1 #ifdef USE_PRAGMA_IDENT_SRC 2 #pragma ident "@(#)memnode.cpp 1.239 08/11/24 12:23:43 JVM" 3 #endif 4 /* 5 * Copyright 1997-2008 Sun Microsystems, Inc. All Rights Reserved. 6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 7 * 8 * This code is free software; you can redistribute it and/or modify it 9 * under the terms of the GNU General Public License version 2 only, as 10 * published by the Free Software Foundation. 11 * 12 * This code is distributed in the hope that it will be useful, but WITHOUT 13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 * version 2 for more details (a copy is included in the LICENSE file that 16 * accompanied this code). 17 * 18 * You should have received a copy of the GNU General Public License version 19 * 2 along with this work; if not, write to the Free Software Foundation, 20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 21 * 22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 23 * CA 95054 USA or visit www.sun.com if you need additional information or 24 * have any questions. 25 * 26 */ 27 28 // Portions of code courtesy of Clifford Click 29 30 // Optimization - Graph Style 31 32 #include "incls/_precompiled.incl" 33 #include "incls/_memnode.cpp.incl" 34 35 //============================================================================= 36 uint MemNode::size_of() const { return sizeof(*this); } 37 38 const TypePtr *MemNode::adr_type() const { 39 Node* adr = in(Address); 40 const TypePtr* cross_check = NULL; 41 DEBUG_ONLY(cross_check = _adr_type); 42 return calculate_adr_type(adr->bottom_type(), cross_check); 43 } 44 45 #ifndef PRODUCT 46 void MemNode::dump_spec(outputStream *st) const { 47 if (in(Address) == NULL) return; // node is dead 48 #ifndef ASSERT 49 // fake the missing field 50 const TypePtr* _adr_type = NULL; 51 if (in(Address) != NULL) 52 _adr_type = in(Address)->bottom_type()->isa_ptr(); 53 #endif 54 dump_adr_type(this, _adr_type, st); 55 56 Compile* C = Compile::current(); 57 if( C->alias_type(_adr_type)->is_volatile() ) 58 st->print(" Volatile!"); 59 } 60 61 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 62 st->print(" @"); 63 if (adr_type == NULL) { 64 st->print("NULL"); 65 } else { 66 adr_type->dump_on(st); 67 Compile* C = Compile::current(); 68 Compile::AliasType* atp = NULL; 69 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 70 if (atp == NULL) 71 st->print(", idx=?\?;"); 72 else if (atp->index() == Compile::AliasIdxBot) 73 st->print(", idx=Bot;"); 74 else if (atp->index() == Compile::AliasIdxTop) 75 st->print(", idx=Top;"); 76 else if (atp->index() == Compile::AliasIdxRaw) 77 st->print(", idx=Raw;"); 78 else { 79 ciField* field = atp->field(); 80 if (field) { 81 st->print(", name="); 82 field->print_name_on(st); 83 } 84 st->print(", idx=%d;", atp->index()); 85 } 86 } 87 } 88 89 extern void print_alias_types(); 90 91 #endif 92 93 //--------------------------Ideal_common--------------------------------------- 94 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 95 // Unhook non-raw memories from complete (macro-expanded) initializations. 96 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 97 // If our control input is a dead region, kill all below the region 98 Node *ctl = in(MemNode::Control); 99 if (ctl && remove_dead_region(phase, can_reshape)) 100 return this; 101 102 // Ignore if memory is dead, or self-loop 103 Node *mem = in(MemNode::Memory); 104 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 105 assert( mem != this, "dead loop in MemNode::Ideal" ); 106 107 Node *address = in(MemNode::Address); 108 const Type *t_adr = phase->type( address ); 109 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 110 111 // Avoid independent memory operations 112 Node* old_mem = mem; 113 114 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 115 InitializeNode* init = mem->in(0)->as_Initialize(); 116 if (init->is_complete()) { // i.e., after macro expansion 117 const TypePtr* tp = t_adr->is_ptr(); 118 uint alias_idx = phase->C->get_alias_index(tp); 119 // Free this slice from the init. It was hooked, temporarily, 120 // by GraphKit::set_output_for_allocation. 121 if (alias_idx > Compile::AliasIdxRaw) { 122 mem = init->memory(alias_idx); 123 // ...but not with the raw-pointer slice. 124 } 125 } 126 } 127 128 if (mem->is_MergeMem()) { 129 MergeMemNode* mmem = mem->as_MergeMem(); 130 const TypePtr *tp = t_adr->is_ptr(); 131 uint alias_idx = phase->C->get_alias_index(tp); 132 #ifdef ASSERT 133 { 134 // Check that current type is consistent with the alias index used during graph construction 135 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 136 const TypePtr *adr_t = adr_type(); 137 bool consistent = adr_t == NULL || adr_t->empty() || phase->C->must_alias(adr_t, alias_idx ); 138 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 139 if( !consistent && adr_t != NULL && !adr_t->empty() && 140 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 141 adr_t->isa_aryptr() && adr_t->offset() != Type::OffsetBot && 142 ( adr_t->offset() == arrayOopDesc::length_offset_in_bytes() || 143 adr_t->offset() == oopDesc::klass_offset_in_bytes() || 144 adr_t->offset() == oopDesc::mark_offset_in_bytes() ) ) { 145 // don't assert if it is dead code. 146 consistent = true; 147 } 148 if( !consistent ) { 149 tty->print("alias_idx==%d, adr_type()==", alias_idx); if( adr_t == NULL ) { tty->print("NULL"); } else { adr_t->dump(); } 150 tty->cr(); 151 print_alias_types(); 152 assert(consistent, "adr_type must match alias idx"); 153 } 154 } 155 #endif 156 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 157 // means an array I have not precisely typed yet. Do not do any 158 // alias stuff with it any time soon. 159 const TypeInstPtr *tinst = tp->isa_instptr(); 160 if( tp->base() != Type::AnyPtr && 161 !(tinst && 162 tinst->klass()->is_java_lang_Object() && 163 tinst->offset() == Type::OffsetBot) ) { 164 // compress paths and change unreachable cycles to TOP 165 // If not, we can update the input infinitely along a MergeMem cycle 166 // Equivalent code in PhiNode::Ideal 167 Node* m = phase->transform(mmem); 168 // If tranformed to a MergeMem, get the desired slice 169 // Otherwise the returned node represents memory for every slice 170 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 171 // Update input if it is progress over what we have now 172 } 173 } 174 175 if (mem != old_mem) { 176 set_req(MemNode::Memory, mem); 177 return this; 178 } 179 180 // let the subclass continue analyzing... 181 return NULL; 182 } 183 184 // Helper function for proving some simple control dominations. 185 // Attempt to prove that control input 'dom' dominates (or equals) 'sub'. 186 // Already assumes that 'dom' is available at 'sub', and that 'sub' 187 // is not a constant (dominated by the method's StartNode). 188 // Used by MemNode::find_previous_store to prove that the 189 // control input of a memory operation predates (dominates) 190 // an allocation it wants to look past. 191 bool MemNode::detect_dominating_control(Node* dom, Node* sub) { 192 if (dom == NULL) return false; 193 if (dom->is_Proj()) dom = dom->in(0); 194 if (dom->is_Start()) return true; // anything inside the method 195 if (dom->is_Root()) return true; // dom 'controls' a constant 196 int cnt = 20; // detect cycle or too much effort 197 while (sub != NULL) { // walk 'sub' up the chain to 'dom' 198 if (--cnt < 0) return false; // in a cycle or too complex 199 if (sub == dom) return true; 200 if (sub->is_Start()) return false; 201 if (sub->is_Root()) return false; 202 Node* up = sub->in(0); 203 if (sub == up && sub->is_Region()) { 204 for (uint i = 1; i < sub->req(); i++) { 205 Node* in = sub->in(i); 206 if (in != NULL && !in->is_top() && in != sub) { 207 up = in; break; // take any path on the way up to 'dom' 208 } 209 } 210 } 211 if (sub == up) return false; // some kind of tight cycle 212 sub = up; 213 } 214 return false; 215 } 216 217 //---------------------detect_ptr_independence--------------------------------- 218 // Used by MemNode::find_previous_store to prove that two base 219 // pointers are never equal. 220 // The pointers are accompanied by their associated allocations, 221 // if any, which have been previously discovered by the caller. 222 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 223 Node* p2, AllocateNode* a2, 224 PhaseTransform* phase) { 225 // Attempt to prove that these two pointers cannot be aliased. 226 // They may both manifestly be allocations, and they should differ. 227 // Or, if they are not both allocations, they can be distinct constants. 228 // Otherwise, one is an allocation and the other a pre-existing value. 229 if (a1 == NULL && a2 == NULL) { // neither an allocation 230 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 231 } else if (a1 != NULL && a2 != NULL) { // both allocations 232 return (a1 != a2); 233 } else if (a1 != NULL) { // one allocation a1 234 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 235 return detect_dominating_control(p2->in(0), a1->in(0)); 236 } else { //(a2 != NULL) // one allocation a2 237 return detect_dominating_control(p1->in(0), a2->in(0)); 238 } 239 return false; 240 } 241 242 243 // The logic for reordering loads and stores uses four steps: 244 // (a) Walk carefully past stores and initializations which we 245 // can prove are independent of this load. 246 // (b) Observe that the next memory state makes an exact match 247 // with self (load or store), and locate the relevant store. 248 // (c) Ensure that, if we were to wire self directly to the store, 249 // the optimizer would fold it up somehow. 250 // (d) Do the rewiring, and return, depending on some other part of 251 // the optimizer to fold up the load. 252 // This routine handles steps (a) and (b). Steps (c) and (d) are 253 // specific to loads and stores, so they are handled by the callers. 254 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 255 // 256 Node* MemNode::find_previous_store(PhaseTransform* phase) { 257 Node* ctrl = in(MemNode::Control); 258 Node* adr = in(MemNode::Address); 259 intptr_t offset = 0; 260 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 261 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 262 263 if (offset == Type::OffsetBot) 264 return NULL; // cannot unalias unless there are precise offsets 265 266 intptr_t size_in_bytes = memory_size(); 267 268 Node* mem = in(MemNode::Memory); // start searching here... 269 270 int cnt = 50; // Cycle limiter 271 for (;;) { // While we can dance past unrelated stores... 272 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 273 274 if (mem->is_Store()) { 275 Node* st_adr = mem->in(MemNode::Address); 276 intptr_t st_offset = 0; 277 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 278 if (st_base == NULL) 279 break; // inscrutable pointer 280 if (st_offset != offset && st_offset != Type::OffsetBot) { 281 const int MAX_STORE = BytesPerLong; 282 if (st_offset >= offset + size_in_bytes || 283 st_offset <= offset - MAX_STORE || 284 st_offset <= offset - mem->as_Store()->memory_size()) { 285 // Success: The offsets are provably independent. 286 // (You may ask, why not just test st_offset != offset and be done? 287 // The answer is that stores of different sizes can co-exist 288 // in the same sequence of RawMem effects. We sometimes initialize 289 // a whole 'tile' of array elements with a single jint or jlong.) 290 mem = mem->in(MemNode::Memory); 291 continue; // (a) advance through independent store memory 292 } 293 } 294 if (st_base != base && 295 detect_ptr_independence(base, alloc, 296 st_base, 297 AllocateNode::Ideal_allocation(st_base, phase), 298 phase)) { 299 // Success: The bases are provably independent. 300 mem = mem->in(MemNode::Memory); 301 continue; // (a) advance through independent store memory 302 } 303 304 // (b) At this point, if the bases or offsets do not agree, we lose, 305 // since we have not managed to prove 'this' and 'mem' independent. 306 if (st_base == base && st_offset == offset) { 307 return mem; // let caller handle steps (c), (d) 308 } 309 310 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 311 InitializeNode* st_init = mem->in(0)->as_Initialize(); 312 AllocateNode* st_alloc = st_init->allocation(); 313 if (st_alloc == NULL) 314 break; // something degenerated 315 bool known_identical = false; 316 bool known_independent = false; 317 if (alloc == st_alloc) 318 known_identical = true; 319 else if (alloc != NULL) 320 known_independent = true; 321 else if (ctrl != NULL && 322 detect_dominating_control(ctrl, st_alloc->in(0))) 323 known_independent = true; 324 325 if (known_independent) { 326 // The bases are provably independent: Either they are 327 // manifestly distinct allocations, or else the control 328 // of this load dominates the store's allocation. 329 int alias_idx = phase->C->get_alias_index(adr_type()); 330 if (alias_idx == Compile::AliasIdxRaw) { 331 mem = st_alloc->in(TypeFunc::Memory); 332 } else { 333 mem = st_init->memory(alias_idx); 334 } 335 continue; // (a) advance through independent store memory 336 } 337 338 // (b) at this point, if we are not looking at a store initializing 339 // the same allocation we are loading from, we lose. 340 if (known_identical) { 341 // From caller, can_see_stored_value will consult find_captured_store. 342 return mem; // let caller handle steps (c), (d) 343 } 344 345 } 346 347 // Unless there is an explicit 'continue', we must bail out here, 348 // because 'mem' is an inscrutable memory state (e.g., a call). 349 break; 350 } 351 352 return NULL; // bail out 353 } 354 355 //----------------------calculate_adr_type------------------------------------- 356 // Helper function. Notices when the given type of address hits top or bottom. 357 // Also, asserts a cross-check of the type against the expected address type. 358 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 359 if (t == Type::TOP) return NULL; // does not touch memory any more? 360 #ifdef PRODUCT 361 cross_check = NULL; 362 #else 363 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 364 #endif 365 const TypePtr* tp = t->isa_ptr(); 366 if (tp == NULL) { 367 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 368 return TypePtr::BOTTOM; // touches lots of memory 369 } else { 370 #ifdef ASSERT 371 // %%%% [phh] We don't check the alias index if cross_check is 372 // TypeRawPtr::BOTTOM. Needs to be investigated. 373 if (cross_check != NULL && 374 cross_check != TypePtr::BOTTOM && 375 cross_check != TypeRawPtr::BOTTOM) { 376 // Recheck the alias index, to see if it has changed (due to a bug). 377 Compile* C = Compile::current(); 378 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 379 "must stay in the original alias category"); 380 // The type of the address must be contained in the adr_type, 381 // disregarding "null"-ness. 382 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 383 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 384 assert(cross_check->meet(tp_notnull) == cross_check, 385 "real address must not escape from expected memory type"); 386 } 387 #endif 388 return tp; 389 } 390 } 391 392 //------------------------adr_phi_is_loop_invariant---------------------------- 393 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 394 // loop is loop invariant. Make a quick traversal of Phi and associated 395 // CastPP nodes, looking to see if they are a closed group within the loop. 396 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 397 // The idea is that the phi-nest must boil down to only CastPP nodes 398 // with the same data. This implies that any path into the loop already 399 // includes such a CastPP, and so the original cast, whatever its input, 400 // must be covered by an equivalent cast, with an earlier control input. 401 ResourceMark rm; 402 403 // The loop entry input of the phi should be the unique dominating 404 // node for every Phi/CastPP in the loop. 405 Unique_Node_List closure; 406 closure.push(adr_phi->in(LoopNode::EntryControl)); 407 408 // Add the phi node and the cast to the worklist. 409 Unique_Node_List worklist; 410 worklist.push(adr_phi); 411 if( cast != NULL ){ 412 if( !cast->is_ConstraintCast() ) return false; 413 worklist.push(cast); 414 } 415 416 // Begin recursive walk of phi nodes. 417 while( worklist.size() ){ 418 // Take a node off the worklist 419 Node *n = worklist.pop(); 420 if( !closure.member(n) ){ 421 // Add it to the closure. 422 closure.push(n); 423 // Make a sanity check to ensure we don't waste too much time here. 424 if( closure.size() > 20) return false; 425 // This node is OK if: 426 // - it is a cast of an identical value 427 // - or it is a phi node (then we add its inputs to the worklist) 428 // Otherwise, the node is not OK, and we presume the cast is not invariant 429 if( n->is_ConstraintCast() ){ 430 worklist.push(n->in(1)); 431 } else if( n->is_Phi() ) { 432 for( uint i = 1; i < n->req(); i++ ) { 433 worklist.push(n->in(i)); 434 } 435 } else { 436 return false; 437 } 438 } 439 } 440 441 // Quit when the worklist is empty, and we've found no offending nodes. 