1 /* 2 * Copyright (c) 1998, 2010, 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 "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/connode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/idealGraphPrinter.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/mulnode.hpp" 37 #include "opto/rootnode.hpp" 38 #include "opto/superword.hpp" 39 40 //============================================================================= 41 //------------------------------is_loop_iv------------------------------------- 42 // Determine if a node is Counted loop induction variable. 43 // The method is declared in node.hpp. 44 const Node* Node::is_loop_iv() const { 45 if (this->is_Phi() && !this->as_Phi()->is_copy() && 46 this->as_Phi()->region()->is_CountedLoop() && 47 this->as_Phi()->region()->as_CountedLoop()->phi() == this) { 48 return this; 49 } else { 50 return NULL; 51 } 52 } 53 54 //============================================================================= 55 //------------------------------dump_spec-------------------------------------- 56 // Dump special per-node info 57 #ifndef PRODUCT 58 void LoopNode::dump_spec(outputStream *st) const { 59 if( is_inner_loop () ) st->print( "inner " ); 60 if( is_partial_peel_loop () ) st->print( "partial_peel " ); 61 if( partial_peel_has_failed () ) st->print( "partial_peel_failed " ); 62 } 63 #endif 64 65 //------------------------------get_early_ctrl--------------------------------- 66 // Compute earliest legal control 67 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { 68 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); 69 uint i; 70 Node *early; 71 if( n->in(0) ) { 72 early = n->in(0); 73 if( !early->is_CFG() ) // Might be a non-CFG multi-def 74 early = get_ctrl(early); // So treat input as a straight data input 75 i = 1; 76 } else { 77 early = get_ctrl(n->in(1)); 78 i = 2; 79 } 80 uint e_d = dom_depth(early); 81 assert( early, "" ); 82 for( ; i < n->req(); i++ ) { 83 Node *cin = get_ctrl(n->in(i)); 84 assert( cin, "" ); 85 // Keep deepest dominator depth 86 uint c_d = dom_depth(cin); 87 if( c_d > e_d ) { // Deeper guy? 88 early = cin; // Keep deepest found so far 89 e_d = c_d; 90 } else if( c_d == e_d && // Same depth? 91 early != cin ) { // If not equal, must use slower algorithm 92 // If same depth but not equal, one _must_ dominate the other 93 // and we want the deeper (i.e., dominated) guy. 94 Node *n1 = early; 95 Node *n2 = cin; 96 while( 1 ) { 97 n1 = idom(n1); // Walk up until break cycle 98 n2 = idom(n2); 99 if( n1 == cin || // Walked early up to cin 100 dom_depth(n2) < c_d ) 101 break; // early is deeper; keep him 102 if( n2 == early || // Walked cin up to early 103 dom_depth(n1) < c_d ) { 104 early = cin; // cin is deeper; keep him 105 break; 106 } 107 } 108 e_d = dom_depth(early); // Reset depth register cache 109 } 110 } 111 112 // Return earliest legal location 113 assert(early == find_non_split_ctrl(early), "unexpected early control"); 114 115 return early; 116 } 117 118 //------------------------------set_early_ctrl--------------------------------- 119 // Set earliest legal control 120 void PhaseIdealLoop::set_early_ctrl( Node *n ) { 121 Node *early = get_early_ctrl(n); 122 123 // Record earliest legal location 124 set_ctrl(n, early); 125 } 126 127 //------------------------------set_subtree_ctrl------------------------------- 128 // set missing _ctrl entries on new nodes 129 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { 130 // Already set? Get out. 131 if( _nodes[n->_idx] ) return; 132 // Recursively set _nodes array to indicate where the Node goes 133 uint i; 134 for( i = 0; i < n->req(); ++i ) { 135 Node *m = n->in(i); 136 if( m && m != C->root() ) 137 set_subtree_ctrl( m ); 138 } 139 140 // Fixup self 141 set_early_ctrl( n ); 142 } 143 144 //------------------------------is_counted_loop-------------------------------- 145 Node *PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) { 146 PhaseGVN *gvn = &_igvn; 147 148 // Counted loop head must be a good RegionNode with only 3 not NULL 149 // control input edges: Self, Entry, LoopBack. 150 if ( x->in(LoopNode::Self) == NULL || x->req() != 3 ) 151 return NULL; 152 153 Node *init_control = x->in(LoopNode::EntryControl); 154 Node *back_control = x->in(LoopNode::LoopBackControl); 155 if( init_control == NULL || back_control == NULL ) // Partially dead 156 return NULL; 157 // Must also check for TOP when looking for a dead loop 158 if( init_control->is_top() || back_control->is_top() ) 159 return NULL; 160 161 // Allow funny placement of Safepoint 162 if( back_control->Opcode() == Op_SafePoint ) 163 back_control = back_control->in(TypeFunc::Control); 164 165 // Controlling test for loop 166 Node *iftrue = back_control; 167 uint iftrue_op = iftrue->Opcode(); 168 if( iftrue_op != Op_IfTrue && 169 iftrue_op != Op_IfFalse ) 170 // I have a weird back-control. Probably the loop-exit test is in 171 // the middle of the loop and I am looking at some trailing control-flow 172 // merge point. To fix this I would have to partially peel the loop. 173 return NULL; // Obscure back-control 174 175 // Get boolean guarding loop-back test 176 Node *iff = iftrue->in(0); 177 if( get_loop(iff) != loop || !iff->in(1)->is_Bool() ) return NULL; 178 BoolNode *test = iff->in(1)->as_Bool(); 179 BoolTest::mask bt = test->_test._test; 180 float cl_prob = iff->as_If()->_prob; 181 if( iftrue_op == Op_IfFalse ) { 182 bt = BoolTest(bt).negate(); 183 cl_prob = 1.0 - cl_prob; 184 } 185 // Get backedge compare 186 Node *cmp = test->in(1); 187 int cmp_op = cmp->Opcode(); 188 if( cmp_op != Op_CmpI ) 189 return NULL; // Avoid pointer & float compares 190 191 // Find the trip-counter increment & limit. Limit must be loop invariant. 192 Node *incr = cmp->in(1); 193 Node *limit = cmp->in(2); 194 195 // --------- 196 // need 'loop()' test to tell if limit is loop invariant 197 // --------- 198 199 if( !is_member( loop, get_ctrl(incr) ) ) { // Swapped trip counter and limit? 200 Node *tmp = incr; // Then reverse order into the CmpI 201 incr = limit; 202 limit = tmp; 203 bt = BoolTest(bt).commute(); // And commute the exit test 204 } 205 if( is_member( loop, get_ctrl(limit) ) ) // Limit must loop-invariant 206 return NULL; 207 208 // Trip-counter increment must be commutative & associative. 209 uint incr_op = incr->Opcode(); 210 if( incr_op == Op_Phi && incr->req() == 3 ) { 211 incr = incr->in(2); // Assume incr is on backedge of Phi 212 incr_op = incr->Opcode(); 213 } 214 Node* trunc1 = NULL; 215 Node* trunc2 = NULL; 216 const TypeInt* iv_trunc_t = NULL; 217 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { 218 return NULL; // Funny increment opcode 219 } 220 221 // Get merge point 222 Node *xphi = incr->in(1); 223 Node *stride = incr->in(2); 224 if( !stride->is_Con() ) { // Oops, swap these 225 if( !xphi->is_Con() ) // Is the other guy a constant? 226 return NULL; // Nope, unknown stride, bail out 227 Node *tmp = xphi; // 'incr' is commutative, so ok to swap 228 xphi = stride; 229 stride = tmp; 230 } 231 //if( loop(xphi) != l) return NULL;// Merge point is in inner loop?? 232 if( !xphi->is_Phi() ) return NULL; // Too much math on the trip counter 233 PhiNode *phi = xphi->as_Phi(); 234 235 // Stride must be constant 236 const Type *stride_t = stride->bottom_type(); 237 int stride_con = stride_t->is_int()->get_con(); 238 assert( stride_con, "missed some peephole opt" ); 239 240 // Phi must be of loop header; backedge must wrap to increment 241 if( phi->region() != x ) return NULL; 242 if( trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr || 243 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1 ) { 244 return NULL; 245 } 246 Node *init_trip = phi->in(LoopNode::EntryControl); 247 //if (!init_trip->is_Con()) return NULL; // avoid rolling over MAXINT/MININT 248 249 // If iv trunc type is smaller than int, check for possible wrap. 250 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { 251 assert(trunc1 != NULL, "must have found some truncation"); 252 253 // Get a better type for the phi (filtered thru if's) 254 const TypeInt* phi_ft = filtered_type(phi); 255 256 // Can iv take on a value that will wrap? 257 // 258 // Ensure iv's limit is not within "stride" of the wrap value. 259 // 260 // Example for "short" type 261 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) 262 // If the stride is +10, then the last value of the induction 263 // variable before the increment (phi_ft->_hi) must be 264 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to 265 // ensure no truncation occurs after the increment. 266 267 if (stride_con > 0) { 268 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || 269 iv_trunc_t->_lo > phi_ft->_lo) { 270 return NULL; // truncation may occur 271 } 272 } else if (stride_con < 0) { 273 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || 274 iv_trunc_t->_hi < phi_ft->_hi) { 275 return NULL; // truncation may occur 276 } 277 } 278 // No possibility of wrap so truncation can be discarded 279 // Promote iv type to Int 280 } else { 281 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); 282 } 283 284 // ================================================= 285 // ---- SUCCESS! Found A Trip-Counted Loop! ----- 286 // 287 // Canonicalize the condition on the test. If we can exactly determine 288 // the trip-counter exit value, then set limit to that value and use 289 // a '!=' test. Otherwise use condition '<' for count-up loops and 290 // '>' for count-down loops. If the condition is inverted and we will 291 // be rolling through MININT to MAXINT, then bail out. 292 293 C->print_method("Before CountedLoop", 3); 294 295 // Check for SafePoint on backedge and remove 296 Node *sfpt = x->in(LoopNode::LoopBackControl); 297 if( sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { 298 lazy_replace( sfpt, iftrue ); 299 loop->_tail = iftrue; 300 } 301 302 303 // If compare points to incr, we are ok. Otherwise the compare 304 // can directly point to the phi; in this case adjust the compare so that 305 // it points to the incr by adjusting the limit. 306 if( cmp->in(1) == phi || cmp->in(2) == phi ) 307 limit = gvn->transform(new (C, 3) AddINode(limit,stride)); 308 309 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride. 310 // Final value for iterator should be: trip_count * stride + init_trip. 311 const Type *limit_t = limit->bottom_type(); 312 const Type *init_t = init_trip->bottom_type(); 313 Node *one_p = gvn->intcon( 1); 314 Node *one_m = gvn->intcon(-1); 315 316 Node *trip_count = NULL; 317 Node *hook = new (C, 6) Node(6); 318 switch( bt ) { 319 case BoolTest::eq: 320 return NULL; // Bail out, but this loop trips at most twice! 321 case BoolTest::ne: // Ahh, the case we desire 322 if( stride_con == 1 ) 323 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip)); 324 else if( stride_con == -1 ) 325 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit)); 326 else 327 return NULL; // Odd stride; must prove we hit limit exactly 328 set_subtree_ctrl( trip_count ); 329 //_loop.map(trip_count->_idx,loop(limit)); 330 break; 331 case BoolTest::le: // Maybe convert to '<' case 332 limit = gvn->transform(new (C, 3) AddINode(limit,one_p)); 333 set_subtree_ctrl( limit ); 334 hook->init_req(4, limit); 335 336 bt = BoolTest::lt; 337 // Make the new limit be in the same loop nest as the old limit 338 //_loop.map(limit->_idx,limit_loop); 339 // Fall into next case 340 case BoolTest::lt: { // Maybe convert to '!=' case 341 if( stride_con < 0 ) return NULL; // Count down loop rolls through MAXINT 342 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); 343 set_subtree_ctrl( range ); 344 hook->init_req(0, range); 345 346 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); 347 set_subtree_ctrl( bias ); 348 hook->init_req(1, bias); 349 350 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m)); 351 set_subtree_ctrl( bias1 ); 352 hook->init_req(2, bias1); 353 354 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); 355 set_subtree_ctrl( trip_count ); 356 hook->init_req(3, trip_count); 357 break; 358 } 359 360 case BoolTest::ge: // Maybe convert to '>' case 361 limit = gvn->transform(new (C, 3) AddINode(limit,one_m)); 362 set_subtree_ctrl( limit ); 363 hook->init_req(4 ,limit); 364 365 bt = BoolTest::gt; 366 // Make the new limit be in the same loop nest as the old limit 367 //_loop.map(limit->_idx,limit_loop); 368 // Fall into next case 369 case BoolTest::gt: { // Maybe convert to '!=' case 370 if( stride_con > 0 ) return NULL; // count up loop rolls through MININT 371 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); 372 set_subtree_ctrl( range ); 373 hook->init_req(0, range); 374 375 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); 376 set_subtree_ctrl( bias ); 377 hook->init_req(1, bias); 378 379 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p)); 380 set_subtree_ctrl( bias1 ); 381 hook->init_req(2, bias1); 382 383 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); 384 set_subtree_ctrl( trip_count ); 385 hook->init_req(3, trip_count); 386 break; 387 } 388 } 389 390 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride)); 391 set_subtree_ctrl( span ); 392 hook->init_req(5, span); 393 394 limit = gvn->transform(new (C, 3) AddINode(span,init_trip)); 395 set_subtree_ctrl( limit ); 396 397 // Build a canonical trip test. 398 // Clone code, as old values may be in use. 399 incr = incr->clone(); 400 incr->set_req(1,phi); 401 incr->set_req(2,stride); 402 incr = _igvn.register_new_node_with_optimizer(incr); 403 set_early_ctrl( incr ); 404 _igvn.hash_delete(phi); 405 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); 406 407 // If phi type is more restrictive than Int, raise to 408 // Int to prevent (almost) infinite recursion in igvn 409 // which can only handle integer types for constants or minint..maxint. 