1 /* 2 * Copyright (c) 1997, 2024, 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 "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/connode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/inlinetypenode.hpp" 37 #include "opto/loopnode.hpp" 38 #include "opto/machnode.hpp" 39 #include "opto/movenode.hpp" 40 #include "opto/narrowptrnode.hpp" 41 #include "opto/mulnode.hpp" 42 #include "opto/phaseX.hpp" 43 #include "opto/regalloc.hpp" 44 #include "opto/regmask.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/subnode.hpp" 47 #include "opto/vectornode.hpp" 48 #include "utilities/vmError.hpp" 49 50 // Portions of code courtesy of Clifford Click 51 52 // Optimization - Graph Style 53 54 //============================================================================= 55 //------------------------------Value------------------------------------------ 56 // Compute the type of the RegionNode. 57 const Type* RegionNode::Value(PhaseGVN* phase) const { 58 for( uint i=1; i<req(); ++i ) { // For all paths in 59 Node *n = in(i); // Get Control source 60 if( !n ) continue; // Missing inputs are TOP 61 if( phase->type(n) == Type::CONTROL ) 62 return Type::CONTROL; 63 } 64 return Type::TOP; // All paths dead? Then so are we 65 } 66 67 //------------------------------Identity--------------------------------------- 68 // Check for Region being Identity. 69 Node* RegionNode::Identity(PhaseGVN* phase) { 70 // Cannot have Region be an identity, even if it has only 1 input. 71 // Phi users cannot have their Region input folded away for them, 72 // since they need to select the proper data input 73 return this; 74 } 75 76 //------------------------------merge_region----------------------------------- 77 // If a Region flows into a Region, merge into one big happy merge. This is 78 // hard to do if there is stuff that has to happen 79 static Node *merge_region(RegionNode *region, PhaseGVN *phase) { 80 if( region->Opcode() != Op_Region ) // Do not do to LoopNodes 81 return nullptr; 82 Node *progress = nullptr; // Progress flag 83 PhaseIterGVN *igvn = phase->is_IterGVN(); 84 85 uint rreq = region->req(); 86 for( uint i = 1; i < rreq; i++ ) { 87 Node *r = region->in(i); 88 if( r && r->Opcode() == Op_Region && // Found a region? 89 r->in(0) == r && // Not already collapsed? 90 r != region && // Avoid stupid situations 91 r->outcnt() == 2 ) { // Self user and 'region' user only? 92 assert(!r->as_Region()->has_phi(), "no phi users"); 93 if( !progress ) { // No progress 94 if (region->has_phi()) { 95 return nullptr; // Only flatten if no Phi users 96 // igvn->hash_delete( phi ); 97 } 98 igvn->hash_delete( region ); 99 progress = region; // Making progress 100 } 101 igvn->hash_delete( r ); 102 103 // Append inputs to 'r' onto 'region' 104 for( uint j = 1; j < r->req(); j++ ) { 105 // Move an input from 'r' to 'region' 106 region->add_req(r->in(j)); 107 r->set_req(j, phase->C->top()); 108 // Update phis of 'region' 109 //for( uint k = 0; k < max; k++ ) { 110 // Node *phi = region->out(k); 111 // if( phi->is_Phi() ) { 112 // phi->add_req(phi->in(i)); 113 // } 114 //} 115 116 rreq++; // One more input to Region 117 } // Found a region to merge into Region 118 igvn->_worklist.push(r); 119 // Clobber pointer to the now dead 'r' 120 region->set_req(i, phase->C->top()); 121 } 122 } 123 124 return progress; 125 } 126 127 128 129 //--------------------------------has_phi-------------------------------------- 130 // Helper function: Return any PhiNode that uses this region or null 131 PhiNode* RegionNode::has_phi() const { 132 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 133 Node* phi = fast_out(i); 134 if (phi->is_Phi()) { // Check for Phi users 135 assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)"); 136 return phi->as_Phi(); // this one is good enough 137 } 138 } 139 140 return nullptr; 141 } 142 143 144 //-----------------------------has_unique_phi---------------------------------- 145 // Helper function: Return the only PhiNode that uses this region or null 146 PhiNode* RegionNode::has_unique_phi() const { 147 // Check that only one use is a Phi 148 PhiNode* only_phi = nullptr; 149 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 150 Node* phi = fast_out(i); 151 if (phi->is_Phi()) { // Check for Phi users 152 assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)"); 153 if (only_phi == nullptr) { 154 only_phi = phi->as_Phi(); 155 } else { 156 return nullptr; // multiple phis 157 } 158 } 159 } 160 161 return only_phi; 162 } 163 164 165 //------------------------------check_phi_clipping----------------------------- 166 // Helper function for RegionNode's identification of FP clipping 167 // Check inputs to the Phi 168 static bool check_phi_clipping( PhiNode *phi, ConNode * &min, uint &min_idx, ConNode * &max, uint &max_idx, Node * &val, uint &val_idx ) { 169 min = nullptr; 170 max = nullptr; 171 val = nullptr; 172 min_idx = 0; 173 max_idx = 0; 174 val_idx = 0; 175 uint phi_max = phi->req(); 176 if( phi_max == 4 ) { 177 for( uint j = 1; j < phi_max; ++j ) { 178 Node *n = phi->in(j); 179 int opcode = n->Opcode(); 180 switch( opcode ) { 181 case Op_ConI: 182 { 183 if( min == nullptr ) { 184 min = n->Opcode() == Op_ConI ? (ConNode*)n : nullptr; 185 min_idx = j; 186 } else { 187 max = n->Opcode() == Op_ConI ? (ConNode*)n : nullptr; 188 max_idx = j; 189 if( min->get_int() > max->get_int() ) { 190 // Swap min and max 191 ConNode *temp; 192 uint temp_idx; 193 temp = min; min = max; max = temp; 194 temp_idx = min_idx; min_idx = max_idx; max_idx = temp_idx; 195 } 196 } 197 } 198 break; 199 default: 200 { 201 val = n; 202 val_idx = j; 203 } 204 break; 205 } 206 } 207 } 208 return ( min && max && val && (min->get_int() <= 0) && (max->get_int() >=0) ); 209 } 210 211 212 //------------------------------check_if_clipping------------------------------ 213 // Helper function for RegionNode's identification of FP clipping 214 // Check that inputs to Region come from two IfNodes, 215 // 216 // If 217 // False True 218 // If | 219 // False True | 220 // | | | 221 // RegionNode_inputs 222 // 223 static bool check_if_clipping( const RegionNode *region, IfNode * &bot_if, IfNode * &top_if ) { 224 top_if = nullptr; 225 bot_if = nullptr; 226 227 // Check control structure above RegionNode for (if ( if ) ) 228 Node *in1 = region->in(1); 229 Node *in2 = region->in(2); 230 Node *in3 = region->in(3); 231 // Check that all inputs are projections 232 if( in1->is_Proj() && in2->is_Proj() && in3->is_Proj() ) { 233 Node *in10 = in1->in(0); 234 Node *in20 = in2->in(0); 235 Node *in30 = in3->in(0); 236 // Check that #1 and #2 are ifTrue and ifFalse from same If 237 if( in10 != nullptr && in10->is_If() && 238 in20 != nullptr && in20->is_If() && 239 in30 != nullptr && in30->is_If() && in10 == in20 && 240 (in1->Opcode() != in2->Opcode()) ) { 241 Node *in100 = in10->in(0); 242 Node *in1000 = (in100 != nullptr && in100->is_Proj()) ? in100->in(0) : nullptr; 243 // Check that control for in10 comes from other branch of IF from in3 244 if( in1000 != nullptr && in1000->is_If() && 245 in30 == in1000 && (in3->Opcode() != in100->Opcode()) ) { 246 // Control pattern checks 247 top_if = (IfNode*)in1000; 248 bot_if = (IfNode*)in10; 249 } 250 } 251 } 252 253 return (top_if != nullptr); 254 } 255 256 257 //------------------------------check_convf2i_clipping------------------------- 258 // Helper function for RegionNode's identification of FP clipping 259 // Verify that the value input to the phi comes from "ConvF2I; LShift; RShift" 260 static bool check_convf2i_clipping( PhiNode *phi, uint idx, ConvF2INode * &convf2i, Node *min, Node *max) { 261 convf2i = nullptr; 262 263 // Check for the RShiftNode 264 Node *rshift = phi->in(idx); 265 assert( rshift, "Previous checks ensure phi input is present"); 266 if( rshift->Opcode() != Op_RShiftI ) { return false; } 267 268 // Check for the LShiftNode 269 Node *lshift = rshift->in(1); 270 assert( lshift, "Previous checks ensure phi input is present"); 271 if( lshift->Opcode() != Op_LShiftI ) { return false; } 272 273 // Check for the ConvF2INode 274 Node *conv = lshift->in(1); 275 if( conv->Opcode() != Op_ConvF2I ) { return false; } 276 277 // Check that shift amounts are only to get sign bits set after F2I 278 jint max_cutoff = max->get_int(); 279 jint min_cutoff = min->get_int(); 280 jint left_shift = lshift->in(2)->get_int(); 281 jint right_shift = rshift->in(2)->get_int(); 282 jint max_post_shift = nth_bit(BitsPerJavaInteger - left_shift - 1); 283 if( left_shift != right_shift || 284 0 > left_shift || left_shift >= BitsPerJavaInteger || 285 max_post_shift < max_cutoff || 286 max_post_shift < -min_cutoff ) { 287 // Shifts are necessary but current transformation eliminates them 288 return false; 289 } 290 291 // OK to return the result of ConvF2I without shifting 292 convf2i = (ConvF2INode*)conv; 293 return true; 294 } 295 296 297 //------------------------------check_compare_clipping------------------------- 298 // Helper function for RegionNode's identification of FP clipping 299 static bool check_compare_clipping( bool less_than, IfNode *iff, ConNode *limit, Node * & input ) { 300 Node *i1 = iff->in(1); 301 if ( !i1->is_Bool() ) { return false; } 302 BoolNode *bool1 = i1->as_Bool(); 303 if( less_than && bool1->_test._test != BoolTest::le ) { return false; } 304 else if( !less_than && bool1->_test._test != BoolTest::lt ) { return false; } 305 const Node *cmpF = bool1->in(1); 306 if( cmpF->Opcode() != Op_CmpF ) { return false; } 307 // Test that the float value being compared against 308 // is equivalent to the int value used as a limit 309 Node *nodef = cmpF->in(2); 310 if( nodef->Opcode() != Op_ConF ) { return false; } 311 jfloat conf = nodef->getf(); 312 jint coni = limit->get_int(); 313 if( ((int)conf) != coni ) { return false; } 314 input = cmpF->in(1); 315 return true; 316 } 317 318 //------------------------------is_unreachable_region-------------------------- 319 // Check if the RegionNode is part of an unsafe loop and unreachable from root. 320 bool RegionNode::is_unreachable_region(const PhaseGVN* phase) { 321 Node* top = phase->C->top(); 322 assert(req() == 2 || (req() == 3 && in(1) != nullptr && in(2) == top), "sanity check arguments"); 323 if (_is_unreachable_region) { 324 // Return cached result from previous evaluation which should still be valid 325 assert(is_unreachable_from_root(phase), "walk the graph again and check if its indeed unreachable"); 326 return true; 327 } 328 329 // First, cut the simple case of fallthrough region when NONE of 330 // region's phis references itself directly or through a data node. 331 if (is_possible_unsafe_loop(phase)) { 332 // If we have a possible unsafe loop, check if the region node is actually unreachable from root. 333 if (is_unreachable_from_root(phase)) { 334 _is_unreachable_region = true; 335 return true; 336 } 337 } 338 return false; 339 } 340 341 bool RegionNode::is_possible_unsafe_loop(const PhaseGVN* phase) const { 342 uint max = outcnt(); 343 uint i; 344 for (i = 0; i < max; i++) { 345 Node* n = raw_out(i); 346 if (n != nullptr && n->is_Phi()) { 347 PhiNode* phi = n->as_Phi(); 348 assert(phi->in(0) == this, "sanity check phi"); 349 if (phi->outcnt() == 0) { 350 continue; // Safe case - no loops 351 } 352 if (phi->outcnt() == 1) { 353 Node* u = phi->raw_out(0); 354 // Skip if only one use is an other Phi or Call or Uncommon trap. 355 // It is safe to consider this case as fallthrough. 356 if (u != nullptr && (u->is_Phi() || u->is_CFG())) { 357 continue; 358 } 359 } 360 // Check when phi references itself directly or through an other node. 361 if (phi->as_Phi()->simple_data_loop_check(phi->in(1)) >= PhiNode::Unsafe) { 362 break; // Found possible unsafe data loop. 363 } 364 } 365 } 366 if (i >= max) { 367 return false; // An unsafe case was NOT found - don't need graph walk. 368 } 369 return true; 370 } 371 372 bool RegionNode::is_unreachable_from_root(const PhaseGVN* phase) const { 373 ResourceMark rm; 374 Node_List nstack; 375 VectorSet visited; 376 377 // Mark all control nodes reachable from root outputs 378 Node* n = (Node*)phase->C->root(); 379 nstack.push(n); 380 visited.set(n->_idx); 381 while (nstack.size() != 0) { 382 n = nstack.pop(); 383 uint max = n->outcnt(); 384 for (uint i = 0; i < max; i++) { 385 Node* m = n->raw_out(i); 386 if (m != nullptr && m->is_CFG()) { 387 if (m == this) { 388 return false; // We reached the Region node - it is not dead. 389 } 390 if (!visited.test_set(m->_idx)) 391 nstack.push(m); 392 } 393 } 394 } 395 return true; // The Region node is unreachable - it is dead. 396 } 397 398 #ifdef ASSERT 399 // Is this region in an infinite subgraph? 400 // (no path to root except through false NeverBranch exit) 401 bool RegionNode::is_in_infinite_subgraph() { 402 ResourceMark rm; 403 Unique_Node_List worklist; 404 worklist.push(this); 405 return RegionNode::are_all_nodes_in_infinite_subgraph(worklist); 406 } 407 408 // Are all nodes in worklist in infinite subgraph? 409 // (no path to root except through false NeverBranch exit) 410 // worklist is directly used for the traversal 411 bool RegionNode::are_all_nodes_in_infinite_subgraph(Unique_Node_List& worklist) { 412 // BFS traversal down the CFG, except through NeverBranch exits 413 for (uint i = 0; i < worklist.size(); ++i) { 414 Node* n = worklist.at(i); 415 assert(n->is_CFG(), "only traverse CFG"); 416 if (n->is_Root()) { 417 // Found root -> there was an exit! 418 return false; 419 } else if (n->is_NeverBranch()) { 420 // Only follow the loop-internal projection, not the NeverBranch exit 421 ProjNode* proj = n->as_NeverBranch()->proj_out_or_null(0); 422 assert(proj != nullptr, "must find loop-internal projection of NeverBranch"); 423 worklist.push(proj); 424 } else { 425 // Traverse all CFG outputs 426 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 427 Node* use = n->fast_out(i); 428 if (use->is_CFG()) { 429 worklist.push(use); 430 } 431 } 432 } 433 } 434 // No exit found for any loop -> all are infinite 435 return true; 436 } 437 #endif //ASSERT 438 439 void RegionNode::set_loop_status(RegionNode::LoopStatus status) { 440 assert(loop_status() == RegionNode::LoopStatus::NeverIrreducibleEntry, "why set our status again?"); 441 _loop_status = status; 442 } 443 444 #ifdef ASSERT 445 void RegionNode::verify_can_be_irreducible_entry() const { 446 assert(loop_status() == RegionNode::LoopStatus::MaybeIrreducibleEntry, "must be marked irreducible"); 447 assert(!is_Loop(), "LoopNode cannot be irreducible loop entry"); 448 } 449 #endif //ASSERT 450 451 void RegionNode::try_clean_mem_phis(PhaseIterGVN* igvn) { 452 // Incremental inlining + PhaseStringOpts sometimes produce: 453 // 454 // cmpP with 1 top input 455 // | 456 // If 457 // / \ 458 // IfFalse IfTrue /- Some Node 459 // \ / / / 460 // Region / /-MergeMem 461 // \---Phi 462 // 463 // 464 // It's expected by PhaseStringOpts that the Region goes away and is 465 // replaced by If's control input but because there's still a Phi, 466 // the Region stays in the graph. The top input from the cmpP is 467 // propagated forward and a subgraph that is useful goes away. The 468 // code in PhiNode::try_clean_memory_phi() replaces the Phi with the 469 // MergeMem in order to remove the Region if its last phi dies. 470 471 if (!is_diamond()) { 472 return; 473 } 474 475 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 476 Node* phi = fast_out(i); 477 if (phi->is_Phi() && phi->as_Phi()->try_clean_memory_phi(igvn)) { 478 --i; 479 --imax; 480 } 481 } 482 } 483 484 // Does this region merge a simple diamond formed by a proper IfNode? 