop? 3410 l = l->_parent; // Move up 1 for me 3411 // If this loop is not properly parented, then this loop 3412 // has no exit path out, i.e. its an infinite loop. 3413 if( !l ) { 3414 // Make loop "reachable" from root so the CFG is reachable. Basically 3415 // insert a bogus loop exit that is never taken. 'm', the loop head, 3416 // points to 'n', one (of possibly many) fall-in paths. There may be 3417 // many backedges as well. 3418 3419 // Here I set the loop to be the root loop. I could have, after 3420 // inserting a bogus loop exit, restarted the recursion and found my 3421 // new loop exit. This would make the infinite loop a first-class 3422 // loop and it would then get properly optimized. What's the use of 3423 // optimizing an infinite loop? 3424 l = _ltree_root; // Oops, found infinite loop 3425 3426 if (!_verify_only) { 3427 // Insert the NeverBranch between 'm' and it's control user. 3428 NeverBranchNode *iff = new NeverBranchNode( m ); 3429 _igvn.register_new_node_with_optimizer(iff); 3430 set_loop(iff, l); 3431 Node *if_t = new CProjNode( iff, 0 ); 3432 _igvn.register_new_node_with_optimizer(if_t); 3433 set_loop(if_t, l); 3434 3435 Node* cfg = NULL; // Find the One True Control User of m 3436 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3437 Node* x = m->fast_out(j); 3438 if (x->is_CFG() && x != m && x != iff) 3439 { cfg = x; break; } 3440 } 3441 assert(cfg != NULL, "must find the control user of m"); 3442 uint k = 0; // Probably cfg->in(0) 3443 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region 3444 cfg->set_req( k, if_t ); // Now point to NeverBranch 3445 _igvn._worklist.push(cfg); 3446 3447 // Now create the never-taken loop exit 3448 Node *if_f = new CProjNode( iff, 1 ); 3449 _igvn.register_new_node_with_optimizer(if_f); 3450 set_loop(if_f, l); 3451 // Find frame ptr for Halt. Relies on the optimizer 3452 // V-N'ing. Easier and quicker than searching through 3453 // the program structure. 3454 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr ); 3455 _igvn.register_new_node_with_optimizer(frame); 3456 // Halt & Catch Fire 3457 Node *halt = new HaltNode( if_f, frame ); 3458 _igvn.register_new_node_with_optimizer(halt); 3459 set_loop(halt, l); 3460 C->root()->add_req(halt); 3461 } 3462 set_loop(C->root(), _ltree_root); 3463 } 3464 } 3465 // Weeny check for irreducible. This child was already visited (this 3466 // IS the post-work phase). Is this child's loop header post-visited 3467 // as well? If so, then I found another entry into the loop. 3468 if (!_verify_only) { 3469 while( is_postvisited(l->_head) ) { 3470 // found irreducible 3471 l->_irreducible = 1; // = true 3472 l = l->_parent; 3473 _has_irreducible_loops = true; 3474 // Check for bad CFG here to prevent crash, and bailout of compile 3475 if (l == NULL) { 3476 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); 3477 return pre_order; 3478 } 3479 } 3480 C->set_has_irreducible_loop(_has_irreducible_loops); 3481 } 3482 3483 // This Node might be a decision point for loops. It is only if 3484 // it's children belong to several different loops. The sort call 3485 // does a trivial amount of work if there is only 1 child or all 3486 // children belong to the same loop. If however, the children 3487 // belong to different loops, the sort call will properly set the 3488 // _parent pointers to show how the loops nest. 3489 // 3490 // In any case, it returns the tightest enclosing loop. 3491 innermost = sort( l, innermost ); 3492 } 3493 3494 // Def-use info will have some dead stuff; dead stuff will have no 3495 // loop decided on. 3496 3497 // Am I a loop header? If so fix up my parent's child and next ptrs. 3498 if( innermost && innermost->_head == n ) { 3499 assert( get_loop(n) == innermost, "" ); 3500 IdealLoopTree *p = innermost->_parent; 3501 IdealLoopTree *l = innermost; 3502 while( p && l->_head == n ) { 3503 l->_next = p->_child; // Put self on parents 'next child' 3504 p->_child = l; // Make self as first child of parent 3505 l = p; // Now walk up the parent chain 3506 p = l->_parent; 3507 } 3508 } else { 3509 // Note that it is possible for a LoopNode to reach here, if the 3510 // backedge has been made unreachable (hence the LoopNode no longer 3511 // denotes a Loop, and will eventually be removed). 