410 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { 411 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); 412 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); 413 nphi = _igvn.register_new_node_with_optimizer(nphi); 414 set_ctrl(nphi, get_ctrl(phi)); 415 _igvn.replace_node(phi, nphi); 416 phi = nphi->as_Phi(); 417 } 418 cmp = cmp->clone(); 419 cmp->set_req(1,incr); 420 cmp->set_req(2,limit); 421 cmp = _igvn.register_new_node_with_optimizer(cmp); 422 set_ctrl(cmp, iff->in(0)); 423 424 Node *tmp = test->clone(); 425 assert( tmp->is_Bool(), "" ); 426 test = (BoolNode*)tmp; 427 (*(BoolTest*)&test->_test)._test = bt; //BoolTest::ne; 428 test->set_req(1,cmp); 429 _igvn.register_new_node_with_optimizer(test); 430 set_ctrl(test, iff->in(0)); 431 // If the exit test is dead, STOP! 432 if( test == NULL ) return NULL; 433 _igvn.hash_delete(iff); 434 iff->set_req_X( 1, test, &_igvn ); 435 436 // Replace the old IfNode with a new LoopEndNode 437 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), iff->in(1), cl_prob, iff->as_If()->_fcnt )); 438 IfNode *le = lex->as_If(); 439 uint dd = dom_depth(iff); 440 set_idom(le, le->in(0), dd); // Update dominance for loop exit 441 set_loop(le, loop); 442 443 // Get the loop-exit control 444 Node *if_f = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); 445 446 // Need to swap loop-exit and loop-back control? 447 if( iftrue_op == Op_IfFalse ) { 448 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le)); 449 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le)); 450 451 loop->_tail = back_control = ift2; 452 set_loop(ift2, loop); 453 set_loop(iff2, get_loop(if_f)); 454 455 // Lazy update of 'get_ctrl' mechanism. 456 lazy_replace_proj( if_f , iff2 ); 457 lazy_replace_proj( iftrue, ift2 ); 458 459 // Swap names 460 if_f = iff2; 461 iftrue = ift2; 462 } else { 463 _igvn.hash_delete(if_f ); 464 _igvn.hash_delete(iftrue); 465 if_f ->set_req_X( 0, le, &_igvn ); 466 iftrue->set_req_X( 0, le, &_igvn ); 467 } 468 469 set_idom(iftrue, le, dd+1); 470 set_idom(if_f, le, dd+1); 471 472 // Now setup a new CountedLoopNode to replace the existing LoopNode 473 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control); 474 // The following assert is approximately true, and defines the intention 475 // of can_be_counted_loop. It fails, however, because phase->type 476 // is not yet initialized for this loop and its parts. 477 //assert(l->can_be_counted_loop(this), "sanity"); 478 _igvn.register_new_node_with_optimizer(l); 479 set_loop(l, loop); 480 loop->_head = l; 481 // Fix all data nodes placed at the old loop head. 482 // Uses the lazy-update mechanism of 'get_ctrl'. 483 lazy_replace( x, l ); 484 set_idom(l, init_control, dom_depth(x)); 485 486 // Check for immediately preceding SafePoint and remove 487 Node *sfpt2 = le->in(0); 488 if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) 489 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); 490 491 // Free up intermediate goo 492 _igvn.remove_dead_node(hook); 493 494 C->print_method("After CountedLoop", 3); 495 496 // Return trip counter 497 return trip_count; 498 } 499 500 501 //------------------------------Ideal------------------------------------------ 502 // Return a node which is more "ideal" than the current node. 503 // Attempt to convert into a counted-loop. 504 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 505 if (!can_be_counted_loop(phase)) { 506 phase->C->set_major_progress(); 507 } 508 return RegionNode::Ideal(phase, can_reshape); 509 } 510 511 512 //============================================================================= 513 //------------------------------Ideal------------------------------------------ 514 // Return a node which is more "ideal" than the current node. 515 // Attempt to convert into a counted-loop. 516 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 517 return RegionNode::Ideal(phase, can_reshape); 518 } 519 520 //------------------------------dump_spec-------------------------------------- 521 // Dump special per-node info 522 #ifndef PRODUCT 523 void CountedLoopNode::dump_spec(outputStream *st) const { 524 LoopNode::dump_spec(st); 525 if( stride_is_con() ) { 526 st->print("stride: %d ",stride_con()); 527 } else { 528 st->print("stride: not constant "); 529 } 530 if( is_pre_loop () ) st->print("pre of N%d" , _main_idx ); 531 if( is_main_loop() ) st->print("main of N%d", _idx ); 532 if( is_post_loop() ) st->print("post of N%d", _main_idx ); 533 } 534 #endif 535 536 //============================================================================= 537 int CountedLoopEndNode::stride_con() const { 538 return stride()->bottom_type()->is_int()->get_con(); 539 } 540 541 542 //----------------------match_incr_with_optional_truncation-------------------- 543 // Match increment with optional truncation: 544 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 545 // Return NULL for failure. Success returns the increment node. 546 Node* CountedLoopNode::match_incr_with_optional_truncation( 547 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { 548 // Quick cutouts: 549 if (expr == NULL || expr->req() != 3) return false; 550 551 Node *t1 = NULL; 552 Node *t2 = NULL; 553 const TypeInt* trunc_t = TypeInt::INT; 554 Node* n1 = expr; 555 int n1op = n1->Opcode(); 556 557 // Try to strip (n1 & M) or (n1 << N >> N) from n1. 558 if (n1op == Op_AndI && 559 n1->in(2)->is_Con() && 560 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { 561 // %%% This check should match any mask of 2**K-1. 562 t1 = n1; 563 n1 = t1->in(1); 564 n1op = n1->Opcode(); 565 trunc_t = TypeInt::CHAR; 566 } else if (n1op == Op_RShiftI && 567 n1->in(1) != NULL && 568 n1->in(1)->Opcode() == Op_LShiftI && 569 n1->in(2) == n1->in(1)->in(2) && 570 n1->in(2)->is_Con()) { 571 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); 572 // %%% This check should match any shift in [1..31]. 573 if (shift == 16 || shift == 8) { 574 t1 = n1; 575 t2 = t1->in(1); 576 n1 = t2->in(1); 577 n1op = n1->Opcode(); 578 if (shift == 16) { 579 trunc_t = TypeInt::SHORT; 580 } else if (shift == 8) { 581 trunc_t = TypeInt::BYTE; 582 } 583 } 584 } 585 586 // If (maybe after stripping) it is an AddI, we won: 587 if (n1op == Op_AddI) { 588 *trunc1 = t1; 589 *trunc2 = t2; 590 *trunc_type = trunc_t; 591 return n1; 592 } 593 594 // failed 595 return NULL; 596 } 597 598 599 //------------------------------filtered_type-------------------------------- 600 // Return a type based on condition control flow 601 // A successful return will be a type that is restricted due 602 // to a series of dominating if-tests, such as: 603 // if (i < 10) { 604 // if (i > 0) { 605 // here: "i" type is [1..10) 606 // } 607 // } 608 // or a control flow merge 609 // if (i < 10) { 610 // do { 611 // phi( , ) -- at top of loop type is [min_int..10) 612 // i = ? 613 // } while ( i < 10) 614 // 615 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { 616 assert(n && n->bottom_type()->is_int(), "must be int"); 617 const TypeInt* filtered_t = NULL; 618 if (!n->is_Phi()) { 619 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); 620 filtered_t = filtered_type_from_dominators(n, n_ctrl); 621 622 } else { 623 Node* phi = n->as_Phi(); 624 Node* region = phi->in(0); 625 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); 626 if (region && region != C->top()) { 627 for (uint i = 1; i < phi->req(); i++) { 628 Node* val = phi->in(i); 629 Node* use_c = region->in(i); 630 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); 631 if (val_t != NULL) { 632 if (filtered_t == NULL) { 633 filtered_t = val_t; 634 } else { 635 filtered_t = filtered_t->meet(val_t)->is_int(); 636 } 637 } 638 } 639 } 640 } 641 const TypeInt* n_t = _igvn.type(n)->is_int(); 642 if (filtered_t != NULL) { 643 n_t = n_t->join(filtered_t)->is_int(); 644 } 645 return n_t; 646 } 647 648 649 //------------------------------filtered_type_from_dominators-------------------------------- 650 // Return a possibly more restrictive type for val based on condition control flow of dominators 651 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { 652 if (val->is_Con()) { 653 return val->bottom_type()->is_int(); 654 } 655 uint if_limit = 10; // Max number of dominating if's visited 656 const TypeInt* rtn_t = NULL; 657 658 if (use_ctrl && use_ctrl != C->top()) { 659 Node* val_ctrl = get_ctrl(val); 660 uint val_dom_depth = dom_depth(val_ctrl); 661 Node* pred = use_ctrl; 662 uint if_cnt = 0; 663 while (if_cnt < if_limit) { 664 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { 665 if_cnt++; 666 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); 667 if (if_t != NULL) { 668 if (rtn_t == NULL) { 669 rtn_t = if_t; 670 } else { 671 rtn_t = rtn_t->join(if_t)->is_int(); 672 } 673 } 674 } 675 pred = idom(pred); 676 if (pred == NULL || pred == C->top()) { 677 break; 678 } 679 // Stop if going beyond definition block of val 680 if (dom_depth(pred) < val_dom_depth) { 681 break; 682 } 683 } 684 } 685 return rtn_t; 686 } 687 688 689 //------------------------------dump_spec-------------------------------------- 690 // Dump special per-node info 691 #ifndef PRODUCT 692 void CountedLoopEndNode::dump_spec(outputStream *st) const { 693 if( in(TestValue)->is_Bool() ) { 694 BoolTest bt( test_trip()); // Added this for g++. 695 696 st->print("["); 697 bt.dump_on(st); 698 st->print("]"); 699 } 700 st->print(" "); 701 IfNode::dump_spec(st); 702 } 703 #endif 704 705 //============================================================================= 706 //------------------------------is_member-------------------------------------- 707 // Is 'l' a member of 'this'? 708 int IdealLoopTree::is_member( const IdealLoopTree *l ) const { 709 while( l->_nest > _nest ) l = l->_parent; 710 return l == this; 711 } 712 713 //------------------------------set_nest--------------------------------------- 714 // Set loop tree nesting depth. Accumulate _has_call bits. 715 int IdealLoopTree::set_nest( uint depth ) { 716 _nest = depth; 717 int bits = _has_call; 718 if( _child ) bits |= _child->set_nest(depth+1); 719 if( bits ) _has_call = 1; 720 if( _next ) bits |= _next ->set_nest(depth ); 721 return bits; 722 } 723 724 //------------------------------split_fall_in---------------------------------- 725 // Split out multiple fall-in edges from the loop header. Move them to a 726 // private RegionNode before the loop. This becomes the loop landing pad. 727 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { 728 PhaseIterGVN &igvn = phase->_igvn; 729 uint i; 730 731 // Make a new RegionNode to be the landing pad. 732 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 ); 733 phase->set_loop(landing_pad,_parent); 734 // Gather all the fall-in control paths into the landing pad 735 uint icnt = fall_in_cnt; 736 uint oreq = _head->req(); 737 for( i = oreq-1; i>0; i-- ) 738 if( !phase->is_member( this, _head->in(i) ) ) 739 landing_pad->set_req(icnt--,_head->in(i)); 740 741 // Peel off PhiNode edges as well 742 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 743 Node *oj = _head->fast_out(j); 744 if( oj->is_Phi() ) { 745 PhiNode* old_phi = oj->as_Phi(); 746 assert( old_phi->region() == _head, "" ); 747 igvn.hash_delete(old_phi); // Yank from hash before hacking edges 748 Node *p = PhiNode::make_blank(landing_pad, old_phi); 749 uint icnt = fall_in_cnt; 750 for( i = oreq-1; i>0; i-- ) { 751 if( !phase->is_member( this, _head->in(i) ) ) { 752 p->init_req(icnt--, old_phi->in(i)); 753 // Go ahead and clean out old edges from old phi 754 old_phi->del_req(i); 755 } 756 } 757 // Search for CSE's here, because ZKM.jar does a lot of 758 // loop hackery and we need to be a little incremental 759 // with the CSE to avoid O(N^2) node blow-up. 760 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE 761 if( p2 ) { // Found CSE 762 p->destruct(); // Recover useless new node 763 p = p2; // Use old node 764 } else { 765 igvn.register_new_node_with_optimizer(p, old_phi); 766 } 767 // Make old Phi refer to new Phi. 768 old_phi->add_req(p); 769 // Check for the special case of making the old phi useless and 770 // disappear it. In JavaGrande I have a case where this useless 771 // Phi is the loop limit and prevents recognizing a CountedLoop 772 // which in turn prevents removing an empty loop. 773 Node *id_old_phi = old_phi->Identity( &igvn ); 774 if( id_old_phi != old_phi ) { // Found a simple identity? 775 // Note that I cannot call 'replace_node' here, because 776 // that will yank the edge from old_phi to the Region and 777 // I'm mid-iteration over the Region's uses. 778 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { 779 Node* use = old_phi->last_out(i); 780 igvn.hash_delete(use); 781 igvn._worklist.push(use); 782 uint uses_found = 0; 783 for (uint j = 0; j < use->len(); j++) { 784 if (use->in(j) == old_phi) { 785 if (j < use->req()) use->set_req (j, id_old_phi); 786 else use->set_prec(j, id_old_phi); 787 uses_found++; 788 } 789 } 790 i -= uses_found; // we deleted 1 or more copies of this edge 791 } 792 } 793 igvn._worklist.push(old_phi); 794 } 795 } 796 // Finally clean out the fall-in edges from the RegionNode 797 for( i = oreq-1; i>0; i-- ) { 798 if( !phase->is_member( this, _head->in(i) ) ) { 799 _head->del_req(i); 800 } 801 } 802 // Transform landing pad 803 igvn.register_new_node_with_optimizer(landing_pad, _head); 804 // Insert landing pad into the header 805 _head->add_req(landing_pad); 806 } 807 808 //------------------------------split_outer_loop------------------------------- 809 // Split out the outermost loop from this shared header. 