442 return true; 443 } 444 445 //------------------------------Ideal_DU_postCCP------------------------------- 446 // Find any cast-away of null-ness and keep its control. Null cast-aways are 447 // going away in this pass and we need to make this memory op depend on the 448 // gating null check. 449 450 // I tried to leave the CastPP's in. This makes the graph more accurate in 451 // some sense; we get to keep around the knowledge that an oop is not-null 452 // after some test. Alas, the CastPP's interfere with GVN (some values are 453 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 454 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 455 // some of the more trivial cases in the optimizer. Removing more useless 456 // Phi's started allowing Loads to illegally float above null checks. I gave 457 // up on this approach. CNC 10/20/2000 458 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 459 Node *ctr = in(MemNode::Control); 460 Node *mem = in(MemNode::Memory); 461 Node *adr = in(MemNode::Address); 462 Node *skipped_cast = NULL; 463 // Need a null check? Regular static accesses do not because they are 464 // from constant addresses. Array ops are gated by the range check (which 465 // always includes a NULL check). Just check field ops. 466 if( !ctr ) { 467 // Scan upwards for the highest location we can place this memory op. 468 while( true ) { 469 switch( adr->Opcode() ) { 470 471 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 472 adr = adr->in(AddPNode::Base); 473 continue; 474 475 case Op_CastPP: 476 // If the CastPP is useless, just peek on through it. 477 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 478 // Remember the cast that we've peeked though. If we peek 479 // through more than one, then we end up remembering the highest 480 // one, that is, if in a loop, the one closest to the top. 481 skipped_cast = adr; 482 adr = adr->in(1); 483 continue; 484 } 485 // CastPP is going away in this pass! We need this memory op to be 486 // control-dependent on the test that is guarding the CastPP. 487 ccp->hash_delete(this); 488 set_req(MemNode::Control, adr->in(0)); 489 ccp->hash_insert(this); 490 return this; 491 492 case Op_Phi: 493 // Attempt to float above a Phi to some dominating point. 494 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 495 // If we've already peeked through a Cast (which could have set the 496 // control), we can't float above a Phi, because the skipped Cast 497 // may not be loop invariant. 498 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 499 adr = adr->in(1); 500 continue; 501 } 502 } 503 504 // Intentional fallthrough! 505 506 // No obvious dominating point. The mem op is pinned below the Phi 507 // by the Phi itself. If the Phi goes away (no true value is merged) 508 // then the mem op can float, but not indefinitely. It must be pinned 509 // behind the controls leading to the Phi. 510 case Op_CheckCastPP: 511 // These usually stick around to change address type, however a 512 // useless one can be elided and we still need to pick up a control edge 513 if (adr->in(0) == NULL) { 514 // This CheckCastPP node has NO control and is likely useless. But we 515 // need check further up the ancestor chain for a control input to keep 516 // the node in place. 4959717. 517 skipped_cast = adr; 518 adr = adr->in(1); 519 continue; 520 } 521 ccp->hash_delete(this); 522 set_req(MemNode::Control, adr->in(0)); 523 ccp->hash_insert(this); 524 return this; 525 526 // List of "safe" opcodes; those that implicitly block the memory 527 // op below any null check. 528 case Op_CastX2P: // no null checks on native pointers 529 case Op_Parm: // 'this' pointer is not null 530 case Op_LoadP: // Loading from within a klass 531 case Op_LoadKlass: // Loading from within a klass 532 case Op_ConP: // Loading from a klass 533 case Op_CreateEx: // Sucking up the guts of an exception oop 534 case Op_Con: // Reading from TLS 535 case Op_CMoveP: // CMoveP is pinned 536 break; // No progress 537 538 case Op_Proj: // Direct call to an allocation routine 539 case Op_SCMemProj: // Memory state from store conditional ops 540 #ifdef ASSERT 541 { 542 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 543 const Node* call = adr->in(0); 544 if (call->is_CallStaticJava()) { 545 const CallStaticJavaNode* call_java = call->as_CallStaticJava(); 546 assert(call_java && call_java->method() == NULL, "must be runtime call"); 547 // We further presume that this is one of 548 // new_instance_Java, new_array_Java, or 549 // the like, but do not assert for this. 550 } else if (call->is_Allocate()) { 551 // similar case to new_instance_Java, etc. 552 } else if (!call->is_CallLeaf()) { 553 // Projections from fetch_oop (OSR) are allowed as well. 554 ShouldNotReachHere(); 555 } 556 } 557 #endif 558 break; 559 default: 560 ShouldNotReachHere(); 561 } 562 break; 563 } 564 } 565 566 return NULL; // No progress 567 } 568 569 570 //============================================================================= 571 uint LoadNode::size_of() const { return sizeof(*this); } 572 uint LoadNode::cmp( const Node &n ) const 573 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 574 const Type *LoadNode::bottom_type() const { return _type; } 575 uint LoadNode::ideal_reg() const { 576 return Matcher::base2reg[_type->base()]; 577 } 578 579 #ifndef PRODUCT 580 void LoadNode::dump_spec(outputStream *st) const { 581 MemNode::dump_spec(st); 582 if( !Verbose && !WizardMode ) { 583 // standard dump does this in Verbose and WizardMode 584 st->print(" #"); _type->dump_on(st); 585 } 586 } 587 #endif 588 589 590 //----------------------------LoadNode::make----------------------------------- 591 // Polymorphic factory method: 592 LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 593 // sanity check the alias category against the created node type 594 assert(!(adr_type->isa_oopptr() && 595 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 596 "use LoadKlassNode instead"); 597 assert(!(adr_type->isa_aryptr() && 598 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 599 "use LoadRangeNode instead"); 600 switch (bt) { 601 case T_BOOLEAN: 602 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() ); 603 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() ); 604 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() ); 605 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() ); 606 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() ); 607 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt ); 608 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt ); 609 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() ); 610 case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 611 } 612 ShouldNotReachHere(); 613 return (LoadNode*)NULL; 614 } 615 616 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 617 bool require_atomic = true; 618 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 619 } 620 621 622 623 624 //------------------------------hash------------------------------------------- 625 uint LoadNode::hash() const { 626 // unroll addition of interesting fields 627 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 628 } 629 630 //---------------------------can_see_stored_value------------------------------ 631 // This routine exists to make sure this set of tests is done the same 632 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 633 // will change the graph shape in a way which makes memory alive twice at the 634 // same time (uses the Oracle model of aliasing), then some 635 // LoadXNode::Identity will fold things back to the equivalence-class model 636 // of aliasing. 637 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 638 Node* ld_adr = in(MemNode::Address); 639 640 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 641 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 642 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 643 atp->field() != NULL && !atp->field()->is_volatile()) { 644 uint alias_idx = atp->index(); 645 bool final = atp->field()->is_final(); 646 Node* result = NULL; 647 Node* current = st; 648 // Skip through chains of MemBarNodes checking the MergeMems for 649 // new states for the slice of this load. Stop once any other 650 // kind of node is encountered. Loads from final memory can skip 651 // through any kind of MemBar but normal loads shouldn't skip 652 // through MemBarAcquire since the could allow them to move out of 653 // a synchronized region. 654 while (current->is_Proj()) { 655 int opc = current->in(0)->Opcode(); 656 if ((final && opc == Op_MemBarAcquire) || 657 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 658 Node* mem = current->in(0)->in(TypeFunc::Memory); 659 if (mem->is_MergeMem()) { 660 MergeMemNode* merge = mem->as_MergeMem(); 661 Node* new_st = merge->memory_at(alias_idx); 662 if (new_st == merge->base_memory()) { 663 // Keep searching 664 current = merge->base_memory(); 665 continue; 666 } 667 // Save the new memory state for the slice and fall through 668 // to exit. 669 result = new_st; 670 } 671 } 672 break; 673 } 674 if (result != NULL) { 675 st = result; 676 } 677 } 678 679 680 // Loop around twice in the case Load -> Initialize -> Store. 681 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 682 for (int trip = 0; trip <= 1; trip++) { 683 684 if (st->is_Store()) { 685 Node* st_adr = st->in(MemNode::Address); 686 if (!phase->eqv(st_adr, ld_adr)) { 687 // Try harder before giving up... Match raw and non-raw pointers. 688 intptr_t st_off = 0; 689 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 690 if (alloc == NULL) return NULL; 691 intptr_t ld_off = 0; 692 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 693 if (alloc != allo2) return NULL; 694 if (ld_off != st_off) return NULL; 695 // At this point we have proven something like this setup: 696 // A = Allocate(...) 697 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 698 // S = StoreQ(, AddP(, A.Parm , #Off), V) 699 // (Actually, we haven't yet proven the Q's are the same.) 700 // In other words, we are loading from a casted version of 701 // the same pointer-and-offset that we stored to. 702 // Thus, we are able to replace L by V. 703 } 704 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 705 if (store_Opcode() != st->Opcode()) 706 return NULL; 707 return st->in(MemNode::ValueIn); 708 } 709 710 intptr_t offset = 0; // scratch 711 712 // A load from a freshly-created object always returns zero. 713 // (This can happen after LoadNode::Ideal resets the load's memory input 714 // to find_captured_store, which returned InitializeNode::zero_memory.) 715 if (st->is_Proj() && st->in(0)->is_Allocate() && 716 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 717 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 718 // return a zero value for the load's basic type 719 // (This is one of the few places where a generic PhaseTransform 720 // can create new nodes. Think of it as lazily manifesting 721 // virtually pre-existing constants.) 722 return phase->zerocon(memory_type()); 723 } 724 725 // A load from an initialization barrier can match a captured store. 726 if (st->is_Proj() && st->in(0)->is_Initialize()) { 727 InitializeNode* init = st->in(0)->as_Initialize(); 728 AllocateNode* alloc = init->allocation(); 729 if (alloc != NULL && 730 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 731 // examine a captured store value 732 st = init->find_captured_store(offset, memory_size(), phase); 733 if (st != NULL) 734 continue; // take one more trip around 735 } 736 } 737 738 break; 739 } 740 741 return NULL; 742 } 743 744 //------------------------------Identity--------------------------------------- 745 // Loads are identity if previous store is to same address 746 Node *LoadNode::Identity( PhaseTransform *phase ) { 747 // If the previous store-maker is the right kind of Store, and the store is 748 // to the same address, then we are equal to the value stored. 749 Node* mem = in(MemNode::Memory); 750 Node* value = can_see_stored_value(mem, phase); 751 if( value ) { 752 // byte, short & char stores truncate naturally. 753 // A load has to load the truncated value which requires 754 // some sort of masking operation and that requires an 755 // Ideal call instead of an Identity call. 756 if (memory_size() < BytesPerInt) { 757 // If the input to the store does not fit with the load's result type, 758 // it must be truncated via an Ideal call. 759 if (!phase->type(value)->higher_equal(phase->type(this))) 760 return this; 761 } 762 // (This works even when value is a Con, but LoadNode::Value 763 // usually runs first, producing the singleton type of the Con.) 764 return value; 765 } 766 return this; 767 } 768 769 770 // Returns true if the AliasType refers to the field that holds the 771 // cached box array. Currently only handles the IntegerCache case. 772 static bool is_autobox_cache(Compile::AliasType* atp) { 773 if (atp != NULL && atp->field() != NULL) { 774 ciField* field = atp->field(); 775 ciSymbol* klass = field->holder()->name(); 776 if (field->name() == ciSymbol::cache_field_name() && 777 field->holder()->uses_default_loader() && 778 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 779 return true; 780 } 781 } 782 return false; 783 } 784 785 // Fetch the base value in the autobox array 786 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 787 if (atp != NULL && atp->field() != NULL) { 788 ciField* field = atp->field(); 789 ciSymbol* klass = field->holder()->name(); 790 if (field->name() == ciSymbol::cache_field_name() && 791 field->holder()->uses_default_loader() && 792 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 793 assert(field->is_constant(), "what?"); 794 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 795 // Fetch the box object at the base of the array and get its value 796 ciInstance* box = array->obj_at(0)->as_instance(); 797 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 798 if (ik->nof_nonstatic_fields() == 1) { 799 // This should be true nonstatic_field_at requires calling 800 // nof_nonstatic_fields so check it anyway 801 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 802 cache_offset = c.as_int(); 803 } 804 return true; 805 } 806 } 807 return false; 808 } 809 810 // Returns true if the AliasType refers to the value field of an 811 // autobox object. Currently only handles Integer. 812 static bool is_autobox_object(Compile::AliasType* atp) { 813 if (atp != NULL && atp->field() != NULL) { 814 ciField* field = atp->field(); 815 ciSymbol* klass = field->holder()->name(); 816 if (field->name() == ciSymbol::value_name() && 817 field->holder()->uses_default_loader() && 818 klass == ciSymbol::java_lang_Integer()) { 819 return true; 820 } 821 } 822 return false; 823 } 824 825 826 // We're loading from an object which has autobox behaviour. 827 // If this object is result of a valueOf call we'll have a phi 828 // merging a newly allocated object and a load from the cache. 829 // We want to replace this load with the original incoming 830 // argument to the valueOf call. 831 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 832 Node* base = in(Address)->in(AddPNode::Base); 833 if (base->is_Phi() && base->req() == 3) { 834 AllocateNode* allocation = NULL; 835 int allocation_index = -1; 836 int load_index = -1; 837 for (uint i = 1; i < base->req(); i++) { 838 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 839 if (allocation != NULL) { 840 allocation_index = i; 841 load_index = 3 - allocation_index; 842 break; 843 } 844 } 845 LoadNode* load = NULL; 846 if (allocation != NULL && base->in(load_index)->is_Load()) { 847 load = base->in(load_index)->as_Load(); 848 } 849 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 850 // Push the loads from the phi that comes from valueOf up 851 // through it to allow elimination of the loads and the recovery 852 // of the original value. 