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 if (be->is_Store() && old_new[be->_idx] != NULL) { 1398 assert(false, "store on the backedge + sunk stores: unsupported"); 1399 // drop outer loop 1400 IfNode* outer_le = outer_loop_end(); 1401 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1402 igvn->replace_node(outer_le, iff); 1403 inner_cl->clear_strip_mined(); 1404 return; 1405 } 1406 if (be == last || be == first->in(MemNode::Memory)) { 1407 assert(phi == NULL, "only one phi"); 1408 phi = uu; 1409 } 1410 } 1411 } 1412 #ifdef ASSERT 1413 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1414 Node* uu = fast_out(j); 1415 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) { 1416 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) { 1417 assert(phi == uu, "what's that phi?"); 1418 } else if (uu->adr_type() == TypePtr::BOTTOM) { 1419 Node* n = uu->in(LoopNode::LoopBackControl); 1420 uint limit = igvn->C->live_nodes(); 1421 uint i = 0; 1422 while (n != uu) { 1423 i++; 1424 assert(i < limit, "infinite loop"); 1425 if (n->is_Proj()) { 1426 n = n->in(0); 1427 } else if (n->is_SafePoint() || n->is_MemBar()) { 1428 n = n->in(TypeFunc::Memory); 1429 } else if (n->is_Phi()) { 1430 n = n->in(1); 1431 } else if (n->is_MergeMem()) { 1432 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type())); 1433 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) { 1434 n = n->in(MemNode::Memory); 1435 } else { 1436 n->dump(); 1437 ShouldNotReachHere(); 1438 } 1439 } 1440 } 1441 } 1442 } 1443 #endif 1444 if (phi == NULL) { 1445 // If the an entire chains was sunk, the 1446 // inner loop has no phi for that memory 1447 // slice, create one for the outer loop 1448 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY, 1449 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))); 1450 phi->set_req(LoopNode::LoopBackControl, last); 1451 phi = igvn->transform(phi); 1452 igvn->replace_input_of(first, MemNode::Memory, phi); 1453 } else { 1454 // Or fix the outer loop fix to include 1455 // that chain of stores. 1456 Node* be = phi->in(LoopNode::LoopBackControl); 1457 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported"); 1458 if (be == first->in(MemNode::Memory)) { 1459 if (be == phi->in(LoopNode::LoopBackControl)) { 1460 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last); 1461 } else { 1462 igvn->replace_input_of(be, MemNode::Memory, last); 1463 } 1464 } else { 1465 #ifdef ASSERT 1466 if (be == phi->in(LoopNode::LoopBackControl)) { 1467 assert(phi->in(LoopNode::LoopBackControl) == last, ""); 1468 } else { 1469 assert(be->in(MemNode::Memory) == last, ""); 1470 } 1471 #endif 1472 } 1473 } 1474 } 1475 } 1476 } 1477 1478 if (iv_phi != NULL) { 1479 // Now adjust the inner loop's exit condition 1480 Node* limit = inner_cl->limit(); 1481 Node* sub = NULL; 1482 if (stride > 0) { 1483 sub = igvn->transform(new SubINode(limit, iv_phi)); 1484 } else { 1485 sub = igvn->transform(new SubINode(iv_phi, limit)); 1486 } 1487 Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters))); 1488 Node* new_limit = NULL; 1489 if (stride > 0) { 1490 new_limit = igvn->transform(new AddINode(min, iv_phi)); 1491 } else { 1492 new_limit = igvn->transform(new SubINode(iv_phi, min)); 1493 } 1494 igvn->replace_input_of(inner_cle->cmp_node(), 2, new_limit); 1495 Node* cmp = inner_cle->cmp_node()->clone(); 1496 Node* bol = inner_cle->in(CountedLoopEndNode::TestValue)->clone(); 1497 cmp->set_req(2, limit); 1498 bol->set_req(1, igvn->transform(cmp)); 1499 igvn->replace_input_of(outer_loop_end(), 1, igvn->transform(bol)); 1500 } else { 1501 assert(false, "should be able to adjust outer loop"); 1502 IfNode* outer_le = outer_loop_end(); 1503 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1504 igvn->replace_node(outer_le, iff); 1505 inner_cl->clear_strip_mined(); 1506 } 1507 } 1508 1509 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const { 1510 if (!in(0)) return Type::TOP; 1511 if (phase->type(in(0)) == Type::TOP) 1512 return Type::TOP; 1513 1514 return TypeTuple::IFBOTH; 1515 } 1516 1517 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1518 if (remove_dead_region(phase, can_reshape)) return this; 1519 1520 return NULL; 1521 } 1522 1523 //------------------------------filtered_type-------------------------------- 1524 // Return a type based on condition control flow 1525 // A successful return will be a type that is restricted due 1526 // to a series of dominating if-tests, such as: 1527 // if (i < 10) { 1528 // if (i > 0) { 1529 // here: "i" type is [1..10) 1530 // } 1531 // } 1532 // or a control flow merge 1533 // if (i < 10) { 1534 // do { 1535 // phi( , ) -- at top of loop type is [min_int..10) 1536 // i = ? 1537 // } while ( i < 10) 1538 // 1539 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { 1540 assert(n && n->bottom_type()->is_int(), "must be int"); 1541 const TypeInt* filtered_t = NULL; 1542 if (!n->is_Phi()) { 1543 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); 1544 filtered_t = filtered_type_from_dominators(n, n_ctrl); 1545 1546 } else { 1547 Node* phi = n->as_Phi(); 1548 Node* region = phi->in(0); 1549 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); 1550 if (region && region != C->top()) { 1551 for (uint i = 1; i < phi->req(); i++) { 1552 Node* val = phi->in(i); 1553 Node* use_c = region->in(i); 1554 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); 1555 if (val_t != NULL) { 1556 if (filtered_t == NULL) { 1557 filtered_t = val_t; 1558 } else { 1559 filtered_t = filtered_t->meet(val_t)->is_int(); 1560 } 1561 } 1562 } 1563 } 1564 } 1565 const TypeInt* n_t = _igvn.type(n)->is_int(); 1566 if (filtered_t != NULL) { 1567 n_t = n_t->join(filtered_t)->is_int(); 1568 } 1569 return n_t; 1570 } 1571 1572 1573 //------------------------------filtered_type_from_dominators-------------------------------- 1574 // Return a possibly more restrictive type for val based on condition control flow of dominators 1575 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { 1576 if (val->is_Con()) { 1577 return val->bottom_type()->is_int(); 1578 } 1579 uint if_limit = 10; // Max number of dominating if's visited 1580 const TypeInt* rtn_t = NULL; 1581 1582 if (use_ctrl && use_ctrl != C->top()) { 1583 Node* val_ctrl = get_ctrl(val); 1584 uint val_dom_depth = dom_depth(val_ctrl); 1585 Node* pred = use_ctrl; 1586 uint if_cnt = 0; 1587 while (if_cnt < if_limit) { 1588 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { 1589 if_cnt++; 1590 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); 1591 if (if_t != NULL) { 1592 if (rtn_t == NULL) { 1593 rtn_t = if_t; 1594 } else { 1595 rtn_t = rtn_t->join(if_t)->is_int(); 1596 } 1597 } 1598 } 1599 pred = idom(pred); 1600 if (pred == NULL || pred == C->top()) { 1601 break; 1602 } 1603 // Stop if going beyond definition block of val 1604 if (dom_depth(pred) < val_dom_depth) { 1605 break; 1606 } 1607 } 1608 } 1609 return rtn_t; 1610 } 1611 1612 1613 //------------------------------dump_spec-------------------------------------- 1614 // Dump special per-node info 1615 #ifndef PRODUCT 1616 void CountedLoopEndNode::dump_spec(outputStream *st) const { 1617 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) { 1618 BoolTest bt( test_trip()); // Added this for g++. 1619 1620 st->print("["); 1621 bt.dump_on(st); 1622 st->print("]"); 1623 } 1624 st->print(" "); 1625 IfNode::dump_spec(st); 1626 } 1627 #endif 1628 1629 //============================================================================= 1630 //------------------------------is_member-------------------------------------- 1631 // Is 'l' a member of 'this'? 1632 bool IdealLoopTree::is_member(const IdealLoopTree *l) const { 1633 while( l->_nest > _nest ) l = l->_parent; 1634 return l == this; 1635 } 1636 1637 //------------------------------set_nest--------------------------------------- 1638 // Set loop tree nesting depth. Accumulate _has_call bits. 1639 int IdealLoopTree::set_nest( uint depth ) { 1640 _nest = depth; 1641 int bits = _has_call; 1642 if( _child ) bits |= _child->set_nest(depth+1); 1643 if( bits ) _has_call = 1; 1644 if( _next ) bits |= _next ->set_nest(depth ); 1645 return bits; 1646 } 1647 1648 //------------------------------split_fall_in---------------------------------- 1649 // Split out multiple fall-in edges from the loop header. Move them to a 1650 // private RegionNode before the loop. This becomes the loop landing pad. 1651 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { 1652 PhaseIterGVN &igvn = phase->_igvn; 1653 uint i; 1654 1655 // Make a new RegionNode to be the landing pad. 1656 Node *landing_pad = new RegionNode( fall_in_cnt+1 ); 1657 phase->set_loop(landing_pad,_parent); 1658 // Gather all the fall-in control paths into the landing pad 1659 uint icnt = fall_in_cnt; 1660 uint oreq = _head->req(); 1661 for( i = oreq-1; i>0; i-- ) 1662 if( !phase->is_member( this, _head->in(i) ) ) 1663 landing_pad->set_req(icnt--,_head->in(i)); 1664 1665 // Peel off PhiNode edges as well 1666 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1667 Node *oj = _head->fast_out(j); 1668 if( oj->is_Phi() ) { 1669 PhiNode* old_phi = oj->as_Phi(); 1670 assert( old_phi->region() == _head, "" ); 1671 igvn.hash_delete(old_phi); // Yank from hash before hacking edges 1672 Node *p = PhiNode::make_blank(landing_pad, old_phi); 1673 uint icnt = fall_in_cnt; 1674 for( i = oreq-1; i>0; i-- ) { 1675 if( !phase->is_member( this, _head->in(i) ) ) { 1676 p->init_req(icnt--, old_phi->in(i)); 1677 // Go ahead and clean out old edges from old phi 1678 old_phi->del_req(i); 1679 } 1680 } 1681 // Search for CSE's here, because ZKM.jar does a lot of 1682 // loop hackery and we need to be a little incremental 1683 // with the CSE to avoid O(N^2) node blow-up. 1684 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE 1685 if( p2 ) { // Found CSE 1686 p->destruct(); // Recover useless new node 1687 p = p2; // Use old node 1688 } else { 1689 igvn.register_new_node_with_optimizer(p, old_phi); 1690 } 1691 // Make old Phi refer to new Phi. 1692 old_phi->add_req(p); 1693 // Check for the special case of making the old phi useless and 1694 // disappear it. In JavaGrande I have a case where this useless 1695 // Phi is the loop limit and prevents recognizing a CountedLoop 1696 // which in turn prevents removing an empty loop. 1697 Node *id_old_phi = old_phi->Identity( &igvn ); 1698 if( id_old_phi != old_phi ) { // Found a simple identity? 1699 // Note that I cannot call 'replace_node' here, because 1700 // that will yank the edge from old_phi to the Region and 1701 // I'm mid-iteration over the Region's uses. 1702 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { 1703 Node* use = old_phi->last_out(i); 1704 igvn.rehash_node_delayed(use); 1705 uint uses_found = 0; 1706 for (uint j = 0; j < use->len(); j++) { 1707 if (use->in(j) == old_phi) { 1708 if (j < use->req()) use->set_req (j, id_old_phi); 1709 else use->set_prec(j, id_old_phi); 1710 uses_found++; 1711 } 1712 } 1713 i -= uses_found; // we deleted 1 or more copies of this edge 1714 } 1715 } 1716 igvn._worklist.push(old_phi); 1717 } 1718 } 1719 // Finally clean out the fall-in edges from the RegionNode 1720 for( i = oreq-1; i>0; i-- ) { 1721 if( !phase->is_member( this, _head->in(i) ) ) { 1722 _head->del_req(i); 1723 } 1724 } 1725 igvn.rehash_node_delayed(_head); 1726 // Transform landing pad 1727 igvn.register_new_node_with_optimizer(landing_pad, _head); 1728 // Insert landing pad into the header 1729 _head->add_req(landing_pad); 1730 } 1731 1732 //------------------------------split_outer_loop------------------------------- 1733 // Split out the outermost loop from this shared header. 1734 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { 1735 PhaseIterGVN &igvn = phase->_igvn; 1736 1737 // Find index of outermost loop; it should also be my tail. 1738 uint outer_idx = 1; 1739 while( _head->in(outer_idx) != _tail ) outer_idx++; 1740 1741 // Make a LoopNode for the outermost loop. 1742 Node *ctl = _head->in(LoopNode::EntryControl); 1743 Node *outer = new LoopNode( ctl, _head->in(outer_idx) ); 1744 outer = igvn.register_new_node_with_optimizer(outer, _head); 1745 phase->set_created_loop_node(); 1746 1747 // Outermost loop falls into '_head' loop 1748 _head->set_req(LoopNode::EntryControl, outer); 1749 _head->del_req(outer_idx); 1750 // Split all the Phis up between '_head' loop and 'outer' loop. 1751 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1752 Node *out = _head->fast_out(j); 1753 if( out->is_Phi() ) { 1754 PhiNode *old_phi = out->as_Phi(); 1755 assert( old_phi->region() == _head, "" ); 1756 Node *phi = PhiNode::make_blank(outer, old_phi); 1757 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); 1758 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); 1759 phi = igvn.register_new_node_with_optimizer(phi, old_phi); 1760 // Make old Phi point to new Phi on the fall-in path 1761 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi); 1762 old_phi->del_req(outer_idx); 1763 } 1764 } 1765 1766 // Use the new loop head instead of the old shared one 1767 _head = outer; 1768 phase->set_loop(_head, this); 1769 } 1770 1771 //------------------------------fix_parent------------------------------------- 1772 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { 1773 loop->_parent = parent; 1774 if( loop->_child ) fix_parent( loop->_child, loop ); 1775 if( loop->_next ) fix_parent( loop->_next , parent ); 1776 } 1777 1778 //------------------------------estimate_path_freq----------------------------- 1779 static float estimate_path_freq( Node *n ) { 1780 // Try to extract some path frequency info 1781 IfNode *iff; 1782 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests 1783 uint nop = n->Opcode(); 1784 if( nop == Op_SafePoint ) { // Skip any safepoint 1785 n = n->in(0); 1786 continue; 1787 } 1788 if( nop == Op_CatchProj ) { // Get count from a prior call 1789 // Assume call does not always throw exceptions: means the call-site 1790 // count is also the frequency of the fall-through path. 1791 assert( n->is_CatchProj(), "" ); 1792 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) 1793 return 0.0f; // Assume call exception path is rare 1794 Node *call = n->in(0)->in(0)->in(0); 1795 assert( call->is_Call(), "expect a call here" ); 1796 const JVMState *jvms = ((CallNode*)call)->jvms(); 1797 ciMethodData* methodData = jvms->method()->method_data(); 1798 if (!methodData->is_mature()) return 0.0f; // No call-site data 1799 ciProfileData* data = methodData->bci_to_data(jvms->bci()); 1800 if ((data == NULL) || !data->is_CounterData()) { 1801 // no call profile available, try call's control input 1802 n = n->in(0); 1803 continue; 1804 } 1805 return data->as_CounterData()->count()/FreqCountInvocations; 1806 } 1807 // See if there's a gating IF test 1808 Node *n_c = n->in(0); 1809 if( !n_c->is_If() ) break; // No estimate available 1810 iff = n_c->as_If(); 1811 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? 