1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 // Portions of code courtesy of Clifford Click 26 27 // Optimization - Graph Style 28 29 #include "incls/_precompiled.incl" 30 #include "incls/_memnode.cpp.incl" 31 32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 33 34 //============================================================================= 35 uint MemNode::size_of() const { return sizeof(*this); } 36 37 const TypePtr *MemNode::adr_type() const { 38 Node* adr = in(Address); 39 const TypePtr* cross_check = NULL; 40 DEBUG_ONLY(cross_check = _adr_type); 41 return calculate_adr_type(adr->bottom_type(), cross_check); 42 } 43 44 #ifndef PRODUCT 45 void MemNode::dump_spec(outputStream *st) const { 46 if (in(Address) == NULL) return; // node is dead 47 #ifndef ASSERT 48 // fake the missing field 49 const TypePtr* _adr_type = NULL; 50 if (in(Address) != NULL) 51 _adr_type = in(Address)->bottom_type()->isa_ptr(); 52 #endif 53 dump_adr_type(this, _adr_type, st); 54 55 Compile* C = Compile::current(); 56 if( C->alias_type(_adr_type)->is_volatile() ) 57 st->print(" Volatile!"); 58 } 59 60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 61 st->print(" @"); 62 if (adr_type == NULL) { 63 st->print("NULL"); 64 } else { 65 adr_type->dump_on(st); 66 Compile* C = Compile::current(); 67 Compile::AliasType* atp = NULL; 68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 69 if (atp == NULL) 70 st->print(", idx=?\?;"); 71 else if (atp->index() == Compile::AliasIdxBot) 72 st->print(", idx=Bot;"); 73 else if (atp->index() == Compile::AliasIdxTop) 74 st->print(", idx=Top;"); 75 else if (atp->index() == Compile::AliasIdxRaw) 76 st->print(", idx=Raw;"); 77 else { 78 ciField* field = atp->field(); 79 if (field) { 80 st->print(", name="); 81 field->print_name_on(st); 82 } 83 st->print(", idx=%d;", atp->index()); 84 } 85 } 86 } 87 88 extern void print_alias_types(); 89 90 #endif 91 92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 93 const TypeOopPtr *tinst = t_adr->isa_oopptr(); 94 if (tinst == NULL || !tinst->is_known_instance_field()) 95 return mchain; // don't try to optimize non-instance types 96 uint instance_id = tinst->instance_id(); 97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 98 Node *prev = NULL; 99 Node *result = mchain; 100 while (prev != result) { 101 prev = result; 102 if (result == start_mem) 103 break; // hit one of our sentinels 104 // skip over a call which does not affect this memory slice 105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 106 Node *proj_in = result->in(0); 107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 108 break; // hit one of our sentinels 109 } else if (proj_in->is_Call()) { 110 CallNode *call = proj_in->as_Call(); 111 if (!call->may_modify(t_adr, phase)) { 112 result = call->in(TypeFunc::Memory); 113 } 114 } else if (proj_in->is_Initialize()) { 115 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 116 // Stop if this is the initialization for the object instance which 117 // which contains this memory slice, otherwise skip over it. 118 if (alloc != NULL && alloc->_idx != instance_id) { 119 result = proj_in->in(TypeFunc::Memory); 120 } 121 } else if (proj_in->is_MemBar()) { 122 result = proj_in->in(TypeFunc::Memory); 123 } else { 124 assert(false, "unexpected projection"); 125 } 126 } else if (result->is_MergeMem()) { 127 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 128 } 129 } 130 return result; 131 } 132 133 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 134 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 135 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); 136 PhaseIterGVN *igvn = phase->is_IterGVN(); 137 Node *result = mchain; 138 result = optimize_simple_memory_chain(result, t_adr, phase); 139 if (is_instance && igvn != NULL && result->is_Phi()) { 140 PhiNode *mphi = result->as_Phi(); 141 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 142 const TypePtr *t = mphi->adr_type(); 143 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 144 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 145 t->is_oopptr()->cast_to_exactness(true) 146 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 147 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 148 // clone the Phi with our address type 149 result = mphi->split_out_instance(t_adr, igvn); 150 } else { 151 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 152 } 153 } 154 return result; 155 } 156 157 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 158 uint alias_idx = phase->C->get_alias_index(tp); 159 Node *mem = mmem; 160 #ifdef ASSERT 161 { 162 // Check that current type is consistent with the alias index used during graph construction 163 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 164 bool consistent = adr_check == NULL || adr_check->empty() || 165 phase->C->must_alias(adr_check, alias_idx ); 166 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 167 if( !consistent && adr_check != NULL && !adr_check->empty() && 168 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 169 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 170 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 171 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 172 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 173 // don't assert if it is dead code. 174 consistent = true; 175 } 176 if( !consistent ) { 177 st->print("alias_idx==%d, adr_check==", alias_idx); 178 if( adr_check == NULL ) { 179 st->print("NULL"); 180 } else { 181 adr_check->dump(); 182 } 183 st->cr(); 184 print_alias_types(); 185 assert(consistent, "adr_check must match alias idx"); 186 } 187 } 188 #endif 189 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 190 // means an array I have not precisely typed yet. Do not do any 191 // alias stuff with it any time soon. 192 const TypeOopPtr *tinst = tp->isa_oopptr(); 193 if( tp->base() != Type::AnyPtr && 194 !(tinst && 195 tinst->klass()->is_java_lang_Object() && 196 tinst->offset() == Type::OffsetBot) ) { 197 // compress paths and change unreachable cycles to TOP 198 // If not, we can update the input infinitely along a MergeMem cycle 199 // Equivalent code in PhiNode::Ideal 200 Node* m = phase->transform(mmem); 201 // If transformed to a MergeMem, get the desired slice 202 // Otherwise the returned node represents memory for every slice 203 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 204 // Update input if it is progress over what we have now 205 } 206 return mem; 207 } 208 209 //--------------------------Ideal_common--------------------------------------- 210 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 211 // Unhook non-raw memories from complete (macro-expanded) initializations. 212 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 213 // If our control input is a dead region, kill all below the region 214 Node *ctl = in(MemNode::Control); 215 if (ctl && remove_dead_region(phase, can_reshape)) 216 return this; 217 ctl = in(MemNode::Control); 218 // Don't bother trying to transform a dead node 219 if( ctl && ctl->is_top() ) return NodeSentinel; 220 221 // Ignore if memory is dead, or self-loop 222 Node *mem = in(MemNode::Memory); 223 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 224 assert( mem != this, "dead loop in MemNode::Ideal" ); 225 226 Node *address = in(MemNode::Address); 227 const Type *t_adr = phase->type( address ); 228 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 229 230 PhaseIterGVN *igvn = phase->is_IterGVN(); 231 if( can_reshape && igvn != NULL && igvn->_worklist.member(address) ) { 232 // The address's base and type may change when the address is processed. 233 // Delay this mem node transformation until the address is processed. 234 phase->is_IterGVN()->_worklist.push(this); 235 return NodeSentinel; // caller will return NULL 236 } 237 238 // Avoid independent memory operations 239 Node* old_mem = mem; 240 241 // The code which unhooks non-raw memories from complete (macro-expanded) 242 // initializations was removed. After macro-expansion all stores catched 243 // by Initialize node became raw stores and there is no information 244 // which memory slices they modify. So it is unsafe to move any memory 245 // operation above these stores. Also in most cases hooked non-raw memories 246 // were already unhooked by using information from detect_ptr_independence() 247 // and find_previous_store(). 248 249 if (mem->is_MergeMem()) { 250 MergeMemNode* mmem = mem->as_MergeMem(); 251 const TypePtr *tp = t_adr->is_ptr(); 252 253 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 254 } 255 256 if (mem != old_mem) { 257 set_req(MemNode::Memory, mem); 258 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 259 return this; 260 } 261 262 // let the subclass continue analyzing... 263 return NULL; 264 } 265 266 // Helper function for proving some simple control dominations. 267 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 268 // Already assumes that 'dom' is available at 'sub', and that 'sub' 269 // is not a constant (dominated by the method's StartNode). 270 // Used by MemNode::find_previous_store to prove that the 271 // control input of a memory operation predates (dominates) 272 // an allocation it wants to look past. 273 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 274 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 275 return false; // Conservative answer for dead code 276 277 // Check 'dom'. Skip Proj and CatchProj nodes. 278 dom = dom->find_exact_control(dom); 279 if (dom == NULL || dom->is_top()) 280 return false; // Conservative answer for dead code 281 282 if (dom == sub) { 283 // For the case when, for example, 'sub' is Initialize and the original 284 // 'dom' is Proj node of the 'sub'. 285 return false; 286 } 287 288 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 289 return true; 290 291 // 'dom' dominates 'sub' if its control edge and control edges 292 // of all its inputs dominate or equal to sub's control edge. 293 294 // Currently 'sub' is either Allocate, Initialize or Start nodes. 295 // Or Region for the check in LoadNode::Ideal(); 296 // 'sub' should have sub->in(0) != NULL. 297 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 298 sub->is_Region(), "expecting only these nodes"); 299 300 // Get control edge of 'sub'. 301 Node* orig_sub = sub; 302 sub = sub->find_exact_control(sub->in(0)); 303 if (sub == NULL || sub->is_top()) 304 return false; // Conservative answer for dead code 305 306 assert(sub->is_CFG(), "expecting control"); 307 308 if (sub == dom) 309 return true; 310 311 if (sub->is_Start() || sub->is_Root()) 312 return false; 313 314 { 315 // Check all control edges of 'dom'. 316 317 ResourceMark rm; 318 Arena* arena = Thread::current()->resource_area(); 319 Node_List nlist(arena); 320 Unique_Node_List dom_list(arena); 321 322 dom_list.push(dom); 323 bool only_dominating_controls = false; 324 325 for (uint next = 0; next < dom_list.size(); next++) { 326 Node* n = dom_list.at(next); 327 if (n == orig_sub) 328 return false; // One of dom's inputs dominated by sub. 329 if (!n->is_CFG() && n->pinned()) { 330 // Check only own control edge for pinned non-control nodes. 331 n = n->find_exact_control(n->in(0)); 332 if (n == NULL || n->is_top()) 333 return false; // Conservative answer for dead code 334 assert(n->is_CFG(), "expecting control"); 335 dom_list.push(n); 336 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 337 only_dominating_controls = true; 338 } else if (n->is_CFG()) { 339 if (n->dominates(sub, nlist)) 340 only_dominating_controls = true; 341 else 342 return false; 343 } else { 344 // First, own control edge. 345 Node* m = n->find_exact_control(n->in(0)); 346 if (m != NULL) { 347 if (m->is_top()) 348 return false; // Conservative answer for dead code 349 dom_list.push(m); 350 } 351 // Now, the rest of edges. 352 uint cnt = n->req(); 353 for (uint i = 1; i < cnt; i++) { 354 m = n->find_exact_control(n->in(i)); 355 if (m == NULL || m->is_top()) 356 continue; 357 dom_list.push(m); 358 } 359 } 360 } 361 return only_dominating_controls; 362 } 363 } 364 365 //---------------------detect_ptr_independence--------------------------------- 366 // Used by MemNode::find_previous_store to prove that two base 367 // pointers are never equal. 368 // The pointers are accompanied by their associated allocations, 369 // if any, which have been previously discovered by the caller. 370 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 371 Node* p2, AllocateNode* a2, 372 PhaseTransform* phase) { 373 // Attempt to prove that these two pointers cannot be aliased. 374 // They may both manifestly be allocations, and they should differ. 375 // Or, if they are not both allocations, they can be distinct constants. 376 // Otherwise, one is an allocation and the other a pre-existing value. 377 if (a1 == NULL && a2 == NULL) { // neither an allocation 378 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 379 } else if (a1 != NULL && a2 != NULL) { // both allocations 380 return (a1 != a2); 381 } else if (a1 != NULL) { // one allocation a1 382 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 383 return all_controls_dominate(p2, a1); 384 } else { //(a2 != NULL) // one allocation a2 385 return all_controls_dominate(p1, a2); 386 } 387 return false; 388 } 389 390 391 // The logic for reordering loads and stores uses four steps: 392 // (a) Walk carefully past stores and initializations which we 393 // can prove are independent of this load. 394 // (b) Observe that the next memory state makes an exact match 395 // with self (load or store), and locate the relevant store. 396 // (c) Ensure that, if we were to wire self directly to the store, 397 // the optimizer would fold it up somehow. 398 // (d) Do the rewiring, and return, depending on some other part of 399 // the optimizer to fold up the load. 400 // This routine handles steps (a) and (b). Steps (c) and (d) are 401 // specific to loads and stores, so they are handled by the callers. 402 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 403 // 404 Node* MemNode::find_previous_store(PhaseTransform* phase) { 405 Node* ctrl = in(MemNode::Control); 406 Node* adr = in(MemNode::Address); 407 intptr_t offset = 0; 408 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 409 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 410 411 if (offset == Type::OffsetBot) 412 return NULL; // cannot unalias unless there are precise offsets 413 414 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 415 416 intptr_t size_in_bytes = memory_size(); 417 418 Node* mem = in(MemNode::Memory); // start searching here... 419 420 int cnt = 50; // Cycle limiter 421 for (;;) { // While we can dance past unrelated stores... 422 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 423 424 if (mem->is_Store()) { 425 Node* st_adr = mem->in(MemNode::Address); 426 intptr_t st_offset = 0; 427 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 428 if (st_base == NULL) 429 break; // inscrutable pointer 430 if (st_offset != offset && st_offset != Type::OffsetBot) { 431 const int MAX_STORE = BytesPerLong; 432 if (st_offset >= offset + size_in_bytes || 433 st_offset <= offset - MAX_STORE || 434 st_offset <= offset - mem->as_Store()->memory_size()) { 435 // Success: The offsets are provably independent. 436 // (You may ask, why not just test st_offset != offset and be done? 437 // The answer is that stores of different sizes can co-exist 438 // in the same sequence of RawMem effects. We sometimes initialize 439 // a whole 'tile' of array elements with a single jint or jlong.) 440 mem = mem->in(MemNode::Memory); 441 continue; // (a) advance through independent store memory 442 } 443 } 444 if (st_base != base && 445 detect_ptr_independence(base, alloc, 446 st_base, 447 AllocateNode::Ideal_allocation(st_base, phase), 448 phase)) { 449 // Success: The bases are provably independent. 450 mem = mem->in(MemNode::Memory); 451 continue; // (a) advance through independent store memory 452 } 453 454 // (b) At this point, if the bases or offsets do not agree, we lose, 455 // since we have not managed to prove 'this' and 'mem' independent. 456 if (st_base == base && st_offset == offset) { 457 return mem; // let caller handle steps (c), (d) 458 } 459 460 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 461 InitializeNode* st_init = mem->in(0)->as_Initialize(); 462 AllocateNode* st_alloc = st_init->allocation(); 463 if (st_alloc == NULL) 464 break; // something degenerated 465 bool known_identical = false; 466 bool known_independent = false; 467 if (alloc == st_alloc) 468 known_identical = true; 469 else if (alloc != NULL) 470 known_independent = true; 471 else if (all_controls_dominate(this, st_alloc)) 472 known_independent = true; 473 474 if (known_independent) { 475 // The bases are provably independent: Either they are 476 // manifestly distinct allocations, or else the control 477 // of this load dominates the store's allocation. 478 int alias_idx = phase->C->get_alias_index(adr_type()); 479 if (alias_idx == Compile::AliasIdxRaw) { 480 mem = st_alloc->in(TypeFunc::Memory); 481 } else { 482 mem = st_init->memory(alias_idx); 483 } 484 continue; // (a) advance through independent store memory 485 } 486 487 // (b) at this point, if we are not looking at a store initializing 488 // the same allocation we are loading from, we lose. 489 if (known_identical) { 490 // From caller, can_see_stored_value will consult find_captured_store. 491 return mem; // let caller handle steps (c), (d) 492 } 493 494 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 495 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 496 if (mem->is_Proj() && mem->in(0)->is_Call()) { 497 CallNode *call = mem->in(0)->as_Call(); 498 if (!call->may_modify(addr_t, phase)) { 499 mem = call->in(TypeFunc::Memory); 500 continue; // (a) advance through independent call memory 501 } 502 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 503 mem = mem->in(0)->in(TypeFunc::Memory); 504 continue; // (a) advance through independent MemBar memory 505 } else if (mem->is_MergeMem()) { 506 int alias_idx = phase->C->get_alias_index(adr_type()); 507 mem = mem->as_MergeMem()->memory_at(alias_idx); 508 continue; // (a) advance through independent MergeMem memory 509 } 510 } 511 512 // Unless there is an explicit 'continue', we must bail out here, 513 // because 'mem' is an inscrutable memory state (e.g., a call). 514 break; 515 } 516 517 return NULL; // bail out 518 } 519 520 //----------------------calculate_adr_type------------------------------------- 521 // Helper function. Notices when the given type of address hits top or bottom. 522 // Also, asserts a cross-check of the type against the expected address type. 523 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 524 if (t == Type::TOP) return NULL; // does not touch memory any more? 525 #ifdef PRODUCT 526 cross_check = NULL; 527 #else 528 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 529 #endif 530 const TypePtr* tp = t->isa_ptr(); 531 if (tp == NULL) { 532 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 533 return TypePtr::BOTTOM; // touches lots of memory 534 } else { 535 #ifdef ASSERT 536 // %%%% [phh] We don't check the alias index if cross_check is 537 // TypeRawPtr::BOTTOM. Needs to be investigated. 538 if (cross_check != NULL && 539 cross_check != TypePtr::BOTTOM && 540 cross_check != TypeRawPtr::BOTTOM) { 541 // Recheck the alias index, to see if it has changed (due to a bug). 