1 /* 2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciMethodData.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "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/callnode.hpp" 35 #include "opto/connode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/divnode.hpp" 38 #include "opto/idealGraphPrinter.hpp" 39 #include "opto/loopnode.hpp" 40 #include "opto/mulnode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/superword.hpp" 43 44 //============================================================================= 45 //------------------------------is_loop_iv------------------------------------- 46 // Determine if a node is Counted loop induction variable. 47 // The method is declared in node.hpp. 48 const Node* Node::is_loop_iv() const { 49 if (this->is_Phi() && !this->as_Phi()->is_copy() && 50 this->as_Phi()->region()->is_CountedLoop() && 51 this->as_Phi()->region()->as_CountedLoop()->phi() == this) { 52 return this; 53 } else { 54 return NULL; 55 } 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 //------------------------------is_counted_loop-------------------------------- 328 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) { 329 PhaseGVN *gvn = &_igvn; 330 331 // Counted loop head must be a good RegionNode with only 3 not NULL 332 // control input edges: Self, Entry, LoopBack. 333 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) { 334 return false; 335 } 336 Node *init_control = x->in(LoopNode::EntryControl); 337 Node *back_control = x->in(LoopNode::LoopBackControl); 338 if (init_control == NULL || back_control == NULL) // Partially dead 339 return false; 340 // Must also check for TOP when looking for a dead loop 341 if (init_control->is_top() || back_control->is_top()) 342 return false; 343 344 // Allow funny placement of Safepoint 345 if (back_control->Opcode() == Op_SafePoint) { 346 if (LoopStripMiningIter != 0) { 347 // Leaving the safepoint on the backedge and creating a 348 // CountedLoop will confuse optimizations. We can't move the 349 // safepoint around because its jvm state wouldn't match a new 350 // location. Give up on that loop. 351 return false; 352 } 353 back_control = back_control->in(TypeFunc::Control); 354 } 355 356 // Controlling test for loop 357 Node *iftrue = back_control; 358 uint iftrue_op = iftrue->Opcode(); 359 if (iftrue_op != Op_IfTrue && 360 iftrue_op != Op_IfFalse) 361 // I have a weird back-control. Probably the loop-exit test is in 362 // the middle of the loop and I am looking at some trailing control-flow 363 // merge point. To fix this I would have to partially peel the loop. 364 return false; // Obscure back-control 365 366 // Get boolean guarding loop-back test 367 Node *iff = iftrue->in(0); 368 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) 369 return false; 370 BoolNode *test = iff->in(1)->as_Bool(); 371 BoolTest::mask bt = test->_test._test; 372 float cl_prob = iff->as_If()->_prob; 373 if (iftrue_op == Op_IfFalse) { 374 bt = BoolTest(bt).negate(); 375 cl_prob = 1.0 - cl_prob; 376 } 377 // Get backedge compare 378 Node *cmp = test->in(1); 379 int cmp_op = cmp->Opcode(); 380 if (cmp_op != Op_CmpI) 381 return false; // Avoid pointer & float compares 382 383 // Find the trip-counter increment & limit. Limit must be loop invariant. 384 Node *incr = cmp->in(1); 385 Node *limit = cmp->in(2); 386 387 // --------- 388 // need 'loop()' test to tell if limit is loop invariant 389 // --------- 390 391 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit? 392 Node *tmp = incr; // Then reverse order into the CmpI 393 incr = limit; 394 limit = tmp; 395 bt = BoolTest(bt).commute(); // And commute the exit test 396 } 397 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant 398 return false; 399 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 400 return false; 401 402 Node* phi_incr = NULL; 403 // Trip-counter increment must be commutative & associative. 404 if (incr->Opcode() == Op_CastII) { 405 incr = incr->in(1); 406 } 407 if (incr->is_Phi()) { 408 if (incr->as_Phi()->region() != x || incr->req() != 3) 409 return false; // Not simple trip counter expression 410 phi_incr = incr; 411 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi 412 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 413 return false; 414 } 415 416 Node* trunc1 = NULL; 417 Node* trunc2 = NULL; 418 const TypeInt* iv_trunc_t = NULL; 419 Node* orig_incr = incr; 420 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { 421 return false; // Funny increment opcode 422 } 423 assert(incr->Opcode() == Op_AddI, "wrong increment code"); 424 425 const TypeInt* limit_t = gvn->type(limit)->is_int(); 426 if (trunc1 != NULL) { 427 // When there is a truncation, we must be sure that after the truncation 428 // the trip counter will end up higher than the limit, otherwise we are looking 429 // at an endless loop. Can happen with range checks. 430 431 // Example: 432 // int i = 0; 433 // while (true) 434 // sum + = array[i]; 435 // i++; 436 // i = i && 0x7fff; 437 // } 438 // 439 // If the array is shorter than 0x8000 this exits through a AIOOB 440 // - Counted loop transformation is ok 441 // If the array is longer then this is an endless loop 442 // - No transformation can be done. 443 444 const TypeInt* incr_t = gvn->type(orig_incr)->is_int(); 445 if (limit_t->_hi > incr_t->_hi) { 446 // if the limit can have a higher value than the increment (before the phi) 447 return false; 448 } 449 } 450 451 // Get merge point 452 Node *xphi = incr->in(1); 453 Node *stride = incr->in(2); 454 if (!stride->is_Con()) { // Oops, swap these 455 if (!xphi->is_Con()) // Is the other guy a constant? 456 return false; // Nope, unknown stride, bail out 457 Node *tmp = xphi; // 'incr' is commutative, so ok to swap 458 xphi = stride; 459 stride = tmp; 460 } 461 if (xphi->Opcode() == Op_CastII) { 462 xphi = xphi->in(1); 463 } 464 // Stride must be constant 465 int stride_con = stride->get_int(); 466 if (stride_con == 0) 467 return false; // missed some peephole opt 468 469 if (!xphi->is_Phi()) 470 return false; // Too much math on the trip counter 471 if (phi_incr != NULL && phi_incr != xphi) 472 return false; 473 PhiNode *phi = xphi->as_Phi(); 474 475 // Phi must be of loop header; backedge must wrap to increment 476 if (phi->region() != x) 477 return false; 478 if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) || 479 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) { 480 return false; 481 } 482 Node *init_trip = phi->in(LoopNode::EntryControl); 483 484 // If iv trunc type is smaller than int, check for possible wrap. 485 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { 486 assert(trunc1 != NULL, "must have found some truncation"); 487 488 // Get a better type for the phi (filtered thru if's) 489 const TypeInt* phi_ft = filtered_type(phi); 490 491 // Can iv take on a value that will wrap? 492 // 493 // Ensure iv's limit is not within "stride" of the wrap value. 494 // 495 // Example for "short" type 496 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) 497 // If the stride is +10, then the last value of the induction 498 // variable before the increment (phi_ft->_hi) must be 499 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to 500 // ensure no truncation occurs after the increment. 501 502 if (stride_con > 0) { 503 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || 504 iv_trunc_t->_lo > phi_ft->_lo) { 505 return false; // truncation may occur 506 } 507 } else if (stride_con < 0) { 508 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || 509 iv_trunc_t->_hi < phi_ft->_hi) { 510 return false; // truncation may occur 511 } 512 } 513 // No possibility of wrap so truncation can be discarded 514 // Promote iv type to Int 515 } else { 516 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); 517 } 518 519 // If the condition is inverted and we will be rolling 520 // through MININT to MAXINT, then bail out. 521 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice! 522 // Odd stride 523 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) || 524 // Count down loop rolls through MAXINT 525 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) || 526 // Count up loop rolls through MININT 527 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) { 528 return false; // Bail out 529 } 530 531 const TypeInt* init_t = gvn->type(init_trip)->is_int(); 532 533 if (stride_con > 0) { 534 jlong init_p = (jlong)init_t->_lo + stride_con; 535 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi) 536 return false; // cyclic loop or this loop trips only once 537 } else { 538 jlong init_p = (jlong)init_t->_hi + stride_con; 539 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo) 540 return false; // cyclic loop or this loop trips only once 541 } 542 543 if (phi_incr != NULL) { 544 // check if there is a possiblity of IV overflowing after the first increment 545 if (stride_con > 0) { 546 if (init_t->_hi > max_jint - stride_con) { 547 return false; 548 } 549 } else { 550 if (init_t->_lo < min_jint - stride_con) { 551 return false; 552 } 553 } 554 } 555 556 // ================================================= 557 // ---- SUCCESS! Found A Trip-Counted Loop! ----- 558 // 559 assert(x->Opcode() == Op_Loop, "regular loops only"); 560 C->print_method(PHASE_BEFORE_CLOOPS, 3); 561 562 Node *hook = new Node(6); 563 564 // =================================================== 565 // Generate loop limit check to avoid integer overflow 566 // in cases like next (cyclic loops): 567 // 568 // for (i=0; i <= max_jint; i++) {} 569 // for (i=0; i < max_jint; i+=2) {} 570 // 571 // 572 // Limit check predicate depends on the loop test: 573 // 574 // for(;i != limit; i++) --> limit <= (max_jint) 575 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1) 576 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride ) 577 // 578 579 // Check if limit is excluded to do more precise int overflow check. 580 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge); 581 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1)); 582 583 // If compare points directly to the phi we need to adjust 584 // the compare so that it points to the incr. Limit have 585 // to be adjusted to keep trip count the same and the 586 // adjusted limit should be checked for int overflow. 587 if (phi_incr != NULL) { 588 stride_m += stride_con; 589 } 590 591 if (limit->is_Con()) { 592 int limit_con = limit->get_int(); 593 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) || 594 (stride_con < 0 && limit_con < (min_jint - stride_m))) { 595 // Bailout: it could be integer overflow. 596 return false; 597 } 598 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) || 599 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) { 600 // Limit's type may satisfy the condition, for example, 601 // when it is an array length. 602 } else { 603 // Generate loop's limit check. 604 // Loop limit check predicate should be near the loop. 605 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); 606 if (!limit_check_proj) { 607 // The limit check predicate is not generated if this method trapped here before. 608 #ifdef ASSERT 609 if (TraceLoopLimitCheck) { 610 tty->print("missing loop limit check:"); 611 loop->dump_head(); 612 x->dump(1); 613 } 614 #endif 615 return false; 616 } 617 618 IfNode* check_iff = limit_check_proj->in(0)->as_If(); 619 Node* cmp_limit; 620 Node* bol; 621 622 if (stride_con > 0) { 623 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m)); 624 bol = new BoolNode(cmp_limit, BoolTest::le); 625 } else { 626 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m)); 627 bol = new BoolNode(cmp_limit, BoolTest::ge); 628 } 629 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit); 630 bol = _igvn.register_new_node_with_optimizer(bol); 631 set_subtree_ctrl(bol); 632 633 // Replace condition in original predicate but preserve Opaque node 634 // so that previous predicates could be found. 635 assert(check_iff->in(1)->Opcode() == Op_Conv2B && 636 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, ""); 637 Node* opq = check_iff->in(1)->in(1); 638 _igvn.replace_input_of(opq, 1, bol); 639 // Update ctrl. 640 set_ctrl(opq, check_iff->in(0)); 641 set_ctrl(check_iff->in(1), check_iff->in(0)); 642 643 #ifndef PRODUCT 644 // report that the loop predication has been actually performed 645 // for this loop 646 if (TraceLoopLimitCheck) { 647 tty->print_cr("Counted Loop Limit Check generated:"); 648 debug_only( bol->dump(2); ) 649 } 650 #endif 651 } 652 653 if (phi_incr != NULL) { 654 // If compare points directly to the phi we need to adjust 655 // the compare so that it points to the incr. Limit have 656 // to be adjusted to keep trip count the same and we 657 // should avoid int overflow. 658 // 659 // i = init; do {} while(i++ < limit); 660 // is converted to 661 // i = init; do {} while(++i < limit+1); 662 // 663 limit = gvn->transform(new AddINode(limit, stride)); 664 } 665 666 // Now we need to canonicalize loop condition. 667 if (bt == BoolTest::ne) { 668 assert(stride_con == 1 || stride_con == -1, "simple increment only"); 669 // 'ne' can be replaced with 'lt' only when init < limit. 670 if (stride_con > 0 && init_t->_hi < limit_t->_lo) 671 bt = BoolTest::lt; 672 // 'ne' can be replaced with 'gt' only when init > limit. 673 if (stride_con < 0 && init_t->_lo > limit_t->_hi) 674 bt = BoolTest::gt; 675 } 676 677 if (incl_limit) { 678 // The limit check guaranties that 'limit <= (max_jint - stride)' so 679 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0. 680 // 681 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1); 682 limit = gvn->transform(new AddINode(limit, one)); 683 if (bt == BoolTest::le) 684 bt = BoolTest::lt; 685 else if (bt == BoolTest::ge) 686 bt = BoolTest::gt; 687 else 688 ShouldNotReachHere(); 689 } 690 set_subtree_ctrl( limit ); 691 692 if (LoopStripMiningIter == 0) { 693 // Check for SafePoint on backedge and remove 694 Node *sfpt = x->in(LoopNode::LoopBackControl); 695 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { 696 lazy_replace( sfpt, iftrue ); 697 if (loop->_safepts != NULL) { 698 loop->_safepts->yank(sfpt); 699 } 700 loop->_tail = iftrue; 701 } 702 } 703 704 // Build a canonical trip test. 705 // Clone code, as old values may be in use. 706 incr = incr->clone(); 707 incr->set_req(1,phi); 708 incr->set_req(2,stride); 709 incr = _igvn.register_new_node_with_optimizer(incr); 710 set_early_ctrl( incr ); 711 _igvn.rehash_node_delayed(phi); 712 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); 713 714 // If phi type is more restrictive than Int, raise to 715 // Int to prevent (almost) infinite recursion in igvn 716 // which can only handle integer types for constants or minint..maxint. 717 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { 718 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); 719 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); 720 nphi = _igvn.register_new_node_with_optimizer(nphi); 721 set_ctrl(nphi, get_ctrl(phi)); 722 _igvn.replace_node(phi, nphi); 723 phi = nphi->as_Phi(); 724 } 725 cmp = cmp->clone(); 726 cmp->set_req(1,incr); 727 cmp->set_req(2,limit); 728 cmp = _igvn.register_new_node_with_optimizer(cmp); 729 set_ctrl(cmp, iff->in(0)); 730 731 test = test->clone()->as_Bool(); 732 (*(BoolTest*)&test->_test)._test = bt; 733 test->set_req(1,cmp); 734 _igvn.register_new_node_with_optimizer(test); 735 set_ctrl(test, iff->in(0)); 736 737 // Replace the old IfNode with a new LoopEndNode 738 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt )); 739 IfNode *le = lex->as_If(); 740 uint dd = dom_depth(iff); 741 set_idom(le, le->in(0), dd); // Update dominance for loop exit 742 set_loop(le, loop); 743 744 // Get the loop-exit control 745 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); 746 747 // Need to swap loop-exit and loop-back control? 748 if (iftrue_op == Op_IfFalse) { 749 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le)); 750 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le)); 751 752 loop->_tail = back_control = ift2; 753 set_loop(ift2, loop); 754 set_loop(iff2, get_loop(iffalse)); 755 756 // Lazy update of 'get_ctrl' mechanism. 757 lazy_replace(iffalse, iff2); 758 lazy_replace(iftrue, ift2); 759 760 // Swap names 761 iffalse = iff2; 762 iftrue = ift2; 763 } else { 764 _igvn.rehash_node_delayed(iffalse); 765 _igvn.rehash_node_delayed(iftrue); 766 iffalse->set_req_X( 0, le, &_igvn ); 767 iftrue ->set_req_X( 0, le, &_igvn ); 768 } 769 770 set_idom(iftrue, le, dd+1); 771 set_idom(iffalse, le, dd+1); 772 assert(iff->outcnt() == 0, "should be dead now"); 773 lazy_replace( iff, le ); // fix 'get_ctrl' 774 775 Node *sfpt2 = le->in(0); 776 777 Node* entry_control = init_control; 778 bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL && 779 sfpt2->Opcode() == Op_SafePoint && !loop->_has_call; 780 IdealLoopTree* outer_ilt = NULL; 781 if (strip_mine_loop) { 782 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop, 783 cl_prob, le->_fcnt, entry_control, 784 iffalse); 785 } 786 787 // Now setup a new CountedLoopNode to replace the existing LoopNode 788 CountedLoopNode *l = new CountedLoopNode(entry_control, back_control); 789 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve 790 // The following assert is approximately true, and defines the intention 791 // of can_be_counted_loop. It fails, however, because phase->type 792 // is not yet initialized for this loop and its parts. 793 //assert(l->can_be_counted_loop(this), "sanity"); 794 _igvn.register_new_node_with_optimizer(l); 795 set_loop(l, loop); 796 loop->_head = l; 797 // Fix all data nodes placed at the old loop head. 798 // Uses the lazy-update mechanism of 'get_ctrl'. 799 lazy_replace( x, l ); 800 set_idom(l, entry_control, dom_depth(entry_control) + 1); 801 802 if (LoopStripMiningIter == 0 || strip_mine_loop) { 803 // Check for immediately preceding SafePoint and remove 804 if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) { 805 if (strip_mine_loop) { 806 Node* outer_le = outer_ilt->_tail->in(0); 807 Node* sfpt = sfpt2->clone(); 808 sfpt->set_req(0, iffalse); 809 outer_le->set_req(0, sfpt); 810 register_control(sfpt, outer_ilt, iffalse); 811 set_idom(outer_le, sfpt, dom_depth(sfpt)); 812 } 813 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); 814 if (loop->_safepts != NULL) { 815 loop->_safepts->yank(sfpt2); 816 } 817 } 818 } 819 820 // Free up intermediate goo 821 _igvn.remove_dead_node(hook); 822 823 #ifdef ASSERT 824 assert(l->is_valid_counted_loop(), "counted loop shape is messed up"); 825 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" ); 826 #endif 827 #ifndef PRODUCT 828 if (TraceLoopOpts) { 829 tty->print("Counted "); 830 loop->dump_head(); 831 } 832 #endif 833 834 C->print_method(PHASE_AFTER_CLOOPS, 3); 835 836 // Capture bounds of the loop in the induction variable Phi before 837 // subsequent transformation (iteration splitting) obscures the 838 // bounds 839 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn)); 840 841 if (strip_mine_loop) { 842 l->mark_strip_mined(); 843 l->verify_strip_mined(1); 844 outer_ilt->_head->as_Loop()->verify_strip_mined(1); 845 loop = outer_ilt; 846 } 847 848 return true; 849 } 850 851 //----------------------exact_limit------------------------------------------- 852 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) { 853 assert(loop->_head->is_CountedLoop(), ""); 854 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 855 assert(cl->is_valid_counted_loop(), ""); 856 857 if (ABS(cl->stride_con()) == 1 || 858 cl->limit()->Opcode() == Op_LoopLimit) { 859 // Old code has exact limit (it could be incorrect in case of int overflow). 