1 /* 2 * Copyright (c) 1998, 2020, 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 "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "libadt/vectset.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/callnode.hpp" 36 #include "opto/compile.inline.hpp" 37 #include "opto/connode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/divnode.hpp" 40 #include "opto/idealGraphPrinter.hpp" 41 #include "opto/loopnode.hpp" 42 #include "opto/movenode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/rootnode.hpp" 45 #include "opto/superword.hpp" 46 #include "utilities/powerOfTwo.hpp" 47 48 //============================================================================= 49 //--------------------------is_cloop_ind_var----------------------------------- 50 // Determine if a node is a counted loop induction variable. 51 // NOTE: The method is declared in "node.hpp". 52 bool Node::is_cloop_ind_var() const { 53 return (is_Phi() && !as_Phi()->is_copy() && 54 as_Phi()->region()->is_CountedLoop() && 55 as_Phi()->region()->as_CountedLoop()->phi() == this); 56 } 57 58 //============================================================================= 59 //------------------------------dump_spec-------------------------------------- 60 // Dump special per-node info 61 #ifndef PRODUCT 62 void LoopNode::dump_spec(outputStream *st) const { 63 if (is_inner_loop()) st->print( "inner " ); 64 if (is_partial_peel_loop()) st->print( "partial_peel " ); 65 if (partial_peel_has_failed()) st->print( "partial_peel_failed " ); 66 } 67 #endif 68 69 //------------------------------is_valid_counted_loop------------------------- 70 bool LoopNode::is_valid_counted_loop() const { 71 if (is_CountedLoop()) { 72 CountedLoopNode* l = as_CountedLoop(); 73 CountedLoopEndNode* le = l->loopexit_or_null(); 74 if (le != NULL && 75 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) { 76 Node* phi = l->phi(); 77 Node* exit = le->proj_out_or_null(0 /* false */); 78 if (exit != NULL && exit->Opcode() == Op_IfFalse && 79 phi != NULL && phi->is_Phi() && 80 phi->in(LoopNode::LoopBackControl) == l->incr() && 81 le->loopnode() == l && le->stride_is_con()) { 82 return true; 83 } 84 } 85 } 86 return false; 87 } 88 89 //------------------------------get_early_ctrl--------------------------------- 90 // Compute earliest legal control 91 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { 92 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); 93 uint i; 94 Node *early; 95 if (n->in(0) && !n->is_expensive()) { 96 early = n->in(0); 97 if (!early->is_CFG()) // Might be a non-CFG multi-def 98 early = get_ctrl(early); // So treat input as a straight data input 99 i = 1; 100 } else { 101 early = get_ctrl(n->in(1)); 102 i = 2; 103 } 104 uint e_d = dom_depth(early); 105 assert( early, "" ); 106 for (; i < n->req(); i++) { 107 Node *cin = get_ctrl(n->in(i)); 108 assert( cin, "" ); 109 // Keep deepest dominator depth 110 uint c_d = dom_depth(cin); 111 if (c_d > e_d) { // Deeper guy? 112 early = cin; // Keep deepest found so far 113 e_d = c_d; 114 } else if (c_d == e_d && // Same depth? 115 early != cin) { // If not equal, must use slower algorithm 116 // If same depth but not equal, one _must_ dominate the other 117 // and we want the deeper (i.e., dominated) guy. 118 Node *n1 = early; 119 Node *n2 = cin; 120 while (1) { 121 n1 = idom(n1); // Walk up until break cycle 122 n2 = idom(n2); 123 if (n1 == cin || // Walked early up to cin 124 dom_depth(n2) < c_d) 125 break; // early is deeper; keep him 126 if (n2 == early || // Walked cin up to early 127 dom_depth(n1) < c_d) { 128 early = cin; // cin is deeper; keep him 129 break; 130 } 131 } 132 e_d = dom_depth(early); // Reset depth register cache 133 } 134 } 135 136 // Return earliest legal location 137 assert(early == find_non_split_ctrl(early), "unexpected early control"); 138 139 if (n->is_expensive() && !_verify_only && !_verify_me) { 140 assert(n->in(0), "should have control input"); 141 early = get_early_ctrl_for_expensive(n, early); 142 } 143 144 return early; 145 } 146 147 //------------------------------get_early_ctrl_for_expensive--------------------------------- 148 // Move node up the dominator tree as high as legal while still beneficial 149 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) { 150 assert(n->in(0) && n->is_expensive(), "expensive node with control input here"); 151 assert(OptimizeExpensiveOps, "optimization off?"); 152 153 Node* ctl = n->in(0); 154 assert(ctl->is_CFG(), "expensive input 0 must be cfg"); 155 uint min_dom_depth = dom_depth(earliest); 156 #ifdef ASSERT 157 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) { 158 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl); 159 assert(false, "Bad graph detected in get_early_ctrl_for_expensive"); 160 } 161 #endif 162 if (dom_depth(ctl) < min_dom_depth) { 163 return earliest; 164 } 165 166 while (1) { 167 Node *next = ctl; 168 // Moving the node out of a loop on the projection of a If 169 // confuses loop predication. So once we hit a Loop in a If branch 170 // that doesn't branch to an UNC, we stop. The code that process 171 // expensive nodes will notice the loop and skip over it to try to 172 // move the node further up. 173 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) { 174 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 175 break; 176 } 177 next = idom(ctl->in(1)->in(0)); 178 } else if (ctl->is_Proj()) { 179 // We only move it up along a projection if the projection is 180 // the single control projection for its parent: same code path, 181 // if it's a If with UNC or fallthrough of a call. 182 Node* parent_ctl = ctl->in(0); 183 if (parent_ctl == NULL) { 184 break; 185 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) { 186 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control(); 187 } else if (parent_ctl->is_If()) { 188 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 189 break; 190 } 191 assert(idom(ctl) == parent_ctl, "strange"); 192 next = idom(parent_ctl); 193 } else if (ctl->is_CatchProj()) { 194 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) { 195 break; 196 } 197 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph"); 198 next = parent_ctl->in(0)->in(0)->in(0); 199 } else { 200 // Check if parent control has a single projection (this 201 // control is the only possible successor of the parent 202 // control). If so, we can try to move the node above the 203 // parent control. 204 int nb_ctl_proj = 0; 205 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) { 206 Node *p = parent_ctl->fast_out(i); 207 if (p->is_Proj() && p->is_CFG()) { 208 nb_ctl_proj++; 209 if (nb_ctl_proj > 1) { 210 break; 211 } 212 } 213 } 214 215 if (nb_ctl_proj > 1) { 216 break; 217 } 218 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() || 219 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node"); 220 assert(idom(ctl) == parent_ctl, "strange"); 221 next = idom(parent_ctl); 222 } 223 } else { 224 next = idom(ctl); 225 } 226 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) { 227 break; 228 } 229 ctl = next; 230 } 231 232 if (ctl != n->in(0)) { 233 _igvn.replace_input_of(n, 0, ctl); 234 _igvn.hash_insert(n); 235 } 236 237 return ctl; 238 } 239 240 241 //------------------------------set_early_ctrl--------------------------------- 242 // Set earliest legal control 243 void PhaseIdealLoop::set_early_ctrl( Node *n ) { 244 Node *early = get_early_ctrl(n); 245 246 // Record earliest legal location 247 set_ctrl(n, early); 248 } 249 250 //------------------------------set_subtree_ctrl------------------------------- 251 // set missing _ctrl entries on new nodes 252 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { 253 // Already set? Get out. 254 if( _nodes[n->_idx] ) return; 255 // Recursively set _nodes array to indicate where the Node goes 256 uint i; 257 for( i = 0; i < n->req(); ++i ) { 258 Node *m = n->in(i); 259 if( m && m != C->root() ) 260 set_subtree_ctrl( m ); 261 } 262 263 // Fixup self 264 set_early_ctrl( n ); 265 } 266 267 // Create a skeleton strip mined outer loop: a Loop head before the 268 // inner strip mined loop, a safepoint and an exit condition guarded 269 // by an opaque node after the inner strip mined loop with a backedge 270 // to the loop head. The inner strip mined loop is left as it is. Only 271 // once loop optimizations are over, do we adjust the inner loop exit 272 // condition to limit its number of iterations, set the outer loop 273 // exit condition and add Phis to the outer loop head. Some loop 274 // optimizations that operate on the inner strip mined loop need to be 275 // aware of the outer strip mined loop: loop unswitching needs to 276 // clone the outer loop as well as the inner, unrolling needs to only 277 // clone the inner loop etc. No optimizations need to change the outer 278 // strip mined loop as it is only a skeleton. 279 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control, 280 IdealLoopTree* loop, float cl_prob, float le_fcnt, 281 Node*& entry_control, Node*& iffalse) { 282 Node* outer_test = _igvn.intcon(0); 283 set_ctrl(outer_test, C->root()); 284 Node *orig = iffalse; 285 iffalse = iffalse->clone(); 286 _igvn.register_new_node_with_optimizer(iffalse); 287 set_idom(iffalse, idom(orig), dom_depth(orig)); 288 289 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt); 290 Node *outer_ift = new IfTrueNode (outer_le); 291 Node* outer_iff = orig; 292 _igvn.replace_input_of(outer_iff, 0, outer_le); 293 294 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift); 295 entry_control = outer_l; 296 297 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift); 298 IdealLoopTree* parent = loop->_parent; 299 IdealLoopTree* sibling = parent->_child; 300 if (sibling == loop) { 301 parent->_child = outer_ilt; 302 } else { 303 while (sibling->_next != loop) { 304 sibling = sibling->_next; 305 } 306 sibling->_next = outer_ilt; 307 } 308 outer_ilt->_next = loop->_next; 309 outer_ilt->_parent = parent; 310 outer_ilt->_child = loop; 311 outer_ilt->_nest = loop->_nest; 312 loop->_parent = outer_ilt; 313 loop->_next = NULL; 314 loop->_nest++; 315 316 set_loop(iffalse, outer_ilt); 317 register_control(outer_le, outer_ilt, iffalse); 318 register_control(outer_ift, outer_ilt, outer_le); 319 set_idom(outer_iff, outer_le, dom_depth(outer_le)); 320 _igvn.register_new_node_with_optimizer(outer_l); 321 set_loop(outer_l, outer_ilt); 322 set_idom(outer_l, init_control, dom_depth(init_control)+1); 323 324 return outer_ilt; 325 } 326 327 void PhaseIdealLoop::insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol) { 328 Node* new_predicate_proj = create_new_if_for_predicate(limit_check_proj, NULL, 329 Deoptimization::Reason_loop_limit_check, 330 Op_If); 331 Node* iff = new_predicate_proj->in(0); 332 assert(iff->Opcode() == Op_If, "bad graph shape"); 333 Node* conv = iff->in(1); 334 assert(conv->Opcode() == Op_Conv2B, "bad graph shape"); 335 Node* opaq = conv->in(1); 336 assert(opaq->Opcode() == Op_Opaque1, "bad graph shape"); 337 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit); 338 bol = _igvn.register_new_node_with_optimizer(bol); 339 set_subtree_ctrl(bol); 340 _igvn.replace_input_of(iff, 1, bol); 341 342 #ifndef PRODUCT 343 // report that the loop predication has been actually performed 344 // for this loop 345 if (TraceLoopLimitCheck) { 346 tty->print_cr("Counted Loop Limit Check generated:"); 347 debug_only( bol->dump(2); ) 348 } 349 #endif 350 } 351 352 //------------------------------is_counted_loop-------------------------------- 353 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) { 354 PhaseGVN *gvn = &_igvn; 355 356 // Counted loop head must be a good RegionNode with only 3 not NULL 357 // control input edges: Self, Entry, LoopBack. 358 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) { 359 return false; 360 } 361 Node *init_control = x->in(LoopNode::EntryControl); 362 Node *back_control = x->in(LoopNode::LoopBackControl); 363 if (init_control == NULL || back_control == NULL) // Partially dead 364 return false; 365 // Must also check for TOP when looking for a dead loop 366 if (init_control->is_top() || back_control->is_top()) 367 return false; 368 369 // Allow funny placement of Safepoint 370 if (back_control->Opcode() == Op_SafePoint) { 371 if (LoopStripMiningIter != 0) { 372 // Leaving the safepoint on the backedge and creating a 373 // CountedLoop will confuse optimizations. We can't move the 374 // safepoint around because its jvm state wouldn't match a new 375 // location. Give up on that loop. 376 return false; 377 } 378 back_control = back_control->in(TypeFunc::Control); 379 } 380 381 // Controlling test for loop 382 Node *iftrue = back_control; 383 uint iftrue_op = iftrue->Opcode(); 384 if (iftrue_op != Op_IfTrue && 385 iftrue_op != Op_IfFalse) 386 // I have a weird back-control. Probably the loop-exit test is in 387 // the middle of the loop and I am looking at some trailing control-flow 388 // merge point. To fix this I would have to partially peel the loop. 389 return false; // Obscure back-control 390 391 // Get boolean guarding loop-back test 392 Node *iff = iftrue->in(0); 393 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) 394 return false; 395 BoolNode *test = iff->in(1)->as_Bool(); 396 BoolTest::mask bt = test->_test._test; 397 float cl_prob = iff->as_If()->_prob; 398 if (iftrue_op == Op_IfFalse) { 399 bt = BoolTest(bt).negate(); 400 cl_prob = 1.0 - cl_prob; 401 } 402 // Get backedge compare 403 Node *cmp = test->in(1); 404 int cmp_op = cmp->Opcode(); 405 if (cmp_op != Op_CmpI) 406 return false; // Avoid pointer & float compares 407 408 // Find the trip-counter increment & limit. Limit must be loop invariant. 409 Node *incr = cmp->in(1); 410 Node *limit = cmp->in(2); 411 412 // --------- 413 // need 'loop()' test to tell if limit is loop invariant 414 // --------- 415 416 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit? 417 Node *tmp = incr; // Then reverse order into the CmpI 418 incr = limit; 419 limit = tmp; 420 bt = BoolTest(bt).commute(); // And commute the exit test 421 } 422 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant 423 return false; 424 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 425 return false; 426 427 Node* phi_incr = NULL; 428 // Trip-counter increment must be commutative & associative. 429 if (incr->Opcode() == Op_CastII) { 430 incr = incr->in(1); 431 } 432 if (incr->is_Phi()) { 433 if (incr->as_Phi()->region() != x || incr->req() != 3) 434 return false; // Not simple trip counter expression 435 phi_incr = incr; 436 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi 437 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 438 return false; 439 } 440 441 Node* trunc1 = NULL; 442 Node* trunc2 = NULL; 443 const TypeInt* iv_trunc_t = NULL; 444 Node* orig_incr = incr; 445 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { 446 return false; // Funny increment opcode 447 } 448 assert(incr->Opcode() == Op_AddI, "wrong increment code"); 449 450 const TypeInt* limit_t = gvn->type(limit)->is_int(); 451 if (trunc1 != NULL) { 452 // When there is a truncation, we must be sure that after the truncation 453 // the trip counter will end up higher than the limit, otherwise we are looking 454 // at an endless loop. Can happen with range checks. 455 456 // Example: 457 // int i = 0; 458 // while (true) 459 // sum + = array[i]; 460 // i++; 461 // i = i && 0x7fff; 462 // } 463 // 464 // If the array is shorter than 0x8000 this exits through a AIOOB 465 // - Counted loop transformation is ok 466 // If the array is longer then this is an endless loop 467 // - No transformation can be done. 468 469 const TypeInt* incr_t = gvn->type(orig_incr)->is_int(); 470 if (limit_t->_hi > incr_t->_hi) { 471 // if the limit can have a higher value than the increment (before the phi) 472 return false; 473 } 474 } 475 476 // Get merge point 477 Node *xphi = incr->in(1); 478 Node *stride = incr->in(2); 479 if (!stride->is_Con()) { // Oops, swap these 480 if (!xphi->is_Con()) // Is the other guy a constant? 481 return false; // Nope, unknown stride, bail out 482 Node *tmp = xphi; // 'incr' is commutative, so ok to swap 483 xphi = stride; 484 stride = tmp; 485 } 486 if (xphi->Opcode() == Op_CastII) { 487 xphi = xphi->in(1); 488 } 489 // Stride must be constant 490 int stride_con = stride->get_int(); 491 if (stride_con == 0) 492 return false; // missed some peephole opt 493 494 if (!xphi->is_Phi()) 495 return false; // Too much math on the trip counter 496 if (phi_incr != NULL && phi_incr != xphi) 497 return false; 498 PhiNode *phi = xphi->as_Phi(); 499 500 // Phi must be of loop header; backedge must wrap to increment 501 if (phi->region() != x) 502 return false; 503 if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) || 504 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) { 505 return false; 506 } 507 Node *init_trip = phi->in(LoopNode::EntryControl); 508 509 // If iv trunc type is smaller than int, check for possible wrap. 510 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { 511 assert(trunc1 != NULL, "must have found some truncation"); 512 513 // Get a better type for the phi (filtered thru if's) 514 const TypeInt* phi_ft = filtered_type(phi); 515 516 // Can iv take on a value that will wrap? 517 // 518 // Ensure iv's limit is not within "stride" of the wrap value. 519 // 520 // Example for "short" type 521 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) 522 // If the stride is +10, then the last value of the induction 523 // variable before the increment (phi_ft->_hi) must be 524 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to 525 // ensure no truncation occurs after the increment. 526 527 if (stride_con > 0) { 528 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || 529 iv_trunc_t->_lo > phi_ft->_lo) { 530 return false; // truncation may occur 531 } 532 } else if (stride_con < 0) { 533 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || 534 iv_trunc_t->_hi < phi_ft->_hi) { 535 return false; // truncation may occur 536 } 537 } 538 // No possibility of wrap so truncation can be discarded 539 // Promote iv type to Int 540 } else { 541 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); 542 } 543 544 // If the condition is inverted and we will be rolling 545 // through MININT to MAXINT, then bail out. 546 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice! 547 // Odd stride 548 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) || 549 // Count down loop rolls through MAXINT 550 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) || 551 // Count up loop rolls through MININT 552 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) { 553 return false; // Bail out 554 } 555 556 const TypeInt* init_t = gvn->type(init_trip)->is_int(); 557 558 if (stride_con > 0) { 559 jlong init_p = (jlong)init_t->_lo + stride_con; 560 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi) 561 return false; // cyclic loop or this loop trips only once 562 } else { 563 jlong init_p = (jlong)init_t->_hi + stride_con; 564 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo) 565 return false; // cyclic loop or this loop trips only once 566 } 567 568 if (phi_incr != NULL && bt != BoolTest::ne) { 569 // check if there is a possiblity of IV overflowing after the first increment 570 if (stride_con > 0) { 571 if (init_t->_hi > max_jint - stride_con) { 572 return false; 573 } 574 } else { 575 if (init_t->_lo < min_jint - stride_con) { 576 return false; 577 } 578 } 579 } 580 581 // ================================================= 582 // ---- SUCCESS! Found A Trip-Counted Loop! ----- 583 // 584 assert(x->Opcode() == Op_Loop, "regular loops only"); 585 C->print_method(PHASE_BEFORE_CLOOPS, 3); 586 587 Node *hook = new Node(6); 588 589 // =================================================== 590 // Generate loop limit check to avoid integer overflow 591 // in cases like next (cyclic loops): 592 // 593 // for (i=0; i <= max_jint; i++) {} 594 // for (i=0; i < max_jint; i+=2) {} 595 // 596 // 597 // Limit check predicate depends on the loop test: 598 // 599 // for(;i != limit; i++) --> limit <= (max_jint) 600 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1) 601 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride ) 602 // 603 604 // Check if limit is excluded to do more precise int overflow check. 605 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge); 606 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1)); 607 608 // If compare points directly to the phi we need to adjust 609 // the compare so that it points to the incr. Limit have 610 // to be adjusted to keep trip count the same and the 611 // adjusted limit should be checked for int overflow. 612 if (phi_incr != NULL) { 613 stride_m += stride_con; 614 } 615 616 if (limit->is_Con()) { 617 int limit_con = limit->get_int(); 618 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) || 619 (stride_con < 0 && limit_con < (min_jint - stride_m))) { 620 // Bailout: it could be integer overflow. 