853 Node* mem_phi = in(Memory); 854 Node* offset = in(Address)->in(AddPNode::Offset); 855 856 Node* in1 = clone(); 857 Node* in1_addr = in1->in(Address)->clone(); 858 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 859 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 860 in1_addr->set_req(AddPNode::Offset, offset); 861 in1->set_req(0, base->in(allocation_index)); 862 in1->set_req(Address, in1_addr); 863 in1->set_req(Memory, mem_phi->in(allocation_index)); 864 865 Node* in2 = clone(); 866 Node* in2_addr = in2->in(Address)->clone(); 867 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 868 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 869 in2_addr->set_req(AddPNode::Offset, offset); 870 in2->set_req(0, base->in(load_index)); 871 in2->set_req(Address, in2_addr); 872 in2->set_req(Memory, mem_phi->in(load_index)); 873 874 in1_addr = phase->transform(in1_addr); 875 in1 = phase->transform(in1); 876 in2_addr = phase->transform(in2_addr); 877 in2 = phase->transform(in2); 878 879 PhiNode* result = PhiNode::make_blank(base->in(0), this); 880 result->set_req(allocation_index, in1); 881 result->set_req(load_index, in2); 882 return result; 883 } 884 } else if (base->is_Load()) { 885 // Eliminate the load of Integer.value for integers from the cache 886 // array by deriving the value from the index into the array. 887 // Capture the offset of the load and then reverse the computation. 888 Node* load_base = base->in(Address)->in(AddPNode::Base); 889 if (load_base != NULL) { 890 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 891 intptr_t cache_offset; 892 int shift = -1; 893 Node* cache = NULL; 894 if (is_autobox_cache(atp)) { 895 shift = exact_log2(type2aelembytes[T_OBJECT]); 896 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 897 } 898 if (cache != NULL && base->in(Address)->is_AddP()) { 899 Node* elements[4]; 900 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 901 int cache_low; 902 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 903 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 904 // Add up all the offsets making of the address of the load 905 Node* result = elements[0]; 906 for (int i = 1; i < count; i++) { 907 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 908 } 909 // Remove the constant offset from the address and then 910 // remove the scaling of the offset to recover the original index. 911 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 912 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 913 // Peel the shift off directly but wrap it in a dummy node 914 // since Ideal can't return existing nodes 915 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 916 } else { 917 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 918 } 919 #ifdef _LP64 920 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 921 #endif 922 return result; 923 } 924 } 925 } 926 } 927 return NULL; 928 } 929 930 931 //------------------------------Ideal------------------------------------------ 932 // If the load is from Field memory and the pointer is non-null, we can 933 // zero out the control input. 934 // If the offset is constant and the base is an object allocation, 935 // try to hook me up to the exact initializing store. 936 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 937 Node* p = MemNode::Ideal_common(phase, can_reshape); 938 if (p) return (p == NodeSentinel) ? NULL : p; 939 940 Node* ctrl = in(MemNode::Control); 941 Node* address = in(MemNode::Address); 942 943 // Skip up past a SafePoint control. Cannot do this for Stores because 944 // pointer stores & cardmarks must stay on the same side of a SafePoint. 945 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 946 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 947 ctrl = ctrl->in(0); 948 set_req(MemNode::Control,ctrl); 949 } 950 951 // Check for useless control edge in some common special cases 952 if (in(MemNode::Control) != NULL) { 953 intptr_t ignore = 0; 954 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 955 if (base != NULL 956 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 957 && detect_dominating_control(base->in(0), phase->C->start())) { 958 // A method-invariant, non-null address (constant or 'this' argument). 959 set_req(MemNode::Control, NULL); 960 } 961 } 962 963 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) { 964 Node* base = in(Address)->in(AddPNode::Base); 965 if (base != NULL) { 966 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 967 if (is_autobox_object(atp)) { 968 Node* result = eliminate_autobox(phase); 969 if (result != NULL) return result; 970 } 971 } 972 } 973 974 // Check for prior store with a different base or offset; make Load 975 // independent. Skip through any number of them. Bail out if the stores 976 // are in an endless dead cycle and report no progress. This is a key 977 // transform for Reflection. However, if after skipping through the Stores 978 // we can't then fold up against a prior store do NOT do the transform as 979 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 980 // array memory alive twice: once for the hoisted Load and again after the 981 // bypassed Store. This situation only works if EVERYBODY who does 982 // anti-dependence work knows how to bypass. I.e. we need all 983 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 984 // the alias index stuff. So instead, peek through Stores and IFF we can 985 // fold up, do so. 986 Node* prev_mem = find_previous_store(phase); 987 // Steps (a), (b): Walk past independent stores to find an exact match. 988 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 989 // (c) See if we can fold up on the spot, but don't fold up here. 990 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or 991 // just return a prior value, which is done by Identity calls. 992 if (can_see_stored_value(prev_mem, phase)) { 993 // Make ready for step (d): 994 set_req(MemNode::Memory, prev_mem); 995 return this; 996 } 997 } 998 999 return NULL; // No further progress 1000 } 1001 1002 // Helper to recognize certain Klass fields which are invariant across 1003 // some group of array types (e.g., int[] or all T[] where T < Object). 1004 const Type* 1005 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1006 ciKlass* klass) const { 1007 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1008 // The field is Klass::_modifier_flags. Return its (constant) value. 1009 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1010 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1011 return TypeInt::make(klass->modifier_flags()); 1012 } 1013 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1014 // The field is Klass::_access_flags. Return its (constant) value. 1015 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1016 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1017 return TypeInt::make(klass->access_flags()); 1018 } 1019 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1020 // The field is Klass::_layout_helper. Return its constant value if known. 1021 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1022 return TypeInt::make(klass->layout_helper()); 1023 } 1024 1025 // No match. 1026 return NULL; 1027 } 1028 1029 //------------------------------Value----------------------------------------- 1030 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1031 // Either input is TOP ==> the result is TOP 1032 Node* mem = in(MemNode::Memory); 1033 const Type *t1 = phase->type(mem); 1034 if (t1 == Type::TOP) return Type::TOP; 1035 Node* adr = in(MemNode::Address); 1036 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1037 if (tp == NULL || tp->empty()) return Type::TOP; 1038 int off = tp->offset(); 1039 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1040 1041 // Try to guess loaded type from pointer type 1042 if (tp->base() == Type::AryPtr) { 1043 const Type *t = tp->is_aryptr()->elem(); 1044 // Don't do this for integer types. There is only potential profit if 1045 // the element type t is lower than _type; that is, for int types, if _type is 1046 // more restrictive than t. This only happens here if one is short and the other 1047 // char (both 16 bits), and in those cases we've made an intentional decision 1048 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1049 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1050 // 1051 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1052 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1053 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1054 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1055 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1056 // In fact, that could have been the original type of p1, and p1 could have 1057 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1058 // expression (LShiftL quux 3) independently optimized to the constant 8. 1059 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1060 && Opcode() != Op_LoadKlass) { 1061 // t might actually be lower than _type, if _type is a unique 1062 // concrete subclass of abstract class t. 1063 // Make sure the reference is not into the header, by comparing 1064 // the offset against the offset of the start of the array's data. 1065 // Different array types begin at slightly different offsets (12 vs. 16). 1066 // We choose T_BYTE as an example base type that is least restrictive 1067 // as to alignment, which will therefore produce the smallest 1068 // possible base offset. 1069 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1070 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1071 const Type* jt = t->join(_type); 1072 // In any case, do not allow the join, per se, to empty out the type. 1073 if (jt->empty() && !t->empty()) { 1074 // This can happen if a interface-typed array narrows to a class type. 1075 jt = _type; 1076 } 1077 1078 if (EliminateAutoBox) { 1079 // The pointers in the autobox arrays are always non-null 1080 Node* base = in(Address)->in(AddPNode::Base); 1081 if (base != NULL) { 1082 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1083 if (is_autobox_cache(atp)) { 1084 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1085 } 1086 } 1087 } 1088 return jt; 1089 } 1090 } 1091 } else if (tp->base() == Type::InstPtr) { 1092 assert( off != Type::OffsetBot || 1093 // arrays can be cast to Objects 1094 tp->is_oopptr()->klass()->is_java_lang_Object() || 1095 // unsafe field access may not have a constant offset 1096 phase->C->has_unsafe_access(), 1097 "Field accesses must be precise" ); 1098 // For oop loads, we expect the _type to be precise 1099 } else if (tp->base() == Type::KlassPtr) { 1100 assert( off != Type::OffsetBot || 1101 // arrays can be cast to Objects 1102 tp->is_klassptr()->klass()->is_java_lang_Object() || 1103 // also allow array-loading from the primary supertype 1104 // array during subtype checks 1105 Opcode() == Op_LoadKlass, 1106 "Field accesses must be precise" ); 1107 // For klass/static loads, we expect the _type to be precise 1108 } 1109 1110 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1111 if (tkls != NULL && !StressReflectiveCode) { 1112 ciKlass* klass = tkls->klass(); 1113 if (klass->is_loaded() && tkls->klass_is_exact()) { 1114 // We are loading a field from a Klass metaobject whose identity 1115 // is known at compile time (the type is "exact" or "precise"). 1116 // Check for fields we know are maintained as constants by the VM. 1117 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1118 // The field is Klass::_super_check_offset. Return its (constant) value. 1119 // (Folds up type checking code.) 1120 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1121 return TypeInt::make(klass->super_check_offset()); 1122 } 1123 // Compute index into primary_supers array 1124 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1125 // Check for overflowing; use unsigned compare to handle the negative case. 1126 if( depth < ciKlass::primary_super_limit() ) { 1127 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1128 // (Folds up type checking code.) 1129 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1130 ciKlass *ss = klass->super_of_depth(depth); 1131 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1132 } 1133 const Type* aift = load_array_final_field(tkls, klass); 1134 if (aift != NULL) return aift; 1135 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1136 && klass->is_array_klass()) { 1137 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1138 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1139 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1140 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1141 } 1142 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1143 // The field is Klass::_java_mirror. Return its (constant) value. 1144 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1145 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1146 return TypeInstPtr::make(klass->java_mirror()); 1147 } 1148 } 1149 1150 // We can still check if we are loading from the primary_supers array at a 1151 // shallow enough depth. Even though the klass is not exact, entries less 1152 // than or equal to its super depth are correct. 1153 if (klass->is_loaded() ) { 1154 ciType *inner = klass->klass(); 1155 while( inner->is_obj_array_klass() ) 1156 inner = inner->as_obj_array_klass()->base_element_type(); 1157 if( inner->is_instance_klass() && 1158 !inner->as_instance_klass()->flags().is_interface() ) { 1159 // Compute index into primary_supers array 1160 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1161 // Check for overflowing; use unsigned compare to handle the negative case. 1162 if( depth < ciKlass::primary_super_limit() && 1163 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1164 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1165 // (Folds up type checking code.) 1166 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1167 ciKlass *ss = klass->super_of_depth(depth); 1168 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1169 } 1170 } 1171 } 1172 1173 // If the type is enough to determine that the thing is not an array, 1174 // we can give the layout_helper a positive interval type. 1175 // This will help short-circuit some reflective code. 1176 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1177 && !klass->is_array_klass() // not directly typed as an array 1178 && !klass->is_interface() // specifically not Serializable & Cloneable 1179 && !klass->is_java_lang_Object() // not the supertype of all T[] 1180 ) { 1181 // Note: When interfaces are reliable, we can narrow the interface 1182 // test to (klass != Serializable && klass != Cloneable). 1183 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1184 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1185 // The key property of this type is that it folds up tests 1186 // for array-ness, since it proves that the layout_helper is positive. 1187 // Thus, a generic value like the basic object layout helper works fine. 1188 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1189 } 1190 } 1191 1192 // If we are loading from a freshly-allocated object, produce a zero, 1193 // if the load is provably beyond the header of the object. 1194 // (Also allow a variable load from a fresh array to produce zero.) 1195 if (ReduceFieldZeroing) { 1196 Node* value = can_see_stored_value(mem,phase); 1197 if (value != NULL && value->is_Con()) 1198 return value->bottom_type(); 1199 } 1200 1201 return _type; 1202 } 1203 1204 //------------------------------match_edge------------------------------------- 1205 // Do we Match on this edge index or not? Match only the address. 1206 uint LoadNode::match_edge(uint idx) const { 1207 return idx == MemNode::Address; 1208 } 1209 1210 //--------------------------LoadBNode::Ideal-------------------------------------- 1211 // 1212 // If the previous store is to the same address as this load, 1213 // and the value stored was larger than a byte, replace this load 1214 // with the value stored truncated to a byte. If no truncation is 1215 // needed, the replacement is done in LoadNode::Identity(). 1216 // 1217 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1218 Node* mem = in(MemNode::Memory); 1219 Node* value = can_see_stored_value(mem,phase); 1220 if( value && !phase->type(value)->higher_equal( _type ) ) { 1221 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1222 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1223 } 1224 // Identity call will handle the case where truncation is not needed. 1225 return LoadNode::Ideal(phase, can_reshape); 1226 } 1227 1228 //--------------------------LoadCNode::Ideal-------------------------------------- 1229 // 1230 // If the previous store is to the same address as this load, 1231 // and the value stored was larger than a char, replace this load 1232 // with the value stored truncated to a char. If no truncation is 1233 // needed, the replacement is done in LoadNode::Identity(). 1234 // 1235 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1236 Node* mem = in(MemNode::Memory); 1237 Node* value = can_see_stored_value(mem,phase); 1238 if( value && !phase->type(value)->higher_equal( _type ) ) 1239 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1240 // Identity call will handle the case where truncation is not needed. 