810 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { 811 PhaseIterGVN &igvn = phase->_igvn; 812 813 // Find index of outermost loop; it should also be my tail. 814 uint outer_idx = 1; 815 while( _head->in(outer_idx) != _tail ) outer_idx++; 816 817 // Make a LoopNode for the outermost loop. 818 Node *ctl = _head->in(LoopNode::EntryControl); 819 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) ); 820 outer = igvn.register_new_node_with_optimizer(outer, _head); 821 phase->set_created_loop_node(); 822 // Outermost loop falls into '_head' loop 823 _head->set_req(LoopNode::EntryControl, outer); 824 _head->del_req(outer_idx); 825 // Split all the Phis up between '_head' loop and 'outer' loop. 826 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 827 Node *out = _head->fast_out(j); 828 if( out->is_Phi() ) { 829 PhiNode *old_phi = out->as_Phi(); 830 assert( old_phi->region() == _head, "" ); 831 Node *phi = PhiNode::make_blank(outer, old_phi); 832 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); 833 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); 834 phi = igvn.register_new_node_with_optimizer(phi, old_phi); 835 // Make old Phi point to new Phi on the fall-in path 836 igvn.hash_delete(old_phi); 837 old_phi->set_req(LoopNode::EntryControl, phi); 838 old_phi->del_req(outer_idx); 839 igvn._worklist.push(old_phi); 840 } 841 } 842 843 // Use the new loop head instead of the old shared one 844 _head = outer; 845 phase->set_loop(_head, this); 846 } 847 848 //------------------------------fix_parent------------------------------------- 849 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { 850 loop->_parent = parent; 851 if( loop->_child ) fix_parent( loop->_child, loop ); 852 if( loop->_next ) fix_parent( loop->_next , parent ); 853 } 854 855 //------------------------------estimate_path_freq----------------------------- 856 static float estimate_path_freq( Node *n ) { 857 // Try to extract some path frequency info 858 IfNode *iff; 859 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests 860 uint nop = n->Opcode(); 861 if( nop == Op_SafePoint ) { // Skip any safepoint 862 n = n->in(0); 863 continue; 864 } 865 if( nop == Op_CatchProj ) { // Get count from a prior call 866 // Assume call does not always throw exceptions: means the call-site 867 // count is also the frequency of the fall-through path. 868 assert( n->is_CatchProj(), "" ); 869 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) 870 return 0.0f; // Assume call exception path is rare 871 Node *call = n->in(0)->in(0)->in(0); 872 assert( call->is_Call(), "expect a call here" ); 873 const JVMState *jvms = ((CallNode*)call)->jvms(); 874 ciMethodData* methodData = jvms->method()->method_data(); 875 if (!methodData->is_mature()) return 0.0f; // No call-site data 876 ciProfileData* data = methodData->bci_to_data(jvms->bci()); 877 if ((data == NULL) || !data->is_CounterData()) { 878 // no call profile available, try call's control input 879 n = n->in(0); 880 continue; 881 } 882 return data->as_CounterData()->count()/FreqCountInvocations; 883 } 884 // See if there's a gating IF test 885 Node *n_c = n->in(0); 886 if( !n_c->is_If() ) break; // No estimate available 887 iff = n_c->as_If(); 888 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? 889 // Compute how much count comes on this path 890 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; 891 // Have no count info. Skip dull uncommon-trap like branches. 892 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || 893 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) 894 break; 895 // Skip through never-taken branch; look for a real loop exit. 896 n = iff->in(0); 897 } 898 return 0.0f; // No estimate available 899 } 900 901 //------------------------------merge_many_backedges--------------------------- 902 // Merge all the backedges from the shared header into a private Region. 903 // Feed that region as the one backedge to this loop. 904 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { 905 uint i; 906 907 // Scan for the top 2 hottest backedges 908 float hotcnt = 0.0f; 909 float warmcnt = 0.0f; 910 uint hot_idx = 0; 911 // Loop starts at 2 because slot 1 is the fall-in path 912 for( i = 2; i < _head->req(); i++ ) { 913 float cnt = estimate_path_freq(_head->in(i)); 914 if( cnt > hotcnt ) { // Grab hottest path 915 warmcnt = hotcnt; 916 hotcnt = cnt; 917 hot_idx = i; 918 } else if( cnt > warmcnt ) { // And 2nd hottest path 919 warmcnt = cnt; 920 } 921 } 922 923 // See if the hottest backedge is worthy of being an inner loop 924 // by being much hotter than the next hottest backedge. 925 if( hotcnt <= 0.0001 || 926 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge 927 928 // Peel out the backedges into a private merge point; peel 929 // them all except optionally hot_idx. 930 PhaseIterGVN &igvn = phase->_igvn; 931 932 Node *hot_tail = NULL; 933 // Make a Region for the merge point 934 Node *r = new (phase->C, 1) RegionNode(1); 935 for( i = 2; i < _head->req(); i++ ) { 936 if( i != hot_idx ) 937 r->add_req( _head->in(i) ); 938 else hot_tail = _head->in(i); 939 } 940 igvn.register_new_node_with_optimizer(r, _head); 941 // Plug region into end of loop _head, followed by hot_tail 942 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); 943 _head->set_req(2, r); 944 if( hot_idx ) _head->add_req(hot_tail); 945 946 // Split all the Phis up between '_head' loop and the Region 'r' 947 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 948 Node *out = _head->fast_out(j); 949 if( out->is_Phi() ) { 950 PhiNode* n = out->as_Phi(); 951 igvn.hash_delete(n); // Delete from hash before hacking edges 952 Node *hot_phi = NULL; 953 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type()); 954 // Check all inputs for the ones to peel out 955 uint j = 1; 956 for( uint i = 2; i < n->req(); i++ ) { 957 if( i != hot_idx ) 958 phi->set_req( j++, n->in(i) ); 959 else hot_phi = n->in(i); 960 } 961 // Register the phi but do not transform until whole place transforms 962 igvn.register_new_node_with_optimizer(phi, n); 963 // Add the merge phi to the old Phi 964 while( n->req() > 3 ) n->del_req( n->req()-1 ); 965 n->set_req(2, phi); 966 if( hot_idx ) n->add_req(hot_phi); 967 } 968 } 969 970 971 // Insert a new IdealLoopTree inserted below me. Turn it into a clone 972 // of self loop tree. Turn self into a loop headed by _head and with 973 // tail being the new merge point. 974 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); 975 phase->set_loop(_tail,ilt); // Adjust tail 976 _tail = r; // Self's tail is new merge point 977 phase->set_loop(r,this); 978 ilt->_child = _child; // New guy has my children 979 _child = ilt; // Self has new guy as only child 980 ilt->_parent = this; // new guy has self for parent 981 ilt->_nest = _nest; // Same nesting depth (for now) 982 983 // Starting with 'ilt', look for child loop trees using the same shared 984 // header. Flatten these out; they will no longer be loops in the end. 985 IdealLoopTree **pilt = &_child; 986 while( ilt ) { 987 if( ilt->_head == _head ) { 988 uint i; 989 for( i = 2; i < _head->req(); i++ ) 990 if( _head->in(i) == ilt->_tail ) 991 break; // Still a loop 992 if( i == _head->req() ) { // No longer a loop 993 // Flatten ilt. Hang ilt's "_next" list from the end of 994 // ilt's '_child' list. Move the ilt's _child up to replace ilt. 995 IdealLoopTree **cp = &ilt->_child; 996 while( *cp ) cp = &(*cp)->_next; // Find end of child list 997 *cp = ilt->_next; // Hang next list at end of child list 998 *pilt = ilt->_child; // Move child up to replace ilt 999 ilt->_head = NULL; // Flag as a loop UNIONED into parent 1000 ilt = ilt->_child; // Repeat using new ilt 1001 continue; // do not advance over ilt->_child 1002 } 1003 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); 1004 phase->set_loop(_head,ilt); 1005 } 1006 pilt = &ilt->_child; // Advance to next 1007 ilt = *pilt; 1008 } 1009 1010 if( _child ) fix_parent( _child, this ); 1011 } 1012 1013 //------------------------------beautify_loops--------------------------------- 1014 // Split shared headers and insert loop landing pads. 1015 // Insert a LoopNode to replace the RegionNode. 1016 // Return TRUE if loop tree is structurally changed. 1017 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { 1018 bool result = false; 1019 // Cache parts in locals for easy 1020 PhaseIterGVN &igvn = phase->_igvn; 1021 1022 phase->C->print_method("Before beautify loops", 3); 1023 1024 igvn.hash_delete(_head); // Yank from hash before hacking edges 1025 1026 // Check for multiple fall-in paths. Peel off a landing pad if need be. 1027 int fall_in_cnt = 0; 1028 for( uint i = 1; i < _head->req(); i++ ) 1029 if( !phase->is_member( this, _head->in(i) ) ) 1030 fall_in_cnt++; 1031 assert( fall_in_cnt, "at least 1 fall-in path" ); 1032 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins 1033 split_fall_in( phase, fall_in_cnt ); 1034 1035 // Swap inputs to the _head and all Phis to move the fall-in edge to 1036 // the left. 1037 fall_in_cnt = 1; 1038 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) 1039 fall_in_cnt++; 1040 if( fall_in_cnt > 1 ) { 1041 // Since I am just swapping inputs I do not need to update def-use info 1042 Node *tmp = _head->in(1); 1043 _head->set_req( 1, _head->in(fall_in_cnt) ); 1044 _head->set_req( fall_in_cnt, tmp ); 1045 // Swap also all Phis 1046 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { 1047 Node* phi = _head->fast_out(i); 1048 if( phi->is_Phi() ) { 1049 igvn.hash_delete(phi); // Yank from hash before hacking edges 1050 tmp = phi->in(1); 1051 phi->set_req( 1, phi->in(fall_in_cnt) ); 1052 phi->set_req( fall_in_cnt, tmp ); 1053 } 1054 } 1055 } 1056 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); 1057 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); 1058 1059 // If I am a shared header (multiple backedges), peel off the many 1060 // backedges into a private merge point and use the merge point as 1061 // the one true backedge. 1062 if( _head->req() > 3 ) { 1063 // Merge the many backedges into a single backedge. 1064 merge_many_backedges( phase ); 1065 result = true; 1066 } 1067 1068 // If I am a shared header (multiple backedges), peel off myself loop. 1069 // I better be the outermost loop. 1070 if( _head->req() > 3 ) { 1071 split_outer_loop( phase ); 1072 result = true; 1073 1074 } else if( !_head->is_Loop() && !_irreducible ) { 1075 // Make a new LoopNode to replace the old loop head 1076 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) ); 1077 l = igvn.register_new_node_with_optimizer(l, _head); 1078 phase->set_created_loop_node(); 1079 // Go ahead and replace _head 1080 phase->_igvn.replace_node( _head, l ); 1081 _head = l; 1082 phase->set_loop(_head, this); 1083 } 1084 1085 // Now recursively beautify nested loops 1086 if( _child ) result |= _child->beautify_loops( phase ); 1087 if( _next ) result |= _next ->beautify_loops( phase ); 1088 return result; 1089 } 1090 1091 //------------------------------allpaths_check_safepts---------------------------- 1092 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 1093 // encountered. Helper for check_safepts. 1094 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { 1095 assert(stack.size() == 0, "empty stack"); 1096 stack.push(_tail); 1097 visited.Clear(); 1098 visited.set(_tail->_idx); 1099 while (stack.size() > 0) { 1100 Node* n = stack.pop(); 1101 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 1102 // Terminate this path 1103 } else if (n->Opcode() == Op_SafePoint) { 1104 if (_phase->get_loop(n) != this) { 1105 if (_required_safept == NULL) _required_safept = new Node_List(); 1106 _required_safept->push(n); // save the one closest to the tail 1107 } 1108 // Terminate this path 1109 } else { 1110 uint start = n->is_Region() ? 1 : 0; 1111 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; 1112 for (uint i = start; i < end; i++) { 1113 Node* in = n->in(i); 1114 assert(in->is_CFG(), "must be"); 1115 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { 1116 stack.push(in); 1117 } 1118 } 1119 } 1120 } 1121 } 1122 1123 //------------------------------check_safepts---------------------------- 1124 // Given dominators, try to find loops with calls that must always be 1125 // executed (call dominates loop tail). These loops do not need non-call 1126 // safepoints (ncsfpt). 1127 // 1128 // A complication is that a safepoint in a inner loop may be needed 1129 // by an outer loop. In the following, the inner loop sees it has a 1130 // call (block 3) on every path from the head (block 2) to the 1131 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) 1132 // in block 2, _but_ this leaves the outer loop without a safepoint. 1133 // 1134 // entry 0 1135 // | 1136 // v 1137 // outer 1,2 +->1 1138 // | | 1139 // | v 1140 // | 2<---+ ncsfpt in 2 1141 // |_/|\ | 1142 // | v | 1143 // inner 2,3 / 3 | call in 3 1144 // / | | 1145 // v +--+ 1146 // exit 4 1147 // 1148 // 1149 // This method creates a list (_required_safept) of ncsfpt nodes that must 1150 // be protected is created for each loop. When a ncsfpt maybe deleted, it 1151 // is first looked for in the lists for the outer loops of the current loop. 1152 // 1153 // The insights into the problem: 1154 // A) counted loops are okay 1155 // B) innermost loops are okay (only an inner loop can delete 1156 // a ncsfpt needed by an outer loop) 1157 // C) a loop is immune from an inner loop deleting a safepoint 1158 // if the loop has a call on the idom-path 1159 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the 1160 // idom-path that is not in a nested loop 1161 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner 1162 // loop needs to be prevented from deletion by an inner loop 1163 // 1164 // There are two analyses: 1165 // 1) The first, and cheaper one, scans the loop body from 1166 // tail to head following the idom (immediate dominator) 1167 // chain, looking for the cases (C,D,E) above. 