621 return false; 622 } 623 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) || 624 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) { 625 // Limit's type may satisfy the condition, for example, 626 // when it is an array length. 627 } else { 628 // Generate loop's limit check. 629 // Loop limit check predicate should be near the loop. 630 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); 631 if (!limit_check_proj) { 632 // The limit check predicate is not generated if this method trapped here before. 633 #ifdef ASSERT 634 if (TraceLoopLimitCheck) { 635 tty->print("missing loop limit check:"); 636 loop->dump_head(); 637 x->dump(1); 638 } 639 #endif 640 return false; 641 } 642 643 IfNode* check_iff = limit_check_proj->in(0)->as_If(); 644 645 if (!is_dominator(get_ctrl(limit), check_iff->in(0))) { 646 return false; 647 } 648 649 Node* cmp_limit; 650 Node* bol; 651 652 if (stride_con > 0) { 653 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m)); 654 bol = new BoolNode(cmp_limit, BoolTest::le); 655 } else { 656 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m)); 657 bol = new BoolNode(cmp_limit, BoolTest::ge); 658 } 659 660 insert_loop_limit_check(limit_check_proj, cmp_limit, bol); 661 } 662 663 // Now we need to canonicalize loop condition. 664 if (bt == BoolTest::ne) { 665 assert(stride_con == 1 || stride_con == -1, "simple increment only"); 666 if (stride_con > 0 && init_t->_hi < limit_t->_lo) { 667 // 'ne' can be replaced with 'lt' only when init < limit. 668 bt = BoolTest::lt; 669 } else if (stride_con < 0 && init_t->_lo > limit_t->_hi) { 670 // 'ne' can be replaced with 'gt' only when init > limit. 671 bt = BoolTest::gt; 672 } else { 673 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); 674 if (!limit_check_proj) { 675 // The limit check predicate is not generated if this method trapped here before. 676 #ifdef ASSERT 677 if (TraceLoopLimitCheck) { 678 tty->print("missing loop limit check:"); 679 loop->dump_head(); 680 x->dump(1); 681 } 682 #endif 683 return false; 684 } 685 IfNode* check_iff = limit_check_proj->in(0)->as_If(); 686 687 if (!is_dominator(get_ctrl(limit), check_iff->in(0)) || 688 !is_dominator(get_ctrl(init_trip), check_iff->in(0))) { 689 return false; 690 } 691 692 Node* cmp_limit; 693 Node* bol; 694 695 if (stride_con > 0) { 696 cmp_limit = new CmpINode(init_trip, limit); 697 bol = new BoolNode(cmp_limit, BoolTest::lt); 698 } else { 699 cmp_limit = new CmpINode(init_trip, limit); 700 bol = new BoolNode(cmp_limit, BoolTest::gt); 701 } 702 703 insert_loop_limit_check(limit_check_proj, cmp_limit, bol); 704 705 if (stride_con > 0) { 706 // 'ne' can be replaced with 'lt' only when init < limit. 707 bt = BoolTest::lt; 708 } else if (stride_con < 0) { 709 // 'ne' can be replaced with 'gt' only when init > limit. 710 bt = BoolTest::gt; 711 } 712 } 713 } 714 715 if (phi_incr != NULL) { 716 // If compare points directly to the phi we need to adjust 717 // the compare so that it points to the incr. Limit have 718 // to be adjusted to keep trip count the same and we 719 // should avoid int overflow. 720 // 721 // i = init; do {} while(i++ < limit); 722 // is converted to 723 // i = init; do {} while(++i < limit+1); 724 // 725 limit = gvn->transform(new AddINode(limit, stride)); 726 } 727 728 if (incl_limit) { 729 // The limit check guaranties that 'limit <= (max_jint - stride)' so 730 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0. 731 // 732 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1); 733 limit = gvn->transform(new AddINode(limit, one)); 734 if (bt == BoolTest::le) 735 bt = BoolTest::lt; 736 else if (bt == BoolTest::ge) 737 bt = BoolTest::gt; 738 else 739 ShouldNotReachHere(); 740 } 741 set_subtree_ctrl( limit ); 742 743 if (LoopStripMiningIter == 0) { 744 // Check for SafePoint on backedge and remove 745 Node *sfpt = x->in(LoopNode::LoopBackControl); 746 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { 747 lazy_replace( sfpt, iftrue ); 748 if (loop->_safepts != NULL) { 749 loop->_safepts->yank(sfpt); 750 } 751 loop->_tail = iftrue; 752 } 753 } 754 755 // Build a canonical trip test. 756 // Clone code, as old values may be in use. 757 incr = incr->clone(); 758 incr->set_req(1,phi); 759 incr->set_req(2,stride); 760 incr = _igvn.register_new_node_with_optimizer(incr); 761 set_early_ctrl( incr ); 762 _igvn.rehash_node_delayed(phi); 763 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); 764 765 // If phi type is more restrictive than Int, raise to 766 // Int to prevent (almost) infinite recursion in igvn 767 // which can only handle integer types for constants or minint..maxint. 768 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { 769 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); 770 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); 771 nphi = _igvn.register_new_node_with_optimizer(nphi); 772 set_ctrl(nphi, get_ctrl(phi)); 773 _igvn.replace_node(phi, nphi); 774 phi = nphi->as_Phi(); 775 } 776 cmp = cmp->clone(); 777 cmp->set_req(1,incr); 778 cmp->set_req(2,limit); 779 cmp = _igvn.register_new_node_with_optimizer(cmp); 780 set_ctrl(cmp, iff->in(0)); 781 782 test = test->clone()->as_Bool(); 783 (*(BoolTest*)&test->_test)._test = bt; 784 test->set_req(1,cmp); 785 _igvn.register_new_node_with_optimizer(test); 786 set_ctrl(test, iff->in(0)); 787 788 // Replace the old IfNode with a new LoopEndNode 789 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt )); 790 IfNode *le = lex->as_If(); 791 uint dd = dom_depth(iff); 792 set_idom(le, le->in(0), dd); // Update dominance for loop exit 793 set_loop(le, loop); 794 795 // Get the loop-exit control 796 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); 797 798 // Need to swap loop-exit and loop-back control? 799 if (iftrue_op == Op_IfFalse) { 800 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le)); 801 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le)); 802 803 loop->_tail = back_control = ift2; 804 set_loop(ift2, loop); 805 set_loop(iff2, get_loop(iffalse)); 806 807 // Lazy update of 'get_ctrl' mechanism. 808 lazy_replace(iffalse, iff2); 809 lazy_replace(iftrue, ift2); 810 811 // Swap names 812 iffalse = iff2; 813 iftrue = ift2; 814 } else { 815 _igvn.rehash_node_delayed(iffalse); 816 _igvn.rehash_node_delayed(iftrue); 817 iffalse->set_req_X( 0, le, &_igvn ); 818 iftrue ->set_req_X( 0, le, &_igvn ); 819 } 820 821 set_idom(iftrue, le, dd+1); 822 set_idom(iffalse, le, dd+1); 823 assert(iff->outcnt() == 0, "should be dead now"); 824 lazy_replace( iff, le ); // fix 'get_ctrl' 825 826 Node *sfpt2 = le->in(0); 827 828 Node* entry_control = init_control; 829 bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL && 830 sfpt2->Opcode() == Op_SafePoint && !loop->_has_call; 831 IdealLoopTree* outer_ilt = NULL; 832 if (strip_mine_loop) { 833 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop, 834 cl_prob, le->_fcnt, entry_control, 835 iffalse); 836 } 837 838 // Now setup a new CountedLoopNode to replace the existing LoopNode 839 CountedLoopNode *l = new CountedLoopNode(entry_control, back_control); 840 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve 841 // The following assert is approximately true, and defines the intention 842 // of can_be_counted_loop. It fails, however, because phase->type 843 // is not yet initialized for this loop and its parts. 844 //assert(l->can_be_counted_loop(this), "sanity"); 845 _igvn.register_new_node_with_optimizer(l); 846 set_loop(l, loop); 847 loop->_head = l; 848 // Fix all data nodes placed at the old loop head. 849 // Uses the lazy-update mechanism of 'get_ctrl'. 850 lazy_replace( x, l ); 851 set_idom(l, entry_control, dom_depth(entry_control) + 1); 852 853 if (LoopStripMiningIter == 0 || strip_mine_loop) { 854 // Check for immediately preceding SafePoint and remove 855 if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) { 856 if (strip_mine_loop) { 857 Node* outer_le = outer_ilt->_tail->in(0); 858 Node* sfpt = sfpt2->clone(); 859 sfpt->set_req(0, iffalse); 860 outer_le->set_req(0, sfpt); 861 register_control(sfpt, outer_ilt, iffalse); 862 set_idom(outer_le, sfpt, dom_depth(sfpt)); 863 } 864 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); 865 if (loop->_safepts != NULL) { 866 loop->_safepts->yank(sfpt2); 867 } 868 } 869 } 870 871 // Free up intermediate goo 872 _igvn.remove_dead_node(hook); 873 874 #ifdef ASSERT 875 assert(l->is_valid_counted_loop(), "counted loop shape is messed up"); 876 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" ); 877 #endif 878 #ifndef PRODUCT 879 if (TraceLoopOpts) { 880 tty->print("Counted "); 881 loop->dump_head(); 882 } 883 #endif 884 885 C->print_method(PHASE_AFTER_CLOOPS, 3); 886 887 // Capture bounds of the loop in the induction variable Phi before 888 // subsequent transformation (iteration splitting) obscures the 889 // bounds 890 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn)); 891 892 if (strip_mine_loop) { 893 l->mark_strip_mined(); 894 l->verify_strip_mined(1); 895 outer_ilt->_head->as_Loop()->verify_strip_mined(1); 896 loop = outer_ilt; 897 } 898 899 return true; 900 } 901 902 //----------------------exact_limit------------------------------------------- 903 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) { 904 assert(loop->_head->is_CountedLoop(), ""); 905 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 906 assert(cl->is_valid_counted_loop(), ""); 907 908 if (ABS(cl->stride_con()) == 1 || 909 cl->limit()->Opcode() == Op_LoopLimit) { 910 // Old code has exact limit (it could be incorrect in case of int overflow). 911 // Loop limit is exact with stride == 1. And loop may already have exact limit. 912 return cl->limit(); 913 } 914 Node *limit = NULL; 915 #ifdef ASSERT 916 BoolTest::mask bt = cl->loopexit()->test_trip(); 917 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 918 #endif 919 if (cl->has_exact_trip_count()) { 920 // Simple case: loop has constant boundaries. 921 // Use jlongs to avoid integer overflow. 922 int stride_con = cl->stride_con(); 923 jlong init_con = cl->init_trip()->get_int(); 924 jlong limit_con = cl->limit()->get_int(); 925 julong trip_cnt = cl->trip_count(); 926 jlong final_con = init_con + trip_cnt*stride_con; 927 int final_int = (int)final_con; 928 // The final value should be in integer range since the loop 929 // is counted and the limit was checked for overflow. 930 assert(final_con == (jlong)final_int, "final value should be integer"); 931 limit = _igvn.intcon(final_int); 932 } else { 933 // Create new LoopLimit node to get exact limit (final iv value). 934 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride()); 935 register_new_node(limit, cl->in(LoopNode::EntryControl)); 936 } 937 assert(limit != NULL, "sanity"); 938 return limit; 939 } 940 941 //------------------------------Ideal------------------------------------------ 942 // Return a node which is more "ideal" than the current node. 943 // Attempt to convert into a counted-loop. 944 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 945 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) { 946 phase->C->set_major_progress(); 947 } 948 return RegionNode::Ideal(phase, can_reshape); 949 } 950 951 #ifdef ASSERT 952 void LoopNode::verify_strip_mined(int expect_skeleton) const { 953 const OuterStripMinedLoopNode* outer = NULL; 954 const CountedLoopNode* inner = NULL; 955 if (is_strip_mined()) { 956 if (!is_valid_counted_loop()) { 957 return; // Skip malformed counted loop 958 } 959 assert(is_CountedLoop(), "no Loop should be marked strip mined"); 960 inner = as_CountedLoop(); 961 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop(); 962 } else if (is_OuterStripMinedLoop()) { 963 outer = this->as_OuterStripMinedLoop(); 964 inner = outer->unique_ctrl_out()->as_CountedLoop(); 965 assert(inner->is_valid_counted_loop() && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed"); 966 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined"); 967 } 968 if (inner != NULL || outer != NULL) { 969 assert(inner != NULL && outer != NULL, "missing loop in strip mined nest"); 970 Node* outer_tail = outer->in(LoopNode::LoopBackControl); 971 Node* outer_le = outer_tail->in(0); 972 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If"); 973 Node* sfpt = outer_le->in(0); 974 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?"); 975 Node* inner_out = sfpt->in(0); 976 if (inner_out->outcnt() != 1) { 977 ResourceMark rm; 978 Unique_Node_List wq; 979 980 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) { 981 Node* u = inner_out->fast_out(i); 982 if (u == sfpt) { 983 continue; 984 } 985 wq.clear(); 986 wq.push(u); 987 bool found_sfpt = false; 988 for (uint next = 0; next < wq.size() && !found_sfpt; next++) { 989 Node* n = wq.at(next); 990 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) { 991 Node* u = n->fast_out(i); 992 if (u == sfpt) { 993 found_sfpt = true; 994 } 995 if (!u->is_CFG()) { 996 wq.push(u); 997 } 998 } 999 } 1000 assert(found_sfpt, "no node in loop that's not input to safepoint"); 1001 } 1002 } 1003 1004 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd(); 1005 assert(cle == inner->loopexit_or_null(), "mismatch"); 1006 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0; 1007 if (has_skeleton) { 1008 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node"); 1009 assert(outer->outcnt() == 2, "only phis"); 1010 } else { 1011 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?"); 1012 uint phis = 0; 1013 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) { 1014 Node* u = inner->fast_out(i); 1015 if (u->is_Phi()) { 1016 phis++; 1017 } 1018 } 1019 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) { 1020 Node* u = outer->fast_out(i); 1021 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop"); 1022 } 1023 uint stores = 0; 1024 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) { 1025 Node* u = inner_out->fast_out(i); 1026 if (u->is_Store()) { 1027 stores++; 1028 } 1029 } 1030 assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis"); 1031 } 1032 assert(sfpt->outcnt() == 1, "no data node"); 1033 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node"); 1034 } 1035 } 1036 #endif 1037 1038 //============================================================================= 1039 //------------------------------Ideal------------------------------------------ 1040 // Return a node which is more "ideal" than the current node. 1041 // Attempt to convert into a counted-loop. 1042 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1043 return RegionNode::Ideal(phase, can_reshape); 1044 } 1045 1046 //------------------------------dump_spec-------------------------------------- 1047 // Dump special per-node info 1048 #ifndef PRODUCT 1049 void CountedLoopNode::dump_spec(outputStream *st) const { 1050 LoopNode::dump_spec(st); 1051 if (stride_is_con()) { 1052 st->print("stride: %d ",stride_con()); 1053 } 1054 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx); 1055 if (is_main_loop()) st->print("main of N%d", _idx); 1056 if (is_post_loop()) st->print("post of N%d", _main_idx); 1057 if (is_strip_mined()) st->print(" strip mined"); 1058 } 1059 #endif 1060 1061 //============================================================================= 1062 int CountedLoopEndNode::stride_con() const { 1063 return stride()->bottom_type()->is_int()->get_con(); 1064 } 1065 1066 //============================================================================= 1067 //------------------------------Value----------------------------------------- 1068 const Type* LoopLimitNode::Value(PhaseGVN* phase) const { 1069 const Type* init_t = phase->type(in(Init)); 1070 const Type* limit_t = phase->type(in(Limit)); 1071 const Type* stride_t = phase->type(in(Stride)); 1072 // Either input is TOP ==> the result is TOP 1073 if (init_t == Type::TOP) return Type::TOP; 1074 if (limit_t == Type::TOP) return Type::TOP; 1075 if (stride_t == Type::TOP) return Type::TOP; 1076 1077 int stride_con = stride_t->is_int()->get_con(); 1078 if (stride_con == 1) 1079 return NULL; // Identity 1080 1081 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) { 1082 // Use jlongs to avoid integer overflow. 1083 jlong init_con = init_t->is_int()->get_con(); 1084 jlong limit_con = limit_t->is_int()->get_con(); 1085 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 1086 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 1087 jlong final_con = init_con + stride_con*trip_count; 1088 int final_int = (int)final_con; 1089 // The final value should be in integer range since the loop 1090 // is counted and the limit was checked for overflow. 1091 assert(final_con == (jlong)final_int, "final value should be integer"); 1092 return TypeInt::make(final_int); 1093 } 1094 1095 return bottom_type(); // TypeInt::INT 1096 } 1097 1098 //------------------------------Ideal------------------------------------------ 1099 // Return a node which is more "ideal" than the current node. 1100 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1101 if (phase->type(in(Init)) == Type::TOP || 1102 phase->type(in(Limit)) == Type::TOP || 1103 phase->type(in(Stride)) == Type::TOP) 1104 return NULL; // Dead 1105 1106 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1107 if (stride_con == 1) 1108 return NULL; // Identity 1109 1110 if (in(Init)->is_Con() && in(Limit)->is_Con()) 1111 return NULL; // Value 1112 1113 // Delay following optimizations until all loop optimizations 1114 // done to keep Ideal graph simple. 1115 if (!can_reshape || phase->C->major_progress()) 1116 return NULL; 1117 1118 const TypeInt* init_t = phase->type(in(Init) )->is_int(); 1119 const TypeInt* limit_t = phase->type(in(Limit))->is_int(); 1120 int stride_p; 1121 jlong lim, ini; 1122 julong max; 1123 if (stride_con > 0) { 1124 stride_p = stride_con; 1125 lim = limit_t->_hi; 1126 ini = init_t->_lo; 1127 max = (julong)max_jint; 1128 } else { 1129 stride_p = -stride_con; 1130 lim = init_t->_hi; 1131 ini = limit_t->_lo; 1132 max = (julong)min_jint; 1133 } 1134 julong range = lim - ini + stride_p; 1135 if (range <= max) { 1136 // Convert to integer expression if it is not overflow. 1137 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1)); 1138 Node *range = phase->transform(new SubINode(in(Limit), in(Init))); 1139 Node *bias = phase->transform(new AddINode(range, stride_m)); 1140 Node *trip = phase->transform(new DivINode(0, bias, in(Stride))); 1141 Node *span = phase->transform(new MulINode(trip, in(Stride))); 1142 return new AddINode(span, in(Init)); // exact limit 1143 } 1144 1145 if (is_power_of_2(stride_p) || // divisor is 2^n 1146 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node? 1147 // Convert to long expression to avoid integer overflow 1148 // and let igvn optimizer convert this division. 1149 // 1150 Node* init = phase->transform( new ConvI2LNode(in(Init))); 1151 Node* limit = phase->transform( new ConvI2LNode(in(Limit))); 1152 Node* stride = phase->longcon(stride_con); 1153 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1)); 1154 1155 Node *range = phase->transform(new SubLNode(limit, init)); 1156 Node *bias = phase->transform(new AddLNode(range, stride_m)); 1157 Node *span; 1158 if (stride_con > 0 && is_power_of_2(stride_p)) { 1159 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride) 1160 // and avoid generating rounding for division. Zero trip guard should 1161 // guarantee that init < limit but sometimes the guard is missing and 1162 // we can get situation when init > limit. Note, for the empty loop 1163 // optimization zero trip guard is generated explicitly which leaves 1164 // only RCE predicate where exact limit is used and the predicate 1165 // will simply fail forcing recompilation. 1166 Node* neg_stride = phase->longcon(-stride_con); 1167 span = phase->transform(new AndLNode(bias, neg_stride)); 1168 } else { 1169 Node *trip = phase->transform(new DivLNode(0, bias, stride)); 1170 span = phase->transform(new MulLNode(trip, stride)); 1171 } 1172 // Convert back to int 1173 Node *span_int = phase->transform(new ConvL2INode(span)); 1174 return new AddINode(span_int, in(Init)); // exact limit 1175 } 1176 1177 return NULL; // No progress 1178 } 1179 1180 //------------------------------Identity--------------------------------------- 1181 // If stride == 1 return limit node. 1182 Node* LoopLimitNode::Identity(PhaseGVN* phase) { 1183 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1184 if (stride_con == 1 || stride_con == -1) 1185 return in(Limit); 1186 return this; 1187 } 1188 1189 //============================================================================= 1190 //----------------------match_incr_with_optional_truncation-------------------- 1191 // Match increment with optional truncation: 1192 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 1193 // Return NULL for failure. Success returns the increment node. 1194 Node* CountedLoopNode::match_incr_with_optional_truncation( 1195 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { 1196 // Quick cutouts: 1197 if (expr == NULL || expr->req() != 3) return NULL; 1198 1199 Node *t1 = NULL; 1200 Node *t2 = NULL; 1201 const TypeInt* trunc_t = TypeInt::INT; 1202 Node* n1 = expr; 1203 int n1op = n1->Opcode(); 1204 1205 // Try to strip (n1 & M) or (n1 << N >> N) from n1. 