1241 return LoadNode::Ideal(phase, can_reshape); 1242 } 1243 1244 //--------------------------LoadSNode::Ideal-------------------------------------- 1245 // 1246 // If the previous store is to the same address as this load, 1247 // and the value stored was larger than a short, replace this load 1248 // with the value stored truncated to a short. If no truncation is 1249 // needed, the replacement is done in LoadNode::Identity(). 1250 // 1251 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1252 Node* mem = in(MemNode::Memory); 1253 Node* value = can_see_stored_value(mem,phase); 1254 if( value && !phase->type(value)->higher_equal( _type ) ) { 1255 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1256 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1257 } 1258 // Identity call will handle the case where truncation is not needed. 1259 return LoadNode::Ideal(phase, can_reshape); 1260 } 1261 1262 //============================================================================= 1263 //------------------------------Value------------------------------------------ 1264 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1265 // Either input is TOP ==> the result is TOP 1266 const Type *t1 = phase->type( in(MemNode::Memory) ); 1267 if (t1 == Type::TOP) return Type::TOP; 1268 Node *adr = in(MemNode::Address); 1269 const Type *t2 = phase->type( adr ); 1270 if (t2 == Type::TOP) return Type::TOP; 1271 const TypePtr *tp = t2->is_ptr(); 1272 if (TypePtr::above_centerline(tp->ptr()) || 1273 tp->ptr() == TypePtr::Null) return Type::TOP; 1274 1275 // Return a more precise klass, if possible 1276 const TypeInstPtr *tinst = tp->isa_instptr(); 1277 if (tinst != NULL) { 1278 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1279 int offset = tinst->offset(); 1280 if (ik == phase->C->env()->Class_klass() 1281 && (offset == java_lang_Class::klass_offset_in_bytes() || 1282 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1283 // We are loading a special hidden field from a Class mirror object, 1284 // the field which points to the VM's Klass metaobject. 1285 ciType* t = tinst->java_mirror_type(); 1286 // java_mirror_type returns non-null for compile-time Class constants. 1287 if (t != NULL) { 1288 // constant oop => constant klass 1289 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1290 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1291 } 1292 if (!t->is_klass()) { 1293 // a primitive Class (e.g., int.class) has NULL for a klass field 1294 return TypePtr::NULL_PTR; 1295 } 1296 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1297 return TypeKlassPtr::make(t->as_klass()); 1298 } 1299 // non-constant mirror, so we can't tell what's going on 1300 } 1301 if( !ik->is_loaded() ) 1302 return _type; // Bail out if not loaded 1303 if (offset == oopDesc::klass_offset_in_bytes()) { 1304 if (tinst->klass_is_exact()) { 1305 return TypeKlassPtr::make(ik); 1306 } 1307 // See if we can become precise: no subklasses and no interface 1308 // (Note: We need to support verified interfaces.) 1309 if (!ik->is_interface() && !ik->has_subklass()) { 1310 //assert(!UseExactTypes, "this code should be useless with exact types"); 1311 // Add a dependence; if any subclass added we need to recompile 1312 if (!ik->is_final()) { 1313 // %%% should use stronger assert_unique_concrete_subtype instead 1314 phase->C->dependencies()->assert_leaf_type(ik); 1315 } 1316 // Return precise klass 1317 return TypeKlassPtr::make(ik); 1318 } 1319 1320 // Return root of possible klass 1321 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1322 } 1323 } 1324 1325 // Check for loading klass from an array 1326 const TypeAryPtr *tary = tp->isa_aryptr(); 1327 if( tary != NULL ) { 1328 ciKlass *tary_klass = tary->klass(); 1329 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1330 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1331 if (tary->klass_is_exact()) { 1332 return TypeKlassPtr::make(tary_klass); 1333 } 1334 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1335 // If the klass is an object array, we defer the question to the 1336 // array component klass. 1337 if( ak->is_obj_array_klass() ) { 1338 assert( ak->is_loaded(), "" ); 1339 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1340 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1341 ciInstanceKlass* ik = base_k->as_instance_klass(); 1342 // See if we can become precise: no subklasses and no interface 1343 if (!ik->is_interface() && !ik->has_subklass()) { 1344 //assert(!UseExactTypes, "this code should be useless with exact types"); 1345 // Add a dependence; if any subclass added we need to recompile 1346 if (!ik->is_final()) { 1347 phase->C->dependencies()->assert_leaf_type(ik); 1348 } 1349 // Return precise array klass 1350 return TypeKlassPtr::make(ak); 1351 } 1352 } 1353 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1354 } else { // Found a type-array? 1355 //assert(!UseExactTypes, "this code should be useless with exact types"); 1356 assert( ak->is_type_array_klass(), "" ); 1357 return TypeKlassPtr::make(ak); // These are always precise 1358 } 1359 } 1360 } 1361 1362 // Check for loading klass from an array klass 1363 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1364 if (tkls != NULL && !StressReflectiveCode) { 1365 ciKlass* klass = tkls->klass(); 1366 if( !klass->is_loaded() ) 1367 return _type; // Bail out if not loaded 1368 if( klass->is_obj_array_klass() && 1369 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1370 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1371 // // Always returning precise element type is incorrect, 1372 // // e.g., element type could be object and array may contain strings 1373 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1374 1375 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1376 // according to the element type's subclassing. 1377 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1378 } 1379 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1380 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1381 ciKlass* sup = klass->as_instance_klass()->super(); 1382 // The field is Klass::_super. Return its (constant) value. 1383 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1384 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1385 } 1386 } 1387 1388 // Bailout case 1389 return LoadNode::Value(phase); 1390 } 1391 1392 //------------------------------Identity--------------------------------------- 1393 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1394 // Also feed through the klass in Allocate(...klass...)._klass. 1395 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1396 Node* x = LoadNode::Identity(phase); 1397 if (x != this) return x; 1398 1399 // Take apart the address into an oop and and offset. 1400 // Return 'this' if we cannot. 1401 Node* adr = in(MemNode::Address); 1402 intptr_t offset = 0; 1403 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1404 if (base == NULL) return this; 1405 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1406 if (toop == NULL) return this; 1407 1408 // We can fetch the klass directly through an AllocateNode. 1409 // This works even if the klass is not constant (clone or newArray). 1410 if (offset == oopDesc::klass_offset_in_bytes()) { 1411 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1412 if (allocated_klass != NULL) { 1413 return allocated_klass; 1414 } 1415 } 1416 1417 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1418 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1419 // See inline_native_Class_query for occurrences of these patterns. 1420 // Java Example: x.getClass().isAssignableFrom(y) 1421 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1422 // 1423 // This improves reflective code, often making the Class 1424 // mirror go completely dead. (Current exception: Class 1425 // mirrors may appear in debug info, but we could clean them out by 1426 // introducing a new debug info operator for klassOop.java_mirror). 1427 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1428 && (offset == java_lang_Class::klass_offset_in_bytes() || 1429 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1430 // We are loading a special hidden field from a Class mirror, 1431 // the field which points to its Klass or arrayKlass metaobject. 1432 if (base->is_Load()) { 1433 Node* adr2 = base->in(MemNode::Address); 1434 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1435 if (tkls != NULL && !tkls->empty() 1436 && (tkls->klass()->is_instance_klass() || 1437 tkls->klass()->is_array_klass()) 1438 && adr2->is_AddP() 1439 ) { 1440 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1441 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1442 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1443 } 1444 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1445 return adr2->in(AddPNode::Base); 1446 } 1447 } 1448 } 1449 } 1450 1451 return this; 1452 } 1453 1454 //------------------------------Value----------------------------------------- 1455 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1456 // Either input is TOP ==> the result is TOP 1457 const Type *t1 = phase->type( in(MemNode::Memory) ); 1458 if( t1 == Type::TOP ) return Type::TOP; 1459 Node *adr = in(MemNode::Address); 1460 const Type *t2 = phase->type( adr ); 1461 if( t2 == Type::TOP ) return Type::TOP; 1462 const TypePtr *tp = t2->is_ptr(); 1463 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 1464 const TypeAryPtr *tap = tp->isa_aryptr(); 1465 if( !tap ) return _type; 1466 return tap->size(); 1467 } 1468 1469 //------------------------------Identity--------------------------------------- 1470 // Feed through the length in AllocateArray(...length...)._length. 1471 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 1472 Node* x = LoadINode::Identity(phase); 1473 if (x != this) return x; 1474 1475 // Take apart the address into an oop and and offset. 1476 // Return 'this' if we cannot. 1477 Node* adr = in(MemNode::Address); 1478 intptr_t offset = 0; 1479 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1480 if (base == NULL) return this; 1481 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1482 if (tary == NULL) return this; 1483 1484 // We can fetch the length directly through an AllocateArrayNode. 1485 // This works even if the length is not constant (clone or newArray). 1486 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1487 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase); 1488 if (allocated_length != NULL) { 1489 return allocated_length; 1490 } 1491 } 1492 1493 return this; 1494 1495 } 1496 //============================================================================= 1497 //---------------------------StoreNode::make----------------------------------- 1498 // Polymorphic factory method: 1499 StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 1500 switch (bt) { 1501 case T_BOOLEAN: 1502 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 1503 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 1504 case T_CHAR: 1505 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 1506 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 1507 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 1508 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 1509 case T_ADDRESS: 1510 case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 1511 } 1512 ShouldNotReachHere(); 1513 return (StoreNode*)NULL; 1514 } 1515 1516 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 1517 bool require_atomic = true; 1518 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 1519 } 1520 1521 1522 //--------------------------bottom_type---------------------------------------- 1523 const Type *StoreNode::bottom_type() const { 1524 return Type::MEMORY; 1525 } 1526 1527 //------------------------------hash------------------------------------------- 1528 uint StoreNode::hash() const { 1529 // unroll addition of interesting fields 1530 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 1531 1532 // Since they are not commoned, do not hash them: 1533 return NO_HASH; 1534 } 1535 1536 //------------------------------Ideal------------------------------------------ 1537 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 1538 // When a store immediately follows a relevant allocation/initialization, 1539 // try to capture it into the initialization, or hoist it above. 1540 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1541 Node* p = MemNode::Ideal_common(phase, can_reshape); 1542 if (p) return (p == NodeSentinel) ? NULL : p; 1543 1544 Node* mem = in(MemNode::Memory); 1545 Node* address = in(MemNode::Address); 1546 1547 // Back-to-back stores to same address? Fold em up. 1548 // Generally unsafe if I have intervening uses... 1549 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 1550 // Looking at a dead closed cycle of memory? 1551 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 1552 1553 assert(Opcode() == mem->Opcode() || 1554 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 1555 "no mismatched stores, except on raw memory"); 1556 1557 if (mem->outcnt() == 1 && // check for intervening uses 1558 mem->as_Store()->memory_size() <= this->memory_size()) { 1559 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 1560 // For example, 'mem' might be the final state at a conditional return. 1561 // Or, 'mem' might be used by some node which is live at the same time 1562 // 'this' is live, which might be unschedulable. So, require exactly 1563 // ONE user, the 'this' store, until such time as we clone 'mem' for 1564 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 1565 if (can_reshape) { // (%%% is this an anachronism?) 1566 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 1567 phase->is_IterGVN()); 1568 } else { 1569 // It's OK to do this in the parser, since DU info is always accurate, 1570 // and the parser always refers to nodes via SafePointNode maps. 1571 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 1572 } 1573 return this; 1574 } 1575 } 1576 1577 // Capture an unaliased, unconditional, simple store into an initializer. 1578 // Or, if it is independent of the allocation, hoist it above the allocation. 1579 if (ReduceFieldZeroing && /*can_reshape &&*/ 1580 mem->is_Proj() && mem->in(0)->is_Initialize()) { 1581 InitializeNode* init = mem->in(0)->as_Initialize(); 1582 intptr_t offset = init->can_capture_store(this, phase); 1583 if (offset > 0) { 1584 Node* moved = init->capture_store(this, offset, phase); 1585 // If the InitializeNode captured me, it made a raw copy of me, 1586 // and I need to disappear. 1587 if (moved != NULL) { 1588 // %%% hack to ensure that Ideal returns a new node: 1589 mem = MergeMemNode::make(phase->C, mem); 1590 return mem; // fold me away 1591 } 1592 } 1593 } 1594 1595 return NULL; // No further progress 1596 } 1597 1598 //------------------------------Value----------------------------------------- 1599 const Type *StoreNode::Value( PhaseTransform *phase ) const { 1600 // Either input is TOP ==> the result is TOP 1601 const Type *t1 = phase->type( in(MemNode::Memory) ); 1602 if( t1 == Type::TOP ) return Type::TOP; 1603 const Type *t2 = phase->type( in(MemNode::Address) ); 1604 if( t2 == Type::TOP ) return Type::TOP; 1605 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 1606 if( t3 == Type::TOP ) return Type::TOP; 1607 return Type::MEMORY; 1608 } 1609 1610 //------------------------------Identity--------------------------------------- 1611 // Remove redundant stores: 1612 // Store(m, p, Load(m, p)) changes to m. 1613 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 1614 Node *StoreNode::Identity( PhaseTransform *phase ) { 1615 Node* mem = in(MemNode::Memory); 1616 Node* adr = in(MemNode::Address); 1617 Node* val = in(MemNode::ValueIn); 1618 1619 // Load then Store? Then the Store is useless 1620 if (val->is_Load() && 1621 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 1622 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 1623 val->as_Load()->store_Opcode() == Opcode()) { 1624 return mem; 1625 } 1626 1627 // Two stores in a row of the same value? 1628 if (mem->is_Store() && 1629 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 1630 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 1631 mem->Opcode() == Opcode()) { 1632 return mem; 1633 } 1634 1635 // Store of zero anywhere into a freshly-allocated object? 1636 // Then the store is useless. 1637 // (It must already have been captured by the InitializeNode.) 1638 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 1639 // a newly allocated object is already all-zeroes everywhere 1640 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 1641 return mem; 1642 } 1643 1644 // the store may also apply to zero-bits in an earlier object 1645 Node* prev_mem = find_previous_store(phase); 1646 // Steps (a), (b): Walk past independent stores to find an exact match. 