3512 3513 // Record tightest enclosing loop for self. Mark as post-visited. 3514 set_loop(n, innermost); 3515 // Also record has_call flag early on 3516 if( innermost ) { 3517 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { 3518 // Do not count uncommon calls 3519 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { 3520 Node *iff = n->in(0)->in(0); 3521 // No any calls for vectorized loops. 3522 if( UseSuperWord || !iff->is_If() || 3523 (n->in(0)->Opcode() == Op_IfFalse && 3524 (1.0 - iff->as_If()->_prob) >= 0.01) || 3525 (iff->as_If()->_prob >= 0.01) ) 3526 innermost->_has_call = 1; 3527 } 3528 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { 3529 // Disable loop optimizations if the loop has a scalar replaceable 3530 // allocation. This disabling may cause a potential performance lost 3531 // if the allocation is not eliminated for some reason. 3532 innermost->_allow_optimizations = false; 3533 innermost->_has_call = 1; // = true 3534 } else if (n->Opcode() == Op_SafePoint) { 3535 // Record all safepoints in this loop. 3536 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List(); 3537 innermost->_safepts->push(n); 3538 } 3539 } 3540 } 3541 3542 // Flag as post-visited now 3543 set_postvisited(n); 3544 return pre_order; 3545 } 3546 3547 3548 //------------------------------build_loop_early------------------------------- 3549 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3550 // First pass computes the earliest controlling node possible. This is the 3551 // controlling input with the deepest dominating depth. 3552 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3553 while (worklist.size() != 0) { 3554 // Use local variables nstack_top_n & nstack_top_i to cache values 3555 // on nstack's top. 3556 Node *nstack_top_n = worklist.pop(); 3557 uint nstack_top_i = 0; 3558 //while_nstack_nonempty: 3559 while (true) { 3560 // Get parent node and next input's index from stack's top. 3561 Node *n = nstack_top_n; 3562 uint i = nstack_top_i; 3563 uint cnt = n->req(); // Count of inputs 3564 if (i == 0) { // Pre-process the node. 3565 if( has_node(n) && // Have either loop or control already? 3566 !has_ctrl(n) ) { // Have loop picked out already? 3567 // During "merge_many_backedges" we fold up several nested loops 3568 // into a single loop. This makes the members of the original 3569 // loop bodies pointing to dead loops; they need to move up 3570 // to the new UNION'd larger loop. I set the _head field of these 3571 // dead loops to NULL and the _parent field points to the owning 3572 // loop. Shades of UNION-FIND algorithm. 3573 IdealLoopTree *ilt; 3574 while( !(ilt = get_loop(n))->_head ) { 3575 // Normally I would use a set_loop here. But in this one special 3576 // case, it is legal (and expected) to change what loop a Node 3577 // belongs to. 3578 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); 3579 } 3580 // Remove safepoints ONLY if I've already seen I don't need one. 3581 // (the old code here would yank a 2nd safepoint after seeing a 3582 // first one, even though the 1st did not dominate in the loop body 3583 // and thus could be avoided indefinitely) 3584 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && 3585 is_deleteable_safept(n)) { 3586 Node *in = n->in(TypeFunc::Control); 3587 lazy_replace(n,in); // Pull safepoint now 3588 if (ilt->_safepts != NULL) { 3589 ilt->_safepts->yank(n); 3590 } 3591 // Carry on with the recursion "as if" we are walking 3592 // only the control input 3593 if( !visited.test_set( in->_idx ) ) { 3594 worklist.push(in); // Visit this guy later, using worklist 3595 } 3596 // Get next node from nstack: 3597 // - skip n's inputs processing by setting i > cnt; 3598 // - we also will not call set_early_ctrl(n) since 3599 // has_node(n) == true (see the condition above). 3600 i = cnt + 1; 3601 } 3602 } 3603 } // if (i == 0) 3604 3605 // Visit all inputs 3606 bool done = true; // Assume all n's inputs will be processed 3607 while (i < cnt) { 3608 Node *in = n->in(i); 3609 ++i; 3610 if (in == NULL) continue; 3611 if (in->pinned() && !in->is_CFG()) 3612 set_ctrl(in, in->in(0)); 3613 int is_visited = visited.test_set( in->_idx ); 3614 if (!has_node(in)) { // No controlling input yet? 3615 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); 3616 assert( !is_visited, "visit only once" ); 3617 nstack.push(n, i); // Save parent node and next input's index. 