485 // 486 // Cmp 487 // / 488 // ctrl Bool 489 // \ / 490 // IfNode 491 // / \ 492 // IfFalse IfTrue 493 // \ / 494 // Region 495 bool RegionNode::is_diamond() const { 496 if (req() != 3) { 497 return false; 498 } 499 500 Node* left_path = in(1); 501 Node* right_path = in(2); 502 if (left_path == nullptr || right_path == nullptr) { 503 return false; 504 } 505 Node* diamond_if = left_path->in(0); 506 if (diamond_if == nullptr || !diamond_if->is_If() || diamond_if != right_path->in(0)) { 507 // Not an IfNode merging a diamond or TOP. 508 return false; 509 } 510 511 // Check for a proper bool/cmp 512 const Node* bol = diamond_if->in(1); 513 if (!bol->is_Bool()) { 514 return false; 515 } 516 const Node* cmp = bol->in(1); 517 if (!cmp->is_Cmp()) { 518 return false; 519 } 520 return true; 521 } 522 523 //------------------------------Ideal------------------------------------------ 524 // Return a node which is more "ideal" than the current node. Must preserve 525 // the CFG, but we can still strip out dead paths. 526 Node *RegionNode::Ideal(PhaseGVN *phase, bool can_reshape) { 527 if( !can_reshape && !in(0) ) return nullptr; // Already degraded to a Copy 528 assert(!in(0) || !in(0)->is_Root(), "not a specially hidden merge"); 529 530 // Check for RegionNode with no Phi users and both inputs come from either 531 // arm of the same IF. If found, then the control-flow split is useless. 532 bool has_phis = false; 533 if (can_reshape) { // Need DU info to check for Phi users 534 try_clean_mem_phis(phase->is_IterGVN()); 535 has_phis = (has_phi() != nullptr); // Cache result 536 537 if (!has_phis) { // No Phi users? Nothing merging? 538 for (uint i = 1; i < req()-1; i++) { 539 Node *if1 = in(i); 540 if( !if1 ) continue; 541 Node *iff = if1->in(0); 542 if( !iff || !iff->is_If() ) continue; 543 for( uint j=i+1; j<req(); j++ ) { 544 if( in(j) && in(j)->in(0) == iff && 545 if1->Opcode() != in(j)->Opcode() ) { 546 // Add the IF Projections to the worklist. They (and the IF itself) 547 // will be eliminated if dead. 548 phase->is_IterGVN()->add_users_to_worklist(iff); 549 set_req(i, iff->in(0));// Skip around the useless IF diamond 550 set_req(j, nullptr); 551 return this; // Record progress 552 } 553 } 554 } 555 } 556 } 557 558 // Remove TOP or null input paths. If only 1 input path remains, this Region 559 // degrades to a copy. 560 bool add_to_worklist = true; 561 bool modified = false; 562 int cnt = 0; // Count of values merging 563 DEBUG_ONLY( int cnt_orig = req(); ) // Save original inputs count 564 DEBUG_ONLY( uint outcnt_orig = outcnt(); ) 565 int del_it = 0; // The last input path we delete 566 bool found_top = false; // irreducible loops need to check reachability if we find TOP 567 // For all inputs... 568 for( uint i=1; i<req(); ++i ){// For all paths in 569 Node *n = in(i); // Get the input 570 if( n != nullptr ) { 571 // Remove useless control copy inputs 572 if( n->is_Region() && n->as_Region()->is_copy() ) { 573 set_req(i, n->nonnull_req()); 574 modified = true; 575 i--; 576 continue; 577 } 578 if( n->is_Proj() ) { // Remove useless rethrows 579 Node *call = n->in(0); 580 if (call->is_Call() && call->as_Call()->entry_point() == OptoRuntime::rethrow_stub()) { 581 set_req(i, call->in(0)); 582 modified = true; 583 i--; 584 continue; 585 } 586 } 587 if( phase->type(n) == Type::TOP ) { 588 set_req_X(i, nullptr, phase); // Ignore TOP inputs 589 modified = true; 590 found_top = true; 591 i--; 592 continue; 593 } 594 cnt++; // One more value merging 595 } else if (can_reshape) { // Else found dead path with DU info 596 PhaseIterGVN *igvn = phase->is_IterGVN(); 597 del_req(i); // Yank path from self 598 del_it = i; 599 600 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 601 Node* use = fast_out(j); 602 603 if (use->req() != req() && use->is_Phi()) { 604 assert(use->in(0) == this, "unexpected control input"); 605 igvn->hash_delete(use); // Yank from hash before hacking edges 606 use->set_req_X(i, nullptr, igvn);// Correct DU info 607 use->del_req(i); // Yank path from Phis 608 } 609 } 610 611 if (add_to_worklist) { 612 igvn->add_users_to_worklist(this); 613 add_to_worklist = false; 614 } 615 616 i--; 617 } 618 } 619 620 assert(outcnt() == outcnt_orig, "not expect to remove any use"); 621 622 if (can_reshape && found_top && loop_status() == RegionNode::LoopStatus::MaybeIrreducibleEntry) { 623 // Is it a dead irreducible loop? 624 // If an irreducible loop loses one of the multiple entries 625 // that went into the loop head, or any secondary entries, 626 // we need to verify if the irreducible loop is still reachable, 627 // as the special logic in is_unreachable_region only works 628 // for reducible loops. 629 if (is_unreachable_from_root(phase)) { 630 // The irreducible loop is dead - must remove it 631 PhaseIterGVN* igvn = phase->is_IterGVN(); 632 remove_unreachable_subgraph(igvn); 633 return nullptr; 634 } 635 } else if (can_reshape && cnt == 1) { 636 // Is it dead loop? 637 // If it is LoopNopde it had 2 (+1 itself) inputs and 638 // one of them was cut. The loop is dead if it was EntryContol. 639 // Loop node may have only one input because entry path 640 // is removed in PhaseIdealLoop::Dominators(). 641 assert(!this->is_Loop() || cnt_orig <= 3, "Loop node should have 3 or less inputs"); 642 if ((this->is_Loop() && (del_it == LoopNode::EntryControl || 643 (del_it == 0 && is_unreachable_region(phase)))) || 644 (!this->is_Loop() && has_phis && is_unreachable_region(phase))) { 645 PhaseIterGVN* igvn = phase->is_IterGVN(); 646 remove_unreachable_subgraph(igvn); 647 return nullptr; 648 } 649 } 650 651 if( cnt <= 1 ) { // Only 1 path in? 652 set_req(0, nullptr); // Null control input for region copy 653 if( cnt == 0 && !can_reshape) { // Parse phase - leave the node as it is. 654 // No inputs or all inputs are null. 655 return nullptr; 656 } else if (can_reshape) { // Optimization phase - remove the node 657 PhaseIterGVN *igvn = phase->is_IterGVN(); 658 // Strip mined (inner) loop is going away, remove outer loop. 659 if (is_CountedLoop() && 660 as_Loop()->is_strip_mined()) { 661 Node* outer_sfpt = as_CountedLoop()->outer_safepoint(); 662 Node* outer_out = as_CountedLoop()->outer_loop_exit(); 663 if (outer_sfpt != nullptr && outer_out != nullptr) { 664 Node* in = outer_sfpt->in(0); 665 igvn->replace_node(outer_out, in); 666 LoopNode* outer = as_CountedLoop()->outer_loop(); 667 igvn->replace_input_of(outer, LoopNode::LoopBackControl, igvn->C->top()); 668 } 669 } 670 if (is_CountedLoop()) { 671 Node* opaq = as_CountedLoop()->is_canonical_loop_entry(); 672 if (opaq != nullptr) { 673 // This is not a loop anymore. No need to keep the Opaque1 node on the test that guards the loop as it won't be 674 // subject to further loop opts. 675 assert(opaq->Opcode() == Op_OpaqueZeroTripGuard, ""); 676 igvn->replace_node(opaq, opaq->in(1)); 677 } 678 } 679 Node *parent_ctrl; 680 if( cnt == 0 ) { 681 assert( req() == 1, "no inputs expected" ); 682 // During IGVN phase such region will be subsumed by TOP node 683 // so region's phis will have TOP as control node. 684 // Kill phis here to avoid it. 685 // Also set other user's input to top. 686 parent_ctrl = phase->C->top(); 687 } else { 688 // The fallthrough case since we already checked dead loops above. 689 parent_ctrl = in(1); 690 assert(parent_ctrl != nullptr, "Region is a copy of some non-null control"); 691 assert(parent_ctrl != this, "Close dead loop"); 692 } 693 if (add_to_worklist) { 694 igvn->add_users_to_worklist(this); // Check for further allowed opts 695 } 696 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) { 697 Node* n = last_out(i); 698 igvn->hash_delete(n); // Remove from worklist before modifying edges 699 if (n->outcnt() == 0) { 700 int uses_found = n->replace_edge(this, phase->C->top(), igvn); 701 if (uses_found > 1) { // (--i) done at the end of the loop. 702 i -= (uses_found - 1); 703 } 704 continue; 705 } 706 if( n->is_Phi() ) { // Collapse all Phis 707 // Eagerly replace phis to avoid regionless phis. 708 Node* in; 709 if( cnt == 0 ) { 710 assert( n->req() == 1, "No data inputs expected" ); 711 in = parent_ctrl; // replaced by top 712 } else { 713 assert( n->req() == 2 && n->in(1) != nullptr, "Only one data input expected" ); 714 in = n->in(1); // replaced by unique input 715 if( n->as_Phi()->is_unsafe_data_reference(in) ) 716 in = phase->C->top(); // replaced by top 717 } 718 igvn->replace_node(n, in); 719 } 720 else if( n->is_Region() ) { // Update all incoming edges 721 assert(n != this, "Must be removed from DefUse edges"); 722 int uses_found = n->replace_edge(this, parent_ctrl, igvn); 723 if (uses_found > 1) { // (--i) done at the end of the loop. 724 i -= (uses_found - 1); 725 } 726 } 727 else { 728 assert(n->in(0) == this, "Expect RegionNode to be control parent"); 729 n->set_req(0, parent_ctrl); 730 } 731 #ifdef ASSERT 732 for( uint k=0; k < n->req(); k++ ) { 733 assert(n->in(k) != this, "All uses of RegionNode should be gone"); 734 } 735 #endif 736 } 737 // Remove the RegionNode itself from DefUse info 738 igvn->remove_dead_node(this); 739 return nullptr; 740 } 741 return this; // Record progress 742 } 743 744 745 // If a Region flows into a Region, merge into one big happy merge. 746 if (can_reshape) { 747 Node *m = merge_region(this, phase); 748 if (m != nullptr) return m; 749 } 750 751 // Check if this region is the root of a clipping idiom on floats 752 if( ConvertFloat2IntClipping && can_reshape && req() == 4 ) { 753 // Check that only one use is a Phi and that it simplifies to two constants + 754 PhiNode* phi = has_unique_phi(); 755 if (phi != nullptr) { // One Phi user 756 // Check inputs to the Phi 757 ConNode *min; 758 ConNode *max; 759 Node *val; 760 uint min_idx; 761 uint max_idx; 762 uint val_idx; 763 if( check_phi_clipping( phi, min, min_idx, max, max_idx, val, val_idx ) ) { 764 IfNode *top_if; 765 IfNode *bot_if; 766 if( check_if_clipping( this, bot_if, top_if ) ) { 767 // Control pattern checks, now verify compares 768 Node *top_in = nullptr; // value being compared against 769 Node *bot_in = nullptr; 770 if( check_compare_clipping( true, bot_if, min, bot_in ) && 771 check_compare_clipping( false, top_if, max, top_in ) ) { 772 if( bot_in == top_in ) { 773 PhaseIterGVN *gvn = phase->is_IterGVN(); 774 assert( gvn != nullptr, "Only had DefUse info in IterGVN"); 775 // Only remaining check is that bot_in == top_in == (Phi's val + mods) 776 777 // Check for the ConvF2INode 778 ConvF2INode *convf2i; 779 if( check_convf2i_clipping( phi, val_idx, convf2i, min, max ) && 780 convf2i->in(1) == bot_in ) { 781 // Matched pattern, including LShiftI; RShiftI, replace with integer compares 782 // max test 783 Node *cmp = gvn->register_new_node_with_optimizer(new CmpINode( convf2i, min )); 784 Node *boo = gvn->register_new_node_with_optimizer(new BoolNode( cmp, BoolTest::lt )); 785 IfNode *iff = (IfNode*)gvn->register_new_node_with_optimizer(new IfNode( top_if->in(0), boo, PROB_UNLIKELY_MAG(5), top_if->_fcnt )); 786 Node *if_min= gvn->register_new_node_with_optimizer(new IfTrueNode (iff)); 787 Node *ifF = gvn->register_new_node_with_optimizer(new IfFalseNode(iff)); 788 // min test 789 cmp = gvn->register_new_node_with_optimizer(new CmpINode( convf2i, max )); 790 boo = gvn->register_new_node_with_optimizer(new BoolNode( cmp, BoolTest::gt )); 791 iff = (IfNode*)gvn->register_new_node_with_optimizer(new IfNode( ifF, boo, PROB_UNLIKELY_MAG(5), bot_if->_fcnt )); 792 Node *if_max= gvn->register_new_node_with_optimizer(new IfTrueNode (iff)); 793 ifF = gvn->register_new_node_with_optimizer(new IfFalseNode(iff)); 794 // update input edges to region node 795 set_req_X( min_idx, if_min, gvn ); 796 set_req_X( max_idx, if_max, gvn ); 797 set_req_X( val_idx, ifF, gvn ); 798 // remove unnecessary 'LShiftI; RShiftI' idiom 799 gvn->hash_delete(phi); 800 phi->set_req_X( val_idx, convf2i, gvn ); 801 gvn->hash_find_insert(phi); 802 // Return transformed region node 803 return this; 804 } 805 } 806 } 807 } 808 } 809 } 810 } 811 812 if (can_reshape) { 813 modified |= optimize_trichotomy(phase->is_IterGVN()); 814 } 815 816 return modified ? this : nullptr; 817 } 818 819 //--------------------------remove_unreachable_subgraph---------------------- 820 // This region and therefore all nodes on the input control path(s) are unreachable 821 // from root. To avoid incomplete removal of unreachable subgraphs, walk up the CFG 822 // and aggressively replace all nodes by top. 823 // If a control node "def" with a single control output "use" has its single output 824 // "use" replaced with top, then "use" removes itself. This has the consequence that 825 // when we visit "use", it already has all inputs removed. They are lost and we cannot 826 // traverse them. This is why we fist find all unreachable nodes, and then remove 827 // them in a second step. 828 void RegionNode::remove_unreachable_subgraph(PhaseIterGVN* igvn) { 829 Node* top = igvn->C->top(); 830 ResourceMark rm; 831 Unique_Node_List unreachable; // visit each only once 832 unreachable.push(this); 833 // Recursively find all control inputs. 834 for (uint i = 0; i < unreachable.size(); i++) { 835 Node* n = unreachable.at(i); 836 for (uint i = 0; i < n->req(); ++i) { 837 Node* m = n->in(i); 838 assert(m == nullptr || !m->is_Root(), "Should be unreachable from root"); 839 if (m != nullptr && m->is_CFG()) { 840 unreachable.push(m); 841 } 842 } 843 } 844 // Remove all unreachable nodes. 845 for (uint i = 0; i < unreachable.size(); i++) { 846 Node* n = unreachable.at(i); 847 if (n->is_Region()) { 848 // Eagerly replace phis with top to avoid regionless phis. 849 n->set_req(0, nullptr); 850 bool progress = true; 851 uint max = n->outcnt(); 852 DUIterator j; 853 while (progress) { 854 progress = false; 855 for (j = n->outs(); n->has_out(j); j++) { 856 Node* u = n->out(j); 857 if (u->is_Phi()) { 858 igvn->replace_node(u, top); 859 if (max != n->outcnt()) { 860 progress = true; 861 j = n->refresh_out_pos(j); 862 max = n->outcnt(); 863 } 864 } 865 } 866 } 867 } 868 igvn->replace_node(n, top); 869 } 870 } 871 872 //------------------------------optimize_trichotomy-------------------------- 873 // Optimize nested comparisons of the following kind: 874 // 875 // int compare(int a, int b) { 876 // return (a < b) ? -1 : (a == b) ? 0 : 1; 877 // } 878 // 879 // Shape 1: 880 // if (compare(a, b) == 1) { ... } -> if (a > b) { ... } 881 // 882 // Shape 2: 883 // if (compare(a, b) == 0) { ... } -> if (a == b) { ... } 884 // 885 // Above code leads to the following IR shapes where both Ifs compare the 886 // same value and two out of three region inputs idx1 and idx2 map to 887 // the same value and control flow. 888 // 889 // (1) If (2) If 890 // / \ / \ 891 // Proj Proj Proj Proj 892 // | \ | \ 893 // | If | If If 894 // | / \ | / \ / \ 895 // | Proj Proj | Proj Proj ==> Proj Proj 896 // | / / \ | / | / 897 // Region / \ | / | / 898 // \ / \ | / | / 899 // Region Region Region 900 // 901 // The method returns true if 'this' is modified and false otherwise. 