1647 if (prev_mem != NULL) { 1648 Node* prev_val = can_see_stored_value(prev_mem, phase); 1649 if (prev_val != NULL && phase->eqv(prev_val, val)) { 1650 // prev_val and val might differ by a cast; it would be good 1651 // to keep the more informative of the two. 1652 return mem; 1653 } 1654 } 1655 } 1656 1657 return this; 1658 } 1659 1660 //------------------------------match_edge------------------------------------- 1661 // Do we Match on this edge index or not? Match only memory & value 1662 uint StoreNode::match_edge(uint idx) const { 1663 return idx == MemNode::Address || idx == MemNode::ValueIn; 1664 } 1665 1666 //------------------------------cmp-------------------------------------------- 1667 // Do not common stores up together. They generally have to be split 1668 // back up anyways, so do not bother. 1669 uint StoreNode::cmp( const Node &n ) const { 1670 return (&n == this); // Always fail except on self 1671 } 1672 1673 //------------------------------Ideal_masked_input----------------------------- 1674 // Check for a useless mask before a partial-word store 1675 // (StoreB ... (AndI valIn conIa) ) 1676 // If (conIa & mask == mask) this simplifies to 1677 // (StoreB ... (valIn) ) 1678 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 1679 Node *val = in(MemNode::ValueIn); 1680 if( val->Opcode() == Op_AndI ) { 1681 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 1682 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 1683 set_req(MemNode::ValueIn, val->in(1)); 1684 return this; 1685 } 1686 } 1687 return NULL; 1688 } 1689 1690 1691 //------------------------------Ideal_sign_extended_input---------------------- 1692 // Check for useless sign-extension before a partial-word store 1693 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 1694 // If (conIL == conIR && conIR <= num_bits) this simplifies to 1695 // (StoreB ... (valIn) ) 1696 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 1697 Node *val = in(MemNode::ValueIn); 1698 if( val->Opcode() == Op_RShiftI ) { 1699 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 1700 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 1701 Node *shl = val->in(1); 1702 if( shl->Opcode() == Op_LShiftI ) { 1703 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 1704 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 1705 set_req(MemNode::ValueIn, shl->in(1)); 1706 return this; 1707 } 1708 } 1709 } 1710 } 1711 return NULL; 1712 } 1713 1714 //------------------------------value_never_loaded----------------------------------- 1715 // Determine whether there are any possible loads of the value stored. 1716 // For simplicity, we actually check if there are any loads from the 1717 // address stored to, not just for loads of the value stored by this node. 1718 // 1719 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 1720 Node *adr = in(Address); 1721 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 1722 if (adr_oop == NULL) 1723 return false; 1724 if (!adr_oop->is_instance()) 1725 return false; // if not a distinct instance, there may be aliases of the address 1726 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 1727 Node *use = adr->fast_out(i); 1728 int opc = use->Opcode(); 1729 if (use->is_Load() || use->is_LoadStore()) { 1730 return false; 1731 } 1732 } 1733 return true; 1734 } 1735 1736 //============================================================================= 1737 //------------------------------Ideal------------------------------------------ 1738 // If the store is from an AND mask that leaves the low bits untouched, then 1739 // we can skip the AND operation. If the store is from a sign-extension 1740 // (a left shift, then right shift) we can skip both. 1741 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1742 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 1743 if( progress != NULL ) return progress; 1744 1745 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 1746 if( progress != NULL ) return progress; 1747 1748 // Finally check the default case 1749 return StoreNode::Ideal(phase, can_reshape); 1750 } 1751 1752 //============================================================================= 1753 //------------------------------Ideal------------------------------------------ 1754 // If the store is from an AND mask that leaves the low bits untouched, then 1755 // we can skip the AND operation 1756 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1757 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 1758 if( progress != NULL ) return progress; 1759 1760 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 1761 if( progress != NULL ) return progress; 1762 1763 // Finally check the default case 1764 return StoreNode::Ideal(phase, can_reshape); 1765 } 1766 1767 //============================================================================= 1768 //------------------------------Identity--------------------------------------- 1769 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 1770 // No need to card mark when storing a null ptr 1771 Node* my_store = in(MemNode::OopStore); 1772 if (my_store->is_Store()) { 1773 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 1774 if( t1 == TypePtr::NULL_PTR ) { 1775 return in(MemNode::Memory); 1776 } 1777 } 1778 return this; 1779 } 1780 1781 //------------------------------Value----------------------------------------- 1782 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 1783 // Either input is TOP ==> the result is TOP 1784 const Type *t = phase->type( in(MemNode::Memory) ); 1785 if( t == Type::TOP ) return Type::TOP; 1786 t = phase->type( in(MemNode::Address) ); 1787 if( t == Type::TOP ) return Type::TOP; 1788 t = phase->type( in(MemNode::ValueIn) ); 1789 if( t == Type::TOP ) return Type::TOP; 1790 // If extra input is TOP ==> the result is TOP 1791 t = phase->type( in(MemNode::OopStore) ); 1792 if( t == Type::TOP ) return Type::TOP; 1793 1794 return StoreNode::Value( phase ); 1795 } 1796 1797 1798 //============================================================================= 1799 //----------------------------------SCMemProjNode------------------------------ 1800 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 1801 { 1802 return bottom_type(); 1803 } 1804 1805 //============================================================================= 1806 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 1807 init_req(MemNode::Control, c ); 1808 init_req(MemNode::Memory , mem); 1809 init_req(MemNode::Address, adr); 1810 init_req(MemNode::ValueIn, val); 1811 init_req( ExpectedIn, ex ); 1812 init_class_id(Class_LoadStore); 1813 1814 } 1815 1816 //============================================================================= 1817 //-------------------------------adr_type-------------------------------------- 1818 // Do we Match on this edge index or not? Do not match memory 1819 const TypePtr* ClearArrayNode::adr_type() const { 1820 Node *adr = in(3); 1821 return MemNode::calculate_adr_type(adr->bottom_type()); 1822 } 1823 1824 //------------------------------match_edge------------------------------------- 1825 // Do we Match on this edge index or not? Do not match memory 1826 uint ClearArrayNode::match_edge(uint idx) const { 1827 return idx > 1; 1828 } 1829 1830 //------------------------------Identity--------------------------------------- 1831 // Clearing a zero length array does nothing 1832 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 1833 return phase->type(in(2))->higher_equal(TypeInt::ZERO) ? in(1) : this; 1834 } 1835 1836 //------------------------------Idealize--------------------------------------- 1837 // Clearing a short array is faster with stores 1838 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1839 const int unit = BytesPerLong; 1840 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 1841 if (!t) return NULL; 1842 if (!t->is_con()) return NULL; 1843 intptr_t raw_count = t->get_con(); 1844 intptr_t size = raw_count; 1845 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 1846 // Clearing nothing uses the Identity call. 1847 // Negative clears are possible on dead ClearArrays 1848 // (see jck test stmt114.stmt11402.val). 1849 if (size <= 0 || size % unit != 0) return NULL; 1850 intptr_t count = size / unit; 1851 // Length too long; use fast hardware clear 1852 if (size > Matcher::init_array_short_size) return NULL; 1853 Node *mem = in(1); 1854 if( phase->type(mem)==Type::TOP ) return NULL; 1855 Node *adr = in(3); 1856 const Type* at = phase->type(adr); 1857 if( at==Type::TOP ) return NULL; 1858 const TypePtr* atp = at->isa_ptr(); 1859 // adjust atp to be the correct array element address type 1860 if (atp == NULL) atp = TypePtr::BOTTOM; 1861 else atp = atp->add_offset(Type::OffsetBot); 1862 // Get base for derived pointer purposes 1863 if( adr->Opcode() != Op_AddP ) Unimplemented(); 1864 Node *base = adr->in(1); 1865 1866 Node *zero = phase->makecon(TypeLong::ZERO); 1867 Node *off = phase->MakeConX(BytesPerLong); 1868 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 1869 count--; 1870 while( count-- ) { 1871 mem = phase->transform(mem); 1872 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 1873 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 1874 } 1875 return mem; 1876 } 1877 1878 //----------------------------clear_memory------------------------------------- 1879 // Generate code to initialize object storage to zero. 1880 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 1881 intptr_t start_offset, 1882 Node* end_offset, 1883 PhaseGVN* phase) { 1884 Compile* C = phase->C; 1885 intptr_t offset = start_offset; 1886 1887 int unit = BytesPerLong; 1888 if ((offset % unit) != 0) { 1889 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 1890 adr = phase->transform(adr); 1891 const TypePtr* atp = TypeRawPtr::BOTTOM; 1892 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 1893 mem = phase->transform(mem); 1894 offset += BytesPerInt; 1895 } 1896 assert((offset % unit) == 0, ""); 1897 1898 // Initialize the remaining stuff, if any, with a ClearArray. 1899 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 1900 } 1901 1902 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 1903 Node* start_offset, 1904 Node* end_offset, 1905 PhaseGVN* phase) { 1906 Compile* C = phase->C; 1907 int unit = BytesPerLong; 1908 Node* zbase = start_offset; 1909 Node* zend = end_offset; 1910 1911 // Scale to the unit required by the CPU: 1912 if (!Matcher::init_array_count_is_in_bytes) { 1913 Node* shift = phase->intcon(exact_log2(unit)); 1914 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 1915 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 1916 } 1917 1918 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 1919 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 1920 1921 // Bulk clear double-words 1922 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 1923 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 1924 return phase->transform(mem); 1925 } 1926 1927 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 1928 intptr_t start_offset, 1929 intptr_t end_offset, 1930 PhaseGVN* phase) { 1931 Compile* C = phase->C; 1932 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 1933 intptr_t done_offset = end_offset; 1934 if ((done_offset % BytesPerLong) != 0) { 1935 done_offset -= BytesPerInt; 1936 } 1937 if (done_offset > start_offset) { 1938 mem = clear_memory(ctl, mem, dest, 1939 start_offset, phase->MakeConX(done_offset), phase); 1940 } 1941 if (done_offset < end_offset) { // emit the final 32-bit store 1942 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 1943 adr = phase->transform(adr); 1944 const TypePtr* atp = TypeRawPtr::BOTTOM; 1945 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 1946 mem = phase->transform(mem); 1947 done_offset += BytesPerInt; 1948 } 1949 assert(done_offset == end_offset, ""); 1950 return mem; 1951 } 1952 1953 //============================================================================= 1954 // Do we match on this edge? No memory edges 1955 uint StrCompNode::match_edge(uint idx) const { 1956 return idx == 5 || idx == 6; 1957 } 1958 1959 //------------------------------Ideal------------------------------------------ 1960 // Return a node which is more "ideal" than the current node. Strip out 1961 // control copies 1962 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1963 return remove_dead_region(phase, can_reshape) ? this : NULL; 1964 } 1965 1966 1967 //============================================================================= 1968 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 1969 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 1970 _adr_type(C->get_adr_type(alias_idx)) 1971 { 1972 init_class_id(Class_MemBar); 1973 Node* top = C->top(); 1974 init_req(TypeFunc::I_O,top); 1975 init_req(TypeFunc::FramePtr,top); 1976 init_req(TypeFunc::ReturnAdr,top); 1977 if (precedent != NULL) 1978 init_req(TypeFunc::Parms, precedent); 1979 } 1980 1981 //------------------------------cmp-------------------------------------------- 1982 uint MemBarNode::hash() const { return NO_HASH; } 1983 uint MemBarNode::cmp( const Node &n ) const { 1984 return (&n == this); // Always fail except on self 1985 } 1986 1987 //------------------------------make------------------------------------------- 1988 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 1989 int len = Precedent + (pn == NULL? 0: 1); 1990 switch (opcode) { 1991 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 1992 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 1993 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 1994 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 1995 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 1996 default: ShouldNotReachHere(); return NULL; 1997 } 1998 } 1999 2000 //------------------------------Ideal------------------------------------------ 2001 // Return a node which is more "ideal" than the current node. Strip out 2002 // control copies 2003 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2004 if (remove_dead_region(phase, can_reshape)) return this; 2005 return NULL; 2006 } 2007 2008 //------------------------------Value------------------------------------------ 2009 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2010 if( !in(0) ) return Type::TOP; 2011 if( phase->type(in(0)) == Type::TOP ) 2012 return Type::TOP; 2013 return TypeTuple::MEMBAR; 2014 } 2015 2016 //------------------------------match------------------------------------------ 2017 // Construct projections for memory. 2018 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2019 switch (proj->_con) { 2020 case TypeFunc::Control: 2021 case TypeFunc::Memory: 2022 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2023 } 2024 ShouldNotReachHere(); 2025 return NULL; 2026 } 2027 2028 //===========================InitializeNode==================================== 2029 // SUMMARY: 2030 // This node acts as a memory barrier on raw memory, after some raw stores. 2031 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2032 // The Initialize can 'capture' suitably constrained stores as raw inits. 2033 // It can coalesce related raw stores into larger units (called 'tiles'). 2034 // It can avoid zeroing new storage for memory units which have raw inits. 2035 // At macro-expansion, it is marked 'complete', and does not optimize further. 2036 // 2037 // EXAMPLE: 2038 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2039 // ctl = incoming control; mem* = incoming memory 2040 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2041 // First allocate uninitialized memory and fill in the header: 2042 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2043 // ctl := alloc.Control; mem* := alloc.Memory* 2044 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2045 // Then initialize to zero the non-header parts of the raw memory block: 2046 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2047 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2048 // After the initialize node executes, the object is ready for service: 2049 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2050 // Suppose its body is immediately initialized as {1,2}: 2051 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2052 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2053 // mem.SLICE(#short[*]) := store2 2054 // 2055 // DETAILS: 2056 // An InitializeNode collects and isolates object initialization after 2057 // an AllocateNode and before the next possible safepoint. As a 2058 // memory barrier (MemBarNode), it keeps critical stores from drifting 2059 // down past any safepoint or any publication of the allocation. 2060 // Before this barrier, a newly-allocated object may have uninitialized bits. 2061 // After this barrier, it may be treated as a real oop, and GC is allowed. 