3618 nstack_top_n = in; // Process current input now. 3619 nstack_top_i = 0; 3620 done = false; // Not all n's inputs processed. 3621 break; // continue while_nstack_nonempty; 3622 } else if (!is_visited) { 3623 // This guy has a location picked out for him, but has not yet 3624 // been visited. Happens to all CFG nodes, for instance. 3625 // Visit him using the worklist instead of recursion, to break 3626 // cycles. Since he has a location already we do not need to 3627 // find his location before proceeding with the current Node. 3628 worklist.push(in); // Visit this guy later, using worklist 3629 } 3630 } 3631 if (done) { 3632 // All of n's inputs have been processed, complete post-processing. 3633 3634 // Compute earliest point this Node can go. 3635 // CFG, Phi, pinned nodes already know their controlling input. 3636 if (!has_node(n)) { 3637 // Record earliest legal location 3638 set_early_ctrl( n ); 3639 } 3640 if (nstack.is_empty()) { 3641 // Finished all nodes on stack. 3642 // Process next node on the worklist. 3643 break; 3644 } 3645 // Get saved parent node and next input's index. 3646 nstack_top_n = nstack.node(); 3647 nstack_top_i = nstack.index(); 3648 nstack.pop(); 3649 } 3650 } // while (true) 3651 } 3652 } 3653 3654 //------------------------------dom_lca_internal-------------------------------- 3655 // Pair-wise LCA 3656 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { 3657 if( !n1 ) return n2; // Handle NULL original LCA 3658 assert( n1->is_CFG(), "" ); 3659 assert( n2->is_CFG(), "" ); 3660 // find LCA of all uses 3661 uint d1 = dom_depth(n1); 3662 uint d2 = dom_depth(n2); 3663 while (n1 != n2) { 3664 if (d1 > d2) { 3665 n1 = idom(n1); 3666 d1 = dom_depth(n1); 3667 } else if (d1 < d2) { 3668 n2 = idom(n2); 3669 d2 = dom_depth(n2); 3670 } else { 3671 // Here d1 == d2. Due to edits of the dominator-tree, sections 3672 // of the tree might have the same depth. These sections have 3673 // to be searched more carefully. 3674 3675 // Scan up all the n1's with equal depth, looking for n2. 3676 Node *t1 = idom(n1); 3677 while (dom_depth(t1) == d1) { 3678 if (t1 == n2) return n2; 3679 t1 = idom(t1); 3680 } 3681 // Scan up all the n2's with equal depth, looking for n1. 3682 Node *t2 = idom(n2); 3683 while (dom_depth(t2) == d2) { 3684 if (t2 == n1) return n1; 3685 t2 = idom(t2); 3686 } 3687 // Move up to a new dominator-depth value as well as up the dom-tree. 3688 n1 = t1; 3689 n2 = t2; 3690 d1 = dom_depth(n1); 3691 d2 = dom_depth(n2); 3692 } 3693 } 3694 return n1; 3695 } 3696 3697 //------------------------------compute_idom----------------------------------- 3698 // Locally compute IDOM using dom_lca call. Correct only if the incoming 3699 // IDOMs are correct. 3700 Node *PhaseIdealLoop::compute_idom( Node *region ) const { 3701 assert( region->is_Region(), "" ); 3702 Node *LCA = NULL; 3703 for( uint i = 1; i < region->req(); i++ ) { 3704 if( region->in(i) != C->top() ) 3705 LCA = dom_lca( LCA, region->in(i) ); 3706 } 3707 return LCA; 3708 } 3709 3710 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { 3711 bool had_error = false; 3712 #ifdef ASSERT 3713 if (early != C->root()) { 3714 // Make sure that there's a dominance path from LCA to early 3715 Node* d = LCA; 3716 while (d != early) { 3717 if (d == C->root()) { 3718 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA); 3719 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx); 3720 had_error = true; 3721 break; 3722 } 3723 d = idom(d); 3724 } 3725 } 3726 #endif 3727 return had_error; 3728 } 3729 3730 3731 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { 3732 // Compute LCA over list of uses 3733 bool had_error = false; 3734 Node *LCA = NULL; 3735 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { 3736 Node* c = n->fast_out(i); 3737 if (_nodes[c->_idx] == NULL) 3738 continue; // Skip the occasional dead node 3739 if( c->is_Phi() ) { // For Phis, we must land above on the path 3740 for( uint j=1; j<c->req(); j++ ) {// For all inputs 3741 if( c->in(j) == n ) { // Found matching input? 