902 bool RegionNode::optimize_trichotomy(PhaseIterGVN* igvn) { 903 int idx1 = 1, idx2 = 2; 904 Node* region = nullptr; 905 if (req() == 3 && in(1) != nullptr && in(2) != nullptr) { 906 // Shape 1: Check if one of the inputs is a region that merges two control 907 // inputs and has no other users (especially no Phi users). 908 region = in(1)->isa_Region() ? in(1) : in(2)->isa_Region(); 909 if (region == nullptr || region->outcnt() != 2 || region->req() != 3) { 910 return false; // No suitable region input found 911 } 912 } else if (req() == 4) { 913 // Shape 2: Check if two control inputs map to the same value of the unique phi 914 // user and treat these as if they would come from another region (shape (1)). 915 PhiNode* phi = has_unique_phi(); 916 if (phi == nullptr) { 917 return false; // No unique phi user 918 } 919 if (phi->in(idx1) != phi->in(idx2)) { 920 idx2 = 3; 921 if (phi->in(idx1) != phi->in(idx2)) { 922 idx1 = 2; 923 if (phi->in(idx1) != phi->in(idx2)) { 924 return false; // No equal phi inputs found 925 } 926 } 927 } 928 assert(phi->in(idx1) == phi->in(idx2), "must be"); // Region is merging same value 929 region = this; 930 } 931 if (region == nullptr || region->in(idx1) == nullptr || region->in(idx2) == nullptr) { 932 return false; // Region does not merge two control inputs 933 } 934 // At this point we know that region->in(idx1) and region->(idx2) map to the same 935 // value and control flow. Now search for ifs that feed into these region inputs. 936 ProjNode* proj1 = region->in(idx1)->isa_Proj(); 937 ProjNode* proj2 = region->in(idx2)->isa_Proj(); 938 if (proj1 == nullptr || proj1->outcnt() != 1 || 939 proj2 == nullptr || proj2->outcnt() != 1) { 940 return false; // No projection inputs with region as unique user found 941 } 942 assert(proj1 != proj2, "should be different projections"); 943 IfNode* iff1 = proj1->in(0)->isa_If(); 944 IfNode* iff2 = proj2->in(0)->isa_If(); 945 if (iff1 == nullptr || iff1->outcnt() != 2 || 946 iff2 == nullptr || iff2->outcnt() != 2) { 947 return false; // No ifs found 948 } 949 if (iff1 == iff2) { 950 igvn->add_users_to_worklist(iff1); // Make sure dead if is eliminated 951 igvn->replace_input_of(region, idx1, iff1->in(0)); 952 igvn->replace_input_of(region, idx2, igvn->C->top()); 953 return (region == this); // Remove useless if (both projections map to the same control/value) 954 } 955 BoolNode* bol1 = iff1->in(1)->isa_Bool(); 956 BoolNode* bol2 = iff2->in(1)->isa_Bool(); 957 if (bol1 == nullptr || bol2 == nullptr) { 958 return false; // No bool inputs found 959 } 960 Node* cmp1 = bol1->in(1); 961 Node* cmp2 = bol2->in(1); 962 bool commute = false; 963 if (!cmp1->is_Cmp() || !cmp2->is_Cmp()) { 964 return false; // No comparison 965 } else if (cmp1->Opcode() == Op_CmpF || cmp1->Opcode() == Op_CmpD || 966 cmp2->Opcode() == Op_CmpF || cmp2->Opcode() == Op_CmpD || 967 cmp1->Opcode() == Op_CmpP || cmp1->Opcode() == Op_CmpN || 968 cmp2->Opcode() == Op_CmpP || cmp2->Opcode() == Op_CmpN || 969 cmp1->is_SubTypeCheck() || cmp2->is_SubTypeCheck() || 970 cmp1->is_FlatArrayCheck() || cmp2->is_FlatArrayCheck()) { 971 // Floats and pointers don't exactly obey trichotomy. To be on the safe side, don't transform their tests. 972 // SubTypeCheck is not commutative 973 return false; 974 } else if (cmp1 != cmp2) { 975 if (cmp1->in(1) == cmp2->in(2) && 976 cmp1->in(2) == cmp2->in(1)) { 977 commute = true; // Same but swapped inputs, commute the test 978 } else { 979 return false; // Ifs are not comparing the same values 980 } 981 } 982 proj1 = proj1->other_if_proj(); 983 proj2 = proj2->other_if_proj(); 984 if (!((proj1->unique_ctrl_out_or_null() == iff2 && 985 proj2->unique_ctrl_out_or_null() == this) || 986 (proj2->unique_ctrl_out_or_null() == iff1 && 987 proj1->unique_ctrl_out_or_null() == this))) { 988 return false; // Ifs are not connected through other projs 989 } 990 // Found 'iff -> proj -> iff -> proj -> this' shape where all other projs are merged 991 // through 'region' and map to the same value. Merge the boolean tests and replace 992 // the ifs by a single comparison. 993 BoolTest test1 = (proj1->_con == 1) ? bol1->_test : bol1->_test.negate(); 994 BoolTest test2 = (proj2->_con == 1) ? bol2->_test : bol2->_test.negate(); 995 test1 = commute ? test1.commute() : test1; 996 // After possibly commuting test1, if we can merge test1 & test2, then proj2/iff2/bol2 are the nodes to refine. 997 BoolTest::mask res = test1.merge(test2); 998 if (res == BoolTest::illegal) { 999 return false; // Unable to merge tests 1000 } 1001 // Adjust iff1 to always pass (only iff2 will remain) 1002 igvn->replace_input_of(iff1, 1, igvn->intcon(proj1->_con)); 1003 if (res == BoolTest::never) { 1004 // Merged test is always false, adjust iff2 to always fail 1005 igvn->replace_input_of(iff2, 1, igvn->intcon(1 - proj2->_con)); 1006 } else { 1007 // Replace bool input of iff2 with merged test 1008 BoolNode* new_bol = new BoolNode(bol2->in(1), res); 1009 igvn->replace_input_of(iff2, 1, igvn->transform((proj2->_con == 1) ? new_bol : new_bol->negate(igvn))); 1010 if (new_bol->outcnt() == 0) { 1011 igvn->remove_dead_node(new_bol); 1012 } 1013 } 1014 return false; 1015 } 1016 1017 const RegMask &RegionNode::out_RegMask() const { 1018 return RegMask::Empty; 1019 } 1020 1021 #ifndef PRODUCT 1022 void RegionNode::dump_spec(outputStream* st) const { 1023 Node::dump_spec(st); 1024 switch (loop_status()) { 1025 case RegionNode::LoopStatus::MaybeIrreducibleEntry: 1026 st->print("#irreducible "); 1027 break; 1028 case RegionNode::LoopStatus::Reducible: 1029 st->print("#reducible "); 1030 break; 1031 case RegionNode::LoopStatus::NeverIrreducibleEntry: 1032 break; // nothing 1033 } 1034 } 1035 #endif 1036 1037 // Find the one non-null required input. RegionNode only 1038 Node *Node::nonnull_req() const { 1039 assert( is_Region(), "" ); 1040 for( uint i = 1; i < _cnt; i++ ) 1041 if( in(i) ) 1042 return in(i); 1043 ShouldNotReachHere(); 1044 return nullptr; 1045 } 1046 1047 1048 //============================================================================= 1049 // note that these functions assume that the _adr_type field is flat 1050 uint PhiNode::hash() const { 1051 const Type* at = _adr_type; 1052 return TypeNode::hash() + (at ? at->hash() : 0); 1053 } 1054 bool PhiNode::cmp( const Node &n ) const { 1055 return TypeNode::cmp(n) && _adr_type == ((PhiNode&)n)._adr_type; 1056 } 1057 static inline 1058 const TypePtr* flatten_phi_adr_type(const TypePtr* at) { 1059 if (at == nullptr || at == TypePtr::BOTTOM) return at; 1060 return Compile::current()->alias_type(at)->adr_type(); 1061 } 1062 1063 //----------------------------make--------------------------------------------- 1064 // create a new phi with edges matching r and set (initially) to x 1065 PhiNode* PhiNode::make(Node* r, Node* x, const Type *t, const TypePtr* at) { 1066 uint preds = r->req(); // Number of predecessor paths 1067 assert(t != Type::MEMORY || at == flatten_phi_adr_type(at) || (flatten_phi_adr_type(at) == TypeAryPtr::INLINES && Compile::current()->flat_accesses_share_alias()), "flatten at"); 1068 PhiNode* p = new PhiNode(r, t, at); 1069 for (uint j = 1; j < preds; j++) { 1070 // Fill in all inputs, except those which the region does not yet have 1071 if (r->in(j) != nullptr) 1072 p->init_req(j, x); 1073 } 1074 return p; 1075 } 1076 PhiNode* PhiNode::make(Node* r, Node* x) { 1077 const Type* t = x->bottom_type(); 1078 const TypePtr* at = nullptr; 1079 if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type()); 1080 return make(r, x, t, at); 1081 } 1082 PhiNode* PhiNode::make_blank(Node* r, Node* x) { 1083 const Type* t = x->bottom_type(); 1084 const TypePtr* at = nullptr; 1085 if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type()); 1086 return new PhiNode(r, t, at); 1087 } 1088 1089 1090 //------------------------slice_memory----------------------------------------- 1091 // create a new phi with narrowed memory type 1092 PhiNode* PhiNode::slice_memory(const TypePtr* adr_type) const { 1093 PhiNode* mem = (PhiNode*) clone(); 1094 *(const TypePtr**)&mem->_adr_type = adr_type; 1095 // convert self-loops, or else we get a bad graph 1096 for (uint i = 1; i < req(); i++) { 1097 if ((const Node*)in(i) == this) mem->set_req(i, mem); 1098 } 1099 mem->verify_adr_type(); 1100 return mem; 1101 } 1102 1103 //------------------------split_out_instance----------------------------------- 1104 // Split out an instance type from a bottom phi. 1105 PhiNode* PhiNode::split_out_instance(const TypePtr* at, PhaseIterGVN *igvn) const { 1106 const TypeOopPtr *t_oop = at->isa_oopptr(); 1107 assert(t_oop != nullptr && t_oop->is_known_instance(), "expecting instance oopptr"); 1108 1109 // Check if an appropriate node already exists. 1110 Node *region = in(0); 1111 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 1112 Node* use = region->fast_out(k); 1113 if( use->is_Phi()) { 1114 PhiNode *phi2 = use->as_Phi(); 1115 if (phi2->type() == Type::MEMORY && phi2->adr_type() == at) { 1116 return phi2; 1117 } 1118 } 1119 } 1120 Compile *C = igvn->C; 1121 Node_Array node_map; 1122 Node_Stack stack(C->live_nodes() >> 4); 1123 PhiNode *nphi = slice_memory(at); 1124 igvn->register_new_node_with_optimizer( nphi ); 1125 node_map.map(_idx, nphi); 1126 stack.push((Node *)this, 1); 1127 while(!stack.is_empty()) { 1128 PhiNode *ophi = stack.node()->as_Phi(); 1129 uint i = stack.index(); 1130 assert(i >= 1, "not control edge"); 1131 stack.pop(); 1132 nphi = node_map[ophi->_idx]->as_Phi(); 1133 for (; i < ophi->req(); i++) { 1134 Node *in = ophi->in(i); 1135 if (in == nullptr || igvn->type(in) == Type::TOP) 1136 continue; 1137 Node *opt = MemNode::optimize_simple_memory_chain(in, t_oop, nullptr, igvn); 1138 PhiNode *optphi = opt->is_Phi() ? opt->as_Phi() : nullptr; 1139 if (optphi != nullptr && optphi->adr_type() == TypePtr::BOTTOM) { 1140 opt = node_map[optphi->_idx]; 1141 if (opt == nullptr) { 1142 stack.push(ophi, i); 1143 nphi = optphi->slice_memory(at); 1144 igvn->register_new_node_with_optimizer( nphi ); 1145 node_map.map(optphi->_idx, nphi); 1146 ophi = optphi; 1147 i = 0; // will get incremented at top of loop 1148 continue; 1149 } 1150 } 1151 nphi->set_req(i, opt); 1152 } 1153 } 1154 return nphi; 1155 } 1156 1157 //------------------------verify_adr_type-------------------------------------- 1158 #ifdef ASSERT 1159 void PhiNode::verify_adr_type(VectorSet& visited, const TypePtr* at) const { 1160 if (visited.test_set(_idx)) return; //already visited 1161 1162 // recheck constructor invariants: 1163 verify_adr_type(false); 1164 1165 // recheck local phi/phi consistency: 1166 assert(_adr_type == at || _adr_type == TypePtr::BOTTOM, 1167 "adr_type must be consistent across phi nest"); 1168 1169 // walk around 1170 for (uint i = 1; i < req(); i++) { 1171 Node* n = in(i); 1172 if (n == nullptr) continue; 1173 const Node* np = in(i); 1174 if (np->is_Phi()) { 1175 np->as_Phi()->verify_adr_type(visited, at); 1176 } else if (n->bottom_type() == Type::TOP 1177 || (n->is_Mem() && n->in(MemNode::Address)->bottom_type() == Type::TOP)) { 1178 // ignore top inputs 1179 } else { 1180 const TypePtr* nat = flatten_phi_adr_type(n->adr_type()); 1181 // recheck phi/non-phi consistency at leaves: 1182 assert((nat != nullptr) == (at != nullptr), ""); 1183 assert(nat == at || nat == TypePtr::BOTTOM, 1184 "adr_type must be consistent at leaves of phi nest"); 1185 } 1186 } 1187 } 1188 1189 // Verify a whole nest of phis rooted at this one. 1190 void PhiNode::verify_adr_type(bool recursive) const { 1191 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 1192 if (Node::in_dump()) return; // muzzle asserts when printing 1193 1194 assert((_type == Type::MEMORY) == (_adr_type != nullptr), "adr_type for memory phis only"); 1195 // Flat array element shouldn't get their own memory slice until flat_accesses_share_alias is cleared. 1196 // It could be the graph has no loads/stores and flat_accesses_share_alias is never cleared. EA could still 1197 // creates per element Phis but that wouldn't be a problem as there are no memory accesses for that array. 1198 assert(_adr_type == nullptr || _adr_type->isa_aryptr() == nullptr || 1199 _adr_type->is_aryptr()->is_known_instance() || 1200 !_adr_type->is_aryptr()->is_flat() || 1201 !Compile::current()->flat_accesses_share_alias() || 1202 _adr_type == TypeAryPtr::INLINES, "flat array element shouldn't get its own slice yet"); 1203 1204 if (!VerifyAliases) return; // verify thoroughly only if requested 1205 1206 assert(_adr_type == flatten_phi_adr_type(_adr_type), 1207 "Phi::adr_type must be pre-normalized"); 1208 1209 if (recursive) { 1210 VectorSet visited; 1211 verify_adr_type(visited, _adr_type); 1212 } 1213 } 1214 #endif 1215 1216 1217 //------------------------------Value------------------------------------------ 1218 // Compute the type of the PhiNode 1219 const Type* PhiNode::Value(PhaseGVN* phase) const { 1220 Node *r = in(0); // RegionNode 1221 if( !r ) // Copy or dead 1222 return in(1) ? phase->type(in(1)) : Type::TOP; 1223 1224 // Note: During parsing, phis are often transformed before their regions. 1225 // This means we have to use type_or_null to defend against untyped regions. 1226 if( phase->type_or_null(r) == Type::TOP ) // Dead code? 1227 return Type::TOP; 1228 1229 // Check for trip-counted loop. If so, be smarter. 1230 BaseCountedLoopNode* l = r->is_BaseCountedLoop() ? r->as_BaseCountedLoop() : nullptr; 1231 if (l && ((const Node*)l->phi() == this)) { // Trip counted loop! 1232 // protect against init_trip() or limit() returning null 1233 if (l->can_be_counted_loop(phase)) { 1234 const Node* init = l->init_trip(); 1235 const Node* limit = l->limit(); 1236 const Node* stride = l->stride(); 1237 if (init != nullptr && limit != nullptr && stride != nullptr) { 1238 const TypeInteger* lo = phase->type(init)->isa_integer(l->bt()); 1239 const TypeInteger* hi = phase->type(limit)->isa_integer(l->bt()); 1240 const TypeInteger* stride_t = phase->type(stride)->isa_integer(l->bt()); 1241 if (lo != nullptr && hi != nullptr && stride_t != nullptr) { // Dying loops might have TOP here 1242 assert(stride_t->is_con(), "bad stride type"); 1243 BoolTest::mask bt = l->loopexit()->test_trip(); 1244 // If the loop exit condition is "not equal", the condition 1245 // would not trigger if init > limit (if stride > 0) or if 1246 // init < limit if (stride > 0) so we can't deduce bounds 1247 // for the iv from the exit condition. 1248 if (bt != BoolTest::ne) { 1249 jlong stride_con = stride_t->get_con_as_long(l->bt()); 1250 if (stride_con < 0) { // Down-counter loop 1251 swap(lo, hi); 1252 jlong iv_range_lower_limit = lo->lo_as_long(); 1253 // Prevent overflow when adding one below 1254 if (iv_range_lower_limit < max_signed_integer(l->bt())) { 1255 // The loop exit condition is: iv + stride > limit (iv is this Phi). So the loop iterates until 1256 // iv + stride <= limit 1257 // We know that: limit >= lo->lo_as_long() and stride <= -1 1258 // So when the loop exits, iv has to be at most lo->lo_as_long() + 1 1259 iv_range_lower_limit += 1; // lo is after decrement 1260 // Exact bounds for the phi can be computed when ABS(stride) greater than 1 if bounds are constant. 