2062 // 2063 // The semantics of the InitializeNode include an implicit zeroing of 2064 // the new object from object header to the end of the object. 2065 // (The object header and end are determined by the AllocateNode.) 2066 // 2067 // Certain stores may be added as direct inputs to the InitializeNode. 2068 // These stores must update raw memory, and they must be to addresses 2069 // derived from the raw address produced by AllocateNode, and with 2070 // a constant offset. They must be ordered by increasing offset. 2071 // The first one is at in(RawStores), the last at in(req()-1). 2072 // Unlike most memory operations, they are not linked in a chain, 2073 // but are displayed in parallel as users of the rawmem output of 2074 // the allocation. 2075 // 2076 // (See comments in InitializeNode::capture_store, which continue 2077 // the example given above.) 2078 // 2079 // When the associated Allocate is macro-expanded, the InitializeNode 2080 // may be rewritten to optimize collected stores. A ClearArrayNode 2081 // may also be created at that point to represent any required zeroing. 2082 // The InitializeNode is then marked 'complete', prohibiting further 2083 // capturing of nearby memory operations. 2084 // 2085 // During macro-expansion, all captured initializations which store 2086 // constant values of 32 bits or smaller are coalesced (if advantagous) 2087 // into larger 'tiles' 32 or 64 bits. This allows an object to be 2088 // initialized in fewer memory operations. Memory words which are 2089 // covered by neither tiles nor non-constant stores are pre-zeroed 2090 // by explicit stores of zero. (The code shape happens to do all 2091 // zeroing first, then all other stores, with both sequences occurring 2092 // in order of ascending offsets.) 2093 // 2094 // Alternatively, code may be inserted between an AllocateNode and its 2095 // InitializeNode, to perform arbitrary initialization of the new object. 2096 // E.g., the object copying intrinsics insert complex data transfers here. 2097 // The initialization must then be marked as 'complete' disable the 2098 // built-in zeroing semantics and the collection of initializing stores. 2099 // 2100 // While an InitializeNode is incomplete, reads from the memory state 2101 // produced by it are optimizable if they match the control edge and 2102 // new oop address associated with the allocation/initialization. 2103 // They return a stored value (if the offset matches) or else zero. 2104 // A write to the memory state, if it matches control and address, 2105 // and if it is to a constant offset, may be 'captured' by the 2106 // InitializeNode. It is cloned as a raw memory operation and rewired 2107 // inside the initialization, to the raw oop produced by the allocation. 2108 // Operations on addresses which are provably distinct (e.g., to 2109 // other AllocateNodes) are allowed to bypass the initialization. 2110 // 2111 // The effect of all this is to consolidate object initialization 2112 // (both arrays and non-arrays, both piecewise and bulk) into a 2113 // single location, where it can be optimized as a unit. 2114 // 2115 // Only stores with an offset less than TrackedInitializationLimit words 2116 // will be considered for capture by an InitializeNode. This puts a 2117 // reasonable limit on the complexity of optimized initializations. 2118 2119 //---------------------------InitializeNode------------------------------------ 2120 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2121 : _is_complete(false), 2122 MemBarNode(C, adr_type, rawoop) 2123 { 2124 init_class_id(Class_Initialize); 2125 2126 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2127 assert(in(RawAddress) == rawoop, "proper init"); 2128 // Note: allocation() can be NULL, for secondary initialization barriers 2129 } 2130 2131 // Since this node is not matched, it will be processed by the 2132 // register allocator. Declare that there are no constraints 2133 // on the allocation of the RawAddress edge. 2134 const RegMask &InitializeNode::in_RegMask(uint idx) const { 2135 // This edge should be set to top, by the set_complete. But be conservative. 2136 if (idx == InitializeNode::RawAddress) 2137 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2138 return RegMask::Empty; 2139 } 2140 2141 Node* InitializeNode::memory(uint alias_idx) { 2142 Node* mem = in(Memory); 2143 if (mem->is_MergeMem()) { 2144 return mem->as_MergeMem()->memory_at(alias_idx); 2145 } else { 2146 // incoming raw memory is not split 2147 return mem; 2148 } 2149 } 2150 2151 bool InitializeNode::is_non_zero() { 2152 if (is_complete()) return false; 2153 remove_extra_zeroes(); 2154 return (req() > RawStores); 2155 } 2156 2157 void InitializeNode::set_complete(PhaseGVN* phase) { 2158 assert(!is_complete(), "caller responsibility"); 2159 _is_complete = true; 2160 2161 // After this node is complete, it contains a bunch of 2162 // raw-memory initializations. There is no need for 2163 // it to have anything to do with non-raw memory effects. 2164 // Therefore, tell all non-raw users to re-optimize themselves, 2165 // after skipping the memory effects of this initialization. 2166 PhaseIterGVN* igvn = phase->is_IterGVN(); 2167 if (igvn) igvn->add_users_to_worklist(this); 2168 } 2169 2170 // convenience function 2171 // return false if the init contains any stores already 2172 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2173 InitializeNode* init = initialization(); 2174 if (init == NULL || init->is_complete()) return false; 2175 init->remove_extra_zeroes(); 2176 // for now, if this allocation has already collected any inits, bail: 2177 if (init->is_non_zero()) return false; 2178 init->set_complete(phase); 2179 return true; 2180 } 2181 2182 void InitializeNode::remove_extra_zeroes() { 2183 if (req() == RawStores) return; 2184 Node* zmem = zero_memory(); 2185 uint fill = RawStores; 2186 for (uint i = fill; i < req(); i++) { 2187 Node* n = in(i); 2188 if (n->is_top() || n == zmem) continue; // skip 2189 if (fill < i) set_req(fill, n); // compact 2190 ++fill; 2191 } 2192 // delete any empty spaces created: 2193 while (fill < req()) { 2194 del_req(fill); 2195 } 2196 } 2197 2198 // Helper for remembering which stores go with which offsets. 2199 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2200 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2201 intptr_t offset = -1; 2202 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2203 phase, offset); 2204 if (base == NULL) return -1; // something is dead, 2205 if (offset < 0) return -1; // dead, dead 2206 return offset; 2207 } 2208 2209 // Helper for proving that an initialization expression is 2210 // "simple enough" to be folded into an object initialization. 2211 // Attempts to prove that a store's initial value 'n' can be captured 2212 // within the initialization without creating a vicious cycle, such as: 2213 // { Foo p = new Foo(); p.next = p; } 2214 // True for constants and parameters and small combinations thereof. 2215 bool InitializeNode::detect_init_independence(Node* n, 2216 bool st_is_pinned, 2217 int& count) { 2218 if (n == NULL) return true; // (can this really happen?) 2219 if (n->is_Proj()) n = n->in(0); 2220 if (n == this) return false; // found a cycle 2221 if (n->is_Con()) return true; 2222 if (n->is_Start()) return true; // params, etc., are OK 2223 if (n->is_Root()) return true; // even better 2224 2225 Node* ctl = n->in(0); 2226 if (ctl != NULL && !ctl->is_top()) { 2227 if (ctl->is_Proj()) ctl = ctl->in(0); 2228 if (ctl == this) return false; 2229 2230 // If we already know that the enclosing memory op is pinned right after 2231 // the init, then any control flow that the store has picked up 2232 // must have preceded the init, or else be equal to the init. 2233 // Even after loop optimizations (which might change control edges) 2234 // a store is never pinned *before* the availability of its inputs. 2235 if (!MemNode::detect_dominating_control(ctl, this->in(0))) 2236 return false; // failed to prove a good control 2237 2238 } 2239 2240 // Check data edges for possible dependencies on 'this'. 2241 if ((count += 1) > 20) return false; // complexity limit 2242 for (uint i = 1; i < n->req(); i++) { 2243 Node* m = n->in(i); 2244 if (m == NULL || m == n || m->is_top()) continue; 2245 uint first_i = n->find_edge(m); 2246 if (i != first_i) continue; // process duplicate edge just once 2247 if (!detect_init_independence(m, st_is_pinned, count)) { 2248 return false; 2249 } 2250 } 2251 2252 return true; 2253 } 2254 2255 // Here are all the checks a Store must pass before it can be moved into 2256 // an initialization. Returns zero if a check fails. 2257 // On success, returns the (constant) offset to which the store applies, 2258 // within the initialized memory. 2259 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2260 const int FAIL = 0; 2261 if (st->req() != MemNode::ValueIn + 1) 2262 return FAIL; // an inscrutable StoreNode (card mark?) 2263 Node* ctl = st->in(MemNode::Control); 2264 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2265 return FAIL; // must be unconditional after the initialization 2266 Node* mem = st->in(MemNode::Memory); 2267 if (!(mem->is_Proj() && mem->in(0) == this)) 2268 return FAIL; // must not be preceded by other stores 2269 Node* adr = st->in(MemNode::Address); 2270 intptr_t offset; 2271 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2272 if (alloc == NULL) 2273 return FAIL; // inscrutable address 2274 if (alloc != allocation()) 2275 return FAIL; // wrong allocation! (store needs to float up) 2276 Node* val = st->in(MemNode::ValueIn); 2277 int complexity_count = 0; 2278 if (!detect_init_independence(val, true, complexity_count)) 2279 return FAIL; // stored value must be 'simple enough' 2280 2281 return offset; // success 2282 } 2283 2284 // Find the captured store in(i) which corresponds to the range 2285 // [start..start+size) in the initialized object. 2286 // If there is one, return its index i. If there isn't, return the 2287 // negative of the index where it should be inserted. 2288 // Return 0 if the queried range overlaps an initialization boundary 2289 // or if dead code is encountered. 2290 // If size_in_bytes is zero, do not bother with overlap checks. 2291 int InitializeNode::captured_store_insertion_point(intptr_t start, 2292 int size_in_bytes, 2293 PhaseTransform* phase) { 2294 const int FAIL = 0, MAX_STORE = BytesPerLong; 2295 2296 if (is_complete()) 2297 return FAIL; // arraycopy got here first; punt 2298 2299 assert(allocation() != NULL, "must be present"); 2300 2301 // no negatives, no header fields: 2302 if (start < (intptr_t) sizeof(oopDesc)) return FAIL; 2303 if (start < (intptr_t) sizeof(arrayOopDesc) && 2304 start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 2305 2306 // after a certain size, we bail out on tracking all the stores: 2307 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2308 if (start >= ti_limit) return FAIL; 2309 2310 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 2311 if (i >= limit) return -(int)i; // not found; here is where to put it 2312 2313 Node* st = in(i); 2314 intptr_t st_off = get_store_offset(st, phase); 2315 if (st_off < 0) { 2316 if (st != zero_memory()) { 2317 return FAIL; // bail out if there is dead garbage 2318 } 2319 } else if (st_off > start) { 2320 // ...we are done, since stores are ordered 2321 if (st_off < start + size_in_bytes) { 2322 return FAIL; // the next store overlaps 2323 } 2324 return -(int)i; // not found; here is where to put it 2325 } else if (st_off < start) { 2326 if (size_in_bytes != 0 && 2327 start < st_off + MAX_STORE && 2328 start < st_off + st->as_Store()->memory_size()) { 2329 return FAIL; // the previous store overlaps 2330 } 2331 } else { 2332 if (size_in_bytes != 0 && 2333 st->as_Store()->memory_size() != size_in_bytes) { 2334 return FAIL; // mismatched store size 2335 } 2336 return i; 2337 } 2338 2339 ++i; 2340 } 2341 } 2342 2343 // Look for a captured store which initializes at the offset 'start' 2344 // with the given size. If there is no such store, and no other 2345 // initialization interferes, then return zero_memory (the memory 2346 // projection of the AllocateNode). 2347 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 2348 PhaseTransform* phase) { 2349 assert(stores_are_sane(phase), ""); 2350 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2351 if (i == 0) { 2352 return NULL; // something is dead 2353 } else if (i < 0) { 2354 return zero_memory(); // just primordial zero bits here 2355 } else { 2356 Node* st = in(i); // here is the store at this position 2357 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 2358 return st; 2359 } 2360 } 2361 2362 // Create, as a raw pointer, an address within my new object at 'offset'. 2363 Node* InitializeNode::make_raw_address(intptr_t offset, 2364 PhaseTransform* phase) { 2365 Node* addr = in(RawAddress); 2366 if (offset != 0) { 2367 Compile* C = phase->C; 2368 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 2369 phase->MakeConX(offset)) ); 2370 } 2371 return addr; 2372 } 2373 2374 // Clone the given store, converting it into a raw store 2375 // initializing a field or element of my new object. 2376 // Caller is responsible for retiring the original store, 2377 // with subsume_node or the like. 2378 // 2379 // From the example above InitializeNode::InitializeNode, 2380 // here are the old stores to be captured: 2381 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2382 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2383 // 2384 // Here is the changed code; note the extra edges on init: 2385 // alloc = (Allocate ...) 2386 // rawoop = alloc.RawAddress 2387 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 2388 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 2389 // init = (Initialize alloc.Control alloc.Memory rawoop 2390 // rawstore1 rawstore2) 2391 // 2392 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 2393 PhaseTransform* phase) { 2394 assert(stores_are_sane(phase), ""); 2395 2396 if (start < 0) return NULL; 2397 assert(can_capture_store(st, phase) == start, "sanity"); 2398 2399 Compile* C = phase->C; 2400 int size_in_bytes = st->memory_size(); 2401 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2402 if (i == 0) return NULL; // bail out 2403 Node* prev_mem = NULL; // raw memory for the captured store 2404 if (i > 0) { 2405 prev_mem = in(i); // there is a pre-existing store under this one 2406 set_req(i, C->top()); // temporarily disconnect it 2407 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 2408 } else { 2409 i = -i; // no pre-existing store 2410 prev_mem = zero_memory(); // a slice of the newly allocated object 2411 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 2412 set_req(--i, C->top()); // reuse this edge; it has been folded away 2413 else 2414 ins_req(i, C->top()); // build a new edge 2415 } 2416 Node* new_st = st->clone(); 2417 new_st->set_req(MemNode::Control, in(Control)); 2418 new_st->set_req(MemNode::Memory, prev_mem); 2419 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 2420 new_st = phase->transform(new_st); 2421 2422 // At this point, new_st might have swallowed a pre-existing store 2423 // at the same offset, or perhaps new_st might have disappeared, 2424 // if it redundantly stored the same value (or zero to fresh memory). 2425 2426 // In any case, wire it in: 2427 set_req(i, new_st); 2428 2429 // The caller may now kill the old guy. 2430 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 2431 assert(check_st == new_st || check_st == NULL, "must be findable"); 2432 assert(!is_complete(), ""); 2433 return new_st; 2434 } 2435 2436 static bool store_constant(jlong* tiles, int num_tiles, 2437 intptr_t st_off, int st_size, 2438 jlong con) { 2439 if ((st_off & (st_size-1)) != 0) 2440 return false; // strange store offset (assume size==2**N) 2441 address addr = (address)tiles + st_off; 2442 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 2443 switch (st_size) { 2444 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 2445 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 2446 case sizeof(jint): *(jint*) addr = (jint) con; break; 2447 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 2448 default: return false; // strange store size (detect size!=2**N here) 2449 } 2450 return true; // return success to caller 2451 } 2452 2453 // Coalesce subword constants into int constants and possibly 2454 // into long constants. The goal, if the CPU permits, 2455 // is to initialize the object with a small number of 64-bit tiles. 2456 // Also, convert floating-point constants to bit patterns. 