3742 Node *use = c->in(0)->in(j); 3743 if (_verify_only && use->is_top()) continue; 3744 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3745 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3746 } 3747 } 3748 } else { 3749 // For CFG data-users, use is in the block just prior 3750 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); 3751 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3752 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3753 } 3754 } 3755 assert(!had_error, "bad dominance"); 3756 return LCA; 3757 } 3758 3759 // Check the shape of the graph at the loop entry. In some cases, 3760 // the shape of the graph does not match the shape outlined below. 3761 // That is caused by the Opaque1 node "protecting" the shape of 3762 // the graph being removed by, for example, the IGVN performed 3763 // in PhaseIdealLoop::build_and_optimize(). 3764 // 3765 // After the Opaque1 node has been removed, optimizations (e.g., split-if, 3766 // loop unswitching, and IGVN, or a combination of them) can freely change 3767 // the graph's shape. As a result, the graph shape outlined below cannot 3768 // be guaranteed anymore. 3769 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) { 3770 if (!cl->is_main_loop() && !cl->is_post_loop()) { 3771 return false; 3772 } 3773 Node* ctrl = cl->skip_strip_mined()->in(LoopNode::EntryControl); 3774 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) { 3775 return false; 3776 } 3777 Node* iffm = ctrl->in(0); 3778 if (iffm == NULL || !iffm->is_If()) { 3779 return false; 3780 } 3781 Node* bolzm = iffm->in(1); 3782 if (bolzm == NULL || !bolzm->is_Bool()) { 3783 return false; 3784 } 3785 Node* cmpzm = bolzm->in(1); 3786 if (cmpzm == NULL || !cmpzm->is_Cmp()) { 3787 return false; 3788 } 3789 // compares can get conditionally flipped 3790 bool found_opaque = false; 3791 for (uint i = 1; i < cmpzm->req(); i++) { 3792 Node* opnd = cmpzm->in(i); 3793 if (opnd && opnd->Opcode() == Op_Opaque1) { 3794 found_opaque = true; 3795 break; 3796 } 3797 } 3798 if (!found_opaque) { 3799 return false; 3800 } 3801 return true; 3802 } 3803 3804 //------------------------------get_late_ctrl---------------------------------- 3805 // Compute latest legal control. 3806 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { 3807 assert(early != NULL, "early control should not be NULL"); 3808 3809 Node* LCA = compute_lca_of_uses(n, early); 3810 #ifdef ASSERT 3811 if (LCA == C->root() && LCA != early) { 3812 // def doesn't dominate uses so print some useful debugging output 3813 compute_lca_of_uses(n, early, true); 3814 } 3815 #endif 3816 3817 // if this is a load, check for anti-dependent stores 3818 // We use a conservative algorithm to identify potential interfering 3819 // instructions and for rescheduling the load. The users of the memory 3820 // input of this load are examined. Any use which is not a load and is 3821 // dominated by early is considered a potentially interfering store. 3822 // This can produce false positives. 3823 if (n->is_Load() && LCA != early) { 3824 Node_List worklist; 3825 3826 Node *mem = n->in(MemNode::Memory); 3827 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 3828 Node* s = mem->fast_out(i); 3829 worklist.push(s); 3830 } 3831 while(worklist.size() != 0 && LCA != early) { 3832 Node* s = worklist.pop(); 3833 if (s->is_Load() || s->Opcode() == Op_SafePoint) { 3834 continue; 3835 } else if (s->is_MergeMem()) { 3836 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 3837 Node* s1 = s->fast_out(i); 3838 worklist.push(s1); 3839 } 3840 } else { 3841 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); 3842 assert(sctrl != NULL || s->outcnt() == 0, "must have control"); 3843 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { 3844 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); 3845 } 3846 } 3847 } 3848 } 3849 3850 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); 3851 return LCA; 3852 } 3853 3854 // true if CFG node d dominates CFG node n 3855 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { 3856 if (d == n) 3857 return true; 3858 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 3859 uint dd = dom_depth(d); 3860 while (dom_depth(n) >= dd) { 3861 if (n == d) 3862 return true; 3863 n = idom(n); 3864 } 3865 return false; 3866 } 3867 3868 //------------------------------dom_lca_for_get_late_ctrl_internal------------- 3869 // Pair-wise LCA with tags. 3870 // Tag each index with the node 'tag' currently being processed 3871 // before advancing up the dominator chain using idom(). 