1261 if (lo->is_con() && hi->is_con() && hi->lo_as_long() > lo->hi_as_long() && stride_con != -1) { 1262 julong uhi = static_cast<julong>(hi->lo_as_long()); 1263 julong ulo = static_cast<julong>(lo->hi_as_long()); 1264 julong diff = ((uhi - ulo - 1) / (-stride_con)) * (-stride_con); 1265 julong ufirst = hi->lo_as_long() - diff; 1266 iv_range_lower_limit = reinterpret_cast<jlong &>(ufirst); 1267 assert(iv_range_lower_limit >= lo->lo_as_long() + 1, "should end up with narrower range"); 1268 } 1269 } 1270 return TypeInteger::make(MIN2(iv_range_lower_limit, hi->lo_as_long()), hi->hi_as_long(), 3, l->bt())->filter_speculative(_type); 1271 } else if (stride_con >= 0) { 1272 jlong iv_range_upper_limit = hi->hi_as_long(); 1273 // Prevent overflow when subtracting one below 1274 if (iv_range_upper_limit > min_signed_integer(l->bt())) { 1275 // The loop exit condition is: iv + stride < limit (iv is this Phi). So the loop iterates until 1276 // iv + stride >= limit 1277 // We know that: limit <= hi->hi_as_long() and stride >= 1 1278 // So when the loop exits, iv has to be at most hi->hi_as_long() - 1 1279 iv_range_upper_limit -= 1; 1280 // Exact bounds for the phi can be computed when ABS(stride) greater than 1 if bounds are constant. 1281 if (lo->is_con() && hi->is_con() && hi->lo_as_long() > lo->hi_as_long() && stride_con != 1) { 1282 julong uhi = static_cast<julong>(hi->lo_as_long()); 1283 julong ulo = static_cast<julong>(lo->hi_as_long()); 1284 julong diff = ((uhi - ulo - 1) / stride_con) * stride_con; 1285 julong ulast = lo->hi_as_long() + diff; 1286 iv_range_upper_limit = reinterpret_cast<jlong &>(ulast); 1287 assert(iv_range_upper_limit <= hi->hi_as_long() - 1, "should end up with narrower range"); 1288 } 1289 } 1290 return TypeInteger::make(lo->lo_as_long(), MAX2(lo->hi_as_long(), iv_range_upper_limit), 3, l->bt())->filter_speculative(_type); 1291 } 1292 } 1293 } 1294 } 1295 } else if (l->in(LoopNode::LoopBackControl) != nullptr && 1296 in(LoopNode::EntryControl) != nullptr && 1297 phase->type(l->in(LoopNode::LoopBackControl)) == Type::TOP) { 1298 // During CCP, if we saturate the type of a counted loop's Phi 1299 // before the special code for counted loop above has a chance 1300 // to run (that is as long as the type of the backedge's control 1301 // is top), we might end up with non monotonic types 1302 return phase->type(in(LoopNode::EntryControl))->filter_speculative(_type); 1303 } 1304 } 1305 1306 // Default case: merge all inputs 1307 const Type *t = Type::TOP; // Merged type starting value 1308 for (uint i = 1; i < req(); ++i) {// For all paths in 1309 // Reachable control path? 1310 if (r->in(i) && phase->type(r->in(i)) == Type::CONTROL) { 1311 const Type* ti = phase->type(in(i)); 1312 t = t->meet_speculative(ti); 1313 } 1314 } 1315 1316 // The worst-case type (from ciTypeFlow) should be consistent with "t". 1317 // That is, we expect that "t->higher_equal(_type)" holds true. 1318 // There are various exceptions: 1319 // - Inputs which are phis might in fact be widened unnecessarily. 1320 // For example, an input might be a widened int while the phi is a short. 1321 // - Inputs might be BotPtrs but this phi is dependent on a null check, 1322 // and postCCP has removed the cast which encodes the result of the check. 1323 // - The type of this phi is an interface, and the inputs are classes. 1324 // - Value calls on inputs might produce fuzzy results. 1325 // (Occurrences of this case suggest improvements to Value methods.) 1326 // 1327 // It is not possible to see Type::BOTTOM values as phi inputs, 1328 // because the ciTypeFlow pre-pass produces verifier-quality types. 1329 const Type* ft = t->filter_speculative(_type); // Worst case type 1330 1331 #ifdef ASSERT 1332 // The following logic has been moved into TypeOopPtr::filter. 1333 const Type* jt = t->join_speculative(_type); 1334 if (jt->empty()) { // Emptied out??? 1335 // Otherwise it's something stupid like non-overlapping int ranges 1336 // found on dying counted loops. 1337 assert(ft == Type::TOP, ""); // Canonical empty value 1338 } 1339 1340 else { 1341 1342 if (jt != ft && jt->base() == ft->base()) { 1343 if (jt->isa_int() && 1344 jt->is_int()->_lo == ft->is_int()->_lo && 1345 jt->is_int()->_hi == ft->is_int()->_hi) 1346 jt = ft; 1347 if (jt->isa_long() && 1348 jt->is_long()->_lo == ft->is_long()->_lo && 1349 jt->is_long()->_hi == ft->is_long()->_hi) 1350 jt = ft; 1351 } 1352 if (jt != ft) { 1353 tty->print("merge type: "); t->dump(); tty->cr(); 1354 tty->print("kill type: "); _type->dump(); tty->cr(); 1355 tty->print("join type: "); jt->dump(); tty->cr(); 1356 tty->print("filter type: "); ft->dump(); tty->cr(); 1357 } 1358 assert(jt == ft, ""); 1359 } 1360 #endif //ASSERT 1361 1362 // Deal with conversion problems found in data loops. 1363 ft = phase->saturate_and_maybe_push_to_igvn_worklist(this, ft); 1364 return ft; 1365 } 1366 1367 // Does this Phi represent a simple well-shaped diamond merge? Return the 1368 // index of the true path or 0 otherwise. 1369 int PhiNode::is_diamond_phi() const { 1370 Node* region = in(0); 1371 assert(region != nullptr && region->is_Region(), "phi must have region"); 1372 if (!region->as_Region()->is_diamond()) { 1373 return 0; 1374 } 1375 1376 if (region->in(1)->is_IfTrue()) { 1377 assert(region->in(2)->is_IfFalse(), "bad If"); 1378 return 1; 1379 } else { 1380 // Flipped projections. 1381 assert(region->in(2)->is_IfTrue(), "bad If"); 1382 return 2; 1383 } 1384 } 1385 1386 // Do the following transformation if we find the corresponding graph shape, remove the involved memory phi and return 1387 // true. Otherwise, return false if the transformation cannot be applied. 1388 // 1389 // If If 1390 // / \ / \ 1391 // IfFalse IfTrue /- Some Node IfFalse IfTrue 1392 // \ / / / \ / Some Node 1393 // Region / /-MergeMem ===> Region | 1394 // / \---Phi | MergeMem 1395 // [other phis] \ [other phis] | 1396 // use use 1397 bool PhiNode::try_clean_memory_phi(PhaseIterGVN* igvn) { 1398 if (_type != Type::MEMORY) { 1399 return false; 1400 } 1401 assert(is_diamond_phi() > 0, "sanity"); 1402 assert(req() == 3, "same as region"); 1403 const Node* region = in(0); 1404 for (uint i = 1; i < 3; i++) { 1405 Node* phi_input = in(i); 1406 if (phi_input != nullptr && phi_input->is_MergeMem() && region->in(i)->outcnt() == 1) { 1407 // Nothing is control-dependent on path #i except the region itself. 1408 MergeMemNode* merge_mem = phi_input->as_MergeMem(); 1409 uint j = 3 - i; 1410 Node* other_phi_input = in(j); 1411 if (other_phi_input != nullptr && other_phi_input == merge_mem->base_memory()) { 1412 // merge_mem is a successor memory to other_phi_input, and is not pinned inside the diamond, so push it out. 1413 // This will allow the diamond to collapse completely if there are no other phis left. 1414 igvn->replace_node(this, merge_mem); 1415 return true; 1416 } 1417 } 1418 } 1419 return false; 1420 } 1421 1422 //----------------------------check_cmove_id----------------------------------- 1423 // Check for CMove'ing a constant after comparing against the constant. 1424 // Happens all the time now, since if we compare equality vs a constant in 1425 // the parser, we "know" the variable is constant on one path and we force 1426 // it. Thus code like "if( x==0 ) {/*EMPTY*/}" ends up inserting a 1427 // conditional move: "x = (x==0)?0:x;". Yucko. This fix is slightly more 1428 // general in that we don't need constants. Since CMove's are only inserted 1429 // in very special circumstances, we do it here on generic Phi's. 1430 Node* PhiNode::is_cmove_id(PhaseTransform* phase, int true_path) { 1431 assert(true_path !=0, "only diamond shape graph expected"); 1432 1433 // is_diamond_phi() has guaranteed the correctness of the nodes sequence: 1434 // phi->region->if_proj->ifnode->bool->cmp 1435 Node* region = in(0); 1436 Node* iff = region->in(1)->in(0); 1437 BoolNode* b = iff->in(1)->as_Bool(); 1438 Node* cmp = b->in(1); 1439 Node* tval = in(true_path); 1440 Node* fval = in(3-true_path); 1441 Node* id = CMoveNode::is_cmove_id(phase, cmp, tval, fval, b); 1442 if (id == nullptr) 1443 return nullptr; 1444 1445 // Either value might be a cast that depends on a branch of 'iff'. 1446 // Since the 'id' value will float free of the diamond, either 1447 // decast or return failure. 1448 Node* ctl = id->in(0); 1449 if (ctl != nullptr && ctl->in(0) == iff) { 1450 if (id->is_ConstraintCast()) { 1451 return id->in(1); 1452 } else { 1453 // Don't know how to disentangle this value. 1454 return nullptr; 1455 } 1456 } 1457 1458 return id; 1459 } 1460 1461 //------------------------------Identity--------------------------------------- 1462 // Check for Region being Identity. 1463 Node* PhiNode::Identity(PhaseGVN* phase) { 1464 if (must_wait_for_region_in_irreducible_loop(phase)) { 1465 return this; 1466 } 1467 // Check for no merging going on 1468 // (There used to be special-case code here when this->region->is_Loop. 1469 // It would check for a tributary phi on the backedge that the main phi 1470 // trivially, perhaps with a single cast. The unique_input method 1471 // does all this and more, by reducing such tributaries to 'this'.) 1472 Node* uin = unique_input(phase, false); 1473 if (uin != nullptr) { 1474 return uin; 1475 } 1476 1477 int true_path = is_diamond_phi(); 1478 // Delay CMove'ing identity if Ideal has not had the chance to handle unsafe cases, yet. 1479 if (true_path != 0 && !(phase->is_IterGVN() && wait_for_region_igvn(phase))) { 1480 Node* id = is_cmove_id(phase, true_path); 1481 if (id != nullptr) { 1482 return id; 1483 } 1484 } 1485 1486 // Looking for phis with identical inputs. If we find one that has 1487 // type TypePtr::BOTTOM, replace the current phi with the bottom phi. 1488 if (phase->is_IterGVN() && type() == Type::MEMORY && adr_type() != 1489 TypePtr::BOTTOM && !adr_type()->is_known_instance()) { 1490 uint phi_len = req(); 1491 Node* phi_reg = region(); 1492 for (DUIterator_Fast imax, i = phi_reg->fast_outs(imax); i < imax; i++) { 1493 Node* u = phi_reg->fast_out(i); 1494 if (u->is_Phi() && u->as_Phi()->type() == Type::MEMORY && 1495 u->adr_type() == TypePtr::BOTTOM && u->in(0) == phi_reg && 1496 u->req() == phi_len) { 1497 for (uint j = 1; j < phi_len; j++) { 1498 if (in(j) != u->in(j)) { 1499 u = nullptr; 1500 break; 1501 } 1502 } 1503 if (u != nullptr) { 1504 return u; 1505 } 1506 } 1507 } 1508 } 1509 1510 return this; // No identity 1511 } 1512 1513 //-----------------------------unique_input------------------------------------ 1514 // Find the unique value, discounting top, self-loops, and casts. 1515 // Return top if there are no inputs, and self if there are multiple. 1516 Node* PhiNode::unique_input(PhaseValues* phase, bool uncast) { 1517 // 1) One unique direct input, 1518 // or if uncast is true: 1519 // 2) some of the inputs have an intervening ConstraintCast 1520 // 3) an input is a self loop 1521 // 1522 // 1) input or 2) input or 3) input __ 1523 // / \ / \ \ / \ 1524 // \ / | cast phi cast 1525 // phi \ / / \ / 1526 // phi / -- 1527 1528 Node* r = in(0); // RegionNode 1529 Node* input = nullptr; // The unique direct input (maybe uncasted = ConstraintCasts removed) 1530 1531 for (uint i = 1, cnt = req(); i < cnt; ++i) { 1532 Node* rc = r->in(i); 1533 if (rc == nullptr || phase->type(rc) == Type::TOP) 1534 continue; // ignore unreachable control path 1535 Node* n = in(i); 1536 if (n == nullptr) 1537 continue; 1538 Node* un = n; 1539 if (uncast) { 1540 #ifdef ASSERT 1541 Node* m = un->uncast(); 1542 #endif 1543 while (un != nullptr && un->req() == 2 && un->is_ConstraintCast()) { 1544 Node* next = un->in(1); 1545 if (phase->type(next)->isa_rawptr() && phase->type(un)->isa_oopptr()) { 1546 // risk exposing raw ptr at safepoint 1547 break; 1548 } 1549 un = next; 1550 } 1551 assert(m == un || un->in(1) == m, "Only expected at CheckCastPP from allocation"); 1552 } 1553 if (un == nullptr || un == this || phase->type(un) == Type::TOP) { 1554 continue; // ignore if top, or in(i) and "this" are in a data cycle 1555 } 1556 // Check for a unique input (maybe uncasted) 1557 if (input == nullptr) { 1558 input = un; 1559 } else if (input != un) { 1560 input = NodeSentinel; // no unique input 1561 } 1562 } 1563 if (input == nullptr) { 1564 return phase->C->top(); // no inputs 1565 } 1566 1567 if (input != NodeSentinel) { 1568 return input; // one unique direct input 1569 } 1570 1571 // Nothing. 1572 return nullptr; 1573 } 1574 1575 //------------------------------is_x2logic------------------------------------- 1576 // Check for simple convert-to-boolean pattern 1577 // If:(C Bool) Region:(IfF IfT) Phi:(Region 0 1) 1578 // Convert Phi to an ConvIB. 1579 static Node *is_x2logic( PhaseGVN *phase, PhiNode *phi, int true_path ) { 1580 assert(true_path !=0, "only diamond shape graph expected"); 1581 1582 // If we're late in the optimization process, we may have already expanded Conv2B nodes 1583 if (phase->C->post_loop_opts_phase() && !Matcher::match_rule_supported(Op_Conv2B)) { 1584 return nullptr; 1585 } 1586 1587 // Convert the true/false index into an expected 0/1 return. 1588 // Map 2->0 and 1->1. 1589 int flipped = 2-true_path; 1590 1591 // is_diamond_phi() has guaranteed the correctness of the nodes sequence: 1592 // phi->region->if_proj->ifnode->bool->cmp 1593 Node *region = phi->in(0); 1594 Node *iff = region->in(1)->in(0); 1595 BoolNode *b = (BoolNode*)iff->in(1); 1596 const CmpNode *cmp = (CmpNode*)b->in(1); 1597 1598 Node *zero = phi->in(1); 1599 Node *one = phi->in(2); 1600 const Type *tzero = phase->type( zero ); 1601 const Type *tone = phase->type( one ); 1602 1603 // Check for compare vs 0 1604 const Type *tcmp = phase->type(cmp->in(2)); 1605 if( tcmp != TypeInt::ZERO && tcmp != TypePtr::NULL_PTR ) { 1606 // Allow cmp-vs-1 if the other input is bounded by 0-1 1607 if( !(tcmp == TypeInt::ONE && phase->type(cmp->in(1)) == TypeInt::BOOL) ) 1608 return nullptr; 1609 flipped = 1-flipped; // Test is vs 1 instead of 0! 1610 } 1611 1612 // Check for setting zero/one opposite expected 1613 if( tzero == TypeInt::ZERO ) { 1614 if( tone == TypeInt::ONE ) { 1615 } else return nullptr; 1616 } else if( tzero == TypeInt::ONE ) { 1617 if( tone == TypeInt::ZERO ) { 1618 flipped = 1-flipped; 1619 } else return nullptr; 1620 } else return nullptr; 1621 1622 // Check for boolean test backwards 1623 if( b->_test._test == BoolTest::ne ) { 1624 } else if( b->_test._test == BoolTest::eq ) { 1625 flipped = 1-flipped; 1626 } else return nullptr; 1627 1628 // Build int->bool conversion 1629 Node* n = new Conv2BNode(cmp->in(1)); 1630 if (flipped) { 1631 n = new XorINode(phase->transform(n), phase->intcon(1)); 1632 } 1633 1634 return n; 1635 } 1636 1637 //------------------------------is_cond_add------------------------------------ 1638 // Check for simple conditional add pattern: "(P < Q) ? X+Y : X;" 1639 // To be profitable the control flow has to disappear; there can be no other 1640 // values merging here. We replace the test-and-branch with: 1641 // "(sgn(P-Q))&Y) + X". Basically, convert "(P < Q)" into 0 or -1 by 1642 // moving the carry bit from (P-Q) into a register with 'sbb EAX,EAX'. 