1168 // Since inner loops are scanned before outer loops, there is summary 1169 // information about inner loops. Inner loops can be skipped over 1170 // when the tail of an inner loop is encountered. 1171 // 1172 // 2) The second, invoked if the first fails to find a call or ncsfpt on 1173 // the idom path (which is rare), scans all predecessor control paths 1174 // from the tail to the head, terminating a path when a call or sfpt 1175 // is encountered, to find the ncsfpt's that are closest to the tail. 1176 // 1177 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { 1178 // Bottom up traversal 1179 IdealLoopTree* ch = _child; 1180 while (ch != NULL) { 1181 ch->check_safepts(visited, stack); 1182 ch = ch->_next; 1183 } 1184 1185 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { 1186 bool has_call = false; // call on dom-path 1187 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth 1188 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth 1189 // Scan the dom-path nodes from tail to head 1190 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { 1191 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 1192 has_call = true; 1193 _has_sfpt = 1; // Then no need for a safept! 1194 break; 1195 } else if (n->Opcode() == Op_SafePoint) { 1196 if (_phase->get_loop(n) == this) { 1197 has_local_ncsfpt = true; 1198 break; 1199 } 1200 if (nonlocal_ncsfpt == NULL) { 1201 nonlocal_ncsfpt = n; // save the one closest to the tail 1202 } 1203 } else { 1204 IdealLoopTree* nlpt = _phase->get_loop(n); 1205 if (this != nlpt) { 1206 // If at an inner loop tail, see if the inner loop has already 1207 // recorded seeing a call on the dom-path (and stop.) If not, 1208 // jump to the head of the inner loop. 1209 assert(is_member(nlpt), "nested loop"); 1210 Node* tail = nlpt->_tail; 1211 if (tail->in(0)->is_If()) tail = tail->in(0); 1212 if (n == tail) { 1213 // If inner loop has call on dom-path, so does outer loop 1214 if (nlpt->_has_sfpt) { 1215 has_call = true; 1216 _has_sfpt = 1; 1217 break; 1218 } 1219 // Skip to head of inner loop 1220 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); 1221 n = nlpt->_head; 1222 } 1223 } 1224 } 1225 } 1226 // Record safept's that this loop needs preserved when an 1227 // inner loop attempts to delete it's safepoints. 1228 if (_child != NULL && !has_call && !has_local_ncsfpt) { 1229 if (nonlocal_ncsfpt != NULL) { 1230 if (_required_safept == NULL) _required_safept = new Node_List(); 1231 _required_safept->push(nonlocal_ncsfpt); 1232 } else { 1233 // Failed to find a suitable safept on the dom-path. Now use 1234 // an all paths walk from tail to head, looking for safepoints to preserve. 1235 allpaths_check_safepts(visited, stack); 1236 } 1237 } 1238 } 1239 } 1240 1241 //---------------------------is_deleteable_safept---------------------------- 1242 // Is safept not required by an outer loop? 1243 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { 1244 assert(sfpt->Opcode() == Op_SafePoint, ""); 1245 IdealLoopTree* lp = get_loop(sfpt)->_parent; 1246 while (lp != NULL) { 1247 Node_List* sfpts = lp->_required_safept; 1248 if (sfpts != NULL) { 1249 for (uint i = 0; i < sfpts->size(); i++) { 1250 if (sfpt == sfpts->at(i)) 1251 return false; 1252 } 1253 } 1254 lp = lp->_parent; 1255 } 1256 return true; 1257 } 1258 1259 //------------------------------counted_loop----------------------------------- 1260 // Convert to counted loops where possible 1261 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { 1262 1263 // For grins, set the inner-loop flag here 1264 if( !_child ) { 1265 if( _head->is_Loop() ) _head->as_Loop()->set_inner_loop(); 1266 } 1267 1268 if( _head->is_CountedLoop() || 1269 phase->is_counted_loop( _head, this ) ) { 1270 _has_sfpt = 1; // Indicate we do not need a safepoint here 1271 1272 // Look for a safepoint to remove 1273 for (Node* n = tail(); n != _head; n = phase->idom(n)) 1274 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this && 1275 phase->is_deleteable_safept(n)) 1276 phase->lazy_replace(n,n->in(TypeFunc::Control)); 1277 1278 CountedLoopNode *cl = _head->as_CountedLoop(); 1279 Node *incr = cl->incr(); 1280 if( !incr ) return; // Dead loop? 1281 Node *init = cl->init_trip(); 1282 Node *phi = cl->phi(); 1283 // protect against stride not being a constant 1284 if( !cl->stride_is_con() ) return; 1285 int stride_con = cl->stride_con(); 1286 1287 // Look for induction variables 1288 1289 // Visit all children, looking for Phis 1290 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { 1291 Node *out = cl->out(i); 1292 // Look for other phis (secondary IVs). Skip dead ones 1293 if (!out->is_Phi() || out == phi || !phase->has_node(out)) continue; 1294 PhiNode* phi2 = out->as_Phi(); 1295 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); 1296 // Look for induction variables of the form: X += constant 1297 if( phi2->region() != _head || 1298 incr2->req() != 3 || 1299 incr2->in(1) != phi2 || 1300 incr2 == incr || 1301 incr2->Opcode() != Op_AddI || 1302 !incr2->in(2)->is_Con() ) 1303 continue; 1304 1305 // Check for parallel induction variable (parallel to trip counter) 1306 // via an affine function. In particular, count-down loops with 1307 // count-up array indices are common. We only RCE references off 1308 // the trip-counter, so we need to convert all these to trip-counter 1309 // expressions. 1310 Node *init2 = phi2->in( LoopNode::EntryControl ); 1311 int stride_con2 = incr2->in(2)->get_int(); 1312 1313 // The general case here gets a little tricky. We want to find the 1314 // GCD of all possible parallel IV's and make a new IV using this 1315 // GCD for the loop. Then all possible IVs are simple multiples of 1316 // the GCD. In practice, this will cover very few extra loops. 1317 // Instead we require 'stride_con2' to be a multiple of 'stride_con', 1318 // where +/-1 is the common case, but other integer multiples are 1319 // also easy to handle. 1320 int ratio_con = stride_con2/stride_con; 1321 1322 if( ratio_con * stride_con == stride_con2 ) { // Check for exact 1323 // Convert to using the trip counter. The parallel induction 1324 // variable differs from the trip counter by a loop-invariant 1325 // amount, the difference between their respective initial values. 1326 // It is scaled by the 'ratio_con'. 1327 Compile* C = phase->C; 1328 Node* ratio = phase->_igvn.intcon(ratio_con); 1329 phase->set_ctrl(ratio, C->root()); 1330 Node* ratio_init = new (C, 3) MulINode(init, ratio); 1331 phase->_igvn.register_new_node_with_optimizer(ratio_init, init); 1332 phase->set_early_ctrl(ratio_init); 1333 Node* diff = new (C, 3) SubINode(init2, ratio_init); 1334 phase->_igvn.register_new_node_with_optimizer(diff, init2); 1335 phase->set_early_ctrl(diff); 1336 Node* ratio_idx = new (C, 3) MulINode(phi, ratio); 1337 phase->_igvn.register_new_node_with_optimizer(ratio_idx, phi); 1338 phase->set_ctrl(ratio_idx, cl); 1339 Node* add = new (C, 3) AddINode(ratio_idx, diff); 1340 phase->_igvn.register_new_node_with_optimizer(add); 1341 phase->set_ctrl(add, cl); 1342 phase->_igvn.replace_node( phi2, add ); 1343 // Sometimes an induction variable is unused 1344 if (add->outcnt() == 0) { 1345 phase->_igvn.remove_dead_node(add); 1346 } 1347 --i; // deleted this phi; rescan starting with next position 1348 continue; 1349 } 1350 } 1351 } else if (_parent != NULL && !_irreducible) { 1352 // Not a counted loop. 1353 // Look for a safepoint on the idom-path to remove, preserving the first one 1354 bool found = false; 1355 Node* n = tail(); 1356 for (; n != _head && !found; n = phase->idom(n)) { 1357 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this) 1358 found = true; // Found one 1359 } 1360 // Skip past it and delete the others 1361 for (; n != _head; n = phase->idom(n)) { 1362 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this && 1363 phase->is_deleteable_safept(n)) 1364 phase->lazy_replace(n,n->in(TypeFunc::Control)); 1365 } 1366 } 1367 1368 // Recursively 1369 if( _child ) _child->counted_loop( phase ); 1370 if( _next ) _next ->counted_loop( phase ); 1371 } 1372 1373 #ifndef PRODUCT 1374 //------------------------------dump_head-------------------------------------- 1375 // Dump 1 liner for loop header info 1376 void IdealLoopTree::dump_head( ) const { 1377 for( uint i=0; i<_nest; i++ ) 1378 tty->print(" "); 1379 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); 1380 if( _irreducible ) tty->print(" IRREDUCIBLE"); 1381 if( _head->is_CountedLoop() ) { 1382 CountedLoopNode *cl = _head->as_CountedLoop(); 1383 tty->print(" counted"); 1384 if( cl->is_pre_loop () ) tty->print(" pre" ); 1385 if( cl->is_main_loop() ) tty->print(" main"); 1386 if( cl->is_post_loop() ) tty->print(" post"); 1387 } 1388 tty->cr(); 1389 } 1390 1391 //------------------------------dump------------------------------------------- 1392 // Dump loops by loop tree 1393 void IdealLoopTree::dump( ) const { 1394 dump_head(); 1395 if( _child ) _child->dump(); 1396 if( _next ) _next ->dump(); 1397 } 1398 1399 #endif 1400 1401 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { 1402 if (loop == root) { 1403 if (loop->_child != NULL) { 1404 log->begin_head("loop_tree"); 1405 log->end_head(); 1406 if( loop->_child ) log_loop_tree(root, loop->_child, log); 1407 log->tail("loop_tree"); 1408 assert(loop->_next == NULL, "what?"); 1409 } 1410 } else { 1411 Node* head = loop->_head; 1412 log->begin_head("loop"); 1413 log->print(" idx='%d' ", head->_idx); 1414 if (loop->_irreducible) log->print("irreducible='1' "); 1415 if (head->is_Loop()) { 1416 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); 1417 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); 1418 } 1419 if (head->is_CountedLoop()) { 1420 CountedLoopNode* cl = head->as_CountedLoop(); 1421 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); 1422 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); 1423 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); 1424 } 1425 log->end_head(); 1426 if( loop->_child ) log_loop_tree(root, loop->_child, log); 1427 log->tail("loop"); 1428 if( loop->_next ) log_loop_tree(root, loop->_next, log); 1429 } 1430 } 1431 1432 //---------------------collect_potentially_useful_predicates----------------------- 1433 // Helper function to collect potentially useful predicates to prevent them from 1434 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates 1435 void PhaseIdealLoop::collect_potentially_useful_predicates( 1436 IdealLoopTree * loop, Unique_Node_List &useful_predicates) { 1437 if (loop->_child) { // child 1438 collect_potentially_useful_predicates(loop->_child, useful_predicates); 1439 } 1440 1441 // self (only loops that we can apply loop predication may use their predicates) 1442 if (loop->_head->is_Loop() && 1443 !loop->_irreducible && 1444 !loop->tail()->is_top()) { 1445 LoopNode *lpn = loop->_head->as_Loop(); 1446 Node* entry = lpn->in(LoopNode::EntryControl); 1447 ProjNode *predicate_proj = find_predicate_insertion_point(entry); 1448 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication 1449 assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); 1450 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 1451 } 1452 } 1453 1454 if ( loop->_next ) { // sibling 1455 collect_potentially_useful_predicates(loop->_next, useful_predicates); 1456 } 1457 } 1458 1459 //------------------------eliminate_useless_predicates----------------------------- 1460 // Eliminate all inserted predicates if they could not be used by loop predication. 1461 void PhaseIdealLoop::eliminate_useless_predicates() { 1462 if (C->predicate_count() == 0) return; // no predicate left 1463 1464 Unique_Node_List useful_predicates; // to store useful predicates 1465 if (C->has_loops()) { 1466 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); 1467 } 1468 1469 for (int i = C->predicate_count(); i > 0; i--) { 1470 Node * n = C->predicate_opaque1_node(i-1); 1471 assert(n->Opcode() == Op_Opaque1, "must be"); 1472 if (!useful_predicates.member(n)) { // not in the useful list 1473 _igvn.replace_node(n, n->in(1)); 1474 } 1475 } 1476 } 1477 1478 //============================================================================= 1479 //----------------------------build_and_optimize------------------------------- 1480 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to 1481 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. 1482 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool do_loop_pred) { 1483 ResourceMark rm; 1484 1485 int old_progress = C->major_progress(); 1486 1487 // Reset major-progress flag for the driver's heuristics 1488 C->clear_major_progress(); 1489 1490 #ifndef PRODUCT 1491 // Capture for later assert 1492 uint unique = C->unique(); 1493 _loop_invokes++; 1494 _loop_work += unique; 1495 #endif 1496 1497 // True if the method has at least 1 irreducible loop 1498 _has_irreducible_loops = false; 1499 1500 _created_loop_node = false; 1501 1502 Arena *a = Thread::current()->resource_area(); 1503 VectorSet visited(a); 1504 // Pre-grow the mapping from Nodes to IdealLoopTrees. 1505 _nodes.map(C->unique(), NULL); 1506 memset(_nodes.adr(), 0, wordSize * C->unique()); 1507 1508 // Pre-build the top-level outermost loop tree entry 1509 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); 1510 // Do not need a safepoint at the top level 1511 _ltree_root->_has_sfpt = 1; 1512 1513 // Empty pre-order array 1514 allocate_preorders(); 1515 1516 // Build a loop tree on the fly. Build a mapping from CFG nodes to 1517 // IdealLoopTree entries. Data nodes are NOT walked. 