1812 // Compute how much count comes on this path 1813 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; 1814 // Have no count info. Skip dull uncommon-trap like branches. 1815 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || 1816 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) 1817 break; 1818 // Skip through never-taken branch; look for a real loop exit. 1819 n = iff->in(0); 1820 } 1821 return 0.0f; // No estimate available 1822 } 1823 1824 //------------------------------merge_many_backedges--------------------------- 1825 // Merge all the backedges from the shared header into a private Region. 1826 // Feed that region as the one backedge to this loop. 1827 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { 1828 uint i; 1829 1830 // Scan for the top 2 hottest backedges 1831 float hotcnt = 0.0f; 1832 float warmcnt = 0.0f; 1833 uint hot_idx = 0; 1834 // Loop starts at 2 because slot 1 is the fall-in path 1835 for( i = 2; i < _head->req(); i++ ) { 1836 float cnt = estimate_path_freq(_head->in(i)); 1837 if( cnt > hotcnt ) { // Grab hottest path 1838 warmcnt = hotcnt; 1839 hotcnt = cnt; 1840 hot_idx = i; 1841 } else if( cnt > warmcnt ) { // And 2nd hottest path 1842 warmcnt = cnt; 1843 } 1844 } 1845 1846 // See if the hottest backedge is worthy of being an inner loop 1847 // by being much hotter than the next hottest backedge. 1848 if( hotcnt <= 0.0001 || 1849 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge 1850 1851 // Peel out the backedges into a private merge point; peel 1852 // them all except optionally hot_idx. 1853 PhaseIterGVN &igvn = phase->_igvn; 1854 1855 Node *hot_tail = NULL; 1856 // Make a Region for the merge point 1857 Node *r = new RegionNode(1); 1858 for( i = 2; i < _head->req(); i++ ) { 1859 if( i != hot_idx ) 1860 r->add_req( _head->in(i) ); 1861 else hot_tail = _head->in(i); 1862 } 1863 igvn.register_new_node_with_optimizer(r, _head); 1864 // Plug region into end of loop _head, followed by hot_tail 1865 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); 1866 igvn.replace_input_of(_head, 2, r); 1867 if( hot_idx ) _head->add_req(hot_tail); 1868 1869 // Split all the Phis up between '_head' loop and the Region 'r' 1870 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1871 Node *out = _head->fast_out(j); 1872 if( out->is_Phi() ) { 1873 PhiNode* n = out->as_Phi(); 1874 igvn.hash_delete(n); // Delete from hash before hacking edges 1875 Node *hot_phi = NULL; 1876 Node *phi = new PhiNode(r, n->type(), n->adr_type()); 1877 // Check all inputs for the ones to peel out 1878 uint j = 1; 1879 for( uint i = 2; i < n->req(); i++ ) { 1880 if( i != hot_idx ) 1881 phi->set_req( j++, n->in(i) ); 1882 else hot_phi = n->in(i); 1883 } 1884 // Register the phi but do not transform until whole place transforms 1885 igvn.register_new_node_with_optimizer(phi, n); 1886 // Add the merge phi to the old Phi 1887 while( n->req() > 3 ) n->del_req( n->req()-1 ); 1888 igvn.replace_input_of(n, 2, phi); 1889 if( hot_idx ) n->add_req(hot_phi); 1890 } 1891 } 1892 1893 1894 // Insert a new IdealLoopTree inserted below me. Turn it into a clone 1895 // of self loop tree. Turn self into a loop headed by _head and with 1896 // tail being the new merge point. 1897 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); 1898 phase->set_loop(_tail,ilt); // Adjust tail 1899 _tail = r; // Self's tail is new merge point 1900 phase->set_loop(r,this); 1901 ilt->_child = _child; // New guy has my children 1902 _child = ilt; // Self has new guy as only child 1903 ilt->_parent = this; // new guy has self for parent 1904 ilt->_nest = _nest; // Same nesting depth (for now) 1905 1906 // Starting with 'ilt', look for child loop trees using the same shared 1907 // header. Flatten these out; they will no longer be loops in the end. 1908 IdealLoopTree **pilt = &_child; 1909 while( ilt ) { 1910 if( ilt->_head == _head ) { 1911 uint i; 1912 for( i = 2; i < _head->req(); i++ ) 1913 if( _head->in(i) == ilt->_tail ) 1914 break; // Still a loop 1915 if( i == _head->req() ) { // No longer a loop 1916 // Flatten ilt. Hang ilt's "_next" list from the end of 1917 // ilt's '_child' list. Move the ilt's _child up to replace ilt. 1918 IdealLoopTree **cp = &ilt->_child; 1919 while( *cp ) cp = &(*cp)->_next; // Find end of child list 1920 *cp = ilt->_next; // Hang next list at end of child list 1921 *pilt = ilt->_child; // Move child up to replace ilt 1922 ilt->_head = NULL; // Flag as a loop UNIONED into parent 1923 ilt = ilt->_child; // Repeat using new ilt 1924 continue; // do not advance over ilt->_child 1925 } 1926 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); 1927 phase->set_loop(_head,ilt); 1928 } 1929 pilt = &ilt->_child; // Advance to next 1930 ilt = *pilt; 1931 } 1932 1933 if( _child ) fix_parent( _child, this ); 1934 } 1935 1936 //------------------------------beautify_loops--------------------------------- 1937 // Split shared headers and insert loop landing pads. 1938 // Insert a LoopNode to replace the RegionNode. 1939 // Return TRUE if loop tree is structurally changed. 1940 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { 1941 bool result = false; 1942 // Cache parts in locals for easy 1943 PhaseIterGVN &igvn = phase->_igvn; 1944 1945 igvn.hash_delete(_head); // Yank from hash before hacking edges 1946 1947 // Check for multiple fall-in paths. Peel off a landing pad if need be. 1948 int fall_in_cnt = 0; 1949 for( uint i = 1; i < _head->req(); i++ ) 1950 if( !phase->is_member( this, _head->in(i) ) ) 1951 fall_in_cnt++; 1952 assert( fall_in_cnt, "at least 1 fall-in path" ); 1953 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins 1954 split_fall_in( phase, fall_in_cnt ); 1955 1956 // Swap inputs to the _head and all Phis to move the fall-in edge to 1957 // the left. 1958 fall_in_cnt = 1; 1959 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) 1960 fall_in_cnt++; 1961 if( fall_in_cnt > 1 ) { 1962 // Since I am just swapping inputs I do not need to update def-use info 1963 Node *tmp = _head->in(1); 1964 igvn.rehash_node_delayed(_head); 1965 _head->set_req( 1, _head->in(fall_in_cnt) ); 1966 _head->set_req( fall_in_cnt, tmp ); 1967 // Swap also all Phis 1968 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { 1969 Node* phi = _head->fast_out(i); 1970 if( phi->is_Phi() ) { 1971 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges 1972 tmp = phi->in(1); 1973 phi->set_req( 1, phi->in(fall_in_cnt) ); 1974 phi->set_req( fall_in_cnt, tmp ); 1975 } 1976 } 1977 } 1978 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); 1979 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); 1980 1981 // If I am a shared header (multiple backedges), peel off the many 1982 // backedges into a private merge point and use the merge point as 1983 // the one true backedge. 1984 if( _head->req() > 3 ) { 1985 // Merge the many backedges into a single backedge but leave 1986 // the hottest backedge as separate edge for the following peel. 1987 merge_many_backedges( phase ); 1988 result = true; 1989 } 1990 1991 // If I have one hot backedge, peel off myself loop. 1992 // I better be the outermost loop. 1993 if (_head->req() > 3 && !_irreducible) { 1994 split_outer_loop( phase ); 1995 result = true; 1996 1997 } else if (!_head->is_Loop() && !_irreducible) { 1998 // Make a new LoopNode to replace the old loop head 1999 Node *l = new LoopNode( _head->in(1), _head->in(2) ); 2000 l = igvn.register_new_node_with_optimizer(l, _head); 2001 phase->set_created_loop_node(); 2002 // Go ahead and replace _head 2003 phase->_igvn.replace_node( _head, l ); 2004 _head = l; 2005 phase->set_loop(_head, this); 2006 } 2007 2008 // Now recursively beautify nested loops 2009 if( _child ) result |= _child->beautify_loops( phase ); 2010 if( _next ) result |= _next ->beautify_loops( phase ); 2011 return result; 2012 } 2013 2014 //------------------------------allpaths_check_safepts---------------------------- 2015 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 2016 // encountered. Helper for check_safepts. 2017 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { 2018 assert(stack.size() == 0, "empty stack"); 2019 stack.push(_tail); 2020 visited.Clear(); 2021 visited.set(_tail->_idx); 2022 while (stack.size() > 0) { 2023 Node* n = stack.pop(); 2024 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2025 // Terminate this path 2026 } else if (n->Opcode() == Op_SafePoint) { 2027 if (_phase->get_loop(n) != this) { 2028 if (_required_safept == NULL) _required_safept = new Node_List(); 2029 _required_safept->push(n); // save the one closest to the tail 2030 } 2031 // Terminate this path 2032 } else { 2033 uint start = n->is_Region() ? 1 : 0; 2034 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; 2035 for (uint i = start; i < end; i++) { 2036 Node* in = n->in(i); 2037 assert(in->is_CFG(), "must be"); 2038 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { 2039 stack.push(in); 2040 } 2041 } 2042 } 2043 } 2044 } 2045 2046 //------------------------------check_safepts---------------------------- 2047 // Given dominators, try to find loops with calls that must always be 2048 // executed (call dominates loop tail). These loops do not need non-call 2049 // safepoints (ncsfpt). 2050 // 2051 // A complication is that a safepoint in a inner loop may be needed 2052 // by an outer loop. In the following, the inner loop sees it has a 2053 // call (block 3) on every path from the head (block 2) to the 2054 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) 2055 // in block 2, _but_ this leaves the outer loop without a safepoint. 2056 // 2057 // entry 0 2058 // | 2059 // v 2060 // outer 1,2 +->1 2061 // | | 2062 // | v 2063 // | 2<---+ ncsfpt in 2 2064 // |_/|\ | 2065 // | v | 2066 // inner 2,3 / 3 | call in 3 2067 // / | | 2068 // v +--+ 2069 // exit 4 2070 // 2071 // 2072 // This method creates a list (_required_safept) of ncsfpt nodes that must 2073 // be protected is created for each loop. When a ncsfpt maybe deleted, it 2074 // is first looked for in the lists for the outer loops of the current loop. 2075 // 2076 // The insights into the problem: 2077 // A) counted loops are okay 2078 // B) innermost loops are okay (only an inner loop can delete 2079 // a ncsfpt needed by an outer loop) 2080 // C) a loop is immune from an inner loop deleting a safepoint 2081 // if the loop has a call on the idom-path 2082 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the 2083 // idom-path that is not in a nested loop 2084 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner 2085 // loop needs to be prevented from deletion by an inner loop 2086 // 2087 // There are two analyses: 2088 // 1) The first, and cheaper one, scans the loop body from 2089 // tail to head following the idom (immediate dominator) 2090 // chain, looking for the cases (C,D,E) above. 2091 // Since inner loops are scanned before outer loops, there is summary 2092 // information about inner loops. Inner loops can be skipped over 2093 // when the tail of an inner loop is encountered. 2094 // 2095 // 2) The second, invoked if the first fails to find a call or ncsfpt on 2096 // the idom path (which is rare), scans all predecessor control paths 2097 // from the tail to the head, terminating a path when a call or sfpt 2098 // is encountered, to find the ncsfpt's that are closest to the tail. 2099 // 2100 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { 2101 // Bottom up traversal 2102 IdealLoopTree* ch = _child; 2103 if (_child) _child->check_safepts(visited, stack); 2104 if (_next) _next ->check_safepts(visited, stack); 2105 2106 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { 2107 bool has_call = false; // call on dom-path 2108 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth 2109 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth 2110 // Scan the dom-path nodes from tail to head 2111 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { 2112 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2113 has_call = true; 2114 _has_sfpt = 1; // Then no need for a safept! 