542 Compile* C = Compile::current(); 543 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 544 "must stay in the original alias category"); 545 // The type of the address must be contained in the adr_type, 546 // disregarding "null"-ness. 547 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 548 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 549 assert(cross_check->meet(tp_notnull) == cross_check, 550 "real address must not escape from expected memory type"); 551 } 552 #endif 553 return tp; 554 } 555 } 556 557 //------------------------adr_phi_is_loop_invariant---------------------------- 558 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 559 // loop is loop invariant. Make a quick traversal of Phi and associated 560 // CastPP nodes, looking to see if they are a closed group within the loop. 561 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 562 // The idea is that the phi-nest must boil down to only CastPP nodes 563 // with the same data. This implies that any path into the loop already 564 // includes such a CastPP, and so the original cast, whatever its input, 565 // must be covered by an equivalent cast, with an earlier control input. 566 ResourceMark rm; 567 568 // The loop entry input of the phi should be the unique dominating 569 // node for every Phi/CastPP in the loop. 570 Unique_Node_List closure; 571 closure.push(adr_phi->in(LoopNode::EntryControl)); 572 573 // Add the phi node and the cast to the worklist. 574 Unique_Node_List worklist; 575 worklist.push(adr_phi); 576 if( cast != NULL ){ 577 if( !cast->is_ConstraintCast() ) return false; 578 worklist.push(cast); 579 } 580 581 // Begin recursive walk of phi nodes. 582 while( worklist.size() ){ 583 // Take a node off the worklist 584 Node *n = worklist.pop(); 585 if( !closure.member(n) ){ 586 // Add it to the closure. 587 closure.push(n); 588 // Make a sanity check to ensure we don't waste too much time here. 589 if( closure.size() > 20) return false; 590 // This node is OK if: 591 // - it is a cast of an identical value 592 // - or it is a phi node (then we add its inputs to the worklist) 593 // Otherwise, the node is not OK, and we presume the cast is not invariant 594 if( n->is_ConstraintCast() ){ 595 worklist.push(n->in(1)); 596 } else if( n->is_Phi() ) { 597 for( uint i = 1; i < n->req(); i++ ) { 598 worklist.push(n->in(i)); 599 } 600 } else { 601 return false; 602 } 603 } 604 } 605 606 // Quit when the worklist is empty, and we've found no offending nodes. 607 return true; 608 } 609 610 //------------------------------Ideal_DU_postCCP------------------------------- 611 // Find any cast-away of null-ness and keep its control. Null cast-aways are 612 // going away in this pass and we need to make this memory op depend on the 613 // gating null check. 614 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 615 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 616 } 617 618 // I tried to leave the CastPP's in. This makes the graph more accurate in 619 // some sense; we get to keep around the knowledge that an oop is not-null 620 // after some test. Alas, the CastPP's interfere with GVN (some values are 621 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 622 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 623 // some of the more trivial cases in the optimizer. Removing more useless 624 // Phi's started allowing Loads to illegally float above null checks. I gave 625 // up on this approach. CNC 10/20/2000 626 // This static method may be called not from MemNode (EncodePNode calls it). 627 // Only the control edge of the node 'n' might be updated. 628 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 629 Node *skipped_cast = NULL; 630 // Need a null check? Regular static accesses do not because they are 631 // from constant addresses. Array ops are gated by the range check (which 632 // always includes a NULL check). Just check field ops. 633 if( n->in(MemNode::Control) == NULL ) { 634 // Scan upwards for the highest location we can place this memory op. 635 while( true ) { 636 switch( adr->Opcode() ) { 637 638 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 639 adr = adr->in(AddPNode::Base); 640 continue; 641 642 case Op_DecodeN: // No change to NULL-ness, so peek thru 643 adr = adr->in(1); 644 continue; 645 646 case Op_CastPP: 647 // If the CastPP is useless, just peek on through it. 648 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 649 // Remember the cast that we've peeked though. If we peek 650 // through more than one, then we end up remembering the highest 651 // one, that is, if in a loop, the one closest to the top. 652 skipped_cast = adr; 653 adr = adr->in(1); 654 continue; 655 } 656 // CastPP is going away in this pass! We need this memory op to be 657 // control-dependent on the test that is guarding the CastPP. 658 ccp->hash_delete(n); 659 n->set_req(MemNode::Control, adr->in(0)); 660 ccp->hash_insert(n); 661 return n; 662 663 case Op_Phi: 664 // Attempt to float above a Phi to some dominating point. 665 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 666 // If we've already peeked through a Cast (which could have set the 667 // control), we can't float above a Phi, because the skipped Cast 668 // may not be loop invariant. 669 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 670 adr = adr->in(1); 671 continue; 672 } 673 } 674 675 // Intentional fallthrough! 676 677 // No obvious dominating point. The mem op is pinned below the Phi 678 // by the Phi itself. If the Phi goes away (no true value is merged) 679 // then the mem op can float, but not indefinitely. It must be pinned 680 // behind the controls leading to the Phi. 681 case Op_CheckCastPP: 682 // These usually stick around to change address type, however a 683 // useless one can be elided and we still need to pick up a control edge 684 if (adr->in(0) == NULL) { 685 // This CheckCastPP node has NO control and is likely useless. But we 686 // need check further up the ancestor chain for a control input to keep 687 // the node in place. 4959717. 688 skipped_cast = adr; 689 adr = adr->in(1); 690 continue; 691 } 692 ccp->hash_delete(n); 693 n->set_req(MemNode::Control, adr->in(0)); 694 ccp->hash_insert(n); 695 return n; 696 697 // List of "safe" opcodes; those that implicitly block the memory 698 // op below any null check. 699 case Op_CastX2P: // no null checks on native pointers 700 case Op_Parm: // 'this' pointer is not null 701 case Op_LoadP: // Loading from within a klass 702 case Op_LoadN: // Loading from within a klass 703 case Op_LoadKlass: // Loading from within a klass 704 case Op_LoadNKlass: // Loading from within a klass 705 case Op_ConP: // Loading from a klass 706 case Op_ConN: // Loading from a klass 707 case Op_CreateEx: // Sucking up the guts of an exception oop 708 case Op_Con: // Reading from TLS 709 case Op_CMoveP: // CMoveP is pinned 710 case Op_CMoveN: // CMoveN is pinned 711 break; // No progress 712 713 case Op_Proj: // Direct call to an allocation routine 714 case Op_SCMemProj: // Memory state from store conditional ops 715 #ifdef ASSERT 716 { 717 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 718 const Node* call = adr->in(0); 719 if (call->is_CallJava()) { 720 const CallJavaNode* call_java = call->as_CallJava(); 721 const TypeTuple *r = call_java->tf()->range(); 722 assert(r->cnt() > TypeFunc::Parms, "must return value"); 723 const Type* ret_type = r->field_at(TypeFunc::Parms); 724 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 725 // We further presume that this is one of 726 // new_instance_Java, new_array_Java, or 727 // the like, but do not assert for this. 728 } else if (call->is_Allocate()) { 729 // similar case to new_instance_Java, etc. 730 } else if (!call->is_CallLeaf()) { 731 // Projections from fetch_oop (OSR) are allowed as well. 732 ShouldNotReachHere(); 733 } 734 } 735 #endif 736 break; 737 default: 738 ShouldNotReachHere(); 739 } 740 break; 741 } 742 } 743 744 return NULL; // No progress 745 } 746 747 748 //============================================================================= 749 uint LoadNode::size_of() const { return sizeof(*this); } 750 uint LoadNode::cmp( const Node &n ) const 751 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 752 const Type *LoadNode::bottom_type() const { return _type; } 753 uint LoadNode::ideal_reg() const { 754 return Matcher::base2reg[_type->base()]; 755 } 756 757 #ifndef PRODUCT 758 void LoadNode::dump_spec(outputStream *st) const { 759 MemNode::dump_spec(st); 760 if( !Verbose && !WizardMode ) { 761 // standard dump does this in Verbose and WizardMode 762 st->print(" #"); _type->dump_on(st); 763 } 764 } 765 #endif 766 767 768 //----------------------------LoadNode::make----------------------------------- 769 // Polymorphic factory method: 770 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 771 Compile* C = gvn.C; 772 773 // sanity check the alias category against the created node type 774 assert(!(adr_type->isa_oopptr() && 775 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 776 "use LoadKlassNode instead"); 777 assert(!(adr_type->isa_aryptr() && 778 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 779 "use LoadRangeNode instead"); 780 switch (bt) { 781 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 782 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 783 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 784 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 785 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 786 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 787 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 788 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 789 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 790 case T_OBJECT: 791 #ifdef _LP64 792 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 793 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 794 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 795 } else 796 #endif 797 { 798 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 799 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 800 } 801 } 802 ShouldNotReachHere(); 803 return (LoadNode*)NULL; 804 } 805 806 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 807 bool require_atomic = true; 808 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 809 } 810 811 812 813 814 //------------------------------hash------------------------------------------- 815 uint LoadNode::hash() const { 816 // unroll addition of interesting fields 817 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 818 } 819 820 //---------------------------can_see_stored_value------------------------------ 821 // This routine exists to make sure this set of tests is done the same 822 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 823 // will change the graph shape in a way which makes memory alive twice at the 824 // same time (uses the Oracle model of aliasing), then some 825 // LoadXNode::Identity will fold things back to the equivalence-class model 826 // of aliasing. 827 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 828 Node* ld_adr = in(MemNode::Address); 829 830 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 831 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 832 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 833 atp->field() != NULL && !atp->field()->is_volatile()) { 834 uint alias_idx = atp->index(); 835 bool final = atp->field()->is_final(); 836 Node* result = NULL; 837 Node* current = st; 838 // Skip through chains of MemBarNodes checking the MergeMems for 839 // new states for the slice of this load. Stop once any other 840 // kind of node is encountered. Loads from final memory can skip 841 // through any kind of MemBar but normal loads shouldn't skip 842 // through MemBarAcquire since the could allow them to move out of 843 // a synchronized region. 844 while (current->is_Proj()) { 845 int opc = current->in(0)->Opcode(); 846 if ((final && opc == Op_MemBarAcquire) || 847 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 848 Node* mem = current->in(0)->in(TypeFunc::Memory); 849 if (mem->is_MergeMem()) { 850 MergeMemNode* merge = mem->as_MergeMem(); 851 Node* new_st = merge->memory_at(alias_idx); 852 if (new_st == merge->base_memory()) { 853 // Keep searching 854 current = merge->base_memory(); 855 continue; 856 } 857 // Save the new memory state for the slice and fall through 858 // to exit. 859 result = new_st; 860 } 861 } 862 break; 863 } 864 if (result != NULL) { 865 st = result; 866 } 867 } 868 869 870 // Loop around twice in the case Load -> Initialize -> Store. 871 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 872 for (int trip = 0; trip <= 1; trip++) { 873 874 if (st->is_Store()) { 875 Node* st_adr = st->in(MemNode::Address); 876 if (!phase->eqv(st_adr, ld_adr)) { 877 // Try harder before giving up... Match raw and non-raw pointers. 878 intptr_t st_off = 0; 879 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 880 if (alloc == NULL) return NULL; 881 intptr_t ld_off = 0; 882 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 883 if (alloc != allo2) return NULL; 884 if (ld_off != st_off) return NULL; 885 // At this point we have proven something like this setup: 886 // A = Allocate(...) 887 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 888 // S = StoreQ(, AddP(, A.Parm , #Off), V) 889 // (Actually, we haven't yet proven the Q's are the same.) 890 // In other words, we are loading from a casted version of 891 // the same pointer-and-offset that we stored to. 892 // Thus, we are able to replace L by V. 893 } 894 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 895 if (store_Opcode() != st->Opcode()) 896 return NULL; 897 return st->in(MemNode::ValueIn); 898 } 899 900 intptr_t offset = 0; // scratch 901 902 // A load from a freshly-created object always returns zero. 903 // (This can happen after LoadNode::Ideal resets the load's memory input 904 // to find_captured_store, which returned InitializeNode::zero_memory.) 905 if (st->is_Proj() && st->in(0)->is_Allocate() && 906 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 907 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 908 // return a zero value for the load's basic type 909 // (This is one of the few places where a generic PhaseTransform 910 // can create new nodes. Think of it as lazily manifesting 911 // virtually pre-existing constants.) 912 return phase->zerocon(memory_type()); 913 } 914 915 // A load from an initialization barrier can match a captured store. 916 if (st->is_Proj() && st->in(0)->is_Initialize()) { 917 InitializeNode* init = st->in(0)->as_Initialize(); 918 AllocateNode* alloc = init->allocation(); 919 if (alloc != NULL && 920 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 921 // examine a captured store value 922 st = init->find_captured_store(offset, memory_size(), phase); 923 if (st != NULL) 924 continue; // take one more trip around 925 } 926 } 927 928 break; 929 } 930 931 return NULL; 932 } 933 934 //----------------------is_instance_field_load_with_local_phi------------------ 935 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 936 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 937 in(MemNode::Address)->is_AddP() ) { 938 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 939 // Only instances. 940 if( t_oop != NULL && t_oop->is_known_instance_field() && 941 t_oop->offset() != Type::OffsetBot && 942 t_oop->offset() != Type::OffsetTop) { 943 return true; 944 } 945 } 946 return false; 947 } 948 949 //------------------------------Identity--------------------------------------- 950 // Loads are identity if previous store is to same address 951 Node *LoadNode::Identity( PhaseTransform *phase ) { 952 // If the previous store-maker is the right kind of Store, and the store is 953 // to the same address, then we are equal to the value stored. 954 Node* mem = in(MemNode::Memory); 955 Node* value = can_see_stored_value(mem, phase); 956 if( value ) { 957 // byte, short & char stores truncate naturally. 958 // A load has to load the truncated value which requires 959 // some sort of masking operation and that requires an 960 // Ideal call instead of an Identity call. 961 if (memory_size() < BytesPerInt) { 962 // If the input to the store does not fit with the load's result type, 963 // it must be truncated via an Ideal call. 964 if (!phase->type(value)->higher_equal(phase->type(this))) 965 return this; 966 } 967 // (This works even when value is a Con, but LoadNode::Value 968 // usually runs first, producing the singleton type of the Con.) 969 return value; 970 } 971 972 // Search for an existing data phi which was generated before for the same 973 // instance's field to avoid infinite generation of phis in a loop. 974 Node *region = mem->in(0); 975 if (is_instance_field_load_with_local_phi(region)) { 976 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 977 int this_index = phase->C->get_alias_index(addr_t); 978 int this_offset = addr_t->offset(); 979 int this_id = addr_t->is_oopptr()->instance_id(); 980 const Type* this_type = bottom_type(); 981 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 982 Node* phi = region->fast_out(i); 983 if (phi->is_Phi() && phi != mem && 984 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 985 return phi; 986 } 987 } 988 } 989 990 return this; 991 } 992 993 994 // Returns true if the AliasType refers to the field that holds the 995 // cached box array. Currently only handles the IntegerCache case. 996 static bool is_autobox_cache(Compile::AliasType* atp) { 997 if (atp != NULL && atp->field() != NULL) { 998 ciField* field = atp->field(); 999 ciSymbol* klass = field->holder()->name(); 1000 if (field->name() == ciSymbol::cache_field_name() && 1001 field->holder()->uses_default_loader() && 1002 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1003 return true; 1004 } 1005 } 1006 return false; 1007 } 1008 1009 // Fetch the base value in the autobox array 1010 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1011 if (atp != NULL && atp->field() != NULL) { 1012 ciField* field = atp->field(); 1013 ciSymbol* klass = field->holder()->name(); 1014 if (field->name() == ciSymbol::cache_field_name() && 1015 field->holder()->uses_default_loader() && 1016 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1017 assert(field->is_constant(), "what?"); 1018 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1019 // Fetch the box object at the base of the array and get its value 1020 ciInstance* box = array->obj_at(0)->as_instance(); 1021 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1022 if (ik->nof_nonstatic_fields() == 1) { 1023 // This should be true nonstatic_field_at requires calling 1024 // nof_nonstatic_fields so check it anyway 1025 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1026 cache_offset = c.as_int(); 1027 } 1028 return true; 1029 } 1030 } 1031 return false; 1032 } 1033 1034 // Returns true if the AliasType refers to the value field of an 1035 // autobox object. Currently only handles Integer. 1036 static bool is_autobox_object(Compile::AliasType* atp) { 1037 if (atp != NULL && atp->field() != NULL) { 1038 ciField* field = atp->field(); 1039 ciSymbol* klass = field->holder()->name(); 1040 if (field->name() == ciSymbol::value_name() && 1041 field->holder()->uses_default_loader() && 1042 klass == ciSymbol::java_lang_Integer()) { 1043 return true; 1044 } 1045 } 1046 return false; 1047 } 1048 1049 1050 // We're loading from an object which has autobox behaviour. 1051 // If this object is result of a valueOf call we'll have a phi 1052 // merging a newly allocated object and a load from the cache. 