2457 // Non-constants are not relevant to this pass. 2458 // 2459 // In terms of the running example on InitializeNode::InitializeNode 2460 // and InitializeNode::capture_store, here is the transformation 2461 // of rawstore1 and rawstore2 into rawstore12: 2462 // alloc = (Allocate ...) 2463 // rawoop = alloc.RawAddress 2464 // tile12 = 0x00010002 2465 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 2466 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 2467 // 2468 void 2469 InitializeNode::coalesce_subword_stores(intptr_t header_size, 2470 Node* size_in_bytes, 2471 PhaseGVN* phase) { 2472 Compile* C = phase->C; 2473 2474 assert(stores_are_sane(phase), ""); 2475 // Note: After this pass, they are not completely sane, 2476 // since there may be some overlaps. 2477 2478 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 2479 2480 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2481 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 2482 size_limit = MIN2(size_limit, ti_limit); 2483 size_limit = align_size_up(size_limit, BytesPerLong); 2484 int num_tiles = size_limit / BytesPerLong; 2485 2486 // allocate space for the tile map: 2487 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 2488 jlong tiles_buf[small_len]; 2489 Node* nodes_buf[small_len]; 2490 jlong inits_buf[small_len]; 2491 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 2492 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2493 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 2494 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 2495 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 2496 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2497 // tiles: exact bitwise model of all primitive constants 2498 // nodes: last constant-storing node subsumed into the tiles model 2499 // inits: which bytes (in each tile) are touched by any initializations 2500 2501 //// Pass A: Fill in the tile model with any relevant stores. 2502 2503 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 2504 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 2505 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 2506 Node* zmem = zero_memory(); // initially zero memory state 2507 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 2508 Node* st = in(i); 2509 intptr_t st_off = get_store_offset(st, phase); 2510 2511 // Figure out the store's offset and constant value: 2512 if (st_off < header_size) continue; //skip (ignore header) 2513 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 2514 int st_size = st->as_Store()->memory_size(); 2515 if (st_off + st_size > size_limit) break; 2516 2517 // Record which bytes are touched, whether by constant or not. 2518 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 2519 continue; // skip (strange store size) 2520 2521 const Type* val = phase->type(st->in(MemNode::ValueIn)); 2522 if (!val->singleton()) continue; //skip (non-con store) 2523 BasicType type = val->basic_type(); 2524 2525 jlong con = 0; 2526 switch (type) { 2527 case T_INT: con = val->is_int()->get_con(); break; 2528 case T_LONG: con = val->is_long()->get_con(); break; 2529 case T_FLOAT: con = jint_cast(val->getf()); break; 2530 case T_DOUBLE: con = jlong_cast(val->getd()); break; 2531 default: continue; //skip (odd store type) 2532 } 2533 2534 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 2535 st->Opcode() == Op_StoreL) { 2536 continue; // This StoreL is already optimal. 2537 } 2538 2539 // Store down the constant. 2540 store_constant(tiles, num_tiles, st_off, st_size, con); 2541 2542 intptr_t j = st_off >> LogBytesPerLong; 2543 2544 if (type == T_INT && st_size == BytesPerInt 2545 && (st_off & BytesPerInt) == BytesPerInt) { 2546 jlong lcon = tiles[j]; 2547 if (!Matcher::isSimpleConstant64(lcon) && 2548 st->Opcode() == Op_StoreI) { 2549 // This StoreI is already optimal by itself. 2550 jint* intcon = (jint*) &tiles[j]; 2551 intcon[1] = 0; // undo the store_constant() 2552 2553 // If the previous store is also optimal by itself, back up and 2554 // undo the action of the previous loop iteration... if we can. 2555 // But if we can't, just let the previous half take care of itself. 2556 st = nodes[j]; 2557 st_off -= BytesPerInt; 2558 con = intcon[0]; 2559 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 2560 assert(st_off >= header_size, "still ignoring header"); 2561 assert(get_store_offset(st, phase) == st_off, "must be"); 2562 assert(in(i-1) == zmem, "must be"); 2563 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 2564 assert(con == tcon->is_int()->get_con(), "must be"); 2565 // Undo the effects of the previous loop trip, which swallowed st: 2566 intcon[0] = 0; // undo store_constant() 2567 set_req(i-1, st); // undo set_req(i, zmem) 2568 nodes[j] = NULL; // undo nodes[j] = st 2569 --old_subword; // undo ++old_subword 2570 } 2571 continue; // This StoreI is already optimal. 2572 } 2573 } 2574 2575 // This store is not needed. 2576 set_req(i, zmem); 2577 nodes[j] = st; // record for the moment 2578 if (st_size < BytesPerLong) // something has changed 2579 ++old_subword; // includes int/float, but who's counting... 2580 else ++old_long; 2581 } 2582 2583 if ((old_subword + old_long) == 0) 2584 return; // nothing more to do 2585 2586 //// Pass B: Convert any non-zero tiles into optimal constant stores. 2587 // Be sure to insert them before overlapping non-constant stores. 2588 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 2589 for (int j = 0; j < num_tiles; j++) { 2590 jlong con = tiles[j]; 2591 jlong init = inits[j]; 2592 if (con == 0) continue; 2593 jint con0, con1; // split the constant, address-wise 2594 jint init0, init1; // split the init map, address-wise 2595 { union { jlong con; jint intcon[2]; } u; 2596 u.con = con; 2597 con0 = u.intcon[0]; 2598 con1 = u.intcon[1]; 2599 u.con = init; 2600 init0 = u.intcon[0]; 2601 init1 = u.intcon[1]; 2602 } 2603 2604 Node* old = nodes[j]; 2605 assert(old != NULL, "need the prior store"); 2606 intptr_t offset = (j * BytesPerLong); 2607 2608 bool split = !Matcher::isSimpleConstant64(con); 2609 2610 if (offset < header_size) { 2611 assert(offset + BytesPerInt >= header_size, "second int counts"); 2612 assert(*(jint*)&tiles[j] == 0, "junk in header"); 2613 split = true; // only the second word counts 2614 // Example: int a[] = { 42 ... } 2615 } else if (con0 == 0 && init0 == -1) { 2616 split = true; // first word is covered by full inits 2617 // Example: int a[] = { ... foo(), 42 ... } 2618 } else if (con1 == 0 && init1 == -1) { 2619 split = true; // second word is covered by full inits 2620 // Example: int a[] = { ... 42, foo() ... } 2621 } 2622 2623 // Here's a case where init0 is neither 0 nor -1: 2624 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 2625 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 2626 // In this case the tile is not split; it is (jlong)42. 2627 // The big tile is stored down, and then the foo() value is inserted. 2628 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 2629 2630 Node* ctl = old->in(MemNode::Control); 2631 Node* adr = make_raw_address(offset, phase); 2632 const TypePtr* atp = TypeRawPtr::BOTTOM; 2633 2634 // One or two coalesced stores to plop down. 2635 Node* st[2]; 2636 intptr_t off[2]; 2637 int nst = 0; 2638 if (!split) { 2639 ++new_long; 2640 off[nst] = offset; 2641 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, 2642 phase->longcon(con), T_LONG); 2643 } else { 2644 // Omit either if it is a zero. 2645 if (con0 != 0) { 2646 ++new_int; 2647 off[nst] = offset; 2648 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, 2649 phase->intcon(con0), T_INT); 2650 } 2651 if (con1 != 0) { 2652 ++new_int; 2653 offset += BytesPerInt; 2654 adr = make_raw_address(offset, phase); 2655 off[nst] = offset; 2656 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, 2657 phase->intcon(con1), T_INT); 2658 } 2659 } 2660 2661 // Insert second store first, then the first before the second. 2662 // Insert each one just before any overlapping non-constant stores. 2663 while (nst > 0) { 2664 Node* st1 = st[--nst]; 2665 C->copy_node_notes_to(st1, old); 2666 st1 = phase->transform(st1); 2667 offset = off[nst]; 2668 assert(offset >= header_size, "do not smash header"); 2669 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 2670 guarantee(ins_idx != 0, "must re-insert constant store"); 2671 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 2672 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 2673 set_req(--ins_idx, st1); 2674 else 2675 ins_req(ins_idx, st1); 2676 } 2677 } 2678 2679 if (PrintCompilation && WizardMode) 2680 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 2681 old_subword, old_long, new_int, new_long); 2682 if (C->log() != NULL) 2683 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 2684 old_subword, old_long, new_int, new_long); 2685 2686 // Clean up any remaining occurrences of zmem: 2687 remove_extra_zeroes(); 2688 } 2689 2690 // Explore forward from in(start) to find the first fully initialized 2691 // word, and return its offset. Skip groups of subword stores which 2692 // together initialize full words. If in(start) is itself part of a 2693 // fully initialized word, return the offset of in(start). If there 2694 // are no following full-word stores, or if something is fishy, return 2695 // a negative value. 2696 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 2697 int int_map = 0; 2698 intptr_t int_map_off = 0; 2699 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 2700 2701 for (uint i = start, limit = req(); i < limit; i++) { 2702 Node* st = in(i); 2703 2704 intptr_t st_off = get_store_offset(st, phase); 2705 if (st_off < 0) break; // return conservative answer 2706 2707 int st_size = st->as_Store()->memory_size(); 2708 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 2709 return st_off; // we found a complete word init 2710 } 2711 2712 // update the map: 2713 2714 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 2715 if (this_int_off != int_map_off) { 2716 // reset the map: 2717 int_map = 0; 2718 int_map_off = this_int_off; 2719 } 2720 2721 int subword_off = st_off - this_int_off; 2722 int_map |= right_n_bits(st_size) << subword_off; 2723 if ((int_map & FULL_MAP) == FULL_MAP) { 2724 return this_int_off; // we found a complete word init 2725 } 2726 2727 // Did this store hit or cross the word boundary? 2728 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 2729 if (next_int_off == this_int_off + BytesPerInt) { 2730 // We passed the current int, without fully initializing it. 2731 int_map_off = next_int_off; 2732 int_map >>= BytesPerInt; 2733 } else if (next_int_off > this_int_off + BytesPerInt) { 2734 // We passed the current and next int. 2735 return this_int_off + BytesPerInt; 2736 } 2737 } 2738 2739 return -1; 2740 } 2741 2742 2743 // Called when the associated AllocateNode is expanded into CFG. 2744 // At this point, we may perform additional optimizations. 2745 // Linearize the stores by ascending offset, to make memory 2746 // activity as coherent as possible. 2747 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 2748 intptr_t header_size, 2749 Node* size_in_bytes, 2750 PhaseGVN* phase) { 2751 assert(!is_complete(), "not already complete"); 2752 assert(stores_are_sane(phase), ""); 2753 assert(allocation() != NULL, "must be present"); 2754 2755 remove_extra_zeroes(); 2756 2757 if (ReduceFieldZeroing || ReduceBulkZeroing) 2758 // reduce instruction count for common initialization patterns 2759 coalesce_subword_stores(header_size, size_in_bytes, phase); 2760 2761 Node* zmem = zero_memory(); // initially zero memory state 2762 Node* inits = zmem; // accumulating a linearized chain of inits 2763 #ifdef ASSERT 2764 intptr_t last_init_off = sizeof(oopDesc); // previous init offset 2765 intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size 2766 intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size 2767 #endif 2768 intptr_t zeroes_done = header_size; 2769 2770 bool do_zeroing = true; // we might give up if inits are very sparse 2771 int big_init_gaps = 0; // how many large gaps have we seen? 2772 2773 if (ZeroTLAB) do_zeroing = false; 2774 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 2775 2776 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 2777 Node* st = in(i); 2778 intptr_t st_off = get_store_offset(st, phase); 2779 if (st_off < 0) 2780 break; // unknown junk in the inits 2781 if (st->in(MemNode::Memory) != zmem) 2782 break; // complicated store chains somehow in list 2783 2784 int st_size = st->as_Store()->memory_size(); 2785 intptr_t next_init_off = st_off + st_size; 2786 2787 if (do_zeroing && zeroes_done < next_init_off) { 2788 // See if this store needs a zero before it or under it. 2789 intptr_t zeroes_needed = st_off; 2790 2791 if (st_size < BytesPerInt) { 2792 // Look for subword stores which only partially initialize words. 2793 // If we find some, we must lay down some word-level zeroes first, 2794 // underneath the subword stores. 2795 // 2796 // Examples: 2797 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 2798 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 2799 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 2800 // 2801 // Note: coalesce_subword_stores may have already done this, 2802 // if it was prompted by constant non-zero subword initializers. 2803 // But this case can still arise with non-constant stores. 2804 2805 intptr_t next_full_store = find_next_fullword_store(i, phase); 2806 2807 // In the examples above: 2808 // in(i) p q r s x y z 2809 // st_off 12 13 14 15 12 13 14 2810 // st_size 1 1 1 1 1 1 1 2811 // next_full_s. 12 16 16 16 16 16 16 2812 // z's_done 12 16 16 16 12 16 12 2813 // z's_needed 12 16 16 16 16 16 16 2814 // zsize 0 0 0 0 4 0 4 2815 if (next_full_store < 0) { 2816 // Conservative tack: Zero to end of current word. 2817 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 2818 } else { 2819 // Zero to beginning of next fully initialized word. 2820 // Or, don't zero at all, if we are already in that word. 2821 assert(next_full_store >= zeroes_needed, "must go forward"); 2822 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 2823 zeroes_needed = next_full_store; 2824 } 2825 } 2826 2827 if (zeroes_needed > zeroes_done) { 2828 intptr_t zsize = zeroes_needed - zeroes_done; 2829 // Do some incremental zeroing on rawmem, in parallel with inits. 2830 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 2831 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 2832 zeroes_done, zeroes_needed, 2833 phase); 2834 zeroes_done = zeroes_needed; 2835 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 2836 do_zeroing = false; // leave the hole, next time 2837 } 2838 } 2839 2840 // Collect the store and move on: 2841 st->set_req(MemNode::Memory, inits); 2842 inits = st; // put it on the linearized chain 2843 set_req(i, zmem); // unhook from previous position 2844 2845 if (zeroes_done == st_off) 2846 zeroes_done = next_init_off; 2847 2848 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 2849 2850 #ifdef ASSERT 2851 // Various order invariants. Weaker than stores_are_sane because 2852 // a large constant tile can be filled in by smaller non-constant stores. 2853 assert(st_off >= last_init_off, "inits do not reverse"); 2854 last_init_off = st_off; 2855 const Type* val = NULL; 2856 if (st_size >= BytesPerInt && 2857 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 2858 (int)val->basic_type() < (int)T_OBJECT) { 2859 assert(st_off >= last_tile_end, "tiles do not overlap"); 2860 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 2861 last_tile_end = MAX2(last_tile_end, next_init_off); 2862 } else { 2863 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 2864 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 2865 assert(st_off >= last_init_end, "inits do not overlap"); 2866 last_init_end = next_init_off; // it's a non-tile 2867 } 2868 #endif //ASSERT 2869 } 2870 2871 remove_extra_zeroes(); // clear out all the zmems left over 2872 add_req(inits); 2873 2874 if (!ZeroTLAB) { 2875 // If anything remains to be zeroed, zero it all now. 2876 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 2877 // if it is the last unused 4 bytes of an instance, forget about it 2878 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 2879 if (zeroes_done + BytesPerLong >= size_limit) { 2880 assert(allocation() != NULL, ""); 2881 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 2882 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 2883 if (zeroes_done == k->layout_helper()) 2884 zeroes_done = size_limit; 2885 } 2886 if (zeroes_done < size_limit) { 2887 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 2888 zeroes_done, size_in_bytes, phase); 2889 } 2890 } 2891 2892 set_complete(phase); 2893 return rawmem; 2894 } 2895 2896 2897 #ifdef ASSERT 2898 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 2899 if (is_complete()) 2900 return true; // stores could be anything at this point 2901 intptr_t last_off = sizeof(oopDesc); 2902 for (uint i = InitializeNode::RawStores; i < req(); i++) { 2903 Node* st = in(i); 2904 intptr_t st_off = get_store_offset(st, phase); 2905 if (st_off < 0) continue; // ignore dead garbage 2906 if (last_off > st_off) { 2907 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 2908 this->dump(2); 2909 assert(false, "ascending store offsets"); 2910 return false; 2911 } 2912 last_off = st_off + st->as_Store()->memory_size(); 2913 } 2914 return true; 2915 } 2916 #endif //ASSERT 2917 2918 2919 2920 2921 //============================MergeMemNode===================================== 2922 // 2923 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 2924 // contributing store or call operations. Each contributor provides the memory 2925 // state for a particular "alias type" (see Compile::alias_type). For example, 2926 // if a MergeMem has an input X for alias category #6, then any memory reference 2927 // to alias category #6 may use X as its memory state input, as an exact equivalent 2928 // to using the MergeMem as a whole. 