2115 break; 2116 } else if (n->Opcode() == Op_SafePoint) { 2117 if (_phase->get_loop(n) == this) { 2118 has_local_ncsfpt = true; 2119 break; 2120 } 2121 if (nonlocal_ncsfpt == NULL) { 2122 nonlocal_ncsfpt = n; // save the one closest to the tail 2123 } 2124 } else { 2125 IdealLoopTree* nlpt = _phase->get_loop(n); 2126 if (this != nlpt) { 2127 // If at an inner loop tail, see if the inner loop has already 2128 // recorded seeing a call on the dom-path (and stop.) If not, 2129 // jump to the head of the inner loop. 2130 assert(is_member(nlpt), "nested loop"); 2131 Node* tail = nlpt->_tail; 2132 if (tail->in(0)->is_If()) tail = tail->in(0); 2133 if (n == tail) { 2134 // If inner loop has call on dom-path, so does outer loop 2135 if (nlpt->_has_sfpt) { 2136 has_call = true; 2137 _has_sfpt = 1; 2138 break; 2139 } 2140 // Skip to head of inner loop 2141 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); 2142 n = nlpt->_head; 2143 } 2144 } 2145 } 2146 } 2147 // Record safept's that this loop needs preserved when an 2148 // inner loop attempts to delete it's safepoints. 2149 if (_child != NULL && !has_call && !has_local_ncsfpt) { 2150 if (nonlocal_ncsfpt != NULL) { 2151 if (_required_safept == NULL) _required_safept = new Node_List(); 2152 _required_safept->push(nonlocal_ncsfpt); 2153 } else { 2154 // Failed to find a suitable safept on the dom-path. Now use 2155 // an all paths walk from tail to head, looking for safepoints to preserve. 2156 allpaths_check_safepts(visited, stack); 2157 } 2158 } 2159 } 2160 } 2161 2162 //---------------------------is_deleteable_safept---------------------------- 2163 // Is safept not required by an outer loop? 2164 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { 2165 assert(sfpt->Opcode() == Op_SafePoint, ""); 2166 IdealLoopTree* lp = get_loop(sfpt)->_parent; 2167 while (lp != NULL) { 2168 Node_List* sfpts = lp->_required_safept; 2169 if (sfpts != NULL) { 2170 for (uint i = 0; i < sfpts->size(); i++) { 2171 if (sfpt == sfpts->at(i)) 2172 return false; 2173 } 2174 } 2175 lp = lp->_parent; 2176 } 2177 return true; 2178 } 2179 2180 //---------------------------replace_parallel_iv------------------------------- 2181 // Replace parallel induction variable (parallel to trip counter) 2182 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) { 2183 assert(loop->_head->is_CountedLoop(), ""); 2184 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2185 if (!cl->is_valid_counted_loop()) 2186 return; // skip malformed counted loop 2187 Node *incr = cl->incr(); 2188 if (incr == NULL) 2189 return; // Dead loop? 2190 Node *init = cl->init_trip(); 2191 Node *phi = cl->phi(); 2192 int stride_con = cl->stride_con(); 2193 2194 // Visit all children, looking for Phis 2195 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { 2196 Node *out = cl->out(i); 2197 // Look for other phis (secondary IVs). Skip dead ones 2198 if (!out->is_Phi() || out == phi || !has_node(out)) 2199 continue; 2200 PhiNode* phi2 = out->as_Phi(); 2201 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); 2202 // Look for induction variables of the form: X += constant 2203 if (phi2->region() != loop->_head || 2204 incr2->req() != 3 || 2205 incr2->in(1) != phi2 || 2206 incr2 == incr || 2207 incr2->Opcode() != Op_AddI || 2208 !incr2->in(2)->is_Con()) 2209 continue; 2210 2211 // Check for parallel induction variable (parallel to trip counter) 2212 // via an affine function. In particular, count-down loops with 2213 // count-up array indices are common. We only RCE references off 2214 // the trip-counter, so we need to convert all these to trip-counter 2215 // expressions. 2216 Node *init2 = phi2->in( LoopNode::EntryControl ); 2217 int stride_con2 = incr2->in(2)->get_int(); 2218 2219 // The ratio of the two strides cannot be represented as an int 2220 // if stride_con2 is min_int and stride_con is -1. 2221 if (stride_con2 == min_jint && stride_con == -1) { 2222 continue; 2223 } 2224 2225 // The general case here gets a little tricky. We want to find the 2226 // GCD of all possible parallel IV's and make a new IV using this 2227 // GCD for the loop. Then all possible IVs are simple multiples of 2228 // the GCD. In practice, this will cover very few extra loops. 2229 // Instead we require 'stride_con2' to be a multiple of 'stride_con', 2230 // where +/-1 is the common case, but other integer multiples are 2231 // also easy to handle. 2232 int ratio_con = stride_con2/stride_con; 2233 2234 if ((ratio_con * stride_con) == stride_con2) { // Check for exact 2235 #ifndef PRODUCT 2236 if (TraceLoopOpts) { 2237 tty->print("Parallel IV: %d ", phi2->_idx); 2238 loop->dump_head(); 2239 } 2240 #endif 2241 // Convert to using the trip counter. The parallel induction 2242 // variable differs from the trip counter by a loop-invariant 2243 // amount, the difference between their respective initial values. 2244 // It is scaled by the 'ratio_con'. 2245 Node* ratio = _igvn.intcon(ratio_con); 2246 set_ctrl(ratio, C->root()); 2247 Node* ratio_init = new MulINode(init, ratio); 2248 _igvn.register_new_node_with_optimizer(ratio_init, init); 2249 set_early_ctrl(ratio_init); 2250 Node* diff = new SubINode(init2, ratio_init); 2251 _igvn.register_new_node_with_optimizer(diff, init2); 2252 set_early_ctrl(diff); 2253 Node* ratio_idx = new MulINode(phi, ratio); 2254 _igvn.register_new_node_with_optimizer(ratio_idx, phi); 2255 set_ctrl(ratio_idx, cl); 2256 Node* add = new AddINode(ratio_idx, diff); 2257 _igvn.register_new_node_with_optimizer(add); 2258 set_ctrl(add, cl); 2259 _igvn.replace_node( phi2, add ); 2260 // Sometimes an induction variable is unused 2261 if (add->outcnt() == 0) { 2262 _igvn.remove_dead_node(add); 2263 } 2264 --i; // deleted this phi; rescan starting with next position 2265 continue; 2266 } 2267 } 2268 } 2269 2270 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) { 2271 Node* keep = NULL; 2272 if (keep_one) { 2273 // Look for a safepoint on the idom-path. 2274 for (Node* i = tail(); i != _head; i = phase->idom(i)) { 2275 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) { 2276 keep = i; 2277 break; // Found one 2278 } 2279 } 2280 } 2281 2282 // Don't remove any safepoints if it is requested to keep a single safepoint and 2283 // no safepoint was found on idom-path. It is not safe to remove any safepoint 2284 // in this case since there's no safepoint dominating all paths in the loop body. 2285 bool prune = !keep_one || keep != NULL; 2286 2287 // Delete other safepoints in this loop. 2288 Node_List* sfpts = _safepts; 2289 if (prune && sfpts != NULL) { 2290 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint"); 2291 for (uint i = 0; i < sfpts->size(); i++) { 2292 Node* n = sfpts->at(i); 2293 assert(phase->get_loop(n) == this, ""); 2294 if (n != keep && phase->is_deleteable_safept(n)) { 2295 phase->lazy_replace(n, n->in(TypeFunc::Control)); 2296 } 2297 } 2298 } 2299 } 2300 2301 //------------------------------counted_loop----------------------------------- 2302 // Convert to counted loops where possible 2303 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { 2304 2305 // For grins, set the inner-loop flag here 2306 if (!_child) { 2307 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop(); 2308 } 2309 2310 IdealLoopTree* loop = this; 2311 if (_head->is_CountedLoop() || 2312 phase->is_counted_loop(_head, loop)) { 2313 2314 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) { 2315 // Indicate we do not need a safepoint here 2316 _has_sfpt = 1; 2317 } 2318 2319 // Remove safepoints 2320 bool keep_one_sfpt = !(_has_call || _has_sfpt); 2321 remove_safepoints(phase, keep_one_sfpt); 2322 2323 // Look for induction variables 2324 phase->replace_parallel_iv(this); 2325 2326 } else if (_parent != NULL && !_irreducible) { 2327 // Not a counted loop. Keep one safepoint. 2328 bool keep_one_sfpt = true; 2329 remove_safepoints(phase, keep_one_sfpt); 2330 } 2331 2332 // Recursively 2333 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?"); 2334 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops"); 2335 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase); 2336 if (loop->_next) loop->_next ->counted_loop(phase); 2337 } 2338 2339 #ifndef PRODUCT 2340 //------------------------------dump_head-------------------------------------- 2341 // Dump 1 liner for loop header info 2342 void IdealLoopTree::dump_head( ) const { 2343 for (uint i=0; i<_nest; i++) 2344 tty->print(" "); 2345 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); 2346 if (_irreducible) tty->print(" IRREDUCIBLE"); 2347 Node* entry = _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl); 2348 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 2349 if (predicate != NULL ) { 2350 tty->print(" limit_check"); 2351 entry = entry->in(0)->in(0); 2352 } 2353 if (UseLoopPredicate) { 2354 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 2355 if (entry != NULL) { 2356 tty->print(" predicated"); 2357 } 2358 } 2359 if (_head->is_CountedLoop()) { 2360 CountedLoopNode *cl = _head->as_CountedLoop(); 2361 tty->print(" counted"); 2362 2363 Node* init_n = cl->init_trip(); 2364 if (init_n != NULL && init_n->is_Con()) 2365 tty->print(" [%d,", cl->init_trip()->get_int()); 2366 else 2367 tty->print(" [int,"); 2368 Node* limit_n = cl->limit(); 2369 if (limit_n != NULL && limit_n->is_Con()) 2370 tty->print("%d),", cl->limit()->get_int()); 2371 else 2372 tty->print("int),"); 2373 int stride_con = cl->stride_con(); 2374 if (stride_con > 0) tty->print("+"); 2375 tty->print("%d", stride_con); 2376 2377 tty->print(" (%0.f iters) ", cl->profile_trip_cnt()); 2378 2379 if (cl->is_pre_loop ()) tty->print(" pre" ); 2380 if (cl->is_main_loop()) tty->print(" main"); 2381 if (cl->is_post_loop()) tty->print(" post"); 2382 if (cl->is_vectorized_loop()) tty->print(" vector"); 2383 if (cl->range_checks_present()) tty->print(" rc "); 2384 if (cl->is_multiversioned()) tty->print(" multi "); 2385 } 2386 if (_has_call) tty->print(" has_call"); 2387 if (_has_sfpt) tty->print(" has_sfpt"); 2388 if (_rce_candidate) tty->print(" rce"); 2389 if (_safepts != NULL && _safepts->size() > 0) { 2390 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }"); 2391 } 2392 if (_required_safept != NULL && _required_safept->size() > 0) { 2393 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }"); 2394 } 2395 if (Verbose) { 2396 tty->print(" body={"); _body.dump_simple(); tty->print(" }"); 2397 } 2398 if (_head->as_Loop()->is_strip_mined()) { 2399 tty->print(" strip_mined"); 2400 } 2401 tty->cr(); 2402 } 2403 2404 //------------------------------dump------------------------------------------- 2405 // Dump loops by loop tree 2406 void IdealLoopTree::dump( ) const { 2407 dump_head(); 2408 if (_child) _child->dump(); 2409 if (_next) _next ->dump(); 2410 } 2411 2412 #endif 2413 2414 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { 2415 if (loop == root) { 2416 if (loop->_child != NULL) { 2417 log->begin_head("loop_tree"); 2418 log->end_head(); 2419 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2420 log->tail("loop_tree"); 2421 assert(loop->_next == NULL, "what?"); 2422 } 2423 } else { 2424 Node* head = loop->_head; 2425 log->begin_head("loop"); 2426 log->print(" idx='%d' ", head->_idx); 2427 if (loop->_irreducible) log->print("irreducible='1' "); 2428 if (head->is_Loop()) { 2429 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); 2430 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); 2431 } 2432 if (head->is_CountedLoop()) { 2433 CountedLoopNode* cl = head->as_CountedLoop(); 2434 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); 2435 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); 2436 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); 2437 } 2438 log->end_head(); 2439 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2440 log->tail("loop"); 2441 if( loop->_next ) log_loop_tree(root, loop->_next, log); 2442 } 2443 } 2444 2445 //---------------------collect_potentially_useful_predicates----------------------- 2446 // Helper function to collect potentially useful predicates to prevent them from 2447 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates 2448 void PhaseIdealLoop::collect_potentially_useful_predicates( 2449 IdealLoopTree * loop, Unique_Node_List &useful_predicates) { 2450 if (loop->_child) { // child 2451 collect_potentially_useful_predicates(loop->_child, useful_predicates); 2452 } 2453 2454 // self (only loops that we can apply loop predication may use their predicates) 2455 if (loop->_head->is_Loop() && 2456 !loop->_irreducible && 2457 !loop->tail()->is_top()) { 2458 LoopNode* lpn = loop->_head->as_Loop(); 2459 Node* entry = lpn->in(LoopNode::EntryControl); 2460 Node* predicate_proj = find_predicate(entry); // loop_limit_check first 2461 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication 2462 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be"); 2463 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2464 entry = entry->in(0)->in(0); 2465 } 2466 predicate_proj = find_predicate(entry); // Predicate 2467 if (predicate_proj != NULL ) { 2468 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2469 } 2470 } 2471 2472 if (loop->_next) { // sibling 2473 collect_potentially_useful_predicates(loop->_next, useful_predicates); 2474 } 2475 } 2476 2477 //------------------------eliminate_useless_predicates----------------------------- 2478 // Eliminate all inserted predicates if they could not be used by loop predication. 2479 // Note: it will also eliminates loop limits check predicate since it also uses 2480 // Opaque1 node (see Parse::add_predicate()). 