1518 build_loop_tree(); 1519 // Check for bailout, and return 1520 if (C->failing()) { 1521 return; 1522 } 1523 1524 // No loops after all 1525 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); 1526 1527 // There should always be an outer loop containing the Root and Return nodes. 1528 // If not, we have a degenerate empty program. Bail out in this case. 1529 if (!has_node(C->root())) { 1530 if (!_verify_only) { 1531 C->clear_major_progress(); 1532 C->record_method_not_compilable("empty program detected during loop optimization"); 1533 } 1534 return; 1535 } 1536 1537 // Nothing to do, so get out 1538 if( !C->has_loops() && !do_split_ifs && !_verify_me && !_verify_only ) { 1539 _igvn.optimize(); // Cleanup NeverBranches 1540 return; 1541 } 1542 1543 // Set loop nesting depth 1544 _ltree_root->set_nest( 0 ); 1545 1546 // Split shared headers and insert loop landing pads. 1547 // Do not bother doing this on the Root loop of course. 1548 if( !_verify_me && !_verify_only && _ltree_root->_child ) { 1549 if( _ltree_root->_child->beautify_loops( this ) ) { 1550 // Re-build loop tree! 1551 _ltree_root->_child = NULL; 1552 _nodes.clear(); 1553 reallocate_preorders(); 1554 build_loop_tree(); 1555 // Check for bailout, and return 1556 if (C->failing()) { 1557 return; 1558 } 1559 // Reset loop nesting depth 1560 _ltree_root->set_nest( 0 ); 1561 1562 C->print_method("After beautify loops", 3); 1563 } 1564 } 1565 1566 // Build Dominators for elision of NULL checks & loop finding. 1567 // Since nodes do not have a slot for immediate dominator, make 1568 // a persistent side array for that info indexed on node->_idx. 1569 _idom_size = C->unique(); 1570 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); 1571 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); 1572 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth 1573 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); 1574 1575 Dominators(); 1576 1577 if (!_verify_only) { 1578 // As a side effect, Dominators removed any unreachable CFG paths 1579 // into RegionNodes. It doesn't do this test against Root, so 1580 // we do it here. 1581 for( uint i = 1; i < C->root()->req(); i++ ) { 1582 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? 1583 _igvn.hash_delete(C->root()); 1584 C->root()->del_req(i); 1585 _igvn._worklist.push(C->root()); 1586 i--; // Rerun same iteration on compressed edges 1587 } 1588 } 1589 1590 // Given dominators, try to find inner loops with calls that must 1591 // always be executed (call dominates loop tail). These loops do 1592 // not need a separate safepoint. 1593 Node_List cisstack(a); 1594 _ltree_root->check_safepts(visited, cisstack); 1595 } 1596 1597 // Walk the DATA nodes and place into loops. Find earliest control 1598 // node. For CFG nodes, the _nodes array starts out and remains 1599 // holding the associated IdealLoopTree pointer. For DATA nodes, the 1600 // _nodes array holds the earliest legal controlling CFG node. 1601 1602 // Allocate stack with enough space to avoid frequent realloc 1603 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats 1604 Node_Stack nstack( a, stack_size ); 1605 1606 visited.Clear(); 1607 Node_List worklist(a); 1608 // Don't need C->root() on worklist since 1609 // it will be processed among C->top() inputs 1610 worklist.push( C->top() ); 1611 visited.set( C->top()->_idx ); // Set C->top() as visited now 1612 build_loop_early( visited, worklist, nstack ); 1613 1614 // Given early legal placement, try finding counted loops. This placement 1615 // is good enough to discover most loop invariants. 1616 if( !_verify_me && !_verify_only ) 1617 _ltree_root->counted_loop( this ); 1618 1619 // Find latest loop placement. Find ideal loop placement. 1620 visited.Clear(); 1621 init_dom_lca_tags(); 1622 // Need C->root() on worklist when processing outs 1623 worklist.push( C->root() ); 1624 NOT_PRODUCT( C->verify_graph_edges(); ) 1625 worklist.push( C->top() ); 1626 build_loop_late( visited, worklist, nstack ); 1627 1628 if (_verify_only) { 1629 // restore major progress flag 1630 for (int i = 0; i < old_progress; i++) 1631 C->set_major_progress(); 1632 assert(C->unique() == unique, "verification mode made Nodes? ? ?"); 1633 assert(_igvn._worklist.size() == 0, "shouldn't push anything"); 1634 return; 1635 } 1636 1637 // some parser-inserted loop predicates could never be used by loop 1638 // predication. Eliminate them before loop optimization 1639 if (UseLoopPredicate) { 1640 eliminate_useless_predicates(); 1641 } 1642 1643 // clear out the dead code 1644 while(_deadlist.size()) { 1645 _igvn.remove_globally_dead_node(_deadlist.pop()); 1646 } 1647 1648 #ifndef PRODUCT 1649 C->verify_graph_edges(); 1650 if( _verify_me ) { // Nested verify pass? 1651 // Check to see if the verify mode is broken 1652 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); 1653 return; 1654 } 1655 if( VerifyLoopOptimizations ) verify(); 1656 #endif 1657 1658 if (ReassociateInvariants) { 1659 // Reassociate invariants and prep for split_thru_phi 1660 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1661 IdealLoopTree* lpt = iter.current(); 1662 if (!lpt->is_counted() || !lpt->is_inner()) continue; 1663 1664 lpt->reassociate_invariants(this); 1665 1666 // Because RCE opportunities can be masked by split_thru_phi, 1667 // look for RCE candidates and inhibit split_thru_phi 1668 // on just their loop-phi's for this pass of loop opts 1669 if (SplitIfBlocks && do_split_ifs) { 1670 if (lpt->policy_range_check(this)) { 1671 lpt->_rce_candidate = 1; // = true 1672 } 1673 } 1674 } 1675 } 1676 1677 // Check for aggressive application of split-if and other transforms 1678 // that require basic-block info (like cloning through Phi's) 1679 if( SplitIfBlocks && do_split_ifs ) { 1680 visited.Clear(); 1681 split_if_with_blocks( visited, nstack ); 1682 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); 1683 } 1684 1685 // Perform loop predication before iteration splitting 1686 if (do_loop_pred && C->has_loops() && !C->major_progress()) { 1687 _ltree_root->_child->loop_predication(this); 1688 } 1689 1690 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { 1691 if (do_intrinsify_fill()) { 1692 C->set_major_progress(); 1693 } 1694 } 1695 1696 // Perform iteration-splitting on inner loops. Split iterations to avoid 1697 // range checks or one-shot null checks. 1698 1699 // If split-if's didn't hack the graph too bad (no CFG changes) 1700 // then do loop opts. 1701 if (C->has_loops() && !C->major_progress()) { 1702 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 1703 _ltree_root->_child->iteration_split( this, worklist ); 1704 // No verify after peeling! GCM has hoisted code out of the loop. 1705 // After peeling, the hoisted code could sink inside the peeled area. 1706 // The peeling code does not try to recompute the best location for 1707 // all the code before the peeled area, so the verify pass will always 1708 // complain about it. 1709 } 1710 // Do verify graph edges in any case 1711 NOT_PRODUCT( C->verify_graph_edges(); ); 1712 1713 if (!do_split_ifs) { 1714 // We saw major progress in Split-If to get here. We forced a 1715 // pass with unrolling and not split-if, however more split-if's 1716 // might make progress. If the unrolling didn't make progress 1717 // then the major-progress flag got cleared and we won't try 1718 // another round of Split-If. In particular the ever-common 1719 // instance-of/check-cast pattern requires at least 2 rounds of 1720 // Split-If to clear out. 1721 C->set_major_progress(); 1722 } 1723 1724 // Repeat loop optimizations if new loops were seen 1725 if (created_loop_node()) { 1726 C->set_major_progress(); 1727 } 1728 1729 // Convert scalar to superword operations 1730 1731 if (UseSuperWord && C->has_loops() && !C->major_progress()) { 1732 // SuperWord transform 1733 SuperWord sw(this); 1734 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1735 IdealLoopTree* lpt = iter.current(); 1736 if (lpt->is_counted()) { 1737 sw.transform_loop(lpt); 1738 } 1739 } 1740 } 1741 1742 // Cleanup any modified bits 1743 _igvn.optimize(); 1744 1745 // disable assert until issue with split_flow_path is resolved (6742111) 1746 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), 1747 // "shouldn't introduce irreducible loops"); 1748 1749 if (C->log() != NULL) { 1750 log_loop_tree(_ltree_root, _ltree_root, C->log()); 1751 } 1752 } 1753 1754 #ifndef PRODUCT 1755 //------------------------------print_statistics------------------------------- 1756 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes 1757 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique 1758 void PhaseIdealLoop::print_statistics() { 1759 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); 1760 } 1761 1762 //------------------------------verify----------------------------------------- 1763 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. 1764 static int fail; // debug only, so its multi-thread dont care 1765 void PhaseIdealLoop::verify() const { 1766 int old_progress = C->major_progress(); 1767 ResourceMark rm; 1768 PhaseIdealLoop loop_verify( _igvn, this ); 1769 VectorSet visited(Thread::current()->resource_area()); 1770 1771 fail = 0; 1772 verify_compare( C->root(), &loop_verify, visited ); 1773 assert( fail == 0, "verify loops failed" ); 1774 // Verify loop structure is the same 1775 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); 1776 // Reset major-progress. It was cleared by creating a verify version of 1777 // PhaseIdealLoop. 1778 for( int i=0; i<old_progress; i++ ) 1779 C->set_major_progress(); 1780 } 1781 1782 //------------------------------verify_compare--------------------------------- 1783 // Make sure me and the given PhaseIdealLoop agree on key data structures 1784 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { 1785 if( !n ) return; 1786 if( visited.test_set( n->_idx ) ) return; 1787 if( !_nodes[n->_idx] ) { // Unreachable 1788 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); 1789 return; 1790 } 1791 1792 uint i; 1793 for( i = 0; i < n->req(); i++ ) 1794 verify_compare( n->in(i), loop_verify, visited ); 1795 1796 // Check the '_nodes' block/loop structure 1797 i = n->_idx; 1798 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl 1799 if( _nodes[i] != loop_verify->_nodes[i] && 1800 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { 1801 tty->print("Mismatched control setting for: "); 1802 n->dump(); 1803 if( fail++ > 10 ) return; 1804 Node *c = get_ctrl_no_update(n); 1805 tty->print("We have it as: "); 1806 if( c->in(0) ) c->dump(); 1807 else tty->print_cr("N%d",c->_idx); 1808 tty->print("Verify thinks: "); 1809 if( loop_verify->has_ctrl(n) ) 1810 loop_verify->get_ctrl_no_update(n)->dump(); 1811 else 1812 loop_verify->get_loop_idx(n)->dump(); 1813 tty->cr(); 1814 } 1815 } else { // We have a loop 1816 IdealLoopTree *us = get_loop_idx(n); 1817 if( loop_verify->has_ctrl(n) ) { 1818 tty->print("Mismatched loop setting for: "); 1819 n->dump(); 1820 if( fail++ > 10 ) return; 1821 tty->print("We have it as: "); 1822 us->dump(); 1823 tty->print("Verify thinks: "); 1824 loop_verify->get_ctrl_no_update(n)->dump(); 1825 tty->cr(); 1826 } else if (!C->major_progress()) { 1827 // Loop selection can be messed up if we did a major progress 1828 // operation, like split-if. Do not verify in that case. 1829 IdealLoopTree *them = loop_verify->get_loop_idx(n); 1830 if( us->_head != them->_head || us->_tail != them->_tail ) { 1831 tty->print("Unequals loops for: "); 1832 n->dump(); 1833 if( fail++ > 10 ) return; 1834 tty->print("We have it as: "); 1835 us->dump(); 1836 tty->print("Verify thinks: "); 1837 them->dump(); 1838 tty->cr(); 1839 } 1840 } 1841 } 1842 1843 // Check for immediate dominators being equal 1844 if( i >= _idom_size ) { 1845 if( !n->is_CFG() ) return; 1846 tty->print("CFG Node with no idom: "); 1847 n->dump(); 1848 return; 1849 } 1850 if( !n->is_CFG() ) return; 1851 if( n == C->root() ) return; // No IDOM here 1852 1853 assert(n->_idx == i, "sanity"); 1854 Node *id = idom_no_update(n); 1855 if( id != loop_verify->idom_no_update(n) ) { 1856 tty->print("Unequals idoms for: "); 1857 n->dump(); 1858 if( fail++ > 10 ) return; 1859 tty->print("We have it as: "); 1860 id->dump(); 1861 tty->print("Verify thinks: "); 1862 loop_verify->idom_no_update(n)->dump(); 1863 tty->cr(); 1864 } 1865 1866 } 1867 1868 //------------------------------verify_tree------------------------------------ 1869 // Verify that tree structures match. Because the CFG can change, siblings 1870 // within the loop tree can be reordered. We attempt to deal with that by 1871 // reordering the verify's loop tree if possible. 1872 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { 1873 assert( _parent == parent, "Badly formed loop tree" ); 1874 1875 // Siblings not in same order? Attempt to re-order. 1876 if( _head != loop->_head ) { 1877 // Find _next pointer to update 1878 IdealLoopTree **pp = &loop->_parent->_child; 1879 while( *pp != loop ) 1880 pp = &((*pp)->_next); 1881 // Find proper sibling to be next 1882 IdealLoopTree **nn = &loop->_next; 1883 while( (*nn) && (*nn)->_head != _head ) 1884 nn = &((*nn)->_next); 1885 1886 // Check for no match. 1887 if( !(*nn) ) { 1888 // Annoyingly, irreducible loops can pick different headers 1889 // after a major_progress operation, so the rest of the loop 1890 // tree cannot be matched. 