860 // Loop limit is exact with stride == 1. And loop may already have exact limit. 861 return cl->limit(); 862 } 863 Node *limit = NULL; 864 #ifdef ASSERT 865 BoolTest::mask bt = cl->loopexit()->test_trip(); 866 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 867 #endif 868 if (cl->has_exact_trip_count()) { 869 // Simple case: loop has constant boundaries. 870 // Use jlongs to avoid integer overflow. 871 int stride_con = cl->stride_con(); 872 jlong init_con = cl->init_trip()->get_int(); 873 jlong limit_con = cl->limit()->get_int(); 874 julong trip_cnt = cl->trip_count(); 875 jlong final_con = init_con + trip_cnt*stride_con; 876 int final_int = (int)final_con; 877 // The final value should be in integer range since the loop 878 // is counted and the limit was checked for overflow. 879 assert(final_con == (jlong)final_int, "final value should be integer"); 880 limit = _igvn.intcon(final_int); 881 } else { 882 // Create new LoopLimit node to get exact limit (final iv value). 883 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride()); 884 register_new_node(limit, cl->in(LoopNode::EntryControl)); 885 } 886 assert(limit != NULL, "sanity"); 887 return limit; 888 } 889 890 //------------------------------Ideal------------------------------------------ 891 // Return a node which is more "ideal" than the current node. 892 // Attempt to convert into a counted-loop. 893 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 894 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) { 895 phase->C->set_major_progress(); 896 } 897 return RegionNode::Ideal(phase, can_reshape); 898 } 899 900 void LoopNode::verify_strip_mined(int expect_skeleton) const { 901 #ifdef ASSERT 902 const OuterStripMinedLoopNode* outer = NULL; 903 const CountedLoopNode* inner = NULL; 904 if (is_strip_mined()) { 905 assert(is_CountedLoop(), "no Loop should be marked strip mined"); 906 inner = as_CountedLoop(); 907 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop(); 908 } else if (is_OuterStripMinedLoop()) { 909 outer = this->as_OuterStripMinedLoop(); 910 inner = outer->unique_ctrl_out()->as_CountedLoop(); 911 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined"); 912 } 913 if (inner != NULL || outer != NULL) { 914 assert(inner != NULL && outer != NULL, "missing loop in strip mined nest"); 915 Node* outer_tail = outer->in(LoopNode::LoopBackControl); 916 Node* outer_le = outer_tail->in(0); 917 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If"); 918 Node* sfpt = outer_le->in(0); 919 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?"); 920 Node* inner_out = sfpt->in(0); 921 if (inner_out->outcnt() != 1) { 922 ResourceMark rm; 923 Unique_Node_List wq; 924 925 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) { 926 Node* u = inner_out->fast_out(i); 927 if (u == sfpt) { 928 continue; 929 } 930 wq.clear(); 931 wq.push(u); 932 bool found_sfpt = false; 933 for (uint next = 0; next < wq.size() && !found_sfpt; next++) { 934 Node *n = wq.at(next); 935 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) { 936 Node* u = n->fast_out(i); 937 if (u == sfpt) { 938 found_sfpt = true; 939 } 940 if (!u->is_CFG()) { 941 wq.push(u); 942 } 943 } 944 } 945 assert(found_sfpt, "no node in loop that's not input to safepoint"); 946 } 947 } 948 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd(); 949 assert(cle == inner->loopexit_or_null(), "mismatch"); 950 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0; 951 if (has_skeleton) { 952 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node"); 953 assert(outer->outcnt() == 2, "only phis"); 954 } else { 955 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?"); 956 uint phis = 0; 957 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) { 958 Node* u = inner->fast_out(i); 959 if (u->is_Phi()) { 960 phis++; 961 } 962 } 963 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) { 964 Node* u = outer->fast_out(i); 965 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop"); 966 } 967 uint stores = 0; 968 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) { 969 Node* u = inner_out->fast_out(i); 970 if (u->is_Store()) { 971 stores++; 972 } 973 } 974 assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis"); 975 } 976 assert(sfpt->outcnt() == 1, "no data node"); 977 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node"); 978 } 979 #endif 980 } 981 982 //============================================================================= 983 //------------------------------Ideal------------------------------------------ 984 // Return a node which is more "ideal" than the current node. 985 // Attempt to convert into a counted-loop. 986 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 987 return RegionNode::Ideal(phase, can_reshape); 988 } 989 990 //------------------------------dump_spec-------------------------------------- 991 // Dump special per-node info 992 #ifndef PRODUCT 993 void CountedLoopNode::dump_spec(outputStream *st) const { 994 LoopNode::dump_spec(st); 995 if (stride_is_con()) { 996 st->print("stride: %d ",stride_con()); 997 } 998 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx); 999 if (is_main_loop()) st->print("main of N%d", _idx); 1000 if (is_post_loop()) st->print("post of N%d", _main_idx); 1001 if (is_strip_mined()) st->print(" strip mined"); 1002 } 1003 #endif 1004 1005 //============================================================================= 1006 int CountedLoopEndNode::stride_con() const { 1007 return stride()->bottom_type()->is_int()->get_con(); 1008 } 1009 1010 //============================================================================= 1011 //------------------------------Value----------------------------------------- 1012 const Type* LoopLimitNode::Value(PhaseGVN* phase) const { 1013 const Type* init_t = phase->type(in(Init)); 1014 const Type* limit_t = phase->type(in(Limit)); 1015 const Type* stride_t = phase->type(in(Stride)); 1016 // Either input is TOP ==> the result is TOP 1017 if (init_t == Type::TOP) return Type::TOP; 1018 if (limit_t == Type::TOP) return Type::TOP; 1019 if (stride_t == Type::TOP) return Type::TOP; 1020 1021 int stride_con = stride_t->is_int()->get_con(); 1022 if (stride_con == 1) 1023 return NULL; // Identity 1024 1025 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) { 1026 // Use jlongs to avoid integer overflow. 1027 jlong init_con = init_t->is_int()->get_con(); 1028 jlong limit_con = limit_t->is_int()->get_con(); 1029 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 1030 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 1031 jlong final_con = init_con + stride_con*trip_count; 1032 int final_int = (int)final_con; 1033 // The final value should be in integer range since the loop 1034 // is counted and the limit was checked for overflow. 1035 assert(final_con == (jlong)final_int, "final value should be integer"); 1036 return TypeInt::make(final_int); 1037 } 1038 1039 return bottom_type(); // TypeInt::INT 1040 } 1041 1042 //------------------------------Ideal------------------------------------------ 1043 // Return a node which is more "ideal" than the current node. 1044 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1045 if (phase->type(in(Init)) == Type::TOP || 1046 phase->type(in(Limit)) == Type::TOP || 1047 phase->type(in(Stride)) == Type::TOP) 1048 return NULL; // Dead 1049 1050 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1051 if (stride_con == 1) 1052 return NULL; // Identity 1053 1054 if (in(Init)->is_Con() && in(Limit)->is_Con()) 1055 return NULL; // Value 1056 1057 // Delay following optimizations until all loop optimizations 1058 // done to keep Ideal graph simple. 1059 if (!can_reshape || phase->C->major_progress()) 1060 return NULL; 1061 1062 const TypeInt* init_t = phase->type(in(Init) )->is_int(); 1063 const TypeInt* limit_t = phase->type(in(Limit))->is_int(); 1064 int stride_p; 1065 jlong lim, ini; 1066 julong max; 1067 if (stride_con > 0) { 1068 stride_p = stride_con; 1069 lim = limit_t->_hi; 1070 ini = init_t->_lo; 1071 max = (julong)max_jint; 1072 } else { 1073 stride_p = -stride_con; 1074 lim = init_t->_hi; 1075 ini = limit_t->_lo; 1076 max = (julong)min_jint; 1077 } 1078 julong range = lim - ini + stride_p; 1079 if (range <= max) { 1080 // Convert to integer expression if it is not overflow. 1081 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1)); 1082 Node *range = phase->transform(new SubINode(in(Limit), in(Init))); 1083 Node *bias = phase->transform(new AddINode(range, stride_m)); 1084 Node *trip = phase->transform(new DivINode(0, bias, in(Stride))); 1085 Node *span = phase->transform(new MulINode(trip, in(Stride))); 1086 return new AddINode(span, in(Init)); // exact limit 1087 } 1088 1089 if (is_power_of_2(stride_p) || // divisor is 2^n 1090 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node? 1091 // Convert to long expression to avoid integer overflow 1092 // and let igvn optimizer convert this division. 1093 // 1094 Node* init = phase->transform( new ConvI2LNode(in(Init))); 1095 Node* limit = phase->transform( new ConvI2LNode(in(Limit))); 1096 Node* stride = phase->longcon(stride_con); 1097 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1)); 1098 1099 Node *range = phase->transform(new SubLNode(limit, init)); 1100 Node *bias = phase->transform(new AddLNode(range, stride_m)); 1101 Node *span; 1102 if (stride_con > 0 && is_power_of_2(stride_p)) { 1103 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride) 1104 // and avoid generating rounding for division. Zero trip guard should 1105 // guarantee that init < limit but sometimes the guard is missing and 1106 // we can get situation when init > limit. Note, for the empty loop 1107 // optimization zero trip guard is generated explicitly which leaves 1108 // only RCE predicate where exact limit is used and the predicate 1109 // will simply fail forcing recompilation. 1110 Node* neg_stride = phase->longcon(-stride_con); 1111 span = phase->transform(new AndLNode(bias, neg_stride)); 1112 } else { 1113 Node *trip = phase->transform(new DivLNode(0, bias, stride)); 1114 span = phase->transform(new MulLNode(trip, stride)); 1115 } 1116 // Convert back to int 1117 Node *span_int = phase->transform(new ConvL2INode(span)); 1118 return new AddINode(span_int, in(Init)); // exact limit 1119 } 1120 1121 return NULL; // No progress 1122 } 1123 1124 //------------------------------Identity--------------------------------------- 1125 // If stride == 1 return limit node. 1126 Node* LoopLimitNode::Identity(PhaseGVN* phase) { 1127 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1128 if (stride_con == 1 || stride_con == -1) 1129 return in(Limit); 1130 return this; 1131 } 1132 1133 //============================================================================= 1134 //----------------------match_incr_with_optional_truncation-------------------- 1135 // Match increment with optional truncation: 1136 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 1137 // Return NULL for failure. Success returns the increment node. 1138 Node* CountedLoopNode::match_incr_with_optional_truncation( 1139 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { 1140 // Quick cutouts: 1141 if (expr == NULL || expr->req() != 3) return NULL; 1142 1143 Node *t1 = NULL; 1144 Node *t2 = NULL; 1145 const TypeInt* trunc_t = TypeInt::INT; 1146 Node* n1 = expr; 1147 int n1op = n1->Opcode(); 1148 1149 // Try to strip (n1 & M) or (n1 << N >> N) from n1. 1150 if (n1op == Op_AndI && 1151 n1->in(2)->is_Con() && 1152 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { 1153 // %%% This check should match any mask of 2**K-1. 1154 t1 = n1; 1155 n1 = t1->in(1); 1156 n1op = n1->Opcode(); 1157 trunc_t = TypeInt::CHAR; 1158 } else if (n1op == Op_RShiftI && 1159 n1->in(1) != NULL && 1160 n1->in(1)->Opcode() == Op_LShiftI && 1161 n1->in(2) == n1->in(1)->in(2) && 1162 n1->in(2)->is_Con()) { 1163 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); 1164 // %%% This check should match any shift in [1..31]. 1165 if (shift == 16 || shift == 8) { 1166 t1 = n1; 1167 t2 = t1->in(1); 1168 n1 = t2->in(1); 1169 n1op = n1->Opcode(); 1170 if (shift == 16) { 1171 trunc_t = TypeInt::SHORT; 1172 } else if (shift == 8) { 1173 trunc_t = TypeInt::BYTE; 1174 } 1175 } 1176 } 1177 1178 // If (maybe after stripping) it is an AddI, we won: 1179 if (n1op == Op_AddI) { 1180 *trunc1 = t1; 1181 *trunc2 = t2; 1182 *trunc_type = trunc_t; 1183 return n1; 1184 } 1185 1186 // failed 1187 return NULL; 1188 } 1189 1190 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) { 1191 if (is_strip_mined()) { 1192 verify_strip_mined(expect_skeleton); 1193 return in(EntryControl)->as_Loop(); 1194 } 1195 return this; 1196 } 1197 1198 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const { 1199 assert(is_strip_mined(), "not a strip mined loop"); 1200 Node* c = in(EntryControl); 1201 if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) { 1202 return NULL; 1203 } 1204 return c->as_OuterStripMinedLoop(); 1205 } 1206 1207 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const { 1208 Node* c = in(LoopBackControl); 1209 if (c == NULL || c->is_top()) { 1210 return NULL; 1211 } 1212 return c->as_IfTrue(); 1213 } 1214 1215 IfTrueNode* CountedLoopNode::outer_loop_tail() const { 1216 LoopNode* l = outer_loop(); 1217 if (l == NULL) { 1218 return NULL; 1219 } 1220 return l->outer_loop_tail(); 1221 } 1222 1223 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const { 1224 IfTrueNode* proj = outer_loop_tail(); 1225 if (proj == NULL) { 1226 return NULL; 1227 } 1228 Node* c = proj->in(0); 1229 if (c == NULL || c->is_top() || c->outcnt() != 2) { 1230 return NULL; 1231 } 1232 return c->as_OuterStripMinedLoopEnd(); 1233 } 1234 1235 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const { 1236 LoopNode* l = outer_loop(); 1237 if (l == NULL) { 1238 return NULL; 1239 } 1240 return l->outer_loop_end(); 1241 } 1242 1243 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const { 1244 IfNode* le = outer_loop_end(); 1245 if (le == NULL) { 1246 return NULL; 1247 } 1248 Node* c = le->proj_out_or_null(false); 1249 if (c == NULL) { 1250 return NULL; 1251 } 1252 return c->as_IfFalse(); 1253 } 1254 1255 IfFalseNode* CountedLoopNode::outer_loop_exit() const { 1256 LoopNode* l = outer_loop(); 1257 if (l == NULL) { 1258 return NULL; 1259 } 1260 return l->outer_loop_exit(); 1261 } 1262 1263 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const { 1264 IfNode* le = outer_loop_end(); 1265 if (le == NULL) { 1266 return NULL; 1267 } 1268 Node* c = le->in(0); 1269 if (c == NULL || c->is_top()) { 1270 return NULL; 1271 } 1272 assert(c->Opcode() == Op_SafePoint, "broken outer loop"); 1273 return c->as_SafePoint(); 1274 } 1275 1276 SafePointNode* CountedLoopNode::outer_safepoint() const { 1277 LoopNode* l = outer_loop(); 1278 if (l == NULL) { 1279 return NULL; 1280 } 1281 return l->outer_safepoint(); 1282 } 1283 1284 Node* CountedLoopNode::skip_predicates() { 1285 if (is_main_loop()) { 1286 Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl); 1287 while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() && 1288 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 && 1289 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) { 1290 ctrl = ctrl->in(0)->in(0); 1291 } 1292 1293 return ctrl; 1294 } 1295 return in(LoopNode::EntryControl); 1296 } 1297 1298 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) { 1299 // Look for the outer & inner strip mined loop, reduce number of 1300 // iterations of the inner loop, set exit condition of outer loop, 1301 // construct required phi nodes for outer loop. 1302 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop(); 1303 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined"); 1304 Node* inner_iv_phi = inner_cl->phi(); 1305 if (inner_iv_phi == NULL) { 1306 IfNode* outer_le = outer_loop_end(); 1307 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1308 igvn->replace_node(outer_le, iff); 1309 inner_cl->clear_strip_mined(); 1310 return; 1311 } 1312 CountedLoopEndNode* inner_cle = inner_cl->loopexit(); 1313 1314 int stride = inner_cl->stride_con(); 1315 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride); 1316 int scaled_iters = (int)scaled_iters_long; 1317 int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride); 1318 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int(); 1319 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo; 1320 assert(iter_estimate > 0, "broken"); 1321 if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) { 1322 // Remove outer loop and safepoint (too few iterations) 1323 Node* outer_sfpt = outer_safepoint(); 1324 Node* outer_out = outer_loop_exit(); 1325 igvn->replace_node(outer_out, outer_sfpt->in(0)); 1326 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top()); 1327 inner_cl->clear_strip_mined(); 1328 return; 1329 } 1330 if (iter_estimate <= scaled_iters_long) { 1331 // We would only go through one iteration of 1332 // the outer loop: drop the outer loop but 1333 // keep the safepoint so we don't run for 1334 // too long without a safepoint 1335 IfNode* outer_le = outer_loop_end(); 1336 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1337 igvn->replace_node(outer_le, iff); 1338 inner_cl->clear_strip_mined(); 1339 return; 1340 } 1341 1342 Node* cle_tail = inner_cle->proj_out(true); 1343 ResourceMark rm; 1344 Node_List old_new; 1345 if (cle_tail->outcnt() > 1) { 1346 // Look for nodes on backedge of inner loop and clone them 1347 Unique_Node_List backedge_nodes; 1348 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) { 1349 Node* u = cle_tail->fast_out(i); 1350 if (u != inner_cl) { 1351 assert(!u->is_CFG(), "control flow on the backedge?"); 1352 backedge_nodes.push(u); 1353 } 1354 } 