2481 void PhaseIdealLoop::eliminate_useless_predicates() { 2482 if (C->predicate_count() == 0) 2483 return; // no predicate left 2484 2485 Unique_Node_List useful_predicates; // to store useful predicates 2486 if (C->has_loops()) { 2487 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); 2488 } 2489 2490 for (int i = C->predicate_count(); i > 0; i--) { 2491 Node * n = C->predicate_opaque1_node(i-1); 2492 assert(n->Opcode() == Op_Opaque1, "must be"); 2493 if (!useful_predicates.member(n)) { // not in the useful list 2494 _igvn.replace_node(n, n->in(1)); 2495 } 2496 } 2497 } 2498 2499 //------------------------process_expensive_nodes----------------------------- 2500 // Expensive nodes have their control input set to prevent the GVN 2501 // from commoning them and as a result forcing the resulting node to 2502 // be in a more frequent path. Use CFG information here, to change the 2503 // control inputs so that some expensive nodes can be commoned while 2504 // not executed more frequently. 2505 bool PhaseIdealLoop::process_expensive_nodes() { 2506 assert(OptimizeExpensiveOps, "optimization off?"); 2507 2508 // Sort nodes to bring similar nodes together 2509 C->sort_expensive_nodes(); 2510 2511 bool progress = false; 2512 2513 for (int i = 0; i < C->expensive_count(); ) { 2514 Node* n = C->expensive_node(i); 2515 int start = i; 2516 // Find nodes similar to n 2517 i++; 2518 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++); 2519 int end = i; 2520 // And compare them two by two 2521 for (int j = start; j < end; j++) { 2522 Node* n1 = C->expensive_node(j); 2523 if (is_node_unreachable(n1)) { 2524 continue; 2525 } 2526 for (int k = j+1; k < end; k++) { 2527 Node* n2 = C->expensive_node(k); 2528 if (is_node_unreachable(n2)) { 2529 continue; 2530 } 2531 2532 assert(n1 != n2, "should be pair of nodes"); 2533 2534 Node* c1 = n1->in(0); 2535 Node* c2 = n2->in(0); 2536 2537 Node* parent_c1 = c1; 2538 Node* parent_c2 = c2; 2539 2540 // The call to get_early_ctrl_for_expensive() moves the 2541 // expensive nodes up but stops at loops that are in a if 2542 // branch. See whether we can exit the loop and move above the 2543 // If. 2544 if (c1->is_Loop()) { 2545 parent_c1 = c1->in(1); 2546 } 2547 if (c2->is_Loop()) { 2548 parent_c2 = c2->in(1); 2549 } 2550 2551 if (parent_c1 == parent_c2) { 2552 _igvn._worklist.push(n1); 2553 _igvn._worklist.push(n2); 2554 continue; 2555 } 2556 2557 // Look for identical expensive node up the dominator chain. 2558 if (is_dominator(c1, c2)) { 2559 c2 = c1; 2560 } else if (is_dominator(c2, c1)) { 2561 c1 = c2; 2562 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() && 2563 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) { 2564 // Both branches have the same expensive node so move it up 2565 // before the if. 2566 c1 = c2 = idom(parent_c1->in(0)); 2567 } 2568 // Do the actual moves 2569 if (n1->in(0) != c1) { 2570 _igvn.hash_delete(n1); 2571 n1->set_req(0, c1); 2572 _igvn.hash_insert(n1); 2573 _igvn._worklist.push(n1); 2574 progress = true; 2575 } 2576 if (n2->in(0) != c2) { 2577 _igvn.hash_delete(n2); 2578 n2->set_req(0, c2); 2579 _igvn.hash_insert(n2); 2580 _igvn._worklist.push(n2); 2581 progress = true; 2582 } 2583 } 2584 } 2585 } 2586 2587 return progress; 2588 } 2589 2590 2591 //============================================================================= 2592 //----------------------------build_and_optimize------------------------------- 2593 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to 2594 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. 2595 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) { 2596 ResourceMark rm; 2597 2598 int old_progress = C->major_progress(); 2599 uint orig_worklist_size = _igvn._worklist.size(); 2600 2601 // Reset major-progress flag for the driver's heuristics 2602 C->clear_major_progress(); 2603 2604 #ifndef PRODUCT 2605 // Capture for later assert 2606 uint unique = C->unique(); 2607 _loop_invokes++; 2608 _loop_work += unique; 2609 #endif 2610 2611 // True if the method has at least 1 irreducible loop 2612 _has_irreducible_loops = false; 2613 2614 _created_loop_node = false; 2615 2616 Arena *a = Thread::current()->resource_area(); 2617 VectorSet visited(a); 2618 // Pre-grow the mapping from Nodes to IdealLoopTrees. 2619 _nodes.map(C->unique(), NULL); 2620 memset(_nodes.adr(), 0, wordSize * C->unique()); 2621 2622 // Pre-build the top-level outermost loop tree entry 2623 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); 2624 // Do not need a safepoint at the top level 2625 _ltree_root->_has_sfpt = 1; 2626 2627 // Initialize Dominators. 2628 // Checked in clone_loop_predicate() during beautify_loops(). 2629 _idom_size = 0; 2630 _idom = NULL; 2631 _dom_depth = NULL; 2632 _dom_stk = NULL; 2633 2634 // Empty pre-order array 2635 allocate_preorders(); 2636 2637 // Build a loop tree on the fly. Build a mapping from CFG nodes to 2638 // IdealLoopTree entries. Data nodes are NOT walked. 2639 build_loop_tree(); 2640 // Check for bailout, and return 2641 if (C->failing()) { 2642 return; 2643 } 2644 2645 // No loops after all 2646 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); 2647 2648 // There should always be an outer loop containing the Root and Return nodes. 2649 // If not, we have a degenerate empty program. Bail out in this case. 2650 if (!has_node(C->root())) { 2651 if (!_verify_only) { 2652 C->clear_major_progress(); 2653 C->record_method_not_compilable("empty program detected during loop optimization"); 2654 } 2655 return; 2656 } 2657 2658 // Nothing to do, so get out 2659 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only; 2660 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn); 2661 if (stop_early && !do_expensive_nodes) { 2662 _igvn.optimize(); // Cleanup NeverBranches 2663 return; 2664 } 2665 2666 // Set loop nesting depth 2667 _ltree_root->set_nest( 0 ); 2668 2669 // Split shared headers and insert loop landing pads. 2670 // Do not bother doing this on the Root loop of course. 2671 if( !_verify_me && !_verify_only && _ltree_root->_child ) { 2672 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3); 2673 if( _ltree_root->_child->beautify_loops( this ) ) { 2674 // Re-build loop tree! 2675 _ltree_root->_child = NULL; 2676 _nodes.clear(); 2677 reallocate_preorders(); 2678 build_loop_tree(); 2679 // Check for bailout, and return 2680 if (C->failing()) { 2681 return; 2682 } 2683 // Reset loop nesting depth 2684 _ltree_root->set_nest( 0 ); 2685 2686 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3); 2687 } 2688 } 2689 2690 // Build Dominators for elision of NULL checks & loop finding. 2691 // Since nodes do not have a slot for immediate dominator, make 2692 // a persistent side array for that info indexed on node->_idx. 2693 _idom_size = C->unique(); 2694 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); 2695 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); 2696 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth 2697 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); 2698 2699 Dominators(); 2700 2701 if (!_verify_only) { 2702 // As a side effect, Dominators removed any unreachable CFG paths 2703 // into RegionNodes. It doesn't do this test against Root, so 2704 // we do it here. 2705 for( uint i = 1; i < C->root()->req(); i++ ) { 2706 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? 2707 _igvn.delete_input_of(C->root(), i); 2708 i--; // Rerun same iteration on compressed edges 2709 } 2710 } 2711 2712 // Given dominators, try to find inner loops with calls that must 2713 // always be executed (call dominates loop tail). These loops do 2714 // not need a separate safepoint. 2715 Node_List cisstack(a); 2716 _ltree_root->check_safepts(visited, cisstack); 2717 } 2718 2719 // Walk the DATA nodes and place into loops. Find earliest control 2720 // node. For CFG nodes, the _nodes array starts out and remains 2721 // holding the associated IdealLoopTree pointer. For DATA nodes, the 2722 // _nodes array holds the earliest legal controlling CFG node. 2723 2724 // Allocate stack with enough space to avoid frequent realloc 2725 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats 2726 Node_Stack nstack( a, stack_size ); 2727 2728 visited.Clear(); 2729 Node_List worklist(a); 2730 // Don't need C->root() on worklist since 2731 // it will be processed among C->top() inputs 2732 worklist.push( C->top() ); 2733 visited.set( C->top()->_idx ); // Set C->top() as visited now 2734 build_loop_early( visited, worklist, nstack ); 2735 2736 // Given early legal placement, try finding counted loops. This placement 2737 // is good enough to discover most loop invariants. 2738 if( !_verify_me && !_verify_only ) 2739 _ltree_root->counted_loop( this ); 2740 2741 // Find latest loop placement. Find ideal loop placement. 2742 visited.Clear(); 2743 init_dom_lca_tags(); 2744 // Need C->root() on worklist when processing outs 2745 worklist.push( C->root() ); 2746 NOT_PRODUCT( C->verify_graph_edges(); ) 2747 worklist.push( C->top() ); 2748 build_loop_late( visited, worklist, nstack ); 2749 2750 if (_verify_only) { 2751 // restore major progress flag 2752 for (int i = 0; i < old_progress; i++) 2753 C->set_major_progress(); 2754 assert(C->unique() == unique, "verification mode made Nodes? ? ?"); 2755 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything"); 2756 return; 2757 } 2758 2759 // clear out the dead code after build_loop_late 2760 while (_deadlist.size()) { 2761 _igvn.remove_globally_dead_node(_deadlist.pop()); 2762 } 2763 2764 if (stop_early) { 2765 assert(do_expensive_nodes, "why are we here?"); 2766 if (process_expensive_nodes()) { 2767 // If we made some progress when processing expensive nodes then 2768 // the IGVN may modify the graph in a way that will allow us to 2769 // make some more progress: we need to try processing expensive 2770 // nodes again. 2771 C->set_major_progress(); 2772 } 2773 _igvn.optimize(); 2774 return; 2775 } 2776 2777 // Some parser-inserted loop predicates could never be used by loop 2778 // predication or they were moved away from loop during some optimizations. 2779 // For example, peeling. Eliminate them before next loop optimizations. 2780 eliminate_useless_predicates(); 2781 2782 #ifndef PRODUCT 2783 C->verify_graph_edges(); 2784 if (_verify_me) { // Nested verify pass? 2785 // Check to see if the verify mode is broken 2786 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); 2787 return; 2788 } 2789 if(VerifyLoopOptimizations) verify(); 2790 if(TraceLoopOpts && C->has_loops()) { 2791 _ltree_root->dump(); 2792 } 2793 #endif 2794 2795 if (skip_loop_opts) { 2796 // restore major progress flag 2797 for (int i = 0; i < old_progress; i++) { 2798 C->set_major_progress(); 2799 } 2800 2801 // Cleanup any modified bits 2802 _igvn.optimize(); 2803 2804 if (C->log() != NULL) { 2805 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2806 } 2807 return; 2808 } 2809 2810 if (ReassociateInvariants) { 2811 // Reassociate invariants and prep for split_thru_phi 2812 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2813 IdealLoopTree* lpt = iter.current(); 2814 bool is_counted = lpt->is_counted(); 2815 if (!is_counted || !lpt->is_inner()) continue; 2816 2817 // check for vectorized loops, any reassociation of invariants was already done 2818 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue; 2819 2820 lpt->reassociate_invariants(this); 2821 2822 // Because RCE opportunities can be masked by split_thru_phi, 2823 // look for RCE candidates and inhibit split_thru_phi 2824 // on just their loop-phi's for this pass of loop opts 2825 if (SplitIfBlocks && do_split_ifs) { 2826 if (lpt->policy_range_check(this)) { 2827 lpt->_rce_candidate = 1; // = true 2828 } 2829 } 2830 } 2831 } 2832 2833 // Check for aggressive application of split-if and other transforms 2834 // that require basic-block info (like cloning through Phi's) 2835 if( SplitIfBlocks && do_split_ifs ) { 2836 visited.Clear(); 2837 split_if_with_blocks( visited, nstack ); 2838 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); 2839 } 2840 2841 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) { 2842 C->set_major_progress(); 2843 } 2844 2845 // Perform loop predication before iteration splitting 2846 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) { 2847 _ltree_root->_child->loop_predication(this); 2848 } 2849 2850 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { 2851 if (do_intrinsify_fill()) { 2852 C->set_major_progress(); 2853 } 2854 } 2855 2856 // Perform iteration-splitting on inner loops. Split iterations to avoid 2857 // range checks or one-shot null checks. 2858 2859 // If split-if's didn't hack the graph too bad (no CFG changes) 2860 // then do loop opts. 2861 if (C->has_loops() && !C->major_progress()) { 2862 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 2863 _ltree_root->_child->iteration_split( this, worklist ); 2864 // No verify after peeling! GCM has hoisted code out of the loop. 2865 // After peeling, the hoisted code could sink inside the peeled area. 2866 // The peeling code does not try to recompute the best location for 2867 // all the code before the peeled area, so the verify pass will always 2868 // complain about it. 2869 } 2870 // Do verify graph edges in any case 2871 NOT_PRODUCT( C->verify_graph_edges(); ); 2872 2873 if (!do_split_ifs) { 2874 // We saw major progress in Split-If to get here. We forced a 2875 // pass with unrolling and not split-if, however more split-if's 2876 // might make progress. If the unrolling didn't make progress 2877 // then the major-progress flag got cleared and we won't try 2878 // another round of Split-If. In particular the ever-common 2879 // instance-of/check-cast pattern requires at least 2 rounds of 2880 // Split-If to clear out. 2881 C->set_major_progress(); 2882 } 2883 2884 // Repeat loop optimizations if new loops were seen 2885 if (created_loop_node()) { 2886 C->set_major_progress(); 2887 } 2888 2889 // Keep loop predicates and perform optimizations with them 2890 // until no more loop optimizations could be done. 2891 // After that switch predicates off and do more loop optimizations. 