1891 if (_irreducible && Compile::current()->major_progress()) return; 1892 assert( 0, "failed to match loop tree" ); 1893 } 1894 1895 // Move (*nn) to (*pp) 1896 IdealLoopTree *hit = *nn; 1897 *nn = hit->_next; 1898 hit->_next = loop; 1899 *pp = loop; 1900 loop = hit; 1901 // Now try again to verify 1902 } 1903 1904 assert( _head == loop->_head , "mismatched loop head" ); 1905 Node *tail = _tail; // Inline a non-updating version of 1906 while( !tail->in(0) ) // the 'tail()' call. 1907 tail = tail->in(1); 1908 assert( tail == loop->_tail, "mismatched loop tail" ); 1909 1910 // Counted loops that are guarded should be able to find their guards 1911 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { 1912 CountedLoopNode *cl = _head->as_CountedLoop(); 1913 Node *init = cl->init_trip(); 1914 Node *ctrl = cl->in(LoopNode::EntryControl); 1915 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1916 Node *iff = ctrl->in(0); 1917 assert( iff->Opcode() == Op_If, "" ); 1918 Node *bol = iff->in(1); 1919 assert( bol->Opcode() == Op_Bool, "" ); 1920 Node *cmp = bol->in(1); 1921 assert( cmp->Opcode() == Op_CmpI, "" ); 1922 Node *add = cmp->in(1); 1923 Node *opaq; 1924 if( add->Opcode() == Op_Opaque1 ) { 1925 opaq = add; 1926 } else { 1927 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); 1928 assert( add == init, "" ); 1929 opaq = cmp->in(2); 1930 } 1931 assert( opaq->Opcode() == Op_Opaque1, "" ); 1932 1933 } 1934 1935 if (_child != NULL) _child->verify_tree(loop->_child, this); 1936 if (_next != NULL) _next ->verify_tree(loop->_next, parent); 1937 // Innermost loops need to verify loop bodies, 1938 // but only if no 'major_progress' 1939 int fail = 0; 1940 if (!Compile::current()->major_progress() && _child == NULL) { 1941 for( uint i = 0; i < _body.size(); i++ ) { 1942 Node *n = _body.at(i); 1943 if (n->outcnt() == 0) continue; // Ignore dead 1944 uint j; 1945 for( j = 0; j < loop->_body.size(); j++ ) 1946 if( loop->_body.at(j) == n ) 1947 break; 1948 if( j == loop->_body.size() ) { // Not found in loop body 1949 // Last ditch effort to avoid assertion: Its possible that we 1950 // have some users (so outcnt not zero) but are still dead. 1951 // Try to find from root. 1952 if (Compile::current()->root()->find(n->_idx)) { 1953 fail++; 1954 tty->print("We have that verify does not: "); 1955 n->dump(); 1956 } 1957 } 1958 } 1959 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { 1960 Node *n = loop->_body.at(i2); 1961 if (n->outcnt() == 0) continue; // Ignore dead 1962 uint j; 1963 for( j = 0; j < _body.size(); j++ ) 1964 if( _body.at(j) == n ) 1965 break; 1966 if( j == _body.size() ) { // Not found in loop body 1967 // Last ditch effort to avoid assertion: Its possible that we 1968 // have some users (so outcnt not zero) but are still dead. 1969 // Try to find from root. 1970 if (Compile::current()->root()->find(n->_idx)) { 1971 fail++; 1972 tty->print("Verify has that we do not: "); 1973 n->dump(); 1974 } 1975 } 1976 } 1977 assert( !fail, "loop body mismatch" ); 1978 } 1979 } 1980 1981 #endif 1982 1983 //------------------------------set_idom--------------------------------------- 1984 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { 1985 uint idx = d->_idx; 1986 if (idx >= _idom_size) { 1987 uint newsize = _idom_size<<1; 1988 while( idx >= newsize ) { 1989 newsize <<= 1; 1990 } 1991 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); 1992 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); 1993 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); 1994 _idom_size = newsize; 1995 } 1996 _idom[idx] = n; 1997 _dom_depth[idx] = dom_depth; 1998 } 1999 2000 //------------------------------recompute_dom_depth--------------------------------------- 2001 // The dominator tree is constructed with only parent pointers. 2002 // This recomputes the depth in the tree by first tagging all 2003 // nodes as "no depth yet" marker. The next pass then runs up 2004 // the dom tree from each node marked "no depth yet", and computes 2005 // the depth on the way back down. 2006 void PhaseIdealLoop::recompute_dom_depth() { 2007 uint no_depth_marker = C->unique(); 2008 uint i; 2009 // Initialize depth to "no depth yet" 2010 for (i = 0; i < _idom_size; i++) { 2011 if (_dom_depth[i] > 0 && _idom[i] != NULL) { 2012 _dom_depth[i] = no_depth_marker; 2013 } 2014 } 2015 if (_dom_stk == NULL) { 2016 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size. 2017 if (init_size < 10) init_size = 10; 2018 _dom_stk = new GrowableArray<uint>(init_size); 2019 } 2020 // Compute new depth for each node. 2021 for (i = 0; i < _idom_size; i++) { 2022 uint j = i; 2023 // Run up the dom tree to find a node with a depth 2024 while (_dom_depth[j] == no_depth_marker) { 2025 _dom_stk->push(j); 2026 j = _idom[j]->_idx; 2027 } 2028 // Compute the depth on the way back down this tree branch 2029 uint dd = _dom_depth[j] + 1; 2030 while (_dom_stk->length() > 0) { 2031 uint j = _dom_stk->pop(); 2032 _dom_depth[j] = dd; 2033 dd++; 2034 } 2035 } 2036 } 2037 2038 //------------------------------sort------------------------------------------- 2039 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the 2040 // loop tree, not the root. 2041 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { 2042 if( !innermost ) return loop; // New innermost loop 2043 2044 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number 2045 assert( loop_preorder, "not yet post-walked loop" ); 2046 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer 2047 IdealLoopTree *l = *pp; // Do I go before or after 'l'? 2048 2049 // Insert at start of list 2050 while( l ) { // Insertion sort based on pre-order 2051 if( l == loop ) return innermost; // Already on list! 2052 int l_preorder = get_preorder(l->_head); // Cache pre-order number 2053 assert( l_preorder, "not yet post-walked l" ); 2054 // Check header pre-order number to figure proper nesting 2055 if( loop_preorder > l_preorder ) 2056 break; // End of insertion 2057 // If headers tie (e.g., shared headers) check tail pre-order numbers. 2058 // Since I split shared headers, you'd think this could not happen. 2059 // BUT: I must first do the preorder numbering before I can discover I 2060 // have shared headers, so the split headers all get the same preorder 2061 // number as the RegionNode they split from. 2062 if( loop_preorder == l_preorder && 2063 get_preorder(loop->_tail) < get_preorder(l->_tail) ) 2064 break; // Also check for shared headers (same pre#) 2065 pp = &l->_parent; // Chain up list 2066 l = *pp; 2067 } 2068 // Link into list 2069 // Point predecessor to me 2070 *pp = loop; 2071 // Point me to successor 2072 IdealLoopTree *p = loop->_parent; 2073 loop->_parent = l; // Point me to successor 2074 if( p ) sort( p, innermost ); // Insert my parents into list as well 2075 return innermost; 2076 } 2077 2078 //------------------------------build_loop_tree-------------------------------- 2079 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit 2080 // bits. The _nodes[] array is mapped by Node index and holds a NULL for 2081 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the 2082 // tightest enclosing IdealLoopTree for post-walked. 2083 // 2084 // During my forward walk I do a short 1-layer lookahead to see if I can find 2085 // a loop backedge with that doesn't have any work on the backedge. This 2086 // helps me construct nested loops with shared headers better. 2087 // 2088 // Once I've done the forward recursion, I do the post-work. For each child 2089 // I check to see if there is a backedge. Backedges define a loop! I 2090 // insert an IdealLoopTree at the target of the backedge. 2091 // 2092 // During the post-work I also check to see if I have several children 2093 // belonging to different loops. If so, then this Node is a decision point 2094 // where control flow can choose to change loop nests. It is at this 2095 // decision point where I can figure out how loops are nested. At this 2096 // time I can properly order the different loop nests from my children. 2097 // Note that there may not be any backedges at the decision point! 2098 // 2099 // Since the decision point can be far removed from the backedges, I can't 2100 // order my loops at the time I discover them. Thus at the decision point 2101 // I need to inspect loop header pre-order numbers to properly nest my 2102 // loops. This means I need to sort my childrens' loops by pre-order. 2103 // The sort is of size number-of-control-children, which generally limits 2104 // it to size 2 (i.e., I just choose between my 2 target loops). 2105 void PhaseIdealLoop::build_loop_tree() { 2106 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2107 GrowableArray <Node *> bltstack(C->unique() >> 1); 2108 Node *n = C->root(); 2109 bltstack.push(n); 2110 int pre_order = 1; 2111 int stack_size; 2112 2113 while ( ( stack_size = bltstack.length() ) != 0 ) { 2114 n = bltstack.top(); // Leave node on stack 2115 if ( !is_visited(n) ) { 2116 // ---- Pre-pass Work ---- 2117 // Pre-walked but not post-walked nodes need a pre_order number. 2118 2119 set_preorder_visited( n, pre_order ); // set as visited 2120 2121 // ---- Scan over children ---- 2122 // Scan first over control projections that lead to loop headers. 2123 // This helps us find inner-to-outer loops with shared headers better. 2124 2125 // Scan children's children for loop headers. 2126 for ( int i = n->outcnt() - 1; i >= 0; --i ) { 2127 Node* m = n->raw_out(i); // Child 2128 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children 2129 // Scan over children's children to find loop 2130 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 2131 Node* l = m->fast_out(j); 2132 if( is_visited(l) && // Been visited? 2133 !is_postvisited(l) && // But not post-visited 2134 get_preorder(l) < pre_order ) { // And smaller pre-order 2135 // Found! Scan the DFS down this path before doing other paths 2136 bltstack.push(m); 2137 break; 2138 } 2139 } 2140 } 2141 } 2142 pre_order++; 2143 } 2144 else if ( !is_postvisited(n) ) { 2145 // Note: build_loop_tree_impl() adds out edges on rare occasions, 2146 // such as com.sun.rsasign.am::a. 2147 // For non-recursive version, first, process current children. 2148 // On next iteration, check if additional children were added. 2149 for ( int k = n->outcnt() - 1; k >= 0; --k ) { 2150 Node* u = n->raw_out(k); 2151 if ( u->is_CFG() && !is_visited(u) ) { 2152 bltstack.push(u); 2153 } 2154 } 2155 if ( bltstack.length() == stack_size ) { 2156 // There were no additional children, post visit node now 2157 (void)bltstack.pop(); // Remove node from stack 2158 pre_order = build_loop_tree_impl( n, pre_order ); 2159 // Check for bailout 2160 if (C->failing()) { 2161 return; 2162 } 2163 // Check to grow _preorders[] array for the case when 2164 // build_loop_tree_impl() adds new nodes. 2165 check_grow_preorders(); 2166 } 2167 } 2168 else { 2169 (void)bltstack.pop(); // Remove post-visited node from stack 2170 } 2171 } 2172 } 2173 2174 //------------------------------build_loop_tree_impl--------------------------- 2175 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { 2176 // ---- Post-pass Work ---- 2177 // Pre-walked but not post-walked nodes need a pre_order number. 2178 2179 // Tightest enclosing loop for this Node 2180 IdealLoopTree *innermost = NULL; 2181 2182 // For all children, see if any edge is a backedge. If so, make a loop 2183 // for it. Then find the tightest enclosing loop for the self Node. 2184 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2185 Node* m = n->fast_out(i); // Child 2186 if( n == m ) continue; // Ignore control self-cycles 2187 if( !m->is_CFG() ) continue;// Ignore non-CFG edges 2188 2189 IdealLoopTree *l; // Child's loop 2190 if( !is_postvisited(m) ) { // Child visited but not post-visited? 2191 // Found a backedge 2192 assert( get_preorder(m) < pre_order, "should be backedge" ); 2193 // Check for the RootNode, which is already a LoopNode and is allowed 2194 // to have multiple "backedges". 2195 if( m == C->root()) { // Found the root? 2196 l = _ltree_root; // Root is the outermost LoopNode 2197 } else { // Else found a nested loop 2198 // Insert a LoopNode to mark this loop. 2199 l = new IdealLoopTree(this, m, n); 2200 } // End of Else found a nested loop 2201 if( !has_loop(m) ) // If 'm' does not already have a loop set 2202 set_loop(m, l); // Set loop header to loop now 2203 2204 } else { // Else not a nested loop 2205 if( !_nodes[m->_idx] ) continue; // Dead code has no loop 2206 l = get_loop(m); // Get previously determined loop 2207 // If successor is header of a loop (nest), move up-loop till it 2208 // is a member of some outer enclosing loop. Since there are no 2209 // shared headers (I've split them already) I only need to go up 2210 // at most 1 level. 2211 while( l && l->_head == m ) // Successor heads loop? 2212 l = l->_parent; // Move up 1 for me 2213 // If this loop is not properly parented, then this loop 2214 // has no exit path out, i.e. its an infinite loop. 2215 if( !l ) { 2216 // Make loop "reachable" from root so the CFG is reachable. Basically 2217 // insert a bogus loop exit that is never taken. 'm', the loop head, 2218 // points to 'n', one (of possibly many) fall-in paths. There may be 2219 // many backedges as well. 2220 2221 // Here I set the loop to be the root loop. I could have, after 2222 // inserting a bogus loop exit, restarted the recursion and found my 2223 // new loop exit. This would make the infinite loop a first-class 2224 // loop and it would then get properly optimized. What's the use of 2225 // optimizing an infinite loop? 2226 l = _ltree_root; // Oops, found infinite loop 2227 2228 if (!_verify_only) { 2229 // Insert the NeverBranch between 'm' and it's control user. 2230 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m ); 2231 _igvn.register_new_node_with_optimizer(iff); 2232 set_loop(iff, l); 2233 Node *if_t = new (C, 1) CProjNode( iff, 0 ); 2234 _igvn.register_new_node_with_optimizer(if_t); 2235 set_loop(if_t, l); 2236 2237 Node* cfg = NULL; // Find the One True Control User of m 2238 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 2239 Node* x = m->fast_out(j); 2240 if (x->is_CFG() && x != m && x != iff) 2241 { cfg = x; break; } 2242 } 2243 assert(cfg != NULL, "must find the control user of m"); 2244 uint k = 0; // Probably cfg->in(0) 2245 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region 2246 cfg->set_req( k, if_t ); // Now point to NeverBranch 2247 2248 // Now create the never-taken loop exit 2249 Node *if_f = new (C, 1) CProjNode( iff, 1 ); 2250 _igvn.register_new_node_with_optimizer(if_f); 2251 set_loop(if_f, l); 2252 // Find frame ptr for Halt. Relies on the optimizer 2253 // V-N'ing. Easier and quicker than searching through 2254 // the program structure. 