1643 // Then convert Y to 0-or-Y and finally add. 1644 // This is a key transform for SpecJava _201_compress. 1645 static Node* is_cond_add(PhaseGVN *phase, PhiNode *phi, int true_path) { 1646 assert(true_path !=0, "only diamond shape graph expected"); 1647 1648 // is_diamond_phi() has guaranteed the correctness of the nodes sequence: 1649 // phi->region->if_proj->ifnode->bool->cmp 1650 RegionNode *region = (RegionNode*)phi->in(0); 1651 Node *iff = region->in(1)->in(0); 1652 BoolNode* b = iff->in(1)->as_Bool(); 1653 const CmpNode *cmp = (CmpNode*)b->in(1); 1654 1655 // Make sure only merging this one phi here 1656 if (region->has_unique_phi() != phi) return nullptr; 1657 1658 // Make sure each arm of the diamond has exactly one output, which we assume 1659 // is the region. Otherwise, the control flow won't disappear. 1660 if (region->in(1)->outcnt() != 1) return nullptr; 1661 if (region->in(2)->outcnt() != 1) return nullptr; 1662 1663 // Check for "(P < Q)" of type signed int 1664 if (b->_test._test != BoolTest::lt) return nullptr; 1665 if (cmp->Opcode() != Op_CmpI) return nullptr; 1666 1667 Node *p = cmp->in(1); 1668 Node *q = cmp->in(2); 1669 Node *n1 = phi->in( true_path); 1670 Node *n2 = phi->in(3-true_path); 1671 1672 int op = n1->Opcode(); 1673 if( op != Op_AddI // Need zero as additive identity 1674 /*&&op != Op_SubI && 1675 op != Op_AddP && 1676 op != Op_XorI && 1677 op != Op_OrI*/ ) 1678 return nullptr; 1679 1680 Node *x = n2; 1681 Node *y = nullptr; 1682 if( x == n1->in(1) ) { 1683 y = n1->in(2); 1684 } else if( x == n1->in(2) ) { 1685 y = n1->in(1); 1686 } else return nullptr; 1687 1688 // Not so profitable if compare and add are constants 1689 if( q->is_Con() && phase->type(q) != TypeInt::ZERO && y->is_Con() ) 1690 return nullptr; 1691 1692 Node *cmplt = phase->transform( new CmpLTMaskNode(p,q) ); 1693 Node *j_and = phase->transform( new AndINode(cmplt,y) ); 1694 return new AddINode(j_and,x); 1695 } 1696 1697 //------------------------------is_absolute------------------------------------ 1698 // Check for absolute value. 1699 static Node* is_absolute( PhaseGVN *phase, PhiNode *phi_root, int true_path) { 1700 assert(true_path !=0, "only diamond shape graph expected"); 1701 1702 int cmp_zero_idx = 0; // Index of compare input where to look for zero 1703 int phi_x_idx = 0; // Index of phi input where to find naked x 1704 1705 // ABS ends with the merge of 2 control flow paths. 1706 // Find the false path from the true path. With only 2 inputs, 3 - x works nicely. 1707 int false_path = 3 - true_path; 1708 1709 // is_diamond_phi() has guaranteed the correctness of the nodes sequence: 1710 // phi->region->if_proj->ifnode->bool->cmp 1711 BoolNode *bol = phi_root->in(0)->in(1)->in(0)->in(1)->as_Bool(); 1712 Node *cmp = bol->in(1); 1713 1714 // Check bool sense 1715 if (cmp->Opcode() == Op_CmpF || cmp->Opcode() == Op_CmpD) { 1716 switch (bol->_test._test) { 1717 case BoolTest::lt: cmp_zero_idx = 1; phi_x_idx = true_path; break; 1718 case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = false_path; break; 1719 case BoolTest::gt: cmp_zero_idx = 2; phi_x_idx = true_path; break; 1720 case BoolTest::ge: cmp_zero_idx = 1; phi_x_idx = false_path; break; 1721 default: return nullptr; break; 1722 } 1723 } else if (cmp->Opcode() == Op_CmpI || cmp->Opcode() == Op_CmpL) { 1724 switch (bol->_test._test) { 1725 case BoolTest::lt: 1726 case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = false_path; break; 1727 case BoolTest::gt: 1728 case BoolTest::ge: cmp_zero_idx = 2; phi_x_idx = true_path; break; 1729 default: return nullptr; break; 1730 } 1731 } 1732 1733 // Test is next 1734 const Type *tzero = nullptr; 1735 switch (cmp->Opcode()) { 1736 case Op_CmpI: tzero = TypeInt::ZERO; break; // Integer ABS 1737 case Op_CmpL: tzero = TypeLong::ZERO; break; // Long ABS 1738 case Op_CmpF: tzero = TypeF::ZERO; break; // Float ABS 1739 case Op_CmpD: tzero = TypeD::ZERO; break; // Double ABS 1740 default: return nullptr; 1741 } 1742 1743 // Find zero input of compare; the other input is being abs'd 1744 Node *x = nullptr; 1745 bool flip = false; 1746 if( phase->type(cmp->in(cmp_zero_idx)) == tzero ) { 1747 x = cmp->in(3 - cmp_zero_idx); 1748 } else if( phase->type(cmp->in(3 - cmp_zero_idx)) == tzero ) { 1749 // The test is inverted, we should invert the result... 1750 x = cmp->in(cmp_zero_idx); 1751 flip = true; 1752 } else { 1753 return nullptr; 1754 } 1755 1756 // Next get the 2 pieces being selected, one is the original value 1757 // and the other is the negated value. 1758 if( phi_root->in(phi_x_idx) != x ) return nullptr; 1759 1760 // Check other phi input for subtract node 1761 Node *sub = phi_root->in(3 - phi_x_idx); 1762 1763 bool is_sub = sub->Opcode() == Op_SubF || sub->Opcode() == Op_SubD || 1764 sub->Opcode() == Op_SubI || sub->Opcode() == Op_SubL; 1765 1766 // Allow only Sub(0,X) and fail out for all others; Neg is not OK 1767 if (!is_sub || phase->type(sub->in(1)) != tzero || sub->in(2) != x) return nullptr; 1768 1769 if (tzero == TypeF::ZERO) { 1770 x = new AbsFNode(x); 1771 if (flip) { 1772 x = new SubFNode(sub->in(1), phase->transform(x)); 1773 } 1774 } else if (tzero == TypeD::ZERO) { 1775 x = new AbsDNode(x); 1776 if (flip) { 1777 x = new SubDNode(sub->in(1), phase->transform(x)); 1778 } 1779 } else if (tzero == TypeInt::ZERO && Matcher::match_rule_supported(Op_AbsI)) { 1780 x = new AbsINode(x); 1781 if (flip) { 1782 x = new SubINode(sub->in(1), phase->transform(x)); 1783 } 1784 } else if (tzero == TypeLong::ZERO && Matcher::match_rule_supported(Op_AbsL)) { 1785 x = new AbsLNode(x); 1786 if (flip) { 1787 x = new SubLNode(sub->in(1), phase->transform(x)); 1788 } 1789 } else return nullptr; 1790 1791 return x; 1792 } 1793 1794 //------------------------------split_once------------------------------------- 1795 // Helper for split_flow_path 1796 static void split_once(PhaseIterGVN *igvn, Node *phi, Node *val, Node *n, Node *newn) { 1797 igvn->hash_delete(n); // Remove from hash before hacking edges 1798 1799 uint j = 1; 1800 for (uint i = phi->req()-1; i > 0; i--) { 1801 if (phi->in(i) == val) { // Found a path with val? 1802 // Add to NEW Region/Phi, no DU info 1803 newn->set_req( j++, n->in(i) ); 1804 // Remove from OLD Region/Phi 1805 n->del_req(i); 1806 } 1807 } 1808 1809 // Register the new node but do not transform it. Cannot transform until the 1810 // entire Region/Phi conglomerate has been hacked as a single huge transform. 1811 igvn->register_new_node_with_optimizer( newn ); 1812 1813 // Now I can point to the new node. 1814 n->add_req(newn); 1815 igvn->_worklist.push(n); 1816 } 1817 1818 //------------------------------split_flow_path-------------------------------- 1819 // Check for merging identical values and split flow paths 1820 static Node* split_flow_path(PhaseGVN *phase, PhiNode *phi) { 1821 // This optimization tries to find two or more inputs of phi with the same constant value 1822 // It then splits them into a separate Phi, and according Region. If this is a loop-entry, 1823 // and the loop entry has multiple fall-in edges, and some of those fall-in edges have that 1824 // constant, and others not, we may split the fall-in edges into separate Phi's, and create 1825 // an irreducible loop. For reducible loops, this never seems to happen, as the multiple 1826 // fall-in edges are already merged before the loop head during parsing. But with irreducible 1827 // loops present the order or merging during parsing can sometimes prevent this. 1828 if (phase->C->has_irreducible_loop()) { 1829 // Avoid this optimization if any irreducible loops are present. Else we may create 1830 // an irreducible loop that we do not detect. 1831 return nullptr; 1832 } 1833 BasicType bt = phi->type()->basic_type(); 1834 if( bt == T_ILLEGAL || type2size[bt] <= 0 ) 1835 return nullptr; // Bail out on funny non-value stuff 1836 if( phi->req() <= 3 ) // Need at least 2 matched inputs and a 1837 return nullptr; // third unequal input to be worth doing 1838 1839 // Scan for a constant 1840 uint i; 1841 for( i = 1; i < phi->req()-1; i++ ) { 1842 Node *n = phi->in(i); 1843 if( !n ) return nullptr; 1844 if( phase->type(n) == Type::TOP ) return nullptr; 1845 if( n->Opcode() == Op_ConP || n->Opcode() == Op_ConN || n->Opcode() == Op_ConNKlass ) 1846 break; 1847 } 1848 if( i >= phi->req() ) // Only split for constants 1849 return nullptr; 1850 1851 Node *val = phi->in(i); // Constant to split for 1852 uint hit = 0; // Number of times it occurs 1853 Node *r = phi->region(); 1854 1855 for( ; i < phi->req(); i++ ){ // Count occurrences of constant 1856 Node *n = phi->in(i); 1857 if( !n ) return nullptr; 1858 if( phase->type(n) == Type::TOP ) return nullptr; 1859 if( phi->in(i) == val ) { 1860 hit++; 1861 if (Node::may_be_loop_entry(r->in(i))) { 1862 return nullptr; // don't split loop entry path 1863 } 1864 } 1865 } 1866 1867 if( hit <= 1 || // Make sure we find 2 or more 1868 hit == phi->req()-1 ) // and not ALL the same value 1869 return nullptr; 1870 1871 // Now start splitting out the flow paths that merge the same value. 1872 // Split first the RegionNode. 1873 PhaseIterGVN *igvn = phase->is_IterGVN(); 1874 RegionNode *newr = new RegionNode(hit+1); 1875 split_once(igvn, phi, val, r, newr); 1876 1877 // Now split all other Phis than this one 1878 for (DUIterator_Fast kmax, k = r->fast_outs(kmax); k < kmax; k++) { 1879 Node* phi2 = r->fast_out(k); 1880 if( phi2->is_Phi() && phi2->as_Phi() != phi ) { 1881 PhiNode *newphi = PhiNode::make_blank(newr, phi2); 1882 split_once(igvn, phi, val, phi2, newphi); 1883 } 1884 } 1885 1886 // Clean up this guy 1887 igvn->hash_delete(phi); 1888 for( i = phi->req()-1; i > 0; i-- ) { 1889 if( phi->in(i) == val ) { 1890 phi->del_req(i); 1891 } 1892 } 1893 phi->add_req(val); 1894 1895 return phi; 1896 } 1897 1898 // Returns the BasicType of a given convert node and a type, with special handling to ensure that conversions to 1899 // and from half float will return the SHORT basic type, as that wouldn't be returned typically from TypeInt. 1900 static BasicType get_convert_type(Node* convert, const Type* type) { 1901 int convert_op = convert->Opcode(); 1902 if (type->isa_int() && (convert_op == Op_ConvHF2F || convert_op == Op_ConvF2HF)) { 1903 return T_SHORT; 1904 } 1905 1906 return type->basic_type(); 1907 } 1908 1909 //============================================================================= 1910 //------------------------------simple_data_loop_check------------------------- 1911 // Try to determining if the phi node in a simple safe/unsafe data loop. 1912 // Returns: 1913 // enum LoopSafety { Safe = 0, Unsafe, UnsafeLoop }; 1914 // Safe - safe case when the phi and it's inputs reference only safe data 1915 // nodes; 1916 // Unsafe - the phi and it's inputs reference unsafe data nodes but there 1917 // is no reference back to the phi - need a graph walk 1918 // to determine if it is in a loop; 1919 // UnsafeLoop - unsafe case when the phi references itself directly or through 1920 // unsafe data node. 1921 // Note: a safe data node is a node which could/never reference itself during 1922 // GVN transformations. For now it is Con, Proj, Phi, CastPP, CheckCastPP. 1923 // I mark Phi nodes as safe node not only because they can reference itself 1924 // but also to prevent mistaking the fallthrough case inside an outer loop 1925 // as dead loop when the phi references itself through an other phi. 1926 PhiNode::LoopSafety PhiNode::simple_data_loop_check(Node *in) const { 1927 // It is unsafe loop if the phi node references itself directly. 1928 if (in == (Node*)this) 1929 return UnsafeLoop; // Unsafe loop 1930 // Unsafe loop if the phi node references itself through an unsafe data node. 1931 // Exclude cases with null inputs or data nodes which could reference 1932 // itself (safe for dead loops). 1933 if (in != nullptr && !in->is_dead_loop_safe()) { 1934 // Check inputs of phi's inputs also. 1935 // It is much less expensive then full graph walk. 1936 uint cnt = in->req(); 1937 uint i = (in->is_Proj() && !in->is_CFG()) ? 0 : 1; 1938 for (; i < cnt; ++i) { 1939 Node* m = in->in(i); 1940 if (m == (Node*)this) 1941 return UnsafeLoop; // Unsafe loop 1942 if (m != nullptr && !m->is_dead_loop_safe()) { 1943 // Check the most common case (about 30% of all cases): 1944 // phi->Load/Store->AddP->(ConP ConP Con)/(Parm Parm Con). 1945 Node *m1 = (m->is_AddP() && m->req() > 3) ? m->in(1) : nullptr; 1946 if (m1 == (Node*)this) 1947 return UnsafeLoop; // Unsafe loop 1948 if (m1 != nullptr && m1 == m->in(2) && 1949 m1->is_dead_loop_safe() && m->in(3)->is_Con()) { 1950 continue; // Safe case 1951 } 1952 // The phi references an unsafe node - need full analysis. 1953 return Unsafe; 1954 } 1955 } 1956 } 1957 return Safe; // Safe case - we can optimize the phi node. 1958 } 1959 1960 //------------------------------is_unsafe_data_reference----------------------- 1961 // If phi can be reached through the data input - it is data loop. 1962 bool PhiNode::is_unsafe_data_reference(Node *in) const { 1963 assert(req() > 1, ""); 1964 // First, check simple cases when phi references itself directly or 1965 // through an other node. 1966 LoopSafety safety = simple_data_loop_check(in); 1967 if (safety == UnsafeLoop) 1968 return true; // phi references itself - unsafe loop 1969 else if (safety == Safe) 1970 return false; // Safe case - phi could be replaced with the unique input. 1971 1972 // Unsafe case when we should go through data graph to determine 1973 // if the phi references itself. 1974 1975 ResourceMark rm; 1976 1977 Node_List nstack; 1978 VectorSet visited; 1979 1980 nstack.push(in); // Start with unique input. 1981 visited.set(in->_idx); 1982 while (nstack.size() != 0) { 1983 Node* n = nstack.pop(); 1984 uint cnt = n->req(); 1985 uint i = (n->is_Proj() && !n->is_CFG()) ? 0 : 1; 1986 for (; i < cnt; i++) { 1987 Node* m = n->in(i); 1988 if (m == (Node*)this) { 1989 return true; // Data loop 1990 } 1991 if (m != nullptr && !m->is_dead_loop_safe()) { // Only look for unsafe cases. 1992 if (!visited.test_set(m->_idx)) 1993 nstack.push(m); 1994 } 1995 } 1996 } 1997 return false; // The phi is not reachable from its inputs 1998 } 1999 2000 // Is this Phi's region or some inputs to the region enqueued for IGVN 2001 // and so could cause the region to be optimized out? 2002 bool PhiNode::wait_for_region_igvn(PhaseGVN* phase) { 2003 PhaseIterGVN* igvn = phase->is_IterGVN(); 2004 Unique_Node_List& worklist = igvn->_worklist; 2005 bool delay = false; 2006 Node* r = in(0); 2007 for (uint j = 1; j < req(); j++) { 2008 Node* rc = r->in(j); 2009 Node* n = in(j); 2010 2011 if (rc == nullptr || !rc->is_Proj()) { continue; } 2012 if (worklist.member(rc)) { 2013 delay = true; 2014 break; 2015 } 2016 2017 if (rc->in(0) == nullptr || !rc->in(0)->is_If()) { continue; } 2018 if (worklist.member(rc->in(0))) { 2019 delay = true; 2020 break; 2021 } 2022 2023 if (rc->in(0)->in(1) == nullptr || !rc->in(0)->in(1)->is_Bool()) { continue; } 2024 if (worklist.member(rc->in(0)->in(1))) { 2025 delay = true; 2026 break; 2027 } 2028 2029 if (rc->in(0)->in(1)->in(1) == nullptr || !rc->in(0)->in(1)->in(1)->is_Cmp()) { continue; } 2030 if (worklist.member(rc->in(0)->in(1)->in(1))) { 2031 delay = true; 2032 break; 2033 } 2034 } 2035 2036 if (delay) { 2037 worklist.push(this); 2038 } 2039 return delay; 2040 } 2041 2042 // Push inline type input nodes (and null) down through the phi recursively (can handle data loops). 