3872 // Later calls that find a match to 'tag' know that this path has already 3873 // been considered in the current LCA (which is input 'n1' by convention). 3874 // Since get_late_ctrl() is only called once for each node, the tag array 3875 // does not need to be cleared between calls to get_late_ctrl(). 3876 // Algorithm trades a larger constant factor for better asymptotic behavior 3877 // 3878 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { 3879 uint d1 = dom_depth(n1); 3880 uint d2 = dom_depth(n2); 3881 3882 do { 3883 if (d1 > d2) { 3884 // current lca is deeper than n2 3885 _dom_lca_tags.map(n1->_idx, tag); 3886 n1 = idom(n1); 3887 d1 = dom_depth(n1); 3888 } else if (d1 < d2) { 3889 // n2 is deeper than current lca 3890 Node *memo = _dom_lca_tags[n2->_idx]; 3891 if( memo == tag ) { 3892 return n1; // Return the current LCA 3893 } 3894 _dom_lca_tags.map(n2->_idx, tag); 3895 n2 = idom(n2); 3896 d2 = dom_depth(n2); 3897 } else { 3898 // Here d1 == d2. Due to edits of the dominator-tree, sections 3899 // of the tree might have the same depth. These sections have 3900 // to be searched more carefully. 3901 3902 // Scan up all the n1's with equal depth, looking for n2. 3903 _dom_lca_tags.map(n1->_idx, tag); 3904 Node *t1 = idom(n1); 3905 while (dom_depth(t1) == d1) { 3906 if (t1 == n2) return n2; 3907 _dom_lca_tags.map(t1->_idx, tag); 3908 t1 = idom(t1); 3909 } 3910 // Scan up all the n2's with equal depth, looking for n1. 3911 _dom_lca_tags.map(n2->_idx, tag); 3912 Node *t2 = idom(n2); 3913 while (dom_depth(t2) == d2) { 3914 if (t2 == n1) return n1; 3915 _dom_lca_tags.map(t2->_idx, tag); 3916 t2 = idom(t2); 3917 } 3918 // Move up to a new dominator-depth value as well as up the dom-tree. 3919 n1 = t1; 3920 n2 = t2; 3921 d1 = dom_depth(n1); 3922 d2 = dom_depth(n2); 3923 } 3924 } while (n1 != n2); 3925 return n1; 3926 } 3927 3928 //------------------------------init_dom_lca_tags------------------------------ 3929 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3930 // Intended use does not involve any growth for the array, so it could 3931 // be of fixed size. 3932 void PhaseIdealLoop::init_dom_lca_tags() { 3933 uint limit = C->unique() + 1; 3934 _dom_lca_tags.map( limit, NULL ); 3935 #ifdef ASSERT 3936 for( uint i = 0; i < limit; ++i ) { 3937 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3938 } 3939 #endif // ASSERT 3940 } 3941 3942 //------------------------------clear_dom_lca_tags------------------------------ 3943 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3944 // Intended use does not involve any growth for the array, so it could 3945 // be of fixed size. 3946 void PhaseIdealLoop::clear_dom_lca_tags() { 3947 uint limit = C->unique() + 1; 3948 _dom_lca_tags.map( limit, NULL ); 3949 _dom_lca_tags.clear(); 3950 #ifdef ASSERT 3951 for( uint i = 0; i < limit; ++i ) { 3952 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3953 } 3954 #endif // ASSERT 3955 } 3956 3957 //------------------------------build_loop_late-------------------------------- 3958 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3959 // Second pass finds latest legal placement, and ideal loop placement. 3960 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3961 while (worklist.size() != 0) { 3962 Node *n = worklist.pop(); 3963 // Only visit once 3964 if (visited.test_set(n->_idx)) continue; 3965 uint cnt = n->outcnt(); 3966 uint i = 0; 3967 while (true) { 3968 assert( _nodes[n->_idx], "no dead nodes" ); 3969 // Visit all children 3970 if (i < cnt) { 3971 Node* use = n->raw_out(i); 3972 ++i; 3973 // Check for dead uses. Aggressively prune such junk. It might be 3974 // dead in the global sense, but still have local uses so I cannot 3975 // easily call 'remove_dead_node'. 3976 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? 3977 // Due to cycles, we might not hit the same fixed point in the verify 3978 // pass as we do in the regular pass. Instead, visit such phis as 3979 // simple uses of the loop head. 3980 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { 3981 if( !visited.test(use->_idx) ) 3982 worklist.push(use); 3983 } else if( !visited.test_set(use->_idx) ) { 3984 nstack.push(n, i); // Save parent and next use's index. 3985 n = use; // Process all children of current use. 3986 cnt = use->outcnt(); 3987 i = 0; 3988 } 3989 } else { 3990 // Do not visit around the backedge of loops via data edges. 