2255 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr ); 2256 _igvn.register_new_node_with_optimizer(frame); 2257 // Halt & Catch Fire 2258 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame ); 2259 _igvn.register_new_node_with_optimizer(halt); 2260 set_loop(halt, l); 2261 C->root()->add_req(halt); 2262 } 2263 set_loop(C->root(), _ltree_root); 2264 } 2265 } 2266 // Weeny check for irreducible. This child was already visited (this 2267 // IS the post-work phase). Is this child's loop header post-visited 2268 // as well? If so, then I found another entry into the loop. 2269 if (!_verify_only) { 2270 while( is_postvisited(l->_head) ) { 2271 // found irreducible 2272 l->_irreducible = 1; // = true 2273 l = l->_parent; 2274 _has_irreducible_loops = true; 2275 // Check for bad CFG here to prevent crash, and bailout of compile 2276 if (l == NULL) { 2277 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); 2278 return pre_order; 2279 } 2280 } 2281 } 2282 2283 // This Node might be a decision point for loops. It is only if 2284 // it's children belong to several different loops. The sort call 2285 // does a trivial amount of work if there is only 1 child or all 2286 // children belong to the same loop. If however, the children 2287 // belong to different loops, the sort call will properly set the 2288 // _parent pointers to show how the loops nest. 2289 // 2290 // In any case, it returns the tightest enclosing loop. 2291 innermost = sort( l, innermost ); 2292 } 2293 2294 // Def-use info will have some dead stuff; dead stuff will have no 2295 // loop decided on. 2296 2297 // Am I a loop header? If so fix up my parent's child and next ptrs. 2298 if( innermost && innermost->_head == n ) { 2299 assert( get_loop(n) == innermost, "" ); 2300 IdealLoopTree *p = innermost->_parent; 2301 IdealLoopTree *l = innermost; 2302 while( p && l->_head == n ) { 2303 l->_next = p->_child; // Put self on parents 'next child' 2304 p->_child = l; // Make self as first child of parent 2305 l = p; // Now walk up the parent chain 2306 p = l->_parent; 2307 } 2308 } else { 2309 // Note that it is possible for a LoopNode to reach here, if the 2310 // backedge has been made unreachable (hence the LoopNode no longer 2311 // denotes a Loop, and will eventually be removed). 2312 2313 // Record tightest enclosing loop for self. Mark as post-visited. 2314 set_loop(n, innermost); 2315 // Also record has_call flag early on 2316 if( innermost ) { 2317 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { 2318 // Do not count uncommon calls 2319 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { 2320 Node *iff = n->in(0)->in(0); 2321 if( !iff->is_If() || 2322 (n->in(0)->Opcode() == Op_IfFalse && 2323 (1.0 - iff->as_If()->_prob) >= 0.01) || 2324 (iff->as_If()->_prob >= 0.01) ) 2325 innermost->_has_call = 1; 2326 } 2327 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { 2328 // Disable loop optimizations if the loop has a scalar replaceable 2329 // allocation. This disabling may cause a potential performance lost 2330 // if the allocation is not eliminated for some reason. 2331 innermost->_allow_optimizations = false; 2332 innermost->_has_call = 1; // = true 2333 } 2334 } 2335 } 2336 2337 // Flag as post-visited now 2338 set_postvisited(n); 2339 return pre_order; 2340 } 2341 2342 2343 //------------------------------build_loop_early------------------------------- 2344 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 2345 // First pass computes the earliest controlling node possible. This is the 2346 // controlling input with the deepest dominating depth. 2347 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 2348 while (worklist.size() != 0) { 2349 // Use local variables nstack_top_n & nstack_top_i to cache values 2350 // on nstack's top. 2351 Node *nstack_top_n = worklist.pop(); 2352 uint nstack_top_i = 0; 2353 //while_nstack_nonempty: 2354 while (true) { 2355 // Get parent node and next input's index from stack's top. 2356 Node *n = nstack_top_n; 2357 uint i = nstack_top_i; 2358 uint cnt = n->req(); // Count of inputs 2359 if (i == 0) { // Pre-process the node. 2360 if( has_node(n) && // Have either loop or control already? 2361 !has_ctrl(n) ) { // Have loop picked out already? 2362 // During "merge_many_backedges" we fold up several nested loops 2363 // into a single loop. This makes the members of the original 2364 // loop bodies pointing to dead loops; they need to move up 2365 // to the new UNION'd larger loop. I set the _head field of these 2366 // dead loops to NULL and the _parent field points to the owning 2367 // loop. Shades of UNION-FIND algorithm. 2368 IdealLoopTree *ilt; 2369 while( !(ilt = get_loop(n))->_head ) { 2370 // Normally I would use a set_loop here. But in this one special 2371 // case, it is legal (and expected) to change what loop a Node 2372 // belongs to. 2373 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); 2374 } 2375 // Remove safepoints ONLY if I've already seen I don't need one. 2376 // (the old code here would yank a 2nd safepoint after seeing a 2377 // first one, even though the 1st did not dominate in the loop body 2378 // and thus could be avoided indefinitely) 2379 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && 2380 is_deleteable_safept(n)) { 2381 Node *in = n->in(TypeFunc::Control); 2382 lazy_replace(n,in); // Pull safepoint now 2383 // Carry on with the recursion "as if" we are walking 2384 // only the control input 2385 if( !visited.test_set( in->_idx ) ) { 2386 worklist.push(in); // Visit this guy later, using worklist 2387 } 2388 // Get next node from nstack: 2389 // - skip n's inputs processing by setting i > cnt; 2390 // - we also will not call set_early_ctrl(n) since 2391 // has_node(n) == true (see the condition above). 2392 i = cnt + 1; 2393 } 2394 } 2395 } // if (i == 0) 2396 2397 // Visit all inputs 2398 bool done = true; // Assume all n's inputs will be processed 2399 while (i < cnt) { 2400 Node *in = n->in(i); 2401 ++i; 2402 if (in == NULL) continue; 2403 if (in->pinned() && !in->is_CFG()) 2404 set_ctrl(in, in->in(0)); 2405 int is_visited = visited.test_set( in->_idx ); 2406 if (!has_node(in)) { // No controlling input yet? 2407 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); 2408 assert( !is_visited, "visit only once" ); 2409 nstack.push(n, i); // Save parent node and next input's index. 2410 nstack_top_n = in; // Process current input now. 2411 nstack_top_i = 0; 2412 done = false; // Not all n's inputs processed. 2413 break; // continue while_nstack_nonempty; 2414 } else if (!is_visited) { 2415 // This guy has a location picked out for him, but has not yet 2416 // been visited. Happens to all CFG nodes, for instance. 2417 // Visit him using the worklist instead of recursion, to break 2418 // cycles. Since he has a location already we do not need to 2419 // find his location before proceeding with the current Node. 2420 worklist.push(in); // Visit this guy later, using worklist 2421 } 2422 } 2423 if (done) { 2424 // All of n's inputs have been processed, complete post-processing. 2425 2426 // Compute earliest point this Node can go. 2427 // CFG, Phi, pinned nodes already know their controlling input. 2428 if (!has_node(n)) { 2429 // Record earliest legal location 2430 set_early_ctrl( n ); 2431 } 2432 if (nstack.is_empty()) { 2433 // Finished all nodes on stack. 2434 // Process next node on the worklist. 2435 break; 2436 } 2437 // Get saved parent node and next input's index. 2438 nstack_top_n = nstack.node(); 2439 nstack_top_i = nstack.index(); 2440 nstack.pop(); 2441 } 2442 } // while (true) 2443 } 2444 } 2445 2446 //------------------------------dom_lca_internal-------------------------------- 2447 // Pair-wise LCA 2448 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { 2449 if( !n1 ) return n2; // Handle NULL original LCA 2450 assert( n1->is_CFG(), "" ); 2451 assert( n2->is_CFG(), "" ); 2452 // find LCA of all uses 2453 uint d1 = dom_depth(n1); 2454 uint d2 = dom_depth(n2); 2455 while (n1 != n2) { 2456 if (d1 > d2) { 2457 n1 = idom(n1); 2458 d1 = dom_depth(n1); 2459 } else if (d1 < d2) { 2460 n2 = idom(n2); 2461 d2 = dom_depth(n2); 2462 } else { 2463 // Here d1 == d2. Due to edits of the dominator-tree, sections 2464 // of the tree might have the same depth. These sections have 2465 // to be searched more carefully. 2466 2467 // Scan up all the n1's with equal depth, looking for n2. 2468 Node *t1 = idom(n1); 2469 while (dom_depth(t1) == d1) { 2470 if (t1 == n2) return n2; 2471 t1 = idom(t1); 2472 } 2473 // Scan up all the n2's with equal depth, looking for n1. 2474 Node *t2 = idom(n2); 2475 while (dom_depth(t2) == d2) { 2476 if (t2 == n1) return n1; 2477 t2 = idom(t2); 2478 } 2479 // Move up to a new dominator-depth value as well as up the dom-tree. 2480 n1 = t1; 2481 n2 = t2; 2482 d1 = dom_depth(n1); 2483 d2 = dom_depth(n2); 2484 } 2485 } 2486 return n1; 2487 } 2488 2489 //------------------------------compute_idom----------------------------------- 2490 // Locally compute IDOM using dom_lca call. Correct only if the incoming 2491 // IDOMs are correct. 2492 Node *PhaseIdealLoop::compute_idom( Node *region ) const { 2493 assert( region->is_Region(), "" ); 2494 Node *LCA = NULL; 2495 for( uint i = 1; i < region->req(); i++ ) { 2496 if( region->in(i) != C->top() ) 2497 LCA = dom_lca( LCA, region->in(i) ); 2498 } 2499 return LCA; 2500 } 2501 2502 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { 2503 bool had_error = false; 2504 #ifdef ASSERT 2505 if (early != C->root()) { 2506 // Make sure that there's a dominance path from use to LCA 2507 Node* d = use; 2508 while (d != LCA) { 2509 d = idom(d); 2510 if (d == C->root()) { 2511 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx); 2512 n->dump(); 2513 use->dump(); 2514 had_error = true; 2515 break; 2516 } 2517 } 2518 } 2519 #endif 2520 return had_error; 2521 } 2522 2523 2524 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { 2525 // Compute LCA over list of uses 2526 bool had_error = false; 2527 Node *LCA = NULL; 2528 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { 2529 Node* c = n->fast_out(i); 2530 if (_nodes[c->_idx] == NULL) 2531 continue; // Skip the occasional dead node 2532 if( c->is_Phi() ) { // For Phis, we must land above on the path 2533 for( uint j=1; j<c->req(); j++ ) {// For all inputs 2534 if( c->in(j) == n ) { // Found matching input? 2535 Node *use = c->in(0)->in(j); 2536 if (_verify_only && use->is_top()) continue; 2537 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 2538 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 2539 } 2540 } 2541 } else { 2542 // For CFG data-users, use is in the block just prior 2543 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); 2544 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 2545 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 2546 } 2547 } 2548 assert(!had_error, "bad dominance"); 2549 return LCA; 2550 } 2551 2552 //------------------------------get_late_ctrl---------------------------------- 2553 // Compute latest legal control. 2554 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { 2555 assert(early != NULL, "early control should not be NULL"); 2556 2557 Node* LCA = compute_lca_of_uses(n, early); 2558 #ifdef ASSERT 2559 if (LCA == C->root() && LCA != early) { 2560 // def doesn't dominate uses so print some useful debugging output 2561 compute_lca_of_uses(n, early, true); 2562 } 2563 #endif 2564 2565 // if this is a load, check for anti-dependent stores 2566 // We use a conservative algorithm to identify potential interfering 2567 // instructions and for rescheduling the load. The users of the memory 2568 // input of this load are examined. Any use which is not a load and is 2569 // dominated by early is considered a potentially interfering store. 2570 // This can produce false positives. 2571 if (n->is_Load() && LCA != early) { 2572 Node_List worklist; 2573 2574 Node *mem = n->in(MemNode::Memory); 2575 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 2576 Node* s = mem->fast_out(i); 2577 worklist.push(s); 2578 } 2579 while(worklist.size() != 0 && LCA != early) { 2580 Node* s = worklist.pop(); 2581 if (s->is_Load()) { 2582 continue; 2583 } else if (s->is_MergeMem()) { 2584 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 2585 Node* s1 = s->fast_out(i); 2586 worklist.push(s1); 2587 } 2588 } else { 2589 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); 2590 assert(sctrl != NULL || s->outcnt() == 0, "must have control"); 2591 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { 2592 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); 2593 } 2594 } 2595 } 2596 } 2597 2598 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); 2599 return LCA; 2600 } 2601 2602 // true if CFG node d dominates CFG node n 2603 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { 2604 if (d == n) 2605 return true; 2606 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 2607 uint dd = dom_depth(d); 2608 while (dom_depth(n) >= dd) { 2609 if (n == d) 2610 return true; 2611 n = idom(n); 2612 } 2613 return false; 2614 } 2615 2616 //------------------------------dom_lca_for_get_late_ctrl_internal------------- 2617 // Pair-wise LCA with tags. 2618 // Tag each index with the node 'tag' currently being processed 2619 // before advancing up the dominator chain using idom(). 