1053 // We want to replace this load with the original incoming 1054 // argument to the valueOf call. 1055 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1056 Node* base = in(Address)->in(AddPNode::Base); 1057 if (base->is_Phi() && base->req() == 3) { 1058 AllocateNode* allocation = NULL; 1059 int allocation_index = -1; 1060 int load_index = -1; 1061 for (uint i = 1; i < base->req(); i++) { 1062 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1063 if (allocation != NULL) { 1064 allocation_index = i; 1065 load_index = 3 - allocation_index; 1066 break; 1067 } 1068 } 1069 bool has_load = ( allocation != NULL && 1070 (base->in(load_index)->is_Load() || 1071 base->in(load_index)->is_DecodeN() && 1072 base->in(load_index)->in(1)->is_Load()) ); 1073 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1074 // Push the loads from the phi that comes from valueOf up 1075 // through it to allow elimination of the loads and the recovery 1076 // of the original value. 1077 Node* mem_phi = in(Memory); 1078 Node* offset = in(Address)->in(AddPNode::Offset); 1079 Node* region = base->in(0); 1080 1081 Node* in1 = clone(); 1082 Node* in1_addr = in1->in(Address)->clone(); 1083 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1084 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1085 in1_addr->set_req(AddPNode::Offset, offset); 1086 in1->set_req(0, region->in(allocation_index)); 1087 in1->set_req(Address, in1_addr); 1088 in1->set_req(Memory, mem_phi->in(allocation_index)); 1089 1090 Node* in2 = clone(); 1091 Node* in2_addr = in2->in(Address)->clone(); 1092 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1093 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1094 in2_addr->set_req(AddPNode::Offset, offset); 1095 in2->set_req(0, region->in(load_index)); 1096 in2->set_req(Address, in2_addr); 1097 in2->set_req(Memory, mem_phi->in(load_index)); 1098 1099 in1_addr = phase->transform(in1_addr); 1100 in1 = phase->transform(in1); 1101 in2_addr = phase->transform(in2_addr); 1102 in2 = phase->transform(in2); 1103 1104 PhiNode* result = PhiNode::make_blank(region, this); 1105 result->set_req(allocation_index, in1); 1106 result->set_req(load_index, in2); 1107 return result; 1108 } 1109 } else if (base->is_Load() || 1110 base->is_DecodeN() && base->in(1)->is_Load()) { 1111 if (base->is_DecodeN()) { 1112 // Get LoadN node which loads cached Integer object 1113 base = base->in(1); 1114 } 1115 // Eliminate the load of Integer.value for integers from the cache 1116 // array by deriving the value from the index into the array. 1117 // Capture the offset of the load and then reverse the computation. 1118 Node* load_base = base->in(Address)->in(AddPNode::Base); 1119 if (load_base->is_DecodeN()) { 1120 // Get LoadN node which loads IntegerCache.cache field 1121 load_base = load_base->in(1); 1122 } 1123 if (load_base != NULL) { 1124 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1125 intptr_t cache_offset; 1126 int shift = -1; 1127 Node* cache = NULL; 1128 if (is_autobox_cache(atp)) { 1129 shift = exact_log2(type2aelembytes(T_OBJECT)); 1130 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1131 } 1132 if (cache != NULL && base->in(Address)->is_AddP()) { 1133 Node* elements[4]; 1134 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1135 int cache_low; 1136 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1137 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1138 // Add up all the offsets making of the address of the load 1139 Node* result = elements[0]; 1140 for (int i = 1; i < count; i++) { 1141 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1142 } 1143 // Remove the constant offset from the address and then 1144 // remove the scaling of the offset to recover the original index. 1145 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1146 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1147 // Peel the shift off directly but wrap it in a dummy node 1148 // since Ideal can't return existing nodes 1149 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1150 } else { 1151 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1152 } 1153 #ifdef _LP64 1154 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1155 #endif 1156 return result; 1157 } 1158 } 1159 } 1160 } 1161 return NULL; 1162 } 1163 1164 //------------------------------split_through_phi------------------------------ 1165 // Split instance field load through Phi. 1166 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1167 Node* mem = in(MemNode::Memory); 1168 Node* address = in(MemNode::Address); 1169 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1170 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1171 1172 assert(mem->is_Phi() && (t_oop != NULL) && 1173 t_oop->is_known_instance_field(), "invalide conditions"); 1174 1175 Node *region = mem->in(0); 1176 if (region == NULL) { 1177 return NULL; // Wait stable graph 1178 } 1179 uint cnt = mem->req(); 1180 for( uint i = 1; i < cnt; i++ ) { 1181 Node *in = mem->in(i); 1182 if( in == NULL ) { 1183 return NULL; // Wait stable graph 1184 } 1185 } 1186 // Check for loop invariant. 1187 if (cnt == 3) { 1188 for( uint i = 1; i < cnt; i++ ) { 1189 Node *in = mem->in(i); 1190 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1191 if (m == mem) { 1192 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1193 return this; 1194 } 1195 } 1196 } 1197 // Split through Phi (see original code in loopopts.cpp). 1198 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1199 1200 // Do nothing here if Identity will find a value 1201 // (to avoid infinite chain of value phis generation). 1202 if ( !phase->eqv(this, this->Identity(phase)) ) 1203 return NULL; 1204 1205 // Skip the split if the region dominates some control edge of the address. 1206 if (cnt == 3 && !MemNode::all_controls_dominate(address, region)) 1207 return NULL; 1208 1209 const Type* this_type = this->bottom_type(); 1210 int this_index = phase->C->get_alias_index(addr_t); 1211 int this_offset = addr_t->offset(); 1212 int this_iid = addr_t->is_oopptr()->instance_id(); 1213 int wins = 0; 1214 PhaseIterGVN *igvn = phase->is_IterGVN(); 1215 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1216 for( uint i = 1; i < region->req(); i++ ) { 1217 Node *x; 1218 Node* the_clone = NULL; 1219 if( region->in(i) == phase->C->top() ) { 1220 x = phase->C->top(); // Dead path? Use a dead data op 1221 } else { 1222 x = this->clone(); // Else clone up the data op 1223 the_clone = x; // Remember for possible deletion. 1224 // Alter data node to use pre-phi inputs 1225 if( this->in(0) == region ) { 1226 x->set_req( 0, region->in(i) ); 1227 } else { 1228 x->set_req( 0, NULL ); 1229 } 1230 for( uint j = 1; j < this->req(); j++ ) { 1231 Node *in = this->in(j); 1232 if( in->is_Phi() && in->in(0) == region ) 1233 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 1234 } 1235 } 1236 // Check for a 'win' on some paths 1237 const Type *t = x->Value(igvn); 1238 1239 bool singleton = t->singleton(); 1240 1241 // See comments in PhaseIdealLoop::split_thru_phi(). 1242 if( singleton && t == Type::TOP ) { 1243 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1244 } 1245 1246 if( singleton ) { 1247 wins++; 1248 x = igvn->makecon(t); 1249 } else { 1250 // We now call Identity to try to simplify the cloned node. 1251 // Note that some Identity methods call phase->type(this). 1252 // Make sure that the type array is big enough for 1253 // our new node, even though we may throw the node away. 1254 // (This tweaking with igvn only works because x is a new node.) 1255 igvn->set_type(x, t); 1256 // If x is a TypeNode, capture any more-precise type permanently into Node 1257 // otherwise it will be not updated during igvn->transform since 1258 // igvn->type(x) is set to x->Value() already. 1259 x->raise_bottom_type(t); 1260 Node *y = x->Identity(igvn); 1261 if( y != x ) { 1262 wins++; 1263 x = y; 1264 } else { 1265 y = igvn->hash_find(x); 1266 if( y ) { 1267 wins++; 1268 x = y; 1269 } else { 1270 // Else x is a new node we are keeping 1271 // We do not need register_new_node_with_optimizer 1272 // because set_type has already been called. 1273 igvn->_worklist.push(x); 1274 } 1275 } 1276 } 1277 if (x != the_clone && the_clone != NULL) 1278 igvn->remove_dead_node(the_clone); 1279 phi->set_req(i, x); 1280 } 1281 if( wins > 0 ) { 1282 // Record Phi 1283 igvn->register_new_node_with_optimizer(phi); 1284 return phi; 1285 } 1286 igvn->remove_dead_node(phi); 1287 return NULL; 1288 } 1289 1290 //------------------------------Ideal------------------------------------------ 1291 // If the load is from Field memory and the pointer is non-null, we can 1292 // zero out the control input. 1293 // If the offset is constant and the base is an object allocation, 1294 // try to hook me up to the exact initializing store. 1295 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1296 Node* p = MemNode::Ideal_common(phase, can_reshape); 1297 if (p) return (p == NodeSentinel) ? NULL : p; 1298 1299 Node* ctrl = in(MemNode::Control); 1300 Node* address = in(MemNode::Address); 1301 1302 // Skip up past a SafePoint control. Cannot do this for Stores because 1303 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1304 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1305 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1306 ctrl = ctrl->in(0); 1307 set_req(MemNode::Control,ctrl); 1308 } 1309 1310 // Check for useless control edge in some common special cases 1311 if (in(MemNode::Control) != NULL) { 1312 intptr_t ignore = 0; 1313 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1314 if (base != NULL 1315 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1316 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw 1317 && all_controls_dominate(base, phase->C->start())) { 1318 // A method-invariant, non-null address (constant or 'this' argument). 1319 set_req(MemNode::Control, NULL); 1320 } 1321 } 1322 1323 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) { 1324 Node* base = in(Address)->in(AddPNode::Base); 1325 if (base != NULL) { 1326 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1327 if (is_autobox_object(atp)) { 1328 Node* result = eliminate_autobox(phase); 1329 if (result != NULL) return result; 1330 } 1331 } 1332 } 1333 1334 Node* mem = in(MemNode::Memory); 1335 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1336 1337 if (addr_t != NULL) { 1338 // try to optimize our memory input 1339 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1340 if (opt_mem != mem) { 1341 set_req(MemNode::Memory, opt_mem); 1342 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1343 return this; 1344 } 1345 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1346 if (can_reshape && opt_mem->is_Phi() && 1347 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1348 // Split instance field load through Phi. 1349 Node* result = split_through_phi(phase); 1350 if (result != NULL) return result; 1351 } 1352 } 1353 1354 // Check for prior store with a different base or offset; make Load 1355 // independent. Skip through any number of them. Bail out if the stores 1356 // are in an endless dead cycle and report no progress. This is a key 1357 // transform for Reflection. However, if after skipping through the Stores 1358 // we can't then fold up against a prior store do NOT do the transform as 1359 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1360 // array memory alive twice: once for the hoisted Load and again after the 1361 // bypassed Store. This situation only works if EVERYBODY who does 1362 // anti-dependence work knows how to bypass. I.e. we need all 1363 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1364 // the alias index stuff. So instead, peek through Stores and IFF we can 1365 // fold up, do so. 1366 Node* prev_mem = find_previous_store(phase); 1367 // Steps (a), (b): Walk past independent stores to find an exact match. 1368 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1369 // (c) See if we can fold up on the spot, but don't fold up here. 1370 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1371 // just return a prior value, which is done by Identity calls. 1372 if (can_see_stored_value(prev_mem, phase)) { 1373 // Make ready for step (d): 1374 set_req(MemNode::Memory, prev_mem); 1375 return this; 1376 } 1377 } 1378 1379 return NULL; // No further progress 1380 } 1381 1382 // Helper to recognize certain Klass fields which are invariant across 1383 // some group of array types (e.g., int[] or all T[] where T < Object). 1384 const Type* 1385 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1386 ciKlass* klass) const { 1387 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1388 // The field is Klass::_modifier_flags. Return its (constant) value. 1389 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1390 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1391 return TypeInt::make(klass->modifier_flags()); 1392 } 1393 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1394 // The field is Klass::_access_flags. Return its (constant) value. 1395 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1396 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1397 return TypeInt::make(klass->access_flags()); 1398 } 1399 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1400 // The field is Klass::_layout_helper. Return its constant value if known. 1401 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1402 return TypeInt::make(klass->layout_helper()); 1403 } 1404 1405 // No match. 1406 return NULL; 1407 } 1408 1409 //------------------------------Value----------------------------------------- 1410 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1411 // Either input is TOP ==> the result is TOP 1412 Node* mem = in(MemNode::Memory); 1413 const Type *t1 = phase->type(mem); 1414 if (t1 == Type::TOP) return Type::TOP; 1415 Node* adr = in(MemNode::Address); 1416 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1417 if (tp == NULL || tp->empty()) return Type::TOP; 1418 int off = tp->offset(); 1419 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1420 1421 // Try to guess loaded type from pointer type 1422 if (tp->base() == Type::AryPtr) { 1423 const Type *t = tp->is_aryptr()->elem(); 1424 // Don't do this for integer types. There is only potential profit if 1425 // the element type t is lower than _type; that is, for int types, if _type is 1426 // more restrictive than t. This only happens here if one is short and the other 1427 // char (both 16 bits), and in those cases we've made an intentional decision 1428 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1429 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1430 // 1431 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1432 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1433 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1434 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1435 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1436 // In fact, that could have been the original type of p1, and p1 could have 1437 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1438 // expression (LShiftL quux 3) independently optimized to the constant 8. 1439 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1440 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1441 // t might actually be lower than _type, if _type is a unique 1442 // concrete subclass of abstract class t. 1443 // Make sure the reference is not into the header, by comparing 1444 // the offset against the offset of the start of the array's data. 1445 // Different array types begin at slightly different offsets (12 vs. 16). 1446 // We choose T_BYTE as an example base type that is least restrictive 1447 // as to alignment, which will therefore produce the smallest 1448 // possible base offset. 1449 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1450 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1451 const Type* jt = t->join(_type); 1452 // In any case, do not allow the join, per se, to empty out the type. 1453 if (jt->empty() && !t->empty()) { 1454 // This can happen if a interface-typed array narrows to a class type. 1455 jt = _type; 1456 } 1457 1458 if (EliminateAutoBox) { 1459 // The pointers in the autobox arrays are always non-null 1460 Node* base = in(Address)->in(AddPNode::Base); 1461 if (base != NULL) { 1462 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1463 if (is_autobox_cache(atp)) { 1464 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1465 } 1466 } 1467 } 1468 return jt; 1469 } 1470 } 1471 } else if (tp->base() == Type::InstPtr) { 1472 assert( off != Type::OffsetBot || 1473 // arrays can be cast to Objects 1474 tp->is_oopptr()->klass()->is_java_lang_Object() || 1475 // unsafe field access may not have a constant offset 1476 phase->C->has_unsafe_access(), 1477 "Field accesses must be precise" ); 1478 // For oop loads, we expect the _type to be precise 1479 } else if (tp->base() == Type::KlassPtr) { 1480 assert( off != Type::OffsetBot || 1481 // arrays can be cast to Objects 1482 tp->is_klassptr()->klass()->is_java_lang_Object() || 1483 // also allow array-loading from the primary supertype 1484 // array during subtype checks 1485 Opcode() == Op_LoadKlass, 1486 "Field accesses must be precise" ); 1487 // For klass/static loads, we expect the _type to be precise 1488 } 1489 1490 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1491 if (tkls != NULL && !StressReflectiveCode) { 1492 ciKlass* klass = tkls->klass(); 1493 if (klass->is_loaded() && tkls->klass_is_exact()) { 1494 // We are loading a field from a Klass metaobject whose identity 1495 // is known at compile time (the type is "exact" or "precise"). 1496 // Check for fields we know are maintained as constants by the VM. 1497 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1498 // The field is Klass::_super_check_offset. Return its (constant) value. 1499 // (Folds up type checking code.) 1500 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1501 return TypeInt::make(klass->super_check_offset()); 1502 } 1503 // Compute index into primary_supers array 1504 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1505 // Check for overflowing; use unsigned compare to handle the negative case. 1506 if( depth < ciKlass::primary_super_limit() ) { 1507 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1508 // (Folds up type checking code.) 1509 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1510 ciKlass *ss = klass->super_of_depth(depth); 1511 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1512 } 1513 const Type* aift = load_array_final_field(tkls, klass); 1514 if (aift != NULL) return aift; 1515 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1516 && klass->is_array_klass()) { 1517 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1518 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1519 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1520 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1521 } 1522 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1523 // The field is Klass::_java_mirror. Return its (constant) value. 1524 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1525 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1526 return TypeInstPtr::make(klass->java_mirror()); 1527 } 1528 } 1529 1530 // We can still check if we are loading from the primary_supers array at a 1531 // shallow enough depth. Even though the klass is not exact, entries less 1532 // than or equal to its super depth are correct. 