3991 // push dead code onto a worklist 3992 _deadlist.push(use); 3993 } 3994 } else { 3995 // All of n's children have been processed, complete post-processing. 3996 build_loop_late_post(n); 3997 if (nstack.is_empty()) { 3998 // Finished all nodes on stack. 3999 // Process next node on the worklist. 4000 break; 4001 } 4002 // Get saved parent node and next use's index. Visit the rest of uses. 4003 n = nstack.node(); 4004 cnt = n->outcnt(); 4005 i = nstack.index(); 4006 nstack.pop(); 4007 } 4008 } 4009 } 4010 } 4011 4012 // Verify that no data node is schedules in the outer loop of a strip 4013 // mined loop. 4014 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) { 4015 #ifdef ASSERT 4016 if (get_loop(least)->_nest == 0) { 4017 return; 4018 } 4019 IdealLoopTree* loop = get_loop(least); 4020 Node* head = loop->_head; 4021 if (head->is_OuterStripMinedLoop()) { 4022 Node* sfpt = head->as_Loop()->outer_safepoint(); 4023 ResourceMark rm; 4024 Unique_Node_List wq; 4025 wq.push(sfpt); 4026 for (uint i = 0; i < wq.size(); i++) { 4027 Node *m = wq.at(i); 4028 for (uint i = 1; i < m->req(); i++) { 4029 Node* nn = m->in(i); 4030 if (nn == n) { 4031 return; 4032 } 4033 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) { 4034 wq.push(nn); 4035 } 4036 } 4037 } 4038 ShouldNotReachHere(); 4039 } 4040 #endif 4041 } 4042 4043 4044 //------------------------------build_loop_late_post--------------------------- 4045 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 4046 // Second pass finds latest legal placement, and ideal loop placement. 4047 void PhaseIdealLoop::build_loop_late_post( Node *n ) { 4048 4049 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) { 4050 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. 4051 } 4052 4053 #ifdef ASSERT 4054 if (_verify_only && !n->is_CFG()) { 4055 // Check def-use domination. 4056 compute_lca_of_uses(n, get_ctrl(n), true /* verify */); 4057 } 4058 #endif 4059 4060 // CFG and pinned nodes already handled 4061 if( n->in(0) ) { 4062 if( n->in(0)->is_top() ) return; // Dead? 4063 4064 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads 4065 // _must_ be pinned (they have to observe their control edge of course). 4066 // Unlike Stores (which modify an unallocable resource, the memory 4067 // state), Mods/Loads can float around. So free them up. 4068 bool pinned = true; 4069 switch( n->Opcode() ) { 4070 case Op_DivI: 4071 case Op_DivF: 4072 case Op_DivD: 4073 case Op_ModI: 4074 case Op_ModF: 4075 case Op_ModD: 4076 case Op_LoadB: // Same with Loads; they can sink 4077 case Op_LoadUB: // during loop optimizations. 4078 case Op_LoadUS: 4079 case Op_LoadD: 4080 case Op_LoadF: 4081 case Op_LoadI: 4082 case Op_LoadKlass: 4083 case Op_LoadNKlass: 4084 case Op_LoadL: 4085 case Op_LoadS: 4086 case Op_LoadP: 4087 case Op_LoadN: 4088 case Op_LoadRange: 4089 case Op_LoadD_unaligned: 4090 case Op_LoadL_unaligned: 4091 case Op_StrComp: // Does a bunch of load-like effects 4092 case Op_StrEquals: 4093 case Op_StrIndexOf: 4094 case Op_StrIndexOfChar: 4095 case Op_AryEq: 4096 case Op_HasNegatives: 4097 pinned = false; 4098 } 4099 if( pinned ) { 4100 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); 4101 if( !chosen_loop->_child ) // Inner loop? 4102 chosen_loop->_body.push(n); // Collect inner loops 4103 return; 4104 } 4105 } else { // No slot zero 4106 if( n->is_CFG() ) { // CFG with no slot 0 is dead 4107 _nodes.map(n->_idx,0); // No block setting, it's globally dead 4108 return; 4109 } 4110 assert(!n->is_CFG() || n->outcnt() == 0, ""); 4111 } 4112 4113 // Do I have a "safe range" I can select over? 4114 Node *early = get_ctrl(n);// Early location already computed 4115 4116 // Compute latest point this Node can go 4117 Node *LCA = get_late_ctrl( n, early ); 4118 // LCA is NULL due to uses being dead 4119 if( LCA == NULL ) { 4120 #ifdef ASSERT 4121 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { 4122 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); 4123 } 4124 #endif 4125 _nodes.map(n->_idx, 0); // This node is useless 4126 _deadlist.push(n); 4127 return; 4128 } 4129 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); 4130 4131 Node *legal = LCA; // Walk 'legal' up the IDOM chain 4132 Node *least = legal; // Best legal