2620 // Later calls that find a match to 'tag' know that this path has already 2621 // been considered in the current LCA (which is input 'n1' by convention). 2622 // Since get_late_ctrl() is only called once for each node, the tag array 2623 // does not need to be cleared between calls to get_late_ctrl(). 2624 // Algorithm trades a larger constant factor for better asymptotic behavior 2625 // 2626 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { 2627 uint d1 = dom_depth(n1); 2628 uint d2 = dom_depth(n2); 2629 2630 do { 2631 if (d1 > d2) { 2632 // current lca is deeper than n2 2633 _dom_lca_tags.map(n1->_idx, tag); 2634 n1 = idom(n1); 2635 d1 = dom_depth(n1); 2636 } else if (d1 < d2) { 2637 // n2 is deeper than current lca 2638 Node *memo = _dom_lca_tags[n2->_idx]; 2639 if( memo == tag ) { 2640 return n1; // Return the current LCA 2641 } 2642 _dom_lca_tags.map(n2->_idx, tag); 2643 n2 = idom(n2); 2644 d2 = dom_depth(n2); 2645 } else { 2646 // Here d1 == d2. Due to edits of the dominator-tree, sections 2647 // of the tree might have the same depth. These sections have 2648 // to be searched more carefully. 2649 2650 // Scan up all the n1's with equal depth, looking for n2. 2651 _dom_lca_tags.map(n1->_idx, tag); 2652 Node *t1 = idom(n1); 2653 while (dom_depth(t1) == d1) { 2654 if (t1 == n2) return n2; 2655 _dom_lca_tags.map(t1->_idx, tag); 2656 t1 = idom(t1); 2657 } 2658 // Scan up all the n2's with equal depth, looking for n1. 2659 _dom_lca_tags.map(n2->_idx, tag); 2660 Node *t2 = idom(n2); 2661 while (dom_depth(t2) == d2) { 2662 if (t2 == n1) return n1; 2663 _dom_lca_tags.map(t2->_idx, tag); 2664 t2 = idom(t2); 2665 } 2666 // Move up to a new dominator-depth value as well as up the dom-tree. 2667 n1 = t1; 2668 n2 = t2; 2669 d1 = dom_depth(n1); 2670 d2 = dom_depth(n2); 2671 } 2672 } while (n1 != n2); 2673 return n1; 2674 } 2675 2676 //------------------------------init_dom_lca_tags------------------------------ 2677 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 2678 // Intended use does not involve any growth for the array, so it could 2679 // be of fixed size. 2680 void PhaseIdealLoop::init_dom_lca_tags() { 2681 uint limit = C->unique() + 1; 2682 _dom_lca_tags.map( limit, NULL ); 2683 #ifdef ASSERT 2684 for( uint i = 0; i < limit; ++i ) { 2685 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 2686 } 2687 #endif // ASSERT 2688 } 2689 2690 //------------------------------clear_dom_lca_tags------------------------------ 2691 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 2692 // Intended use does not involve any growth for the array, so it could 2693 // be of fixed size. 2694 void PhaseIdealLoop::clear_dom_lca_tags() { 2695 uint limit = C->unique() + 1; 2696 _dom_lca_tags.map( limit, NULL ); 2697 _dom_lca_tags.clear(); 2698 #ifdef ASSERT 2699 for( uint i = 0; i < limit; ++i ) { 2700 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 2701 } 2702 #endif // ASSERT 2703 } 2704 2705 //------------------------------build_loop_late-------------------------------- 2706 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 2707 // Second pass finds latest legal placement, and ideal loop placement. 2708 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 2709 while (worklist.size() != 0) { 2710 Node *n = worklist.pop(); 2711 // Only visit once 2712 if (visited.test_set(n->_idx)) continue; 2713 uint cnt = n->outcnt(); 2714 uint i = 0; 2715 while (true) { 2716 assert( _nodes[n->_idx], "no dead nodes" ); 2717 // Visit all children 2718 if (i < cnt) { 2719 Node* use = n->raw_out(i); 2720 ++i; 2721 // Check for dead uses. Aggressively prune such junk. It might be 2722 // dead in the global sense, but still have local uses so I cannot 2723 // easily call 'remove_dead_node'. 2724 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? 2725 // Due to cycles, we might not hit the same fixed point in the verify 2726 // pass as we do in the regular pass. Instead, visit such phis as 2727 // simple uses of the loop head. 2728 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { 2729 if( !visited.test(use->_idx) ) 2730 worklist.push(use); 2731 } else if( !visited.test_set(use->_idx) ) { 2732 nstack.push(n, i); // Save parent and next use's index. 2733 n = use; // Process all children of current use. 2734 cnt = use->outcnt(); 2735 i = 0; 2736 } 2737 } else { 2738 // Do not visit around the backedge of loops via data edges. 2739 // push dead code onto a worklist 2740 _deadlist.push(use); 2741 } 2742 } else { 2743 // All of n's children have been processed, complete post-processing. 2744 build_loop_late_post(n); 2745 if (nstack.is_empty()) { 2746 // Finished all nodes on stack. 2747 // Process next node on the worklist. 2748 break; 2749 } 2750 // Get saved parent node and next use's index. Visit the rest of uses. 2751 n = nstack.node(); 2752 cnt = n->outcnt(); 2753 i = nstack.index(); 2754 nstack.pop(); 2755 } 2756 } 2757 } 2758 } 2759 2760 //------------------------------build_loop_late_post--------------------------- 2761 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 2762 // Second pass finds latest legal placement, and ideal loop placement. 2763 void PhaseIdealLoop::build_loop_late_post( Node *n ) { 2764 2765 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) { 2766 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. 2767 } 2768 2769 // CFG and pinned nodes already handled 2770 if( n->in(0) ) { 2771 if( n->in(0)->is_top() ) return; // Dead? 2772 2773 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads 2774 // _must_ be pinned (they have to observe their control edge of course). 2775 // Unlike Stores (which modify an unallocable resource, the memory 2776 // state), Mods/Loads can float around. So free them up. 2777 bool pinned = true; 2778 switch( n->Opcode() ) { 2779 case Op_DivI: 2780 case Op_DivF: 2781 case Op_DivD: 2782 case Op_ModI: 2783 case Op_ModF: 2784 case Op_ModD: 2785 case Op_LoadB: // Same with Loads; they can sink 2786 case Op_LoadUS: // during loop optimizations. 2787 case Op_LoadD: 2788 case Op_LoadF: 2789 case Op_LoadI: 2790 case Op_LoadKlass: 2791 case Op_LoadNKlass: 2792 case Op_LoadL: 2793 case Op_LoadS: 2794 case Op_LoadP: 2795 case Op_LoadN: 2796 case Op_LoadRange: 2797 case Op_LoadD_unaligned: 2798 case Op_LoadL_unaligned: 2799 case Op_StrComp: // Does a bunch of load-like effects 2800 case Op_StrEquals: 2801 case Op_StrIndexOf: 2802 case Op_AryEq: 2803 pinned = false; 2804 } 2805 if( pinned ) { 2806 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); 2807 if( !chosen_loop->_child ) // Inner loop? 2808 chosen_loop->_body.push(n); // Collect inner loops 2809 return; 2810 } 2811 } else { // No slot zero 2812 if( n->is_CFG() ) { // CFG with no slot 0 is dead 2813 _nodes.map(n->_idx,0); // No block setting, it's globally dead 2814 return; 2815 } 2816 assert(!n->is_CFG() || n->outcnt() == 0, ""); 2817 } 2818 2819 // Do I have a "safe range" I can select over? 2820 Node *early = get_ctrl(n);// Early location already computed 2821 2822 // Compute latest point this Node can go 2823 Node *LCA = get_late_ctrl( n, early ); 2824 // LCA is NULL due to uses being dead 2825 if( LCA == NULL ) { 2826 #ifdef ASSERT 2827 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { 2828 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); 2829 } 2830 #endif 2831 _nodes.map(n->_idx, 0); // This node is useless 2832 _deadlist.push(n); 2833 return; 2834 } 2835 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); 2836 2837 Node *legal = LCA; // Walk 'legal' up the IDOM chain 2838 Node *least = legal; // Best legal position so far 2839 while( early != legal ) { // While not at earliest legal 2840 #ifdef ASSERT 2841 if (legal->is_Start() && !early->is_Root()) { 2842 // Bad graph. Print idom path and fail. 2843 tty->print_cr( "Bad graph detected in build_loop_late"); 2844 tty->print("n: ");n->dump(); tty->cr(); 2845 tty->print("early: ");early->dump(); tty->cr(); 2846 int ct = 0; 2847 Node *dbg_legal = LCA; 2848 while(!dbg_legal->is_Start() && ct < 100) { 2849 tty->print("idom[%d] ",ct); dbg_legal->dump(); tty->cr(); 2850 ct++; 2851 dbg_legal = idom(dbg_legal); 2852 } 2853 assert(false, "Bad graph detected in build_loop_late"); 2854 } 2855 #endif 2856 // Find least loop nesting depth 2857 legal = idom(legal); // Bump up the IDOM tree 2858 // Check for lower nesting depth 2859 if( get_loop(legal)->_nest < get_loop(least)->_nest ) 2860 least = legal; 2861 } 2862 assert(early == legal || legal != C->root(), "bad dominance of inputs"); 2863 2864 // Try not to place code on a loop entry projection 2865 // which can inhibit range check elimination. 2866 if (least != early) { 2867 Node* ctrl_out = least->unique_ctrl_out(); 2868 if (ctrl_out && ctrl_out->is_CountedLoop() && 2869 least == ctrl_out->in(LoopNode::EntryControl)) { 2870 Node* least_dom = idom(least); 2871 if (get_loop(least_dom)->is_member(get_loop(least))) { 2872 least = least_dom; 2873 } 2874 } 2875 } 2876 2877 #ifdef ASSERT 2878 // If verifying, verify that 'verify_me' has a legal location 2879 // and choose it as our location. 2880 if( _verify_me ) { 2881 Node *v_ctrl = _verify_me->get_ctrl_no_update(n); 2882 Node *legal = LCA; 2883 while( early != legal ) { // While not at earliest legal 2884 if( legal == v_ctrl ) break; // Check for prior good location 2885 legal = idom(legal) ;// Bump up the IDOM tree 2886 } 2887 // Check for prior good location 2888 if( legal == v_ctrl ) least = legal; // Keep prior if found 2889 } 2890 #endif 2891 2892 // Assign discovered "here or above" point 2893 least = find_non_split_ctrl(least); 2894 set_ctrl(n, least); 2895 2896 // Collect inner loop bodies 2897 IdealLoopTree *chosen_loop = get_loop(least); 2898 if( !chosen_loop->_child ) // Inner loop? 2899 chosen_loop->_body.push(n);// Collect inner loops 2900 } 2901 2902 #ifndef PRODUCT 2903 //------------------------------dump------------------------------------------- 2904 void PhaseIdealLoop::dump( ) const { 2905 ResourceMark rm; 2906 Arena* arena = Thread::current()->resource_area(); 2907 Node_Stack stack(arena, C->unique() >> 2); 2908 Node_List rpo_list; 2909 VectorSet visited(arena); 2910 visited.set(C->top()->_idx); 2911 rpo( C->root(), stack, visited, rpo_list ); 2912 // Dump root loop indexed by last element in PO order 2913 dump( _ltree_root, rpo_list.size(), rpo_list ); 2914 } 2915 2916 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { 2917 loop->dump_head(); 2918 2919 // Now scan for CFG nodes in the same loop 2920 for( uint j=idx; j > 0; j-- ) { 2921 Node *n = rpo_list[j-1]; 2922 if( !_nodes[n->_idx] ) // Skip dead nodes 2923 continue; 2924 if( get_loop(n) != loop ) { // Wrong loop nest 2925 if( get_loop(n)->_head == n && // Found nested loop? 2926 get_loop(n)->_parent == loop ) 2927 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly 2928 continue; 2929 } 2930 2931 // Dump controlling node 2932 for( uint x = 0; x < loop->_nest; x++ ) 2933 tty->print(" "); 2934 tty->print("C"); 2935 if( n == C->root() ) { 2936 n->dump(); 2937 } else { 2938 Node* cached_idom = idom_no_update(n); 2939 Node *computed_idom = n->in(0); 2940 if( n->is_Region() ) { 2941 computed_idom = compute_idom(n); 2942 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or 2943 // any MultiBranch ctrl node), so apply a similar transform to 2944 // the cached idom returned from idom_no_update. 2945 cached_idom = find_non_split_ctrl(cached_idom); 2946 } 2947 tty->print(" ID:%d",computed_idom->_idx); 2948 n->dump(); 2949 if( cached_idom != computed_idom ) { 2950 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", 2951 computed_idom->_idx, cached_idom->_idx); 2952 } 2953 } 2954 // Dump nodes it controls 2955 for( uint k = 0; k < _nodes.Size(); k++ ) { 2956 // (k < C->unique() && get_ctrl(find(k)) == n) 2957 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { 2958 Node *m = C->root()->find(k); 2959 if( m && m->outcnt() > 0 ) { 2960 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { 2961 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", 2962 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); 2963 } 2964 for( uint j = 0; j < loop->_nest; j++ ) 2965 tty->print(" "); 2966 tty->print(" "); 2967 m->dump(); 2968 } 2969 } 2970 } 2971 } 2972 } 2973 2974 // Collect a R-P-O for the whole CFG. 2975 // Result list is in post-order (scan backwards for RPO) 2976 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { 2977 stk.push(start, 0); 2978 visited.set(start->_idx); 2979 2980 while (stk.is_nonempty()) { 2981 Node* m = stk.node(); 2982 uint idx = stk.index(); 2983 if (idx < m->outcnt()) { 2984 stk.set_index(idx + 1); 2985 Node* n = m->raw_out(idx); 2986 if (n->is_CFG() && !visited.test_set(n->_idx)) { 2987 stk.push(n, 0); 2988 } 2989 } else { 2990 rpo_list.push(m); 2991 stk.pop(); 2992 } 2993 } 2994 } 2995 #endif 2996 2997 2998 //============================================================================= 2999 //------------------------------LoopTreeIterator----------------------------------- 3000 3001 // Advance to next loop tree using a preorder, left-to-right traversal. 3002 void LoopTreeIterator::next() { 3003 assert(!done(), "must not be done."); 3004 if (_curnt->_child != NULL) { 3005 _curnt = _curnt->_child; 3006 } else if (_curnt->_next != NULL) { 3007 _curnt = _curnt->_next; 3008 } else { 3009 while (_curnt != _root && _curnt->_next == NULL) { 3010 _curnt = _curnt->_parent; 3011 } 3012 if (_curnt == _root) { 3013 _curnt = NULL; 3014 assert(done(), "must be done."); 3015 } else { 3016 assert(_curnt->_next != NULL, "must be more to do"); 3017 _curnt = _curnt->_next; 3018 } 3019 } 3020 }