2892 if (!C->major_progress() && (C->predicate_count() > 0)) { 2893 C->cleanup_loop_predicates(_igvn); 2894 if (TraceLoopOpts) { 2895 tty->print_cr("PredicatesOff"); 2896 } 2897 C->set_major_progress(); 2898 } 2899 2900 // Convert scalar to superword operations at the end of all loop opts. 2901 if (UseSuperWord && C->has_loops() && !C->major_progress()) { 2902 // SuperWord transform 2903 SuperWord sw(this); 2904 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2905 IdealLoopTree* lpt = iter.current(); 2906 if (lpt->is_counted()) { 2907 CountedLoopNode *cl = lpt->_head->as_CountedLoop(); 2908 2909 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) { 2910 // Check that the rce'd post loop is encountered first, multiversion after all 2911 // major main loop optimization are concluded 2912 if (!C->major_progress()) { 2913 IdealLoopTree *lpt_next = lpt->_next; 2914 if (lpt_next && lpt_next->is_counted()) { 2915 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop(); 2916 has_range_checks(lpt_next); 2917 if (cl->is_post_loop() && cl->range_checks_present()) { 2918 if (!cl->is_multiversioned()) { 2919 if (multi_version_post_loops(lpt, lpt_next) == false) { 2920 // Cause the rce loop to be optimized away if we fail 2921 cl->mark_is_multiversioned(); 2922 cl->set_slp_max_unroll(0); 2923 poison_rce_post_loop(lpt); 2924 } 2925 } 2926 } 2927 } 2928 sw.transform_loop(lpt, true); 2929 } 2930 } else if (cl->is_main_loop()) { 2931 sw.transform_loop(lpt, true); 2932 } 2933 } 2934 } 2935 } 2936 2937 // Cleanup any modified bits 2938 _igvn.optimize(); 2939 2940 // disable assert until issue with split_flow_path is resolved (6742111) 2941 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), 2942 // "shouldn't introduce irreducible loops"); 2943 2944 if (C->log() != NULL) { 2945 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2946 } 2947 } 2948 2949 #ifndef PRODUCT 2950 //------------------------------print_statistics------------------------------- 2951 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes 2952 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique 2953 void PhaseIdealLoop::print_statistics() { 2954 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); 2955 } 2956 2957 //------------------------------verify----------------------------------------- 2958 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. 2959 static int fail; // debug only, so its multi-thread dont care 2960 void PhaseIdealLoop::verify() const { 2961 int old_progress = C->major_progress(); 2962 ResourceMark rm; 2963 PhaseIdealLoop loop_verify( _igvn, this ); 2964 VectorSet visited(Thread::current()->resource_area()); 2965 2966 fail = 0; 2967 verify_compare( C->root(), &loop_verify, visited ); 2968 assert( fail == 0, "verify loops failed" ); 2969 // Verify loop structure is the same 2970 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); 2971 // Reset major-progress. It was cleared by creating a verify version of 2972 // PhaseIdealLoop. 2973 for( int i=0; i<old_progress; i++ ) 2974 C->set_major_progress(); 2975 } 2976 2977 //------------------------------verify_compare--------------------------------- 2978 // Make sure me and the given PhaseIdealLoop agree on key data structures 2979 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { 2980 if( !n ) return; 2981 if( visited.test_set( n->_idx ) ) return; 2982 if( !_nodes[n->_idx] ) { // Unreachable 2983 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); 2984 return; 2985 } 2986 2987 uint i; 2988 for( i = 0; i < n->req(); i++ ) 2989 verify_compare( n->in(i), loop_verify, visited ); 2990 2991 // Check the '_nodes' block/loop structure 2992 i = n->_idx; 2993 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl 2994 if( _nodes[i] != loop_verify->_nodes[i] && 2995 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { 2996 tty->print("Mismatched control setting for: "); 2997 n->dump(); 2998 if( fail++ > 10 ) return; 2999 Node *c = get_ctrl_no_update(n); 3000 tty->print("We have it as: "); 3001 if( c->in(0) ) c->dump(); 3002 else tty->print_cr("N%d",c->_idx); 3003 tty->print("Verify thinks: "); 3004 if( loop_verify->has_ctrl(n) ) 3005 loop_verify->get_ctrl_no_update(n)->dump(); 3006 else 3007 loop_verify->get_loop_idx(n)->dump(); 3008 tty->cr(); 3009 } 3010 } else { // We have a loop 3011 IdealLoopTree *us = get_loop_idx(n); 3012 if( loop_verify->has_ctrl(n) ) { 3013 tty->print("Mismatched loop setting for: "); 3014 n->dump(); 3015 if( fail++ > 10 ) return; 3016 tty->print("We have it as: "); 3017 us->dump(); 3018 tty->print("Verify thinks: "); 3019 loop_verify->get_ctrl_no_update(n)->dump(); 3020 tty->cr(); 3021 } else if (!C->major_progress()) { 3022 // Loop selection can be messed up if we did a major progress 3023 // operation, like split-if. Do not verify in that case. 3024 IdealLoopTree *them = loop_verify->get_loop_idx(n); 3025 if( us->_head != them->_head || us->_tail != them->_tail ) { 3026 tty->print("Unequals loops for: "); 3027 n->dump(); 3028 if( fail++ > 10 ) return; 3029 tty->print("We have it as: "); 3030 us->dump(); 3031 tty->print("Verify thinks: "); 3032 them->dump(); 3033 tty->cr(); 3034 } 3035 } 3036 } 3037 3038 // Check for immediate dominators being equal 3039 if( i >= _idom_size ) { 3040 if( !n->is_CFG() ) return; 3041 tty->print("CFG Node with no idom: "); 3042 n->dump(); 3043 return; 3044 } 3045 if( !n->is_CFG() ) return; 3046 if( n == C->root() ) return; // No IDOM here 3047 3048 assert(n->_idx == i, "sanity"); 3049 Node *id = idom_no_update(n); 3050 if( id != loop_verify->idom_no_update(n) ) { 3051 tty->print("Unequals idoms for: "); 3052 n->dump(); 3053 if( fail++ > 10 ) return; 3054 tty->print("We have it as: "); 3055 id->dump(); 3056 tty->print("Verify thinks: "); 3057 loop_verify->idom_no_update(n)->dump(); 3058 tty->cr(); 3059 } 3060 3061 } 3062 3063 //------------------------------verify_tree------------------------------------ 3064 // Verify that tree structures match. Because the CFG can change, siblings 3065 // within the loop tree can be reordered. We attempt to deal with that by 3066 // reordering the verify's loop tree if possible. 3067 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { 3068 assert( _parent == parent, "Badly formed loop tree" ); 3069 3070 // Siblings not in same order? Attempt to re-order. 3071 if( _head != loop->_head ) { 3072 // Find _next pointer to update 3073 IdealLoopTree **pp = &loop->_parent->_child; 3074 while( *pp != loop ) 3075 pp = &((*pp)->_next); 3076 // Find proper sibling to be next 3077 IdealLoopTree **nn = &loop->_next; 3078 while( (*nn) && (*nn)->_head != _head ) 3079 nn = &((*nn)->_next); 3080 3081 // Check for no match. 3082 if( !(*nn) ) { 3083 // Annoyingly, irreducible loops can pick different headers 3084 // after a major_progress operation, so the rest of the loop 3085 // tree cannot be matched. 3086 if (_irreducible && Compile::current()->major_progress()) return; 3087 assert( 0, "failed to match loop tree" ); 3088 } 3089 3090 // Move (*nn) to (*pp) 3091 IdealLoopTree *hit = *nn; 3092 *nn = hit->_next; 3093 hit->_next = loop; 3094 *pp = loop; 3095 loop = hit; 3096 // Now try again to verify 3097 } 3098 3099 assert( _head == loop->_head , "mismatched loop head" ); 3100 Node *tail = _tail; // Inline a non-updating version of 3101 while( !tail->in(0) ) // the 'tail()' call. 3102 tail = tail->in(1); 3103 assert( tail == loop->_tail, "mismatched loop tail" ); 3104 3105 // Counted loops that are guarded should be able to find their guards 3106 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { 3107 CountedLoopNode *cl = _head->as_CountedLoop(); 3108 Node *init = cl->init_trip(); 3109 Node *ctrl = cl->in(LoopNode::EntryControl); 3110 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 3111 Node *iff = ctrl->in(0); 3112 assert( iff->Opcode() == Op_If, "" ); 3113 Node *bol = iff->in(1); 3114 assert( bol->Opcode() == Op_Bool, "" ); 3115 Node *cmp = bol->in(1); 3116 assert( cmp->Opcode() == Op_CmpI, "" ); 3117 Node *add = cmp->in(1); 3118 Node *opaq; 3119 if( add->Opcode() == Op_Opaque1 ) { 3120 opaq = add; 3121 } else { 3122 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); 3123 assert( add == init, "" ); 3124 opaq = cmp->in(2); 3125 } 3126 assert( opaq->Opcode() == Op_Opaque1, "" ); 3127 3128 } 3129 3130 if (_child != NULL) _child->verify_tree(loop->_child, this); 3131 if (_next != NULL) _next ->verify_tree(loop->_next, parent); 3132 // Innermost loops need to verify loop bodies, 3133 // but only if no 'major_progress' 3134 int fail = 0; 3135 if (!Compile::current()->major_progress() && _child == NULL) { 3136 for( uint i = 0; i < _body.size(); i++ ) { 3137 Node *n = _body.at(i); 3138 if (n->outcnt() == 0) continue; // Ignore dead 3139 uint j; 3140 for( j = 0; j < loop->_body.size(); j++ ) 3141 if( loop->_body.at(j) == n ) 3142 break; 3143 if( j == loop->_body.size() ) { // Not found in loop body 3144 // Last ditch effort to avoid assertion: Its possible that we 3145 // have some users (so outcnt not zero) but are still dead. 3146 // Try to find from root. 3147 if (Compile::current()->root()->find(n->_idx)) { 3148 fail++; 3149 tty->print("We have that verify does not: "); 3150 n->dump(); 3151 } 3152 } 3153 } 3154 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { 3155 Node *n = loop->_body.at(i2); 3156 if (n->outcnt() == 0) continue; // Ignore dead 3157 uint j; 3158 for( j = 0; j < _body.size(); j++ ) 3159 if( _body.at(j) == n ) 3160 break; 3161 if( j == _body.size() ) { // Not found in loop body 3162 // Last ditch effort to avoid assertion: Its possible that we 3163 // have some users (so outcnt not zero) but are still dead. 3164 // Try to find from root. 3165 if (Compile::current()->root()->find(n->_idx)) { 3166 fail++; 3167 tty->print("Verify has that we do not: "); 3168 n->dump(); 3169 } 3170 } 3171 } 3172 assert( !fail, "loop body mismatch" ); 3173 } 3174 } 3175 3176 #endif 3177 3178 //------------------------------set_idom--------------------------------------- 3179 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { 3180 uint idx = d->_idx; 3181 if (idx >= _idom_size) { 3182 uint newsize = _idom_size<<1; 3183 while( idx >= newsize ) { 3184 newsize <<= 1; 3185 } 3186 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); 3187 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); 3188 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); 3189 _idom_size = newsize; 3190 } 3191 _idom[idx] = n; 3192 _dom_depth[idx] = dom_depth; 3193 } 3194 3195 //------------------------------recompute_dom_depth--------------------------------------- 3196 // The dominator tree is constructed with only parent pointers. 3197 // This recomputes the depth in the tree by first tagging all 3198 // nodes as "no depth yet" marker. The next pass then runs up 3199 // the dom tree from each node marked "no depth yet", and computes 3200 // the depth on the way back down. 3201 void PhaseIdealLoop::recompute_dom_depth() { 3202 uint no_depth_marker = C->unique(); 3203 uint i; 3204 // Initialize depth to "no depth yet" 3205 for (i = 0; i < _idom_size; i++) { 3206 if (_dom_depth[i] > 0 && _idom[i] != NULL) { 3207 _dom_depth[i] = no_depth_marker; 3208 } 3209 } 3210 if (_dom_stk == NULL) { 3211 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size. 3212 if (init_size < 10) init_size = 10; 3213 _dom_stk = new GrowableArray<uint>(init_size); 3214 } 3215 // Compute new depth for each node. 3216 for (i = 0; i < _idom_size; i++) { 3217 uint j = i; 3218 // Run up the dom tree to find a node with a depth 3219 while (_dom_depth[j] == no_depth_marker) { 3220 _dom_stk->push(j); 3221 j = _idom[j]->_idx; 3222 } 3223 // Compute the depth on the way back down this tree branch 3224 uint dd = _dom_depth[j] + 1; 3225 while (_dom_stk->length() > 0) { 3226 uint j = _dom_stk->pop(); 3227 _dom_depth[j] = dd; 3228 dd++; 3229 } 3230 } 3231 } 3232 3233 //------------------------------sort------------------------------------------- 3234 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the 3235 // loop tree, not the root. 3236 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { 3237 if( !innermost ) return loop; // New innermost loop 3238 3239 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number 3240 assert( loop_preorder, "not yet post-walked loop" ); 3241 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer 3242 IdealLoopTree *l = *pp; // Do I go before or after 'l'? 3243 3244 // Insert at start of list 3245 while( l ) { // Insertion sort based on pre-order 3246 if( l == loop ) return innermost; // Already on list! 3247 int l_preorder = get_preorder(l->_head); // Cache pre-order number 3248 assert( l_preorder, "not yet post-walked l" ); 3249 // Check header pre-order number to figure proper nesting 3250 if( loop_preorder > l_preorder ) 3251 break; // End of insertion 3252 // If headers tie (e.g., shared headers) check tail pre-order numbers. 3253 // Since I split shared headers, you'd think this could not happen. 