2043 InlineTypeNode* PhiNode::push_inline_types_down(PhaseGVN* phase, bool can_reshape, ciInlineKlass* inline_klass) { 2044 assert(inline_klass != nullptr, "must be"); 2045 InlineTypeNode* vt = InlineTypeNode::make_null(*phase, inline_klass, /* transform = */ false)->clone_with_phis(phase, in(0), nullptr, !_type->maybe_null()); 2046 if (can_reshape) { 2047 // Replace phi right away to be able to use the inline 2048 // type node when reaching the phi again through data loops. 2049 PhaseIterGVN* igvn = phase->is_IterGVN(); 2050 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2051 Node* u = fast_out(i); 2052 igvn->rehash_node_delayed(u); 2053 imax -= u->replace_edge(this, vt); 2054 --i; 2055 } 2056 igvn->rehash_node_delayed(this); 2057 assert(outcnt() == 0, "should be dead now"); 2058 } 2059 ResourceMark rm; 2060 Node_List casts; 2061 for (uint i = 1; i < req(); ++i) { 2062 Node* n = in(i); 2063 while (n->is_ConstraintCast()) { 2064 casts.push(n); 2065 n = n->in(1); 2066 } 2067 if (phase->type(n)->is_zero_type()) { 2068 n = InlineTypeNode::make_null(*phase, inline_klass); 2069 } else if (n->is_Phi()) { 2070 assert(can_reshape, "can only handle phis during IGVN"); 2071 n = phase->transform(n->as_Phi()->push_inline_types_down(phase, can_reshape, inline_klass)); 2072 } 2073 while (casts.size() != 0) { 2074 // Push the cast(s) through the InlineTypeNode 2075 // TODO 8302217 Can we avoid cloning? See InlineTypeNode::clone_if_required 2076 Node* cast = casts.pop()->clone(); 2077 cast->set_req_X(1, n->as_InlineType()->get_oop(), phase); 2078 n = n->clone(); 2079 n->as_InlineType()->set_oop(*phase, phase->transform(cast)); 2080 n = phase->transform(n); 2081 } 2082 bool transform = !can_reshape && (i == (req()-1)); // Transform phis on last merge 2083 vt->merge_with(phase, n->as_InlineType(), i, transform); 2084 } 2085 return vt; 2086 } 2087 2088 // If the Phi's Region is in an irreducible loop, and the Region 2089 // has had an input removed, but not yet transformed, it could be 2090 // that the Region (and this Phi) are not reachable from Root. 2091 // If we allow the Phi to collapse before the Region, this may lead 2092 // to dead-loop data. Wait for the Region to check for reachability, 2093 // and potentially remove the dead code. 2094 bool PhiNode::must_wait_for_region_in_irreducible_loop(PhaseGVN* phase) const { 2095 RegionNode* region = in(0)->as_Region(); 2096 if (region->loop_status() == RegionNode::LoopStatus::MaybeIrreducibleEntry) { 2097 Node* top = phase->C->top(); 2098 for (uint j = 1; j < req(); j++) { 2099 Node* rc = region->in(j); // for each control input 2100 if (rc == nullptr || phase->type(rc) == Type::TOP) { 2101 // Region is missing a control input 2102 Node* n = in(j); 2103 if (n != nullptr && n != top) { 2104 // Phi still has its input, so region just lost its input 2105 return true; 2106 } 2107 } 2108 } 2109 } 2110 return false; 2111 } 2112 2113 //------------------------------Ideal------------------------------------------ 2114 // Return a node which is more "ideal" than the current node. Must preserve 2115 // the CFG, but we can still strip out dead paths. 2116 Node *PhiNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2117 Node *r = in(0); // RegionNode 2118 assert(r != nullptr && r->is_Region(), "this phi must have a region"); 2119 assert(r->in(0) == nullptr || !r->in(0)->is_Root(), "not a specially hidden merge"); 2120 2121 // Note: During parsing, phis are often transformed before their regions. 2122 // This means we have to use type_or_null to defend against untyped regions. 2123 if( phase->type_or_null(r) == Type::TOP ) // Dead code? 2124 return nullptr; // No change 2125 2126 Node *top = phase->C->top(); 2127 bool new_phi = (outcnt() == 0); // transforming new Phi 2128 // No change for igvn if new phi is not hooked 2129 if (new_phi && can_reshape) 2130 return nullptr; 2131 2132 if (must_wait_for_region_in_irreducible_loop(phase)) { 2133 return nullptr; 2134 } 2135 2136 // The are 2 situations when only one valid phi's input is left 2137 // (in addition to Region input). 2138 // One: region is not loop - replace phi with this input. 2139 // Two: region is loop - replace phi with top since this data path is dead 2140 // and we need to break the dead data loop. 2141 Node* progress = nullptr; // Record if any progress made 2142 for( uint j = 1; j < req(); ++j ){ // For all paths in 2143 // Check unreachable control paths 2144 Node* rc = r->in(j); 2145 Node* n = in(j); // Get the input 2146 if (rc == nullptr || phase->type(rc) == Type::TOP) { 2147 if (n != top) { // Not already top? 2148 PhaseIterGVN *igvn = phase->is_IterGVN(); 2149 if (can_reshape && igvn != nullptr) { 2150 igvn->_worklist.push(r); 2151 } 2152 // Nuke it down 2153 set_req_X(j, top, phase); 2154 progress = this; // Record progress 2155 } 2156 } 2157 } 2158 2159 if (can_reshape && outcnt() == 0) { 2160 // set_req() above may kill outputs if Phi is referenced 2161 // only by itself on the dead (top) control path. 2162 return top; 2163 } 2164 2165 bool uncasted = false; 2166 Node* uin = unique_input(phase, false); 2167 if (uin == nullptr && can_reshape && 2168 // If there is a chance that the region can be optimized out do 2169 // not add a cast node that we can't remove yet. 2170 !wait_for_region_igvn(phase)) { 2171 uncasted = true; 2172 uin = unique_input(phase, true); 2173 } 2174 if (uin == top) { // Simplest case: no alive inputs. 2175 if (can_reshape) // IGVN transformation 2176 return top; 2177 else 2178 return nullptr; // Identity will return TOP 2179 } else if (uin != nullptr) { 2180 // Only one not-null unique input path is left. 2181 // Determine if this input is backedge of a loop. 2182 // (Skip new phis which have no uses and dead regions). 2183 if (outcnt() > 0 && r->in(0) != nullptr) { 2184 if (is_data_loop(r->as_Region(), uin, phase)) { 2185 // Break this data loop to avoid creation of a dead loop. 2186 if (can_reshape) { 2187 return top; 2188 } else { 2189 // We can't return top if we are in Parse phase - cut inputs only 2190 // let Identity to handle the case. 2191 replace_edge(uin, top, phase); 2192 return nullptr; 2193 } 2194 } 2195 } 2196 2197 if (uncasted) { 2198 // Add cast nodes between the phi to be removed and its unique input. 2199 // Wait until after parsing for the type information to propagate from the casts. 2200 assert(can_reshape, "Invalid during parsing"); 2201 const Type* phi_type = bottom_type(); 2202 // Add casts to carry the control dependency of the Phi that is 2203 // going away 2204 Node* cast = nullptr; 2205 const TypeTuple* extra_types = collect_types(phase); 2206 if (phi_type->isa_ptr()) { 2207 const Type* uin_type = phase->type(uin); 2208 if (!phi_type->isa_oopptr() && !uin_type->isa_oopptr()) { 2209 cast = new CastPPNode(r, uin, phi_type, ConstraintCastNode::StrongDependency, extra_types); 2210 } else { 2211 // Use a CastPP for a cast to not null and a CheckCastPP for 2212 // a cast to a new klass (and both if both null-ness and 2213 // klass change). 2214 2215 // If the type of phi is not null but the type of uin may be 2216 // null, uin's type must be casted to not null 2217 if (phi_type->join(TypePtr::NOTNULL) == phi_type->remove_speculative() && 2218 uin_type->join(TypePtr::NOTNULL) != uin_type->remove_speculative()) { 2219 cast = new CastPPNode(r, uin, TypePtr::NOTNULL, ConstraintCastNode::StrongDependency, extra_types); 2220 } 2221 2222 // If the type of phi and uin, both casted to not null, 2223 // differ the klass of uin must be (check)cast'ed to match 2224 // that of phi 2225 if (phi_type->join_speculative(TypePtr::NOTNULL) != uin_type->join_speculative(TypePtr::NOTNULL)) { 2226 Node* n = uin; 2227 if (cast != nullptr) { 2228 cast = phase->transform(cast); 2229 n = cast; 2230 } 2231 cast = new CheckCastPPNode(r, n, phi_type, ConstraintCastNode::StrongDependency, extra_types); 2232 } 2233 if (cast == nullptr) { 2234 cast = new CastPPNode(r, uin, phi_type, ConstraintCastNode::StrongDependency, extra_types); 2235 } 2236 } 2237 } else { 2238 cast = ConstraintCastNode::make_cast_for_type(r, uin, phi_type, ConstraintCastNode::StrongDependency, extra_types); 2239 } 2240 assert(cast != nullptr, "cast should be set"); 2241 cast = phase->transform(cast); 2242 // set all inputs to the new cast(s) so the Phi is removed by Identity 2243 PhaseIterGVN* igvn = phase->is_IterGVN(); 2244 for (uint i = 1; i < req(); i++) { 2245 set_req_X(i, cast, igvn); 2246 } 2247 uin = cast; 2248 } 2249 2250 // One unique input. 2251 debug_only(Node* ident = Identity(phase)); 2252 // The unique input must eventually be detected by the Identity call. 2253 #ifdef ASSERT 2254 if (ident != uin && !ident->is_top() && !must_wait_for_region_in_irreducible_loop(phase)) { 2255 // print this output before failing assert 2256 r->dump(3); 2257 this->dump(3); 2258 ident->dump(); 2259 uin->dump(); 2260 } 2261 #endif 2262 // Identity may not return the expected uin, if it has to wait for the region, in irreducible case 2263 assert(ident == uin || ident->is_top() || must_wait_for_region_in_irreducible_loop(phase), "Identity must clean this up"); 2264 return nullptr; 2265 } 2266 2267 Node* opt = nullptr; 2268 int true_path = is_diamond_phi(); 2269 if (true_path != 0 && 2270 // If one of the diamond's branch is in the process of dying then, the Phi's input for that branch might transform 2271 // to top. If that happens replacing the Phi with an operation that consumes the Phi's inputs will cause the Phi 2272 // to be replaced by top. To prevent that, delay the transformation until the branch has a chance to be removed. 2273 !(can_reshape && wait_for_region_igvn(phase))) { 2274 // Check for CMove'ing identity. If it would be unsafe, 2275 // handle it here. In the safe case, let Identity handle it. 2276 Node* unsafe_id = is_cmove_id(phase, true_path); 2277 if( unsafe_id != nullptr && is_unsafe_data_reference(unsafe_id) ) 2278 opt = unsafe_id; 2279 2280 // Check for simple convert-to-boolean pattern 2281 if( opt == nullptr ) 2282 opt = is_x2logic(phase, this, true_path); 2283 2284 // Check for absolute value 2285 if( opt == nullptr ) 2286 opt = is_absolute(phase, this, true_path); 2287 2288 // Check for conditional add 2289 if( opt == nullptr && can_reshape ) 2290 opt = is_cond_add(phase, this, true_path); 2291 2292 // These 4 optimizations could subsume the phi: 2293 // have to check for a dead data loop creation. 2294 if( opt != nullptr ) { 2295 if( opt == unsafe_id || is_unsafe_data_reference(opt) ) { 2296 // Found dead loop. 2297 if( can_reshape ) 2298 return top; 2299 // We can't return top if we are in Parse phase - cut inputs only 2300 // to stop further optimizations for this phi. Identity will return TOP. 2301 assert(req() == 3, "only diamond merge phi here"); 2302 set_req(1, top); 2303 set_req(2, top); 2304 return nullptr; 2305 } else { 2306 return opt; 2307 } 2308 } 2309 } 2310 2311 // Check for merging identical values and split flow paths 2312 if (can_reshape) { 2313 opt = split_flow_path(phase, this); 2314 // This optimization only modifies phi - don't need to check for dead loop. 2315 assert(opt == nullptr || opt == this, "do not elide phi"); 2316 if (opt != nullptr) return opt; 2317 } 2318 2319 if (in(1) != nullptr && in(1)->Opcode() == Op_AddP && can_reshape) { 2320 // Try to undo Phi of AddP: 2321 // (Phi (AddP base address offset) (AddP base2 address2 offset2)) 2322 // becomes: 2323 // newbase := (Phi base base2) 2324 // newaddress := (Phi address address2) 2325 // newoffset := (Phi offset offset2) 2326 // (AddP newbase newaddress newoffset) 2327 // 2328 // This occurs as a result of unsuccessful split_thru_phi and 2329 // interferes with taking advantage of addressing modes. See the 2330 // clone_shift_expressions code in matcher.cpp 2331 Node* addp = in(1); 2332 Node* base = addp->in(AddPNode::Base); 2333 Node* address = addp->in(AddPNode::Address); 2334 Node* offset = addp->in(AddPNode::Offset); 2335 if (base != nullptr && address != nullptr && offset != nullptr && 2336 !base->is_top() && !address->is_top() && !offset->is_top()) { 2337 const Type* base_type = base->bottom_type(); 2338 const Type* address_type = address->bottom_type(); 2339 // make sure that all the inputs are similar to the first one, 2340 // i.e. AddP with base == address and same offset as first AddP 2341 bool doit = true; 2342 for (uint i = 2; i < req(); i++) { 2343 if (in(i) == nullptr || 2344 in(i)->Opcode() != Op_AddP || 2345 in(i)->in(AddPNode::Base) == nullptr || 2346 in(i)->in(AddPNode::Address) == nullptr || 2347 in(i)->in(AddPNode::Offset) == nullptr || 2348 in(i)->in(AddPNode::Base)->is_top() || 2349 in(i)->in(AddPNode::Address)->is_top() || 2350 in(i)->in(AddPNode::Offset)->is_top()) { 2351 doit = false; 2352 break; 2353 } 2354 if (in(i)->in(AddPNode::Base) != base) { 2355 base = nullptr; 2356 } 2357 if (in(i)->in(AddPNode::Offset) != offset) { 2358 offset = nullptr; 2359 } 2360 if (in(i)->in(AddPNode::Address) != address) { 2361 address = nullptr; 2362 } 2363 // Accumulate type for resulting Phi 2364 base_type = base_type->meet_speculative(in(i)->in(AddPNode::Base)->bottom_type()); 2365 address_type = address_type->meet_speculative(in(i)->in(AddPNode::Address)->bottom_type()); 2366 } 2367 if (doit && base == nullptr) { 2368 // Check for neighboring AddP nodes in a tree. 2369 // If they have a base, use that it. 2370 for (DUIterator_Fast kmax, k = this->fast_outs(kmax); k < kmax; k++) { 2371 Node* u = this->fast_out(k); 2372 if (u->is_AddP()) { 2373 Node* base2 = u->in(AddPNode::Base); 2374 if (base2 != nullptr && !base2->is_top()) { 2375 if (base == nullptr) 2376 base = base2; 2377 else if (base != base2) 2378 { doit = false; break; } 2379 } 2380 } 2381 } 2382 } 2383 if (doit) { 2384 if (base == nullptr) { 2385 base = new PhiNode(in(0), base_type, nullptr); 2386 for (uint i = 1; i < req(); i++) { 2387 base->init_req(i, in(i)->in(AddPNode::Base)); 2388 } 2389 phase->is_IterGVN()->register_new_node_with_optimizer(base); 2390 } 2391 if (address == nullptr) { 2392 address = new PhiNode(in(0), address_type, nullptr); 2393 for (uint i = 1; i < req(); i++) { 2394 address->init_req(i, in(i)->in(AddPNode::Address)); 2395 } 2396 phase->is_IterGVN()->register_new_node_with_optimizer(address); 2397 } 2398 if (offset == nullptr) { 2399 offset = new PhiNode(in(0), TypeX_X, nullptr); 2400 for (uint i = 1; i < req(); i++) { 2401 offset->init_req(i, in(i)->in(AddPNode::Offset)); 2402 } 2403 phase->is_IterGVN()->register_new_node_with_optimizer(offset); 2404 } 2405 return new AddPNode(base, address, offset); 2406 } 2407 } 2408 } 2409 2410 // Split phis through memory merges, so that the memory merges will go away. 2411 // Piggy-back this transformation on the search for a unique input.... 