1206 if (n1op == Op_AndI && 1207 n1->in(2)->is_Con() && 1208 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { 1209 // %%% This check should match any mask of 2**K-1. 1210 t1 = n1; 1211 n1 = t1->in(1); 1212 n1op = n1->Opcode(); 1213 trunc_t = TypeInt::CHAR; 1214 } else if (n1op == Op_RShiftI && 1215 n1->in(1) != NULL && 1216 n1->in(1)->Opcode() == Op_LShiftI && 1217 n1->in(2) == n1->in(1)->in(2) && 1218 n1->in(2)->is_Con()) { 1219 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); 1220 // %%% This check should match any shift in [1..31]. 1221 if (shift == 16 || shift == 8) { 1222 t1 = n1; 1223 t2 = t1->in(1); 1224 n1 = t2->in(1); 1225 n1op = n1->Opcode(); 1226 if (shift == 16) { 1227 trunc_t = TypeInt::SHORT; 1228 } else if (shift == 8) { 1229 trunc_t = TypeInt::BYTE; 1230 } 1231 } 1232 } 1233 1234 // If (maybe after stripping) it is an AddI, we won: 1235 if (n1op == Op_AddI) { 1236 *trunc1 = t1; 1237 *trunc2 = t2; 1238 *trunc_type = trunc_t; 1239 return n1; 1240 } 1241 1242 // failed 1243 return NULL; 1244 } 1245 1246 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) { 1247 if (is_strip_mined() && is_valid_counted_loop()) { 1248 verify_strip_mined(expect_skeleton); 1249 return in(EntryControl)->as_Loop(); 1250 } 1251 return this; 1252 } 1253 1254 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const { 1255 assert(is_strip_mined(), "not a strip mined loop"); 1256 Node* c = in(EntryControl); 1257 if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) { 1258 return NULL; 1259 } 1260 return c->as_OuterStripMinedLoop(); 1261 } 1262 1263 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const { 1264 Node* c = in(LoopBackControl); 1265 if (c == NULL || c->is_top()) { 1266 return NULL; 1267 } 1268 return c->as_IfTrue(); 1269 } 1270 1271 IfTrueNode* CountedLoopNode::outer_loop_tail() const { 1272 LoopNode* l = outer_loop(); 1273 if (l == NULL) { 1274 return NULL; 1275 } 1276 return l->outer_loop_tail(); 1277 } 1278 1279 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const { 1280 IfTrueNode* proj = outer_loop_tail(); 1281 if (proj == NULL) { 1282 return NULL; 1283 } 1284 Node* c = proj->in(0); 1285 if (c == NULL || c->is_top() || c->outcnt() != 2) { 1286 return NULL; 1287 } 1288 return c->as_OuterStripMinedLoopEnd(); 1289 } 1290 1291 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const { 1292 LoopNode* l = outer_loop(); 1293 if (l == NULL) { 1294 return NULL; 1295 } 1296 return l->outer_loop_end(); 1297 } 1298 1299 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const { 1300 IfNode* le = outer_loop_end(); 1301 if (le == NULL) { 1302 return NULL; 1303 } 1304 Node* c = le->proj_out_or_null(false); 1305 if (c == NULL) { 1306 return NULL; 1307 } 1308 return c->as_IfFalse(); 1309 } 1310 1311 IfFalseNode* CountedLoopNode::outer_loop_exit() const { 1312 LoopNode* l = outer_loop(); 1313 if (l == NULL) { 1314 return NULL; 1315 } 1316 return l->outer_loop_exit(); 1317 } 1318 1319 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const { 1320 IfNode* le = outer_loop_end(); 1321 if (le == NULL) { 1322 return NULL; 1323 } 1324 Node* c = le->in(0); 1325 if (c == NULL || c->is_top()) { 1326 return NULL; 1327 } 1328 assert(c->Opcode() == Op_SafePoint, "broken outer loop"); 1329 return c->as_SafePoint(); 1330 } 1331 1332 SafePointNode* CountedLoopNode::outer_safepoint() const { 1333 LoopNode* l = outer_loop(); 1334 if (l == NULL) { 1335 return NULL; 1336 } 1337 return l->outer_safepoint(); 1338 } 1339 1340 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) { 1341 while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() && 1342 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 && 1343 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) { 1344 ctrl = ctrl->in(0)->in(0); 1345 } 1346 1347 return ctrl; 1348 } 1349 1350 Node* CountedLoopNode::skip_predicates() { 1351 if (is_main_loop()) { 1352 Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl); 1353 1354 return skip_predicates_from_entry(ctrl); 1355 } 1356 return in(LoopNode::EntryControl); 1357 } 1358 1359 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) { 1360 // Look for the outer & inner strip mined loop, reduce number of 1361 // iterations of the inner loop, set exit condition of outer loop, 1362 // construct required phi nodes for outer loop. 1363 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop(); 1364 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined"); 1365 Node* inner_iv_phi = inner_cl->phi(); 1366 if (inner_iv_phi == NULL) { 1367 IfNode* outer_le = outer_loop_end(); 1368 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1369 igvn->replace_node(outer_le, iff); 1370 inner_cl->clear_strip_mined(); 1371 return; 1372 } 1373 CountedLoopEndNode* inner_cle = inner_cl->loopexit(); 1374 1375 int stride = inner_cl->stride_con(); 1376 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride); 1377 int scaled_iters = (int)scaled_iters_long; 1378 int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride); 1379 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int(); 1380 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo; 1381 assert(iter_estimate > 0, "broken"); 1382 if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) { 1383 // Remove outer loop and safepoint (too few iterations) 1384 Node* outer_sfpt = outer_safepoint(); 1385 Node* outer_out = outer_loop_exit(); 1386 igvn->replace_node(outer_out, outer_sfpt->in(0)); 1387 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top()); 1388 inner_cl->clear_strip_mined(); 1389 return; 1390 } 1391 if (iter_estimate <= scaled_iters_long) { 1392 // We would only go through one iteration of 1393 // the outer loop: drop the outer loop but 1394 // keep the safepoint so we don't run for 1395 // too long without a safepoint 1396 IfNode* outer_le = outer_loop_end(); 1397 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1398 igvn->replace_node(outer_le, iff); 1399 inner_cl->clear_strip_mined(); 1400 return; 1401 } 1402 1403 Node* cle_tail = inner_cle->proj_out(true); 1404 ResourceMark rm; 1405 Node_List old_new; 1406 if (cle_tail->outcnt() > 1) { 1407 // Look for nodes on backedge of inner loop and clone them 1408 Unique_Node_List backedge_nodes; 1409 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) { 1410 Node* u = cle_tail->fast_out(i); 1411 if (u != inner_cl) { 1412 assert(!u->is_CFG(), "control flow on the backedge?"); 1413 backedge_nodes.push(u); 1414 } 1415 } 1416 uint last = igvn->C->unique(); 1417 for (uint next = 0; next < backedge_nodes.size(); next++) { 1418 Node* n = backedge_nodes.at(next); 1419 old_new.map(n->_idx, n->clone()); 1420 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1421 Node* u = n->fast_out(i); 1422 assert(!u->is_CFG(), "broken"); 1423 if (u->_idx >= last) { 1424 continue; 1425 } 1426 if (!u->is_Phi()) { 1427 backedge_nodes.push(u); 1428 } else { 1429 assert(u->in(0) == inner_cl, "strange phi on the backedge"); 1430 } 1431 } 1432 } 1433 // Put the clones on the outer loop backedge 1434 Node* le_tail = outer_loop_tail(); 1435 for (uint next = 0; next < backedge_nodes.size(); next++) { 1436 Node *n = old_new[backedge_nodes.at(next)->_idx]; 1437 for (uint i = 1; i < n->req(); i++) { 1438 if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) { 1439 n->set_req(i, old_new[n->in(i)->_idx]); 1440 } 1441 } 1442 if (n->in(0) != NULL && n->in(0) == cle_tail) { 1443 n->set_req(0, le_tail); 1444 } 1445 igvn->register_new_node_with_optimizer(n); 1446 } 1447 } 1448 1449 Node* iv_phi = NULL; 1450 // Make a clone of each phi in the inner loop 1451 // for the outer loop 1452 for (uint i = 0; i < inner_cl->outcnt(); i++) { 1453 Node* u = inner_cl->raw_out(i); 1454 if (u->is_Phi()) { 1455 assert(u->in(0) == inner_cl, "inconsistent"); 1456 Node* phi = u->clone(); 1457 phi->set_req(0, this); 1458 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx]; 1459 if (be != NULL) { 1460 phi->set_req(LoopNode::LoopBackControl, be); 1461 } 1462 phi = igvn->transform(phi); 1463 igvn->replace_input_of(u, LoopNode::EntryControl, phi); 1464 if (u == inner_iv_phi) { 1465 iv_phi = phi; 1466 } 1467 } 1468 } 1469 Node* cle_out = inner_cle->proj_out(false); 1470 if (cle_out->outcnt() > 1) { 1471 // Look for chains of stores that were sunk 1472 // out of the inner loop and are in the outer loop 1473 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) { 1474 Node* u = cle_out->fast_out(i); 1475 if (u->is_Store()) { 1476 Node* first = u; 1477 for(;;) { 1478 Node* next = first->in(MemNode::Memory); 1479 if (!next->is_Store() || next->in(0) != cle_out) { 1480 break; 1481 } 1482 first = next; 1483 } 1484 Node* last = u; 1485 for(;;) { 1486 Node* next = NULL; 1487 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) { 1488 Node* uu = last->fast_out(j); 1489 if (uu->is_Store() && uu->in(0) == cle_out) { 1490 assert(next == NULL, "only one in the outer loop"); 1491 next = uu; 1492 } 1493 } 1494 if (next == NULL) { 1495 break; 1496 } 1497 last = next; 1498 } 1499 Node* phi = NULL; 1500 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1501 Node* uu = fast_out(j); 1502 if (uu->is_Phi()) { 1503 Node* be = uu->in(LoopNode::LoopBackControl); 1504 if (be->is_Store() && old_new[be->_idx] != NULL) { 1505 assert(false, "store on the backedge + sunk stores: unsupported"); 1506 // drop outer loop 1507 IfNode* outer_le = outer_loop_end(); 1508 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1509 igvn->replace_node(outer_le, iff); 1510 inner_cl->clear_strip_mined(); 1511 return; 1512 } 1513 if (be == last || be == first->in(MemNode::Memory)) { 1514 assert(phi == NULL, "only one phi"); 1515 phi = uu; 1516 } 1517 } 1518 } 1519 #ifdef ASSERT 1520 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1521 Node* uu = fast_out(j); 1522 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) { 1523 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) { 1524 assert(phi == uu, "what's that phi?"); 1525 } else if (uu->adr_type() == TypePtr::BOTTOM) { 1526 Node* n = uu->in(LoopNode::LoopBackControl); 1527 uint limit = igvn->C->live_nodes(); 1528 uint i = 0; 1529 while (n != uu) { 1530 i++; 1531 assert(i < limit, "infinite loop"); 1532 if (n->is_Proj()) { 1533 n = n->in(0); 1534 } else if (n->is_SafePoint() || n->is_MemBar()) { 1535 n = n->in(TypeFunc::Memory); 1536 } else if (n->is_Phi()) { 1537 n = n->in(1); 1538 } else if (n->is_MergeMem()) { 1539 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type())); 1540 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) { 1541 n = n->in(MemNode::Memory); 1542 } else { 1543 n->dump(); 1544 ShouldNotReachHere(); 1545 } 1546 } 1547 } 1548 } 1549 } 1550 #endif 1551 if (phi == NULL) { 1552 // If the an entire chains was sunk, the 1553 // inner loop has no phi for that memory 1554 // slice, create one for the outer loop 1555 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY, 1556 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))); 1557 phi->set_req(LoopNode::LoopBackControl, last); 1558 phi = igvn->transform(phi); 1559 igvn->replace_input_of(first, MemNode::Memory, phi); 1560 } else { 1561 // Or fix the outer loop fix to include 1562 // that chain of stores. 1563 Node* be = phi->in(LoopNode::LoopBackControl); 1564 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported"); 1565 if (be == first->in(MemNode::Memory)) { 1566 if (be == phi->in(LoopNode::LoopBackControl)) { 1567 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last); 1568 } else { 1569 igvn->replace_input_of(be, MemNode::Memory, last); 1570 } 1571 } else { 1572 #ifdef ASSERT 1573 if (be == phi->in(LoopNode::LoopBackControl)) { 1574 assert(phi->in(LoopNode::LoopBackControl) == last, ""); 1575 } else { 1576 assert(be->in(MemNode::Memory) == last, ""); 1577 } 1578 #endif 1579 } 1580 } 1581 } 1582 } 1583 } 1584 1585 if (iv_phi != NULL) { 1586 // Now adjust the inner loop's exit condition 1587 Node* limit = inner_cl->limit(); 1588 Node* sub = NULL; 1589 if (stride > 0) { 1590 sub = igvn->transform(new SubINode(limit, iv_phi)); 1591 } else { 1592 sub = igvn->transform(new SubINode(iv_phi, limit)); 1593 } 1594 // sub is positive and can be larger than the max signed int 1595 // value. Use an unsigned min. 1596 Node* const_iters = igvn->intcon(scaled_iters); 1597 Node* cmp = igvn->transform(new CmpUNode(sub, const_iters)); 1598 Node* bol = igvn->transform(new BoolNode(cmp, BoolTest::lt)); 1599 Node* min = igvn->transform(new CMoveINode(bol, const_iters, sub, TypeInt::make(0, scaled_iters, Type::WidenMin))); 1600 1601 Node* new_limit = NULL; 1602 if (stride > 0) { 1603 new_limit = igvn->transform(new AddINode(min, iv_phi)); 1604 } else { 1605 new_limit = igvn->transform(new SubINode(iv_phi, min)); 1606 } 1607 Node* inner_cmp = inner_cle->cmp_node(); 1608 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue); 1609 Node* outer_bol = inner_bol; 1610 // cmp node for inner loop may be shared 1611 inner_cmp = inner_cmp->clone(); 1612 inner_cmp->set_req(2, new_limit); 1613 inner_bol = inner_bol->clone(); 1614 inner_bol->set_req(1, igvn->transform(inner_cmp)); 1615 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol)); 1616 // Set the outer loop's exit condition too 1617 igvn->replace_input_of(outer_loop_end(), 1, outer_bol); 1618 } else { 1619 assert(false, "should be able to adjust outer loop"); 1620 IfNode* outer_le = outer_loop_end(); 1621 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1622 igvn->replace_node(outer_le, iff); 1623 inner_cl->clear_strip_mined(); 1624 } 1625 } 1626 1627 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const { 1628 if (!in(0)) return Type::TOP; 1629 if (phase->type(in(0)) == Type::TOP) 1630 return Type::TOP; 1631 1632 return TypeTuple::IFBOTH; 1633 } 1634 1635 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1636 if (remove_dead_region(phase, can_reshape)) return this; 1637 1638 return NULL; 1639 } 1640 1641 //------------------------------filtered_type-------------------------------- 1642 // Return a type based on condition control flow 1643 // A successful return will be a type that is restricted due 1644 // to a series of dominating if-tests, such as: 1645 // if (i < 10) { 1646 // if (i > 0) { 1647 // here: "i" type is [1..10) 1648 // } 1649 // } 1650 // or a control flow merge 1651 // if (i < 10) { 1652 // do { 1653 // phi( , ) -- at top of loop type is [min_int..10) 1654 // i = ? 1655 // } while ( i < 10) 1656 // 1657 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { 1658 assert(n && n->bottom_type()->is_int(), "must be int"); 1659 const TypeInt* filtered_t = NULL; 1660 if (!n->is_Phi()) { 1661 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); 1662 filtered_t = filtered_type_from_dominators(n, n_ctrl); 1663 1664 } else { 1665 Node* phi = n->as_Phi(); 1666 Node* region = phi->in(0); 1667 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); 1668 if (region && region != C->top()) { 1669 for (uint i = 1; i < phi->req(); i++) { 1670 Node* val = phi->in(i); 1671 Node* use_c = region->in(i); 1672 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); 1673 if (val_t != NULL) { 1674 if (filtered_t == NULL) { 1675 filtered_t = val_t; 1676 } else { 1677 filtered_t = filtered_t->meet(val_t)->is_int(); 1678 } 1679 } 1680 } 1681 } 1682 } 1683 const TypeInt* n_t = _igvn.type(n)->is_int(); 1684 if (filtered_t != NULL) { 1685 n_t = n_t->join(filtered_t)->is_int(); 1686 } 1687 return n_t; 1688 } 1689 1690 1691 //------------------------------filtered_type_from_dominators-------------------------------- 1692 // Return a possibly more restrictive type for val based on condition control flow of dominators 1693 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { 1694 if (val->is_Con()) { 1695 return val->bottom_type()->is_int(); 1696 } 1697 uint if_limit = 10; // Max number of dominating if's visited 1698 const TypeInt* rtn_t = NULL; 1699 1700 if (use_ctrl && use_ctrl != C->top()) { 1701 Node* val_ctrl = get_ctrl(val); 1702 uint val_dom_depth = dom_depth(val_ctrl); 1703 Node* pred = use_ctrl; 1704 uint if_cnt = 0; 1705 while (if_cnt < if_limit) { 1706 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { 1707 if_cnt++; 1708 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); 1709 if (if_t != NULL) { 1710 if (rtn_t == NULL) { 1711 rtn_t = if_t; 1712 } else { 1713 rtn_t = rtn_t->join(if_t)->is_int(); 1714 } 1715 } 1716 } 1717 pred = idom(pred); 1718 if (pred == NULL || pred == C->top()) { 1719 break; 1720 } 1721 // Stop if going beyond definition block of val 1722 if (dom_depth(pred) < val_dom_depth) { 1723 break; 1724 } 1725 } 1726 } 1727 return rtn_t; 1728 } 1729 1730 1731 //------------------------------dump_spec-------------------------------------- 1732 // Dump special per-node info 1733 #ifndef PRODUCT 1734 void CountedLoopEndNode::dump_spec(outputStream *st) const { 1735 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) { 1736 BoolTest bt( test_trip()); // Added this for g++. 1737 1738 st->print("["); 1739 bt.dump_on(st); 1740 st->print("]"); 1741 } 1742 st->print(" "); 1743 IfNode::dump_spec(st); 1744 } 1745 #endif 1746 1747 //============================================================================= 1748 //------------------------------is_member-------------------------------------- 1749 // Is 'l' a member of 'this'? 1750 bool IdealLoopTree::is_member(const IdealLoopTree *l) const { 1751 while( l->_nest > _nest ) l = l->_parent; 1752 return l == this; 1753 } 1754 1755 //------------------------------set_nest--------------------------------------- 1756 // Set loop tree nesting depth. Accumulate _has_call bits. 1757 int IdealLoopTree::set_nest( uint depth ) { 1758 _nest = depth; 1759 int bits = _has_call; 1760 if( _child ) bits |= _child->set_nest(depth+1); 1761 if( bits ) _has_call = 1; 1762 if( _next ) bits |= _next ->set_nest(depth ); 1763 return bits; 1764 } 1765 1766 //------------------------------split_fall_in---------------------------------- 1767 // Split out multiple fall-in edges from the loop header. Move them to a 1768 // private RegionNode before the loop. This becomes the loop landing pad. 1769 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { 1770 PhaseIterGVN &igvn = phase->_igvn; 1771 uint i; 1772 1773 // Make a new RegionNode to be the landing pad. 1774 Node *landing_pad = new RegionNode( fall_in_cnt+1 ); 1775 phase->set_loop(landing_pad,_parent); 1776 // Gather all the fall-in control paths into the landing pad 1777 uint icnt = fall_in_cnt; 1778 uint oreq = _head->req(); 1779 for( i = oreq-1; i>0; i-- ) 1780 if( !phase->is_member( this, _head->in(i) ) ) 1781 landing_pad->set_req(icnt--,_head->in(i)); 1782 1783 // Peel off PhiNode edges as well 1784 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1785 Node *oj = _head->fast_out(j); 1786 if( oj->is_Phi() ) { 1787 PhiNode* old_phi = oj->as_Phi(); 1788 assert( old_phi->region() == _head, "" ); 1789 igvn.hash_delete(old_phi); // Yank from hash before hacking edges 1790 Node *p = PhiNode::make_blank(landing_pad, old_phi); 1791 uint icnt = fall_in_cnt; 1792 for( i = oreq-1; i>0; i-- ) { 1793 if( !phase->is_member( this, _head->in(i) ) ) { 1794 p->init_req(icnt--, old_phi->in(i)); 1795 // Go ahead and clean out old edges from old phi 1796 old_phi->del_req(i); 1797 } 1798 } 1799 // Search for CSE's here, because ZKM.jar does a lot of 1800 // loop hackery and we need to be a little incremental 1801 // with the CSE to avoid O(N^2) node blow-up. 1802 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE 1803 if( p2 ) { // Found CSE 1804 p->destruct(); // Recover useless new node 1805 p = p2; // Use old node 1806 } else { 1807 igvn.register_new_node_with_optimizer(p, old_phi); 1808 } 1809 // Make old Phi refer to new Phi. 1810 old_phi->add_req(p); 1811 // Check for the special case of making the old phi useless and 1812 // disappear it. In JavaGrande I have a case where this useless 1813 // Phi is the loop limit and prevents recognizing a CountedLoop 1814 // which in turn prevents removing an empty loop. 1815 Node *id_old_phi = old_phi->Identity(&igvn); 1816 if( id_old_phi != old_phi ) { // Found a simple identity? 