1533 if (klass->is_loaded() ) { 1534 ciType *inner = klass->klass(); 1535 while( inner->is_obj_array_klass() ) 1536 inner = inner->as_obj_array_klass()->base_element_type(); 1537 if( inner->is_instance_klass() && 1538 !inner->as_instance_klass()->flags().is_interface() ) { 1539 // Compute index into primary_supers array 1540 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1541 // Check for overflowing; use unsigned compare to handle the negative case. 1542 if( depth < ciKlass::primary_super_limit() && 1543 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1544 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1545 // (Folds up type checking code.) 1546 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1547 ciKlass *ss = klass->super_of_depth(depth); 1548 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1549 } 1550 } 1551 } 1552 1553 // If the type is enough to determine that the thing is not an array, 1554 // we can give the layout_helper a positive interval type. 1555 // This will help short-circuit some reflective code. 1556 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1557 && !klass->is_array_klass() // not directly typed as an array 1558 && !klass->is_interface() // specifically not Serializable & Cloneable 1559 && !klass->is_java_lang_Object() // not the supertype of all T[] 1560 ) { 1561 // Note: When interfaces are reliable, we can narrow the interface 1562 // test to (klass != Serializable && klass != Cloneable). 1563 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1564 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1565 // The key property of this type is that it folds up tests 1566 // for array-ness, since it proves that the layout_helper is positive. 1567 // Thus, a generic value like the basic object layout helper works fine. 1568 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1569 } 1570 } 1571 1572 // If we are loading from a freshly-allocated object, produce a zero, 1573 // if the load is provably beyond the header of the object. 1574 // (Also allow a variable load from a fresh array to produce zero.) 1575 if (ReduceFieldZeroing) { 1576 Node* value = can_see_stored_value(mem,phase); 1577 if (value != NULL && value->is_Con()) 1578 return value->bottom_type(); 1579 } 1580 1581 const TypeOopPtr *tinst = tp->isa_oopptr(); 1582 if (tinst != NULL && tinst->is_known_instance_field()) { 1583 // If we have an instance type and our memory input is the 1584 // programs's initial memory state, there is no matching store, 1585 // so just return a zero of the appropriate type 1586 Node *mem = in(MemNode::Memory); 1587 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1588 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1589 return Type::get_zero_type(_type->basic_type()); 1590 } 1591 } 1592 return _type; 1593 } 1594 1595 //------------------------------match_edge------------------------------------- 1596 // Do we Match on this edge index or not? Match only the address. 1597 uint LoadNode::match_edge(uint idx) const { 1598 return idx == MemNode::Address; 1599 } 1600 1601 //--------------------------LoadBNode::Ideal-------------------------------------- 1602 // 1603 // If the previous store is to the same address as this load, 1604 // and the value stored was larger than a byte, replace this load 1605 // with the value stored truncated to a byte. If no truncation is 1606 // needed, the replacement is done in LoadNode::Identity(). 1607 // 1608 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1609 Node* mem = in(MemNode::Memory); 1610 Node* value = can_see_stored_value(mem,phase); 1611 if( value && !phase->type(value)->higher_equal( _type ) ) { 1612 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1613 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1614 } 1615 // Identity call will handle the case where truncation is not needed. 1616 return LoadNode::Ideal(phase, can_reshape); 1617 } 1618 1619 //--------------------------LoadUBNode::Ideal------------------------------------- 1620 // 1621 // If the previous store is to the same address as this load, 1622 // and the value stored was larger than a byte, replace this load 1623 // with the value stored truncated to a byte. If no truncation is 1624 // needed, the replacement is done in LoadNode::Identity(). 1625 // 1626 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1627 Node* mem = in(MemNode::Memory); 1628 Node* value = can_see_stored_value(mem, phase); 1629 if (value && !phase->type(value)->higher_equal(_type)) 1630 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); 1631 // Identity call will handle the case where truncation is not needed. 1632 return LoadNode::Ideal(phase, can_reshape); 1633 } 1634 1635 //--------------------------LoadUSNode::Ideal------------------------------------- 1636 // 1637 // If the previous store is to the same address as this load, 1638 // and the value stored was larger than a char, replace this load 1639 // with the value stored truncated to a char. If no truncation is 1640 // needed, the replacement is done in LoadNode::Identity(). 1641 // 1642 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1643 Node* mem = in(MemNode::Memory); 1644 Node* value = can_see_stored_value(mem,phase); 1645 if( value && !phase->type(value)->higher_equal( _type ) ) 1646 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1647 // Identity call will handle the case where truncation is not needed. 1648 return LoadNode::Ideal(phase, can_reshape); 1649 } 1650 1651 //--------------------------LoadSNode::Ideal-------------------------------------- 1652 // 1653 // If the previous store is to the same address as this load, 1654 // and the value stored was larger than a short, replace this load 1655 // with the value stored truncated to a short. If no truncation is 1656 // needed, the replacement is done in LoadNode::Identity(). 1657 // 1658 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1659 Node* mem = in(MemNode::Memory); 1660 Node* value = can_see_stored_value(mem,phase); 1661 if( value && !phase->type(value)->higher_equal( _type ) ) { 1662 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1663 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1664 } 1665 // Identity call will handle the case where truncation is not needed. 1666 return LoadNode::Ideal(phase, can_reshape); 1667 } 1668 1669 //============================================================================= 1670 //----------------------------LoadKlassNode::make------------------------------ 1671 // Polymorphic factory method: 1672 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1673 Compile* C = gvn.C; 1674 Node *ctl = NULL; 1675 // sanity check the alias category against the created node type 1676 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1677 assert(adr_type != NULL, "expecting TypeOopPtr"); 1678 #ifdef _LP64 1679 if (adr_type->is_ptr_to_narrowoop()) { 1680 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1681 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1682 } 1683 #endif 1684 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1685 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1686 } 1687 1688 //------------------------------Value------------------------------------------ 1689 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1690 return klass_value_common(phase); 1691 } 1692 1693 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1694 // Either input is TOP ==> the result is TOP 1695 const Type *t1 = phase->type( in(MemNode::Memory) ); 1696 if (t1 == Type::TOP) return Type::TOP; 1697 Node *adr = in(MemNode::Address); 1698 const Type *t2 = phase->type( adr ); 1699 if (t2 == Type::TOP) return Type::TOP; 1700 const TypePtr *tp = t2->is_ptr(); 1701 if (TypePtr::above_centerline(tp->ptr()) || 1702 tp->ptr() == TypePtr::Null) return Type::TOP; 1703 1704 // Return a more precise klass, if possible 1705 const TypeInstPtr *tinst = tp->isa_instptr(); 1706 if (tinst != NULL) { 1707 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1708 int offset = tinst->offset(); 1709 if (ik == phase->C->env()->Class_klass() 1710 && (offset == java_lang_Class::klass_offset_in_bytes() || 1711 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1712 // We are loading a special hidden field from a Class mirror object, 1713 // the field which points to the VM's Klass metaobject. 1714 ciType* t = tinst->java_mirror_type(); 1715 // java_mirror_type returns non-null for compile-time Class constants. 1716 if (t != NULL) { 1717 // constant oop => constant klass 1718 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1719 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1720 } 1721 if (!t->is_klass()) { 1722 // a primitive Class (e.g., int.class) has NULL for a klass field 1723 return TypePtr::NULL_PTR; 1724 } 1725 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1726 return TypeKlassPtr::make(t->as_klass()); 1727 } 1728 // non-constant mirror, so we can't tell what's going on 1729 } 1730 if( !ik->is_loaded() ) 1731 return _type; // Bail out if not loaded 1732 if (offset == oopDesc::klass_offset_in_bytes()) { 1733 if (tinst->klass_is_exact()) { 1734 return TypeKlassPtr::make(ik); 1735 } 1736 // See if we can become precise: no subklasses and no interface 1737 // (Note: We need to support verified interfaces.) 1738 if (!ik->is_interface() && !ik->has_subklass()) { 1739 //assert(!UseExactTypes, "this code should be useless with exact types"); 1740 // Add a dependence; if any subclass added we need to recompile 1741 if (!ik->is_final()) { 1742 // %%% should use stronger assert_unique_concrete_subtype instead 1743 phase->C->dependencies()->assert_leaf_type(ik); 1744 } 1745 // Return precise klass 1746 return TypeKlassPtr::make(ik); 1747 } 1748 1749 // Return root of possible klass 1750 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1751 } 1752 } 1753 1754 // Check for loading klass from an array 1755 const TypeAryPtr *tary = tp->isa_aryptr(); 1756 if( tary != NULL ) { 1757 ciKlass *tary_klass = tary->klass(); 1758 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1759 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1760 if (tary->klass_is_exact()) { 1761 return TypeKlassPtr::make(tary_klass); 1762 } 1763 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1764 // If the klass is an object array, we defer the question to the 1765 // array component klass. 1766 if( ak->is_obj_array_klass() ) { 1767 assert( ak->is_loaded(), "" ); 1768 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1769 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1770 ciInstanceKlass* ik = base_k->as_instance_klass(); 1771 // See if we can become precise: no subklasses and no interface 1772 if (!ik->is_interface() && !ik->has_subklass()) { 1773 //assert(!UseExactTypes, "this code should be useless with exact types"); 1774 // Add a dependence; if any subclass added we need to recompile 1775 if (!ik->is_final()) { 1776 phase->C->dependencies()->assert_leaf_type(ik); 1777 } 1778 // Return precise array klass 1779 return TypeKlassPtr::make(ak); 1780 } 1781 } 1782 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1783 } else { // Found a type-array? 1784 //assert(!UseExactTypes, "this code should be useless with exact types"); 1785 assert( ak->is_type_array_klass(), "" ); 1786 return TypeKlassPtr::make(ak); // These are always precise 1787 } 1788 } 1789 } 1790 1791 // Check for loading klass from an array klass 1792 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1793 if (tkls != NULL && !StressReflectiveCode) { 1794 ciKlass* klass = tkls->klass(); 1795 if( !klass->is_loaded() ) 1796 return _type; // Bail out if not loaded 1797 if( klass->is_obj_array_klass() && 1798 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1799 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1800 // // Always returning precise element type is incorrect, 1801 // // e.g., element type could be object and array may contain strings 1802 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1803 1804 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1805 // according to the element type's subclassing. 1806 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1807 } 1808 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1809 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1810 ciKlass* sup = klass->as_instance_klass()->super(); 1811 // The field is Klass::_super. Return its (constant) value. 1812 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1813 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1814 } 1815 } 1816 1817 // Bailout case 1818 return LoadNode::Value(phase); 1819 } 1820 1821 //------------------------------Identity--------------------------------------- 1822 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1823 // Also feed through the klass in Allocate(...klass...)._klass. 1824 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1825 return klass_identity_common(phase); 1826 } 1827 1828 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 1829 Node* x = LoadNode::Identity(phase); 1830 if (x != this) return x; 1831 1832 // Take apart the address into an oop and and offset. 1833 // Return 'this' if we cannot. 1834 Node* adr = in(MemNode::Address); 1835 intptr_t offset = 0; 1836 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1837 if (base == NULL) return this; 1838 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1839 if (toop == NULL) return this; 1840 1841 // We can fetch the klass directly through an AllocateNode. 1842 // This works even if the klass is not constant (clone or newArray). 1843 if (offset == oopDesc::klass_offset_in_bytes()) { 1844 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1845 if (allocated_klass != NULL) { 1846 return allocated_klass; 1847 } 1848 } 1849 1850 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1851 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1852 // See inline_native_Class_query for occurrences of these patterns. 1853 // Java Example: x.getClass().isAssignableFrom(y) 1854 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1855 // 1856 // This improves reflective code, often making the Class 1857 // mirror go completely dead. (Current exception: Class 1858 // mirrors may appear in debug info, but we could clean them out by 1859 // introducing a new debug info operator for klassOop.java_mirror). 1860 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1861 && (offset == java_lang_Class::klass_offset_in_bytes() || 1862 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1863 // We are loading a special hidden field from a Class mirror, 1864 // the field which points to its Klass or arrayKlass metaobject. 1865 if (base->is_Load()) { 1866 Node* adr2 = base->in(MemNode::Address); 1867 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1868 if (tkls != NULL && !tkls->empty() 1869 && (tkls->klass()->is_instance_klass() || 1870 tkls->klass()->is_array_klass()) 1871 && adr2->is_AddP() 1872 ) { 1873 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1874 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1875 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1876 } 1877 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1878 return adr2->in(AddPNode::Base); 1879 } 1880 } 1881 } 1882 } 1883 1884 return this; 1885 } 1886 1887 1888 //------------------------------Value------------------------------------------ 1889 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 1890 const Type *t = klass_value_common(phase); 1891 if (t == Type::TOP) 1892 return t; 1893 1894 return t->make_narrowoop(); 1895 } 1896 1897 //------------------------------Identity--------------------------------------- 1898 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 1899 // Also feed through the klass in Allocate(...klass...)._klass. 1900 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 1901 Node *x = klass_identity_common(phase); 1902 1903 const Type *t = phase->type( x ); 1904 if( t == Type::TOP ) return x; 1905 if( t->isa_narrowoop()) return x; 1906 1907 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 1908 } 1909 1910 //------------------------------Value----------------------------------------- 1911 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1912 // Either input is TOP ==> the result is TOP 1913 const Type *t1 = phase->type( in(MemNode::Memory) ); 1914 if( t1 == Type::TOP ) return Type::TOP; 1915 Node *adr = in(MemNode::Address); 1916 const Type *t2 = phase->type( adr ); 1917 if( t2 == Type::TOP ) return Type::TOP; 1918 const TypePtr *tp = t2->is_ptr(); 1919 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 1920 const TypeAryPtr *tap = tp->isa_aryptr(); 1921 if( !tap ) return _type; 1922 return tap->size(); 1923 } 1924 1925 //-------------------------------Ideal--------------------------------------- 1926 // Feed through the length in AllocateArray(...length...)._length. 1927 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1928 Node* p = MemNode::Ideal_common(phase, can_reshape); 1929 if (p) return (p == NodeSentinel) ? NULL : p; 1930 1931 // Take apart the address into an oop and and offset. 1932 // Return 'this' if we cannot. 1933 Node* adr = in(MemNode::Address); 1934 intptr_t offset = 0; 1935 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1936 if (base == NULL) return NULL; 1937 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1938 if (tary == NULL) return NULL; 1939 1940 // We can fetch the length directly through an AllocateArrayNode. 1941 // This works even if the length is not constant (clone or newArray). 1942 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1943 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 1944 if (alloc != NULL) { 1945 Node* allocated_length = alloc->Ideal_length(); 1946 Node* len = alloc->make_ideal_length(tary, phase); 1947 if (allocated_length != len) { 1948 // New CastII improves on this. 1949 return len; 1950 } 1951 } 1952 } 1953 1954 return NULL; 1955 } 1956 1957 //------------------------------Identity--------------------------------------- 1958 // Feed through the length in AllocateArray(...length...)._length. 1959 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 1960 Node* x = LoadINode::Identity(phase); 1961 if (x != this) return x; 1962 1963 // Take apart the address into an oop and and offset. 1964 // Return 'this' if we cannot. 1965 Node* adr = in(MemNode::Address); 1966 intptr_t offset = 0; 1967 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1968 if (base == NULL) return this; 1969 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1970 if (tary == NULL) return this; 1971 1972 // We can fetch the length directly through an AllocateArrayNode. 1973 // This works even if the length is not constant (clone or newArray). 1974 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1975 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 1976 if (alloc != NULL) { 1977 Node* allocated_length = alloc->Ideal_length(); 1978 // Do not allow make_ideal_length to allocate a CastII node. 1979 Node* len = alloc->make_ideal_length(tary, phase, false); 1980 if (allocated_length == len) { 1981 // Return allocated_length only if it would not be improved by a CastII. 