1355 uint last = igvn->C->unique(); 1356 for (uint next = 0; next < backedge_nodes.size(); next++) { 1357 Node* n = backedge_nodes.at(next); 1358 old_new.map(n->_idx, n->clone()); 1359 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1360 Node* u = n->fast_out(i); 1361 assert(!u->is_CFG(), "broken"); 1362 if (u->_idx >= last) { 1363 continue; 1364 } 1365 if (!u->is_Phi()) { 1366 backedge_nodes.push(u); 1367 } else { 1368 assert(u->in(0) == inner_cl, "strange phi on the backedge"); 1369 } 1370 } 1371 } 1372 // Put the clones on the outer loop backedge 1373 Node* le_tail = outer_loop_tail(); 1374 for (uint next = 0; next < backedge_nodes.size(); next++) { 1375 Node *n = old_new[backedge_nodes.at(next)->_idx]; 1376 for (uint i = 1; i < n->req(); i++) { 1377 if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) { 1378 n->set_req(i, old_new[n->in(i)->_idx]); 1379 } 1380 } 1381 if (n->in(0) != NULL && n->in(0) == cle_tail) { 1382 n->set_req(0, le_tail); 1383 } 1384 igvn->register_new_node_with_optimizer(n); 1385 } 1386 } 1387 1388 Node* iv_phi = NULL; 1389 // Make a clone of each phi in the inner loop 1390 // for the outer loop 1391 for (uint i = 0; i < inner_cl->outcnt(); i++) { 1392 Node* u = inner_cl->raw_out(i); 1393 if (u->is_Phi()) { 1394 assert(u->in(0) == inner_cl, "inconsistent"); 1395 Node* phi = u->clone(); 1396 phi->set_req(0, this); 1397 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx]; 1398 if (be != NULL) { 1399 phi->set_req(LoopNode::LoopBackControl, be); 1400 } 1401 phi = igvn->transform(phi); 1402 igvn->replace_input_of(u, LoopNode::EntryControl, phi); 1403 if (u == inner_iv_phi) { 1404 iv_phi = phi; 1405 } 1406 } 1407 } 1408 Node* cle_out = inner_cle->proj_out(false); 1409 if (cle_out->outcnt() > 1) { 1410 // Look for chains of stores that were sunk 1411 // out of the inner loop and are in the outer loop 1412 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) { 1413 Node* u = cle_out->fast_out(i); 1414 if (u->is_Store()) { 1415 Node* first = u; 1416 for(;;) { 1417 Node* next = first->in(MemNode::Memory); 1418 if (!next->is_Store() || next->in(0) != cle_out) { 1419 break; 1420 } 1421 first = next; 1422 } 1423 Node* last = u; 1424 for(;;) { 1425 Node* next = NULL; 1426 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) { 1427 Node* uu = last->fast_out(j); 1428 if (uu->is_Store() && uu->in(0) == cle_out) { 1429 assert(next == NULL, "only one in the outer loop"); 1430 next = uu; 1431 } 1432 } 1433 if (next == NULL) { 1434 break; 1435 } 1436 last = next; 1437 } 1438 Node* phi = NULL; 1439 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1440 Node* uu = fast_out(j); 1441 if (uu->is_Phi()) { 1442 Node* be = uu->in(LoopNode::LoopBackControl); 1443 if (be->is_Store() && old_new[be->_idx] != NULL) { 1444 assert(false, "store on the backedge + sunk stores: unsupported"); 1445 // drop outer loop 1446 IfNode* outer_le = outer_loop_end(); 1447 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1448 igvn->replace_node(outer_le, iff); 1449 inner_cl->clear_strip_mined(); 1450 return; 1451 } 1452 if (be == last || be == first->in(MemNode::Memory)) { 1453 assert(phi == NULL, "only one phi"); 1454 phi = uu; 1455 } 1456 } 1457 } 1458 #ifdef ASSERT 1459 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1460 Node* uu = fast_out(j); 1461 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) { 1462 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) { 1463 assert(phi == uu, "what's that phi?"); 1464 } else if (uu->adr_type() == TypePtr::BOTTOM) { 1465 Node* n = uu->in(LoopNode::LoopBackControl); 1466 uint limit = igvn->C->live_nodes(); 1467 uint i = 0; 1468 while (n != uu) { 1469 i++; 1470 assert(i < limit, "infinite loop"); 1471 if (n->is_Proj()) { 1472 n = n->in(0); 1473 } else if (n->is_SafePoint() || n->is_MemBar()) { 1474 n = n->in(TypeFunc::Memory); 1475 } else if (n->is_Phi()) { 1476 n = n->in(1); 1477 } else if (n->is_MergeMem()) { 1478 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type())); 1479 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) { 1480 n = n->in(MemNode::Memory); 1481 } else { 1482 n->dump(); 1483 ShouldNotReachHere(); 1484 } 1485 } 1486 } 1487 } 1488 } 1489 #endif 1490 if (phi == NULL) { 1491 // If the an entire chains was sunk, the 1492 // inner loop has no phi for that memory 1493 // slice, create one for the outer loop 1494 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY, 1495 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))); 1496 phi->set_req(LoopNode::LoopBackControl, last); 1497 phi = igvn->transform(phi); 1498 igvn->replace_input_of(first, MemNode::Memory, phi); 1499 } else { 1500 // Or fix the outer loop fix to include 1501 // that chain of stores. 1502 Node* be = phi->in(LoopNode::LoopBackControl); 1503 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported"); 1504 if (be == first->in(MemNode::Memory)) { 1505 if (be == phi->in(LoopNode::LoopBackControl)) { 1506 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last); 1507 } else { 1508 igvn->replace_input_of(be, MemNode::Memory, last); 1509 } 1510 } else { 1511 #ifdef ASSERT 1512 if (be == phi->in(LoopNode::LoopBackControl)) { 1513 assert(phi->in(LoopNode::LoopBackControl) == last, ""); 1514 } else { 1515 assert(be->in(MemNode::Memory) == last, ""); 1516 } 1517 #endif 1518 } 1519 } 1520 } 1521 } 1522 } 1523 1524 if (iv_phi != NULL) { 1525 // Now adjust the inner loop's exit condition 1526 Node* limit = inner_cl->limit(); 1527 Node* sub = NULL; 1528 if (stride > 0) { 1529 sub = igvn->transform(new SubINode(limit, iv_phi)); 1530 } else { 1531 sub = igvn->transform(new SubINode(iv_phi, limit)); 1532 } 1533 Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters))); 1534 Node* new_limit = NULL; 1535 if (stride > 0) { 1536 new_limit = igvn->transform(new AddINode(min, iv_phi)); 1537 } else { 1538 new_limit = igvn->transform(new SubINode(iv_phi, min)); 1539 } 1540 Node* cmp = inner_cle->cmp_node()->clone(); 1541 igvn->replace_input_of(cmp, 2, new_limit); 1542 Node* bol = inner_cle->in(CountedLoopEndNode::TestValue)->clone(); 1543 cmp->set_req(2, limit); 1544 bol->set_req(1, igvn->transform(cmp)); 1545 igvn->replace_input_of(outer_loop_end(), 1, igvn->transform(bol)); 1546 } else { 1547 assert(false, "should be able to adjust outer loop"); 1548 IfNode* outer_le = outer_loop_end(); 1549 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1550 igvn->replace_node(outer_le, iff); 1551 inner_cl->clear_strip_mined(); 1552 } 1553 } 1554 1555 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const { 1556 if (!in(0)) return Type::TOP; 1557 if (phase->type(in(0)) == Type::TOP) 1558 return Type::TOP; 1559 1560 return TypeTuple::IFBOTH; 1561 } 1562 1563 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1564 if (remove_dead_region(phase, can_reshape)) return this; 1565 1566 return NULL; 1567 } 1568 1569 //------------------------------filtered_type-------------------------------- 1570 // Return a type based on condition control flow 1571 // A successful return will be a type that is restricted due 1572 // to a series of dominating if-tests, such as: 1573 // if (i < 10) { 1574 // if (i > 0) { 1575 // here: "i" type is [1..10) 1576 // } 1577 // } 1578 // or a control flow merge 1579 // if (i < 10) { 1580 // do { 1581 // phi( , ) -- at top of loop type is [min_int..10) 1582 // i = ? 1583 // } while ( i < 10) 1584 // 1585 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { 1586 assert(n && n->bottom_type()->is_int(), "must be int"); 1587 const TypeInt* filtered_t = NULL; 1588 if (!n->is_Phi()) { 1589 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); 1590 filtered_t = filtered_type_from_dominators(n, n_ctrl); 1591 1592 } else { 1593 Node* phi = n->as_Phi(); 1594 Node* region = phi->in(0); 1595 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); 1596 if (region && region != C->top()) { 1597 for (uint i = 1; i < phi->req(); i++) { 1598 Node* val = phi->in(i); 1599 Node* use_c = region->in(i); 1600 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); 1601 if (val_t != NULL) { 1602 if (filtered_t == NULL) { 1603 filtered_t = val_t; 1604 } else { 1605 filtered_t = filtered_t->meet(val_t)->is_int(); 1606 } 1607 } 1608 } 1609 } 1610 } 1611 const TypeInt* n_t = _igvn.type(n)->is_int(); 1612 if (filtered_t != NULL) { 1613 n_t = n_t->join(filtered_t)->is_int(); 1614 } 1615 return n_t; 1616 } 1617 1618 1619 //------------------------------filtered_type_from_dominators-------------------------------- 1620 // Return a possibly more restrictive type for val based on condition control flow of dominators 1621 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { 1622 if (val->is_Con()) { 1623 return val->bottom_type()->is_int(); 1624 } 1625 uint if_limit = 10; // Max number of dominating if's visited 1626 const TypeInt* rtn_t = NULL; 1627 1628 if (use_ctrl && use_ctrl != C->top()) { 1629 Node* val_ctrl = get_ctrl(val); 1630 uint val_dom_depth = dom_depth(val_ctrl); 1631 Node* pred = use_ctrl; 1632 uint if_cnt = 0; 1633 while (if_cnt < if_limit) { 1634 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { 1635 if_cnt++; 1636 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); 1637 if (if_t != NULL) { 1638 if (rtn_t == NULL) { 1639 rtn_t = if_t; 1640 } else { 1641 rtn_t = rtn_t->join(if_t)->is_int(); 1642 } 1643 } 1644 } 1645 pred = idom(pred); 1646 if (pred == NULL || pred == C->top()) { 1647 break; 1648 } 1649 // Stop if going beyond definition block of val 1650 if (dom_depth(pred) < val_dom_depth) { 1651 break; 1652 } 1653 } 1654 } 1655 return rtn_t; 1656 } 1657 1658 1659 //------------------------------dump_spec-------------------------------------- 1660 // Dump special per-node info 1661 #ifndef PRODUCT 1662 void CountedLoopEndNode::dump_spec(outputStream *st) const { 1663 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) { 1664 BoolTest bt( test_trip()); // Added this for g++. 1665 1666 st->print("["); 1667 bt.dump_on(st); 1668 st->print("]"); 1669 } 1670 st->print(" "); 1671 IfNode::dump_spec(st); 1672 } 1673 #endif 1674 1675 //============================================================================= 1676 //------------------------------is_member-------------------------------------- 1677 // Is 'l' a member of 'this'? 1678 bool IdealLoopTree::is_member(const IdealLoopTree *l) const { 1679 while( l->_nest > _nest ) l = l->_parent; 1680 return l == this; 1681 } 1682 1683 //------------------------------set_nest--------------------------------------- 1684 // Set loop tree nesting depth. Accumulate _has_call bits. 1685 int IdealLoopTree::set_nest( uint depth ) { 1686 _nest = depth; 1687 int bits = _has_call; 1688 if( _child ) bits |= _child->set_nest(depth+1); 1689 if( bits ) _has_call = 1; 1690 if( _next ) bits |= _next ->set_nest(depth ); 1691 return bits; 1692 } 1693 1694 //------------------------------split_fall_in---------------------------------- 1695 // Split out multiple fall-in edges from the loop header. Move them to a 1696 // private RegionNode before the loop. This becomes the loop landing pad. 1697 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { 1698 PhaseIterGVN &igvn = phase->_igvn; 1699 uint i; 1700 1701 // Make a new RegionNode to be the landing pad. 1702 Node *landing_pad = new RegionNode( fall_in_cnt+1 ); 1703 phase->set_loop(landing_pad,_parent); 1704 // Gather all the fall-in control paths into the landing pad 1705 uint icnt = fall_in_cnt; 1706 uint oreq = _head->req(); 1707 for( i = oreq-1; i>0; i-- ) 1708 if( !phase->is_member( this, _head->in(i) ) ) 1709 landing_pad->set_req(icnt--,_head->in(i)); 1710 1711 // Peel off PhiNode edges as well 1712 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1713 Node *oj = _head->fast_out(j); 1714 if( oj->is_Phi() ) { 1715 PhiNode* old_phi = oj->as_Phi(); 1716 assert( old_phi->region() == _head, "" ); 1717 igvn.hash_delete(old_phi); // Yank from hash before hacking edges 1718 Node *p = PhiNode::make_blank(landing_pad, old_phi); 1719 uint icnt = fall_in_cnt; 1720 for( i = oreq-1; i>0; i-- ) { 1721 if( !phase->is_member( this, _head->in(i) ) ) { 1722 p->init_req(icnt--, old_phi->in(i)); 1723 // Go ahead and clean out old edges from old phi 1724 old_phi->del_req(i); 1725 } 1726 } 1727 // Search for CSE's here, because ZKM.jar does a lot of 1728 // loop hackery and we need to be a little incremental 1729 // with the CSE to avoid O(N^2) node blow-up. 1730 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE 1731 if( p2 ) { // Found CSE 1732 p->destruct(); // Recover useless new node 1733 p = p2; // Use old node 1734 } else { 1735 igvn.register_new_node_with_optimizer(p, old_phi); 1736 } 1737 // Make old Phi refer to new Phi. 1738 old_phi->add_req(p); 1739 // Check for the special case of making the old phi useless and 1740 // disappear it. In JavaGrande I have a case where this useless 1741 // Phi is the loop limit and prevents recognizing a CountedLoop 1742 // which in turn prevents removing an empty loop. 1743 Node *id_old_phi = old_phi->Identity( &igvn ); 1744 if( id_old_phi != old_phi ) { // Found a simple identity? 1745 // Note that I cannot call 'replace_node' here, because 1746 // that will yank the edge from old_phi to the Region and 1747 // I'm mid-iteration over the Region's uses. 1748 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { 1749 Node* use = old_phi->last_out(i); 1750 igvn.rehash_node_delayed(use); 1751 uint uses_found = 0; 1752 for (uint j = 0; j < use->len(); j++) { 1753 if (use->in(j) == old_phi) { 1754 if (j < use->req()) use->set_req (j, id_old_phi); 1755 else use->set_prec(j, id_old_phi); 1756 uses_found++; 1757 } 1758 } 1759 i -= uses_found; // we deleted 1 or more copies of this edge 1760 } 1761 } 1762 igvn._worklist.push(old_phi); 1763 } 1764 } 1765 // Finally clean out the fall-in edges from the RegionNode 1766 for( i = oreq-1; i>0; i-- ) { 1767 if( !phase->is_member( this, _head->in(i) ) ) { 1768 _head->del_req(i); 1769 } 1770 } 1771 igvn.rehash_node_delayed(_head); 1772 // Transform landing pad 1773 igvn.register_new_node_with_optimizer(landing_pad, _head); 1774 // Insert landing pad into the header 1775 _head->add_req(landing_pad); 1776 } 1777 1778 //------------------------------split_outer_loop------------------------------- 1779 // Split out the outermost loop from this shared header. 1780 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { 1781 PhaseIterGVN &igvn = phase->_igvn; 1782 1783 // Find index of outermost loop; it should also be my tail. 1784 uint outer_idx = 1; 1785 while( _head->in(outer_idx) != _tail ) outer_idx++; 1786 1787 // Make a LoopNode for the outermost loop. 1788 Node *ctl = _head->in(LoopNode::EntryControl); 1789 Node *outer = new LoopNode( ctl, _head->in(outer_idx) ); 1790 outer = igvn.register_new_node_with_optimizer(outer, _head); 1791 phase->set_created_loop_node(); 1792 1793 // Outermost loop falls into '_head' loop 1794 _head->set_req(LoopNode::EntryControl, outer); 1795 _head->del_req(outer_idx); 1796 // Split all the Phis up between '_head' loop and 'outer' loop. 1797 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1798 Node *out = _head->fast_out(j); 1799 if( out->is_Phi() ) { 1800 PhiNode *old_phi = out->as_Phi(); 1801 assert( old_phi->region() == _head, "" ); 1802 Node *phi = PhiNode::make_blank(outer, old_phi); 1803 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); 1804 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); 1805 phi = igvn.register_new_node_with_optimizer(phi, old_phi); 1806 // Make old Phi point to new Phi on the fall-in path 1807 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi); 1808 old_phi->del_req(outer_idx); 1809 } 1810 } 1811 1812 // Use the new loop head instead of the old shared one 1813 _head = outer; 1814 phase->set_loop(_head, this); 1815 } 1816 1817 //------------------------------fix_parent------------------------------------- 1818 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { 1819 loop->_parent = parent; 1820 if( loop->_child ) fix_parent( loop->_child, loop ); 1821 if( loop->_next ) fix_parent( loop->_next , parent ); 1822 } 1823 1824 //------------------------------estimate_path_freq----------------------------- 1825 static float estimate_path_freq( Node *n ) { 1826 // Try to extract some path frequency info 1827 IfNode *iff; 1828 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests 1829 uint nop = n->Opcode(); 1830 if( nop == Op_SafePoint ) { // Skip any safepoint 1831 n = n->in(0); 1832 continue; 1833 } 1834 if( nop == Op_CatchProj ) { // Get count from a prior call 1835 // Assume call does not always throw exceptions: means the call-site 1836 // count is also the frequency of the fall-through path. 1837 assert( n->is_CatchProj(), "" ); 1838 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) 1839 return 0.0f; // Assume call exception path is rare 1840 Node *call = n->in(0)->in(0)->in(0); 1841 assert( call->is_Call(), "expect a call here" ); 1842 const JVMState *jvms = ((CallNode*)call)->jvms(); 1843 ciMethodData* methodData = jvms->method()->method_data(); 1844 if (!methodData->is_mature()) return 0.0f; // No call-site data 1845 ciProfileData* data = methodData->bci_to_data(jvms->bci()); 1846 if ((data == NULL) || !data->is_CounterData()) { 1847 // no call profile available, try call's control input 1848 n = n->in(0); 1849 continue; 1850 } 1851 return data->as_CounterData()->count()/FreqCountInvocations; 1852 } 1853 // See if there's a gating IF test 1854 Node *n_c = n->in(0); 1855 if( !n_c->is_If() ) break; // No estimate available 1856 iff = n_c->as_If(); 1857 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? 1858 // Compute how much count comes on this path 1859 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; 1860 // Have no count info. Skip dull uncommon-trap like branches. 1861 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || 1862 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) 1863 break; 1864 // Skip through never-taken branch; look for a real loop exit. 1865 n = iff->in(0); 1866 } 1867 return 0.0f; // No estimate available 1868 } 1869 1870 //------------------------------merge_many_backedges--------------------------- 1871 // Merge all the backedges from the shared header into a private Region. 1872 // Feed that region as the one backedge to this loop. 1873 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { 1874 uint i; 1875 1876 // Scan for the top 2 hottest backedges 1877 float hotcnt = 0.0f; 1878 float warmcnt = 0.0f; 1879 uint hot_idx = 0; 1880 // Loop starts at 2 because slot 1 is the fall-in path 1881 for( i = 2; i < _head->req(); i++ ) { 1882 float cnt = estimate_path_freq(_head->in(i)); 1883 if( cnt > hotcnt ) { // Grab hottest path 1884 warmcnt = hotcnt; 1885 hotcnt = cnt; 1886 hot_idx = i; 1887 } else if( cnt > warmcnt ) { // And 2nd hottest path 1888 warmcnt = cnt; 1889 } 1890 } 1891 1892 // See if the hottest backedge is worthy of being an inner loop 1893 // by being much hotter than the next hottest backedge. 