3254 // BUT: I must first do the preorder numbering before I can discover I 3255 // have shared headers, so the split headers all get the same preorder 3256 // number as the RegionNode they split from. 3257 if( loop_preorder == l_preorder && 3258 get_preorder(loop->_tail) < get_preorder(l->_tail) ) 3259 break; // Also check for shared headers (same pre#) 3260 pp = &l->_parent; // Chain up list 3261 l = *pp; 3262 } 3263 // Link into list 3264 // Point predecessor to me 3265 *pp = loop; 3266 // Point me to successor 3267 IdealLoopTree *p = loop->_parent; 3268 loop->_parent = l; // Point me to successor 3269 if( p ) sort( p, innermost ); // Insert my parents into list as well 3270 return innermost; 3271 } 3272 3273 //------------------------------build_loop_tree-------------------------------- 3274 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit 3275 // bits. The _nodes[] array is mapped by Node index and holds a NULL for 3276 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the 3277 // tightest enclosing IdealLoopTree for post-walked. 3278 // 3279 // During my forward walk I do a short 1-layer lookahead to see if I can find 3280 // a loop backedge with that doesn't have any work on the backedge. This 3281 // helps me construct nested loops with shared headers better. 3282 // 3283 // Once I've done the forward recursion, I do the post-work. For each child 3284 // I check to see if there is a backedge. Backedges define a loop! I 3285 // insert an IdealLoopTree at the target of the backedge. 3286 // 3287 // During the post-work I also check to see if I have several children 3288 // belonging to different loops. If so, then this Node is a decision point 3289 // where control flow can choose to change loop nests. It is at this 3290 // decision point where I can figure out how loops are nested. At this 3291 // time I can properly order the different loop nests from my children. 3292 // Note that there may not be any backedges at the decision point! 3293 // 3294 // Since the decision point can be far removed from the backedges, I can't 3295 // order my loops at the time I discover them. Thus at the decision point 3296 // I need to inspect loop header pre-order numbers to properly nest my 3297 // loops. This means I need to sort my childrens' loops by pre-order. 3298 // The sort is of size number-of-control-children, which generally limits 3299 // it to size 2 (i.e., I just choose between my 2 target loops). 3300 void PhaseIdealLoop::build_loop_tree() { 3301 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3302 GrowableArray <Node *> bltstack(C->live_nodes() >> 1); 3303 Node *n = C->root(); 3304 bltstack.push(n); 3305 int pre_order = 1; 3306 int stack_size; 3307 3308 while ( ( stack_size = bltstack.length() ) != 0 ) { 3309 n = bltstack.top(); // Leave node on stack 3310 if ( !is_visited(n) ) { 3311 // ---- Pre-pass Work ---- 3312 // Pre-walked but not post-walked nodes need a pre_order number. 3313 3314 set_preorder_visited( n, pre_order ); // set as visited 3315 3316 // ---- Scan over children ---- 3317 // Scan first over control projections that lead to loop headers. 3318 // This helps us find inner-to-outer loops with shared headers better. 3319 3320 // Scan children's children for loop headers. 3321 for ( int i = n->outcnt() - 1; i >= 0; --i ) { 3322 Node* m = n->raw_out(i); // Child 3323 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children 3324 // Scan over children's children to find loop 3325 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3326 Node* l = m->fast_out(j); 3327 if( is_visited(l) && // Been visited? 3328 !is_postvisited(l) && // But not post-visited 3329 get_preorder(l) < pre_order ) { // And smaller pre-order 3330 // Found! Scan the DFS down this path before doing other paths 3331 bltstack.push(m); 3332 break; 3333 } 3334 } 3335 } 3336 } 3337 pre_order++; 3338 } 3339 else if ( !is_postvisited(n) ) { 3340 // Note: build_loop_tree_impl() adds out edges on rare occasions, 3341 // such as com.sun.rsasign.am::a. 3342 // For non-recursive version, first, process current children. 3343 // On next iteration, check if additional children were added. 3344 for ( int k = n->outcnt() - 1; k >= 0; --k ) { 3345 Node* u = n->raw_out(k); 3346 if ( u->is_CFG() && !is_visited(u) ) { 3347 bltstack.push(u); 3348 } 3349 } 3350 if ( bltstack.length() == stack_size ) { 3351 // There were no additional children, post visit node now 3352 (void)bltstack.pop(); // Remove node from stack 3353 pre_order = build_loop_tree_impl( n, pre_order ); 3354 // Check for bailout 3355 if (C->failing()) { 3356 return; 3357 } 3358 // Check to grow _preorders[] array for the case when 3359 // build_loop_tree_impl() adds new nodes. 3360 check_grow_preorders(); 3361 } 3362 } 3363 else { 3364 (void)bltstack.pop(); // Remove post-visited node from stack 3365 } 3366 } 3367 } 3368 3369 //------------------------------build_loop_tree_impl--------------------------- 3370 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { 3371 // ---- Post-pass Work ---- 3372 // Pre-walked but not post-walked nodes need a pre_order number. 3373 3374 // Tightest enclosing loop for this Node 3375 IdealLoopTree *innermost = NULL; 3376 3377 // For all children, see if any edge is a backedge. If so, make a loop 3378 // for it. Then find the tightest enclosing loop for the self Node. 3379 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3380 Node* m = n->fast_out(i); // Child 3381 if( n == m ) continue; // Ignore control self-cycles 3382 if( !m->is_CFG() ) continue;// Ignore non-CFG edges 3383 3384 IdealLoopTree *l; // Child's loop 3385 if( !is_postvisited(m) ) { // Child visited but not post-visited? 3386 // Found a backedge 3387 assert( get_preorder(m) < pre_order, "should be backedge" ); 3388 // Check for the RootNode, which is already a LoopNode and is allowed 3389 // to have multiple "backedges". 3390 if( m == C->root()) { // Found the root? 3391 l = _ltree_root; // Root is the outermost LoopNode 3392 } else { // Else found a nested loop 3393 // Insert a LoopNode to mark this loop. 3394 l = new IdealLoopTree(this, m, n); 3395 } // End of Else found a nested loop 3396 if( !has_loop(m) ) // If 'm' does not already have a loop set 3397 set_loop(m, l); // Set loop header to loop now 3398 3399 } else { // Else not a nested loop 3400 if( !_nodes[m->_idx] ) continue; // Dead code has no loop 3401 l = get_loop(m); // Get previously determined loop 3402 // If successor is header of a loop (nest), move up-loop till it 3403 // is a member of some outer enclosing loop. Since there are no 3404 // shared headers (I've split them already) I only need to go up 3405 // at most 1 level. 3406 while( l && l->_head == m ) // Successor heads loop? 3407 l = l->_parent; // Move up 1 for me 3408 // If this loop is not properly parented, then this loop 3409 // has no exit path out, i.e. its an infinite loop. 3410 if( !l ) { 3411 // Make loop "reachable" from root so the CFG is reachable. Basically 3412 // insert a bogus loop exit that is never taken. 'm', the loop head, 3413 // points to 'n', one (of possibly many) fall-in paths. There may be 3414 // many backedges as well. 3415 3416 // Here I set the loop to be the root loop. I could have, after 3417 // inserting a bogus loop exit, restarted the recursion and found my 3418 // new loop exit. This would make the infinite loop a first-class 3419 // loop and it would then get properly optimized. What's the use of 3420 // optimizing an infinite loop? 3421 l = _ltree_root; // Oops, found infinite loop 3422 3423 if (!_verify_only) { 3424 // Insert the NeverBranch between 'm' and it's control user. 3425 NeverBranchNode *iff = new NeverBranchNode( m ); 3426 _igvn.register_new_node_with_optimizer(iff); 3427 set_loop(iff, l); 3428 Node *if_t = new CProjNode( iff, 0 ); 3429 _igvn.register_new_node_with_optimizer(if_t); 3430 set_loop(if_t, l); 3431 3432 Node* cfg = NULL; // Find the One True Control User of m 3433 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3434 Node* x = m->fast_out(j); 3435 if (x->is_CFG() && x != m && x != iff) 3436 { cfg = x; break; } 3437 } 3438 assert(cfg != NULL, "must find the control user of m"); 3439 uint k = 0; // Probably cfg->in(0) 3440 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region 3441 cfg->set_req( k, if_t ); // Now point to NeverBranch 3442 _igvn._worklist.push(cfg); 3443 3444 // Now create the never-taken loop exit 3445 Node *if_f = new CProjNode( iff, 1 ); 3446 _igvn.register_new_node_with_optimizer(if_f); 3447 set_loop(if_f, l); 3448 // Find frame ptr for Halt. Relies on the optimizer 3449 // V-N'ing. Easier and quicker than searching through 3450 // the program structure. 3451 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr ); 3452 _igvn.register_new_node_with_optimizer(frame); 3453 // Halt & Catch Fire 3454 Node *halt = new HaltNode( if_f, frame ); 3455 _igvn.register_new_node_with_optimizer(halt); 3456 set_loop(halt, l); 3457 C->root()->add_req(halt); 3458 } 3459 set_loop(C->root(), _ltree_root); 3460 } 3461 } 3462 // Weeny check for irreducible. This child was already visited (this 3463 // IS the post-work phase). Is this child's loop header post-visited 3464 // as well? If so, then I found another entry into the loop. 3465 if (!_verify_only) { 3466 while( is_postvisited(l->_head) ) { 3467 // found irreducible 3468 l->_irreducible = 1; // = true 3469 l = l->_parent; 3470 _has_irreducible_loops = true; 3471 // Check for bad CFG here to prevent crash, and bailout of compile 3472 if (l == NULL) { 3473 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); 3474 return pre_order; 3475 } 3476 } 3477 C->set_has_irreducible_loop(_has_irreducible_loops); 3478 } 3479 3480 // This Node might be a decision point for loops. It is only if 3481 // it's children belong to several different loops. The sort call 3482 // does a trivial amount of work if there is only 1 child or all 3483 // children belong to the same loop. If however, the children 3484 // belong to different loops, the sort call will properly set the 3485 // _parent pointers to show how the loops nest. 3486 // 3487 // In any case, it returns the tightest enclosing loop. 3488 innermost = sort( l, innermost ); 3489 } 3490 3491 // Def-use info will have some dead stuff; dead stuff will have no 3492 // loop decided on. 3493 3494 // Am I a loop header? If so fix up my parent's child and next ptrs. 3495 if( innermost && innermost->_head == n ) { 3496 assert( get_loop(n) == innermost, "" ); 3497 IdealLoopTree *p = innermost->_parent; 3498 IdealLoopTree *l = innermost; 3499 while( p && l->_head == n ) { 3500 l->_next = p->_child; // Put self on parents 'next child' 3501 p->_child = l; // Make self as first child of parent 3502 l = p; // Now walk up the parent chain 3503 p = l->_parent; 3504 } 3505 } else { 3506 // Note that it is possible for a LoopNode to reach here, if the 3507 // backedge has been made unreachable (hence the LoopNode no longer 3508 // denotes a Loop, and will eventually be removed). 3509 3510 // Record tightest enclosing loop for self. Mark as post-visited. 3511 set_loop(n, innermost); 3512 // Also record has_call flag early on 3513 if( innermost ) { 3514 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { 3515 // Do not count uncommon calls 3516 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { 3517 Node *iff = n->in(0)->in(0); 3518 // No any calls for vectorized loops. 3519 if( UseSuperWord || !iff->is_If() || 3520 (n->in(0)->Opcode() == Op_IfFalse && 3521 (1.0 - iff->as_If()->_prob) >= 0.01) || 3522 (iff->as_If()->_prob >= 0.01) ) 3523 innermost->_has_call = 1; 3524 } 3525 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { 3526 // Disable loop optimizations if the loop has a scalar replaceable 3527 // allocation. This disabling may cause a potential performance lost 3528 // if the allocation is not eliminated for some reason. 3529 innermost->_allow_optimizations = false; 3530 innermost->_has_call = 1; // = true 3531 } else if (n->Opcode() == Op_SafePoint) { 3532 // Record all safepoints in this loop. 3533 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List(); 3534 innermost->_safepts->push(n); 3535 } 3536 } 3537 } 3538 3539 // Flag as post-visited now 3540 set_postvisited(n); 3541 return pre_order; 3542 } 3543 3544 3545 //------------------------------build_loop_early------------------------------- 3546 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3547 // First pass computes the earliest controlling node possible. This is the 3548 // controlling input with the deepest dominating depth. 3549 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3550 while (worklist.size() != 0) { 3551 // Use local variables nstack_top_n & nstack_top_i to cache values 3552 // on nstack's top. 3553 Node *nstack_top_n = worklist.pop(); 3554 uint nstack_top_i = 0; 3555 //while_nstack_nonempty: 3556 while (true) { 3557 // Get parent node and next input's index from stack's top. 3558 Node *n = nstack_top_n; 3559 uint i = nstack_top_i; 3560 uint cnt = n->req(); // Count of inputs 3561 if (i == 0) { // Pre-process the node. 3562 if( has_node(n) && // Have either loop or control already? 3563 !has_ctrl(n) ) { // Have loop picked out already? 3564 // During "merge_many_backedges" we fold up several nested loops 3565 // into a single loop. This makes the members of the original 3566 // loop bodies pointing to dead loops; they need to move up 3567 // to the new UNION'd larger loop. I set the _head field of these 3568 // dead loops to NULL and the _parent field points to the owning 3569 // loop. Shades of UNION-FIND algorithm. 3570 IdealLoopTree *ilt; 3571 while( !