1817 // Note that I cannot call 'replace_node' here, because 1818 // that will yank the edge from old_phi to the Region and 1819 // I'm mid-iteration over the Region's uses. 1820 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { 1821 Node* use = old_phi->last_out(i); 1822 igvn.rehash_node_delayed(use); 1823 uint uses_found = 0; 1824 for (uint j = 0; j < use->len(); j++) { 1825 if (use->in(j) == old_phi) { 1826 if (j < use->req()) use->set_req (j, id_old_phi); 1827 else use->set_prec(j, id_old_phi); 1828 uses_found++; 1829 } 1830 } 1831 i -= uses_found; // we deleted 1 or more copies of this edge 1832 } 1833 } 1834 igvn._worklist.push(old_phi); 1835 } 1836 } 1837 // Finally clean out the fall-in edges from the RegionNode 1838 for( i = oreq-1; i>0; i-- ) { 1839 if( !phase->is_member( this, _head->in(i) ) ) { 1840 _head->del_req(i); 1841 } 1842 } 1843 igvn.rehash_node_delayed(_head); 1844 // Transform landing pad 1845 igvn.register_new_node_with_optimizer(landing_pad, _head); 1846 // Insert landing pad into the header 1847 _head->add_req(landing_pad); 1848 } 1849 1850 //------------------------------split_outer_loop------------------------------- 1851 // Split out the outermost loop from this shared header. 1852 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { 1853 PhaseIterGVN &igvn = phase->_igvn; 1854 1855 // Find index of outermost loop; it should also be my tail. 1856 uint outer_idx = 1; 1857 while( _head->in(outer_idx) != _tail ) outer_idx++; 1858 1859 // Make a LoopNode for the outermost loop. 1860 Node *ctl = _head->in(LoopNode::EntryControl); 1861 Node *outer = new LoopNode( ctl, _head->in(outer_idx) ); 1862 outer = igvn.register_new_node_with_optimizer(outer, _head); 1863 phase->set_created_loop_node(); 1864 1865 // Outermost loop falls into '_head' loop 1866 _head->set_req(LoopNode::EntryControl, outer); 1867 _head->del_req(outer_idx); 1868 // Split all the Phis up between '_head' loop and 'outer' loop. 1869 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1870 Node *out = _head->fast_out(j); 1871 if( out->is_Phi() ) { 1872 PhiNode *old_phi = out->as_Phi(); 1873 assert( old_phi->region() == _head, "" ); 1874 Node *phi = PhiNode::make_blank(outer, old_phi); 1875 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); 1876 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); 1877 phi = igvn.register_new_node_with_optimizer(phi, old_phi); 1878 // Make old Phi point to new Phi on the fall-in path 1879 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi); 1880 old_phi->del_req(outer_idx); 1881 } 1882 } 1883 1884 // Use the new loop head instead of the old shared one 1885 _head = outer; 1886 phase->set_loop(_head, this); 1887 } 1888 1889 //------------------------------fix_parent------------------------------------- 1890 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { 1891 loop->_parent = parent; 1892 if( loop->_child ) fix_parent( loop->_child, loop ); 1893 if( loop->_next ) fix_parent( loop->_next , parent ); 1894 } 1895 1896 //------------------------------estimate_path_freq----------------------------- 1897 static float estimate_path_freq( Node *n ) { 1898 // Try to extract some path frequency info 1899 IfNode *iff; 1900 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests 1901 uint nop = n->Opcode(); 1902 if( nop == Op_SafePoint ) { // Skip any safepoint 1903 n = n->in(0); 1904 continue; 1905 } 1906 if( nop == Op_CatchProj ) { // Get count from a prior call 1907 // Assume call does not always throw exceptions: means the call-site 1908 // count is also the frequency of the fall-through path. 1909 assert( n->is_CatchProj(), "" ); 1910 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) 1911 return 0.0f; // Assume call exception path is rare 1912 Node *call = n->in(0)->in(0)->in(0); 1913 assert( call->is_Call(), "expect a call here" ); 1914 const JVMState *jvms = ((CallNode*)call)->jvms(); 1915 ciMethodData* methodData = jvms->method()->method_data(); 1916 if (!methodData->is_mature()) return 0.0f; // No call-site data 1917 ciProfileData* data = methodData->bci_to_data(jvms->bci()); 1918 if ((data == NULL) || !data->is_CounterData()) { 1919 // no call profile available, try call's control input 1920 n = n->in(0); 1921 continue; 1922 } 1923 return data->as_CounterData()->count()/FreqCountInvocations; 1924 } 1925 // See if there's a gating IF test 1926 Node *n_c = n->in(0); 1927 if( !n_c->is_If() ) break; // No estimate available 1928 iff = n_c->as_If(); 1929 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? 1930 // Compute how much count comes on this path 1931 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; 1932 // Have no count info. Skip dull uncommon-trap like branches. 1933 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || 1934 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) 1935 break; 1936 // Skip through never-taken branch; look for a real loop exit. 1937 n = iff->in(0); 1938 } 1939 return 0.0f; // No estimate available 1940 } 1941 1942 //------------------------------merge_many_backedges--------------------------- 1943 // Merge all the backedges from the shared header into a private Region. 1944 // Feed that region as the one backedge to this loop. 1945 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { 1946 uint i; 1947 1948 // Scan for the top 2 hottest backedges 1949 float hotcnt = 0.0f; 1950 float warmcnt = 0.0f; 1951 uint hot_idx = 0; 1952 // Loop starts at 2 because slot 1 is the fall-in path 1953 for( i = 2; i < _head->req(); i++ ) { 1954 float cnt = estimate_path_freq(_head->in(i)); 1955 if( cnt > hotcnt ) { // Grab hottest path 1956 warmcnt = hotcnt; 1957 hotcnt = cnt; 1958 hot_idx = i; 1959 } else if( cnt > warmcnt ) { // And 2nd hottest path 1960 warmcnt = cnt; 1961 } 1962 } 1963 1964 // See if the hottest backedge is worthy of being an inner loop 1965 // by being much hotter than the next hottest backedge. 1966 if( hotcnt <= 0.0001 || 1967 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge 1968 1969 // Peel out the backedges into a private merge point; peel 1970 // them all except optionally hot_idx. 1971 PhaseIterGVN &igvn = phase->_igvn; 1972 1973 Node *hot_tail = NULL; 1974 // Make a Region for the merge point 1975 Node *r = new RegionNode(1); 1976 for( i = 2; i < _head->req(); i++ ) { 1977 if( i != hot_idx ) 1978 r->add_req( _head->in(i) ); 1979 else hot_tail = _head->in(i); 1980 } 1981 igvn.register_new_node_with_optimizer(r, _head); 1982 // Plug region into end of loop _head, followed by hot_tail 1983 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); 1984 igvn.replace_input_of(_head, 2, r); 1985 if( hot_idx ) _head->add_req(hot_tail); 1986 1987 // Split all the Phis up between '_head' loop and the Region 'r' 1988 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1989 Node *out = _head->fast_out(j); 1990 if( out->is_Phi() ) { 1991 PhiNode* n = out->as_Phi(); 1992 igvn.hash_delete(n); // Delete from hash before hacking edges 1993 Node *hot_phi = NULL; 1994 Node *phi = new PhiNode(r, n->type(), n->adr_type()); 1995 // Check all inputs for the ones to peel out 1996 uint j = 1; 1997 for( uint i = 2; i < n->req(); i++ ) { 1998 if( i != hot_idx ) 1999 phi->set_req( j++, n->in(i) ); 2000 else hot_phi = n->in(i); 2001 } 2002 // Register the phi but do not transform until whole place transforms 2003 igvn.register_new_node_with_optimizer(phi, n); 2004 // Add the merge phi to the old Phi 2005 while( n->req() > 3 ) n->del_req( n->req()-1 ); 2006 igvn.replace_input_of(n, 2, phi); 2007 if( hot_idx ) n->add_req(hot_phi); 2008 } 2009 } 2010 2011 2012 // Insert a new IdealLoopTree inserted below me. Turn it into a clone 2013 // of self loop tree. Turn self into a loop headed by _head and with 2014 // tail being the new merge point. 2015 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); 2016 phase->set_loop(_tail,ilt); // Adjust tail 2017 _tail = r; // Self's tail is new merge point 2018 phase->set_loop(r,this); 2019 ilt->_child = _child; // New guy has my children 2020 _child = ilt; // Self has new guy as only child 2021 ilt->_parent = this; // new guy has self for parent 2022 ilt->_nest = _nest; // Same nesting depth (for now) 2023 2024 // Starting with 'ilt', look for child loop trees using the same shared 2025 // header. Flatten these out; they will no longer be loops in the end. 2026 IdealLoopTree **pilt = &_child; 2027 while( ilt ) { 2028 if( ilt->_head == _head ) { 2029 uint i; 2030 for( i = 2; i < _head->req(); i++ ) 2031 if( _head->in(i) == ilt->_tail ) 2032 break; // Still a loop 2033 if( i == _head->req() ) { // No longer a loop 2034 // Flatten ilt. Hang ilt's "_next" list from the end of 2035 // ilt's '_child' list. Move the ilt's _child up to replace ilt. 2036 IdealLoopTree **cp = &ilt->_child; 2037 while( *cp ) cp = &(*cp)->_next; // Find end of child list 2038 *cp = ilt->_next; // Hang next list at end of child list 2039 *pilt = ilt->_child; // Move child up to replace ilt 2040 ilt->_head = NULL; // Flag as a loop UNIONED into parent 2041 ilt = ilt->_child; // Repeat using new ilt 2042 continue; // do not advance over ilt->_child 2043 } 2044 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); 2045 phase->set_loop(_head,ilt); 2046 } 2047 pilt = &ilt->_child; // Advance to next 2048 ilt = *pilt; 2049 } 2050 2051 if( _child ) fix_parent( _child, this ); 2052 } 2053 2054 //------------------------------beautify_loops--------------------------------- 2055 // Split shared headers and insert loop landing pads. 2056 // Insert a LoopNode to replace the RegionNode. 2057 // Return TRUE if loop tree is structurally changed. 2058 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { 2059 bool result = false; 2060 // Cache parts in locals for easy 2061 PhaseIterGVN &igvn = phase->_igvn; 2062 2063 igvn.hash_delete(_head); // Yank from hash before hacking edges 2064 2065 // Check for multiple fall-in paths. Peel off a landing pad if need be. 2066 int fall_in_cnt = 0; 2067 for( uint i = 1; i < _head->req(); i++ ) 2068 if( !phase->is_member( this, _head->in(i) ) ) 2069 fall_in_cnt++; 2070 assert( fall_in_cnt, "at least 1 fall-in path" ); 2071 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins 2072 split_fall_in( phase, fall_in_cnt ); 2073 2074 // Swap inputs to the _head and all Phis to move the fall-in edge to 2075 // the left. 2076 fall_in_cnt = 1; 2077 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) 2078 fall_in_cnt++; 2079 if( fall_in_cnt > 1 ) { 2080 // Since I am just swapping inputs I do not need to update def-use info 2081 Node *tmp = _head->in(1); 2082 igvn.rehash_node_delayed(_head); 2083 _head->set_req( 1, _head->in(fall_in_cnt) ); 2084 _head->set_req( fall_in_cnt, tmp ); 2085 // Swap also all Phis 2086 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { 2087 Node* phi = _head->fast_out(i); 2088 if( phi->is_Phi() ) { 2089 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges 2090 tmp = phi->in(1); 2091 phi->set_req( 1, phi->in(fall_in_cnt) ); 2092 phi->set_req( fall_in_cnt, tmp ); 2093 } 2094 } 2095 } 2096 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); 2097 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); 2098 2099 // If I am a shared header (multiple backedges), peel off the many 2100 // backedges into a private merge point and use the merge point as 2101 // the one true backedge. 2102 if (_head->req() > 3 && !_irreducible) { 2103 // Merge the many backedges into a single backedge but leave 2104 // the hottest backedge as separate edge for the following peel. 2105 merge_many_backedges( phase ); 2106 result = true; 2107 } 2108 2109 // If I have one hot backedge, peel off myself loop. 2110 // I better be the outermost loop. 2111 if (_head->req() > 3 && !_irreducible) { 2112 split_outer_loop( phase ); 2113 result = true; 2114 2115 } else if (!_head->is_Loop() && !_irreducible) { 2116 // Make a new LoopNode to replace the old loop head 2117 Node *l = new LoopNode( _head->in(1), _head->in(2) ); 2118 l = igvn.register_new_node_with_optimizer(l, _head); 2119 phase->set_created_loop_node(); 2120 // Go ahead and replace _head 2121 phase->_igvn.replace_node( _head, l ); 2122 _head = l; 2123 phase->set_loop(_head, this); 2124 } 2125 2126 // Now recursively beautify nested loops 2127 if( _child ) result |= _child->beautify_loops( phase ); 2128 if( _next ) result |= _next ->beautify_loops( phase ); 2129 return result; 2130 } 2131 2132 //------------------------------allpaths_check_safepts---------------------------- 2133 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 2134 // encountered. Helper for check_safepts. 2135 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { 2136 assert(stack.size() == 0, "empty stack"); 2137 stack.push(_tail); 2138 visited.clear(); 2139 visited.set(_tail->_idx); 2140 while (stack.size() > 0) { 2141 Node* n = stack.pop(); 2142 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2143 // Terminate this path 2144 } else if (n->Opcode() == Op_SafePoint) { 2145 if (_phase->get_loop(n) != this) { 2146 if (_required_safept == NULL) _required_safept = new Node_List(); 2147 _required_safept->push(n); // save the one closest to the tail 2148 } 2149 // Terminate this path 2150 } else { 2151 uint start = n->is_Region() ? 1 : 0; 2152 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; 2153 for (uint i = start; i < end; i++) { 2154 Node* in = n->in(i); 2155 assert(in->is_CFG(), "must be"); 2156 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { 2157 stack.push(in); 2158 } 2159 } 2160 } 2161 } 2162 } 2163 2164 //------------------------------check_safepts---------------------------- 2165 // Given dominators, try to find loops with calls that must always be 2166 // executed (call dominates loop tail). These loops do not need non-call 2167 // safepoints (ncsfpt). 2168 // 2169 // A complication is that a safepoint in a inner loop may be needed 2170 // by an outer loop. In the following, the inner loop sees it has a 2171 // call (block 3) on every path from the head (block 2) to the 2172 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) 2173 // in block 2, _but_ this leaves the outer loop without a safepoint. 2174 // 2175 // entry 0 2176 // | 2177 // v 2178 // outer 1,2 +->1 2179 // | | 2180 // | v 2181 // | 2<---+ ncsfpt in 2 2182 // |_/|\ | 2183 // | v | 2184 // inner 2,3 / 3 | call in 3 2185 // / | | 2186 // v +--+ 2187 // exit 4 2188 // 2189 // 2190 // This method creates a list (_required_safept) of ncsfpt nodes that must 2191 // be protected is created for each loop. When a ncsfpt maybe deleted, it 2192 // is first looked for in the lists for the outer loops of the current loop. 2193 // 2194 // The insights into the problem: 2195 // A) counted loops are okay 2196 // B) innermost loops are okay (only an inner loop can delete 2197 // a ncsfpt needed by an outer loop) 2198 // C) a loop is immune from an inner loop deleting a safepoint 2199 // if the loop has a call on the idom-path 2200 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the 2201 // idom-path that is not in a nested loop 2202 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner 2203 // loop needs to be prevented from deletion by an inner loop 2204 // 2205 // There are two analyses: 2206 // 1) The first, and cheaper one, scans the loop body from 2207 // tail to head following the idom (immediate dominator) 2208 // chain, looking for the cases (C,D,E) above. 2209 // Since inner loops are scanned before outer loops, there is summary 2210 // information about inner loops. Inner loops can be skipped over 2211 // when the tail of an inner loop is encountered. 2212 // 2213 // 2) The second, invoked if the first fails to find a call or ncsfpt on 2214 // the idom path (which is rare), scans all predecessor control paths 2215 // from the tail to the head, terminating a path when a call or sfpt 2216 // is encountered, to find the ncsfpt's that are closest to the tail. 2217 // 2218 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { 2219 // Bottom up traversal 2220 IdealLoopTree* ch = _child; 2221 if (_child) _child->check_safepts(visited, stack); 2222 if (_next) _next ->check_safepts(visited, stack); 2223 2224 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { 2225 bool has_call = false; // call on dom-path 2226 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth 2227 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth 2228 // Scan the dom-path nodes from tail to head 2229 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { 2230 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2231 has_call = true; 2232 _has_sfpt = 1; // Then no need for a safept! 2233 break; 2234 } else if (n->Opcode() == Op_SafePoint) { 2235 if (_phase->get_loop(n) == this) { 2236 has_local_ncsfpt = true; 2237 break; 2238 } 2239 if (nonlocal_ncsfpt == NULL) { 2240 nonlocal_ncsfpt = n; // save the one closest to the tail 2241 } 2242 } else { 2243 IdealLoopTree* nlpt = _phase->get_loop(n); 2244 if (this != nlpt) { 2245 // If at an inner loop tail, see if the inner loop has already 2246 // recorded seeing a call on the dom-path (and stop.) If not, 2247 // jump to the head of the inner loop. 2248 assert(is_member(nlpt), "nested loop"); 2249 Node* tail = nlpt->_tail; 2250 if (tail->in(0)->is_If()) tail = tail->in(0); 2251 if (n == tail) { 2252 // If inner loop has call on dom-path, so does outer loop 2253 if (nlpt->_has_sfpt) { 2254 has_call = true; 2255 _has_sfpt = 1; 2256 break; 2257 } 2258 // Skip to head of inner loop 2259 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); 2260 n = nlpt->_head; 2261 } 2262 } 2263 } 2264 } 2265 // Record safept's that this loop needs preserved when an 2266 // inner loop attempts to delete it's safepoints. 2267 if (_child != NULL && !has_call && !has_local_ncsfpt) { 2268 if (nonlocal_ncsfpt != NULL) { 2269 if (_required_safept == NULL) _required_safept = new Node_List(); 2270 _required_safept->push(nonlocal_ncsfpt); 2271 } else { 2272 // Failed to find a suitable safept on the dom-path. Now use 2273 // an all paths walk from tail to head, looking for safepoints to preserve. 2274 allpaths_check_safepts(visited, stack); 2275 } 2276 } 2277 } 2278 } 2279 2280 //---------------------------is_deleteable_safept---------------------------- 2281 // Is safept not required by an outer loop? 2282 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { 2283 assert(sfpt->Opcode() == Op_SafePoint, ""); 2284 IdealLoopTree* lp = get_loop(sfpt)->_parent; 2285 while (lp != NULL) { 2286 Node_List* sfpts = lp->_required_safept; 2287 if (sfpts != NULL) { 2288 for (uint i = 0; i < sfpts->size(); i++) { 2289 if (sfpt == sfpts->at(i)) 2290 return false; 2291 } 2292 } 2293 lp = lp->_parent; 2294 } 2295 return true; 2296 } 2297 2298 //---------------------------replace_parallel_iv------------------------------- 2299 // Replace parallel induction variable (parallel to trip counter) 2300 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) { 2301 assert(loop->_head->is_CountedLoop(), ""); 2302 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2303 if (!cl->is_valid_counted_loop()) 2304 return; // skip malformed counted loop 2305 Node *incr = cl->incr(); 2306 if (incr == NULL) 2307 return; // Dead loop? 2308 Node *init = cl->init_trip(); 2309 Node *phi = cl->phi(); 2310 int stride_con = cl->stride_con(); 2311 2312 // Visit all children, looking for Phis 2313 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { 2314 Node *out = cl->out(i); 2315 // Look for other phis (secondary IVs). Skip dead ones 2316 if (!out->is_Phi() || out == phi || !has_node(out)) 2317 continue; 2318 PhiNode* phi2 = out->as_Phi(); 2319 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); 2320 // Look for induction variables of the form: X += constant 2321 if (phi2->region() != loop->_head || 2322 incr2->req() != 3 || 2323 incr2->in(1) != phi2 || 2324 incr2 == incr || 2325 incr2->Opcode() != Op_AddI || 2326 !incr2->in(2)->is_Con()) 2327 continue; 2328 2329 // Check for parallel induction variable (parallel to trip counter) 2330 // via an affine function. In particular, count-down loops with 2331 // count-up array indices are common. We only RCE references off 2332 // the trip-counter, so we need to convert all these to trip-counter 2333 // expressions. 2334 Node *init2 = phi2->in( LoopNode::EntryControl ); 2335 int stride_con2 = incr2->in(2)->get_int(); 2336 2337 // The ratio of the two strides cannot be represented as an int 2338 // if stride_con2 is min_int and stride_con is -1. 2339 if (stride_con2 == min_jint && stride_con == -1) { 2340 continue; 2341 } 2342 2343 // The general case here gets a little tricky. We want to find the 2344 // GCD of all possible parallel IV's and make a new IV using this 2345 // GCD for the loop. Then all possible IVs are simple multiples of 2346 // the GCD. In practice, this will cover very few extra loops. 2347 // Instead we require 'stride_con2' to be a multiple of 'stride_con', 2348 // where +/-1 is the common case, but other integer multiples are 2349 // also easy to handle. 2350 int ratio_con = stride_con2/stride_con; 2351 2352 if ((ratio_con * stride_con) == stride_con2) { // Check for exact 2353 #ifndef PRODUCT 2354 if (TraceLoopOpts) { 2355 tty->print("Parallel IV: %d ", phi2->_idx); 2356 loop->dump_head(); 2357 } 2358 #endif 2359 // Convert to using the trip counter. The parallel induction 2360 // variable differs from the trip counter by a loop-invariant 2361 // amount, the difference between their respective initial values. 