1894 if( hotcnt <= 0.0001 || 1895 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge 1896 1897 // Peel out the backedges into a private merge point; peel 1898 // them all except optionally hot_idx. 1899 PhaseIterGVN &igvn = phase->_igvn; 1900 1901 Node *hot_tail = NULL; 1902 // Make a Region for the merge point 1903 Node *r = new RegionNode(1); 1904 for( i = 2; i < _head->req(); i++ ) { 1905 if( i != hot_idx ) 1906 r->add_req( _head->in(i) ); 1907 else hot_tail = _head->in(i); 1908 } 1909 igvn.register_new_node_with_optimizer(r, _head); 1910 // Plug region into end of loop _head, followed by hot_tail 1911 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); 1912 igvn.replace_input_of(_head, 2, r); 1913 if( hot_idx ) _head->add_req(hot_tail); 1914 1915 // Split all the Phis up between '_head' loop and the Region 'r' 1916 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1917 Node *out = _head->fast_out(j); 1918 if( out->is_Phi() ) { 1919 PhiNode* n = out->as_Phi(); 1920 igvn.hash_delete(n); // Delete from hash before hacking edges 1921 Node *hot_phi = NULL; 1922 Node *phi = new PhiNode(r, n->type(), n->adr_type()); 1923 // Check all inputs for the ones to peel out 1924 uint j = 1; 1925 for( uint i = 2; i < n->req(); i++ ) { 1926 if( i != hot_idx ) 1927 phi->set_req( j++, n->in(i) ); 1928 else hot_phi = n->in(i); 1929 } 1930 // Register the phi but do not transform until whole place transforms 1931 igvn.register_new_node_with_optimizer(phi, n); 1932 // Add the merge phi to the old Phi 1933 while( n->req() > 3 ) n->del_req( n->req()-1 ); 1934 igvn.replace_input_of(n, 2, phi); 1935 if( hot_idx ) n->add_req(hot_phi); 1936 } 1937 } 1938 1939 1940 // Insert a new IdealLoopTree inserted below me. Turn it into a clone 1941 // of self loop tree. Turn self into a loop headed by _head and with 1942 // tail being the new merge point. 1943 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); 1944 phase->set_loop(_tail,ilt); // Adjust tail 1945 _tail = r; // Self's tail is new merge point 1946 phase->set_loop(r,this); 1947 ilt->_child = _child; // New guy has my children 1948 _child = ilt; // Self has new guy as only child 1949 ilt->_parent = this; // new guy has self for parent 1950 ilt->_nest = _nest; // Same nesting depth (for now) 1951 1952 // Starting with 'ilt', look for child loop trees using the same shared 1953 // header. Flatten these out; they will no longer be loops in the end. 1954 IdealLoopTree **pilt = &_child; 1955 while( ilt ) { 1956 if( ilt->_head == _head ) { 1957 uint i; 1958 for( i = 2; i < _head->req(); i++ ) 1959 if( _head->in(i) == ilt->_tail ) 1960 break; // Still a loop 1961 if( i == _head->req() ) { // No longer a loop 1962 // Flatten ilt. Hang ilt's "_next" list from the end of 1963 // ilt's '_child' list. Move the ilt's _child up to replace ilt. 1964 IdealLoopTree **cp = &ilt->_child; 1965 while( *cp ) cp = &(*cp)->_next; // Find end of child list 1966 *cp = ilt->_next; // Hang next list at end of child list 1967 *pilt = ilt->_child; // Move child up to replace ilt 1968 ilt->_head = NULL; // Flag as a loop UNIONED into parent 1969 ilt = ilt->_child; // Repeat using new ilt 1970 continue; // do not advance over ilt->_child 1971 } 1972 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); 1973 phase->set_loop(_head,ilt); 1974 } 1975 pilt = &ilt->_child; // Advance to next 1976 ilt = *pilt; 1977 } 1978 1979 if( _child ) fix_parent( _child, this ); 1980 } 1981 1982 //------------------------------beautify_loops--------------------------------- 1983 // Split shared headers and insert loop landing pads. 1984 // Insert a LoopNode to replace the RegionNode. 1985 // Return TRUE if loop tree is structurally changed. 1986 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { 1987 bool result = false; 1988 // Cache parts in locals for easy 1989 PhaseIterGVN &igvn = phase->_igvn; 1990 1991 igvn.hash_delete(_head); // Yank from hash before hacking edges 1992 1993 // Check for multiple fall-in paths. Peel off a landing pad if need be. 1994 int fall_in_cnt = 0; 1995 for( uint i = 1; i < _head->req(); i++ ) 1996 if( !phase->is_member( this, _head->in(i) ) ) 1997 fall_in_cnt++; 1998 assert( fall_in_cnt, "at least 1 fall-in path" ); 1999 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins 2000 split_fall_in( phase, fall_in_cnt ); 2001 2002 // Swap inputs to the _head and all Phis to move the fall-in edge to 2003 // the left. 2004 fall_in_cnt = 1; 2005 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) 2006 fall_in_cnt++; 2007 if( fall_in_cnt > 1 ) { 2008 // Since I am just swapping inputs I do not need to update def-use info 2009 Node *tmp = _head->in(1); 2010 igvn.rehash_node_delayed(_head); 2011 _head->set_req( 1, _head->in(fall_in_cnt) ); 2012 _head->set_req( fall_in_cnt, tmp ); 2013 // Swap also all Phis 2014 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { 2015 Node* phi = _head->fast_out(i); 2016 if( phi->is_Phi() ) { 2017 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges 2018 tmp = phi->in(1); 2019 phi->set_req( 1, phi->in(fall_in_cnt) ); 2020 phi->set_req( fall_in_cnt, tmp ); 2021 } 2022 } 2023 } 2024 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); 2025 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); 2026 2027 // If I am a shared header (multiple backedges), peel off the many 2028 // backedges into a private merge point and use the merge point as 2029 // the one true backedge. 2030 if( _head->req() > 3 ) { 2031 // Merge the many backedges into a single backedge but leave 2032 // the hottest backedge as separate edge for the following peel. 2033 merge_many_backedges( phase ); 2034 result = true; 2035 } 2036 2037 // If I have one hot backedge, peel off myself loop. 2038 // I better be the outermost loop. 2039 if (_head->req() > 3 && !_irreducible) { 2040 split_outer_loop( phase ); 2041 result = true; 2042 2043 } else if (!_head->is_Loop() && !_irreducible) { 2044 // Make a new LoopNode to replace the old loop head 2045 Node *l = new LoopNode( _head->in(1), _head->in(2) ); 2046 l = igvn.register_new_node_with_optimizer(l, _head); 2047 phase->set_created_loop_node(); 2048 // Go ahead and replace _head 2049 phase->_igvn.replace_node( _head, l ); 2050 _head = l; 2051 phase->set_loop(_head, this); 2052 } 2053 2054 // Now recursively beautify nested loops 2055 if( _child ) result |= _child->beautify_loops( phase ); 2056 if( _next ) result |= _next ->beautify_loops( phase ); 2057 return result; 2058 } 2059 2060 //------------------------------allpaths_check_safepts---------------------------- 2061 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 2062 // encountered. Helper for check_safepts. 2063 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { 2064 assert(stack.size() == 0, "empty stack"); 2065 stack.push(_tail); 2066 visited.Clear(); 2067 visited.set(_tail->_idx); 2068 while (stack.size() > 0) { 2069 Node* n = stack.pop(); 2070 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2071 // Terminate this path 2072 } else if (n->Opcode() == Op_SafePoint) { 2073 if (_phase->get_loop(n) != this) { 2074 if (_required_safept == NULL) _required_safept = new Node_List(); 2075 _required_safept->push(n); // save the one closest to the tail 2076 } 2077 // Terminate this path 2078 } else { 2079 uint start = n->is_Region() ? 1 : 0; 2080 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; 2081 for (uint i = start; i < end; i++) { 2082 Node* in = n->in(i); 2083 assert(in->is_CFG(), "must be"); 2084 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { 2085 stack.push(in); 2086 } 2087 } 2088 } 2089 } 2090 } 2091 2092 //------------------------------check_safepts---------------------------- 2093 // Given dominators, try to find loops with calls that must always be 2094 // executed (call dominates loop tail). These loops do not need non-call 2095 // safepoints (ncsfpt). 2096 // 2097 // A complication is that a safepoint in a inner loop may be needed 2098 // by an outer loop. In the following, the inner loop sees it has a 2099 // call (block 3) on every path from the head (block 2) to the 2100 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) 2101 // in block 2, _but_ this leaves the outer loop without a safepoint. 2102 // 2103 // entry 0 2104 // | 2105 // v 2106 // outer 1,2 +->1 2107 // | | 2108 // | v 2109 // | 2<---+ ncsfpt in 2 2110 // |_/|\ | 2111 // | v | 2112 // inner 2,3 / 3 | call in 3 2113 // / | | 2114 // v +--+ 2115 // exit 4 2116 // 2117 // 2118 // This method creates a list (_required_safept) of ncsfpt nodes that must 2119 // be protected is created for each loop. When a ncsfpt maybe deleted, it 2120 // is first looked for in the lists for the outer loops of the current loop. 2121 // 2122 // The insights into the problem: 2123 // A) counted loops are okay 2124 // B) innermost loops are okay (only an inner loop can delete 2125 // a ncsfpt needed by an outer loop) 2126 // C) a loop is immune from an inner loop deleting a safepoint 2127 // if the loop has a call on the idom-path 2128 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the 2129 // idom-path that is not in a nested loop 2130 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner 2131 // loop needs to be prevented from deletion by an inner loop 2132 // 2133 // There are two analyses: 2134 // 1) The first, and cheaper one, scans the loop body from 2135 // tail to head following the idom (immediate dominator) 2136 // chain, looking for the cases (C,D,E) above. 2137 // Since inner loops are scanned before outer loops, there is summary 2138 // information about inner loops. Inner loops can be skipped over 2139 // when the tail of an inner loop is encountered. 2140 // 2141 // 2) The second, invoked if the first fails to find a call or ncsfpt on 2142 // the idom path (which is rare), scans all predecessor control paths 2143 // from the tail to the head, terminating a path when a call or sfpt 2144 // is encountered, to find the ncsfpt's that are closest to the tail. 2145 // 2146 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { 2147 // Bottom up traversal 2148 IdealLoopTree* ch = _child; 2149 if (_child) _child->check_safepts(visited, stack); 2150 if (_next) _next ->check_safepts(visited, stack); 2151 2152 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { 2153 bool has_call = false; // call on dom-path 2154 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth 2155 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth 2156 // Scan the dom-path nodes from tail to head 2157 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { 2158 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2159 has_call = true; 2160 _has_sfpt = 1; // Then no need for a safept! 2161 break; 2162 } else if (n->Opcode() == Op_SafePoint) { 2163 if (_phase->get_loop(n) == this) { 2164 has_local_ncsfpt = true; 2165 break; 2166 } 2167 if (nonlocal_ncsfpt == NULL) { 2168 nonlocal_ncsfpt = n; // save the one closest to the tail 2169 } 2170 } else { 2171 IdealLoopTree* nlpt = _phase->get_loop(n); 2172 if (this != nlpt) { 2173 // If at an inner loop tail, see if the inner loop has already 2174 // recorded seeing a call on the dom-path (and stop.) If not, 2175 // jump to the head of the inner loop. 2176 assert(is_member(nlpt), "nested loop"); 2177 Node* tail = nlpt->_tail; 2178 if (tail->in(0)->is_If()) tail = tail->in(0); 2179 if (n == tail) { 2180 // If inner loop has call on dom-path, so does outer loop 2181 if (nlpt->_has_sfpt) { 2182 has_call = true; 2183 _has_sfpt = 1; 2184 break; 2185 } 2186 // Skip to head of inner loop 2187 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); 2188 n = nlpt->_head; 2189 } 2190 } 2191 } 2192 } 2193 // Record safept's that this loop needs preserved when an 2194 // inner loop attempts to delete it's safepoints. 2195 if (_child != NULL && !has_call && !has_local_ncsfpt) { 2196 if (nonlocal_ncsfpt != NULL) { 2197 if (_required_safept == NULL) _required_safept = new Node_List(); 2198 _required_safept->push(nonlocal_ncsfpt); 2199 } else { 2200 // Failed to find a suitable safept on the dom-path. Now use 2201 // an all paths walk from tail to head, looking for safepoints to preserve. 2202 allpaths_check_safepts(visited, stack); 2203 } 2204 } 2205 } 2206 } 2207 2208 //---------------------------is_deleteable_safept---------------------------- 2209 // Is safept not required by an outer loop? 2210 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { 2211 assert(sfpt->Opcode() == Op_SafePoint, ""); 2212 IdealLoopTree* lp = get_loop(sfpt)->_parent; 2213 while (lp != NULL) { 2214 Node_List* sfpts = lp->_required_safept; 2215 if (sfpts != NULL) { 2216 for (uint i = 0; i < sfpts->size(); i++) { 2217 if (sfpt == sfpts->at(i)) 2218 return false; 2219 } 2220 } 2221 lp = lp->_parent; 2222 } 2223 return true; 2224 } 2225 2226 //---------------------------replace_parallel_iv------------------------------- 2227 // Replace parallel induction variable (parallel to trip counter) 2228 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) { 2229 assert(loop->_head->is_CountedLoop(), ""); 2230 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2231 if (!cl->is_valid_counted_loop()) 2232 return; // skip malformed counted loop 2233 Node *incr = cl->incr(); 2234 if (incr == NULL) 2235 return; // Dead loop? 2236 Node *init = cl->init_trip(); 2237 Node *phi = cl->phi(); 2238 int stride_con = cl->stride_con(); 2239 2240 // Visit all children, looking for Phis 2241 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { 2242 Node *out = cl->out(i); 2243 // Look for other phis (secondary IVs). Skip dead ones 2244 if (!out->is_Phi() || out == phi || !has_node(out)) 2245 continue; 2246 PhiNode* phi2 = out->as_Phi(); 2247 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); 2248 // Look for induction variables of the form: X += constant 2249 if (phi2->region() != loop->_head || 2250 incr2->req() != 3 || 2251 incr2->in(1) != phi2 || 2252 incr2 == incr || 2253 incr2->Opcode() != Op_AddI || 2254 !incr2->in(2)->is_Con()) 2255 continue; 2256 2257 // Check for parallel induction variable (parallel to trip counter) 2258 // via an affine function. In particular, count-down loops with 2259 // count-up array indices are common. We only RCE references off 2260 // the trip-counter, so we need to convert all these to trip-counter 2261 // expressions. 2262 Node *init2 = phi2->in( LoopNode::EntryControl ); 2263 int stride_con2 = incr2->in(2)->get_int(); 2264 2265 // The ratio of the two strides cannot be represented as an int 2266 // if stride_con2 is min_int and stride_con is -1. 2267 if (stride_con2 == min_jint && stride_con == -1) { 2268 continue; 2269 } 2270 2271 // The general case here gets a little tricky. We want to find the 2272 // GCD of all possible parallel IV's and make a new IV using this 2273 // GCD for the loop. Then all possible IVs are simple multiples of 2274 // the GCD. In practice, this will cover very few extra loops. 2275 // Instead we require 'stride_con2' to be a multiple of 'stride_con', 2276 // where +/-1 is the common case, but other integer multiples are 2277 // also easy to handle. 2278 int ratio_con = stride_con2/stride_con; 2279 2280 if ((ratio_con * stride_con) == stride_con2) { // Check for exact 2281 #ifndef PRODUCT 2282 if (TraceLoopOpts) { 2283 tty->print("Parallel IV: %d ", phi2->_idx); 2284 loop->dump_head(); 2285 } 2286 #endif 2287 // Convert to using the trip counter. The parallel induction 2288 // variable differs from the trip counter by a loop-invariant 2289 // amount, the difference between their respective initial values. 2290 // It is scaled by the 'ratio_con'. 2291 Node* ratio = _igvn.intcon(ratio_con); 2292 set_ctrl(ratio, C->root()); 2293 Node* ratio_init = new MulINode(init, ratio); 2294 _igvn.register_new_node_with_optimizer(ratio_init, init); 2295 set_early_ctrl(ratio_init); 2296 Node* diff = new SubINode(init2, ratio_init); 2297 _igvn.register_new_node_with_optimizer(diff, init2); 2298 set_early_ctrl(diff); 2299 Node* ratio_idx = new MulINode(phi, ratio); 2300 _igvn.register_new_node_with_optimizer(ratio_idx, phi); 2301 set_ctrl(ratio_idx, cl); 2302 Node* add = new AddINode(ratio_idx, diff); 2303 _igvn.register_new_node_with_optimizer(add); 2304 set_ctrl(add, cl); 2305 _igvn.replace_node( phi2, add ); 2306 // Sometimes an induction variable is unused 2307 if (add->outcnt() == 0) { 2308 _igvn.remove_dead_node(add); 2309 } 2310 --i; // deleted this phi; rescan starting with next position 2311 continue; 2312 } 2313 } 2314 } 2315 2316 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) { 2317 Node* keep = NULL; 2318 if (keep_one) { 2319 // Look for a safepoint on the idom-path. 2320 for (Node* i = tail(); i != _head; i = phase->idom(i)) { 2321 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) { 2322 keep = i; 2323 break; // Found one 2324 } 2325 } 2326 } 2327 2328 // Don't remove any safepoints if it is requested to keep a single safepoint and 2329 // no safepoint was found on idom-path. It is not safe to remove any safepoint 2330 // in this case since there's no safepoint dominating all paths in the loop body. 2331 bool prune = !keep_one || keep != NULL; 2332 2333 // Delete other safepoints in this loop. 2334 Node_List* sfpts = _safepts; 2335 if (prune && sfpts != NULL) { 2336 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint"); 2337 for (uint i = 0; i < sfpts->size(); i++) { 2338 Node* n = sfpts->at(i); 2339 assert(phase->get_loop(n) == this, ""); 2340 if (n != keep && phase->is_deleteable_safept(n)) { 2341 phase->lazy_replace(n, n->in(TypeFunc::Control)); 2342 } 2343 } 2344 } 2345 } 2346 2347 //------------------------------counted_loop----------------------------------- 2348 // Convert to counted loops where possible 2349 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { 2350 2351 // For grins, set the inner-loop flag here 2352 if (!_child) { 2353 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop(); 2354 } 2355 2356 IdealLoopTree* loop = this; 2357 if (_head->is_CountedLoop() || 2358 phase->is_counted_loop(_head, loop)) { 2359 2360 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) { 2361 // Indicate we do not need a safepoint here 2362 _has_sfpt = 1; 2363 } 2364 2365 // Remove safepoints 2366 bool keep_one_sfpt = !(_has_call || _has_sfpt); 2367 remove_safepoints(phase, keep_one_sfpt); 2368 2369 // Look for induction variables 2370 phase->replace_parallel_iv(this); 2371 2372 } else if (_parent != NULL && !