(ilt = get_loop(n))->_head ) { 3572 // Normally I would use a set_loop here. But in this one special 3573 // case, it is legal (and expected) to change what loop a Node 3574 // belongs to. 3575 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); 3576 } 3577 // Remove safepoints ONLY if I've already seen I don't need one. 3578 // (the old code here would yank a 2nd safepoint after seeing a 3579 // first one, even though the 1st did not dominate in the loop body 3580 // and thus could be avoided indefinitely) 3581 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && 3582 is_deleteable_safept(n)) { 3583 Node *in = n->in(TypeFunc::Control); 3584 lazy_replace(n,in); // Pull safepoint now 3585 if (ilt->_safepts != NULL) { 3586 ilt->_safepts->yank(n); 3587 } 3588 // Carry on with the recursion "as if" we are walking 3589 // only the control input 3590 if( !visited.test_set( in->_idx ) ) { 3591 worklist.push(in); // Visit this guy later, using worklist 3592 } 3593 // Get next node from nstack: 3594 // - skip n's inputs processing by setting i > cnt; 3595 // - we also will not call set_early_ctrl(n) since 3596 // has_node(n) == true (see the condition above). 3597 i = cnt + 1; 3598 } 3599 } 3600 } // if (i == 0) 3601 3602 // Visit all inputs 3603 bool done = true; // Assume all n's inputs will be processed 3604 while (i < cnt) { 3605 Node *in = n->in(i); 3606 ++i; 3607 if (in == NULL) continue; 3608 if (in->pinned() && !in->is_CFG()) 3609 set_ctrl(in, in->in(0)); 3610 int is_visited = visited.test_set( in->_idx ); 3611 if (!has_node(in)) { // No controlling input yet? 3612 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); 3613 assert( !is_visited, "visit only once" ); 3614 nstack.push(n, i); // Save parent node and next input's index. 3615 nstack_top_n = in; // Process current input now. 3616 nstack_top_i = 0; 3617 done = false; // Not all n's inputs processed. 3618 break; // continue while_nstack_nonempty; 3619 } else if (!is_visited) { 3620 // This guy has a location picked out for him, but has not yet 3621 // been visited. Happens to all CFG nodes, for instance. 3622 // Visit him using the worklist instead of recursion, to break 3623 // cycles. Since he has a location already we do not need to 3624 // find his location before proceeding with the current Node. 3625 worklist.push(in); // Visit this guy later, using worklist 3626 } 3627 } 3628 if (done) { 3629 // All of n's inputs have been processed, complete post-processing. 3630 3631 // Compute earliest point this Node can go. 3632 // CFG, Phi, pinned nodes already know their controlling input. 3633 if (!has_node(n)) { 3634 // Record earliest legal location 3635 set_early_ctrl( n ); 3636 } 3637 if (nstack.is_empty()) { 3638 // Finished all nodes on stack. 3639 // Process next node on the worklist. 3640 break; 3641 } 3642 // Get saved parent node and next input's index. 3643 nstack_top_n = nstack.node(); 3644 nstack_top_i = nstack.index(); 3645 nstack.pop(); 3646 } 3647 } // while (true) 3648 } 3649 } 3650 3651 //------------------------------dom_lca_internal-------------------------------- 3652 // Pair-wise LCA 3653 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { 3654 if( !n1 ) return n2; // Handle NULL original LCA 3655 assert( n1->is_CFG(), "" ); 3656 assert( n2->is_CFG(), "" ); 3657 // find LCA of all uses 3658 uint d1 = dom_depth(n1); 3659 uint d2 = dom_depth(n2); 3660 while (n1 != n2) { 3661 if (d1 > d2) { 3662 n1 = idom(n1); 3663 d1 = dom_depth(n1); 3664 } else if (d1 < d2) { 3665 n2 = idom(n2); 3666 d2 = dom_depth(n2); 3667 } else { 3668 // Here d1 == d2. Due to edits of the dominator-tree, sections 3669 // of the tree might have the same depth. These sections have 3670 // to be searched more carefully. 3671 3672 // Scan up all the n1's with equal depth, looking for n2. 3673 Node *t1 = idom(n1); 3674 while (dom_depth(t1) == d1) { 3675 if (t1 == n2) return n2; 3676 t1 = idom(t1); 3677 } 3678 // Scan up all the n2's with equal depth, looking for n1. 3679 Node *t2 = idom(n2); 3680 while (dom_depth(t2) == d2) { 3681 if (t2 == n1) return n1; 3682 t2 = idom(t2); 3683 } 3684 // Move up to a new dominator-depth value as well as up the dom-tree. 3685 n1 = t1; 3686 n2 = t2; 3687 d1 = dom_depth(n1); 3688 d2 = dom_depth(n2); 3689 } 3690 } 3691 return n1; 3692 } 3693 3694 //------------------------------compute_idom----------------------------------- 3695 // Locally compute IDOM using dom_lca call. Correct only if the incoming 3696 // IDOMs are correct. 3697 Node *PhaseIdealLoop::compute_idom( Node *region ) const { 3698 assert( region->is_Region(), "" ); 3699 Node *LCA = NULL; 3700 for( uint i = 1; i < region->req(); i++ ) { 3701 if( region->in(i) != C->top() ) 3702 LCA = dom_lca( LCA, region->in(i) ); 3703 } 3704 return LCA; 3705 } 3706 3707 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { 3708 bool had_error = false; 3709 #ifdef ASSERT 3710 if (early != C->root()) { 3711 // Make sure that there's a dominance path from LCA to early 3712 Node* d = LCA; 3713 while (d != early) { 3714 if (d == C->root()) { 3715 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA); 3716 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx); 3717 had_error = true; 3718 break; 3719 } 3720 d = idom(d); 3721 } 3722 } 3723 #endif 3724 return had_error; 3725 } 3726 3727 3728 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { 3729 // Compute LCA over list of uses 3730 bool had_error = false; 3731 Node *LCA = NULL; 3732 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { 3733 Node* c = n->fast_out(i); 3734 if (_nodes[c->_idx] == NULL) 3735 continue; // Skip the occasional dead node 3736 if( c->is_Phi() ) { // For Phis, we must land above on the path 3737 for( uint j=1; j<c->req(); j++ ) {// For all inputs 3738 if( c->in(j) == n ) { // Found matching input? 3739 Node *use = c->in(0)->in(j); 3740 if (_verify_only && use->is_top()) continue; 3741 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3742 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3743 } 3744 } 3745 } else { 3746 // For CFG data-users, use is in the block just prior 3747 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); 3748 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3749 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3750 } 3751 } 3752 assert(!had_error, "bad dominance"); 3753 return LCA; 3754 } 3755 3756 // Check the shape of the graph at the loop entry. In some cases, 3757 // the shape of the graph does not match the shape outlined below. 3758 // That is caused by the Opaque1 node "protecting" the shape of 3759 // the graph being removed by, for example, the IGVN performed 3760 // in PhaseIdealLoop::build_and_optimize(). 3761 // 3762 // After the Opaque1 node has been removed, optimizations (e.g., split-if, 3763 // loop unswitching, and IGVN, or a combination of them) can freely change 3764 // the graph's shape. As a result, the graph shape outlined below cannot 3765 // be guaranteed anymore. 3766 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) { 3767 if (!cl->is_main_loop() && !cl->is_post_loop()) { 3768 return false; 3769 } 3770 Node* ctrl = cl->skip_strip_mined()->in(LoopNode::EntryControl); 3771 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) { 3772 return false; 3773 } 3774 Node* iffm = ctrl->in(0); 3775 if (iffm == NULL || !iffm->is_If()) { 3776 return false; 3777 } 3778 Node* bolzm = iffm->in(1); 3779 if (bolzm == NULL || !bolzm->is_Bool()) { 3780 return false; 3781 } 3782 Node* cmpzm = bolzm->in(1); 3783 if (cmpzm == NULL || !cmpzm->is_Cmp()) { 3784 return false; 3785 } 3786 // compares can get conditionally flipped 3787 bool found_opaque = false; 3788 for (uint i = 1; i < cmpzm->req(); i++) { 3789 Node* opnd = cmpzm->in(i); 3790 if (opnd && opnd->Opcode() == Op_Opaque1) { 3791 found_opaque = true; 3792 break; 3793 } 3794 } 3795 if (!found_opaque) { 3796 return false; 3797 } 3798 return true; 3799 } 3800 3801 //------------------------------get_late_ctrl---------------------------------- 3802 // Compute latest legal control. 3803 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { 3804 assert(early != NULL, "early control should not be NULL"); 3805 3806 Node* LCA = compute_lca_of_uses(n, early); 3807 #ifdef ASSERT 3808 if (LCA == C->root() && LCA != early) { 3809 // def doesn't dominate uses so print some useful debugging output 3810 compute_lca_of_uses(n, early, true); 3811 } 3812 #endif 3813 3814 // if this is a load, check for anti-dependent stores 3815 // We use a conservative algorithm to identify potential interfering 3816 // instructions and for rescheduling the load. The users of the memory 3817 // input of this load are examined. Any use which is not a load and is 3818 // dominated by early is considered a potentially interfering store. 3819 // This can produce false positives. 3820 if (n->is_Load() && LCA != early) { 3821 Node_List worklist; 3822 3823 Node *mem = n->in(MemNode::Memory); 3824 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 3825 Node* s = mem->fast_out(i); 3826 worklist.push(s); 3827 } 3828 while(worklist.size() != 0 && LCA != early) { 3829 Node* s = worklist.pop(); 3830 if (s->is_Load() || s->Opcode() == Op_SafePoint) { 3831 continue; 3832 } else if (s->is_MergeMem()) { 3833 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 3834 Node* s1 = s->fast_out(i); 3835 worklist.push(s1); 3836 } 3837 } else { 3838 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); 3839 assert(sctrl != NULL || s->outcnt() == 0, "must have control"); 3840 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { 3841 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); 3842 } 3843 } 3844 } 3845 } 3846 3847 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); 3848 return LCA; 3849 } 3850 3851 // true if CFG node d dominates CFG node n 3852 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { 3853 if (d == n) 3854 return true; 3855 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 3856 uint dd = dom_depth(d); 3857 while (dom_depth(n) >= dd) { 3858 if (n == d) 3859 return true; 3860 n = idom(n); 3861 } 3862 return false; 3863 } 3864 3865 //------------------------------dom_lca_for_get_late_ctrl_internal------------- 3866 // Pair-wise LCA with tags. 3867 // Tag each index with the node 'tag' currently being processed 3868 // before advancing up the dominator chain using idom(). 3869 // Later calls that find a match to 'tag' know that this path has already 3870 // been considered in the current LCA (which is input 'n1' by convention). 3871 // Since get_late_ctrl() is only called once for each node, the tag array 3872 // does not need to be cleared between calls to get_late_ctrl(). 3873 // Algorithm trades a larger constant factor for better asymptotic behavior 3874 // 3875 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { 3876 uint d1 = dom_depth(n1); 3877 uint d2 = dom_depth(n2); 3878 3879 do { 3880 if (d1 > d2) { 3881 // current lca is deeper than n2 3882 _dom_lca_tags.map(n1->_idx, tag); 3883 n1 = idom(n1); 3884 d1 = dom_depth(n1); 3885 } else if (d1 < d2) { 3886 // n2 is deeper than current lca 3887 Node *memo = _dom_lca_tags[n2->_idx]; 3888 if( memo == tag ) { 3889 return n1; // Return the current LCA 3890 } 3891 _dom_lca_tags.map(n2->_idx, tag); 3892 n2 = idom(n2); 3893 d2 = dom_depth(n2); 3894 } else { 3895 // Here d1 == d2. Due to edits of the dominator-tree, sections 3896 // of the tree might have the same depth. These sections have 3897 // to be searched more carefully. 3898 3899 // Scan up all the n1's with equal depth, looking for n2. 3900 _dom_lca_tags.map(n1->_idx, tag); 3901 Node *t1 = idom(n1); 3902 while (dom_depth(t1) == d1) { 3903 if (t1 == n2) return n2; 3904 _dom_lca_tags.map(t1->_idx, tag); 3905 t1 = idom(t1); 3906 } 3907 // Scan up all the n2's with equal depth, looking for n1. 3908 _dom_lca_tags.map(n2->_idx, tag); 3909 Node *t2 = idom(n2); 3910 while (dom_depth(t2) == d2) { 3911 if (t2 == n1) return n1; 3912 _dom_lca_tags.map(t2->_idx, tag); 3913 t2 = idom(t2); 3914 } 3915 // Move up to a new dominator-depth value as well as up the dom-tree. 3916 n1 = t1; 3917 n2 = t2; 3918 d1 = dom_depth(n1); 3919 d2 = dom_depth(n2); 3920 } 3921 } while (n1 != n2); 3922 return n1; 3923 } 3924 3925 //------------------------------init_dom_lca_tags------------------------------ 3926 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3927 // Intended use does not involve any growth for the array, so it could 3928 // be of fixed size. 3929 void PhaseIdealLoop::init_dom_lca_tags() { 3930 uint limit = C->unique() + 1; 3931 _dom_lca_tags.map( limit, NULL ); 3932 #ifdef ASSERT 3933 for( uint i = 0; i < limit; ++i ) { 3934 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3935 } 3936 #endif // ASSERT 3937 } 3938 3939 //------------------------------clear_dom_lca_tags------------------------------ 3940 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 3941 // Intended use does not involve any growth for the array, so it could 3942 // be of fixed size. 3943 void PhaseIdealLoop::clear_dom_lca_tags() { 3944 uint limit = C->unique() + 1; 3945 _dom_lca_tags.map( limit, NULL ); 3946 _dom_lca_tags.clear(); 3947 #ifdef ASSERT 3948 for( uint i = 0; i < limit; ++i ) { 3949 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 3950 } 3951 #endif // ASSERT 3952 } 3953 3954 //------------------------------build_loop_late-------------------------------- 3955 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3956 // Second pass finds latest legal placement, and ideal loop placement. 3957 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3958 while (worklist.size() != 0) { 3959 Node *n = worklist.pop(); 3960 // Only visit once 3961 if (visited.test_set(n->_idx)) continue; 3962 uint cnt = n->outcnt(); 3963 uint i = 0; 3964 while (true) { 3965 assert( _nodes[n->_idx], "no dead nodes" ); 3966 // Visit all children 3967 if (i < cnt) { 3968 Node* use = n->raw_out(i); 3969 ++i; 3970 // Check for dead uses. Aggressively prune such junk. It might be 3971 // dead in the global sense, but still have local uses so I cannot 3972 // easily call 'remove_dead_node'. 