position so far 4133 while( early != legal ) { // While not at earliest legal 4134 #ifdef ASSERT 4135 if (legal->is_Start() && !early->is_Root()) { 4136 // Bad graph. Print idom path and fail. 4137 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA); 4138 assert(false, "Bad graph detected in build_loop_late"); 4139 } 4140 #endif 4141 // Find least loop nesting depth 4142 legal = idom(legal); // Bump up the IDOM tree 4143 // Check for lower nesting depth 4144 if( get_loop(legal)->_nest < get_loop(least)->_nest ) 4145 least = legal; 4146 } 4147 assert(early == legal || legal != C->root(), "bad dominance of inputs"); 4148 4149 // Try not to place code on a loop entry projection 4150 // which can inhibit range check elimination. 4151 if (least != early) { 4152 Node* ctrl_out = least->unique_ctrl_out(); 4153 if (ctrl_out && ctrl_out->is_Loop() && 4154 least == ctrl_out->in(LoopNode::EntryControl) && 4155 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) { 4156 Node* least_dom = idom(least); 4157 if (get_loop(least_dom)->is_member(get_loop(least))) { 4158 least = least_dom; 4159 } 4160 } 4161 } 4162 4163 #ifdef ASSERT 4164 // If verifying, verify that 'verify_me' has a legal location 4165 // and choose it as our location. 4166 if( _verify_me ) { 4167 Node *v_ctrl = _verify_me->get_ctrl_no_update(n); 4168 Node *legal = LCA; 4169 while( early != legal ) { // While not at earliest legal 4170 if( legal == v_ctrl ) break; // Check for prior good location 4171 legal = idom(legal) ;// Bump up the IDOM tree 4172 } 4173 // Check for prior good location 4174 if( legal == v_ctrl ) least = legal; // Keep prior if found 4175 } 4176 #endif 4177 4178 // Assign discovered "here or above" point 4179 least = find_non_split_ctrl(least); 4180 verify_strip_mined_scheduling(n, least); 4181 set_ctrl(n, least); 4182 4183 // Collect inner loop bodies 4184 IdealLoopTree *chosen_loop = get_loop(least); 4185 if( !chosen_loop->_child ) // Inner loop? 4186 chosen_loop->_body.push(n);// Collect inner loops 4187 } 4188 4189 #ifdef ASSERT 4190 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) { 4191 tty->print_cr("%s", msg); 4192 tty->print("n: "); n->dump(); 4193 tty->print("early(n): "); early->dump(); 4194 if (n->in(0) != NULL && !n->in(0)->is_top() && 4195 n->in(0) != early && !n->in(0)->is_Root()) { 4196 tty->print("n->in(0): "); n->in(0)->dump(); 4197 } 4198 for (uint i = 1; i < n->req(); i++) { 4199 Node* in1 = n->in(i); 4200 if (in1 != NULL && in1 != n && !in1->is_top()) { 4201 tty->print("n->in(%d): ", i); in1->dump(); 4202 Node* in1_early = get_ctrl(in1); 4203 tty->print("early(n->in(%d)): ", i); in1_early->dump(); 4204 if (in1->in(0) != NULL && !in1->in(0)->is_top() && 4205 in1->in(0) != in1_early && !in1->in(0)->is_Root()) { 4206 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump(); 4207 } 4208 for (uint j = 1; j < in1->req(); j++) { 4209 Node* in2 = in1->in(j); 4210 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) { 4211 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump(); 4212 Node* in2_early = get_ctrl(in2); 4213 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); 4214 if (in2->in(0) != NULL && !in2->in(0)->is_top() && 4215 in2->in(0) != in2_early && !in2->in(0)->is_Root()) { 4216 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump(); 4217 } 4218 } 4219 } 4220 } 4221 } 4222 tty->cr(); 4223 tty->print("LCA(n): "); LCA->dump(); 4224 for (uint i = 0; i < n->outcnt(); i++) { 4225 Node* u1 = n->raw_out(i); 4226 if (u1 == n) 4227 continue; 4228 tty->print("n->out(%d): ", i); u1->dump(); 4229 if (u1->is_CFG()) { 4230 for (uint j = 0; j < u1->outcnt(); j++) { 4231 Node* u2 = u1->raw_out(j); 4232 if (u2 != u1 && u2 != n && u2->is_CFG()) { 4233 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4234 } 4235 } 4236 } else { 4237 Node* u1_later = get_ctrl(u1); 4238 tty->print("later(n->out(%d)): ", i); u1_later->dump(); 4239 if (u1->in(0) != NULL && !u1->in(0)->is_top() && 4240 u1->in(0) != u1_later && !u1->in(0)->is_Root()) { 4241 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump(); 4242 } 4243 for (uint j = 0; j < u1->outcnt(); j++) { 4244 Node* u2 = u1->raw_out(j); 4245 if (u2 == n || u2 == u1) 4246 continue; 4247 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4248 if (!u2->is_CFG()) { 4249 Node* u2_later = get_ctrl(u2); 