_irreducible) { 2373 // Not a counted loop. Keep one safepoint. 2374 bool keep_one_sfpt = true; 2375 remove_safepoints(phase, keep_one_sfpt); 2376 } 2377 2378 // Recursively 2379 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?"); 2380 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops"); 2381 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase); 2382 if (loop->_next) loop->_next ->counted_loop(phase); 2383 } 2384 2385 #ifndef PRODUCT 2386 //------------------------------dump_head-------------------------------------- 2387 // Dump 1 liner for loop header info 2388 void IdealLoopTree::dump_head( ) const { 2389 for (uint i=0; i<_nest; i++) 2390 tty->print(" "); 2391 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); 2392 if (_irreducible) tty->print(" IRREDUCIBLE"); 2393 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl); 2394 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 2395 if (predicate != NULL ) { 2396 tty->print(" limit_check"); 2397 entry = entry->in(0)->in(0); 2398 } 2399 if (UseLoopPredicate) { 2400 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 2401 if (entry != NULL) { 2402 tty->print(" predicated"); 2403 } 2404 } 2405 if (_head->is_CountedLoop()) { 2406 CountedLoopNode *cl = _head->as_CountedLoop(); 2407 tty->print(" counted"); 2408 2409 Node* init_n = cl->init_trip(); 2410 if (init_n != NULL && init_n->is_Con()) 2411 tty->print(" [%d,", cl->init_trip()->get_int()); 2412 else 2413 tty->print(" [int,"); 2414 Node* limit_n = cl->limit(); 2415 if (limit_n != NULL && limit_n->is_Con()) 2416 tty->print("%d),", cl->limit()->get_int()); 2417 else 2418 tty->print("int),"); 2419 int stride_con = cl->stride_con(); 2420 if (stride_con > 0) tty->print("+"); 2421 tty->print("%d", stride_con); 2422 2423 tty->print(" (%0.f iters) ", cl->profile_trip_cnt()); 2424 2425 if (cl->is_pre_loop ()) tty->print(" pre" ); 2426 if (cl->is_main_loop()) tty->print(" main"); 2427 if (cl->is_post_loop()) tty->print(" post"); 2428 if (cl->is_vectorized_loop()) tty->print(" vector"); 2429 if (cl->range_checks_present()) tty->print(" rc "); 2430 if (cl->is_multiversioned()) tty->print(" multi "); 2431 } 2432 if (_has_call) tty->print(" has_call"); 2433 if (_has_sfpt) tty->print(" has_sfpt"); 2434 if (_rce_candidate) tty->print(" rce"); 2435 if (_safepts != NULL && _safepts->size() > 0) { 2436 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }"); 2437 } 2438 if (_required_safept != NULL && _required_safept->size() > 0) { 2439 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }"); 2440 } 2441 if (Verbose) { 2442 tty->print(" body={"); _body.dump_simple(); tty->print(" }"); 2443 } 2444 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) { 2445 tty->print(" strip_mined"); 2446 } 2447 tty->cr(); 2448 } 2449 2450 //------------------------------dump------------------------------------------- 2451 // Dump loops by loop tree 2452 void IdealLoopTree::dump( ) const { 2453 dump_head(); 2454 if (_child) _child->dump(); 2455 if (_next) _next ->dump(); 2456 } 2457 2458 #endif 2459 2460 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { 2461 if (loop == root) { 2462 if (loop->_child != NULL) { 2463 log->begin_head("loop_tree"); 2464 log->end_head(); 2465 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2466 log->tail("loop_tree"); 2467 assert(loop->_next == NULL, "what?"); 2468 } 2469 } else { 2470 Node* head = loop->_head; 2471 log->begin_head("loop"); 2472 log->print(" idx='%d' ", head->_idx); 2473 if (loop->_irreducible) log->print("irreducible='1' "); 2474 if (head->is_Loop()) { 2475 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); 2476 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); 2477 } 2478 if (head->is_CountedLoop()) { 2479 CountedLoopNode* cl = head->as_CountedLoop(); 2480 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); 2481 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); 2482 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); 2483 } 2484 log->end_head(); 2485 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2486 log->tail("loop"); 2487 if( loop->_next ) log_loop_tree(root, loop->_next, log); 2488 } 2489 } 2490 2491 //---------------------collect_potentially_useful_predicates----------------------- 2492 // Helper function to collect potentially useful predicates to prevent them from 2493 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates 2494 void PhaseIdealLoop::collect_potentially_useful_predicates( 2495 IdealLoopTree * loop, Unique_Node_List &useful_predicates) { 2496 if (loop->_child) { // child 2497 collect_potentially_useful_predicates(loop->_child, useful_predicates); 2498 } 2499 2500 // self (only loops that we can apply loop predication may use their predicates) 2501 if (loop->_head->is_Loop() && 2502 !loop->_irreducible && 2503 !loop->tail()->is_top()) { 2504 LoopNode* lpn = loop->_head->as_Loop(); 2505 Node* entry = lpn->in(LoopNode::EntryControl); 2506 Node* predicate_proj = find_predicate(entry); // loop_limit_check first 2507 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication 2508 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be"); 2509 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2510 entry = entry->in(0)->in(0); 2511 } 2512 predicate_proj = find_predicate(entry); // Predicate 2513 if (predicate_proj != NULL ) { 2514 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2515 } 2516 } 2517 2518 if (loop->_next) { // sibling 2519 collect_potentially_useful_predicates(loop->_next, useful_predicates); 2520 } 2521 } 2522 2523 //------------------------eliminate_useless_predicates----------------------------- 2524 // Eliminate all inserted predicates if they could not be used by loop predication. 2525 // Note: it will also eliminates loop limits check predicate since it also uses 2526 // Opaque1 node (see Parse::add_predicate()). 2527 void PhaseIdealLoop::eliminate_useless_predicates() { 2528 if (C->predicate_count() == 0) 2529 return; // no predicate left 2530 2531 Unique_Node_List useful_predicates; // to store useful predicates 2532 if (C->has_loops()) { 2533 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); 2534 } 2535 2536 for (int i = C->predicate_count(); i > 0; i--) { 2537 Node * n = C->predicate_opaque1_node(i-1); 2538 assert(n->Opcode() == Op_Opaque1, "must be"); 2539 if (!useful_predicates.member(n)) { // not in the useful list 2540 _igvn.replace_node(n, n->in(1)); 2541 } 2542 } 2543 } 2544 2545 //------------------------process_expensive_nodes----------------------------- 2546 // Expensive nodes have their control input set to prevent the GVN 2547 // from commoning them and as a result forcing the resulting node to 2548 // be in a more frequent path. Use CFG information here, to change the 2549 // control inputs so that some expensive nodes can be commoned while 2550 // not executed more frequently. 2551 bool PhaseIdealLoop::process_expensive_nodes() { 2552 assert(OptimizeExpensiveOps, "optimization off?"); 2553 2554 // Sort nodes to bring similar nodes together 2555 C->sort_expensive_nodes(); 2556 2557 bool progress = false; 2558 2559 for (int i = 0; i < C->expensive_count(); ) { 2560 Node* n = C->expensive_node(i); 2561 int start = i; 2562 // Find nodes similar to n 2563 i++; 2564 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++); 2565 int end = i; 2566 // And compare them two by two 2567 for (int j = start; j < end; j++) { 2568 Node* n1 = C->expensive_node(j); 2569 if (is_node_unreachable(n1)) { 2570 continue; 2571 } 2572 for (int k = j+1; k < end; k++) { 2573 Node* n2 = C->expensive_node(k); 2574 if (is_node_unreachable(n2)) { 2575 continue; 2576 } 2577 2578 assert(n1 != n2, "should be pair of nodes"); 2579 2580 Node* c1 = n1->in(0); 2581 Node* c2 = n2->in(0); 2582 2583 Node* parent_c1 = c1; 2584 Node* parent_c2 = c2; 2585 2586 // The call to get_early_ctrl_for_expensive() moves the 2587 // expensive nodes up but stops at loops that are in a if 2588 // branch. See whether we can exit the loop and move above the 2589 // If. 2590 if (c1->is_Loop()) { 2591 parent_c1 = c1->in(1); 2592 } 2593 if (c2->is_Loop()) { 2594 parent_c2 = c2->in(1); 2595 } 2596 2597 if (parent_c1 == parent_c2) { 2598 _igvn._worklist.push(n1); 2599 _igvn._worklist.push(n2); 2600 continue; 2601 } 2602 2603 // Look for identical expensive node up the dominator chain. 2604 if (is_dominator(c1, c2)) { 2605 c2 = c1; 2606 } else if (is_dominator(c2, c1)) { 2607 c1 = c2; 2608 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() && 2609 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) { 2610 // Both branches have the same expensive node so move it up 2611 // before the if. 2612 c1 = c2 = idom(parent_c1->in(0)); 2613 } 2614 // Do the actual moves 2615 if (n1->in(0) != c1) { 2616 _igvn.hash_delete(n1); 2617 n1->set_req(0, c1); 2618 _igvn.hash_insert(n1); 2619 _igvn._worklist.push(n1); 2620 progress = true; 2621 } 2622 if (n2->in(0) != c2) { 2623 _igvn.hash_delete(n2); 2624 n2->set_req(0, c2); 2625 _igvn.hash_insert(n2); 2626 _igvn._worklist.push(n2); 2627 progress = true; 2628 } 2629 } 2630 } 2631 } 2632 2633 return progress; 2634 } 2635 2636 2637 //============================================================================= 2638 //----------------------------build_and_optimize------------------------------- 2639 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to 2640 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. 2641 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts, bool last_round) { 2642 ResourceMark rm; 2643 2644 int old_progress = C->major_progress(); 2645 uint orig_worklist_size = _igvn._worklist.size(); 2646 2647 #ifndef PRODUCT 2648 // Capture for later assert 2649 uint unique = C->unique(); 2650 _loop_invokes++; 2651 _loop_work += unique; 2652 #endif 2653 2654 // True if the method has at least 1 irreducible loop 2655 _has_irreducible_loops = false; 2656 2657 _created_loop_node = false; 2658 2659 Arena *a = Thread::current()->resource_area(); 2660 VectorSet visited(a); 2661 // Pre-grow the mapping from Nodes to IdealLoopTrees. 2662 _nodes.map(C->unique(), NULL); 2663 memset(_nodes.adr(), 0, wordSize * C->unique()); 2664 2665 // Pre-build the top-level outermost loop tree entry 2666 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); 2667 // Do not need a safepoint at the top level 2668 _ltree_root->_has_sfpt = 1; 2669 2670 // Initialize Dominators. 2671 // Checked in clone_loop_predicate() during beautify_loops(). 2672 _idom_size = 0; 2673 _idom = NULL; 2674 _dom_depth = NULL; 2675 _dom_stk = NULL; 2676 2677 // Empty pre-order array 2678 allocate_preorders(); 2679 2680 // Build a loop tree on the fly. Build a mapping from CFG nodes to 2681 // IdealLoopTree entries. Data nodes are NOT walked. 2682 build_loop_tree(); 2683 // Check for bailout, and return 2684 if (C->failing()) { 2685 return; 2686 } 2687 2688 // No loops after all 2689 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); 2690 2691 // There should always be an outer loop containing the Root and Return nodes. 2692 // If not, we have a degenerate empty program. Bail out in this case. 2693 if (!has_node(C->root())) { 2694 if (!_verify_only) { 2695 C->clear_major_progress(); 2696 C->record_method_not_compilable("empty program detected during loop optimization"); 2697 } 2698 return; 2699 } 2700 2701 // Nothing to do, so get out 2702 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only; 2703 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn); 2704 if (stop_early && !do_expensive_nodes) { 2705 _igvn.optimize(); // Cleanup NeverBranches 2706 return; 2707 } 2708 2709 // Set loop nesting depth 2710 _ltree_root->set_nest( 0 ); 2711 2712 // Split shared headers and insert loop landing pads. 2713 // Do not bother doing this on the Root loop of course. 2714 if( !_verify_me && !_verify_only && _ltree_root->_child ) { 2715 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3); 2716 if( _ltree_root->_child->beautify_loops( this ) ) { 2717 // IdealLoopTree::split_outer_loop may produce phi-nodes with a single in edge. 2718 // Transform them away. 2719 _igvn.optimize(); 2720 2721 // Re-build loop tree! 2722 _ltree_root->_child = NULL; 2723 _nodes.clear(); 2724 reallocate_preorders(); 2725 build_loop_tree(); 2726 2727 // Check for bailout, and return 2728 if (C->failing()) { 2729 return; 2730 } 2731 // Reset loop nesting depth 2732 _ltree_root->set_nest( 0 ); 2733 2734 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3); 2735 } 2736 } 2737 2738 // Reset major-progress flag for the driver's heuristics 2739 C->clear_major_progress(); 2740 2741 // Build Dominators for elision of NULL checks & loop finding. 2742 // Since nodes do not have a slot for immediate dominator, make 2743 // a persistent side array for that info indexed on node->_idx. 2744 _idom_size = C->unique(); 2745 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); 2746 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); 2747 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth 2748 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); 2749 2750 Dominators(); 2751 2752 if (!_verify_only) { 2753 // As a side effect, Dominators removed any unreachable CFG paths 2754 // into RegionNodes. It doesn't do this test against Root, so 2755 // we do it here. 2756 for( uint i = 1; i < C->root()->req(); i++ ) { 2757 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? 2758 _igvn.delete_input_of(C->root(), i); 2759 i--; // Rerun same iteration on compressed edges 2760 } 2761 } 2762 2763 // Given dominators, try to find inner loops with calls that must 2764 // always be executed (call dominates loop tail). These loops do 2765 // not need a separate safepoint. 2766 Node_List cisstack(a); 2767 _ltree_root->check_safepts(visited, cisstack); 2768 } 2769 2770 // Walk the DATA nodes and place into loops. Find earliest control 2771 // node. For CFG nodes, the _nodes array starts out and remains 2772 // holding the associated IdealLoopTree pointer. For DATA nodes, the 2773 // _nodes array holds the earliest legal controlling CFG node. 2774 2775 // Allocate stack with enough space to avoid frequent realloc 2776 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats 2777 Node_Stack nstack( a, stack_size ); 2778 2779 visited.Clear(); 2780 Node_List worklist(a); 2781 // Don't need C->root() on worklist since 2782 // it will be processed among C->top() inputs 2783 worklist.push( C->top() ); 2784 visited.set( C->top()->_idx ); // Set C->top() as visited now 2785 build_loop_early( visited, worklist, nstack ); 2786 2787 // Given early legal placement, try finding counted loops. This placement 2788 // is good enough to discover most loop invariants. 2789 if( !_verify_me && !_verify_only ) 2790 _ltree_root->counted_loop( this ); 2791 2792 // Find latest loop placement. Find ideal loop placement. 2793 visited.Clear(); 2794 init_dom_lca_tags(); 2795 // Need C->root() on worklist when processing outs 2796 worklist.push( C->root() ); 2797 NOT_PRODUCT( C->verify_graph_edges(); ) 2798 worklist.push( C->top() ); 2799 build_loop_late( visited, worklist, nstack ); 2800 2801 if (_verify_only) { 2802 // restore major progress flag 2803 for (int i = 0; i < old_progress; i++) 2804 C->set_major_progress(); 2805 assert(C->unique() == unique, "verification mode made Nodes? ? ?"); 2806 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything"); 2807 return; 2808 } 2809 2810 // clear out the dead code after build_loop_late 2811 while (_deadlist.size()) { 2812 _igvn.remove_globally_dead_node(_deadlist.pop()); 2813 } 2814 2815 if (stop_early) { 2816 assert(do_expensive_nodes, "why are we here?"); 2817 if (process_expensive_nodes()) { 2818 // If we made some progress when processing expensive nodes then 2819 // the IGVN may modify the graph in a way that will allow us to 2820 // make some more progress: we need to try processing expensive 2821 // nodes again. 2822 C->set_major_progress(); 2823 } 2824 _igvn.optimize(); 2825 return; 2826 } 2827 2828 // Some parser-inserted loop predicates could never be used by loop 2829 // predication or they were moved away from loop during some optimizations. 2830 // For example, peeling. Eliminate them before next loop optimizations. 2831 eliminate_useless_predicates(); 2832 2833 #ifndef PRODUCT 2834 C->verify_graph_edges(); 2835 if (_verify_me) { // Nested verify pass? 2836 // Check to see if the verify mode is broken 2837 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); 2838 return; 2839 } 2840 if(VerifyLoopOptimizations) verify(); 2841 if(TraceLoopOpts && C->has_loops()) { 2842 _ltree_root->dump(); 2843 } 2844 #endif 2845 2846 if (skip_loop_opts) { 2847 // restore major progress flag 2848 for (int i = 0; i < old_progress; i++) { 2849 C->set_major_progress(); 2850 } 2851 2852 // Cleanup any modified bits 2853 _igvn.optimize(); 2854 2855 if (C->log() != NULL) { 2856 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2857 } 2858 return; 2859 } 2860 2861 if (ReassociateInvariants) { 2862 // Reassociate invariants and prep for split_thru_phi 2863 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2864 IdealLoopTree* lpt = iter.current(); 2865 bool is_counted = lpt->is_counted(); 2866 if (!is_counted || !lpt->is_inner()) continue; 2867 2868 // check for vectorized loops, any reassociation of invariants was already done 2869 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue; 2870 2871 lpt->reassociate_invariants(this); 2872 2873 // Because RCE opportunities can be masked by split_thru_phi, 2874 // look for RCE candidates and inhibit split_thru_phi 2875 // on just their loop-phi's for this pass of loop opts 2876 if (SplitIfBlocks && do_split_ifs) { 2877 if (lpt->policy_range_check(this)) { 2878 lpt->_rce_candidate = 1; // = true 2879 } 2880 } 2881 } 2882 } 2883 2884 // Check for aggressive application of split-if and other transforms 2885 // that require basic-block info (like cloning through Phi's) 2886 if( SplitIfBlocks && do_split_ifs ) { 2887 visited.Clear(); 2888 split_if_with_blocks( visited, nstack, last_round ); 2889 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); 2890 if (last_round) { 2891 C->set_major_progress(); 2892 } 2893 } 2894 2895 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) { 2896 C->set_major_progress(); 2897 } 2898 2899 // Perform loop predication before iteration splitting 2900 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) { 2901 _ltree_root->_child->loop_predication(this); 2902 } 2903 2904 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { 2905 if (do_intrinsify_fill()) { 2906 C->set_major_progress(); 2907 } 2908 } 2909 2910 // Perform iteration-splitting on inner loops. Split iterations to avoid 2911 // range checks or one-shot null checks. 