2362 // It is scaled by the 'ratio_con'. 2363 Node* ratio = _igvn.intcon(ratio_con); 2364 set_ctrl(ratio, C->root()); 2365 Node* ratio_init = new MulINode(init, ratio); 2366 _igvn.register_new_node_with_optimizer(ratio_init, init); 2367 set_early_ctrl(ratio_init); 2368 Node* diff = new SubINode(init2, ratio_init); 2369 _igvn.register_new_node_with_optimizer(diff, init2); 2370 set_early_ctrl(diff); 2371 Node* ratio_idx = new MulINode(phi, ratio); 2372 _igvn.register_new_node_with_optimizer(ratio_idx, phi); 2373 set_ctrl(ratio_idx, cl); 2374 Node* add = new AddINode(ratio_idx, diff); 2375 _igvn.register_new_node_with_optimizer(add); 2376 set_ctrl(add, cl); 2377 _igvn.replace_node( phi2, add ); 2378 // Sometimes an induction variable is unused 2379 if (add->outcnt() == 0) { 2380 _igvn.remove_dead_node(add); 2381 } 2382 --i; // deleted this phi; rescan starting with next position 2383 continue; 2384 } 2385 } 2386 } 2387 2388 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) { 2389 Node* keep = NULL; 2390 if (keep_one) { 2391 // Look for a safepoint on the idom-path. 2392 for (Node* i = tail(); i != _head; i = phase->idom(i)) { 2393 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) { 2394 keep = i; 2395 break; // Found one 2396 } 2397 } 2398 } 2399 2400 // Don't remove any safepoints if it is requested to keep a single safepoint and 2401 // no safepoint was found on idom-path. It is not safe to remove any safepoint 2402 // in this case since there's no safepoint dominating all paths in the loop body. 2403 bool prune = !keep_one || keep != NULL; 2404 2405 // Delete other safepoints in this loop. 2406 Node_List* sfpts = _safepts; 2407 if (prune && sfpts != NULL) { 2408 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint"); 2409 for (uint i = 0; i < sfpts->size(); i++) { 2410 Node* n = sfpts->at(i); 2411 assert(phase->get_loop(n) == this, ""); 2412 if (n != keep && phase->is_deleteable_safept(n)) { 2413 phase->lazy_replace(n, n->in(TypeFunc::Control)); 2414 } 2415 } 2416 } 2417 } 2418 2419 //------------------------------counted_loop----------------------------------- 2420 // Convert to counted loops where possible 2421 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { 2422 2423 // For grins, set the inner-loop flag here 2424 if (!_child) { 2425 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop(); 2426 } 2427 2428 IdealLoopTree* loop = this; 2429 if (_head->is_CountedLoop() || 2430 phase->is_counted_loop(_head, loop)) { 2431 2432 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) { 2433 // Indicate we do not need a safepoint here 2434 _has_sfpt = 1; 2435 } 2436 2437 // Remove safepoints 2438 bool keep_one_sfpt = !(_has_call || _has_sfpt); 2439 remove_safepoints(phase, keep_one_sfpt); 2440 2441 // Look for induction variables 2442 phase->replace_parallel_iv(this); 2443 2444 } else if (_parent != NULL && !_irreducible) { 2445 // Not a counted loop. Keep one safepoint. 2446 bool keep_one_sfpt = true; 2447 remove_safepoints(phase, keep_one_sfpt); 2448 } 2449 2450 // Recursively 2451 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?"); 2452 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops"); 2453 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase); 2454 if (loop->_next) loop->_next ->counted_loop(phase); 2455 } 2456 2457 2458 // The Estimated Loop Clone Size: 2459 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm, 2460 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants, 2461 // respectively, used to ensure that the node usage estimates made are on the 2462 // safe side, for the most part. The FanOutTerm is an attempt to estimate the 2463 // possible additional/excessive nodes generated due to data and control flow 2464 // merging, for edges reaching outside the loop. 2465 uint IdealLoopTree::est_loop_clone_sz(uint factor) const { 2466 2467 precond(0 < factor && factor < 16); 2468 2469 uint const bc = 13; 2470 uint const cc = 17; 2471 uint const sz = _body.size() + (_body.size() + 7) / 8; 2472 uint estimate = factor * (sz + bc) + cc; 2473 2474 assert((estimate - cc) / factor == sz + bc, "overflow"); 2475 2476 return estimate + est_loop_flow_merge_sz(); 2477 } 2478 2479 // The Estimated Loop (full-) Unroll Size: 2480 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm, 2481 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that 2482 // node usage estimates made are on the safe side, for the most part. This is 2483 // a "light" version of the loop clone size calculation (above), based on the 2484 // assumption that most of the loop-construct overhead will be unraveled when 2485 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1), 2486 // including an overflow check and returning UINT_MAX in case of an overflow. 2487 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const { 2488 2489 precond(factor > 0); 2490 2491 // Take into account that after unroll conjoined heads and tails will fold. 2492 uint const b0 = _body.size() - EMPTY_LOOP_SIZE; 2493 uint const cc = 7; 2494 uint const sz = b0 + (b0 + 15) / 16; 2495 uint estimate = factor * sz + cc; 2496 2497 if ((estimate - cc) / factor != sz) { 2498 return UINT_MAX; 2499 } 2500 2501 return estimate + est_loop_flow_merge_sz(); 2502 } 2503 2504 // Estimate the growth effect (in nodes) of merging control and data flow when 2505 // cloning a loop body, based on the amount of control and data flow reaching 2506 // outside of the (current) loop body. 2507 uint IdealLoopTree::est_loop_flow_merge_sz() const { 2508 2509 uint ctrl_edge_out_cnt = 0; 2510 uint data_edge_out_cnt = 0; 2511 2512 for (uint i = 0; i < _body.size(); i++) { 2513 Node* node = _body.at(i); 2514 uint outcnt = node->outcnt(); 2515 2516 for (uint k = 0; k < outcnt; k++) { 2517 Node* out = node->raw_out(k); 2518 if (out == NULL) continue; 2519 if (out->is_CFG()) { 2520 if (!is_member(_phase->get_loop(out))) { 2521 ctrl_edge_out_cnt++; 2522 } 2523 } else if (_phase->has_ctrl(out)) { 2524 Node* ctrl = _phase->get_ctrl(out); 2525 assert(ctrl != NULL, "must be"); 2526 assert(ctrl->is_CFG(), "must be"); 2527 if (!is_member(_phase->get_loop(ctrl))) { 2528 data_edge_out_cnt++; 2529 } 2530 } 2531 } 2532 } 2533 // Use data and control count (x2.0) in estimate iff both are > 0. This is 2534 // a rather pessimistic estimate for the most part, in particular for some 2535 // complex loops, but still not enough to capture all loops. 2536 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) { 2537 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt); 2538 } 2539 return 0; 2540 } 2541 2542 #ifndef PRODUCT 2543 //------------------------------dump_head-------------------------------------- 2544 // Dump 1 liner for loop header info 2545 void IdealLoopTree::dump_head() const { 2546 tty->sp(2 * _nest); 2547 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx); 2548 if (_irreducible) tty->print(" IRREDUCIBLE"); 2549 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl); 2550 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 2551 if (predicate != NULL ) { 2552 tty->print(" limit_check"); 2553 entry = PhaseIdealLoop::skip_loop_predicates(entry); 2554 } 2555 if (UseProfiledLoopPredicate) { 2556 predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate); 2557 if (predicate != NULL) { 2558 tty->print(" profile_predicated"); 2559 entry = PhaseIdealLoop::skip_loop_predicates(entry); 2560 } 2561 } 2562 if (UseLoopPredicate) { 2563 predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 2564 if (predicate != NULL) { 2565 tty->print(" predicated"); 2566 } 2567 } 2568 if (_head->is_CountedLoop()) { 2569 CountedLoopNode *cl = _head->as_CountedLoop(); 2570 tty->print(" counted"); 2571 2572 Node* init_n = cl->init_trip(); 2573 if (init_n != NULL && init_n->is_Con()) 2574 tty->print(" [%d,", cl->init_trip()->get_int()); 2575 else 2576 tty->print(" [int,"); 2577 Node* limit_n = cl->limit(); 2578 if (limit_n != NULL && limit_n->is_Con()) 2579 tty->print("%d),", cl->limit()->get_int()); 2580 else 2581 tty->print("int),"); 2582 int stride_con = cl->stride_con(); 2583 if (stride_con > 0) tty->print("+"); 2584 tty->print("%d", stride_con); 2585 2586 tty->print(" (%0.f iters) ", cl->profile_trip_cnt()); 2587 2588 if (cl->is_pre_loop ()) tty->print(" pre" ); 2589 if (cl->is_main_loop()) tty->print(" main"); 2590 if (cl->is_post_loop()) tty->print(" post"); 2591 if (cl->is_vectorized_loop()) tty->print(" vector"); 2592 if (cl->range_checks_present()) tty->print(" rc "); 2593 if (cl->is_multiversioned()) tty->print(" multi "); 2594 } 2595 if (_has_call) tty->print(" has_call"); 2596 if (_has_sfpt) tty->print(" has_sfpt"); 2597 if (_rce_candidate) tty->print(" rce"); 2598 if (_safepts != NULL && _safepts->size() > 0) { 2599 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }"); 2600 } 2601 if (_required_safept != NULL && _required_safept->size() > 0) { 2602 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }"); 2603 } 2604 if (Verbose) { 2605 tty->print(" body={"); _body.dump_simple(); tty->print(" }"); 2606 } 2607 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) { 2608 tty->print(" strip_mined"); 2609 } 2610 tty->cr(); 2611 } 2612 2613 //------------------------------dump------------------------------------------- 2614 // Dump loops by loop tree 2615 void IdealLoopTree::dump() const { 2616 dump_head(); 2617 if (_child) _child->dump(); 2618 if (_next) _next ->dump(); 2619 } 2620 2621 #endif 2622 2623 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { 2624 if (loop == root) { 2625 if (loop->_child != NULL) { 2626 log->begin_head("loop_tree"); 2627 log->end_head(); 2628 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2629 log->tail("loop_tree"); 2630 assert(loop->_next == NULL, "what?"); 2631 } 2632 } else { 2633 Node* head = loop->_head; 2634 log->begin_head("loop"); 2635 log->print(" idx='%d' ", head->_idx); 2636 if (loop->_irreducible) log->print("irreducible='1' "); 2637 if (head->is_Loop()) { 2638 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); 2639 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); 2640 } 2641 if (head->is_CountedLoop()) { 2642 CountedLoopNode* cl = head->as_CountedLoop(); 2643 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); 2644 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); 2645 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); 2646 } 2647 log->end_head(); 2648 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2649 log->tail("loop"); 2650 if( loop->_next ) log_loop_tree(root, loop->_next, log); 2651 } 2652 } 2653 2654 //---------------------collect_potentially_useful_predicates----------------------- 2655 // Helper function to collect potentially useful predicates to prevent them from 2656 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates 2657 void PhaseIdealLoop::collect_potentially_useful_predicates( 2658 IdealLoopTree * loop, Unique_Node_List &useful_predicates) { 2659 if (loop->_child) { // child 2660 collect_potentially_useful_predicates(loop->_child, useful_predicates); 2661 } 2662 2663 // self (only loops that we can apply loop predication may use their predicates) 2664 if (loop->_head->is_Loop() && 2665 !loop->_irreducible && 2666 !loop->tail()->is_top()) { 2667 LoopNode* lpn = loop->_head->as_Loop(); 2668 Node* entry = lpn->in(LoopNode::EntryControl); 2669 Node* predicate_proj = find_predicate(entry); // loop_limit_check first 2670 if (predicate_proj != NULL) { // right pattern that can be used by loop predication 2671 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be"); 2672 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2673 entry = skip_loop_predicates(entry); 2674 } 2675 if (UseProfiledLoopPredicate) { 2676 predicate_proj = find_predicate(entry); // Predicate 2677 if (predicate_proj != NULL) { 2678 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2679 entry = skip_loop_predicates(entry); 2680 } 2681 } 2682 predicate_proj = find_predicate(entry); // Predicate 2683 if (predicate_proj != NULL) { 2684 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2685 } 2686 } 2687 2688 if (loop->_next) { // sibling 2689 collect_potentially_useful_predicates(loop->_next, useful_predicates); 2690 } 2691 } 2692 2693 //------------------------eliminate_useless_predicates----------------------------- 2694 // Eliminate all inserted predicates if they could not be used by loop predication. 2695 // Note: it will also eliminates loop limits check predicate since it also uses 2696 // Opaque1 node (see Parse::add_predicate()). 2697 void PhaseIdealLoop::eliminate_useless_predicates() { 2698 if (C->predicate_count() == 0) 2699 return; // no predicate left 2700 2701 Unique_Node_List useful_predicates; // to store useful predicates 2702 if (C->has_loops()) { 2703 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); 2704 } 2705 2706 for (int i = C->predicate_count(); i > 0; i--) { 2707 Node * n = C->predicate_opaque1_node(i-1); 2708 assert(n->Opcode() == Op_Opaque1, "must be"); 2709 if (!useful_predicates.member(n)) { // not in the useful list 2710 _igvn.replace_node(n, n->in(1)); 2711 } 2712 } 2713 } 2714 2715 //------------------------process_expensive_nodes----------------------------- 2716 // Expensive nodes have their control input set to prevent the GVN 2717 // from commoning them and as a result forcing the resulting node to 2718 // be in a more frequent path. Use CFG information here, to change the 2719 // control inputs so that some expensive nodes can be commoned while 2720 // not executed more frequently. 2721 bool PhaseIdealLoop::process_expensive_nodes() { 2722 assert(OptimizeExpensiveOps, "optimization off?"); 2723 2724 // Sort nodes to bring similar nodes together 2725 C->sort_expensive_nodes(); 2726 2727 bool progress = false; 2728 2729 for (int i = 0; i < C->expensive_count(); ) { 2730 Node* n = C->expensive_node(i); 2731 int start = i; 2732 // Find nodes similar to n 2733 i++; 2734 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++); 2735 int end = i; 2736 // And compare them two by two 2737 for (int j = start; j < end; j++) { 2738 Node* n1 = C->expensive_node(j); 2739 if (is_node_unreachable(n1)) { 2740 continue; 2741 } 2742 for (int k = j+1; k < end; k++) { 2743 Node* n2 = C->expensive_node(k); 2744 if (is_node_unreachable(n2)) { 2745 continue; 2746 } 2747 2748 assert(n1 != n2, "should be pair of nodes"); 2749 2750 Node* c1 = n1->in(0); 2751 Node* c2 = n2->in(0); 2752 2753 Node* parent_c1 = c1; 2754 Node* parent_c2 = c2; 2755 2756 // The call to get_early_ctrl_for_expensive() moves the 2757 // expensive nodes up but stops at loops that are in a if 2758 // branch. See whether we can exit the loop and move above the 2759 // If. 2760 if (c1->is_Loop()) { 2761 parent_c1 = c1->in(1); 2762 } 2763 if (c2->is_Loop()) { 2764 parent_c2 = c2->in(1); 2765 } 2766 2767 if (parent_c1 == parent_c2) { 2768 _igvn._worklist.push(n1); 2769 _igvn._worklist.push(n2); 2770 continue; 2771 } 2772 2773 // Look for identical expensive node up the dominator chain. 2774 if (is_dominator(c1, c2)) { 2775 c2 = c1; 2776 } else if (is_dominator(c2, c1)) { 2777 c1 = c2; 2778 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() && 2779 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) { 2780 // Both branches have the same expensive node so move it up 2781 // before the if. 2782 c1 = c2 = idom(parent_c1->in(0)); 2783 } 2784 // Do the actual moves 2785 if (n1->in(0) != c1) { 2786 _igvn.hash_delete(n1); 2787 n1->set_req(0, c1); 2788 _igvn.hash_insert(n1); 2789 _igvn._worklist.push(n1); 2790 progress = true; 2791 } 2792 if (n2->in(0) != c2) { 2793 _igvn.hash_delete(n2); 2794 n2->set_req(0, c2); 2795 _igvn.hash_insert(n2); 2796 _igvn._worklist.push(n2); 2797 progress = true; 2798 } 2799 } 2800 } 2801 } 2802 2803 return progress; 2804 } 2805 2806 2807 //============================================================================= 2808 //----------------------------build_and_optimize------------------------------- 2809 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to 2810 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. 2811 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) { 2812 bool do_split_ifs = (mode == LoopOptsDefault); 2813 bool skip_loop_opts = (mode == LoopOptsNone); 2814 2815 int old_progress = C->major_progress(); 2816 uint orig_worklist_size = _igvn._worklist.size(); 2817 2818 // Reset major-progress flag for the driver's heuristics 2819 C->clear_major_progress(); 2820 2821 #ifndef PRODUCT 2822 // Capture for later assert 2823 uint unique = C->unique(); 2824 _loop_invokes++; 2825 _loop_work += unique; 2826 #endif 2827 2828 // True if the method has at least 1 irreducible loop 2829 _has_irreducible_loops = false; 2830 2831 _created_loop_node = false; 2832 2833 Arena *a = Thread::current()->resource_area(); 2834 VectorSet visited(a); 2835 // Pre-grow the mapping from Nodes to IdealLoopTrees. 2836 _nodes.map(C->unique(), NULL); 2837 memset(_nodes.adr(), 0, wordSize * C->unique()); 2838 2839 // Pre-build the top-level outermost loop tree entry 2840 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); 2841 // Do not need a safepoint at the top level 2842 _ltree_root->_has_sfpt = 1; 2843 2844 // Initialize Dominators. 2845 // Checked in clone_loop_predicate() during beautify_loops(). 2846 _idom_size = 0; 2847 _idom = NULL; 2848 _dom_depth = NULL; 2849 _dom_stk = NULL; 2850 2851 // Empty pre-order array 2852 allocate_preorders(); 2853 2854 // Build a loop tree on the fly. Build a mapping from CFG nodes to 2855 // IdealLoopTree entries. Data nodes are NOT walked. 2856 build_loop_tree(); 2857 // Check for bailout, and return 2858 if (C->failing()) { 2859 return; 2860 } 2861 2862 // No loops after all 2863 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); 2864 2865 // There should always be an outer loop containing the Root and Return nodes. 2866 // If not, we have a degenerate empty program. Bail out in this case. 2867 if (!has_node(C->root())) { 2868 if (!_verify_only) { 2869 C->clear_major_progress(); 2870 C->record_method_not_compilable("empty program detected during loop optimization"); 2871 } 2872 return; 2873 } 2874 2875 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2876 // Nothing to do, so get out 2877 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only && 2878 !bs->is_gc_specific_loop_opts_pass(mode); 2879 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn); 2880 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(mode); 2881 if (stop_early && !do_expensive_nodes) { 2882 _igvn.optimize(); // Cleanup NeverBranches 2883 return; 2884 } 2885 2886 // Set loop nesting depth 2887 _ltree_root->set_nest( 0 ); 2888 2889 // Split shared headers and insert loop landing pads. 2890 // Do not bother doing this on the Root loop of course. 2891 if( !_verify_me && !_verify_only && _ltree_root->_child ) { 2892 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3); 2893 if( _ltree_root->_child->beautify_loops( this ) ) { 2894 // Re-build loop tree! 2895 _ltree_root->_child = NULL; 2896 _nodes.clear(); 2897 reallocate_preorders(); 2898 build_loop_tree(); 2899 // Check for bailout, and return 2900 if (C->failing()) { 2901 return; 2902 } 2903 // Reset loop nesting depth 2904 _ltree_root->set_nest( 0 ); 2905 2906 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3); 2907 } 2908 } 2909 2910 // Build Dominators for elision of NULL checks & loop finding. 2911 // Since nodes do not have a slot for immediate dominator, make 2912 // a persistent side array for that info indexed on node->_idx. 2913 _idom_size = C->unique(); 2914 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); 2915 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); 2916 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth 2917 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); 2918 2919 Dominators(); 2920 2921 if (!_verify_only) { 2922 // As a side effect, Dominators removed any unreachable CFG paths 2923 // into RegionNodes. It doesn't do this test against Root, so 2924 // we do it here. 2925 for( uint i = 1; i < C->root()->req(); i++ ) { 2926 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? 2927 _igvn.delete_input_of(C->root(), i); 2928 i--; // Rerun same iteration on compressed edges 2929 } 2930 } 2931 2932 // Given dominators, try to find inner loops with calls that must 2933 // always be executed (call dominates loop tail). These loops do 2934 // not need a separate safepoint. 2935 Node_List cisstack(a); 2936 _ltree_root->check_safepts(visited, cisstack); 2937 } 2938 2939 // Walk the DATA nodes and place into loops. Find earliest control 2940 // node. For CFG nodes, the _nodes array starts out and remains 2941 // holding the associated IdealLoopTree pointer. For DATA nodes, the 2942 // _nodes array holds the earliest legal controlling CFG node. 2943 2944 // Allocate stack with enough space to avoid frequent realloc 2945 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats 2946 Node_Stack nstack( a, stack_size ); 2947 2948 visited.clear(); 2949 Node_List worklist(a); 2950 // Don't need C->root() on worklist since 2951 // it will be processed among C->top() inputs 2952 worklist.push(C->top()); 2953 visited.set(C->top()->_idx); // Set C->top() as visited now 2954 build_loop_early( visited, worklist, nstack ); 2955 2956 // Given early legal placement, try finding counted loops. This placement 2957 // is good enough to discover most loop invariants. 