4250 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); 4251 if (u2->in(0) != NULL && !u2->in(0)->is_top() && 4252 u2->in(0) != u2_later && !u2->in(0)->is_Root()) { 4253 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump(); 4254 } 4255 } 4256 } 4257 } 4258 } 4259 tty->cr(); 4260 int ct = 0; 4261 Node *dbg_legal = LCA; 4262 while(!dbg_legal->is_Start() && ct < 100) { 4263 tty->print("idom[%d] ",ct); dbg_legal->dump(); 4264 ct++; 4265 dbg_legal = idom(dbg_legal); 4266 } 4267 tty->cr(); 4268 } 4269 #endif 4270 4271 #ifndef PRODUCT 4272 //------------------------------dump------------------------------------------- 4273 void PhaseIdealLoop::dump( ) const { 4274 ResourceMark rm; 4275 Arena* arena = Thread::current()->resource_area(); 4276 Node_Stack stack(arena, C->live_nodes() >> 2); 4277 Node_List rpo_list; 4278 VectorSet visited(arena); 4279 visited.set(C->top()->_idx); 4280 rpo( C->root(), stack, visited, rpo_list ); 4281 // Dump root loop indexed by last element in PO order 4282 dump( _ltree_root, rpo_list.size(), rpo_list ); 4283 } 4284 4285 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { 4286 loop->dump_head(); 4287 4288 // Now scan for CFG nodes in the same loop 4289 for( uint j=idx; j > 0; j-- ) { 4290 Node *n = rpo_list[j-1]; 4291 if( !_nodes[n->_idx] ) // Skip dead nodes 4292 continue; 4293 if( get_loop(n) != loop ) { // Wrong loop nest 4294 if( get_loop(n)->_head == n && // Found nested loop? 4295 get_loop(n)->_parent == loop ) 4296 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly 4297 continue; 4298 } 4299 4300 // Dump controlling node 4301 for( uint x = 0; x < loop->_nest; x++ ) 4302 tty->print(" "); 4303 tty->print("C"); 4304 if( n == C->root() ) { 4305 n->dump(); 4306 } else { 4307 Node* cached_idom = idom_no_update(n); 4308 Node *computed_idom = n->in(0); 4309 if( n->is_Region() ) { 4310 computed_idom = compute_idom(n); 4311 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or 4312 // any MultiBranch ctrl node), so apply a similar transform to 4313 // the cached idom returned from idom_no_update. 4314 cached_idom = find_non_split_ctrl(cached_idom); 4315 } 4316 tty->print(" ID:%d",computed_idom->_idx); 4317 n->dump(); 4318 if( cached_idom != computed_idom ) { 4319 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", 4320 computed_idom->_idx, cached_idom->_idx); 4321 } 4322 } 4323 // Dump nodes it controls 4324 for( uint k = 0; k < _nodes.Size(); k++ ) { 4325 // (k < C->unique() && get_ctrl(find(k)) == n) 4326 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { 4327 Node *m = C->root()->find(k); 4328 if( m && m->outcnt() > 0 ) { 4329 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { 4330 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", 4331 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); 4332 } 4333 for( uint j = 0; j < loop->_nest; j++ ) 4334 tty->print(" "); 4335 tty->print(" "); 4336 m->dump(); 4337 } 4338 } 4339 } 4340 } 4341 } 4342 4343 // Collect a R-P-O for the whole CFG. 4344 // Result list is in post-order (scan backwards for RPO) 4345 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { 4346 stk.push(start, 0); 4347 visited.set(start->_idx); 4348 4349 while (stk.is_nonempty()) { 4350 Node* m = stk.node(); 4351 uint idx = stk.index(); 4352 if (idx < m->outcnt()) { 4353 stk.set_index(idx + 1); 4354 Node* n = m->raw_out(idx); 4355 if (n->is_CFG() && !visited.test_set(n->_idx)) { 4356 stk.push(n, 0); 4357 } 4358 } else { 4359 rpo_list.push(m); 4360 stk.pop(); 4361 } 4362 } 4363 } 4364 #endif 4365 4366 4367 //============================================================================= 4368 //------------------------------LoopTreeIterator----------------------------------- 4369 4370 // Advance to next loop tree using a preorder, left-to-right traversal. 4371 void LoopTreeIterator::next() { 4372 assert(!done(), "must not be done."); 4373 if (_curnt->_child != NULL) { 4374 _curnt = _curnt->_child; 4375 } else if (_curnt->_next != NULL) { 4376 _curnt = _curnt->_next; 4377 } else { 4378 while (_curnt != _root && _curnt->_next == NULL) { 4379 _curnt = _curnt->_parent; 4380 } 4381 if (_curnt == _root) { 4382 _curnt = NULL; 4383 assert(done(), "must be done."); 4384 } else { 4385 assert(_curnt->_next != NULL, "must be more to do"); 4386 _curnt = _curnt->_next; 4387 } 4388 } 4389 }