2912 2913 // If split-if's didn't hack the graph too bad (no CFG changes) 2914 // then do loop opts. 2915 if (C->has_loops() && !C->major_progress()) { 2916 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 2917 _ltree_root->_child->iteration_split( this, worklist ); 2918 // No verify after peeling! GCM has hoisted code out of the loop. 2919 // After peeling, the hoisted code could sink inside the peeled area. 2920 // The peeling code does not try to recompute the best location for 2921 // all the code before the peeled area, so the verify pass will always 2922 // complain about it. 2923 } 2924 // Do verify graph edges in any case 2925 NOT_PRODUCT( C->verify_graph_edges(); ); 2926 2927 if (!do_split_ifs) { 2928 // We saw major progress in Split-If to get here. We forced a 2929 // pass with unrolling and not split-if, however more split-if's 2930 // might make progress. If the unrolling didn't make progress 2931 // then the major-progress flag got cleared and we won't try 2932 // another round of Split-If. In particular the ever-common 2933 // instance-of/check-cast pattern requires at least 2 rounds of 2934 // Split-If to clear out. 2935 C->set_major_progress(); 2936 } 2937 2938 // Repeat loop optimizations if new loops were seen 2939 if (created_loop_node()) { 2940 C->set_major_progress(); 2941 } 2942 2943 // Keep loop predicates and perform optimizations with them 2944 // until no more loop optimizations could be done. 2945 // After that switch predicates off and do more loop optimizations. 2946 if (!C->major_progress() && (C->predicate_count() > 0)) { 2947 C->cleanup_loop_predicates(_igvn); 2948 if (TraceLoopOpts) { 2949 tty->print_cr("PredicatesOff"); 2950 } 2951 C->set_major_progress(); 2952 } 2953 2954 // Convert scalar to superword operations at the end of all loop opts. 2955 if (UseSuperWord && C->has_loops() && !C->major_progress()) { 2956 // SuperWord transform 2957 SuperWord sw(this); 2958 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2959 IdealLoopTree* lpt = iter.current(); 2960 if (lpt->is_counted()) { 2961 CountedLoopNode *cl = lpt->_head->as_CountedLoop(); 2962 2963 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) { 2964 // Check that the rce'd post loop is encountered first, multiversion after all 2965 // major main loop optimization are concluded 2966 if (!C->major_progress()) { 2967 IdealLoopTree *lpt_next = lpt->_next; 2968 if (lpt_next && lpt_next->is_counted()) { 2969 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop(); 2970 has_range_checks(lpt_next); 2971 if (cl->is_post_loop() && cl->range_checks_present()) { 2972 if (!cl->is_multiversioned()) { 2973 if (multi_version_post_loops(lpt, lpt_next) == false) { 2974 // Cause the rce loop to be optimized away if we fail 2975 cl->mark_is_multiversioned(); 2976 cl->set_slp_max_unroll(0); 2977 poison_rce_post_loop(lpt); 2978 } 2979 } 2980 } 2981 } 2982 sw.transform_loop(lpt, true); 2983 } 2984 } else if (cl->is_main_loop()) { 2985 sw.transform_loop(lpt, true); 2986 } 2987 } 2988 } 2989 } 2990 2991 // Cleanup any modified bits 2992 _igvn.optimize(); 2993 2994 // disable assert until issue with split_flow_path is resolved (6742111) 2995 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), 2996 // "shouldn't introduce irreducible loops"); 2997 2998 if (C->log() != NULL) { 2999 log_loop_tree(_ltree_root, _ltree_root, C->log()); 3000 } 3001 } 3002 3003 #ifndef PRODUCT 3004 //------------------------------print_statistics------------------------------- 3005 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes 3006 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique 3007 void PhaseIdealLoop::print_statistics() { 3008 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); 3009 } 3010 3011 //------------------------------verify----------------------------------------- 3012 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. 3013 static int fail; // debug only, so its multi-thread dont care 3014 void PhaseIdealLoop::verify() const { 3015 int old_progress = C->major_progress(); 3016 ResourceMark rm; 3017 PhaseIdealLoop loop_verify( _igvn, this ); 3018 VectorSet visited(Thread::current()->resource_area()); 3019 3020 fail = 0; 3021 verify_compare( C->root(), &loop_verify, visited ); 3022 assert( fail == 0, "verify loops failed" ); 3023 // Verify loop structure is the same 3024 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); 3025 // Reset major-progress. It was cleared by creating a verify version of 3026 // PhaseIdealLoop. 3027 for( int i=0; i<old_progress; i++ ) 3028 C->set_major_progress(); 3029 } 3030 3031 //------------------------------verify_compare--------------------------------- 3032 // Make sure me and the given PhaseIdealLoop agree on key data structures 3033 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { 3034 if( !n ) return; 3035 if( visited.test_set( n->_idx ) ) return; 3036 if( !_nodes[n->_idx] ) { // Unreachable 3037 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); 3038 return; 3039 } 3040 3041 uint i; 3042 for( i = 0; i < n->req(); i++ ) 3043 verify_compare( n->in(i), loop_verify, visited ); 3044 3045 // Check the '_nodes' block/loop structure 3046 i = n->_idx; 3047 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl 3048 if( _nodes[i] != loop_verify->_nodes[i] && 3049 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { 3050 tty->print("Mismatched control setting for: "); 3051 n->dump(); 3052 if( fail++ > 10 ) return; 3053 Node *c = get_ctrl_no_update(n); 3054 tty->print("We have it as: "); 3055 if( c->in(0) ) c->dump(); 3056 else tty->print_cr("N%d",c->_idx); 3057 tty->print("Verify thinks: "); 3058 if( loop_verify->has_ctrl(n) ) 3059 loop_verify->get_ctrl_no_update(n)->dump(); 3060 else 3061 loop_verify->get_loop_idx(n)->dump(); 3062 tty->cr(); 3063 } 3064 } else { // We have a loop 3065 IdealLoopTree *us = get_loop_idx(n); 3066 if( loop_verify->has_ctrl(n) ) { 3067 tty->print("Mismatched loop setting for: "); 3068 n->dump(); 3069 if( fail++ > 10 ) return; 3070 tty->print("We have it as: "); 3071 us->dump(); 3072 tty->print("Verify thinks: "); 3073 loop_verify->get_ctrl_no_update(n)->dump(); 3074 tty->cr(); 3075 } else if (!C->major_progress()) { 3076 // Loop selection can be messed up if we did a major progress 3077 // operation, like split-if. Do not verify in that case. 3078 IdealLoopTree *them = loop_verify->get_loop_idx(n); 3079 if( us->_head != them->_head || us->_tail != them->_tail ) { 3080 tty->print("Unequals loops for: "); 3081 n->dump(); 3082 if( fail++ > 10 ) return; 3083 tty->print("We have it as: "); 3084 us->dump(); 3085 tty->print("Verify thinks: "); 3086 them->dump(); 3087 tty->cr(); 3088 } 3089 } 3090 } 3091 3092 // Check for immediate dominators being equal 3093 if( i >= _idom_size ) { 3094 if( !n->is_CFG() ) return; 3095 tty->print("CFG Node with no idom: "); 3096 n->dump(); 3097 return; 3098 } 3099 if( !n->is_CFG() ) return; 3100 if( n == C->root() ) return; // No IDOM here 3101 3102 assert(n->_idx == i, "sanity"); 3103 Node *id = idom_no_update(n); 3104 if( id != loop_verify->idom_no_update(n) ) { 3105 tty->print("Unequals idoms for: "); 3106 n->dump(); 3107 if( fail++ > 10 ) return; 3108 tty->print("We have it as: "); 3109 id->dump(); 3110 tty->print("Verify thinks: "); 3111 loop_verify->idom_no_update(n)->dump(); 3112 tty->cr(); 3113 } 3114 3115 } 3116 3117 //------------------------------verify_tree------------------------------------ 3118 // Verify that tree structures match. Because the CFG can change, siblings 3119 // within the loop tree can be reordered. We attempt to deal with that by 3120 // reordering the verify's loop tree if possible. 3121 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { 3122 assert( _parent == parent, "Badly formed loop tree" ); 3123 3124 // Siblings not in same order? Attempt to re-order. 3125 if( _head != loop->_head ) { 3126 // Find _next pointer to update 3127 IdealLoopTree **pp = &loop->_parent->_child; 3128 while( *pp != loop ) 3129 pp = &((*pp)->_next); 3130 // Find proper sibling to be next 3131 IdealLoopTree **nn = &loop->_next; 3132 while( (*nn) && (*nn)->_head != _head ) 3133 nn = &((*nn)->_next); 3134 3135 // Check for no match. 3136 if( !(*nn) ) { 3137 // Annoyingly, irreducible loops can pick different headers 3138 // after a major_progress operation, so the rest of the loop 3139 // tree cannot be matched. 3140 if (_irreducible && Compile::current()->major_progress()) return; 3141 assert( 0, "failed to match loop tree" ); 3142 } 3143 3144 // Move (*nn) to (*pp) 3145 IdealLoopTree *hit = *nn; 3146 *nn = hit->_next; 3147 hit->_next = loop; 3148 *pp = loop; 3149 loop = hit; 3150 // Now try again to verify 3151 } 3152 3153 assert( _head == loop->_head , "mismatched loop head" ); 3154 Node *tail = _tail; // Inline a non-updating version of 3155 while( !tail->in(0) ) // the 'tail()' call. 3156 tail = tail->in(1); 3157 assert( tail == loop->_tail, "mismatched loop tail" ); 3158 3159 // Counted loops that are guarded should be able to find their guards 3160 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { 3161 CountedLoopNode *cl = _head->as_CountedLoop(); 3162 Node *init = cl->init_trip(); 3163 Node *ctrl = cl->in(LoopNode::EntryControl); 3164 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 3165 Node *iff = ctrl->in(0); 3166 assert( iff->Opcode() == Op_If, "" ); 3167 Node *bol = iff->in(1); 3168 assert( bol->Opcode() == Op_Bool, "" ); 3169 Node *cmp = bol->in(1); 3170 assert( cmp->Opcode() == Op_CmpI, "" ); 3171 Node *add = cmp->in(1); 3172 Node *opaq; 3173 if( add->Opcode() == Op_Opaque1 ) { 3174 opaq = add; 3175 } else { 3176 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); 3177 assert( add == init, "" ); 3178 opaq = cmp->in(2); 3179 } 3180 assert( opaq->Opcode() == Op_Opaque1, "" ); 3181 3182 } 3183 3184 if (_child != NULL) _child->verify_tree(loop->_child, this); 3185 if (_next != NULL) _next ->verify_tree(loop->_next, parent); 3186 // Innermost loops need to verify loop bodies, 3187 // but only if no 'major_progress' 3188 int fail = 0; 3189 if (!Compile::current()->major_progress() && _child == NULL) { 3190 for( uint i = 0; i < _body.size(); i++ ) { 3191 Node *n = _body.at(i); 3192 if (n->outcnt() == 0) continue; // Ignore dead 3193 uint j; 3194 for( j = 0; j < loop->_body.size(); j++ ) 3195 if( loop->_body.at(j) == n ) 3196 break; 3197 if( j == loop->_body.size() ) { // Not found in loop body 3198 // Last ditch effort to avoid assertion: Its possible that we 3199 // have some users (so outcnt not zero) but are still dead. 3200 // Try to find from root. 3201 if (Compile::current()->root()->find(n->_idx)) { 3202 fail++; 3203 tty->print("We have that verify does not: "); 3204 n->dump(); 3205 } 3206 } 3207 } 3208 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { 3209 Node *n = loop->_body.at(i2); 3210 if (n->outcnt() == 0) continue; // Ignore dead 3211 uint j; 3212 for( j = 0; j < _body.size(); j++ ) 3213 if( _body.at(j) == n ) 3214 break; 3215 if( j == _body.size() ) { // Not found in loop body 3216 // Last ditch effort to avoid assertion: Its possible that we 3217 // have some users (so outcnt not zero) but are still dead. 3218 // Try to find from root. 3219 if (Compile::current()->root()->find(n->_idx)) { 3220 fail++; 3221 tty->print("Verify has that we do not: "); 3222 n->dump(); 3223 } 3224 } 3225 } 3226 assert( !fail, "loop body mismatch" ); 3227 } 3228 } 3229 3230 #endif 3231 3232 //------------------------------set_idom--------------------------------------- 3233 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { 3234 uint idx = d->_idx; 3235 if (idx >= _idom_size) { 3236 uint newsize = _idom_size<<1; 3237 while( idx >= newsize ) { 3238 newsize <<= 1; 3239 } 3240 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); 3241 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); 3242 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); 3243 _idom_size = newsize; 3244 } 3245 _idom[idx] = n; 3246 _dom_depth[idx] = dom_depth; 3247 } 3248 3249 //------------------------------recompute_dom_depth--------------------------------------- 3250 // The dominator tree is constructed with only parent pointers. 3251 // This recomputes the depth in the tree by first tagging all 3252 // nodes as "no depth yet" marker. The next pass then runs up 3253 // the dom tree from each node marked "no depth yet", and computes 3254 // the depth on the way back down. 3255 void PhaseIdealLoop::recompute_dom_depth() { 3256 uint no_depth_marker = C->unique(); 3257 uint i; 3258 // Initialize depth to "no depth yet" and realize all lazy updates 3259 for (i = 0; i < _idom_size; i++) { 3260 // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized). 3261 if (_dom_depth[i] > 0 && _idom[i] != NULL) { 3262 _dom_depth[i] = no_depth_marker; 3263 3264 // heal _idom if it has a fwd mapping in _nodes 3265 if (_idom[i]->in(0) == NULL) { 3266 idom(i); 3267 } 3268 } 3269 } 3270 if (_dom_stk == NULL) { 3271 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size. 3272 if (init_size < 10) init_size = 10; 3273 _dom_stk = new GrowableArray<uint>(init_size); 3274 } 3275 // Compute new depth for each node. 3276 for (i = 0; i < _idom_size; i++) { 3277 uint j = i; 3278 // Run up the dom tree to find a node with a depth 3279 while (_dom_depth[j] == no_depth_marker) { 3280 _dom_stk->push(j); 3281 j = _idom[j]->_idx; 3282 } 3283 // Compute the depth on the way back down this tree branch 3284 uint dd = _dom_depth[j] + 1; 3285 while (_dom_stk->length() > 0) { 3286 uint j = _dom_stk->pop(); 3287 _dom_depth[j] = dd; 3288 dd++; 3289 } 3290 } 3291 } 3292 3293 //------------------------------sort------------------------------------------- 3294 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the 3295 // loop tree, not the root. 3296 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { 3297 if( !innermost ) return loop; // New innermost loop 3298 3299 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number 3300 assert( loop_preorder, "not yet post-walked loop" ); 3301 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer 3302 IdealLoopTree *l = *pp; // Do I go before or after 'l'? 3303 3304 // Insert at start of list 3305 while( l ) { // Insertion sort based on pre-order 3306 if( l == loop ) return innermost; // Already on list! 3307 int l_preorder = get_preorder(l->_head); // Cache pre-order number 3308 assert( l_preorder, "not yet post-walked l" ); 3309 // Check header pre-order number to figure proper nesting 3310 if( loop_preorder > l_preorder ) 3311 break; // End of insertion 3312 // If headers tie (e.g., shared headers) check tail pre-order numbers. 3313 // Since I split shared headers, you'd think this could not happen. 3314 // BUT: I must first do the preorder numbering before I can discover I 3315 // have shared headers, so the split headers all get the same preorder 3316 // number as the RegionNode they split from. 3317 if( loop_preorder == l_preorder && 3318 get_preorder(loop->_tail) < get_preorder(l->_tail) ) 3319 break; // Also check for shared headers (same pre#) 3320 pp = &l->_parent; // Chain up list 3321 l = *pp; 3322 } 3323 // Link into list 3324 // Point predecessor to me 3325 *pp = loop; 3326 // Point me to successor 3327 IdealLoopTree *p = loop->_parent; 3328 loop->_parent = l; // Point me to successor 3329 if( p ) sort( p, innermost ); // Insert my parents into list as well 3330 return innermost; 3331 } 3332 3333 //------------------------------build_loop_tree-------------------------------- 3334 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit 3335 // bits. The _nodes[] array is mapped by Node index and holds a NULL for 3336 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the 3337 // tightest enclosing IdealLoopTree for post-walked. 3338 // 3339 // During my forward walk I do a short 1-layer lookahead to see if I can find 3340 // a loop backedge with that doesn't have any work on the backedge. This 3341 // helps me construct nested loops with shared headers better. 3342 // 3343 // Once I've done the forward recursion, I do the post-work. For each child 3344 // I check to see if there is a backedge. Backedges define a loop! I 3345 // insert an IdealLoopTree at the target of the backedge. 3346 // 3347 // During the post-work I also check to see if I have several children 3348 // belonging to different loops. If so, then this Node is a decision point 3349 // where control flow can choose to change loop nests. It is at this 3350 // decision point where I can figure out how loops are nested. At this 3351 // time I can properly order the different loop nests from my children. 3352 // Note that there may not be any backedges at the decision point! 3353 // 3354 // Since the decision point can be far removed from the backedges, I can't 3355 // order my loops at the time I discover them. Thus at the decision point 3356 // I need to inspect loop header pre-order numbers to properly nest my 3357 // loops. This means I need to sort my childrens' loops by pre-order. 3358 // The sort is of size number-of-control-children, which generally limits 3359 // it to size 2 (i.e., I just choose between my 2 target loops). 3360 void PhaseIdealLoop::build_loop_tree() { 3361 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3362 GrowableArray <Node *> bltstack(C->live_nodes() >> 1); 3363 Node *n = C->root(); 3364 bltstack.push(n); 3365 int pre_order = 1; 3366 int stack_size; 3367 3368 while ( ( stack_size = bltstack.length() ) != 0 ) { 3369 n = bltstack.top(); // Leave node on stack 3370 if ( !is_visited(n) ) { 3371 // ---- Pre-pass Work ---- 3372 // Pre-walked but not post-walked nodes need a pre_order number. 3373 3374 set_preorder_visited( n, pre_order ); // set as visited 3375 3376 // ---- Scan over children ---- 3377 // Scan first over control projections that lead to loop headers. 3378 // This helps us find inner-to-outer loops with shared headers better. 3379 3380 // Scan children's children for loop headers. 3381 for ( int i = n->outcnt() - 1; i >= 0; --i ) { 3382 Node* m = n->raw_out(i); // Child 3383 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children 3384 // Scan over children's children to find loop 3385 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3386 Node* l = m->fast_out(j); 3387 if( is_visited(l) && // Been visited? 3388 !is_postvisited(l) && // But not post-visited 3389 get_preorder(l) < pre_order ) { // And smaller pre-order 3390 // Found! Scan the DFS down this path before doing other paths 3391 bltstack.push(m); 3392 break; 3393 } 3394 } 3395 } 3396 } 3397 pre_order++; 3398 } 3399 else if ( !is_postvisited(n) ) { 3400 // Note: build_loop_tree_impl() adds out edges on rare occasions, 3401 // such as com.sun.rsasign.am::a. 3402 // For non-recursive version, first, process current children. 