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