1982 return allocated_length; 1983 } 1984 } 1985 } 1986 1987 return this; 1988 1989 } 1990 1991 //============================================================================= 1992 //---------------------------StoreNode::make----------------------------------- 1993 // Polymorphic factory method: 1994 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 1995 Compile* C = gvn.C; 1996 1997 switch (bt) { 1998 case T_BOOLEAN: 1999 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 2000 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 2001 case T_CHAR: 2002 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 2003 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 2004 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 2005 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 2006 case T_ADDRESS: 2007 case T_OBJECT: 2008 #ifdef _LP64 2009 if (adr->bottom_type()->is_ptr_to_narrowoop() || 2010 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 2011 adr->bottom_type()->isa_rawptr())) { 2012 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2013 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 2014 } else 2015 #endif 2016 { 2017 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 2018 } 2019 } 2020 ShouldNotReachHere(); 2021 return (StoreNode*)NULL; 2022 } 2023 2024 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2025 bool require_atomic = true; 2026 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2027 } 2028 2029 2030 //--------------------------bottom_type---------------------------------------- 2031 const Type *StoreNode::bottom_type() const { 2032 return Type::MEMORY; 2033 } 2034 2035 //------------------------------hash------------------------------------------- 2036 uint StoreNode::hash() const { 2037 // unroll addition of interesting fields 2038 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2039 2040 // Since they are not commoned, do not hash them: 2041 return NO_HASH; 2042 } 2043 2044 //------------------------------Ideal------------------------------------------ 2045 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2046 // When a store immediately follows a relevant allocation/initialization, 2047 // try to capture it into the initialization, or hoist it above. 2048 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2049 Node* p = MemNode::Ideal_common(phase, can_reshape); 2050 if (p) return (p == NodeSentinel) ? NULL : p; 2051 2052 Node* mem = in(MemNode::Memory); 2053 Node* address = in(MemNode::Address); 2054 2055 // Back-to-back stores to same address? Fold em up. 2056 // Generally unsafe if I have intervening uses... 2057 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 2058 // Looking at a dead closed cycle of memory? 2059 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2060 2061 assert(Opcode() == mem->Opcode() || 2062 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2063 "no mismatched stores, except on raw memory"); 2064 2065 if (mem->outcnt() == 1 && // check for intervening uses 2066 mem->as_Store()->memory_size() <= this->memory_size()) { 2067 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2068 // For example, 'mem' might be the final state at a conditional return. 2069 // Or, 'mem' might be used by some node which is live at the same time 2070 // 'this' is live, which might be unschedulable. So, require exactly 2071 // ONE user, the 'this' store, until such time as we clone 'mem' for 2072 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2073 if (can_reshape) { // (%%% is this an anachronism?) 2074 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2075 phase->is_IterGVN()); 2076 } else { 2077 // It's OK to do this in the parser, since DU info is always accurate, 2078 // and the parser always refers to nodes via SafePointNode maps. 2079 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2080 } 2081 return this; 2082 } 2083 } 2084 2085 // Capture an unaliased, unconditional, simple store into an initializer. 2086 // Or, if it is independent of the allocation, hoist it above the allocation. 2087 if (ReduceFieldZeroing && /*can_reshape &&*/ 2088 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2089 InitializeNode* init = mem->in(0)->as_Initialize(); 2090 intptr_t offset = init->can_capture_store(this, phase); 2091 if (offset > 0) { 2092 Node* moved = init->capture_store(this, offset, phase); 2093 // If the InitializeNode captured me, it made a raw copy of me, 2094 // and I need to disappear. 2095 if (moved != NULL) { 2096 // %%% hack to ensure that Ideal returns a new node: 2097 mem = MergeMemNode::make(phase->C, mem); 2098 return mem; // fold me away 2099 } 2100 } 2101 } 2102 2103 return NULL; // No further progress 2104 } 2105 2106 //------------------------------Value----------------------------------------- 2107 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2108 // Either input is TOP ==> the result is TOP 2109 const Type *t1 = phase->type( in(MemNode::Memory) ); 2110 if( t1 == Type::TOP ) return Type::TOP; 2111 const Type *t2 = phase->type( in(MemNode::Address) ); 2112 if( t2 == Type::TOP ) return Type::TOP; 2113 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2114 if( t3 == Type::TOP ) return Type::TOP; 2115 return Type::MEMORY; 2116 } 2117 2118 //------------------------------Identity--------------------------------------- 2119 // Remove redundant stores: 2120 // Store(m, p, Load(m, p)) changes to m. 2121 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2122 Node *StoreNode::Identity( PhaseTransform *phase ) { 2123 Node* mem = in(MemNode::Memory); 2124 Node* adr = in(MemNode::Address); 2125 Node* val = in(MemNode::ValueIn); 2126 2127 // Load then Store? Then the Store is useless 2128 if (val->is_Load() && 2129 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 2130 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 2131 val->as_Load()->store_Opcode() == Opcode()) { 2132 return mem; 2133 } 2134 2135 // Two stores in a row of the same value? 2136 if (mem->is_Store() && 2137 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 2138 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 2139 mem->Opcode() == Opcode()) { 2140 return mem; 2141 } 2142 2143 // Store of zero anywhere into a freshly-allocated object? 2144 // Then the store is useless. 2145 // (It must already have been captured by the InitializeNode.) 2146 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2147 // a newly allocated object is already all-zeroes everywhere 2148 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2149 return mem; 2150 } 2151 2152 // the store may also apply to zero-bits in an earlier object 2153 Node* prev_mem = find_previous_store(phase); 2154 // Steps (a), (b): Walk past independent stores to find an exact match. 2155 if (prev_mem != NULL) { 2156 Node* prev_val = can_see_stored_value(prev_mem, phase); 2157 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2158 // prev_val and val might differ by a cast; it would be good 2159 // to keep the more informative of the two. 2160 return mem; 2161 } 2162 } 2163 } 2164 2165 return this; 2166 } 2167 2168 //------------------------------match_edge------------------------------------- 2169 // Do we Match on this edge index or not? Match only memory & value 2170 uint StoreNode::match_edge(uint idx) const { 2171 return idx == MemNode::Address || idx == MemNode::ValueIn; 2172 } 2173 2174 //------------------------------cmp-------------------------------------------- 2175 // Do not common stores up together. They generally have to be split 2176 // back up anyways, so do not bother. 2177 uint StoreNode::cmp( const Node &n ) const { 2178 return (&n == this); // Always fail except on self 2179 } 2180 2181 //------------------------------Ideal_masked_input----------------------------- 2182 // Check for a useless mask before a partial-word store 2183 // (StoreB ... (AndI valIn conIa) ) 2184 // If (conIa & mask == mask) this simplifies to 2185 // (StoreB ... (valIn) ) 2186 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2187 Node *val = in(MemNode::ValueIn); 2188 if( val->Opcode() == Op_AndI ) { 2189 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2190 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2191 set_req(MemNode::ValueIn, val->in(1)); 2192 return this; 2193 } 2194 } 2195 return NULL; 2196 } 2197 2198 2199 //------------------------------Ideal_sign_extended_input---------------------- 2200 // Check for useless sign-extension before a partial-word store 2201 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2202 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2203 // (StoreB ... (valIn) ) 2204 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2205 Node *val = in(MemNode::ValueIn); 2206 if( val->Opcode() == Op_RShiftI ) { 2207 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2208 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2209 Node *shl = val->in(1); 2210 if( shl->Opcode() == Op_LShiftI ) { 2211 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2212 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2213 set_req(MemNode::ValueIn, shl->in(1)); 2214 return this; 2215 } 2216 } 2217 } 2218 } 2219 return NULL; 2220 } 2221 2222 //------------------------------value_never_loaded----------------------------------- 2223 // Determine whether there are any possible loads of the value stored. 2224 // For simplicity, we actually check if there are any loads from the 2225 // address stored to, not just for loads of the value stored by this node. 2226 // 2227 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2228 Node *adr = in(Address); 2229 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2230 if (adr_oop == NULL) 2231 return false; 2232 if (!adr_oop->is_known_instance_field()) 2233 return false; // if not a distinct instance, there may be aliases of the address 2234 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2235 Node *use = adr->fast_out(i); 2236 int opc = use->Opcode(); 2237 if (use->is_Load() || use->is_LoadStore()) { 2238 return false; 2239 } 2240 } 2241 return true; 2242 } 2243 2244 //============================================================================= 2245 //------------------------------Ideal------------------------------------------ 2246 // If the store is from an AND mask that leaves the low bits untouched, then 2247 // we can skip the AND operation. If the store is from a sign-extension 2248 // (a left shift, then right shift) we can skip both. 2249 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2250 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2251 if( progress != NULL ) return progress; 2252 2253 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2254 if( progress != NULL ) return progress; 2255 2256 // Finally check the default case 2257 return StoreNode::Ideal(phase, can_reshape); 2258 } 2259 2260 //============================================================================= 2261 //------------------------------Ideal------------------------------------------ 2262 // If the store is from an AND mask that leaves the low bits untouched, then 2263 // we can skip the AND operation 2264 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2265 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2266 if( progress != NULL ) return progress; 2267 2268 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2269 if( progress != NULL ) return progress; 2270 2271 // Finally check the default case 2272 return StoreNode::Ideal(phase, can_reshape); 2273 } 2274 2275 //============================================================================= 2276 //------------------------------Identity--------------------------------------- 2277 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2278 // No need to card mark when storing a null ptr 2279 Node* my_store = in(MemNode::OopStore); 2280 if (my_store->is_Store()) { 2281 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2282 if( t1 == TypePtr::NULL_PTR ) { 2283 return in(MemNode::Memory); 2284 } 2285 } 2286 return this; 2287 } 2288 2289 //------------------------------Value----------------------------------------- 2290 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2291 // Either input is TOP ==> the result is TOP 2292 const Type *t = phase->type( in(MemNode::Memory) ); 2293 if( t == Type::TOP ) return Type::TOP; 2294 t = phase->type( in(MemNode::Address) ); 2295 if( t == Type::TOP ) return Type::TOP; 2296 t = phase->type( in(MemNode::ValueIn) ); 2297 if( t == Type::TOP ) return Type::TOP; 2298 // If extra input is TOP ==> the result is TOP 2299 t = phase->type( in(MemNode::OopStore) ); 2300 if( t == Type::TOP ) return Type::TOP; 2301 2302 return StoreNode::Value( phase ); 2303 } 2304 2305 2306 //============================================================================= 2307 //----------------------------------SCMemProjNode------------------------------ 2308 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2309 { 2310 return bottom_type(); 2311 } 2312 2313 //============================================================================= 2314 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2315 init_req(MemNode::Control, c ); 2316 init_req(MemNode::Memory , mem); 2317 init_req(MemNode::Address, adr); 2318 init_req(MemNode::ValueIn, val); 2319 init_req( ExpectedIn, ex ); 2320 init_class_id(Class_LoadStore); 2321 2322 } 2323 2324 //============================================================================= 2325 //-------------------------------adr_type-------------------------------------- 2326 // Do we Match on this edge index or not? Do not match memory 2327 const TypePtr* ClearArrayNode::adr_type() const { 2328 Node *adr = in(3); 2329 return MemNode::calculate_adr_type(adr->bottom_type()); 2330 } 2331 2332 //------------------------------match_edge------------------------------------- 2333 // Do we Match on this edge index or not? Do not match memory 2334 uint ClearArrayNode::match_edge(uint idx) const { 2335 return idx > 1; 2336 } 2337 2338 //------------------------------Identity--------------------------------------- 2339 // Clearing a zero length array does nothing 2340 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2341 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2342 } 2343 2344 //------------------------------Idealize--------------------------------------- 2345 // Clearing a short array is faster with stores 2346 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2347 const int unit = BytesPerLong; 2348 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2349 if (!t) return NULL; 2350 if (!t->is_con()) return NULL; 2351 intptr_t raw_count = t->get_con(); 2352 intptr_t size = raw_count; 2353 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2354 // Clearing nothing uses the Identity call. 2355 // Negative clears are possible on dead ClearArrays 2356 // (see jck test stmt114.stmt11402.val). 2357 if (size <= 0 || size % unit != 0) return NULL; 2358 intptr_t count = size / unit; 2359 // Length too long; use fast hardware clear 2360 if (size > Matcher::init_array_short_size) return NULL; 2361 Node *mem = in(1); 2362 if( phase->type(mem)==Type::TOP ) return NULL; 2363 Node *adr = in(3); 2364 const Type* at = phase->type(adr); 2365 if( at==Type::TOP ) return NULL; 2366 const TypePtr* atp = at->isa_ptr(); 2367 // adjust atp to be the correct array element address type 2368 if (atp == NULL) atp = TypePtr::BOTTOM; 2369 else atp = atp->add_offset(Type::OffsetBot); 2370 // Get base for derived pointer purposes 2371 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2372 Node *base = adr->in(1); 2373 2374 Node *zero = phase->makecon(TypeLong::ZERO); 2375 Node *off = phase->MakeConX(BytesPerLong); 2376 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2377 count--; 2378 while( count-- ) { 2379 mem = phase->transform(mem); 2380 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2381 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2382 } 2383 return mem; 2384 } 2385 2386 //----------------------------clear_memory------------------------------------- 2387 // Generate code to initialize object storage to zero. 2388 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2389 intptr_t start_offset, 2390 Node* end_offset, 2391 PhaseGVN* phase) { 2392 Compile* C = phase->C; 2393 intptr_t offset = start_offset; 2394 2395 int unit = BytesPerLong; 2396 if ((offset % unit) != 0) { 2397 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2398 adr = phase->transform(adr); 2399 const TypePtr* atp = TypeRawPtr::BOTTOM; 2400 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2401 mem = phase->transform(mem); 2402 offset += BytesPerInt; 2403 } 2404 assert((offset % unit) == 0, ""); 2405 2406 // Initialize the remaining stuff, if any, with a ClearArray. 2407 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2408 } 2409 2410 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2411 Node* start_offset, 2412 Node* end_offset, 2413 PhaseGVN* phase) { 2414 if (start_offset == end_offset) { 2415 // nothing to do 2416 return mem; 2417 } 2418 2419 Compile* C = phase->C; 2420 int unit = BytesPerLong; 2421 Node* zbase = start_offset; 2422 Node* zend = end_offset; 2423 2424 // Scale to the unit required by the CPU: 2425 if (!Matcher::init_array_count_is_in_bytes) { 2426 Node* shift = phase->intcon(exact_log2(unit)); 2427 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2428 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2429 } 2430 2431 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2432 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2433 2434 // Bulk clear double-words 2435 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2436 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2437 return phase->transform(mem); 2438 } 2439 2440 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2441 intptr_t start_offset, 2442 intptr_t end_offset, 2443 PhaseGVN* phase) { 2444 if (start_offset == end_offset) { 2445 // nothing to do 2446 return mem; 2447 } 2448 2449 Compile* C = phase->C; 2450 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2451 intptr_t done_offset = end_offset; 2452 if ((done_offset % BytesPerLong) != 0) { 2453 done_offset -= BytesPerInt; 2454 } 2455 if (done_offset > start_offset) { 2456 mem = clear_memory(ctl, mem, dest, 2457 start_offset, phase->MakeConX(done_offset), phase); 2458 } 2459 if (done_offset < end_offset) { // emit the final 32-bit store 2460 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2461 adr = phase->transform(adr); 2462 const TypePtr* atp = TypeRawPtr::BOTTOM; 2463 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2464 mem = phase->transform(mem); 2465 done_offset += BytesPerInt; 2466 } 2467 assert(done_offset == end_offset, ""); 2468 return mem; 2469 } 2470 2471 //============================================================================= 2472 // Do we match on this edge? No memory edges 2473 uint StrCompNode::match_edge(uint idx) const { 2474 return idx == 5 || idx == 6; 2475 } 2476 2477 //------------------------------Ideal------------------------------------------ 2478 // Return a node which is more "ideal" than the current node. Strip out 2479 // control copies 2480 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2481 return remove_dead_region(phase, can_reshape) ? this : NULL; 2482 } 2483 2484 // Do we match on this edge? No memory edges 2485 uint StrEqualsNode::match_edge(uint idx) const { 2486 return idx == 5 || idx == 6; 2487 } 2488 2489 //------------------------------Ideal------------------------------------------ 2490 // Return a node which is more "ideal" than the current node. Strip out 2491 // control copies 2492 Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2493 return remove_dead_region(phase, can_reshape) ? this : NULL; 2494 } 2495 2496 //============================================================================= 2497 // Do we match on this edge? No memory edges 2498 uint StrIndexOfNode::match_edge(uint idx) const { 2499 return idx == 5 || idx == 6; 2500 } 2501 2502 //------------------------------Ideal------------------------------------------ 2503 // Return a node which is more "ideal" than the current node. Strip out 2504 // control copies 2505 Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2506 return remove_dead_region(phase, can_reshape) ? this : NULL; 2507 } 2508 2509 //------------------------------Ideal------------------------------------------ 2510 // Return a node which is more "ideal" than the current node. Strip out 2511 // control copies 2512 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2513 return remove_dead_region(phase, can_reshape) ? this : NULL; 2514 } 2515 2516 //============================================================================= 2517 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2518 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2519 _adr_type(C->get_adr_type(alias_idx)) 2520 { 2521 init_class_id(Class_MemBar); 2522 Node* top = C->top(); 2523 init_req(TypeFunc::I_O,top); 2524 init_req(TypeFunc::FramePtr,top); 2525 init_req(TypeFunc::ReturnAdr,top); 2526 if (precedent != NULL) 2527 init_req(TypeFunc::Parms, precedent); 2528 } 2529 2530 //------------------------------cmp-------------------------------------------- 2531 uint MemBarNode::hash() const { return NO_HASH; } 2532 uint MemBarNode::cmp( const Node &n ) const { 2533 return (&n == this); // Always fail except on self 2534 } 2535 2536 //------------------------------make------------------------------------------- 2537 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2538 int len = Precedent + (pn == NULL? 0: 1); 2539 switch (opcode) { 2540 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2541 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2542 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2543 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2544 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2545 default: ShouldNotReachHere(); return NULL; 2546 } 2547 } 2548 2549 //------------------------------Ideal------------------------------------------ 2550 // Return a node which is more "ideal" than the current node. Strip out 2551 // control copies 2552 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2553 return remove_dead_region(phase, can_reshape) ? this : NULL; 2554 } 2555 2556 //------------------------------Value------------------------------------------ 2557 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2558 if( !in(0) ) return Type::TOP; 2559 if( phase->type(in(0)) == Type::TOP ) 2560 return Type::TOP; 2561 return TypeTuple::MEMBAR; 2562 } 2563 2564 //------------------------------match------------------------------------------ 2565 // Construct projections for memory. 