3973 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? 3974 // Due to cycles, we might not hit the same fixed point in the verify 3975 // pass as we do in the regular pass. Instead, visit such phis as 3976 // simple uses of the loop head. 3977 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { 3978 if( !visited.test(use->_idx) ) 3979 worklist.push(use); 3980 } else if( !visited.test_set(use->_idx) ) { 3981 nstack.push(n, i); // Save parent and next use's index. 3982 n = use; // Process all children of current use. 3983 cnt = use->outcnt(); 3984 i = 0; 3985 } 3986 } else { 3987 // Do not visit around the backedge of loops via data edges. 3988 // push dead code onto a worklist 3989 _deadlist.push(use); 3990 } 3991 } else { 3992 // All of n's children have been processed, complete post-processing. 3993 build_loop_late_post(n); 3994 if (nstack.is_empty()) { 3995 // Finished all nodes on stack. 3996 // Process next node on the worklist. 3997 break; 3998 } 3999 // Get saved parent node and next use's index. Visit the rest of uses. 4000 n = nstack.node(); 4001 cnt = n->outcnt(); 4002 i = nstack.index(); 4003 nstack.pop(); 4004 } 4005 } 4006 } 4007 } 4008 4009 // Verify that no data node is schedules in the outer loop of a strip 4010 // mined loop. 4011 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) { 4012 #ifdef ASSERT 4013 if (get_loop(least)->_nest == 0) { 4014 return; 4015 } 4016 IdealLoopTree* loop = get_loop(least); 4017 Node* head = loop->_head; 4018 if (head->is_OuterStripMinedLoop()) { 4019 Node* sfpt = head->as_Loop()->outer_safepoint(); 4020 ResourceMark rm; 4021 Unique_Node_List wq; 4022 wq.push(sfpt); 4023 for (uint i = 0; i < wq.size(); i++) { 4024 Node *m = wq.at(i); 4025 for (uint i = 1; i < m->req(); i++) { 4026 Node* nn = m->in(i); 4027 if (nn == n) { 4028 return; 4029 } 4030 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) { 4031 wq.push(nn); 4032 } 4033 } 4034 } 4035 ShouldNotReachHere(); 4036 } 4037 #endif 4038 } 4039 4040 4041 //------------------------------build_loop_late_post--------------------------- 4042 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 4043 // Second pass finds latest legal placement, and ideal loop placement. 4044 void PhaseIdealLoop::build_loop_late_post( Node *n ) { 4045 4046 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) { 4047 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. 4048 } 4049 4050 #ifdef ASSERT 4051 if (_verify_only && !n->is_CFG()) { 4052 // Check def-use domination. 4053 compute_lca_of_uses(n, get_ctrl(n), true /* verify */); 4054 } 4055 #endif 4056 4057 // CFG and pinned nodes already handled 4058 if( n->in(0) ) { 4059 if( n->in(0)->is_top() ) return; // Dead? 4060 4061 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads 4062 // _must_ be pinned (they have to observe their control edge of course). 4063 // Unlike Stores (which modify an unallocable resource, the memory 4064 // state), Mods/Loads can float around. So free them up. 4065 bool pinned = true; 4066 switch( n->Opcode() ) { 4067 case Op_DivI: 4068 case Op_DivF: 4069 case Op_DivD: 4070 case Op_ModI: 4071 case Op_ModF: 4072 case Op_ModD: 4073 case Op_LoadB: // Same with Loads; they can sink 4074 case Op_LoadUB: // during loop optimizations. 4075 case Op_LoadUS: 4076 case Op_LoadD: 4077 case Op_LoadF: 4078 case Op_LoadI: 4079 case Op_LoadKlass: 4080 case Op_LoadNKlass: 4081 case Op_LoadL: 4082 case Op_LoadS: 4083 case Op_LoadP: 4084 case Op_LoadN: 4085 case Op_LoadRange: 4086 case Op_LoadD_unaligned: 4087 case Op_LoadL_unaligned: 4088 case Op_StrComp: // Does a bunch of load-like effects 4089 case Op_StrEquals: 4090 case Op_StrIndexOf: 4091 case Op_StrIndexOfChar: 4092 case Op_AryEq: 4093 case Op_HasNegatives: 4094 pinned = false; 4095 } 4096 if( pinned ) { 4097 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); 4098 if( !chosen_loop->_child ) // Inner loop? 4099 chosen_loop->_body.push(n); // Collect inner loops 4100 return; 4101 } 4102 } else { // No slot zero 4103 if( n->is_CFG() ) { // CFG with no slot 0 is dead 4104 _nodes.map(n->_idx,0); // No block setting, it's globally dead 4105 return; 4106 } 4107 assert(!n->is_CFG() || n->outcnt() == 0, ""); 4108 } 4109 4110 // Do I have a "safe range" I can select over? 4111 Node *early = get_ctrl(n);// Early location already computed 4112 4113 // Compute latest point this Node can go 4114 Node *LCA = get_late_ctrl( n, early ); 4115 // LCA is NULL due to uses being dead 4116 if( LCA == NULL ) { 4117 #ifdef ASSERT 4118 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { 4119 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); 4120 } 4121 #endif 4122 _nodes.map(n->_idx, 0); // This node is useless 4123 _deadlist.push(n); 4124 return; 4125 } 4126 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); 4127 4128 Node *legal = LCA; // Walk 'legal' up the IDOM chain 4129 Node *least = legal; // Best legal position so far 4130 while( early != legal ) { // While not at earliest legal 4131 #ifdef ASSERT 4132 if (legal->is_Start() && !early->is_Root()) { 4133 // Bad graph. Print idom path and fail. 4134 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA); 4135 assert(false, "Bad graph detected in build_loop_late"); 4136 } 4137 #endif 4138 // Find least loop nesting depth 4139 legal = idom(legal); // Bump up the IDOM tree 4140 // Check for lower nesting depth 4141 if( get_loop(legal)->_nest < get_loop(least)->_nest ) 4142 least = legal; 4143 } 4144 assert(early == legal || legal != C->root(), "bad dominance of inputs"); 4145 4146 // Try not to place code on a loop entry projection 4147 // which can inhibit range check elimination. 4148 if (least != early) { 4149 Node* ctrl_out = least->unique_ctrl_out(); 4150 if (ctrl_out && ctrl_out->is_Loop() && 4151 least == ctrl_out->in(LoopNode::EntryControl) && 4152 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) { 4153 Node* least_dom = idom(least); 4154 if (get_loop(least_dom)->is_member(get_loop(least))) { 4155 least = least_dom; 4156 } 4157 } 4158 } 4159 4160 #ifdef ASSERT 4161 // If verifying, verify that 'verify_me' has a legal location 4162 // and choose it as our location. 4163 if( _verify_me ) { 4164 Node *v_ctrl = _verify_me->get_ctrl_no_update(n); 4165 Node *legal = LCA; 4166 while( early != legal ) { // While not at earliest legal 4167 if( legal == v_ctrl ) break; // Check for prior good location 4168 legal = idom(legal) ;// Bump up the IDOM tree 4169 } 4170 // Check for prior good location 4171 if( legal == v_ctrl ) least = legal; // Keep prior if found 4172 } 4173 #endif 4174 4175 // Assign discovered "here or above" point 4176 least = find_non_split_ctrl(least); 4177 verify_strip_mined_scheduling(n, least); 4178 set_ctrl(n, least); 4179 4180 // Collect inner loop bodies 4181 IdealLoopTree *chosen_loop = get_loop(least); 4182 if( !chosen_loop->_child ) // Inner loop? 4183 chosen_loop->_body.push(n);// Collect inner loops 4184 } 4185 4186 #ifdef ASSERT 4187 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) { 4188 tty->print_cr("%s", msg); 4189 tty->print("n: "); n->dump(); 4190 tty->print("early(n): "); early->dump(); 4191 if (n->in(0) != NULL && !n->in(0)->is_top() && 4192 n->in(0) != early && !n->in(0)->is_Root()) { 4193 tty->print("n->in(0): "); n->in(0)->dump(); 4194 } 4195 for (uint i = 1; i < n->req(); i++) { 4196 Node* in1 = n->in(i); 4197 if (in1 != NULL && in1 != n && !in1->is_top()) { 4198 tty->print("n->in(%d): ", i); in1->dump(); 4199 Node* in1_early = get_ctrl(in1); 4200 tty->print("early(n->in(%d)): ", i); in1_early->dump(); 4201 if (in1->in(0) != NULL && !in1->in(0)->is_top() && 4202 in1->in(0) != in1_early && !in1->in(0)->is_Root()) { 4203 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump(); 4204 } 4205 for (uint j = 1; j < in1->req(); j++) { 4206 Node* in2 = in1->in(j); 4207 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) { 4208 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump(); 4209 Node* in2_early = get_ctrl(in2); 4210 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); 4211 if (in2->in(0) != NULL && !in2->in(0)->is_top() && 4212 in2->in(0) != in2_early && !in2->in(0)->is_Root()) { 4213 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump(); 4214 } 4215 } 4216 } 4217 } 4218 } 4219 tty->cr(); 4220 tty->print("LCA(n): "); LCA->dump(); 4221 for (uint i = 0; i < n->outcnt(); i++) { 4222 Node* u1 = n->raw_out(i); 4223 if (u1 == n) 4224 continue; 4225 tty->print("n->out(%d): ", i); u1->dump(); 4226 if (u1->is_CFG()) { 4227 for (uint j = 0; j < u1->outcnt(); j++) { 4228 Node* u2 = u1->raw_out(j); 4229 if (u2 != u1 && u2 != n && u2->is_CFG()) { 4230 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4231 } 4232 } 4233 } else { 4234 Node* u1_later = get_ctrl(u1); 4235 tty->print("later(n->out(%d)): ", i); u1_later->dump(); 4236 if (u1->in(0) != NULL && !u1->in(0)->is_top() && 4237 u1->in(0) != u1_later && !u1->in(0)->is_Root()) { 4238 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump(); 4239 } 4240 for (uint j = 0; j < u1->outcnt(); j++) { 4241 Node* u2 = u1->raw_out(j); 4242 if (u2 == n || u2 == u1) 4243 continue; 4244 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4245 if (!u2->is_CFG()) { 4246 Node* u2_later = get_ctrl(u2); 4247 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); 4248 if (u2->in(0) != NULL && !u2->in(0)->is_top() && 4249 u2->in(0) != u2_later && !u2->in(0)->is_Root()) { 4250 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump(); 4251 } 4252 } 4253 } 4254 } 4255 } 4256 tty->cr(); 4257 int ct = 0; 4258 Node *dbg_legal = LCA; 4259 while(!dbg_legal->is_Start() && ct < 100) { 4260 tty->print("idom[%d] ",ct); dbg_legal->dump(); 4261 ct++; 4262 dbg_legal = idom(dbg_legal); 4263 } 4264 tty->cr(); 4265 } 4266 #endif 4267 4268 #ifndef PRODUCT 4269 //------------------------------dump------------------------------------------- 4270 void PhaseIdealLoop::dump( ) const { 4271 ResourceMark rm; 4272 Arena* arena = Thread::current()->resource_area(); 4273 Node_Stack stack(arena, C->live_nodes() >> 2); 4274 Node_List rpo_list; 4275 VectorSet visited(arena); 4276 visited.set(C->top()->_idx); 4277 rpo( C->root(), stack, visited, rpo_list ); 4278 // Dump root loop indexed by last element in PO order 4279 dump( _ltree_root, rpo_list.size(), rpo_list ); 4280 } 4281 4282 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { 4283 loop->dump_head(); 4284 4285 // Now scan for CFG nodes in the same loop 4286 for( uint j=idx; j > 0; j-- ) { 4287 Node *n = rpo_list[j-1]; 4288 if( !_nodes[n->_idx] ) // Skip dead nodes 4289 continue; 4290 if( get_loop(n) != loop ) { // Wrong loop nest 4291 if( get_loop(n)->_head == n && // Found nested loop? 4292 get_loop(n)->_parent == loop ) 4293 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly 4294 continue; 4295 } 4296 4297 // Dump controlling node 4298 for( uint x = 0; x < loop->_nest; x++ ) 4299 tty->print(" "); 4300 tty->print("C"); 4301 if( n == C->root() ) { 4302 n->dump(); 4303 } else { 4304 Node* cached_idom = idom_no_update(n); 4305 Node *computed_idom = n->in(0); 4306 if( n->is_Region() ) { 4307 computed_idom = compute_idom(n); 4308 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or 4309 // any MultiBranch ctrl node), so apply a similar transform to 4310 // the cached idom returned from idom_no_update. 4311 cached_idom = find_non_split_ctrl(cached_idom); 4312 } 4313 tty->print(" ID:%d",computed_idom->_idx); 4314 n->dump(); 4315 if( cached_idom != computed_idom ) { 4316 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", 4317 computed_idom->_idx, cached_idom->_idx); 4318 } 4319 } 4320 // Dump nodes it controls 4321 for( uint k = 0; k < _nodes.Size(); k++ ) { 4322 // (k < C->unique() && get_ctrl(find(k)) == n) 4323 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { 4324 Node *m = C->root()->find(k); 4325 if( m && m->outcnt() > 0 ) { 4326 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { 4327 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", 4328 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); 4329 } 4330 for( uint j = 0; j < loop->_nest; j++ ) 4331 tty->print(" "); 4332 tty->print(" "); 4333 m->dump(); 4334 } 4335 } 4336 } 4337 } 4338 } 4339 4340 // Collect a R-P-O for the whole CFG. 4341 // Result list is in post-order (scan backwards for RPO) 4342 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { 4343 stk.push(start, 0); 4344 visited.set(start->_idx); 4345 4346 while (stk.is_nonempty()) { 4347 Node* m = stk.node(); 4348 uint idx = stk.index(); 4349 if (idx < m->outcnt()) { 4350 stk.set_index(idx + 1); 4351 Node* n = m->raw_out(idx); 4352 if (n->is_CFG() && !visited.test_set(n->_idx)) { 4353 stk.push(n, 0); 4354 } 4355 } else { 4356 rpo_list.push(m); 4357 stk.pop(); 4358 } 4359 } 4360 } 4361 #endif 4362 4363 4364 //============================================================================= 4365 //------------------------------LoopTreeIterator----------------------------------- 4366 4367 // Advance to next loop tree using a preorder, left-to-right traversal. 4368 void LoopTreeIterator::next() { 4369 assert(!done(), "must not be done."); 4370 if (_curnt->_child != NULL) { 4371 _curnt = _curnt->_child; 4372 } else if (_curnt->_next != NULL) { 4373 _curnt = _curnt->_next; 4374 } else { 4375 while (_curnt != _root && _curnt->_next == NULL) { 4376 _curnt = _curnt->_parent; 4377 } 4378 if (_curnt == _root) { 4379 _curnt = NULL; 4380 assert(done(), "must be done."); 4381 } else { 4382 assert(_curnt->_next != NULL, "must be more to do"); 4383 _curnt = _curnt->_next; 4384 } 4385 } 4386 }