2958 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) { 2959 _ltree_root->counted_loop( this ); 2960 } 2961 2962 // Find latest loop placement. Find ideal loop placement. 2963 visited.clear(); 2964 init_dom_lca_tags(); 2965 // Need C->root() on worklist when processing outs 2966 worklist.push(C->root()); 2967 NOT_PRODUCT( C->verify_graph_edges(); ) 2968 worklist.push(C->top()); 2969 build_loop_late( visited, worklist, nstack ); 2970 2971 if (_verify_only) { 2972 C->restore_major_progress(old_progress); 2973 assert(C->unique() == unique, "verification mode made Nodes? ? ?"); 2974 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything"); 2975 return; 2976 } 2977 2978 // clear out the dead code after build_loop_late 2979 while (_deadlist.size()) { 2980 _igvn.remove_globally_dead_node(_deadlist.pop()); 2981 } 2982 2983 if (stop_early) { 2984 assert(do_expensive_nodes, "why are we here?"); 2985 if (process_expensive_nodes()) { 2986 // If we made some progress when processing expensive nodes then 2987 // the IGVN may modify the graph in a way that will allow us to 2988 // make some more progress: we need to try processing expensive 2989 // nodes again. 2990 C->set_major_progress(); 2991 } 2992 _igvn.optimize(); 2993 return; 2994 } 2995 2996 // Some parser-inserted loop predicates could never be used by loop 2997 // predication or they were moved away from loop during some optimizations. 2998 // For example, peeling. Eliminate them before next loop optimizations. 2999 eliminate_useless_predicates(); 3000 3001 #ifndef PRODUCT 3002 C->verify_graph_edges(); 3003 if (_verify_me) { // Nested verify pass? 3004 // Check to see if the verify mode is broken 3005 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); 3006 return; 3007 } 3008 if (VerifyLoopOptimizations) verify(); 3009 if (TraceLoopOpts && C->has_loops()) { 3010 _ltree_root->dump(); 3011 } 3012 #endif 3013 3014 if (skip_loop_opts) { 3015 // restore major progress flag 3016 C->restore_major_progress(old_progress); 3017 3018 // Cleanup any modified bits 3019 _igvn.optimize(); 3020 3021 if (C->log() != NULL) { 3022 log_loop_tree(_ltree_root, _ltree_root, C->log()); 3023 } 3024 return; 3025 } 3026 3027 if (mode == LoopOptsMaxUnroll) { 3028 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 3029 IdealLoopTree* lpt = iter.current(); 3030 if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) { 3031 lpt->compute_trip_count(this); 3032 if (!lpt->do_one_iteration_loop(this) && 3033 !lpt->do_remove_empty_loop(this)) { 3034 AutoNodeBudget node_budget(this); 3035 if (lpt->_head->as_CountedLoop()->is_normal_loop() && 3036 lpt->policy_maximally_unroll(this)) { 3037 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 3038 do_maximally_unroll(lpt, worklist); 3039 } 3040 } 3041 } 3042 } 3043 3044 C->restore_major_progress(old_progress); 3045 3046 _igvn.optimize(); 3047 3048 if (C->log() != NULL) { 3049 log_loop_tree(_ltree_root, _ltree_root, C->log()); 3050 } 3051 return; 3052 } 3053 3054 if (bs->optimize_loops(this, mode, visited, nstack, worklist)) { 3055 _igvn.optimize(); 3056 if (C->log() != NULL) { 3057 log_loop_tree(_ltree_root, _ltree_root, C->log()); 3058 } 3059 return; 3060 } 3061 3062 if (ReassociateInvariants) { 3063 // Reassociate invariants and prep for split_thru_phi 3064 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 3065 IdealLoopTree* lpt = iter.current(); 3066 bool is_counted = lpt->is_counted(); 3067 if (!is_counted || !lpt->is_innermost()) continue; 3068 3069 // check for vectorized loops, any reassociation of invariants was already done 3070 if (is_counted && lpt->_head->as_CountedLoop()->is_unroll_only()) { 3071 continue; 3072 } else { 3073 AutoNodeBudget node_budget(this); 3074 lpt->reassociate_invariants(this); 3075 } 3076 // Because RCE opportunities can be masked by split_thru_phi, 3077 // look for RCE candidates and inhibit split_thru_phi 3078 // on just their loop-phi's for this pass of loop opts 3079 if (SplitIfBlocks && do_split_ifs) { 3080 AutoNodeBudget node_budget(this, AutoNodeBudget::NO_BUDGET_CHECK); 3081 if (lpt->policy_range_check(this)) { 3082 lpt->_rce_candidate = 1; // = true 3083 } 3084 } 3085 } 3086 } 3087 3088 // Check for aggressive application of split-if and other transforms 3089 // that require basic-block info (like cloning through Phi's) 3090 if( SplitIfBlocks && do_split_ifs ) { 3091 visited.clear(); 3092 split_if_with_blocks( visited, nstack); 3093 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); 3094 } 3095 3096 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) { 3097 C->set_major_progress(); 3098 } 3099 3100 // Perform loop predication before iteration splitting 3101 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) { 3102 _ltree_root->_child->loop_predication(this); 3103 } 3104 3105 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { 3106 if (do_intrinsify_fill()) { 3107 C->set_major_progress(); 3108 } 3109 } 3110 3111 // Perform iteration-splitting on inner loops. Split iterations to avoid 3112 // range checks or one-shot null checks. 3113 3114 // If split-if's didn't hack the graph too bad (no CFG changes) 3115 // then do loop opts. 3116 if (C->has_loops() && !C->major_progress()) { 3117 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 3118 _ltree_root->_child->iteration_split( this, worklist ); 3119 // No verify after peeling! GCM has hoisted code out of the loop. 3120 // After peeling, the hoisted code could sink inside the peeled area. 3121 // The peeling code does not try to recompute the best location for 3122 // all the code before the peeled area, so the verify pass will always 3123 // complain about it. 3124 } 3125 // Do verify graph edges in any case 3126 NOT_PRODUCT( C->verify_graph_edges(); ); 3127 3128 if (!do_split_ifs) { 3129 // We saw major progress in Split-If to get here. We forced a 3130 // pass with unrolling and not split-if, however more split-if's 3131 // might make progress. If the unrolling didn't make progress 3132 // then the major-progress flag got cleared and we won't try 3133 // another round of Split-If. In particular the ever-common 3134 // instance-of/check-cast pattern requires at least 2 rounds of 3135 // Split-If to clear out. 3136 C->set_major_progress(); 3137 } 3138 3139 // Repeat loop optimizations if new loops were seen 3140 if (created_loop_node()) { 3141 C->set_major_progress(); 3142 } 3143 3144 // Keep loop predicates and perform optimizations with them 3145 // until no more loop optimizations could be done. 3146 // After that switch predicates off and do more loop optimizations. 3147 if (!C->major_progress() && (C->predicate_count() > 0)) { 3148 C->cleanup_loop_predicates(_igvn); 3149 if (TraceLoopOpts) { 3150 tty->print_cr("PredicatesOff"); 3151 } 3152 C->set_major_progress(); 3153 } 3154 3155 // Convert scalar to superword operations at the end of all loop opts. 3156 if (UseSuperWord && C->has_loops() && !C->major_progress()) { 3157 // SuperWord transform 3158 SuperWord sw(this); 3159 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 3160 IdealLoopTree* lpt = iter.current(); 3161 if (lpt->is_counted()) { 3162 CountedLoopNode *cl = lpt->_head->as_CountedLoop(); 3163 3164 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) { 3165 // Check that the rce'd post loop is encountered first, multiversion after all 3166 // major main loop optimization are concluded 3167 if (!C->major_progress()) { 3168 IdealLoopTree *lpt_next = lpt->_next; 3169 if (lpt_next && lpt_next->is_counted()) { 3170 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop(); 3171 has_range_checks(lpt_next); 3172 if (cl->is_post_loop() && cl->range_checks_present()) { 3173 if (!cl->is_multiversioned()) { 3174 if (multi_version_post_loops(lpt, lpt_next) == false) { 3175 // Cause the rce loop to be optimized away if we fail 3176 cl->mark_is_multiversioned(); 3177 cl->set_slp_max_unroll(0); 3178 poison_rce_post_loop(lpt); 3179 } 3180 } 3181 } 3182 } 3183 sw.transform_loop(lpt, true); 3184 } 3185 } else if (cl->is_main_loop()) { 3186 sw.transform_loop(lpt, true); 3187 } 3188 } 3189 } 3190 } 3191 3192 // Cleanup any modified bits 3193 _igvn.optimize(); 3194 3195 // disable assert until issue with split_flow_path is resolved (6742111) 3196 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), 3197 // "shouldn't introduce irreducible loops"); 3198 3199 if (C->log() != NULL) { 3200 log_loop_tree(_ltree_root, _ltree_root, C->log()); 3201 } 3202 } 3203 3204 #ifndef PRODUCT 3205 //------------------------------print_statistics------------------------------- 3206 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes 3207 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique 3208 void PhaseIdealLoop::print_statistics() { 3209 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); 3210 } 3211 3212 //------------------------------verify----------------------------------------- 3213 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. 3214 static int fail; // debug only, so its multi-thread dont care 3215 void PhaseIdealLoop::verify() const { 3216 int old_progress = C->major_progress(); 3217 ResourceMark rm; 3218 PhaseIdealLoop loop_verify( _igvn, this ); 3219 VectorSet visited(Thread::current()->resource_area()); 3220 3221 fail = 0; 3222 verify_compare( C->root(), &loop_verify, visited ); 3223 assert( fail == 0, "verify loops failed" ); 3224 // Verify loop structure is the same 3225 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); 3226 // Reset major-progress. It was cleared by creating a verify version of 3227 // PhaseIdealLoop. 3228 C->restore_major_progress(old_progress); 3229 } 3230 3231 //------------------------------verify_compare--------------------------------- 3232 // Make sure me and the given PhaseIdealLoop agree on key data structures 3233 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { 3234 if( !n ) return; 3235 if( visited.test_set( n->_idx ) ) return; 3236 if( !_nodes[n->_idx] ) { // Unreachable 3237 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); 3238 return; 3239 } 3240 3241 uint i; 3242 for( i = 0; i < n->req(); i++ ) 3243 verify_compare( n->in(i), loop_verify, visited ); 3244 3245 // Check the '_nodes' block/loop structure 3246 i = n->_idx; 3247 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl 3248 if( _nodes[i] != loop_verify->_nodes[i] && 3249 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { 3250 tty->print("Mismatched control setting for: "); 3251 n->dump(); 3252 if( fail++ > 10 ) return; 3253 Node *c = get_ctrl_no_update(n); 3254 tty->print("We have it as: "); 3255 if( c->in(0) ) c->dump(); 3256 else tty->print_cr("N%d",c->_idx); 3257 tty->print("Verify thinks: "); 3258 if( loop_verify->has_ctrl(n) ) 3259 loop_verify->get_ctrl_no_update(n)->dump(); 3260 else 3261 loop_verify->get_loop_idx(n)->dump(); 3262 tty->cr(); 3263 } 3264 } else { // We have a loop 3265 IdealLoopTree *us = get_loop_idx(n); 3266 if( loop_verify->has_ctrl(n) ) { 3267 tty->print("Mismatched loop setting for: "); 3268 n->dump(); 3269 if( fail++ > 10 ) return; 3270 tty->print("We have it as: "); 3271 us->dump(); 3272 tty->print("Verify thinks: "); 3273 loop_verify->get_ctrl_no_update(n)->dump(); 3274 tty->cr(); 3275 } else if (!C->major_progress()) { 3276 // Loop selection can be messed up if we did a major progress 3277 // operation, like split-if. Do not verify in that case. 3278 IdealLoopTree *them = loop_verify->get_loop_idx(n); 3279 if( us->_head != them->_head || us->_tail != them->_tail ) { 3280 tty->print("Unequals loops for: "); 3281 n->dump(); 3282 if( fail++ > 10 ) return; 3283 tty->print("We have it as: "); 3284 us->dump(); 3285 tty->print("Verify thinks: "); 3286 them->dump(); 3287 tty->cr(); 3288 } 3289 } 3290 } 3291 3292 // Check for immediate dominators being equal 3293 if( i >= _idom_size ) { 3294 if( !n->is_CFG() ) return; 3295 tty->print("CFG Node with no idom: "); 3296 n->dump(); 3297 return; 3298 } 3299 if( !n->is_CFG() ) return; 3300 if( n == C->root() ) return; // No IDOM here 3301 3302 assert(n->_idx == i, "sanity"); 3303 Node *id = idom_no_update(n); 3304 if( id != loop_verify->idom_no_update(n) ) { 3305 tty->print("Unequals idoms for: "); 3306 n->dump(); 3307 if( fail++ > 10 ) return; 3308 tty->print("We have it as: "); 3309 id->dump(); 3310 tty->print("Verify thinks: "); 3311 loop_verify->idom_no_update(n)->dump(); 3312 tty->cr(); 3313 } 3314 3315 } 3316 3317 //------------------------------verify_tree------------------------------------ 3318 // Verify that tree structures match. Because the CFG can change, siblings 3319 // within the loop tree can be reordered. We attempt to deal with that by 3320 // reordering the verify's loop tree if possible. 3321 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { 3322 assert( _parent == parent, "Badly formed loop tree" ); 3323 3324 // Siblings not in same order? Attempt to re-order. 3325 if( _head != loop->_head ) { 3326 // Find _next pointer to update 3327 IdealLoopTree **pp = &loop->_parent->_child; 3328 while( *pp != loop ) 3329 pp = &((*pp)->_next); 3330 // Find proper sibling to be next 3331 IdealLoopTree **nn = &loop->_next; 3332 while( (*nn) && (*nn)->_head != _head ) 3333 nn = &((*nn)->_next); 3334 3335 // Check for no match. 3336 if( !(*nn) ) { 3337 // Annoyingly, irreducible loops can pick different headers 3338 // after a major_progress operation, so the rest of the loop 3339 // tree cannot be matched. 3340 if (_irreducible && Compile::current()->major_progress()) return; 3341 assert( 0, "failed to match loop tree" ); 3342 } 3343 3344 // Move (*nn) to (*pp) 3345 IdealLoopTree *hit = *nn; 3346 *nn = hit->_next; 3347 hit->_next = loop; 3348 *pp = loop; 3349 loop = hit; 3350 // Now try again to verify 3351 } 3352 3353 assert( _head == loop->_head , "mismatched loop head" ); 3354 Node *tail = _tail; // Inline a non-updating version of 3355 while( !tail->in(0) ) // the 'tail()' call. 3356 tail = tail->in(1); 3357 assert( tail == loop->_tail, "mismatched loop tail" ); 3358 3359 // Counted loops that are guarded should be able to find their guards 3360 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { 3361 CountedLoopNode *cl = _head->as_CountedLoop(); 3362 Node *init = cl->init_trip(); 3363 Node *ctrl = cl->in(LoopNode::EntryControl); 3364 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 3365 Node *iff = ctrl->in(0); 3366 assert( iff->Opcode() == Op_If, "" ); 3367 Node *bol = iff->in(1); 3368 assert( bol->Opcode() == Op_Bool, "" ); 3369 Node *cmp = bol->in(1); 3370 assert( cmp->Opcode() == Op_CmpI, "" ); 3371 Node *add = cmp->in(1); 3372 Node *opaq; 3373 if( add->Opcode() == Op_Opaque1 ) { 3374 opaq = add; 3375 } else { 3376 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); 3377 assert( add == init, "" ); 3378 opaq = cmp->in(2); 3379 } 3380 assert( opaq->Opcode() == Op_Opaque1, "" ); 3381 3382 } 3383 3384 if (_child != NULL) _child->verify_tree(loop->_child, this); 3385 if (_next != NULL) _next ->verify_tree(loop->_next, parent); 3386 // Innermost loops need to verify loop bodies, 3387 // but only if no 'major_progress' 3388 int fail = 0; 3389 if (!Compile::current()->major_progress() && _child == NULL) { 3390 for( uint i = 0; i < _body.size(); i++ ) { 3391 Node *n = _body.at(i); 3392 if (n->outcnt() == 0) continue; // Ignore dead 3393 uint j; 3394 for( j = 0; j < loop->_body.size(); j++ ) 3395 if( loop->_body.at(j) == n ) 3396 break; 3397 if( j == loop->_body.size() ) { // Not found in loop body 3398 // Last ditch effort to avoid assertion: Its possible that we 3399 // have some users (so outcnt not zero) but are still dead. 3400 // Try to find from root. 3401 if (Compile::current()->root()->find(n->_idx)) { 3402 fail++; 3403 tty->print("We have that verify does not: "); 3404 n->dump(); 3405 } 3406 } 3407 } 3408 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { 3409 Node *n = loop->_body.at(i2); 3410 if (n->outcnt() == 0) continue; // Ignore dead 3411 uint j; 3412 for( j = 0; j < _body.size(); j++ ) 3413 if( _body.at(j) == n ) 3414 break; 3415 if( j == _body.size() ) { // Not found in loop body 3416 // Last ditch effort to avoid assertion: Its possible that we 3417 // have some users (so outcnt not zero) but are still dead. 3418 // Try to find from root. 3419 if (Compile::current()->root()->find(n->_idx)) { 3420 fail++; 3421 tty->print("Verify has that we do not: "); 3422 n->dump(); 3423 } 3424 } 3425 } 3426 assert( !fail, "loop body mismatch" ); 3427 } 3428 } 3429 3430 #endif 3431 3432 //------------------------------set_idom--------------------------------------- 3433 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { 3434 uint idx = d->_idx; 3435 if (idx >= _idom_size) { 3436 uint newsize = next_power_of_2(idx); 3437 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); 3438 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); 3439 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); 3440 _idom_size = newsize; 3441 } 3442 _idom[idx] = n; 3443 _dom_depth[idx] = dom_depth; 3444 } 3445 3446 //------------------------------recompute_dom_depth--------------------------------------- 3447 // The dominator tree is constructed with only parent pointers. 3448 // This recomputes the depth in the tree by first tagging all 3449 // nodes as "no depth yet" marker. The next pass then runs up 3450 // the dom tree from each node marked "no depth yet", and computes 3451 // the depth on the way back down. 3452 void PhaseIdealLoop::recompute_dom_depth() { 3453 uint no_depth_marker = C->unique(); 3454 uint i; 3455 // Initialize depth to "no depth yet" and realize all lazy updates 3456 for (i = 0; i < _idom_size; i++) { 3457 // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized). 3458 if (_dom_depth[i] > 0 && _idom[i] != NULL) { 3459 _dom_depth[i] = no_depth_marker; 3460 3461 // heal _idom if it has a fwd mapping in _nodes 3462 if (_idom[i]->in(0) == NULL) { 3463 idom(i); 3464 } 3465 } 3466 } 3467 if (_dom_stk == NULL) { 3468 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size. 3469 if (init_size < 10) init_size = 10; 3470 _dom_stk = new GrowableArray<uint>(init_size); 3471 } 3472 // Compute new depth for each node. 3473 for (i = 0; i < _idom_size; i++) { 3474 uint j = i; 3475 // Run up the dom tree to find a node with a depth 3476 while (_dom_depth[j] == no_depth_marker) { 3477 _dom_stk->push(j); 3478 j = _idom[j]->_idx; 3479 } 3480 // Compute the depth on the way back down this tree branch 3481 uint dd = _dom_depth[j] + 1; 3482 while (_dom_stk->length() > 0) { 3483 uint j = _dom_stk->pop(); 3484 _dom_depth[j] = dd; 3485 dd++; 3486 } 3487 } 3488 } 3489 3490 //------------------------------sort------------------------------------------- 3491 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the 3492 // loop tree, not the root. 3493 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { 3494 if( !innermost ) return loop; // New innermost loop 3495 3496 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number 3497 assert( loop_preorder, "not yet post-walked loop" ); 3498 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer 3499 IdealLoopTree *l = *pp; // Do I go before or after 'l'? 3500 3501 // Insert at start of list 3502 while( l ) { // Insertion sort based on pre-order 3503 if( l == loop ) return innermost; // Already on list! 3504 int l_preorder = get_preorder(l->_head); // Cache pre-order number 3505 assert( l_preorder, "not yet post-walked l" ); 3506 // Check header pre-order number to figure proper nesting 3507 if( loop_preorder > l_preorder ) 3508 break; // End of insertion 3509 // If headers tie (e.g., shared headers) check tail pre-order numbers. 3510 // Since I split shared headers, you'd think this could not happen. 3511 // BUT: I must first do the preorder numbering before I can discover I 3512 // have shared headers, so the split headers all get the same preorder 3513 // number as the RegionNode they split from. 3514 if( loop_preorder == l_preorder && 3515 get_preorder(loop->_tail) < get_preorder(l->_tail) ) 3516 break; // Also check for shared headers (same pre#) 3517 pp = &l->_parent; // Chain up list 3518 l = *pp; 3519 } 3520 // Link into list 3521 // Point predecessor to me 3522 *pp = loop; 3523 // Point me to successor 3524 IdealLoopTree *p = loop->_parent; 3525 loop->_parent = l; // Point me to successor 3526 if( p ) sort( p, innermost ); // Insert my parents into list as well 3527 return innermost; 3528 } 3529 3530 //------------------------------build_loop_tree-------------------------------- 3531 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit 3532 // bits. The _nodes[] array is mapped by Node index and holds a NULL for 3533 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the 3534 // tightest enclosing IdealLoopTree for post-walked. 