2566 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2567 switch (proj->_con) { 2568 case TypeFunc::Control: 2569 case TypeFunc::Memory: 2570 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2571 } 2572 ShouldNotReachHere(); 2573 return NULL; 2574 } 2575 2576 //===========================InitializeNode==================================== 2577 // SUMMARY: 2578 // This node acts as a memory barrier on raw memory, after some raw stores. 2579 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2580 // The Initialize can 'capture' suitably constrained stores as raw inits. 2581 // It can coalesce related raw stores into larger units (called 'tiles'). 2582 // It can avoid zeroing new storage for memory units which have raw inits. 2583 // At macro-expansion, it is marked 'complete', and does not optimize further. 2584 // 2585 // EXAMPLE: 2586 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2587 // ctl = incoming control; mem* = incoming memory 2588 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2589 // First allocate uninitialized memory and fill in the header: 2590 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2591 // ctl := alloc.Control; mem* := alloc.Memory* 2592 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2593 // Then initialize to zero the non-header parts of the raw memory block: 2594 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2595 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2596 // After the initialize node executes, the object is ready for service: 2597 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2598 // Suppose its body is immediately initialized as {1,2}: 2599 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2600 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2601 // mem.SLICE(#short[*]) := store2 2602 // 2603 // DETAILS: 2604 // An InitializeNode collects and isolates object initialization after 2605 // an AllocateNode and before the next possible safepoint. As a 2606 // memory barrier (MemBarNode), it keeps critical stores from drifting 2607 // down past any safepoint or any publication of the allocation. 2608 // Before this barrier, a newly-allocated object may have uninitialized bits. 2609 // After this barrier, it may be treated as a real oop, and GC is allowed. 2610 // 2611 // The semantics of the InitializeNode include an implicit zeroing of 2612 // the new object from object header to the end of the object. 2613 // (The object header and end are determined by the AllocateNode.) 2614 // 2615 // Certain stores may be added as direct inputs to the InitializeNode. 2616 // These stores must update raw memory, and they must be to addresses 2617 // derived from the raw address produced by AllocateNode, and with 2618 // a constant offset. They must be ordered by increasing offset. 2619 // The first one is at in(RawStores), the last at in(req()-1). 2620 // Unlike most memory operations, they are not linked in a chain, 2621 // but are displayed in parallel as users of the rawmem output of 2622 // the allocation. 2623 // 2624 // (See comments in InitializeNode::capture_store, which continue 2625 // the example given above.) 2626 // 2627 // When the associated Allocate is macro-expanded, the InitializeNode 2628 // may be rewritten to optimize collected stores. A ClearArrayNode 2629 // may also be created at that point to represent any required zeroing. 2630 // The InitializeNode is then marked 'complete', prohibiting further 2631 // capturing of nearby memory operations. 2632 // 2633 // During macro-expansion, all captured initializations which store 2634 // constant values of 32 bits or smaller are coalesced (if advantageous) 2635 // into larger 'tiles' 32 or 64 bits. This allows an object to be 2636 // initialized in fewer memory operations. Memory words which are 2637 // covered by neither tiles nor non-constant stores are pre-zeroed 2638 // by explicit stores of zero. (The code shape happens to do all 2639 // zeroing first, then all other stores, with both sequences occurring 2640 // in order of ascending offsets.) 2641 // 2642 // Alternatively, code may be inserted between an AllocateNode and its 2643 // InitializeNode, to perform arbitrary initialization of the new object. 2644 // E.g., the object copying intrinsics insert complex data transfers here. 2645 // The initialization must then be marked as 'complete' disable the 2646 // built-in zeroing semantics and the collection of initializing stores. 2647 // 2648 // While an InitializeNode is incomplete, reads from the memory state 2649 // produced by it are optimizable if they match the control edge and 2650 // new oop address associated with the allocation/initialization. 2651 // They return a stored value (if the offset matches) or else zero. 2652 // A write to the memory state, if it matches control and address, 2653 // and if it is to a constant offset, may be 'captured' by the 2654 // InitializeNode. It is cloned as a raw memory operation and rewired 2655 // inside the initialization, to the raw oop produced by the allocation. 2656 // Operations on addresses which are provably distinct (e.g., to 2657 // other AllocateNodes) are allowed to bypass the initialization. 2658 // 2659 // The effect of all this is to consolidate object initialization 2660 // (both arrays and non-arrays, both piecewise and bulk) into a 2661 // single location, where it can be optimized as a unit. 2662 // 2663 // Only stores with an offset less than TrackedInitializationLimit words 2664 // will be considered for capture by an InitializeNode. This puts a 2665 // reasonable limit on the complexity of optimized initializations. 2666 2667 //---------------------------InitializeNode------------------------------------ 2668 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2669 : _is_complete(false), 2670 MemBarNode(C, adr_type, rawoop) 2671 { 2672 init_class_id(Class_Initialize); 2673 2674 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2675 assert(in(RawAddress) == rawoop, "proper init"); 2676 // Note: allocation() can be NULL, for secondary initialization barriers 2677 } 2678 2679 // Since this node is not matched, it will be processed by the 2680 // register allocator. Declare that there are no constraints 2681 // on the allocation of the RawAddress edge. 2682 const RegMask &InitializeNode::in_RegMask(uint idx) const { 2683 // This edge should be set to top, by the set_complete. But be conservative. 2684 if (idx == InitializeNode::RawAddress) 2685 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2686 return RegMask::Empty; 2687 } 2688 2689 Node* InitializeNode::memory(uint alias_idx) { 2690 Node* mem = in(Memory); 2691 if (mem->is_MergeMem()) { 2692 return mem->as_MergeMem()->memory_at(alias_idx); 2693 } else { 2694 // incoming raw memory is not split 2695 return mem; 2696 } 2697 } 2698 2699 bool InitializeNode::is_non_zero() { 2700 if (is_complete()) return false; 2701 remove_extra_zeroes(); 2702 return (req() > RawStores); 2703 } 2704 2705 void InitializeNode::set_complete(PhaseGVN* phase) { 2706 assert(!is_complete(), "caller responsibility"); 2707 _is_complete = true; 2708 2709 // After this node is complete, it contains a bunch of 2710 // raw-memory initializations. There is no need for 2711 // it to have anything to do with non-raw memory effects. 2712 // Therefore, tell all non-raw users to re-optimize themselves, 2713 // after skipping the memory effects of this initialization. 2714 PhaseIterGVN* igvn = phase->is_IterGVN(); 2715 if (igvn) igvn->add_users_to_worklist(this); 2716 } 2717 2718 // convenience function 2719 // return false if the init contains any stores already 2720 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2721 InitializeNode* init = initialization(); 2722 if (init == NULL || init->is_complete()) return false; 2723 init->remove_extra_zeroes(); 2724 // for now, if this allocation has already collected any inits, bail: 2725 if (init->is_non_zero()) return false; 2726 init->set_complete(phase); 2727 return true; 2728 } 2729 2730 void InitializeNode::remove_extra_zeroes() { 2731 if (req() == RawStores) return; 2732 Node* zmem = zero_memory(); 2733 uint fill = RawStores; 2734 for (uint i = fill; i < req(); i++) { 2735 Node* n = in(i); 2736 if (n->is_top() || n == zmem) continue; // skip 2737 if (fill < i) set_req(fill, n); // compact 2738 ++fill; 2739 } 2740 // delete any empty spaces created: 2741 while (fill < req()) { 2742 del_req(fill); 2743 } 2744 } 2745 2746 // Helper for remembering which stores go with which offsets. 2747 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2748 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2749 intptr_t offset = -1; 2750 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2751 phase, offset); 2752 if (base == NULL) return -1; // something is dead, 2753 if (offset < 0) return -1; // dead, dead 2754 return offset; 2755 } 2756 2757 // Helper for proving that an initialization expression is 2758 // "simple enough" to be folded into an object initialization. 2759 // Attempts to prove that a store's initial value 'n' can be captured 2760 // within the initialization without creating a vicious cycle, such as: 2761 // { Foo p = new Foo(); p.next = p; } 2762 // True for constants and parameters and small combinations thereof. 2763 bool InitializeNode::detect_init_independence(Node* n, 2764 bool st_is_pinned, 2765 int& count) { 2766 if (n == NULL) return true; // (can this really happen?) 2767 if (n->is_Proj()) n = n->in(0); 2768 if (n == this) return false; // found a cycle 2769 if (n->is_Con()) return true; 2770 if (n->is_Start()) return true; // params, etc., are OK 2771 if (n->is_Root()) return true; // even better 2772 2773 Node* ctl = n->in(0); 2774 if (ctl != NULL && !ctl->is_top()) { 2775 if (ctl->is_Proj()) ctl = ctl->in(0); 2776 if (ctl == this) return false; 2777 2778 // If we already know that the enclosing memory op is pinned right after 2779 // the init, then any control flow that the store has picked up 2780 // must have preceded the init, or else be equal to the init. 2781 // Even after loop optimizations (which might change control edges) 2782 // a store is never pinned *before* the availability of its inputs. 2783 if (!MemNode::all_controls_dominate(n, this)) 2784 return false; // failed to prove a good control 2785 2786 } 2787 2788 // Check data edges for possible dependencies on 'this'. 2789 if ((count += 1) > 20) return false; // complexity limit 2790 for (uint i = 1; i < n->req(); i++) { 2791 Node* m = n->in(i); 2792 if (m == NULL || m == n || m->is_top()) continue; 2793 uint first_i = n->find_edge(m); 2794 if (i != first_i) continue; // process duplicate edge just once 2795 if (!detect_init_independence(m, st_is_pinned, count)) { 2796 return false; 2797 } 2798 } 2799 2800 return true; 2801 } 2802 2803 // Here are all the checks a Store must pass before it can be moved into 2804 // an initialization. Returns zero if a check fails. 2805 // On success, returns the (constant) offset to which the store applies, 2806 // within the initialized memory. 2807 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2808 const int FAIL = 0; 2809 if (st->req() != MemNode::ValueIn + 1) 2810 return FAIL; // an inscrutable StoreNode (card mark?) 2811 Node* ctl = st->in(MemNode::Control); 2812 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2813 return FAIL; // must be unconditional after the initialization 2814 Node* mem = st->in(MemNode::Memory); 2815 if (!(mem->is_Proj() && mem->in(0) == this)) 2816 return FAIL; // must not be preceded by other stores 2817 Node* adr = st->in(MemNode::Address); 2818 intptr_t offset; 2819 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2820 if (alloc == NULL) 2821 return FAIL; // inscrutable address 2822 if (alloc != allocation()) 2823 return FAIL; // wrong allocation! (store needs to float up) 2824 Node* val = st->in(MemNode::ValueIn); 2825 int complexity_count = 0; 2826 if (!detect_init_independence(val, true, complexity_count)) 2827 return FAIL; // stored value must be 'simple enough' 2828 2829 return offset; // success 2830 } 2831 2832 // Find the captured store in(i) which corresponds to the range 2833 // [start..start+size) in the initialized object. 2834 // If there is one, return its index i. If there isn't, return the 2835 // negative of the index where it should be inserted. 2836 // Return 0 if the queried range overlaps an initialization boundary 2837 // or if dead code is encountered. 2838 // If size_in_bytes is zero, do not bother with overlap checks. 2839 int InitializeNode::captured_store_insertion_point(intptr_t start, 2840 int size_in_bytes, 2841 PhaseTransform* phase) { 2842 const int FAIL = 0, MAX_STORE = BytesPerLong; 2843 2844 if (is_complete()) 2845 return FAIL; // arraycopy got here first; punt 2846 2847 assert(allocation() != NULL, "must be present"); 2848 2849 // no negatives, no header fields: 2850 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 2851 2852 // after a certain size, we bail out on tracking all the stores: 2853 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2854 if (start >= ti_limit) return FAIL; 2855 2856 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 2857 if (i >= limit) return -(int)i; // not found; here is where to put it 2858 2859 Node* st = in(i); 2860 intptr_t st_off = get_store_offset(st, phase); 2861 if (st_off < 0) { 2862 if (st != zero_memory()) { 2863 return FAIL; // bail out if there is dead garbage 2864 } 2865 } else if (st_off > start) { 2866 // ...we are done, since stores are ordered 2867 if (st_off < start + size_in_bytes) { 2868 return FAIL; // the next store overlaps 2869 } 2870 return -(int)i; // not found; here is where to put it 2871 } else if (st_off < start) { 2872 if (size_in_bytes != 0 && 2873 start < st_off + MAX_STORE && 2874 start < st_off + st->as_Store()->memory_size()) { 2875 return FAIL; // the previous store overlaps 2876 } 2877 } else { 2878 if (size_in_bytes != 0 && 2879 st->as_Store()->memory_size() != size_in_bytes) { 2880 return FAIL; // mismatched store size 2881 } 2882 return i; 2883 } 2884 2885 ++i; 2886 } 2887 } 2888 2889 // Look for a captured store which initializes at the offset 'start' 2890 // with the given size. If there is no such store, and no other 2891 // initialization interferes, then return zero_memory (the memory 2892 // projection of the AllocateNode). 2893 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 2894 PhaseTransform* phase) { 2895 assert(stores_are_sane(phase), ""); 2896 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2897 if (i == 0) { 2898 return NULL; // something is dead 2899 } else if (i < 0) { 2900 return zero_memory(); // just primordial zero bits here 2901 } else { 2902 Node* st = in(i); // here is the store at this position 2903 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 2904 return st; 2905 } 2906 } 2907 2908 // Create, as a raw pointer, an address within my new object at 'offset'. 2909 Node* InitializeNode::make_raw_address(intptr_t offset, 2910 PhaseTransform* phase) { 2911 Node* addr = in(RawAddress); 2912 if (offset != 0) { 2913 Compile* C = phase->C; 2914 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 2915 phase->MakeConX(offset)) ); 2916 } 2917 return addr; 2918 } 2919 2920 // Clone the given store, converting it into a raw store 2921 // initializing a field or element of my new object. 2922 // Caller is responsible for retiring the original store, 2923 // with subsume_node or the like. 2924 // 2925 // From the example above InitializeNode::InitializeNode, 2926 // here are the old stores to be captured: 2927 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2928 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2929 // 2930 // Here is the changed code; note the extra edges on init: 2931 // alloc = (Allocate ...) 2932 // rawoop = alloc.RawAddress 2933 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 2934 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 2935 // init = (Initialize alloc.Control alloc.Memory rawoop 2936 // rawstore1 rawstore2) 2937 // 2938 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 2939 PhaseTransform* phase) { 2940 assert(stores_are_sane(phase), ""); 2941 2942 if (start < 0) return NULL; 2943 assert(can_capture_store(st, phase) == start, "sanity"); 2944 2945 Compile* C = phase->C; 2946 int size_in_bytes = st->memory_size(); 2947 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2948 if (i == 0) return NULL; // bail out 2949 Node* prev_mem = NULL; // raw memory for the captured store 2950 if (i > 0) { 2951 prev_mem = in(i); // there is a pre-existing store under this one 2952 set_req(i, C->top()); // temporarily disconnect it 2953 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 2954 } else { 2955 i = -i; // no pre-existing store 2956 prev_mem = zero_memory(); // a slice of the newly allocated object 2957 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 2958 set_req(--i, C->top()); // reuse this edge; it has been folded away 2959 else 2960 ins_req(i, C->top()); // build a new edge 2961 } 2962 Node* new_st = st->clone(); 2963 new_st->set_req(MemNode::Control, in(Control)); 2964 new_st->set_req(MemNode::Memory, prev_mem); 2965 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 2966 new_st = phase->transform(new_st); 2967 2968 // At this point, new_st might have swallowed a pre-existing store 2969 // at the same offset, or perhaps new_st might have disappeared, 2970 // if it redundantly stored the same value (or zero to fresh memory). 