2412 // It will be as if the merged memory is the unique value of the phi. 2413 // (Do not attempt this optimization unless parsing is complete. 2414 // It would make the parser's memory-merge logic sick.) 2415 // (MergeMemNode is not dead_loop_safe - need to check for dead loop.) 2416 if (progress == nullptr && can_reshape && type() == Type::MEMORY) { 2417 // see if this phi should be sliced 2418 uint merge_width = 0; 2419 bool saw_self = false; 2420 // TODO revisit this with JDK-8247216 2421 bool mergemem_only = true; 2422 for( uint i=1; i<req(); ++i ) {// For all paths in 2423 Node *ii = in(i); 2424 // TOP inputs should not be counted as safe inputs because if the 2425 // Phi references itself through all other inputs then splitting the 2426 // Phi through memory merges would create dead loop at later stage. 2427 if (ii == top) { 2428 return nullptr; // Delay optimization until graph is cleaned. 2429 } 2430 if (ii->is_MergeMem()) { 2431 MergeMemNode* n = ii->as_MergeMem(); 2432 merge_width = MAX2(merge_width, n->req()); 2433 saw_self = saw_self || (n->base_memory() == this); 2434 } else { 2435 mergemem_only = false; 2436 } 2437 } 2438 2439 // This restriction is temporarily necessary to ensure termination: 2440 if (!mergemem_only && !saw_self && adr_type() == TypePtr::BOTTOM) merge_width = 0; 2441 2442 if (merge_width > Compile::AliasIdxRaw) { 2443 // found at least one non-empty MergeMem 2444 const TypePtr* at = adr_type(); 2445 if (at != TypePtr::BOTTOM) { 2446 // Patch the existing phi to select an input from the merge: 2447 // Phi:AT1(...MergeMem(m0, m1, m2)...) into 2448 // Phi:AT1(...m1...) 2449 int alias_idx = phase->C->get_alias_index(at); 2450 for (uint i=1; i<req(); ++i) { 2451 Node *ii = in(i); 2452 if (ii->is_MergeMem()) { 2453 MergeMemNode* n = ii->as_MergeMem(); 2454 // compress paths and change unreachable cycles to TOP 2455 // If not, we can update the input infinitely along a MergeMem cycle 2456 // Equivalent code is in MemNode::Ideal_common 2457 Node *m = phase->transform(n); 2458 if (outcnt() == 0) { // Above transform() may kill us! 2459 return top; 2460 } 2461 // If transformed to a MergeMem, get the desired slice 2462 // Otherwise the returned node represents memory for every slice 2463 Node *new_mem = (m->is_MergeMem()) ? 2464 m->as_MergeMem()->memory_at(alias_idx) : m; 2465 // Update input if it is progress over what we have now 2466 if (new_mem != ii) { 2467 set_req_X(i, new_mem, phase->is_IterGVN()); 2468 progress = this; 2469 } 2470 } 2471 } 2472 } else { 2473 // We know that at least one MergeMem->base_memory() == this 2474 // (saw_self == true). If all other inputs also references this phi 2475 // (directly or through data nodes) - it is a dead loop. 2476 bool saw_safe_input = false; 2477 for (uint j = 1; j < req(); ++j) { 2478 Node* n = in(j); 2479 if (n->is_MergeMem()) { 2480 MergeMemNode* mm = n->as_MergeMem(); 2481 if (mm->base_memory() == this || mm->base_memory() == mm->empty_memory()) { 2482 // Skip this input if it references back to this phi or if the memory path is dead 2483 continue; 2484 } 2485 } 2486 if (!is_unsafe_data_reference(n)) { 2487 saw_safe_input = true; // found safe input 2488 break; 2489 } 2490 } 2491 if (!saw_safe_input) { 2492 // There is a dead loop: All inputs are either dead or reference back to this phi 2493 return top; 2494 } 2495 2496 // Phi(...MergeMem(m0, m1:AT1, m2:AT2)...) into 2497 // MergeMem(Phi(...m0...), Phi:AT1(...m1...), Phi:AT2(...m2...)) 2498 PhaseIterGVN* igvn = phase->is_IterGVN(); 2499 assert(igvn != nullptr, "sanity check"); 2500 Node* hook = new Node(1); 2501 PhiNode* new_base = (PhiNode*) clone(); 2502 // Must eagerly register phis, since they participate in loops. 2503 igvn->register_new_node_with_optimizer(new_base); 2504 hook->add_req(new_base); 2505 2506 MergeMemNode* result = MergeMemNode::make(new_base); 2507 for (uint i = 1; i < req(); ++i) { 2508 Node *ii = in(i); 2509 if (ii->is_MergeMem()) { 2510 MergeMemNode* n = ii->as_MergeMem(); 2511 if (igvn) { 2512 // TODO revisit this with JDK-8247216 2513 // Put 'n' on the worklist because it might be modified by MergeMemStream::iteration_setup 2514 igvn->_worklist.push(n); 2515 } 2516 for (MergeMemStream mms(result, n); mms.next_non_empty2(); ) { 2517 // If we have not seen this slice yet, make a phi for it. 2518 bool made_new_phi = false; 2519 if (mms.is_empty()) { 2520 Node* new_phi = new_base->slice_memory(mms.adr_type(phase->C)); 2521 made_new_phi = true; 2522 igvn->register_new_node_with_optimizer(new_phi); 2523 hook->add_req(new_phi); 2524 mms.set_memory(new_phi); 2525 } 2526 Node* phi = mms.memory(); 2527 assert(made_new_phi || phi->in(i) == n, "replace the i-th merge by a slice"); 2528 phi->set_req(i, mms.memory2()); 2529 } 2530 } 2531 } 2532 // Distribute all self-loops. 2533 { // (Extra braces to hide mms.) 2534 for (MergeMemStream mms(result); mms.next_non_empty(); ) { 2535 Node* phi = mms.memory(); 2536 for (uint i = 1; i < req(); ++i) { 2537 if (phi->in(i) == this) phi->set_req(i, phi); 2538 } 2539 } 2540 } 2541 // Already replace this phi node to cut it off from the graph to not interfere in dead loop checks during the 2542 // transformations of the new phi nodes below. Otherwise, we could wrongly conclude that there is no dead loop 2543 // because we are finding this phi node again. Also set the type of the new MergeMem node in case we are also 2544 // visiting it in the transformations below. 2545 igvn->replace_node(this, result); 2546 igvn->set_type(result, result->bottom_type()); 2547 2548 // now transform the new nodes, and return the mergemem 2549 for (MergeMemStream mms(result); mms.next_non_empty(); ) { 2550 Node* phi = mms.memory(); 2551 mms.set_memory(phase->transform(phi)); 2552 } 2553 hook->destruct(igvn); 2554 // Replace self with the result. 2555 return result; 2556 } 2557 } 2558 // 2559 // Other optimizations on the memory chain 2560 // 2561 const TypePtr* at = adr_type(); 2562 for( uint i=1; i<req(); ++i ) {// For all paths in 2563 Node *ii = in(i); 2564 Node *new_in = MemNode::optimize_memory_chain(ii, at, nullptr, phase); 2565 if (ii != new_in ) { 2566 set_req(i, new_in); 2567 progress = this; 2568 } 2569 } 2570 } 2571 2572 #ifdef _LP64 2573 // Push DecodeN/DecodeNKlass down through phi. 2574 // The rest of phi graph will transform by split EncodeP node though phis up. 2575 if ((UseCompressedOops || UseCompressedClassPointers) && can_reshape && progress == nullptr) { 2576 bool may_push = true; 2577 bool has_decodeN = false; 2578 bool is_decodeN = false; 2579 for (uint i=1; i<req(); ++i) {// For all paths in 2580 Node *ii = in(i); 2581 if (ii->is_DecodeNarrowPtr() && ii->bottom_type() == bottom_type()) { 2582 // Do optimization if a non dead path exist. 2583 if (ii->in(1)->bottom_type() != Type::TOP) { 2584 has_decodeN = true; 2585 is_decodeN = ii->is_DecodeN(); 2586 } 2587 } else if (!ii->is_Phi()) { 2588 may_push = false; 2589 } 2590 } 2591 2592 if (has_decodeN && may_push) { 2593 PhaseIterGVN *igvn = phase->is_IterGVN(); 2594 // Make narrow type for new phi. 2595 const Type* narrow_t; 2596 if (is_decodeN) { 2597 narrow_t = TypeNarrowOop::make(this->bottom_type()->is_ptr()); 2598 } else { 2599 narrow_t = TypeNarrowKlass::make(this->bottom_type()->is_ptr()); 2600 } 2601 PhiNode* new_phi = new PhiNode(r, narrow_t); 2602 uint orig_cnt = req(); 2603 for (uint i=1; i<req(); ++i) {// For all paths in 2604 Node *ii = in(i); 2605 Node* new_ii = nullptr; 2606 if (ii->is_DecodeNarrowPtr()) { 2607 assert(ii->bottom_type() == bottom_type(), "sanity"); 2608 new_ii = ii->in(1); 2609 } else { 2610 assert(ii->is_Phi(), "sanity"); 2611 if (ii->as_Phi() == this) { 2612 new_ii = new_phi; 2613 } else { 2614 if (is_decodeN) { 2615 new_ii = new EncodePNode(ii, narrow_t); 2616 } else { 2617 new_ii = new EncodePKlassNode(ii, narrow_t); 2618 } 2619 igvn->register_new_node_with_optimizer(new_ii); 2620 } 2621 } 2622 new_phi->set_req(i, new_ii); 2623 } 2624 igvn->register_new_node_with_optimizer(new_phi, this); 2625 if (is_decodeN) { 2626 progress = new DecodeNNode(new_phi, bottom_type()); 2627 } else { 2628 progress = new DecodeNKlassNode(new_phi, bottom_type()); 2629 } 2630 } 2631 } 2632 #endif 2633 2634 Node* inline_type = try_push_inline_types_down(phase, can_reshape); 2635 if (inline_type != this) { 2636 return inline_type; 2637 } 2638 2639 // Try to convert a Phi with two duplicated convert nodes into a phi of the pre-conversion type and the convert node 2640 // proceeding the phi, to de-duplicate the convert node and compact the IR. 2641 if (can_reshape && progress == nullptr) { 2642 ConvertNode* convert = in(1)->isa_Convert(); 2643 if (convert != nullptr) { 2644 int conv_op = convert->Opcode(); 2645 bool ok = true; 2646 2647 // Check the rest of the inputs 2648 for (uint i = 2; i < req(); i++) { 2649 // Make sure that all inputs are of the same type of convert node 2650 if (in(i)->Opcode() != conv_op) { 2651 ok = false; 2652 break; 2653 } 2654 } 2655 2656 if (ok) { 2657 // Find the local bottom type to set as the type of the phi 2658 const Type* source_type = Type::get_const_basic_type(convert->in_type()->basic_type()); 2659 const Type* dest_type = convert->bottom_type(); 2660 2661 PhiNode* newphi = new PhiNode(in(0), source_type, nullptr); 2662 // Set inputs to the new phi be the inputs of the convert 2663 for (uint i = 1; i < req(); i++) { 2664 newphi->init_req(i, in(i)->in(1)); 2665 } 2666 2667 phase->is_IterGVN()->register_new_node_with_optimizer(newphi, this); 2668 2669 return ConvertNode::create_convert(get_convert_type(convert, source_type), get_convert_type(convert, dest_type), newphi); 2670 } 2671 } 2672 } 2673 2674 // Phi (VB ... VB) => VB (Phi ...) (Phi ...) 2675 if (EnableVectorReboxing && can_reshape && progress == nullptr && type()->isa_oopptr()) { 2676 progress = merge_through_phi(this, phase->is_IterGVN()); 2677 } 2678 2679 return progress; // Return any progress 2680 } 2681 2682 // Check recursively if inputs are either an inline type, constant null 2683 // or another Phi (including self references through data loops). If so, 2684 // push the inline types down through the phis to enable folding of loads. 2685 Node* PhiNode::try_push_inline_types_down(PhaseGVN* phase, const bool can_reshape) { 2686 if (!can_be_inline_type()) { 2687 return this; 2688 } 2689 2690 ciInlineKlass* inline_klass; 2691 if (can_push_inline_types_down(phase, can_reshape, inline_klass)) { 2692 assert(inline_klass != nullptr, "must be"); 2693 return push_inline_types_down(phase, can_reshape, inline_klass); 2694 } 2695 return this; 2696 } 2697 2698 bool PhiNode::can_push_inline_types_down(PhaseGVN* phase, const bool can_reshape, ciInlineKlass*& inline_klass) { 2699 if (req() <= 2) { 2700 // Dead phi. 2701 return false; 2702 } 2703 inline_klass = nullptr; 2704 2705 // TODO 8302217 We need to prevent endless pushing through 2706 bool only_phi = (outcnt() != 0); 2707 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2708 Node* n = fast_out(i); 2709 if (n->is_InlineType() && n->in(1) == this) { 2710 return false; 2711 } 2712 if (!n->is_Phi()) { 2713 only_phi = false; 2714 } 2715 } 2716 if (only_phi) { 2717 return false; 2718 } 2719 2720 ResourceMark rm; 2721 Unique_Node_List worklist; 2722 worklist.push(this); 2723 Node_List casts; 2724 2725 for (uint next = 0; next < worklist.size(); next++) { 2726 Node* phi = worklist.at(next); 2727 for (uint i = 1; i < phi->req(); i++) { 2728 Node* n = phi->in(i); 2729 if (n == nullptr) { 2730 return false; 2731 } 2732 while (n->is_ConstraintCast()) { 2733 if (n->in(0) != nullptr && n->in(0)->is_top()) { 2734 // Will die, don't optimize 2735 return false; 2736 } 2737 casts.push(n); 2738 n = n->in(1); 2739 } 2740 const Type* type = phase->type(n); 2741 if (n->is_InlineType() && (inline_klass == nullptr || inline_klass == type->inline_klass())) { 2742 inline_klass = type->inline_klass(); 2743 } else if (n->is_Phi() && can_reshape && n->bottom_type()->isa_ptr()) { 2744 worklist.push(n); 2745 } else if (!type->is_zero_type()) { 2746 return false; 2747 } 2748 } 2749 } 2750 if (inline_klass == nullptr) { 2751 return false; 2752 } 2753 2754 // Check if cast nodes can be pushed through 2755 const Type* t = Type::get_const_type(inline_klass); 2756 while (casts.size() != 0 && t != nullptr) { 2757 Node* cast = casts.pop(); 2758 if (t->filter(cast->bottom_type()) == Type::TOP) { 2759 return false; 2760 } 2761 } 2762 2763 return true; 2764 } 2765 2766 static int compare_types(const Type* const& e1, const Type* const& e2) { 2767 return (intptr_t)e1 - (intptr_t)e2; 2768 } 2769 2770 // Collect types at casts that are going to be eliminated at that Phi and store them in a TypeTuple. 