3403 // On next iteration, check if additional children were added. 3404 for ( int k = n->outcnt() - 1; k >= 0; --k ) { 3405 Node* u = n->raw_out(k); 3406 if ( u->is_CFG() && !is_visited(u) ) { 3407 bltstack.push(u); 3408 } 3409 } 3410 if ( bltstack.length() == stack_size ) { 3411 // There were no additional children, post visit node now 3412 (void)bltstack.pop(); // Remove node from stack 3413 pre_order = build_loop_tree_impl( n, pre_order ); 3414 // Check for bailout 3415 if (C->failing()) { 3416 return; 3417 } 3418 // Check to grow _preorders[] array for the case when 3419 // build_loop_tree_impl() adds new nodes. 3420 check_grow_preorders(); 3421 } 3422 } 3423 else { 3424 (void)bltstack.pop(); // Remove post-visited node from stack 3425 } 3426 } 3427 } 3428 3429 //------------------------------build_loop_tree_impl--------------------------- 3430 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { 3431 // ---- Post-pass Work ---- 3432 // Pre-walked but not post-walked nodes need a pre_order number. 3433 3434 // Tightest enclosing loop for this Node 3435 IdealLoopTree *innermost = NULL; 3436 3437 // For all children, see if any edge is a backedge. If so, make a loop 3438 // for it. Then find the tightest enclosing loop for the self Node. 3439 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3440 Node* m = n->fast_out(i); // Child 3441 if( n == m ) continue; // Ignore control self-cycles 3442 if( !m->is_CFG() ) continue;// Ignore non-CFG edges 3443 3444 IdealLoopTree *l; // Child's loop 3445 if( !is_postvisited(m) ) { // Child visited but not post-visited? 3446 // Found a backedge 3447 assert( get_preorder(m) < pre_order, "should be backedge" ); 3448 // Check for the RootNode, which is already a LoopNode and is allowed 3449 // to have multiple "backedges". 3450 if( m == C->root()) { // Found the root? 3451 l = _ltree_root; // Root is the outermost LoopNode 3452 } else { // Else found a nested loop 3453 // Insert a LoopNode to mark this loop. 3454 l = new IdealLoopTree(this, m, n); 3455 } // End of Else found a nested loop 3456 if( !has_loop(m) ) // If 'm' does not already have a loop set 3457 set_loop(m, l); // Set loop header to loop now 3458 3459 } else { // Else not a nested loop 3460 if( !_nodes[m->_idx] ) continue; // Dead code has no loop 3461 l = get_loop(m); // Get previously determined loop 3462 // If successor is header of a loop (nest), move up-loop till it 3463 // is a member of some outer enclosing loop. Since there are no 3464 // shared headers (I've split them already) I only need to go up 3465 // at most 1 level. 3466 while( l && l->_head == m ) // Successor heads loop? 3467 l = l->_parent; // Move up 1 for me 3468 // If this loop is not properly parented, then this loop 3469 // has no exit path out, i.e. its an infinite loop. 3470 if( !l ) { 3471 // Make loop "reachable" from root so the CFG is reachable. Basically 3472 // insert a bogus loop exit that is never taken. 'm', the loop head, 3473 // points to 'n', one (of possibly many) fall-in paths. There may be 3474 // many backedges as well. 3475 3476 // Here I set the loop to be the root loop. I could have, after 3477 // inserting a bogus loop exit, restarted the recursion and found my 3478 // new loop exit. This would make the infinite loop a first-class 3479 // loop and it would then get properly optimized. What's the use of 3480 // optimizing an infinite loop? 3481 l = _ltree_root; // Oops, found infinite loop 3482 3483 if (!_verify_only) { 3484 // Insert the NeverBranch between 'm' and it's control user. 3485 NeverBranchNode *iff = new NeverBranchNode( m ); 3486 _igvn.register_new_node_with_optimizer(iff); 3487 set_loop(iff, l); 3488 Node *if_t = new CProjNode( iff, 0 ); 3489 _igvn.register_new_node_with_optimizer(if_t); 3490 set_loop(if_t, l); 3491 3492 Node* cfg = NULL; // Find the One True Control User of m 3493 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3494 Node* x = m->fast_out(j); 3495 if (x->is_CFG() && x != m && x != iff) 3496 { cfg = x; break; } 3497 } 3498 assert(cfg != NULL, "must find the control user of m"); 3499 uint k = 0; // Probably cfg->in(0) 3500 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region 3501 cfg->set_req( k, if_t ); // Now point to NeverBranch 3502 _igvn._worklist.push(cfg); 3503 3504 // Now create the never-taken loop exit 3505 Node *if_f = new CProjNode( iff, 1 ); 3506 _igvn.register_new_node_with_optimizer(if_f); 3507 set_loop(if_f, l); 3508 // Find frame ptr for Halt. Relies on the optimizer 3509 // V-N'ing. Easier and quicker than searching through 3510 // the program structure. 3511 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr ); 3512 _igvn.register_new_node_with_optimizer(frame); 3513 // Halt & Catch Fire 3514 Node *halt = new HaltNode( if_f, frame ); 3515 _igvn.register_new_node_with_optimizer(halt); 3516 set_loop(halt, l); 3517 C->root()->add_req(halt); 3518 } 3519 set_loop(C->root(), _ltree_root); 3520 } 3521 } 3522 // Weeny check for irreducible. This child was already visited (this 3523 // IS the post-work phase). Is this child's loop header post-visited 3524 // as well? If so, then I found another entry into the loop. 3525 if (!_verify_only) { 3526 while( is_postvisited(l->_head) ) { 3527 // found irreducible 3528 l->_irreducible = 1; // = true 3529 l = l->_parent; 3530 _has_irreducible_loops = true; 3531 // Check for bad CFG here to prevent crash, and bailout of compile 3532 if (l == NULL) { 3533 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); 3534 return pre_order; 3535 } 3536 } 3537 C->set_has_irreducible_loop(_has_irreducible_loops); 3538 } 3539 3540 // This Node might be a decision point for loops. It is only if 3541 // it's children belong to several different loops. The sort call 3542 // does a trivial amount of work if there is only 1 child or all 3543 // children belong to the same loop. If however, the children 3544 // belong to different loops, the sort call will properly set the 3545 // _parent pointers to show how the loops nest. 3546 // 3547 // In any case, it returns the tightest enclosing loop. 3548 innermost = sort( l, innermost ); 3549 } 3550 3551 // Def-use info will have some dead stuff; dead stuff will have no 3552 // loop decided on. 3553 3554 // Am I a loop header? If so fix up my parent's child and next ptrs. 3555 if( innermost && innermost->_head == n ) { 3556 assert( get_loop(n) == innermost, "" ); 3557 IdealLoopTree *p = innermost->_parent; 3558 IdealLoopTree *l = innermost; 3559 while( p && l->_head == n ) { 3560 l->_next = p->_child; // Put self on parents 'next child' 3561 p->_child = l; // Make self as first child of parent 3562 l = p; // Now walk up the parent chain 3563 p = l->_parent; 3564 } 3565 } else { 3566 // Note that it is possible for a LoopNode to reach here, if the 3567 // backedge has been made unreachable (hence the LoopNode no longer 3568 // denotes a Loop, and will eventually be removed). 3569 3570 // Record tightest enclosing loop for self. Mark as post-visited. 3571 set_loop(n, innermost); 3572 // Also record has_call flag early on 3573 if( innermost ) { 3574 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { 3575 // Do not count uncommon calls 3576 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { 3577 Node *iff = n->in(0)->in(0); 3578 // No any calls for vectorized loops. 3579 if( UseSuperWord || !iff->is_If() || 3580 (n->in(0)->Opcode() == Op_IfFalse && 3581 (1.0 - iff->as_If()->_prob) >= 0.01) || 3582 (iff->as_If()->_prob >= 0.01) ) 3583 innermost->_has_call = 1; 3584 } 3585 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { 3586 // Disable loop optimizations if the loop has a scalar replaceable 3587 // allocation. This disabling may cause a potential performance lost 3588 // if the allocation is not eliminated for some reason. 3589 innermost->_allow_optimizations = false; 3590 innermost->_has_call = 1; // = true 3591 } else if (n->Opcode() == Op_SafePoint) { 3592 // Record all safepoints in this loop. 3593 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List(); 3594 innermost->_safepts->push(n); 3595 } 3596 } 3597 } 3598 3599 // Flag as post-visited now 3600 set_postvisited(n); 3601 return pre_order; 3602 } 3603 3604 3605 //------------------------------build_loop_early------------------------------- 3606 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3607 // First pass computes the earliest controlling node possible. This is the 3608 // controlling input with the deepest dominating depth. 3609 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3610 while (worklist.size() != 0) { 3611 // Use local variables nstack_top_n & nstack_top_i to cache values 3612 // on nstack's top. 3613 Node *nstack_top_n = worklist.pop(); 3614 uint nstack_top_i = 0; 3615 //while_nstack_nonempty: 3616 while (true) { 3617 // Get parent node and next input's index from stack's top. 3618 Node *n = nstack_top_n; 3619 uint i = nstack_top_i; 3620 uint cnt = n->req(); // Count of inputs 3621 if (i == 0) { // Pre-process the node. 3622 if( has_node(n) && // Have either loop or control already? 3623 !has_ctrl(n) ) { // Have loop picked out already? 3624 // During "merge_many_backedges" we fold up several nested loops 3625 // into a single loop. This makes the members of the original 3626 // loop bodies pointing to dead loops; they need to move up 3627 // to the new UNION'd larger loop. I set the _head field of these 3628 // dead loops to NULL and the _parent field points to the owning 3629 // loop. Shades of UNION-FIND algorithm. 3630 IdealLoopTree *ilt; 3631 while( !(ilt = get_loop(n))->_head ) { 3632 // Normally I would use a set_loop here. But in this one special 3633 // case, it is legal (and expected) to change what loop a Node 3634 // belongs to. 3635 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); 3636 } 3637 // Remove safepoints ONLY if I've already seen I don't need one. 3638 // (the old code here would yank a 2nd safepoint after seeing a 3639 // first one, even though the 1st did not dominate in the loop body 3640 // and thus could be avoided indefinitely) 3641 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && 3642 is_deleteable_safept(n)) { 3643 Node *in = n->in(TypeFunc::Control); 3644 lazy_replace(n,in); // Pull safepoint now 3645 if (ilt->_safepts != NULL) { 3646 ilt->_safepts->yank(n); 3647 } 3648 // Carry on with the recursion "as if" we are walking 3649 // only the control input 3650 if( !visited.test_set( in->_idx ) ) { 3651 worklist.push(in); // Visit this guy later, using worklist 3652 } 3653 // Get next node from nstack: 3654 // - skip n's inputs processing by setting i > cnt; 3655 // - we also will not call set_early_ctrl(n) since 3656 // has_node(n) == true (see the condition above). 3657 i = cnt + 1; 3658 } 3659 } 3660 } // if (i == 0) 3661 3662 // Visit all inputs 3663 bool done = true; // Assume all n's inputs will be processed 3664 while (i < cnt) { 3665 Node *in = n->in(i); 3666 ++i; 3667 if (in == NULL) continue; 3668 if (in->pinned() && !in->is_CFG()) 3669 set_ctrl(in, in->in(0)); 3670 int is_visited = visited.test_set( in->_idx ); 3671 if (!has_node(in)) { // No controlling input yet? 3672 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); 3673 assert( !is_visited, "visit only once" ); 3674 nstack.push(n, i); // Save parent node and next input's index. 3675 nstack_top_n = in; // Process current input now. 3676 nstack_top_i = 0; 3677 done = false; // Not all n's inputs processed. 3678 break; // continue while_nstack_nonempty; 3679 } else if (!is_visited) { 3680 // This guy has a location picked out for him, but has not yet 3681 // been visited. Happens to all CFG nodes, for instance. 3682 // Visit him using the worklist instead of recursion, to break 3683 // cycles. Since he has a location already we do not need to 3684 // find his location before proceeding with the current Node. 3685 worklist.push(in); // Visit this guy later, using worklist 3686 } 3687 } 3688 if (done) { 3689 // All of n's inputs have been processed, complete post-processing. 3690 3691 // Compute earliest point this Node can go. 3692 // CFG, Phi, pinned nodes already know their controlling input. 3693 if (!has_node(n)) { 3694 // Record earliest legal location 3695 set_early_ctrl( n ); 3696 } 3697 if (nstack.is_empty()) { 3698 // Finished all nodes on stack. 3699 // Process next node on the worklist. 3700 break; 3701 } 3702 // Get saved parent node and next input's index. 3703 nstack_top_n = nstack.node(); 3704 nstack_top_i = nstack.index(); 3705 nstack.pop(); 3706 } 3707 } // while (true) 3708 } 3709 } 3710 3711 //------------------------------dom_lca_internal-------------------------------- 3712 // Pair-wise LCA 3713 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { 3714 if( !n1 ) return n2; // Handle NULL original LCA 3715 assert( n1->is_CFG(), "" ); 3716 assert( n2->is_CFG(), "" ); 3717 // find LCA of all uses 3718 uint d1 = dom_depth(n1); 3719 uint d2 = dom_depth(n2); 3720 while (n1 != n2) { 3721 if (d1 > d2) { 3722 n1 = idom(n1); 3723 d1 = dom_depth(n1); 3724 } else if (d1 < d2) { 3725 n2 = idom(n2); 3726 d2 = dom_depth(n2); 3727 } else { 3728 // Here d1 == d2. Due to edits of the dominator-tree, sections 3729 // of the tree might have the same depth. These sections have 3730 // to be searched more carefully. 3731 3732 // Scan up all the n1's with equal depth, looking for n2. 3733 Node *t1 = idom(n1); 3734 while (dom_depth(t1) == d1) { 3735 if (t1 == n2) return n2; 3736 t1 = idom(t1); 3737 } 3738 // Scan up all the n2's with equal depth, looking for n1. 3739 Node *t2 = idom(n2); 3740 while (dom_depth(t2) == d2) { 3741 if (t2 == n1) return n1; 3742 t2 = idom(t2); 3743 } 3744 // Move up to a new dominator-depth value as well as up the dom-tree. 3745 n1 = t1; 3746 n2 = t2; 3747 d1 = dom_depth(n1); 3748 d2 = dom_depth(n2); 3749 } 3750 } 3751 return n1; 3752 } 3753 3754 //------------------------------compute_idom----------------------------------- 3755 // Locally compute IDOM using dom_lca call. Correct only if the incoming 3756 // IDOMs are correct. 3757 Node *PhaseIdealLoop::compute_idom( Node *region ) const { 3758 assert( region->is_Region(), "" ); 3759 Node *LCA = NULL; 3760 for( uint i = 1; i < region->req(); i++ ) { 3761 if( region->in(i) != C->top() ) 3762 LCA = dom_lca( LCA, region->in(i) ); 3763 } 3764 return LCA; 3765 } 3766 3767 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { 3768 bool had_error = false; 3769 #ifdef ASSERT 3770 if (early != C->root()) { 3771 // Make sure that there's a dominance path from LCA to early 3772 Node* d = LCA; 3773 while (d != early) { 3774 if (d == C->root()) { 3775 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA); 3776 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx); 3777 had_error = true; 3778 break; 3779 } 3780 d = idom(d); 3781 } 3782 } 3783 #endif 3784 return had_error; 3785 } 3786 3787 3788 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { 3789 // Compute LCA over list of uses 3790 bool had_error = false; 3791 Node *LCA = NULL; 3792 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { 3793 Node* c = n->fast_out(i); 3794 if (_nodes[c->_idx] == NULL) 3795 continue; // Skip the occasional dead node 3796 if( c->is_Phi() ) { // For Phis, we must land above on the path 3797 for( uint j=1; j<c->req(); j++ ) {// For all inputs 3798 if( c->in(j) == n ) { // Found matching input? 3799 Node *use = c->in(0)->in(j); 3800 if (_verify_only && use->is_top()) continue; 3801 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3802 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3803 } 3804 } 3805 } else { 3806 // For CFG data-users, use is in the block just prior 3807 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); 3808 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3809 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3810 } 3811 } 3812 assert(!had_error, "bad dominance"); 3813 return LCA; 3814 } 3815 3816 // Check the shape of the graph at the loop entry. In some cases, 3817 // the shape of the graph does not match the shape outlined below. 3818 // That is caused by the Opaque1 node "protecting" the shape of 3819 // the graph being removed by, for example, the IGVN performed 3820 // in PhaseIdealLoop::build_and_optimize(). 3821 // 3822 // After the Opaque1 node has been removed, optimizations (e.g., split-if, 3823 // loop unswitching, and IGVN, or a combination of them) can freely change 3824 // the graph's shape. As a result, the graph shape outlined below cannot 3825 // be guaranteed anymore. 3826 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) { 3827 if (!cl->is_main_loop() && !cl->is_post_loop()) { 3828 return false; 3829 } 3830 Node* ctrl = cl->skip_predicates(); 3831 3832 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) { 3833 return false; 3834 } 3835 Node* iffm = ctrl->in(0); 3836 if (iffm == NULL || !iffm->is_If()) { 3837 return false; 3838 } 3839 Node* bolzm = iffm->in(1); 3840 if (bolzm == NULL || !bolzm->is_Bool()) { 3841 return false; 3842 } 3843 Node* cmpzm = bolzm->in(1); 3844 if (cmpzm == NULL || !cmpzm->is_Cmp()) { 3845 return false; 3846 } 3847 // compares can get conditionally flipped 3848 bool found_opaque = false; 3849 for (uint i = 1; i < cmpzm->req(); i++) { 3850 Node* opnd = cmpzm->in(i); 3851 if (opnd && opnd->Opcode() == Op_Opaque1) { 3852 found_opaque = true; 3853 break; 3854 } 3855 } 3856 if (!found_opaque) { 3857 return false; 3858 } 3859 return true; 3860 } 3861 3862 //------------------------------get_late_ctrl---------------------------------- 3863 // Compute latest legal control. 3864 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { 3865 assert(early != NULL, "early control should not be NULL"); 3866 3867 Node* LCA = compute_lca_of_uses(n, early); 3868 #ifdef ASSERT 3869 if (LCA == C->root() && LCA != early) { 3870 // def doesn't dominate uses so print some useful debugging output 3871 compute_lca_of_uses(n, early, true); 3872 } 3873 #endif 3874 3875 // if this is a load, check for anti-dependent stores 3876 // We use a conservative algorithm to identify potential interfering 3877 // instructions and for rescheduling the load. The users of the memory 3878 // input of this load are examined. Any use which is not a load and is 3879 // dominated by early is considered a potentially interfering store. 3880 // This can produce false positives. 