2971 2972 // In any case, wire it in: 2973 set_req(i, new_st); 2974 2975 // The caller may now kill the old guy. 2976 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 2977 assert(check_st == new_st || check_st == NULL, "must be findable"); 2978 assert(!is_complete(), ""); 2979 return new_st; 2980 } 2981 2982 static bool store_constant(jlong* tiles, int num_tiles, 2983 intptr_t st_off, int st_size, 2984 jlong con) { 2985 if ((st_off & (st_size-1)) != 0) 2986 return false; // strange store offset (assume size==2**N) 2987 address addr = (address)tiles + st_off; 2988 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 2989 switch (st_size) { 2990 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 2991 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 2992 case sizeof(jint): *(jint*) addr = (jint) con; break; 2993 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 2994 default: return false; // strange store size (detect size!=2**N here) 2995 } 2996 return true; // return success to caller 2997 } 2998 2999 // Coalesce subword constants into int constants and possibly 3000 // into long constants. The goal, if the CPU permits, 3001 // is to initialize the object with a small number of 64-bit tiles. 3002 // Also, convert floating-point constants to bit patterns. 3003 // Non-constants are not relevant to this pass. 3004 // 3005 // In terms of the running example on InitializeNode::InitializeNode 3006 // and InitializeNode::capture_store, here is the transformation 3007 // of rawstore1 and rawstore2 into rawstore12: 3008 // alloc = (Allocate ...) 3009 // rawoop = alloc.RawAddress 3010 // tile12 = 0x00010002 3011 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3012 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3013 // 3014 void 3015 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3016 Node* size_in_bytes, 3017 PhaseGVN* phase) { 3018 Compile* C = phase->C; 3019 3020 assert(stores_are_sane(phase), ""); 3021 // Note: After this pass, they are not completely sane, 3022 // since there may be some overlaps. 3023 3024 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3025 3026 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3027 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3028 size_limit = MIN2(size_limit, ti_limit); 3029 size_limit = align_size_up(size_limit, BytesPerLong); 3030 int num_tiles = size_limit / BytesPerLong; 3031 3032 // allocate space for the tile map: 3033 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3034 jlong tiles_buf[small_len]; 3035 Node* nodes_buf[small_len]; 3036 jlong inits_buf[small_len]; 3037 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3038 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3039 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3040 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3041 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3042 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3043 // tiles: exact bitwise model of all primitive constants 3044 // nodes: last constant-storing node subsumed into the tiles model 3045 // inits: which bytes (in each tile) are touched by any initializations 3046 3047 //// Pass A: Fill in the tile model with any relevant stores. 3048 3049 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3050 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3051 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3052 Node* zmem = zero_memory(); // initially zero memory state 3053 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3054 Node* st = in(i); 3055 intptr_t st_off = get_store_offset(st, phase); 3056 3057 // Figure out the store's offset and constant value: 3058 if (st_off < header_size) continue; //skip (ignore header) 3059 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3060 int st_size = st->as_Store()->memory_size(); 3061 if (st_off + st_size > size_limit) break; 3062 3063 // Record which bytes are touched, whether by constant or not. 3064 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3065 continue; // skip (strange store size) 3066 3067 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3068 if (!val->singleton()) continue; //skip (non-con store) 3069 BasicType type = val->basic_type(); 3070 3071 jlong con = 0; 3072 switch (type) { 3073 case T_INT: con = val->is_int()->get_con(); break; 3074 case T_LONG: con = val->is_long()->get_con(); break; 3075 case T_FLOAT: con = jint_cast(val->getf()); break; 3076 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3077 default: continue; //skip (odd store type) 3078 } 3079 3080 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3081 st->Opcode() == Op_StoreL) { 3082 continue; // This StoreL is already optimal. 3083 } 3084 3085 // Store down the constant. 3086 store_constant(tiles, num_tiles, st_off, st_size, con); 3087 3088 intptr_t j = st_off >> LogBytesPerLong; 3089 3090 if (type == T_INT && st_size == BytesPerInt 3091 && (st_off & BytesPerInt) == BytesPerInt) { 3092 jlong lcon = tiles[j]; 3093 if (!Matcher::isSimpleConstant64(lcon) && 3094 st->Opcode() == Op_StoreI) { 3095 // This StoreI is already optimal by itself. 3096 jint* intcon = (jint*) &tiles[j]; 3097 intcon[1] = 0; // undo the store_constant() 3098 3099 // If the previous store is also optimal by itself, back up and 3100 // undo the action of the previous loop iteration... if we can. 3101 // But if we can't, just let the previous half take care of itself. 3102 st = nodes[j]; 3103 st_off -= BytesPerInt; 3104 con = intcon[0]; 3105 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3106 assert(st_off >= header_size, "still ignoring header"); 3107 assert(get_store_offset(st, phase) == st_off, "must be"); 3108 assert(in(i-1) == zmem, "must be"); 3109 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3110 assert(con == tcon->is_int()->get_con(), "must be"); 3111 // Undo the effects of the previous loop trip, which swallowed st: 3112 intcon[0] = 0; // undo store_constant() 3113 set_req(i-1, st); // undo set_req(i, zmem) 3114 nodes[j] = NULL; // undo nodes[j] = st 3115 --old_subword; // undo ++old_subword 3116 } 3117 continue; // This StoreI is already optimal. 3118 } 3119 } 3120 3121 // This store is not needed. 3122 set_req(i, zmem); 3123 nodes[j] = st; // record for the moment 3124 if (st_size < BytesPerLong) // something has changed 3125 ++old_subword; // includes int/float, but who's counting... 3126 else ++old_long; 3127 } 3128 3129 if ((old_subword + old_long) == 0) 3130 return; // nothing more to do 3131 3132 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3133 // Be sure to insert them before overlapping non-constant stores. 3134 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3135 for (int j = 0; j < num_tiles; j++) { 3136 jlong con = tiles[j]; 3137 jlong init = inits[j]; 3138 if (con == 0) continue; 3139 jint con0, con1; // split the constant, address-wise 3140 jint init0, init1; // split the init map, address-wise 3141 { union { jlong con; jint intcon[2]; } u; 3142 u.con = con; 3143 con0 = u.intcon[0]; 3144 con1 = u.intcon[1]; 3145 u.con = init; 3146 init0 = u.intcon[0]; 3147 init1 = u.intcon[1]; 3148 } 3149 3150 Node* old = nodes[j]; 3151 assert(old != NULL, "need the prior store"); 3152 intptr_t offset = (j * BytesPerLong); 3153 3154 bool split = !Matcher::isSimpleConstant64(con); 3155 3156 if (offset < header_size) { 3157 assert(offset + BytesPerInt >= header_size, "second int counts"); 3158 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3159 split = true; // only the second word counts 3160 // Example: int a[] = { 42 ... } 3161 } else if (con0 == 0 && init0 == -1) { 3162 split = true; // first word is covered by full inits 3163 // Example: int a[] = { ... foo(), 42 ... } 3164 } else if (con1 == 0 && init1 == -1) { 3165 split = true; // second word is covered by full inits 3166 // Example: int a[] = { ... 42, foo() ... } 3167 } 3168 3169 // Here's a case where init0 is neither 0 nor -1: 3170 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3171 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3172 // In this case the tile is not split; it is (jlong)42. 3173 // The big tile is stored down, and then the foo() value is inserted. 3174 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3175 3176 Node* ctl = old->in(MemNode::Control); 3177 Node* adr = make_raw_address(offset, phase); 3178 const TypePtr* atp = TypeRawPtr::BOTTOM; 3179 3180 // One or two coalesced stores to plop down. 3181 Node* st[2]; 3182 intptr_t off[2]; 3183 int nst = 0; 3184 if (!split) { 3185 ++new_long; 3186 off[nst] = offset; 3187 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3188 phase->longcon(con), T_LONG); 3189 } else { 3190 // Omit either if it is a zero. 3191 if (con0 != 0) { 3192 ++new_int; 3193 off[nst] = offset; 3194 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3195 phase->intcon(con0), T_INT); 3196 } 3197 if (con1 != 0) { 3198 ++new_int; 3199 offset += BytesPerInt; 3200 adr = make_raw_address(offset, phase); 3201 off[nst] = offset; 3202 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3203 phase->intcon(con1), T_INT); 3204 } 3205 } 3206 3207 // Insert second store first, then the first before the second. 3208 // Insert each one just before any overlapping non-constant stores. 3209 while (nst > 0) { 3210 Node* st1 = st[--nst]; 3211 C->copy_node_notes_to(st1, old); 3212 st1 = phase->transform(st1); 3213 offset = off[nst]; 3214 assert(offset >= header_size, "do not smash header"); 3215 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3216 guarantee(ins_idx != 0, "must re-insert constant store"); 3217 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3218 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3219 set_req(--ins_idx, st1); 3220 else 3221 ins_req(ins_idx, st1); 3222 } 3223 } 3224 3225 if (PrintCompilation && WizardMode) 3226 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3227 old_subword, old_long, new_int, new_long); 3228 if (C->log() != NULL) 3229 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3230 old_subword, old_long, new_int, new_long); 3231 3232 // Clean up any remaining occurrences of zmem: 3233 remove_extra_zeroes(); 3234 } 3235 3236 // Explore forward from in(start) to find the first fully initialized 3237 // word, and return its offset. Skip groups of subword stores which 3238 // together initialize full words. If in(start) is itself part of a 3239 // fully initialized word, return the offset of in(start). If there 3240 // are no following full-word stores, or if something is fishy, return 3241 // a negative value. 3242 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3243 int int_map = 0; 3244 intptr_t int_map_off = 0; 3245 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3246 3247 for (uint i = start, limit = req(); i < limit; i++) { 3248 Node* st = in(i); 3249 3250 intptr_t st_off = get_store_offset(st, phase); 3251 if (st_off < 0) break; // return conservative answer 3252 3253 int st_size = st->as_Store()->memory_size(); 3254 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3255 return st_off; // we found a complete word init 3256 } 3257 3258 // update the map: 3259 3260 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3261 if (this_int_off != int_map_off) { 3262 // reset the map: 3263 int_map = 0; 3264 int_map_off = this_int_off; 3265 } 3266 3267 int subword_off = st_off - this_int_off; 3268 int_map |= right_n_bits(st_size) << subword_off; 3269 if ((int_map & FULL_MAP) == FULL_MAP) { 3270 return this_int_off; // we found a complete word init 3271 } 3272 3273 // Did this store hit or cross the word boundary? 3274 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3275 if (next_int_off == this_int_off + BytesPerInt) { 3276 // We passed the current int, without fully initializing it. 3277 int_map_off = next_int_off; 3278 int_map >>= BytesPerInt; 3279 } else if (next_int_off > this_int_off + BytesPerInt) { 3280 // We passed the current and next int. 3281 return this_int_off + BytesPerInt; 3282 } 3283 } 3284 3285 return -1; 3286 } 3287 3288 3289 // Called when the associated AllocateNode is expanded into CFG. 3290 // At this point, we may perform additional optimizations. 3291 // Linearize the stores by ascending offset, to make memory 3292 // activity as coherent as possible. 3293 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3294 intptr_t header_size, 3295 Node* size_in_bytes, 3296 PhaseGVN* phase) { 3297 assert(!is_complete(), "not already complete"); 3298 assert(stores_are_sane(phase), ""); 3299 assert(allocation() != NULL, "must be present"); 3300 3301 remove_extra_zeroes(); 3302 3303 if (ReduceFieldZeroing || ReduceBulkZeroing) 3304 // reduce instruction count for common initialization patterns 3305 coalesce_subword_stores(header_size, size_in_bytes, phase); 3306 3307 Node* zmem = zero_memory(); // initially zero memory state 3308 Node* inits = zmem; // accumulating a linearized chain of inits 3309 #ifdef ASSERT 3310 intptr_t first_offset = allocation()->minimum_header_size(); 3311 intptr_t last_init_off = first_offset; // previous init offset 3312 intptr_t last_init_end = first_offset; // previous init offset+size 3313 intptr_t last_tile_end = first_offset; // previous tile offset+size 3314 #endif 3315 intptr_t zeroes_done = header_size; 3316 3317 bool do_zeroing = true; // we might give up if inits are very sparse 3318 int big_init_gaps = 0; // how many large gaps have we seen? 3319 3320 if (ZeroTLAB) do_zeroing = false; 3321 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3322 3323 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3324 Node* st = in(i); 3325 intptr_t st_off = get_store_offset(st, phase); 3326 if (st_off < 0) 3327 break; // unknown junk in the inits 3328 if (st->in(MemNode::Memory) != zmem) 3329 break; // complicated store chains somehow in list 3330 3331 int st_size = st->as_Store()->memory_size(); 3332 intptr_t next_init_off = st_off + st_size; 3333 3334 if (do_zeroing && zeroes_done < next_init_off) { 3335 // See if this store needs a zero before it or under it. 3336 intptr_t zeroes_needed = st_off; 3337 3338 if (st_size < BytesPerInt) { 3339 // Look for subword stores which only partially initialize words. 3340 // If we find some, we must lay down some word-level zeroes first, 3341 // underneath the subword stores. 3342 // 3343 // Examples: 3344 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3345 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3346 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3347 // 3348 // Note: coalesce_subword_stores may have already done this, 3349 // if it was prompted by constant non-zero subword initializers. 3350 // But this case can still arise with non-constant stores. 3351 3352 intptr_t next_full_store = find_next_fullword_store(i, phase); 3353 3354 // In the examples above: 3355 // in(i) p q r s x y z 3356 // st_off 12 13 14 15 12 13 14 3357 // st_size 1 1 1 1 1 1 1 3358 // next_full_s. 12 16 16 16 16 16 16 3359 // z's_done 12 16 16 16 12 16 12 3360 // z's_needed 12 16 16 16 16 16 16 3361 // zsize 0 0 0 0 4 0 4 3362 if (next_full_store < 0) { 3363 // Conservative tack: Zero to end of current word. 3364 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3365 } else { 3366 // Zero to beginning of next fully initialized word. 3367 // Or, don't zero at all, if we are already in that word. 3368 assert(next_full_store >= zeroes_needed, "must go forward"); 3369 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3370 zeroes_needed = next_full_store; 3371 } 3372 } 3373 3374 if (zeroes_needed > zeroes_done) { 3375 intptr_t zsize = zeroes_needed - zeroes_done; 3376 // Do some incremental zeroing on rawmem, in parallel with inits. 3377 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3378 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3379 zeroes_done, zeroes_needed, 3380 phase); 3381 zeroes_done = zeroes_needed; 3382 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3383 do_zeroing = false; // leave the hole, next time 3384 } 3385 } 3386 3387 // Collect the store and move on: 3388 st->set_req(MemNode::Memory, inits); 3389 inits = st; // put it on the linearized chain 3390 set_req(i, zmem); // unhook from previous position 3391 3392 if (zeroes_done == st_off) 3393 zeroes_done = next_init_off; 3394 3395 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3396 3397 #ifdef ASSERT 3398 // Various order invariants. Weaker than stores_are_sane because 3399 // a large constant tile can be filled in by smaller non-constant stores. 3400 assert(st_off >= last_init_off, "inits do not reverse"); 3401 last_init_off = st_off; 3402 const Type* val = NULL; 3403 if (st_size >= BytesPerInt && 3404 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3405 (int)val->basic_type() < (int)T_OBJECT) { 3406 assert(st_off >= last_tile_end, "tiles do not overlap"); 3407 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3408 last_tile_end = MAX2(last_tile_end, next_init_off); 3409 } else { 3410 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3411 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3412 assert(st_off >= last_init_end, "inits do not overlap"); 3413 last_init_end = next_init_off; // it's a non-tile 3414 } 3415 #endif //ASSERT 3416 } 3417 3418 remove_extra_zeroes(); // clear out all the zmems left over 3419 add_req(inits); 3420 3421 if (!ZeroTLAB) { 3422 // If anything remains to be zeroed, zero it all now. 3423 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3424 // if it is the last unused 4 bytes of an instance, forget about it 3425 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3426 if (zeroes_done + BytesPerLong >= size_limit) { 3427 assert(allocation() != NULL, ""); 3428 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3429 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3430 if (zeroes_done == k->layout_helper()) 3431 zeroes_done = size_limit; 3432 } 3433 if (zeroes_done < size_limit) { 3434 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3435 zeroes_done, size_in_bytes, phase); 3436 } 3437 } 3438 3439 set_complete(phase); 3440 return rawmem; 3441 } 3442 3443 3444 #ifdef ASSERT 3445 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3446 if (is_complete()) 3447 return true; // stores could be anything at this point 3448 assert(allocation() != NULL, "must be present"); 3449 intptr_t last_off = allocation()->minimum_header_size(); 3450 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3451 Node* st = in(i); 3452 intptr_t st_off = get_store_offset(st, phase); 3453 if (st_off < 0) continue; // ignore dead garbage 3454 if (last_off > st_off) { 3455 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3456 this->dump(2); 3457 assert(false, "ascending store offsets"); 3458 return false; 3459 } 3460 last_off = st_off + st->as_Store()->memory_size(); 3461 } 3462 return true; 3463 } 3464 #endif //ASSERT 3465 3466 3467 3468 3469 //============================MergeMemNode===================================== 3470 // 3471 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3472 // contributing store or call operations. Each contributor provides the memory 3473 // state for a particular "alias type" (see Compile::alias_type). For example, 3474 // if a MergeMem has an input X for alias category #6, then any memory reference 3475 // to alias category #6 may use X as its memory state input, as an exact equivalent 3476 // to using the MergeMem as a whole. 3477 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3478 // 3479 // (Here, the <N> notation gives the index of the relevant adr_type.) 3480 // 3481 // In one special case (and more cases in the future), alias categories overlap. 3482 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3483 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3484 // it is exactly equivalent to that state W: 3485 // MergeMem(<Bot>: W) <==> W 3486 // 3487 // Usually, the merge has more than one input. In that case, where inputs 3488 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3489 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3490 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3491 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3492 // 3493 // A merge can take a "wide" memory state as one of its narrow inputs. 