3535 // 3536 // During my forward walk I do a short 1-layer lookahead to see if I can find 3537 // a loop backedge with that doesn't have any work on the backedge. This 3538 // helps me construct nested loops with shared headers better. 3539 // 3540 // Once I've done the forward recursion, I do the post-work. For each child 3541 // I check to see if there is a backedge. Backedges define a loop! I 3542 // insert an IdealLoopTree at the target of the backedge. 3543 // 3544 // During the post-work I also check to see if I have several children 3545 // belonging to different loops. If so, then this Node is a decision point 3546 // where control flow can choose to change loop nests. It is at this 3547 // decision point where I can figure out how loops are nested. At this 3548 // time I can properly order the different loop nests from my children. 3549 // Note that there may not be any backedges at the decision point! 3550 // 3551 // Since the decision point can be far removed from the backedges, I can't 3552 // order my loops at the time I discover them. Thus at the decision point 3553 // I need to inspect loop header pre-order numbers to properly nest my 3554 // loops. This means I need to sort my childrens' loops by pre-order. 3555 // The sort is of size number-of-control-children, which generally limits 3556 // it to size 2 (i.e., I just choose between my 2 target loops). 3557 void PhaseIdealLoop::build_loop_tree() { 3558 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3559 GrowableArray <Node *> bltstack(C->live_nodes() >> 1); 3560 Node *n = C->root(); 3561 bltstack.push(n); 3562 int pre_order = 1; 3563 int stack_size; 3564 3565 while ( ( stack_size = bltstack.length() ) != 0 ) { 3566 n = bltstack.top(); // Leave node on stack 3567 if ( !is_visited(n) ) { 3568 // ---- Pre-pass Work ---- 3569 // Pre-walked but not post-walked nodes need a pre_order number. 3570 3571 set_preorder_visited( n, pre_order ); // set as visited 3572 3573 // ---- Scan over children ---- 3574 // Scan first over control projections that lead to loop headers. 3575 // This helps us find inner-to-outer loops with shared headers better. 3576 3577 // Scan children's children for loop headers. 3578 for ( int i = n->outcnt() - 1; i >= 0; --i ) { 3579 Node* m = n->raw_out(i); // Child 3580 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children 3581 // Scan over children's children to find loop 3582 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3583 Node* l = m->fast_out(j); 3584 if( is_visited(l) && // Been visited? 3585 !is_postvisited(l) && // But not post-visited 3586 get_preorder(l) < pre_order ) { // And smaller pre-order 3587 // Found! Scan the DFS down this path before doing other paths 3588 bltstack.push(m); 3589 break; 3590 } 3591 } 3592 } 3593 } 3594 pre_order++; 3595 } 3596 else if ( !is_postvisited(n) ) { 3597 // Note: build_loop_tree_impl() adds out edges on rare occasions, 3598 // such as com.sun.rsasign.am::a. 3599 // For non-recursive version, first, process current children. 3600 // On next iteration, check if additional children were added. 3601 for ( int k = n->outcnt() - 1; k >= 0; --k ) { 3602 Node* u = n->raw_out(k); 3603 if ( u->is_CFG() && !is_visited(u) ) { 3604 bltstack.push(u); 3605 } 3606 } 3607 if ( bltstack.length() == stack_size ) { 3608 // There were no additional children, post visit node now 3609 (void)bltstack.pop(); // Remove node from stack 3610 pre_order = build_loop_tree_impl( n, pre_order ); 3611 // Check for bailout 3612 if (C->failing()) { 3613 return; 3614 } 3615 // Check to grow _preorders[] array for the case when 3616 // build_loop_tree_impl() adds new nodes. 3617 check_grow_preorders(); 3618 } 3619 } 3620 else { 3621 (void)bltstack.pop(); // Remove post-visited node from stack 3622 } 3623 } 3624 } 3625 3626 //------------------------------build_loop_tree_impl--------------------------- 3627 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { 3628 // ---- Post-pass Work ---- 3629 // Pre-walked but not post-walked nodes need a pre_order number. 3630 3631 // Tightest enclosing loop for this Node 3632 IdealLoopTree *innermost = NULL; 3633 3634 // For all children, see if any edge is a backedge. If so, make a loop 3635 // for it. Then find the tightest enclosing loop for the self Node. 3636 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3637 Node* m = n->fast_out(i); // Child 3638 if( n == m ) continue; // Ignore control self-cycles 3639 if( !m->is_CFG() ) continue;// Ignore non-CFG edges 3640 3641 IdealLoopTree *l; // Child's loop 3642 if( !is_postvisited(m) ) { // Child visited but not post-visited? 3643 // Found a backedge 3644 assert( get_preorder(m) < pre_order, "should be backedge" ); 3645 // Check for the RootNode, which is already a LoopNode and is allowed 3646 // to have multiple "backedges". 3647 if( m == C->root()) { // Found the root? 3648 l = _ltree_root; // Root is the outermost LoopNode 3649 } else { // Else found a nested loop 3650 // Insert a LoopNode to mark this loop. 3651 l = new IdealLoopTree(this, m, n); 3652 } // End of Else found a nested loop 3653 if( !has_loop(m) ) // If 'm' does not already have a loop set 3654 set_loop(m, l); // Set loop header to loop now 3655 3656 } else { // Else not a nested loop 3657 if( !_nodes[m->_idx] ) continue; // Dead code has no loop 3658 l = get_loop(m); // Get previously determined loop 3659 // If successor is header of a loop (nest), move up-loop till it 3660 // is a member of some outer enclosing loop. Since there are no 3661 // shared headers (I've split them already) I only need to go up 3662 // at most 1 level. 3663 while( l && l->_head == m ) // Successor heads loop? 3664 l = l->_parent; // Move up 1 for me 3665 // If this loop is not properly parented, then this loop 3666 // has no exit path out, i.e. its an infinite loop. 3667 if( !l ) { 3668 // Make loop "reachable" from root so the CFG is reachable. Basically 3669 // insert a bogus loop exit that is never taken. 'm', the loop head, 3670 // points to 'n', one (of possibly many) fall-in paths. There may be 3671 // many backedges as well. 3672 3673 // Here I set the loop to be the root loop. I could have, after 3674 // inserting a bogus loop exit, restarted the recursion and found my 3675 // new loop exit. This would make the infinite loop a first-class 3676 // loop and it would then get properly optimized. What's the use of 3677 // optimizing an infinite loop? 3678 l = _ltree_root; // Oops, found infinite loop 3679 3680 if (!_verify_only) { 3681 // Insert the NeverBranch between 'm' and it's control user. 3682 NeverBranchNode *iff = new NeverBranchNode( m ); 3683 _igvn.register_new_node_with_optimizer(iff); 3684 set_loop(iff, l); 3685 Node *if_t = new CProjNode( iff, 0 ); 3686 _igvn.register_new_node_with_optimizer(if_t); 3687 set_loop(if_t, l); 3688 3689 Node* cfg = NULL; // Find the One True Control User of m 3690 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3691 Node* x = m->fast_out(j); 3692 if (x->is_CFG() && x != m && x != iff) 3693 { cfg = x; break; } 3694 } 3695 assert(cfg != NULL, "must find the control user of m"); 3696 uint k = 0; // Probably cfg->in(0) 3697 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region 3698 cfg->set_req( k, if_t ); // Now point to NeverBranch 3699 _igvn._worklist.push(cfg); 3700 3701 // Now create the never-taken loop exit 3702 Node *if_f = new CProjNode( iff, 1 ); 3703 _igvn.register_new_node_with_optimizer(if_f); 3704 set_loop(if_f, l); 3705 // Find frame ptr for Halt. Relies on the optimizer 3706 // V-N'ing. Easier and quicker than searching through 3707 // the program structure. 3708 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr ); 3709 _igvn.register_new_node_with_optimizer(frame); 3710 // Halt & Catch Fire 3711 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached"); 3712 _igvn.register_new_node_with_optimizer(halt); 3713 set_loop(halt, l); 3714 C->root()->add_req(halt); 3715 } 3716 set_loop(C->root(), _ltree_root); 3717 } 3718 } 3719 // Weeny check for irreducible. This child was already visited (this 3720 // IS the post-work phase). Is this child's loop header post-visited 3721 // as well? If so, then I found another entry into the loop. 3722 if (!_verify_only) { 3723 while( is_postvisited(l->_head) ) { 3724 // found irreducible 3725 l->_irreducible = 1; // = true 3726 l = l->_parent; 3727 _has_irreducible_loops = true; 3728 // Check for bad CFG here to prevent crash, and bailout of compile 3729 if (l == NULL) { 3730 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); 3731 return pre_order; 3732 } 3733 } 3734 C->set_has_irreducible_loop(_has_irreducible_loops); 3735 } 3736 3737 // This Node might be a decision point for loops. It is only if 3738 // it's children belong to several different loops. The sort call 3739 // does a trivial amount of work if there is only 1 child or all 3740 // children belong to the same loop. If however, the children 3741 // belong to different loops, the sort call will properly set the 3742 // _parent pointers to show how the loops nest. 3743 // 3744 // In any case, it returns the tightest enclosing loop. 3745 innermost = sort( l, innermost ); 3746 } 3747 3748 // Def-use info will have some dead stuff; dead stuff will have no 3749 // loop decided on. 3750 3751 // Am I a loop header? If so fix up my parent's child and next ptrs. 3752 if( innermost && innermost->_head == n ) { 3753 assert( get_loop(n) == innermost, "" ); 3754 IdealLoopTree *p = innermost->_parent; 3755 IdealLoopTree *l = innermost; 3756 while( p && l->_head == n ) { 3757 l->_next = p->_child; // Put self on parents 'next child' 3758 p->_child = l; // Make self as first child of parent 3759 l = p; // Now walk up the parent chain 3760 p = l->_parent; 3761 } 3762 } else { 3763 // Note that it is possible for a LoopNode to reach here, if the 3764 // backedge has been made unreachable (hence the LoopNode no longer 3765 // denotes a Loop, and will eventually be removed). 3766 3767 // Record tightest enclosing loop for self. Mark as post-visited. 3768 set_loop(n, innermost); 3769 // Also record has_call flag early on 3770 if( innermost ) { 3771 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { 3772 // Do not count uncommon calls 3773 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { 3774 Node *iff = n->in(0)->in(0); 3775 // No any calls for vectorized loops. 3776 if( UseSuperWord || !iff->is_If() || 3777 (n->in(0)->Opcode() == Op_IfFalse && 3778 (1.0 - iff->as_If()->_prob) >= 0.01) || 3779 (iff->as_If()->_prob >= 0.01) ) 3780 innermost->_has_call = 1; 3781 } 3782 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { 3783 // Disable loop optimizations if the loop has a scalar replaceable 3784 // allocation. This disabling may cause a potential performance lost 3785 // if the allocation is not eliminated for some reason. 3786 innermost->_allow_optimizations = false; 3787 innermost->_has_call = 1; // = true 3788 } else if (n->Opcode() == Op_SafePoint) { 3789 // Record all safepoints in this loop. 3790 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List(); 3791 innermost->_safepts->push(n); 3792 } 3793 } 3794 } 3795 3796 // Flag as post-visited now 3797 set_postvisited(n); 3798 return pre_order; 3799 } 3800 3801 3802 //------------------------------build_loop_early------------------------------- 3803 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3804 // First pass computes the earliest controlling node possible. This is the 3805 // controlling input with the deepest dominating depth. 3806 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3807 while (worklist.size() != 0) { 3808 // Use local variables nstack_top_n & nstack_top_i to cache values 3809 // on nstack's top. 3810 Node *nstack_top_n = worklist.pop(); 3811 uint nstack_top_i = 0; 3812 //while_nstack_nonempty: 3813 while (true) { 3814 // Get parent node and next input's index from stack's top. 3815 Node *n = nstack_top_n; 3816 uint i = nstack_top_i; 3817 uint cnt = n->req(); // Count of inputs 3818 if (i == 0) { // Pre-process the node. 3819 if( has_node(n) && // Have either loop or control already? 3820 !has_ctrl(n) ) { // Have loop picked out already? 3821 // During "merge_many_backedges" we fold up several nested loops 3822 // into a single loop. This makes the members of the original 3823 // loop bodies pointing to dead loops; they need to move up 3824 // to the new UNION'd larger loop. I set the _head field of these 3825 // dead loops to NULL and the _parent field points to the owning 3826 // loop. Shades of UNION-FIND algorithm. 3827 IdealLoopTree *ilt; 3828 while( !(ilt = get_loop(n))->_head ) { 3829 // Normally I would use a set_loop here. But in this one special 3830 // case, it is legal (and expected) to change what loop a Node 3831 // belongs to. 3832 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); 3833 } 3834 // Remove safepoints ONLY if I've already seen I don't need one. 3835 // (the old code here would yank a 2nd safepoint after seeing a 3836 // first one, even though the 1st did not dominate in the loop body 3837 // and thus could be avoided indefinitely) 3838 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && 3839 is_deleteable_safept(n)) { 3840 Node *in = n->in(TypeFunc::Control); 3841 lazy_replace(n,in); // Pull safepoint now 3842 if (ilt->_safepts != NULL) { 3843 ilt->_safepts->yank(n); 3844 } 3845 // Carry on with the recursion "as if" we are walking 3846 // only the control input 3847 if( !visited.test_set( in->_idx ) ) { 3848 worklist.push(in); // Visit this guy later, using worklist 3849 } 3850 // Get next node from nstack: 3851 // - skip n's inputs processing by setting i > cnt; 3852 // - we also will not call set_early_ctrl(n) since 3853 // has_node(n) == true (see the condition above). 3854 i = cnt + 1; 3855 } 3856 } 3857 } // if (i == 0) 3858 3859 // Visit all inputs 3860 bool done = true; // Assume all n's inputs will be processed 3861 while (i < cnt) { 3862 Node *in = n->in(i); 3863 ++i; 3864 if (in == NULL) continue; 3865 if (in->pinned() && !in->is_CFG()) 3866 set_ctrl(in, in->in(0)); 3867 int is_visited = visited.test_set( in->_idx ); 3868 if (!has_node(in)) { // No controlling input yet? 3869 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); 3870 assert( !is_visited, "visit only once" ); 3871 nstack.push(n, i); // Save parent node and next input's index. 3872 nstack_top_n = in; // Process current input now. 3873 nstack_top_i = 0; 3874 done = false; // Not all n's inputs processed. 3875 break; // continue while_nstack_nonempty; 3876 } else if (!is_visited) { 3877 // This guy has a location picked out for him, but has not yet 3878 // been visited. Happens to all CFG nodes, for instance. 3879 // Visit him using the worklist instead of recursion, to break 3880 // cycles. Since he has a location already we do not need to 3881 // find his location before proceeding with the current Node. 3882 worklist.push(in); // Visit this guy later, using worklist 3883 } 3884 } 3885 if (done) { 3886 // All of n's inputs have been processed, complete post-processing. 3887 3888 // Compute earliest point this Node can go. 3889 // CFG, Phi, pinned nodes already know their controlling input. 3890 if (!has_node(n)) { 3891 // Record earliest legal location 3892 set_early_ctrl( n ); 3893 } 3894 if (nstack.is_empty()) { 3895 // Finished all nodes on stack. 3896 // Process next node on the worklist. 3897 break; 3898 } 3899 // Get saved parent node and next input's index. 3900 nstack_top_n = nstack.node(); 3901 nstack_top_i = nstack.index(); 3902 nstack.pop(); 3903 } 3904 } // while (true) 3905 } 3906 } 3907 3908 //------------------------------dom_lca_internal-------------------------------- 3909 // Pair-wise LCA 3910 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { 3911 if( !n1 ) return n2; // Handle NULL original LCA 3912 assert( n1->is_CFG(), "" ); 3913 assert( n2->is_CFG(), "" ); 3914 // find LCA of all uses 3915 uint d1 = dom_depth(n1); 3916 uint d2 = dom_depth(n2); 3917 while (n1 != n2) { 3918 if (d1 > d2) { 3919 n1 = idom(n1); 3920 d1 = dom_depth(n1); 3921 } else if (d1 < d2) { 3922 n2 = idom(n2); 3923 d2 = dom_depth(n2); 3924 } else { 3925 // Here d1 == d2. Due to edits of the dominator-tree, sections 3926 // of the tree might have the same depth. These sections have 3927 // to be searched more carefully. 3928 3929 // Scan up all the n1's with equal depth, looking for n2. 3930 Node *t1 = idom(n1); 3931 while (dom_depth(t1) == d1) { 3932 if (t1 == n2) return n2; 3933 t1 = idom(t1); 3934 } 3935 // Scan up all the n2's with equal depth, looking for n1. 3936 Node *t2 = idom(n2); 3937 while (dom_depth(t2) == d2) { 3938 if (t2 == n1) return n1; 3939 t2 = idom(t2); 3940 } 3941 // Move up to a new dominator-depth value as well as up the dom-tree. 3942 n1 = t1; 3943 n2 = t2; 3944 d1 = dom_depth(n1); 3945 d2 = dom_depth(n2); 3946 } 3947 } 3948 return n1; 3949 } 3950 3951 //------------------------------compute_idom----------------------------------- 3952 // Locally compute IDOM using dom_lca call. Correct only if the incoming 3953 // IDOMs are correct. 3954 Node *PhaseIdealLoop::compute_idom( Node *region ) const { 3955 assert( region->is_Region(), "" ); 3956 Node *LCA = NULL; 3957 for( uint i = 1; i < region->req(); i++ ) { 3958 if( region->in(i) != C->top() ) 3959 LCA = dom_lca( LCA, region->in(i) ); 3960 } 3961 return LCA; 3962 } 3963 3964 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { 3965 bool had_error = false; 3966 #ifdef ASSERT 3967 if (early != C->root()) { 3968 // Make sure that there's a dominance path from LCA to early 3969 Node* d = LCA; 3970 while (d != early) { 3971 if (d == C->root()) { 3972 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA); 3973 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx); 3974 had_error = true; 3975 break; 3976 } 3977 d = idom(d); 3978 } 3979 } 3980 #endif 3981 return had_error; 3982 } 3983 3984 3985 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { 3986 // Compute LCA over list of uses 3987 bool had_error = false; 3988 Node *LCA = NULL; 3989 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { 3990 Node* c = n->fast_out(i); 3991 if (_nodes[c->_idx] == NULL) 3992 continue; // Skip the occasional dead node 3993 if( c->is_Phi() ) { // For Phis, we must land above on the path 3994 for( uint j=1; j<c->req(); j++ ) {// For all inputs 3995 if( c->in(j) == n ) { // Found matching input? 3996 Node *use = c->in(0)->in(j); 3997 if (_verify_only && use->is_top()) continue; 3998 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3999 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 4000 } 4001 } 4002 } else { 4003 // For CFG data-users, use is in the block just prior 4004 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); 4005 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 4006 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 4007 } 4008 } 4009 assert(!had_error, "bad dominance"); 4010 return LCA; 4011 } 4012 4013 // Check the shape of the graph at the loop entry. In some cases, 4014 // the shape of the graph does not match the shape outlined below. 4015 // That is caused by the Opaque1 node "protecting" the shape of 4016 // the graph being removed by, for example, the IGVN performed 4017 // in PhaseIdealLoop::build_and_optimize(). 4018 // 4019 // After the Opaque1 node has been removed, optimizations (e.g., split-if, 4020 // loop unswitching, and IGVN, or a combination of them) can freely change 4021 // the graph's shape. As a result, the graph shape outlined below cannot 4022 // be guaranteed anymore. 4023 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) { 4024 if (!cl->is_main_loop() && !cl->is_post_loop()) { 4025 return false; 4026 } 4027 Node* ctrl = cl->skip_predicates(); 4028 4029 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) { 4030 return false; 4031 } 4032 Node* iffm = ctrl->in(0); 4033 if (iffm == NULL || !iffm->is_If()) { 4034 return false; 4035 } 4036 Node* bolzm = iffm->in(1); 4037 if (bolzm == NULL || !bolzm->is_Bool()) { 4038 return false; 4039 } 4040 Node* cmpzm = bolzm->in(1); 4041 if (cmpzm == NULL || !cmpzm->is_Cmp()) { 4042 return false; 4043 } 4044 // compares can get conditionally flipped 4045 bool found_opaque = false; 4046 for (uint i = 1; i < cmpzm->req(); i++) { 4047 Node* opnd = cmpzm->in(i); 4048 if (opnd && opnd->Opcode() == Op_Opaque1) { 4049 found_opaque = true; 4050 break; 4051 } 4052 } 4053 if (!found_opaque) { 4054 return false; 4055 } 4056 return true; 4057 } 4058 4059 //------------------------------get_late_ctrl---------------------------------- 4060 // Compute latest legal control. 