3881 if (n->is_Load() && LCA != early) { 3882 Node_List worklist; 3883 3884 Node *mem = n->in(MemNode::Memory); 3885 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 3886 Node* s = mem->fast_out(i); 3887 worklist.push(s); 3888 } 3889 while(worklist.size() != 0 && LCA != early) { 3890 Node* s = worklist.pop(); 3891 if (s->is_Load() || s->Opcode() == Op_SafePoint) { 3892 continue; 3893 } else if (s->is_MergeMem()) { 3894 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 3895 Node* s1 = s->fast_out(i); 3896 worklist.push(s1); 3897 } 3898 } else { 3899 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); 3900 assert(sctrl != NULL || s->outcnt() == 0, "must have control"); 3901 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { 3902 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); 3903 } 3904 } 3905 } 3906 } 3907 3908 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); 3909 return LCA; 3910 } 3911 3912 // true if CFG node d dominates CFG node n 3913 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { 3914 if (d == n) 3915 return true; 3916 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 3917 uint dd = dom_depth(d); 3918 while (dom_depth(n) >= dd) { 3919 if (n == d) 3920 return true; 3921 n = idom(n); 3922 } 3923 return false; 3924 } 3925 3926 //------------------------------dom_lca_for_get_late_ctrl_internal------------- 3927 // Pair-wise LCA with tags. 3928 // Tag each index with the node 'tag' currently being processed 3929 // before advancing up the dominator chain using idom(). 3930 // Later calls that find a match to 'tag' know that this path has already 3931 // been considered in the current LCA (which is input 'n1' by convention). 3932 // Since get_late_ctrl() is only called once for each node, the tag array 3933 // does not need to be cleared between calls to get_late_ctrl(). 3934 // Algorithm trades a larger constant factor for better asymptotic behavior 3935 // 3936 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { 3937 uint d1 = dom_depth(n1); 3938 uint d2 = dom_depth(n2); 3939 3940 do { 3941 if (d1 > d2) { 3942 // current lca is deeper than n2 3943 _dom_lca_tags.map(n1->_idx, tag); 3944 n1 = idom(n1); 3945 d1 = dom_depth(n1); 3946 } else if (d1 < d2) { 3947 // n2 is deeper than current lca 3948 Node *memo = _dom_lca_tags[n2->_idx]; 3949 if( memo == tag ) { 3950 return n1; // Return the current LCA 3951 } 3952 _dom_lca_tags.map(n2->_idx, tag); 3953 n2 = idom(n2); 3954 d2 = dom_depth(n2); 3955 } else { 3956 // Here d1 == d2. Due to edits of the dominator-tree, sections 3957 // of the tree might have the same depth. These sections have 3958 // to be searched more carefully. 3959 3960 // Scan up all the n1's with equal depth, looking for n2. 3961 _dom_lca_tags.map(n1->_idx, tag); 3962 Node *t1 = idom(n1); 3963 while (dom_depth(t1) == d1) { 3964 if (t1 == n2) return n2; 3965 _dom_lca_tags.map(t1->_idx, tag); 3966 t1 = idom(t1); 3967 } 3968 // Scan up all the n2's with equal depth, looking for n1. 3969 _dom_lca_tags.map(n2->_idx, tag); 3970 Node *t2 = idom(n2); 3971 while (dom_depth(t2) == d2) { 3972 if (t2 == n1) return n1; 3973 _dom_lca_tags.map(t2->_idx, tag); 3974 t2 = idom(t2); 3975 } 3976 // Move up to a new dominator-depth value as well as up the dom-tree. 3977 n1 = t1; 3978 n2 = t2; 3979 d1 = dom_depth(n1); 3980 d2 = dom_depth(n2); 3981 } 3982 } while (n1 != n2); 3983 return n1; 3984 } 3985 3986 //------------------------------init_dom_lca_tags------------------------------ 3987 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3988 // Intended use does not involve any growth for the array, so it could 3989 // be of fixed size. 3990 void PhaseIdealLoop::init_dom_lca_tags() { 3991 uint limit = C->unique() + 1; 3992 _dom_lca_tags.map( limit, NULL ); 3993 #ifdef ASSERT 3994 for( uint i = 0; i < limit; ++i ) { 3995 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3996 } 3997 #endif // ASSERT 3998 } 3999 4000 //------------------------------clear_dom_lca_tags------------------------------ 4001 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 4002 // Intended use does not involve any growth for the array, so it could 4003 // be of fixed size. 4004 void PhaseIdealLoop::clear_dom_lca_tags() { 4005 uint limit = C->unique() + 1; 4006 _dom_lca_tags.map( limit, NULL ); 4007 _dom_lca_tags.clear(); 4008 #ifdef ASSERT 4009 for( uint i = 0; i < limit; ++i ) { 4010 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 4011 } 4012 #endif // ASSERT 4013 } 4014 4015 //------------------------------build_loop_late-------------------------------- 4016 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 4017 // Second pass finds latest legal placement, and ideal loop placement. 4018 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 4019 while (worklist.size() != 0) { 4020 Node *n = worklist.pop(); 4021 // Only visit once 4022 if (visited.test_set(n->_idx)) continue; 4023 uint cnt = n->outcnt(); 4024 uint i = 0; 4025 while (true) { 4026 assert( _nodes[n->_idx], "no dead nodes" ); 4027 // Visit all children 4028 if (i < cnt) { 4029 Node* use = n->raw_out(i); 4030 ++i; 4031 // Check for dead uses. Aggressively prune such junk. It might be 4032 // dead in the global sense, but still have local uses so I cannot 4033 // easily call 'remove_dead_node'. 4034 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? 4035 // Due to cycles, we might not hit the same fixed point in the verify 4036 // pass as we do in the regular pass. Instead, visit such phis as 4037 // simple uses of the loop head. 4038 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { 4039 if( !visited.test(use->_idx) ) 4040 worklist.push(use); 4041 } else if( !visited.test_set(use->_idx) ) { 4042 nstack.push(n, i); // Save parent and next use's index. 4043 n = use; // Process all children of current use. 4044 cnt = use->outcnt(); 4045 i = 0; 4046 } 4047 } else { 4048 // Do not visit around the backedge of loops via data edges. 4049 // push dead code onto a worklist 4050 _deadlist.push(use); 4051 } 4052 } else { 4053 // All of n's children have been processed, complete post-processing. 4054 build_loop_late_post(n); 4055 if (nstack.is_empty()) { 4056 // Finished all nodes on stack. 4057 // Process next node on the worklist. 4058 break; 4059 } 4060 // Get saved parent node and next use's index. Visit the rest of uses. 4061 n = nstack.node(); 4062 cnt = n->outcnt(); 4063 i = nstack.index(); 4064 nstack.pop(); 4065 } 4066 } 4067 } 4068 } 4069 4070 // Verify that no data node is schedules in the outer loop of a strip 4071 // mined loop. 4072 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) { 4073 #ifdef ASSERT 4074 if (get_loop(least)->_nest == 0) { 4075 return; 4076 } 4077 IdealLoopTree* loop = get_loop(least); 4078 Node* head = loop->_head; 4079 if (head->is_OuterStripMinedLoop()) { 4080 Node* sfpt = head->as_Loop()->outer_safepoint(); 4081 ResourceMark rm; 4082 Unique_Node_List wq; 4083 wq.push(sfpt); 4084 for (uint i = 0; i < wq.size(); i++) { 4085 Node *m = wq.at(i); 4086 for (uint i = 1; i < m->req(); i++) { 4087 Node* nn = m->in(i); 4088 if (nn == n) { 4089 return; 4090 } 4091 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) { 4092 wq.push(nn); 4093 } 4094 } 4095 } 4096 ShouldNotReachHere(); 4097 } 4098 #endif 4099 } 4100 4101 4102 //------------------------------build_loop_late_post--------------------------- 4103 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 4104 // Second pass finds latest legal placement, and ideal loop placement. 4105 void PhaseIdealLoop::build_loop_late_post( Node *n ) { 4106 4107 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) { 4108 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. 4109 } 4110 4111 #ifdef ASSERT 4112 if (_verify_only && !n->is_CFG()) { 4113 // Check def-use domination. 4114 compute_lca_of_uses(n, get_ctrl(n), true /* verify */); 4115 } 4116 #endif 4117 4118 // CFG and pinned nodes already handled 4119 if( n->in(0) ) { 4120 if( n->in(0)->is_top() ) return; // Dead? 4121 4122 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads 4123 // _must_ be pinned (they have to observe their control edge of course). 4124 // Unlike Stores (which modify an unallocable resource, the memory 4125 // state), Mods/Loads can float around. So free them up. 4126 bool pinned = true; 4127 switch( n->Opcode() ) { 4128 case Op_DivI: 4129 case Op_DivF: 4130 case Op_DivD: 4131 case Op_ModI: 4132 case Op_ModF: 4133 case Op_ModD: 4134 case Op_LoadB: // Same with Loads; they can sink 4135 case Op_LoadUB: // during loop optimizations. 4136 case Op_LoadUS: 4137 case Op_LoadD: 4138 case Op_LoadF: 4139 case Op_LoadI: 4140 case Op_LoadKlass: 4141 case Op_LoadNKlass: 4142 case Op_LoadL: 4143 case Op_LoadS: 4144 case Op_LoadP: 4145 case Op_LoadBarrierSlowReg: 4146 case Op_LoadBarrierWeakSlowReg: 4147 case Op_LoadN: 4148 case Op_LoadRange: 4149 case Op_LoadD_unaligned: 4150 case Op_LoadL_unaligned: 4151 case Op_StrComp: // Does a bunch of load-like effects 4152 case Op_StrEquals: 4153 case Op_StrIndexOf: 4154 case Op_StrIndexOfChar: 4155 case Op_AryEq: 4156 case Op_HasNegatives: 4157 pinned = false; 4158 } 4159 if( pinned ) { 4160 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); 4161 if( !chosen_loop->_child ) // Inner loop? 4162 chosen_loop->_body.push(n); // Collect inner loops 4163 return; 4164 } 4165 } else { // No slot zero 4166 if( n->is_CFG() ) { // CFG with no slot 0 is dead 4167 _nodes.map(n->_idx,0); // No block setting, it's globally dead 4168 return; 4169 } 4170 assert(!n->is_CFG() || n->outcnt() == 0, ""); 4171 } 4172 4173 // Do I have a "safe range" I can select over? 4174 Node *early = get_ctrl(n);// Early location already computed 4175 4176 // Compute latest point this Node can go 4177 Node *LCA = get_late_ctrl( n, early ); 4178 // LCA is NULL due to uses being dead 4179 if( LCA == NULL ) { 4180 #ifdef ASSERT 4181 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { 4182 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); 4183 } 4184 #endif 4185 _nodes.map(n->_idx, 0); // This node is useless 4186 _deadlist.push(n); 4187 return; 4188 } 4189 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); 4190 4191 Node *legal = LCA; // Walk 'legal' up the IDOM chain 4192 Node *least = legal; // Best legal position so far 4193 while( early != legal ) { // While not at earliest legal 4194 #ifdef ASSERT 4195 if (legal->is_Start() && !early->is_Root()) { 4196 // Bad graph. Print idom path and fail. 4197 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA); 4198 assert(false, "Bad graph detected in build_loop_late"); 4199 } 4200 #endif 4201 // Find least loop nesting depth 4202 legal = idom(legal); // Bump up the IDOM tree 4203 // Check for lower nesting depth 4204 if( get_loop(legal)->_nest < get_loop(least)->_nest ) 4205 least = legal; 4206 } 4207 assert(early == legal || legal != C->root(), "bad dominance of inputs"); 4208 4209 // Try not to place code on a loop entry projection 4210 // which can inhibit range check elimination. 4211 if (least != early) { 4212 Node* ctrl_out = least->unique_ctrl_out(); 4213 if (ctrl_out && ctrl_out->is_Loop() && 4214 least == ctrl_out->in(LoopNode::EntryControl) && 4215 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) { 4216 Node* least_dom = idom(least); 4217 if (get_loop(least_dom)->is_member(get_loop(least))) { 4218 least = least_dom; 4219 } 4220 } 4221 } 4222 4223 #ifdef ASSERT 4224 // If verifying, verify that 'verify_me' has a legal location 4225 // and choose it as our location. 4226 if( _verify_me ) { 4227 Node *v_ctrl = _verify_me->get_ctrl_no_update(n); 4228 Node *legal = LCA; 4229 while( early != legal ) { // While not at earliest legal 4230 if( legal == v_ctrl ) break; // Check for prior good location 4231 legal = idom(legal) ;// Bump up the IDOM tree 4232 } 4233 // Check for prior good location 4234 if( legal == v_ctrl ) least = legal; // Keep prior if found 4235 } 4236 #endif 4237 4238 // Assign discovered "here or above" point 4239 least = find_non_split_ctrl(least); 4240 verify_strip_mined_scheduling(n, least); 4241 set_ctrl(n, least); 4242 4243 // Collect inner loop bodies 4244 IdealLoopTree *chosen_loop = get_loop(least); 4245 if( !chosen_loop->_child ) // Inner loop? 4246 chosen_loop->_body.push(n);// Collect inner loops 4247 } 4248 4249 #ifdef ASSERT 4250 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) { 4251 tty->print_cr("%s", msg); 4252 tty->print("n: "); n->dump(); 4253 tty->print("early(n): "); early->dump(); 4254 if (n->in(0) != NULL && !n->in(0)->is_top() && 4255 n->in(0) != early && !n->in(0)->is_Root()) { 4256 tty->print("n->in(0): "); n->in(0)->dump(); 4257 } 4258 for (uint i = 1; i < n->req(); i++) { 4259 Node* in1 = n->in(i); 4260 if (in1 != NULL && in1 != n && !in1->is_top()) { 4261 tty->print("n->in(%d): ", i); in1->dump(); 4262 Node* in1_early = get_ctrl(in1); 4263 tty->print("early(n->in(%d)): ", i); in1_early->dump(); 4264 if (in1->in(0) != NULL && !in1->in(0)->is_top() && 4265 in1->in(0) != in1_early && !in1->in(0)->is_Root()) { 4266 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump(); 4267 } 4268 for (uint j = 1; j < in1->req(); j++) { 4269 Node* in2 = in1->in(j); 4270 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) { 4271 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump(); 4272 Node* in2_early = get_ctrl(in2); 4273 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); 4274 if (in2->in(0) != NULL && !in2->in(0)->is_top() && 4275 in2->in(0) != in2_early && !in2->in(0)->is_Root()) { 4276 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump(); 4277 } 4278 } 4279 } 4280 } 4281 } 4282 tty->cr(); 4283 tty->print("LCA(n): "); LCA->dump(); 4284 for (uint i = 0; i < n->outcnt(); i++) { 4285 Node* u1 = n->raw_out(i); 4286 if (u1 == n) 4287 continue; 4288 tty->print("n->out(%d): ", i); u1->dump(); 4289 if (u1->is_CFG()) { 4290 for (uint j = 0; j < u1->outcnt(); j++) { 4291 Node* u2 = u1->raw_out(j); 4292 if (u2 != u1 && u2 != n && u2->is_CFG()) { 4293 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4294 } 4295 } 4296 } else { 4297 Node* u1_later = get_ctrl(u1); 4298 tty->print("later(n->out(%d)): ", i); u1_later->dump(); 4299 if (u1->in(0) != NULL && !u1->in(0)->is_top() && 4300 u1->in(0) != u1_later && !u1->in(0)->is_Root()) { 4301 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump(); 4302 } 4303 for (uint j = 0; j < u1->outcnt(); j++) { 4304 Node* u2 = u1->raw_out(j); 4305 if (u2 == n || u2 == u1) 4306 continue; 4307 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4308 if (!u2->is_CFG()) { 4309 Node* u2_later = get_ctrl(u2); 4310 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); 4311 if (u2->in(0) != NULL && !u2->in(0)->is_top() && 4312 u2->in(0) != u2_later && !u2->in(0)->is_Root()) { 4313 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump(); 4314 } 4315 } 4316 } 4317 } 4318 } 4319 tty->cr(); 4320 int ct = 0; 4321 Node *dbg_legal = LCA; 4322 while(!dbg_legal->is_Start() && ct < 100) { 4323 tty->print("idom[%d] ",ct); dbg_legal->dump(); 4324 ct++; 4325 dbg_legal = idom(dbg_legal); 4326 } 4327 tty->cr(); 4328 } 4329 #endif 4330 4331 #ifndef PRODUCT 4332 //------------------------------dump------------------------------------------- 4333 void PhaseIdealLoop::dump( ) const { 4334 ResourceMark rm; 4335 Arena* arena = Thread::current()->resource_area(); 4336 Node_Stack stack(arena, C->live_nodes() >> 2); 4337 Node_List rpo_list; 4338 VectorSet visited(arena); 4339 visited.set(C->top()->_idx); 4340 rpo( C->root(), stack, visited, rpo_list ); 4341 // Dump root loop indexed by last element in PO order 4342 dump( _ltree_root, rpo_list.size(), rpo_list ); 4343 } 4344 4345 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { 4346 loop->dump_head(); 4347 4348 // Now scan for CFG nodes in the same loop 4349 for( uint j=idx; j > 0; j-- ) { 4350 Node *n = rpo_list[j-1]; 4351 if( !_nodes[n->_idx] ) // Skip dead nodes 4352 continue; 4353 if( get_loop(n) != loop ) { // Wrong loop nest 4354 if( get_loop(n)->_head == n && // Found nested loop? 4355 get_loop(n)->_parent == loop ) 4356 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly 4357 continue; 4358 } 4359 4360 // Dump controlling node 4361 for( uint x = 0; x < loop->_nest; x++ ) 4362 tty->print(" "); 4363 tty->print("C"); 4364 if( n == C->root() ) { 4365 n->dump(); 4366 } else { 4367 Node* cached_idom = idom_no_update(n); 4368 Node *computed_idom = n->in(0); 4369 if( n->is_Region() ) { 4370 computed_idom = compute_idom(n); 4371 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or 4372 // any MultiBranch ctrl node), so apply a similar transform to 4373 // the cached idom returned from idom_no_update. 4374 cached_idom = find_non_split_ctrl(cached_idom); 4375 } 4376 tty->print(" ID:%d",computed_idom->_idx); 4377 n->dump(); 4378 if( cached_idom != computed_idom ) { 4379 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", 4380 computed_idom->_idx, cached_idom->_idx); 4381 } 4382 } 4383 // Dump nodes it controls 4384 for( uint k = 0; k < _nodes.Size(); k++ ) { 4385 // (k < C->unique() && get_ctrl(find(k)) == n) 4386 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { 4387 Node *m = C->root()->find(k); 4388 if( m && m->outcnt() > 0 ) { 4389 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { 4390 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", 4391 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); 4392 } 4393 for( uint j = 0; j < loop->_nest; j++ ) 4394 tty->print(" "); 4395 tty->print(" "); 4396 m->dump(); 4397 } 4398 } 4399 } 4400 } 4401 } 4402 4403 // Collect a R-P-O for the whole CFG. 4404 // Result list is in post-order (scan backwards for RPO) 4405 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { 4406 stk.push(start, 0); 4407 visited.set(start->_idx); 4408 4409 while (stk.is_nonempty()) { 4410 Node* m = stk.node(); 4411 uint idx = stk.index(); 4412 if (idx < m->outcnt()) { 4413 stk.set_index(idx + 1); 4414 Node* n = m->raw_out(idx); 4415 if (n->is_CFG() && !visited.test_set(n->_idx)) { 4416 stk.push(n, 0); 4417 } 4418 } else { 4419 rpo_list.push(m); 4420 stk.pop(); 4421 } 4422 } 4423 } 4424 #endif 4425 4426 4427 //============================================================================= 4428 //------------------------------LoopTreeIterator----------------------------------- 4429 4430 // Advance to next loop tree using a preorder, left-to-right traversal. 4431 void LoopTreeIterator::next() { 4432 assert(!done(), "must not be done."); 4433 if (_curnt->_child != NULL) { 4434 _curnt = _curnt->_child; 4435 } else if (_curnt->_next != NULL) { 4436 _curnt = _curnt->_next; 4437 } else { 4438 while (_curnt != _root && _curnt->_next == NULL) { 4439 _curnt = _curnt->_parent; 4440 } 4441 if (_curnt == _root) { 4442 _curnt = NULL; 4443 assert(done(), "must be done."); 4444 } else { 4445 assert(_curnt->_next != NULL, "must be more to do"); 4446 _curnt = _curnt->_next; 4447 } 4448 } 4449 }