2771 // Sort the types using an arbitrary order so a list of some types always hashes to the same TypeTuple (and TypeTuple 2772 // pointer comparison is enough to tell if 2 list of types are the same or not) 2773 const TypeTuple* PhiNode::collect_types(PhaseGVN* phase) const { 2774 const Node* region = in(0); 2775 const Type* phi_type = bottom_type(); 2776 ResourceMark rm; 2777 GrowableArray<const Type*> types; 2778 for (uint i = 1; i < req(); i++) { 2779 if (region->in(i) == nullptr || phase->type(region->in(i)) == Type::TOP) { 2780 continue; 2781 } 2782 Node* in = Node::in(i); 2783 const Type* t = phase->type(in); 2784 if (in == nullptr || in == this || t == Type::TOP) { 2785 continue; 2786 } 2787 if (t != phi_type && t->higher_equal_speculative(phi_type)) { 2788 types.insert_sorted<compare_types>(t); 2789 } 2790 while (in != nullptr && in->is_ConstraintCast()) { 2791 Node* next = in->in(1); 2792 if (phase->type(next)->isa_rawptr() && phase->type(in)->isa_oopptr()) { 2793 break; 2794 } 2795 ConstraintCastNode* cast = in->as_ConstraintCast(); 2796 for (int j = 0; j < cast->extra_types_count(); ++j) { 2797 const Type* extra_t = cast->extra_type_at(j); 2798 if (extra_t != phi_type && extra_t->higher_equal_speculative(phi_type)) { 2799 types.insert_sorted<compare_types>(extra_t); 2800 } 2801 } 2802 in = next; 2803 } 2804 } 2805 const Type **flds = (const Type **)(phase->C->type_arena()->AmallocWords(types.length()*sizeof(Type*))); 2806 for (int i = 0; i < types.length(); ++i) { 2807 flds[i] = types.at(i); 2808 } 2809 return TypeTuple::make(types.length(), flds); 2810 } 2811 2812 Node* PhiNode::clone_through_phi(Node* root_phi, const Type* t, uint c, PhaseIterGVN* igvn) { 2813 Node_Stack stack(1); 2814 VectorSet visited; 2815 Node_List node_map; 2816 2817 stack.push(root_phi, 1); // ignore control 2818 visited.set(root_phi->_idx); 2819 2820 Node* new_phi = new PhiNode(root_phi->in(0), t); 2821 node_map.map(root_phi->_idx, new_phi); 2822 2823 while (stack.is_nonempty()) { 2824 Node* n = stack.node(); 2825 uint idx = stack.index(); 2826 assert(n->is_Phi(), "not a phi"); 2827 if (idx < n->req()) { 2828 stack.set_index(idx + 1); 2829 Node* def = n->in(idx); 2830 if (def == nullptr) { 2831 continue; // ignore dead path 2832 } else if (def->is_Phi()) { // inner node 2833 Node* new_phi = node_map[n->_idx]; 2834 if (!visited.test_set(def->_idx)) { // not visited yet 2835 node_map.map(def->_idx, new PhiNode(def->in(0), t)); 2836 stack.push(def, 1); // ignore control 2837 } 2838 Node* new_in = node_map[def->_idx]; 2839 new_phi->set_req(idx, new_in); 2840 } else if (def->Opcode() == Op_VectorBox) { // leaf 2841 assert(n->is_Phi(), "not a phi"); 2842 Node* new_phi = node_map[n->_idx]; 2843 new_phi->set_req(idx, def->in(c)); 2844 } else { 2845 assert(false, "not optimizeable"); 2846 return nullptr; 2847 } 2848 } else { 2849 Node* new_phi = node_map[n->_idx]; 2850 igvn->register_new_node_with_optimizer(new_phi, n); 2851 stack.pop(); 2852 } 2853 } 2854 return new_phi; 2855 } 2856 2857 Node* PhiNode::merge_through_phi(Node* root_phi, PhaseIterGVN* igvn) { 2858 Node_Stack stack(1); 2859 VectorSet visited; 2860 2861 stack.push(root_phi, 1); // ignore control 2862 visited.set(root_phi->_idx); 2863 2864 VectorBoxNode* cached_vbox = nullptr; 2865 while (stack.is_nonempty()) { 2866 Node* n = stack.node(); 2867 uint idx = stack.index(); 2868 if (idx < n->req()) { 2869 stack.set_index(idx + 1); 2870 Node* in = n->in(idx); 2871 if (in == nullptr) { 2872 continue; // ignore dead path 2873 } else if (in->isa_Phi()) { 2874 if (!visited.test_set(in->_idx)) { 2875 stack.push(in, 1); // ignore control 2876 } 2877 } else if (in->Opcode() == Op_VectorBox) { 2878 VectorBoxNode* vbox = static_cast<VectorBoxNode*>(in); 2879 if (cached_vbox == nullptr) { 2880 cached_vbox = vbox; 2881 } else if (vbox->vec_type() != cached_vbox->vec_type()) { 2882 // TODO: vector type mismatch can be handled with additional reinterpret casts 2883 assert(!Type::equals(vbox->vec_type(), cached_vbox->vec_type()), "inconsistent"); 2884 return nullptr; // not optimizable: vector type mismatch 2885 } else if (vbox->box_type() != cached_vbox->box_type()) { 2886 assert(!Type::equals(vbox->box_type(), cached_vbox->box_type()), "inconsistent"); 2887 return nullptr; // not optimizable: box type mismatch 2888 } 2889 } else { 2890 return nullptr; // not optimizable: neither Phi nor VectorBox 2891 } 2892 } else { 2893 stack.pop(); 2894 } 2895 } 2896 if (cached_vbox == nullptr) { 2897 // We have a Phi dead-loop (no data-input). Phi nodes are considered safe, 2898 // so just avoid this optimization. 2899 return nullptr; 2900 } 2901 const TypeInstPtr* btype = cached_vbox->box_type(); 2902 const TypeVect* vtype = cached_vbox->vec_type(); 2903 Node* new_vbox_phi = clone_through_phi(root_phi, btype, VectorBoxNode::Box, igvn); 2904 Node* new_vect_phi = clone_through_phi(root_phi, vtype, VectorBoxNode::Value, igvn); 2905 return new VectorBoxNode(igvn->C, new_vbox_phi, new_vect_phi, btype, vtype); 2906 } 2907 2908 bool PhiNode::is_data_loop(RegionNode* r, Node* uin, const PhaseGVN* phase) { 2909 // First, take the short cut when we know it is a loop and the EntryControl data path is dead. 2910 // The loop node may only have one input because the entry path was removed in PhaseIdealLoop::Dominators(). 2911 // Then, check if there is a data loop when the phi references itself directly or through other data nodes. 2912 assert(!r->is_Loop() || r->req() <= 3, "Loop node should have 3 or less inputs"); 2913 const bool is_loop = (r->is_Loop() && r->req() == 3); 2914 const Node* top = phase->C->top(); 2915 if (is_loop) { 2916 return !uin->eqv_uncast(in(LoopNode::EntryControl)); 2917 } else { 2918 // We have a data loop either with an unsafe data reference or if a region is unreachable. 2919 return is_unsafe_data_reference(uin) 2920 || (r->req() == 3 && (r->in(1) != top && r->in(2) == top && r->is_unreachable_region(phase))); 2921 } 2922 } 2923 2924 //------------------------------is_tripcount----------------------------------- 2925 bool PhiNode::is_tripcount(BasicType bt) const { 2926 return (in(0) != nullptr && in(0)->is_BaseCountedLoop() && 2927 in(0)->as_BaseCountedLoop()->bt() == bt && 2928 in(0)->as_BaseCountedLoop()->phi() == this); 2929 } 2930 2931 //------------------------------out_RegMask------------------------------------ 2932 const RegMask &PhiNode::in_RegMask(uint i) const { 2933 return i ? out_RegMask() : RegMask::Empty; 2934 } 2935 2936 const RegMask &PhiNode::out_RegMask() const { 2937 uint ideal_reg = _type->ideal_reg(); 2938 assert( ideal_reg != Node::NotAMachineReg, "invalid type at Phi" ); 2939 if( ideal_reg == 0 ) return RegMask::Empty; 2940 assert(ideal_reg != Op_RegFlags, "flags register is not spillable"); 2941 return *(Compile::current()->matcher()->idealreg2spillmask[ideal_reg]); 2942 } 2943 2944 #ifndef PRODUCT 2945 void PhiNode::dump_spec(outputStream *st) const { 2946 TypeNode::dump_spec(st); 2947 if (is_tripcount(T_INT) || is_tripcount(T_LONG)) { 2948 st->print(" #tripcount"); 2949 } 2950 } 2951 #endif 2952 2953 2954 //============================================================================= 2955 const Type* GotoNode::Value(PhaseGVN* phase) const { 2956 // If the input is reachable, then we are executed. 2957 // If the input is not reachable, then we are not executed. 2958 return phase->type(in(0)); 2959 } 2960 2961 Node* GotoNode::Identity(PhaseGVN* phase) { 2962 return in(0); // Simple copy of incoming control 2963 } 2964 2965 const RegMask &GotoNode::out_RegMask() const { 2966 return RegMask::Empty; 2967 } 2968 2969 //============================================================================= 2970 const RegMask &JumpNode::out_RegMask() const { 2971 return RegMask::Empty; 2972 } 2973 2974 //============================================================================= 2975 const RegMask &JProjNode::out_RegMask() const { 2976 return RegMask::Empty; 2977 } 2978 2979 //============================================================================= 2980 const RegMask &CProjNode::out_RegMask() const { 2981 return RegMask::Empty; 2982 } 2983 2984 2985 2986 //============================================================================= 2987 2988 uint PCTableNode::hash() const { return Node::hash() + _size; } 2989 bool PCTableNode::cmp( const Node &n ) const 2990 { return _size == ((PCTableNode&)n)._size; } 2991 2992 const Type *PCTableNode::bottom_type() const { 2993 const Type** f = TypeTuple::fields(_size); 2994 for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL; 2995 return TypeTuple::make(_size, f); 2996 } 2997 2998 //------------------------------Value------------------------------------------ 2999 // Compute the type of the PCTableNode. If reachable it is a tuple of 3000 // Control, otherwise the table targets are not reachable 3001 const Type* PCTableNode::Value(PhaseGVN* phase) const { 3002 if( phase->type(in(0)) == Type::CONTROL ) 3003 return bottom_type(); 3004 return Type::TOP; // All paths dead? Then so are we 3005 } 3006 3007 //------------------------------Ideal------------------------------------------ 3008 // Return a node which is more "ideal" than the current node. Strip out 3009 // control copies 3010 Node *PCTableNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3011 return remove_dead_region(phase, can_reshape) ? this : nullptr; 3012 } 3013 3014 //============================================================================= 3015 uint JumpProjNode::hash() const { 3016 return Node::hash() + _dest_bci; 3017 } 3018 3019 bool JumpProjNode::cmp( const Node &n ) const { 3020 return ProjNode::cmp(n) && 3021 _dest_bci == ((JumpProjNode&)n)._dest_bci; 3022 } 3023 3024 #ifndef PRODUCT 3025 void JumpProjNode::dump_spec(outputStream *st) const { 3026 ProjNode::dump_spec(st); 3027 st->print("@bci %d ",_dest_bci); 3028 } 3029 3030 void JumpProjNode::dump_compact_spec(outputStream *st) const { 3031 ProjNode::dump_compact_spec(st); 3032 st->print("(%d)%d@%d", _switch_val, _proj_no, _dest_bci); 3033 } 3034 #endif 3035 3036 //============================================================================= 3037 //------------------------------Value------------------------------------------ 3038 // Check for being unreachable, or for coming from a Rethrow. Rethrow's cannot 3039 // have the default "fall_through_index" path. 3040 const Type* CatchNode::Value(PhaseGVN* phase) const { 3041 // Unreachable? Then so are all paths from here. 3042 if( phase->type(in(0)) == Type::TOP ) return Type::TOP; 3043 // First assume all paths are reachable 3044 const Type** f = TypeTuple::fields(_size); 3045 for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL; 3046 // Identify cases that will always throw an exception 3047 // () rethrow call 3048 // () virtual or interface call with null receiver 3049 // () call is a check cast with incompatible arguments 3050 if( in(1)->is_Proj() ) { 3051 Node *i10 = in(1)->in(0); 3052 if( i10->is_Call() ) { 3053 CallNode *call = i10->as_Call(); 3054 // Rethrows always throw exceptions, never return 3055 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3056 f[CatchProjNode::fall_through_index] = Type::TOP; 3057 } else if (call->is_AllocateArray()) { 3058 Node* klass_node = call->in(AllocateNode::KlassNode); 3059 Node* length = call->in(AllocateNode::ALength); 3060 const Type* length_type = phase->type(length); 3061 const Type* klass_type = phase->type(klass_node); 3062 Node* valid_length_test = call->in(AllocateNode::ValidLengthTest); 3063 const Type* valid_length_test_t = phase->type(valid_length_test); 3064 if (length_type == Type::TOP || klass_type == Type::TOP || valid_length_test_t == Type::TOP || 3065 valid_length_test_t->is_int()->is_con(0)) { 3066 f[CatchProjNode::fall_through_index] = Type::TOP; 3067 } 3068 } else if( call->req() > TypeFunc::Parms ) { 3069 const Type *arg0 = phase->type( call->in(TypeFunc::Parms) ); 3070 // Check for null receiver to virtual or interface calls 3071 if( call->is_CallDynamicJava() && 3072 arg0->higher_equal(TypePtr::NULL_PTR) ) { 3073 f[CatchProjNode::fall_through_index] = Type::TOP; 3074 } 3075 } // End of if not a runtime stub 3076 } // End of if have call above me 3077 } // End of slot 1 is not a projection 3078 return TypeTuple::make(_size, f); 3079 } 3080 3081 //============================================================================= 3082 uint CatchProjNode::hash() const { 3083 return Node::hash() + _handler_bci; 3084 } 3085 3086 3087 bool CatchProjNode::cmp( const Node &n ) const { 3088 return ProjNode::cmp(n) && 3089 _handler_bci == ((CatchProjNode&)n)._handler_bci; 3090 } 3091 3092 3093 //------------------------------Identity--------------------------------------- 3094 // If only 1 target is possible, choose it if it is the main control 3095 Node* CatchProjNode::Identity(PhaseGVN* phase) { 3096 // If my value is control and no other value is, then treat as ID 3097 const TypeTuple *t = phase->type(in(0))->is_tuple(); 3098 if (t->field_at(_con) != Type::CONTROL) return this; 3099 // If we remove the last CatchProj and elide the Catch/CatchProj, then we 3100 // also remove any exception table entry. Thus we must know the call 3101 // feeding the Catch will not really throw an exception. This is ok for 3102 // the main fall-thru control (happens when we know a call can never throw 3103 // an exception) or for "rethrow", because a further optimization will 3104 // yank the rethrow (happens when we inline a function that can throw an 3105 // exception and the caller has no handler). Not legal, e.g., for passing 3106 // a null receiver to a v-call, or passing bad types to a slow-check-cast. 3107 // These cases MUST throw an exception via the runtime system, so the VM 3108 // will be looking for a table entry. 3109 Node *proj = in(0)->in(1); // Expect a proj feeding CatchNode 3110 CallNode *call; 3111 if (_con != TypeFunc::Control && // Bail out if not the main control. 3112 !(proj->is_Proj() && // AND NOT a rethrow 3113 proj->in(0)->is_Call() && 3114 (call = proj->in(0)->as_Call()) && 3115 call->entry_point() == OptoRuntime::rethrow_stub())) 3116 return this; 3117 3118 // Search for any other path being control 3119 for (uint i = 0; i < t->cnt(); i++) { 3120 if (i != _con && t->field_at(i) == Type::CONTROL) 3121 return this; 3122 } 3123 // Only my path is possible; I am identity on control to the jump 3124 return in(0)->in(0); 3125 } 3126 3127 3128 #ifndef PRODUCT 3129 void CatchProjNode::dump_spec(outputStream *st) const { 3130 ProjNode::dump_spec(st); 3131 st->print("@bci %d ",_handler_bci); 3132 } 3133 #endif 3134 3135 //============================================================================= 3136 //------------------------------Identity--------------------------------------- 3137 // Check for CreateEx being Identity. 3138 Node* CreateExNode::Identity(PhaseGVN* phase) { 3139 if( phase->type(in(1)) == Type::TOP ) return in(1); 3140 if( phase->type(in(0)) == Type::TOP ) return in(0); 3141 if (phase->type(in(0)->in(0)) == Type::TOP) { 3142 assert(in(0)->is_CatchProj(), "control is CatchProj"); 3143 return phase->C->top(); // dead code 3144 } 3145 // We only come from CatchProj, unless the CatchProj goes away. 3146 // If the CatchProj is optimized away, then we just carry the 3147 // exception oop through. 3148 3149 // CheckCastPPNode::Ideal() for inline types reuses the exception 3150 // paths of a call to perform an allocation: we can see a Phi here. 3151 if (in(1)->is_Phi()) { 3152 return this; 3153 } 3154 CallNode *call = in(1)->in(0)->as_Call(); 3155 3156 return (in(0)->is_CatchProj() && in(0)->in(0)->is_Catch() && 3157 in(0)->in(0)->in(1) == in(1)) ? this : call->in(TypeFunc::Parms); 3158 } 3159 3160 //============================================================================= 3161 //------------------------------Value------------------------------------------ 3162 // Check for being unreachable. 3163 const Type* NeverBranchNode::Value(PhaseGVN* phase) const { 3164 if (!in(0) || in(0)->is_top()) return Type::TOP; 3165 return bottom_type(); 3166 } 3167 3168 //------------------------------Ideal------------------------------------------ 3169 // Check for no longer being part of a loop 3170 Node *NeverBranchNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3171 if (can_reshape && !in(0)->is_Region()) { 3172 // Dead code elimination can sometimes delete this projection so 3173 // if it's not there, there's nothing to do. 3174 Node* fallthru = proj_out_or_null(0); 3175 if (fallthru != nullptr) { 3176 phase->is_IterGVN()->replace_node(fallthru, in(0)); 3177 } 3178 return phase->C->top(); 3179 } 3180 return nullptr; 3181 } 3182 3183 #ifndef PRODUCT 3184 void NeverBranchNode::format( PhaseRegAlloc *ra_, outputStream *st) const { 3185 st->print("%s", Name()); 3186 } 3187 #endif 3188 3189 #ifndef PRODUCT 3190 void BlackholeNode::format(PhaseRegAlloc* ra, outputStream* st) const { 3191 st->print("blackhole "); 3192 bool first = true; 3193 for (uint i = 0; i < req(); i++) { 3194 Node* n = in(i); 3195 if (n != nullptr && OptoReg::is_valid(ra->get_reg_first(n))) { 3196 if (first) { 3197 first = false; 3198 } else { 3199 st->print(", "); 3200 } 3201 char buf[128]; 3202 ra->dump_register(n, buf, sizeof(buf)); 3203 st->print("%s", buf); 3204 } 3205 } 3206 st->cr(); 3207 } 3208 #endif 3209