4061 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { 4062 assert(early != NULL, "early control should not be NULL"); 4063 4064 Node* LCA = compute_lca_of_uses(n, early); 4065 #ifdef ASSERT 4066 if (LCA == C->root() && LCA != early) { 4067 // def doesn't dominate uses so print some useful debugging output 4068 compute_lca_of_uses(n, early, true); 4069 } 4070 #endif 4071 4072 // if this is a load, check for anti-dependent stores 4073 // We use a conservative algorithm to identify potential interfering 4074 // instructions and for rescheduling the load. The users of the memory 4075 // input of this load are examined. Any use which is not a load and is 4076 // dominated by early is considered a potentially interfering store. 4077 // This can produce false positives. 4078 if (n->is_Load() && LCA != early) { 4079 int load_alias_idx = C->get_alias_index(n->adr_type()); 4080 if (C->alias_type(load_alias_idx)->is_rewritable()) { 4081 4082 Node_List worklist; 4083 4084 Node *mem = n->in(MemNode::Memory); 4085 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 4086 Node* s = mem->fast_out(i); 4087 worklist.push(s); 4088 } 4089 while(worklist.size() != 0 && LCA != early) { 4090 Node* s = worklist.pop(); 4091 if (s->is_Load() || s->Opcode() == Op_SafePoint || 4092 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) { 4093 continue; 4094 } else if (s->is_MergeMem()) { 4095 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 4096 Node* s1 = s->fast_out(i); 4097 worklist.push(s1); 4098 } 4099 } else { 4100 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); 4101 const TypePtr* adr_type = s->adr_type(); 4102 if (s->is_ArrayCopy()) { 4103 // Copy to known instance needs destination type to test for aliasing 4104 const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type; 4105 if (dest_type != TypeOopPtr::BOTTOM) { 4106 adr_type = dest_type; 4107 } 4108 } 4109 assert(sctrl != NULL || !s->is_reachable_from_root(), "must have control"); 4110 if (sctrl != NULL && !sctrl->is_top() && C->can_alias(adr_type, load_alias_idx) && is_dominator(early, sctrl)) { 4111 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); 4112 } 4113 } 4114 } 4115 } 4116 } 4117 4118 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); 4119 return LCA; 4120 } 4121 4122 // true if CFG node d dominates CFG node n 4123 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { 4124 if (d == n) 4125 return true; 4126 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 4127 uint dd = dom_depth(d); 4128 while (dom_depth(n) >= dd) { 4129 if (n == d) 4130 return true; 4131 n = idom(n); 4132 } 4133 return false; 4134 } 4135 4136 //------------------------------dom_lca_for_get_late_ctrl_internal------------- 4137 // Pair-wise LCA with tags. 4138 // Tag each index with the node 'tag' currently being processed 4139 // before advancing up the dominator chain using idom(). 4140 // Later calls that find a match to 'tag' know that this path has already 4141 // been considered in the current LCA (which is input 'n1' by convention). 4142 // Since get_late_ctrl() is only called once for each node, the tag array 4143 // does not need to be cleared between calls to get_late_ctrl(). 4144 // Algorithm trades a larger constant factor for better asymptotic behavior 4145 // 4146 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { 4147 uint d1 = dom_depth(n1); 4148 uint d2 = dom_depth(n2); 4149 4150 do { 4151 if (d1 > d2) { 4152 // current lca is deeper than n2 4153 _dom_lca_tags.map(n1->_idx, tag); 4154 n1 = idom(n1); 4155 d1 = dom_depth(n1); 4156 } else if (d1 < d2) { 4157 // n2 is deeper than current lca 4158 Node *memo = _dom_lca_tags[n2->_idx]; 4159 if( memo == tag ) { 4160 return n1; // Return the current LCA 4161 } 4162 _dom_lca_tags.map(n2->_idx, tag); 4163 n2 = idom(n2); 4164 d2 = dom_depth(n2); 4165 } else { 4166 // Here d1 == d2. Due to edits of the dominator-tree, sections 4167 // of the tree might have the same depth. These sections have 4168 // to be searched more carefully. 4169 4170 // Scan up all the n1's with equal depth, looking for n2. 4171 _dom_lca_tags.map(n1->_idx, tag); 4172 Node *t1 = idom(n1); 4173 while (dom_depth(t1) == d1) { 4174 if (t1 == n2) return n2; 4175 _dom_lca_tags.map(t1->_idx, tag); 4176 t1 = idom(t1); 4177 } 4178 // Scan up all the n2's with equal depth, looking for n1. 4179 _dom_lca_tags.map(n2->_idx, tag); 4180 Node *t2 = idom(n2); 4181 while (dom_depth(t2) == d2) { 4182 if (t2 == n1) return n1; 4183 _dom_lca_tags.map(t2->_idx, tag); 4184 t2 = idom(t2); 4185 } 4186 // Move up to a new dominator-depth value as well as up the dom-tree. 4187 n1 = t1; 4188 n2 = t2; 4189 d1 = dom_depth(n1); 4190 d2 = dom_depth(n2); 4191 } 4192 } while (n1 != n2); 4193 return n1; 4194 } 4195 4196 //------------------------------init_dom_lca_tags------------------------------ 4197 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 4198 // Intended use does not involve any growth for the array, so it could 4199 // be of fixed size. 4200 void PhaseIdealLoop::init_dom_lca_tags() { 4201 uint limit = C->unique() + 1; 4202 _dom_lca_tags.map( limit, NULL ); 4203 #ifdef ASSERT 4204 for( uint i = 0; i < limit; ++i ) { 4205 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 4206 } 4207 #endif // ASSERT 4208 } 4209 4210 //------------------------------clear_dom_lca_tags------------------------------ 4211 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 4212 // Intended use does not involve any growth for the array, so it could 4213 // be of fixed size. 4214 void PhaseIdealLoop::clear_dom_lca_tags() { 4215 uint limit = C->unique() + 1; 4216 _dom_lca_tags.map( limit, NULL ); 4217 _dom_lca_tags.clear(); 4218 #ifdef ASSERT 4219 for( uint i = 0; i < limit; ++i ) { 4220 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 4221 } 4222 #endif // ASSERT 4223 } 4224 4225 //------------------------------build_loop_late-------------------------------- 4226 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 4227 // Second pass finds latest legal placement, and ideal loop placement. 4228 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 4229 while (worklist.size() != 0) { 4230 Node *n = worklist.pop(); 4231 // Only visit once 4232 if (visited.test_set(n->_idx)) continue; 4233 uint cnt = n->outcnt(); 4234 uint i = 0; 4235 while (true) { 4236 assert( _nodes[n->_idx], "no dead nodes" ); 4237 // Visit all children 4238 if (i < cnt) { 4239 Node* use = n->raw_out(i); 4240 ++i; 4241 // Check for dead uses. Aggressively prune such junk. It might be 4242 // dead in the global sense, but still have local uses so I cannot 4243 // easily call 'remove_dead_node'. 4244 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? 4245 // Due to cycles, we might not hit the same fixed point in the verify 4246 // pass as we do in the regular pass. Instead, visit such phis as 4247 // simple uses of the loop head. 4248 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { 4249 if( !visited.test(use->_idx) ) 4250 worklist.push(use); 4251 } else if( !visited.test_set(use->_idx) ) { 4252 nstack.push(n, i); // Save parent and next use's index. 4253 n = use; // Process all children of current use. 4254 cnt = use->outcnt(); 4255 i = 0; 4256 } 4257 } else { 4258 // Do not visit around the backedge of loops via data edges. 4259 // push dead code onto a worklist 4260 _deadlist.push(use); 4261 } 4262 } else { 4263 // All of n's children have been processed, complete post-processing. 4264 build_loop_late_post(n); 4265 if (nstack.is_empty()) { 4266 // Finished all nodes on stack. 4267 // Process next node on the worklist. 4268 break; 4269 } 4270 // Get saved parent node and next use's index. Visit the rest of uses. 4271 n = nstack.node(); 4272 cnt = n->outcnt(); 4273 i = nstack.index(); 4274 nstack.pop(); 4275 } 4276 } 4277 } 4278 } 4279 4280 // Verify that no data node is scheduled in the outer loop of a strip 4281 // mined loop. 4282 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) { 4283 #ifdef ASSERT 4284 if (get_loop(least)->_nest == 0) { 4285 return; 4286 } 4287 IdealLoopTree* loop = get_loop(least); 4288 Node* head = loop->_head; 4289 if (head->is_OuterStripMinedLoop() && 4290 // Verification can't be applied to fully built strip mined loops 4291 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) { 4292 Node* sfpt = head->as_Loop()->outer_safepoint(); 4293 ResourceMark rm; 4294 Unique_Node_List wq; 4295 wq.push(sfpt); 4296 for (uint i = 0; i < wq.size(); i++) { 4297 Node *m = wq.at(i); 4298 for (uint i = 1; i < m->req(); i++) { 4299 Node* nn = m->in(i); 4300 if (nn == n) { 4301 return; 4302 } 4303 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) { 4304 wq.push(nn); 4305 } 4306 } 4307 } 4308 ShouldNotReachHere(); 4309 } 4310 #endif 4311 } 4312 4313 4314 //------------------------------build_loop_late_post--------------------------- 4315 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 4316 // Second pass finds latest legal placement, and ideal loop placement. 4317 void PhaseIdealLoop::build_loop_late_post(Node *n) { 4318 build_loop_late_post_work(n, true); 4319 } 4320 4321 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) { 4322 4323 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) { 4324 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. 4325 } 4326 4327 #ifdef ASSERT 4328 if (_verify_only && !n->is_CFG()) { 4329 // Check def-use domination. 4330 compute_lca_of_uses(n, get_ctrl(n), true /* verify */); 4331 } 4332 #endif 4333 4334 // CFG and pinned nodes already handled 4335 if( n->in(0) ) { 4336 if( n->in(0)->is_top() ) return; // Dead? 4337 4338 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads 4339 // _must_ be pinned (they have to observe their control edge of course). 4340 // Unlike Stores (which modify an unallocable resource, the memory 4341 // state), Mods/Loads can float around. So free them up. 4342 switch( n->Opcode() ) { 4343 case Op_DivI: 4344 case Op_DivF: 4345 case Op_DivD: 4346 case Op_ModI: 4347 case Op_ModF: 4348 case Op_ModD: 4349 case Op_LoadB: // Same with Loads; they can sink 4350 case Op_LoadUB: // during loop optimizations. 4351 case Op_LoadUS: 4352 case Op_LoadD: 4353 case Op_LoadF: 4354 case Op_LoadI: 4355 case Op_LoadKlass: 4356 case Op_LoadNKlass: 4357 case Op_LoadL: 4358 case Op_LoadS: 4359 case Op_LoadP: 4360 case Op_LoadN: 4361 case Op_LoadRange: 4362 case Op_LoadD_unaligned: 4363 case Op_LoadL_unaligned: 4364 case Op_StrComp: // Does a bunch of load-like effects 4365 case Op_StrEquals: 4366 case Op_StrIndexOf: 4367 case Op_StrIndexOfChar: 4368 case Op_AryEq: 4369 case Op_HasNegatives: 4370 pinned = false; 4371 } 4372 if (n->is_CMove()) { 4373 pinned = false; 4374 } 4375 if( pinned ) { 4376 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); 4377 if( !chosen_loop->_child ) // Inner loop? 4378 chosen_loop->_body.push(n); // Collect inner loops 4379 return; 4380 } 4381 } else { // No slot zero 4382 if( n->is_CFG() ) { // CFG with no slot 0 is dead 4383 _nodes.map(n->_idx,0); // No block setting, it's globally dead 4384 return; 4385 } 4386 assert(!n->is_CFG() || n->outcnt() == 0, ""); 4387 } 4388 4389 // Do I have a "safe range" I can select over? 4390 Node *early = get_ctrl(n);// Early location already computed 4391 4392 // Compute latest point this Node can go 4393 Node *LCA = get_late_ctrl( n, early ); 4394 // LCA is NULL due to uses being dead 4395 if( LCA == NULL ) { 4396 #ifdef ASSERT 4397 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { 4398 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); 4399 } 4400 #endif 4401 _nodes.map(n->_idx, 0); // This node is useless 4402 _deadlist.push(n); 4403 return; 4404 } 4405 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); 4406 4407 Node *legal = LCA; // Walk 'legal' up the IDOM chain 4408 Node *least = legal; // Best legal position so far 4409 while( early != legal ) { // While not at earliest legal 4410 #ifdef ASSERT 4411 if (legal->is_Start() && !early->is_Root()) { 4412 // Bad graph. Print idom path and fail. 4413 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA); 4414 assert(false, "Bad graph detected in build_loop_late"); 4415 } 4416 #endif 4417 // Find least loop nesting depth 4418 legal = idom(legal); // Bump up the IDOM tree 4419 // Check for lower nesting depth 4420 if( get_loop(legal)->_nest < get_loop(least)->_nest ) 4421 least = legal; 4422 } 4423 assert(early == legal || legal != C->root(), "bad dominance of inputs"); 4424 4425 // Try not to place code on a loop entry projection 4426 // which can inhibit range check elimination. 4427 if (least != early) { 4428 Node* ctrl_out = least->unique_ctrl_out(); 4429 if (ctrl_out && ctrl_out->is_Loop() && 4430 least == ctrl_out->in(LoopNode::EntryControl)) { 4431 // Move the node above predicates as far up as possible so a 4432 // following pass of loop predication doesn't hoist a predicate 4433 // that depends on it above that node. 4434 Node* new_ctrl = least; 4435 for (;;) { 4436 if (!new_ctrl->is_Proj()) { 4437 break; 4438 } 4439 CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 4440 if (call == NULL) { 4441 break; 4442 } 4443 int req = call->uncommon_trap_request(); 4444 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req); 4445 if (trap_reason != Deoptimization::Reason_loop_limit_check && 4446 trap_reason != Deoptimization::Reason_predicate && 4447 trap_reason != Deoptimization::Reason_profile_predicate) { 4448 break; 4449 } 4450 Node* c = new_ctrl->in(0)->in(0); 4451 if (is_dominator(c, early) && c != early) { 4452 break; 4453 } 4454 new_ctrl = c; 4455 } 4456 least = new_ctrl; 4457 } 4458 } 4459 4460 #ifdef ASSERT 4461 // If verifying, verify that 'verify_me' has a legal location 4462 // and choose it as our location. 4463 if( _verify_me ) { 4464 Node *v_ctrl = _verify_me->get_ctrl_no_update(n); 4465 Node *legal = LCA; 4466 while( early != legal ) { // While not at earliest legal 4467 if( legal == v_ctrl ) break; // Check for prior good location 4468 legal = idom(legal) ;// Bump up the IDOM tree 4469 } 4470 // Check for prior good location 4471 if( legal == v_ctrl ) least = legal; // Keep prior if found 4472 } 4473 #endif 4474 4475 // Assign discovered "here or above" point 4476 least = find_non_split_ctrl(least); 4477 verify_strip_mined_scheduling(n, least); 4478 set_ctrl(n, least); 4479 4480 // Collect inner loop bodies 4481 IdealLoopTree *chosen_loop = get_loop(least); 4482 if( !chosen_loop->_child ) // Inner loop? 4483 chosen_loop->_body.push(n);// Collect inner loops 4484 } 4485 4486 #ifdef ASSERT 4487 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) { 4488 tty->print_cr("%s", msg); 4489 tty->print("n: "); n->dump(); 4490 tty->print("early(n): "); early->dump(); 4491 if (n->in(0) != NULL && !n->in(0)->is_top() && 4492 n->in(0) != early && !n->in(0)->is_Root()) { 4493 tty->print("n->in(0): "); n->in(0)->dump(); 4494 } 4495 for (uint i = 1; i < n->req(); i++) { 4496 Node* in1 = n->in(i); 4497 if (in1 != NULL && in1 != n && !in1->is_top()) { 4498 tty->print("n->in(%d): ", i); in1->dump(); 4499 Node* in1_early = get_ctrl(in1); 4500 tty->print("early(n->in(%d)): ", i); in1_early->dump(); 4501 if (in1->in(0) != NULL && !in1->in(0)->is_top() && 4502 in1->in(0) != in1_early && !in1->in(0)->is_Root()) { 4503 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump(); 4504 } 4505 for (uint j = 1; j < in1->req(); j++) { 4506 Node* in2 = in1->in(j); 4507 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) { 4508 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump(); 4509 Node* in2_early = get_ctrl(in2); 4510 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); 4511 if (in2->in(0) != NULL && !in2->in(0)->is_top() && 4512 in2->in(0) != in2_early && !in2->in(0)->is_Root()) { 4513 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump(); 4514 } 4515 } 4516 } 4517 } 4518 } 4519 tty->cr(); 4520 tty->print("LCA(n): "); LCA->dump(); 4521 for (uint i = 0; i < n->outcnt(); i++) { 4522 Node* u1 = n->raw_out(i); 4523 if (u1 == n) 4524 continue; 4525 tty->print("n->out(%d): ", i); u1->dump(); 4526 if (u1->is_CFG()) { 4527 for (uint j = 0; j < u1->outcnt(); j++) { 4528 Node* u2 = u1->raw_out(j); 4529 if (u2 != u1 && u2 != n && u2->is_CFG()) { 4530 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4531 } 4532 } 4533 } else { 4534 Node* u1_later = get_ctrl(u1); 4535 tty->print("later(n->out(%d)): ", i); u1_later->dump(); 4536 if (u1->in(0) != NULL && !u1->in(0)->is_top() && 4537 u1->in(0) != u1_later && !u1->in(0)->is_Root()) { 4538 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump(); 4539 } 4540 for (uint j = 0; j < u1->outcnt(); j++) { 4541 Node* u2 = u1->raw_out(j); 4542 if (u2 == n || u2 == u1) 4543 continue; 4544 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4545 if (!u2->is_CFG()) { 4546 Node* u2_later = get_ctrl(u2); 4547 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); 4548 if (u2->in(0) != NULL && !u2->in(0)->is_top() && 4549 u2->in(0) != u2_later && !u2->in(0)->is_Root()) { 4550 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump(); 4551 } 4552 } 4553 } 4554 } 4555 } 4556 tty->cr(); 4557 int ct = 0; 4558 Node *dbg_legal = LCA; 4559 while(!dbg_legal->is_Start() && ct < 100) { 4560 tty->print("idom[%d] ",ct); dbg_legal->dump(); 4561 ct++; 4562 dbg_legal = idom(dbg_legal); 4563 } 4564 tty->cr(); 4565 } 4566 #endif 4567 4568 #ifndef PRODUCT 4569 //------------------------------dump------------------------------------------- 4570 void PhaseIdealLoop::dump() const { 4571 ResourceMark rm; 4572 Arena* arena = Thread::current()->resource_area(); 4573 Node_Stack stack(arena, C->live_nodes() >> 2); 4574 Node_List rpo_list; 4575 VectorSet visited(arena); 4576 visited.set(C->top()->_idx); 4577 rpo(C->root(), stack, visited, rpo_list); 4578 // Dump root loop indexed by last element in PO order 4579 dump(_ltree_root, rpo_list.size(), rpo_list); 4580 } 4581 4582 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const { 4583 loop->dump_head(); 4584 4585 // Now scan for CFG nodes in the same loop 4586 for (uint j = idx; j > 0; j--) { 4587 Node* n = rpo_list[j-1]; 4588 if (!_nodes[n->_idx]) // Skip dead nodes 4589 continue; 4590 4591 if (get_loop(n) != loop) { // Wrong loop nest 4592 if (get_loop(n)->_head == n && // Found nested loop? 4593 get_loop(n)->_parent == loop) 4594 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly 4595 continue; 4596 } 4597 4598 // Dump controlling node 4599 tty->sp(2 * loop->_nest); 4600 tty->print("C"); 4601 if (n == C->root()) { 4602 n->dump(); 4603 } else { 4604 Node* cached_idom = idom_no_update(n); 4605 Node* computed_idom = n->in(0); 4606 if (n->is_Region()) { 4607 computed_idom = compute_idom(n); 4608 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or 4609 // any MultiBranch ctrl node), so apply a similar transform to 4610 // the cached idom returned from idom_no_update. 4611 cached_idom = find_non_split_ctrl(cached_idom); 4612 } 4613 tty->print(" ID:%d", computed_idom->_idx); 4614 n->dump(); 4615 if (cached_idom != computed_idom) { 4616 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", 4617 computed_idom->_idx, cached_idom->_idx); 4618 } 4619 } 4620 // Dump nodes it controls 4621 for (uint k = 0; k < _nodes.Size(); k++) { 4622 // (k < C->unique() && get_ctrl(find(k)) == n) 4623 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { 4624 Node* m = C->root()->find(k); 4625 if (m && m->outcnt() > 0) { 4626 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { 4627 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", 4628 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); 4629 } 4630 tty->sp(2 * loop->_nest + 1); 4631 m->dump(); 4632 } 4633 } 4634 } 4635 } 4636 } 4637 #endif 4638 4639 // Collect a R-P-O for the whole CFG. 4640 // Result list is in post-order (scan backwards for RPO) 4641 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const { 4642 stk.push(start, 0); 4643 visited.set(start->_idx); 4644 4645 while (stk.is_nonempty()) { 4646 Node* m = stk.node(); 4647 uint idx = stk.index(); 4648 if (idx < m->outcnt()) { 4649 stk.set_index(idx + 1); 4650 Node* n = m->raw_out(idx); 4651 if (n->is_CFG() && !visited.test_set(n->_idx)) { 4652 stk.push(n, 0); 4653 } 4654 } else { 4655 rpo_list.push(m); 4656 stk.pop(); 4657 } 4658 } 4659 } 4660 4661 4662 //============================================================================= 4663 //------------------------------LoopTreeIterator------------------------------- 4664 4665 // Advance to next loop tree using a preorder, left-to-right traversal. 4666 void LoopTreeIterator::next() { 4667 assert(!done(), "must not be done."); 4668 if (_curnt->_child != NULL) { 4669 _curnt = _curnt->_child; 4670 } else if (_curnt->_next != NULL) { 4671 _curnt = _curnt->_next; 4672 } else { 4673 while (_curnt != _root && _curnt->_next == NULL) { 4674 _curnt = _curnt->_parent; 4675 } 4676 if (_curnt == _root) { 4677 _curnt = NULL; 4678 assert(done(), "must be done."); 4679 } else { 4680 assert(_curnt->_next != NULL, "must be more to do"); 4681 _curnt = _curnt->_next; 4682 } 4683 } 4684 }