3494 // This simply means that the merge observes out only the relevant parts of 3495 // the wide input. That is, wide memory states arriving at narrow merge inputs 3496 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3497 // 3498 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3499 // and that memory slices "leak through": 3500 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3501 // 3502 // But, in such a cascade, repeated memory slices can "block the leak": 3503 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3504 // 3505 // In the last example, Y is not part of the combined memory state of the 3506 // outermost MergeMem. The system must, of course, prevent unschedulable 3507 // memory states from arising, so you can be sure that the state Y is somehow 3508 // a precursor to state Y'. 3509 // 3510 // 3511 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3512 // of each MergeMemNode array are exactly the numerical alias indexes, including 3513 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3514 // Compile::alias_type (and kin) produce and manage these indexes. 3515 // 3516 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3517 // (Note that this provides quick access to the top node inside MergeMem methods, 3518 // without the need to reach out via TLS to Compile::current.) 3519 // 3520 // As a consequence of what was just described, a MergeMem that represents a full 3521 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3522 // containing all alias categories. 3523 // 3524 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3525 // 3526 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3527 // a memory state for the alias type <N>, or else the top node, meaning that 3528 // there is no particular input for that alias type. Note that the length of 3529 // a MergeMem is variable, and may be extended at any time to accommodate new 3530 // memory states at larger alias indexes. When merges grow, they are of course 3531 // filled with "top" in the unused in() positions. 3532 // 3533 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3534 // (Top was chosen because it works smoothly with passes like GCM.) 3535 // 3536 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3537 // the type of random VM bits like TLS references.) Since it is always the 3538 // first non-Bot memory slice, some low-level loops use it to initialize an 3539 // index variable: for (i = AliasIdxRaw; i < req(); i++). 3540 // 3541 // 3542 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3543 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3544 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 3545 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3546 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3547 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3548 // 3549 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3550 // really that different from the other memory inputs. An abbreviation called 3551 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3552 // 3553 // 3554 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3555 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3556 // that "emerges though" the base memory will be marked as excluding the alias types 3557 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 3558 // 3559 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3560 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3561 // 3562 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 3563 // (It is currently unimplemented.) As you can see, the resulting merge is 3564 // actually a disjoint union of memory states, rather than an overlay. 3565 // 3566 3567 //------------------------------MergeMemNode----------------------------------- 3568 Node* MergeMemNode::make_empty_memory() { 3569 Node* empty_memory = (Node*) Compile::current()->top(); 3570 assert(empty_memory->is_top(), "correct sentinel identity"); 3571 return empty_memory; 3572 } 3573 3574 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3575 init_class_id(Class_MergeMem); 3576 // all inputs are nullified in Node::Node(int) 3577 // set_input(0, NULL); // no control input 3578 3579 // Initialize the edges uniformly to top, for starters. 3580 Node* empty_mem = make_empty_memory(); 3581 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3582 init_req(i,empty_mem); 3583 } 3584 assert(empty_memory() == empty_mem, ""); 3585 3586 if( new_base != NULL && new_base->is_MergeMem() ) { 3587 MergeMemNode* mdef = new_base->as_MergeMem(); 3588 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3589 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3590 mms.set_memory(mms.memory2()); 3591 } 3592 assert(base_memory() == mdef->base_memory(), ""); 3593 } else { 3594 set_base_memory(new_base); 3595 } 3596 } 3597 3598 // Make a new, untransformed MergeMem with the same base as 'mem'. 3599 // If mem is itself a MergeMem, populate the result with the same edges. 3600 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3601 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3602 } 3603 3604 //------------------------------cmp-------------------------------------------- 3605 uint MergeMemNode::hash() const { return NO_HASH; } 3606 uint MergeMemNode::cmp( const Node &n ) const { 3607 return (&n == this); // Always fail except on self 3608 } 3609 3610 //------------------------------Identity--------------------------------------- 3611 Node* MergeMemNode::Identity(PhaseTransform *phase) { 3612 // Identity if this merge point does not record any interesting memory 3613 // disambiguations. 3614 Node* base_mem = base_memory(); 3615 Node* empty_mem = empty_memory(); 3616 if (base_mem != empty_mem) { // Memory path is not dead? 3617 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3618 Node* mem = in(i); 3619 if (mem != empty_mem && mem != base_mem) { 3620 return this; // Many memory splits; no change 3621 } 3622 } 3623 } 3624 return base_mem; // No memory splits; ID on the one true input 3625 } 3626 3627 //------------------------------Ideal------------------------------------------ 3628 // This method is invoked recursively on chains of MergeMem nodes 3629 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3630 // Remove chain'd MergeMems 3631 // 3632 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3633 // relative to the "in(Bot)". Since we are patching both at the same time, 3634 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3635 // but rewrite each "in(i)" relative to the new "in(Bot)". 3636 Node *progress = NULL; 3637 3638 3639 Node* old_base = base_memory(); 3640 Node* empty_mem = empty_memory(); 3641 if (old_base == empty_mem) 3642 return NULL; // Dead memory path. 3643 3644 MergeMemNode* old_mbase; 3645 if (old_base != NULL && old_base->is_MergeMem()) 3646 old_mbase = old_base->as_MergeMem(); 3647 else 3648 old_mbase = NULL; 3649 Node* new_base = old_base; 3650 3651 // simplify stacked MergeMems in base memory 3652 if (old_mbase) new_base = old_mbase->base_memory(); 3653 3654 // the base memory might contribute new slices beyond my req() 3655 if (old_mbase) grow_to_match(old_mbase); 3656 3657 // Look carefully at the base node if it is a phi. 3658 PhiNode* phi_base; 3659 if (new_base != NULL && new_base->is_Phi()) 3660 phi_base = new_base->as_Phi(); 3661 else 3662 phi_base = NULL; 3663 3664 Node* phi_reg = NULL; 3665 uint phi_len = (uint)-1; 3666 if (phi_base != NULL && !phi_base->is_copy()) { 3667 // do not examine phi if degraded to a copy 3668 phi_reg = phi_base->region(); 3669 phi_len = phi_base->req(); 3670 // see if the phi is unfinished 3671 for (uint i = 1; i < phi_len; i++) { 3672 if (phi_base->in(i) == NULL) { 3673 // incomplete phi; do not look at it yet! 3674 phi_reg = NULL; 3675 phi_len = (uint)-1; 3676 break; 3677 } 3678 } 3679 } 3680 3681 // Note: We do not call verify_sparse on entry, because inputs 3682 // can normalize to the base_memory via subsume_node or similar 3683 // mechanisms. This method repairs that damage. 3684 3685 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3686 3687 // Look at each slice. 3688 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3689 Node* old_in = in(i); 3690 // calculate the old memory value 3691 Node* old_mem = old_in; 3692 if (old_mem == empty_mem) old_mem = old_base; 3693 assert(old_mem == memory_at(i), ""); 3694 3695 // maybe update (reslice) the old memory value 3696 3697 // simplify stacked MergeMems 3698 Node* new_mem = old_mem; 3699 MergeMemNode* old_mmem; 3700 if (old_mem != NULL && old_mem->is_MergeMem()) 3701 old_mmem = old_mem->as_MergeMem(); 3702 else 3703 old_mmem = NULL; 3704 if (old_mmem == this) { 3705 // This can happen if loops break up and safepoints disappear. 3706 // A merge of BotPtr (default) with a RawPtr memory derived from a 3707 // safepoint can be rewritten to a merge of the same BotPtr with 3708 // the BotPtr phi coming into the loop. If that phi disappears 3709 // also, we can end up with a self-loop of the mergemem. 3710 // In general, if loops degenerate and memory effects disappear, 3711 // a mergemem can be left looking at itself. This simply means 3712 // that the mergemem's default should be used, since there is 3713 // no longer any apparent effect on this slice. 3714 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3715 // from start. Update the input to TOP. 3716 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3717 } 3718 else if (old_mmem != NULL) { 3719 new_mem = old_mmem->memory_at(i); 3720 } 3721 // else preceding memory was not a MergeMem 3722 3723 // replace equivalent phis (unfortunately, they do not GVN together) 3724 if (new_mem != NULL && new_mem != new_base && 3725 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3726 if (new_mem->is_Phi()) { 3727 PhiNode* phi_mem = new_mem->as_Phi(); 3728 for (uint i = 1; i < phi_len; i++) { 3729 if (phi_base->in(i) != phi_mem->in(i)) { 3730 phi_mem = NULL; 3731 break; 3732 } 3733 } 3734 if (phi_mem != NULL) { 3735 // equivalent phi nodes; revert to the def 3736 new_mem = new_base; 3737 } 3738 } 3739 } 3740 3741 // maybe store down a new value 3742 Node* new_in = new_mem; 3743 if (new_in == new_base) new_in = empty_mem; 3744 3745 if (new_in != old_in) { 3746 // Warning: Do not combine this "if" with the previous "if" 3747 // A memory slice might have be be rewritten even if it is semantically 3748 // unchanged, if the base_memory value has changed. 3749 set_req(i, new_in); 3750 progress = this; // Report progress 3751 } 3752 } 3753 3754 if (new_base != old_base) { 3755 set_req(Compile::AliasIdxBot, new_base); 3756 // Don't use set_base_memory(new_base), because we need to update du. 3757 assert(base_memory() == new_base, ""); 3758 progress = this; 3759 } 3760 3761 if( base_memory() == this ) { 3762 // a self cycle indicates this memory path is dead 3763 set_req(Compile::AliasIdxBot, empty_mem); 3764 } 3765 3766 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3767 // Recursion must occur after the self cycle check above 3768 if( base_memory()->is_MergeMem() ) { 3769 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3770 Node *m = phase->transform(new_mbase); // Rollup any cycles 3771 if( m != NULL && (m->is_top() || 3772 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3773 // propagate rollup of dead cycle to self 3774 set_req(Compile::AliasIdxBot, empty_mem); 3775 } 3776 } 3777 3778 if( base_memory() == empty_mem ) { 3779 progress = this; 3780 // Cut inputs during Parse phase only. 3781 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3782 if( !can_reshape ) { 3783 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3784 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3785 } 3786 } 3787 } 3788 3789 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3790 // Check if PhiNode::Ideal's "Split phis through memory merges" 3791 // transform should be attempted. Look for this->phi->this cycle. 3792 uint merge_width = req(); 3793 if (merge_width > Compile::AliasIdxRaw) { 3794 PhiNode* phi = base_memory()->as_Phi(); 3795 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3796 if (phi->in(i) == this) { 3797 phase->is_IterGVN()->_worklist.push(phi); 3798 break; 3799 } 3800 } 3801 } 3802 } 3803 3804 assert(progress || verify_sparse(), "please, no dups of base"); 3805 return progress; 3806 } 3807 3808 //-------------------------set_base_memory------------------------------------- 3809 void MergeMemNode::set_base_memory(Node *new_base) { 3810 Node* empty_mem = empty_memory(); 3811 set_req(Compile::AliasIdxBot, new_base); 3812 assert(memory_at(req()) == new_base, "must set default memory"); 3813 // Clear out other occurrences of new_base: 3814 if (new_base != empty_mem) { 3815 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3816 if (in(i) == new_base) set_req(i, empty_mem); 3817 } 3818 } 3819 } 3820 3821 //------------------------------out_RegMask------------------------------------ 3822 const RegMask &MergeMemNode::out_RegMask() const { 3823 return RegMask::Empty; 3824 } 3825 3826 //------------------------------dump_spec-------------------------------------- 3827 #ifndef PRODUCT 3828 void MergeMemNode::dump_spec(outputStream *st) const { 3829 st->print(" {"); 3830 Node* base_mem = base_memory(); 3831 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3832 Node* mem = memory_at(i); 3833 if (mem == base_mem) { st->print(" -"); continue; } 3834 st->print( " N%d:", mem->_idx ); 3835 Compile::current()->get_adr_type(i)->dump_on(st); 3836 } 3837 st->print(" }"); 3838 } 3839 #endif // !PRODUCT 3840 3841 3842 #ifdef ASSERT 3843 static bool might_be_same(Node* a, Node* b) { 3844 if (a == b) return true; 3845 if (!(a->is_Phi() || b->is_Phi())) return false; 3846 // phis shift around during optimization 3847 return true; // pretty stupid... 3848 } 3849 3850 // verify a narrow slice (either incoming or outgoing) 3851 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 3852 if (!VerifyAliases) return; // don't bother to verify unless requested 3853 if (is_error_reported()) return; // muzzle asserts when debugging an error 3854 if (Node::in_dump()) return; // muzzle asserts when printing 3855 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 3856 assert(n != NULL, ""); 3857 // Elide intervening MergeMem's 3858 while (n->is_MergeMem()) { 3859 n = n->as_MergeMem()->memory_at(alias_idx); 3860 } 3861 Compile* C = Compile::current(); 3862 const TypePtr* n_adr_type = n->adr_type(); 3863 if (n == m->empty_memory()) { 3864 // Implicit copy of base_memory() 3865 } else if (n_adr_type != TypePtr::BOTTOM) { 3866 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 3867 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 3868 } else { 3869 // A few places like make_runtime_call "know" that VM calls are narrow, 3870 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 3871 bool expected_wide_mem = false; 3872 if (n == m->base_memory()) { 3873 expected_wide_mem = true; 3874 } else if (alias_idx == Compile::AliasIdxRaw || 3875 n == m->memory_at(Compile::AliasIdxRaw)) { 3876 expected_wide_mem = true; 3877 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 3878 // memory can "leak through" calls on channels that 3879 // are write-once. Allow this also. 3880 expected_wide_mem = true; 3881 } 3882 assert(expected_wide_mem, "expected narrow slice replacement"); 3883 } 3884 } 3885 #else // !ASSERT 3886 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 3887 #endif 3888 3889 3890 //-----------------------------memory_at--------------------------------------- 3891 Node* MergeMemNode::memory_at(uint alias_idx) const { 3892 assert(alias_idx >= Compile::AliasIdxRaw || 3893 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 3894 "must avoid base_memory and AliasIdxTop"); 3895 3896 // Otherwise, it is a narrow slice. 3897 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 3898 Compile *C = Compile::current(); 3899 if (is_empty_memory(n)) { 3900 // the array is sparse; empty slots are the "top" node 3901 n = base_memory(); 3902 assert(Node::in_dump() 3903 || n == NULL || n->bottom_type() == Type::TOP 3904 || n->adr_type() == TypePtr::BOTTOM 3905 || n->adr_type() == TypeRawPtr::BOTTOM 3906 || Compile::current()->AliasLevel() == 0, 3907 "must be a wide memory"); 3908 // AliasLevel == 0 if we are organizing the memory states manually. 3909 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 3910 } else { 3911 // make sure the stored slice is sane 3912 #ifdef ASSERT 3913 if (is_error_reported() || Node::in_dump()) { 3914 } else if (might_be_same(n, base_memory())) { 3915 // Give it a pass: It is a mostly harmless repetition of the base. 3916 // This can arise normally from node subsumption during optimization. 3917 } else { 3918 verify_memory_slice(this, alias_idx, n); 3919 } 3920 #endif 3921 } 3922 return n; 3923 } 3924 3925 //---------------------------set_memory_at------------------------------------- 3926 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 3927 verify_memory_slice(this, alias_idx, n); 3928 Node* empty_mem = empty_memory(); 3929 if (n == base_memory()) n = empty_mem; // collapse default 3930 uint need_req = alias_idx+1; 3931 if (req() < need_req) { 3932 if (n == empty_mem) return; // already the default, so do not grow me 3933 // grow the sparse array 3934 do { 3935 add_req(empty_mem); 3936 } while (req() < need_req); 3937 } 3938 set_req( alias_idx, n ); 3939 } 3940 3941 3942 3943 //--------------------------iteration_setup------------------------------------ 3944 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 3945 if (other != NULL) { 3946 grow_to_match(other); 3947 // invariant: the finite support of mm2 is within mm->req() 3948 #ifdef ASSERT 3949 for (uint i = req(); i < other->req(); i++) { 3950 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 3951 } 3952 #endif 3953 } 3954 // Replace spurious copies of base_memory by top. 3955 Node* base_mem = base_memory(); 3956 if (base_mem != NULL && !base_mem->is_top()) { 3957 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 3958 if (in(i) == base_mem) 3959 set_req(i, empty_memory()); 3960 } 3961 } 3962 } 3963 3964 //---------------------------grow_to_match------------------------------------- 3965 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 3966 Node* empty_mem = empty_memory(); 3967 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 3968 // look for the finite support of the other memory 3969 for (uint i = other->req(); --i >= req(); ) { 3970 if (other->in(i) != empty_mem) { 3971 uint new_len = i+1; 3972 while (req() < new_len) add_req(empty_mem); 3973 break; 3974 } 3975 } 3976 } 3977 3978 //---------------------------verify_sparse------------------------------------- 3979 #ifndef PRODUCT 3980 bool MergeMemNode::verify_sparse() const { 3981 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 3982 Node* base_mem = base_memory(); 3983 // The following can happen in degenerate cases, since empty==top. 3984 if (is_empty_memory(base_mem)) return true; 3985 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3986 assert(in(i) != NULL, "sane slice"); 3987 if (in(i) == base_mem) return false; // should have been the sentinel value! 3988 } 3989 return true; 3990 } 3991 3992 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 3993 Node* n; 3994 n = mm->in(idx); 3995 if (mem == n) return true; // might be empty_memory() 3996 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 3997 if (mem == n) return true; 3998 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 3999 if (mem == n) return true; 4000 if (n == NULL) break; 4001 } 4002 return false; 4003 } 4004 #endif // !PRODUCT