2929 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 2930 // 2931 // (Here, the <N> notation gives the index of the relevant adr_type.) 2932 // 2933 // In one special case (and more cases in the future), alias categories overlap. 2934 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 2935 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 2936 // it is exactly equivalent to that state W: 2937 // MergeMem(<Bot>: W) <==> W 2938 // 2939 // Usually, the merge has more than one input. In that case, where inputs 2940 // overlap (i.e., one is Bot), the narrower alias type determines the memory 2941 // state for that type, and the wider alias type (Bot) fills in everywhere else: 2942 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 2943 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 2944 // 2945 // A merge can take a "wide" memory state as one of its narrow inputs. 2946 // This simply means that the merge observes out only the relevant parts of 2947 // the wide input. That is, wide memory states arriving at narrow merge inputs 2948 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 2949 // 2950 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 2951 // and that memory slices "leak through": 2952 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 2953 // 2954 // But, in such a cascade, repeated memory slices can "block the leak": 2955 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 2956 // 2957 // In the last example, Y is not part of the combined memory state of the 2958 // outermost MergeMem. The system must, of course, prevent unschedulable 2959 // memory states from arising, so you can be sure that the state Y is somehow 2960 // a precursor to state Y'. 2961 // 2962 // 2963 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 2964 // of each MergeMemNode array are exactly the numerical alias indexes, including 2965 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 2966 // Compile::alias_type (and kin) produce and manage these indexes. 2967 // 2968 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 2969 // (Note that this provides quick access to the top node inside MergeMem methods, 2970 // without the need to reach out via TLS to Compile::current.) 2971 // 2972 // As a consequence of what was just described, a MergeMem that represents a full 2973 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 2974 // containing all alias categories. 2975 // 2976 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 2977 // 2978 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 2979 // a memory state for the alias type <N>, or else the top node, meaning that 2980 // there is no particular input for that alias type. Note that the length of 2981 // a MergeMem is variable, and may be extended at any time to accommodate new 2982 // memory states at larger alias indexes. When merges grow, they are of course 2983 // filled with "top" in the unused in() positions. 2984 // 2985 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 2986 // (Top was chosen because it works smoothly with passes like GCM.) 2987 // 2988 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 2989 // the type of random VM bits like TLS references.) Since it is always the 2990 // first non-Bot memory slice, some low-level loops use it to initialize an 2991 // index variable: for (i = AliasIdxRaw; i < req(); i++). 2992 // 2993 // 2994 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 2995 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 2996 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 2997 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 2998 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 2999 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3000 // 3001 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3002 // really that different from the other memory inputs. An abbreviation called 3003 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3004 // 3005 // 3006 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3007 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3008 // that "emerges though" the base memory will be marked as excluding the alias types 3009 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 3010 // 3011 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3012 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3013 // 3014 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 3015 // (It is currently unimplemented.) As you can see, the resulting merge is 3016 // actually a disjoint union of memory states, rather than an overlay. 3017 // 3018 3019 //------------------------------MergeMemNode----------------------------------- 3020 Node* MergeMemNode::make_empty_memory() { 3021 Node* empty_memory = (Node*) Compile::current()->top(); 3022 assert(empty_memory->is_top(), "correct sentinel identity"); 3023 return empty_memory; 3024 } 3025 3026 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3027 init_class_id(Class_MergeMem); 3028 // all inputs are nullified in Node::Node(int) 3029 // set_input(0, NULL); // no control input 3030 3031 // Initialize the edges uniformly to top, for starters. 3032 Node* empty_mem = make_empty_memory(); 3033 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3034 init_req(i,empty_mem); 3035 } 3036 assert(empty_memory() == empty_mem, ""); 3037 3038 if( new_base != NULL && new_base->is_MergeMem() ) { 3039 MergeMemNode* mdef = new_base->as_MergeMem(); 3040 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3041 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3042 mms.set_memory(mms.memory2()); 3043 } 3044 assert(base_memory() == mdef->base_memory(), ""); 3045 } else { 3046 set_base_memory(new_base); 3047 } 3048 } 3049 3050 // Make a new, untransformed MergeMem with the same base as 'mem'. 3051 // If mem is itself a MergeMem, populate the result with the same edges. 3052 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3053 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3054 } 3055 3056 //------------------------------cmp-------------------------------------------- 3057 uint MergeMemNode::hash() const { return NO_HASH; } 3058 uint MergeMemNode::cmp( const Node &n ) const { 3059 return (&n == this); // Always fail except on self 3060 } 3061 3062 //------------------------------Identity--------------------------------------- 3063 Node* MergeMemNode::Identity(PhaseTransform *phase) { 3064 // Identity if this merge point does not record any interesting memory 3065 // disambiguations. 3066 Node* base_mem = base_memory(); 3067 Node* empty_mem = empty_memory(); 3068 if (base_mem != empty_mem) { // Memory path is not dead? 3069 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3070 Node* mem = in(i); 3071 if (mem != empty_mem && mem != base_mem) { 3072 return this; // Many memory splits; no change 3073 } 3074 } 3075 } 3076 return base_mem; // No memory splits; ID on the one true input 3077 } 3078 3079 //------------------------------Ideal------------------------------------------ 3080 // This method is invoked recursively on chains of MergeMem nodes 3081 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3082 // Remove chain'd MergeMems 3083 // 3084 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3085 // relative to the "in(Bot)". Since we are patching both at the same time, 3086 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3087 // but rewrite each "in(i)" relative to the new "in(Bot)". 3088 Node *progress = NULL; 3089 3090 3091 Node* old_base = base_memory(); 3092 Node* empty_mem = empty_memory(); 3093 if (old_base == empty_mem) 3094 return NULL; // Dead memory path. 3095 3096 MergeMemNode* old_mbase; 3097 if (old_base != NULL && old_base->is_MergeMem()) 3098 old_mbase = old_base->as_MergeMem(); 3099 else 3100 old_mbase = NULL; 3101 Node* new_base = old_base; 3102 3103 // simplify stacked MergeMems in base memory 3104 if (old_mbase) new_base = old_mbase->base_memory(); 3105 3106 // the base memory might contribute new slices beyond my req() 3107 if (old_mbase) grow_to_match(old_mbase); 3108 3109 // Look carefully at the base node if it is a phi. 3110 PhiNode* phi_base; 3111 if (new_base != NULL && new_base->is_Phi()) 3112 phi_base = new_base->as_Phi(); 3113 else 3114 phi_base = NULL; 3115 3116 Node* phi_reg = NULL; 3117 uint phi_len = (uint)-1; 3118 if (phi_base != NULL && !phi_base->is_copy()) { 3119 // do not examine phi if degraded to a copy 3120 phi_reg = phi_base->region(); 3121 phi_len = phi_base->req(); 3122 // see if the phi is unfinished 3123 for (uint i = 1; i < phi_len; i++) { 3124 if (phi_base->in(i) == NULL) { 3125 // incomplete phi; do not look at it yet! 3126 phi_reg = NULL; 3127 phi_len = (uint)-1; 3128 break; 3129 } 3130 } 3131 } 3132 3133 // Note: We do not call verify_sparse on entry, because inputs 3134 // can normalize to the base_memory via subsume_node or similar 3135 // mechanisms. This method repairs that damage. 3136 3137 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3138 3139 // Look at each slice. 3140 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3141 Node* old_in = in(i); 3142 // calculate the old memory value 3143 Node* old_mem = old_in; 3144 if (old_mem == empty_mem) old_mem = old_base; 3145 assert(old_mem == memory_at(i), ""); 3146 3147 // maybe update (reslice) the old memory value 3148 3149 // simplify stacked MergeMems 3150 Node* new_mem = old_mem; 3151 MergeMemNode* old_mmem; 3152 if (old_mem != NULL && old_mem->is_MergeMem()) 3153 old_mmem = old_mem->as_MergeMem(); 3154 else 3155 old_mmem = NULL; 3156 if (old_mmem == this) { 3157 // This can happen if loops break up and safepoints disappear. 3158 // A merge of BotPtr (default) with a RawPtr memory derived from a 3159 // safepoint can be rewritten to a merge of the same BotPtr with 3160 // the BotPtr phi coming into the loop. If that phi disappears 3161 // also, we can end up with a self-loop of the mergemem. 3162 // In general, if loops degenerate and memory effects disappear, 3163 // a mergemem can be left looking at itself. This simply means 3164 // that the mergemem's default should be used, since there is 3165 // no longer any apparent effect on this slice. 3166 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3167 // from start. Update the input to TOP. 3168 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3169 } 3170 else if (old_mmem != NULL) { 3171 new_mem = old_mmem->memory_at(i); 3172 } 3173 // else preceeding memory was not a MergeMem 3174 3175 // replace equivalent phis (unfortunately, they do not GVN together) 3176 if (new_mem != NULL && new_mem != new_base && 3177 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3178 if (new_mem->is_Phi()) { 3179 PhiNode* phi_mem = new_mem->as_Phi(); 3180 for (uint i = 1; i < phi_len; i++) { 3181 if (phi_base->in(i) != phi_mem->in(i)) { 3182 phi_mem = NULL; 3183 break; 3184 } 3185 } 3186 if (phi_mem != NULL) { 3187 // equivalent phi nodes; revert to the def 3188 new_mem = new_base; 3189 } 3190 } 3191 } 3192 3193 // maybe store down a new value 3194 Node* new_in = new_mem; 3195 if (new_in == new_base) new_in = empty_mem; 3196 3197 if (new_in != old_in) { 3198 // Warning: Do not combine this "if" with the previous "if" 3199 // A memory slice might have be be rewritten even if it is semantically 3200 // unchanged, if the base_memory value has changed. 3201 set_req(i, new_in); 3202 progress = this; // Report progress 3203 } 3204 } 3205 3206 if (new_base != old_base) { 3207 set_req(Compile::AliasIdxBot, new_base); 3208 // Don't use set_base_memory(new_base), because we need to update du. 3209 assert(base_memory() == new_base, ""); 3210 progress = this; 3211 } 3212 3213 if( base_memory() == this ) { 3214 // a self cycle indicates this memory path is dead 3215 set_req(Compile::AliasIdxBot, empty_mem); 3216 } 3217 3218 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3219 // Recursion must occur after the self cycle check above 3220 if( base_memory()->is_MergeMem() ) { 3221 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3222 Node *m = phase->transform(new_mbase); // Rollup any cycles 3223 if( m != NULL && (m->is_top() || 3224 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3225 // propagate rollup of dead cycle to self 3226 set_req(Compile::AliasIdxBot, empty_mem); 3227 } 3228 } 3229 3230 if( base_memory() == empty_mem ) { 3231 progress = this; 3232 // Cut inputs during Parse phase only. 3233 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3234 if( !can_reshape ) { 3235 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3236 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3237 } 3238 } 3239 } 3240 3241 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3242 // Check if PhiNode::Ideal's "Split phis through memory merges" 3243 // transform should be attempted. Look for this->phi->this cycle. 3244 uint merge_width = req(); 3245 if (merge_width > Compile::AliasIdxRaw) { 3246 PhiNode* phi = base_memory()->as_Phi(); 3247 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3248 if (phi->in(i) == this) { 3249 phase->is_IterGVN()->_worklist.push(phi); 3250 break; 3251 } 3252 } 3253 } 3254 } 3255 3256 assert(verify_sparse(), "please, no dups of base"); 3257 return progress; 3258 } 3259 3260 //-------------------------set_base_memory------------------------------------- 3261 void MergeMemNode::set_base_memory(Node *new_base) { 3262 Node* empty_mem = empty_memory(); 3263 set_req(Compile::AliasIdxBot, new_base); 3264 assert(memory_at(req()) == new_base, "must set default memory"); 3265 // Clear out other occurrences of new_base: 3266 if (new_base != empty_mem) { 3267 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3268 if (in(i) == new_base) set_req(i, empty_mem); 3269 } 3270 } 3271 } 3272 3273 //------------------------------out_RegMask------------------------------------ 3274 const RegMask &MergeMemNode::out_RegMask() const { 3275 return RegMask::Empty; 3276 } 3277 3278 //------------------------------dump_spec-------------------------------------- 3279 #ifndef PRODUCT 3280 void MergeMemNode::dump_spec(outputStream *st) const { 3281 st->print(" {"); 3282 Node* base_mem = base_memory(); 3283 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3284 Node* mem = memory_at(i); 3285 if (mem == base_mem) { st->print(" -"); continue; } 3286 st->print( " N%d:", mem->_idx ); 3287 Compile::current()->get_adr_type(i)->dump_on(st); 3288 } 3289 st->print(" }"); 3290 } 3291 #endif // !PRODUCT 3292 3293 3294 #ifdef ASSERT 3295 static bool might_be_same(Node* a, Node* b) { 3296 if (a == b) return true; 3297 if (!(a->is_Phi() || b->is_Phi())) return false; 3298 // phis shift around during optimization 3299 return true; // pretty stupid... 3300 } 3301 3302 // verify a narrow slice (either incoming or outgoing) 3303 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 3304 if (!VerifyAliases) return; // don't bother to verify unless requested 3305 if (is_error_reported()) return; // muzzle asserts when debugging an error 3306 if (Node::in_dump()) return; // muzzle asserts when printing 3307 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 3308 assert(n != NULL, ""); 3309 // Elide intervening MergeMem's 3310 while (n->is_MergeMem()) { 3311 n = n->as_MergeMem()->memory_at(alias_idx); 3312 } 3313 Compile* C = Compile::current(); 3314 const TypePtr* n_adr_type = n->adr_type(); 3315 if (n == m->empty_memory()) { 3316 // Implicit copy of base_memory() 3317 } else if (n_adr_type != TypePtr::BOTTOM) { 3318 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 3319 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 3320 } else { 3321 // A few places like make_runtime_call "know" that VM calls are narrow, 3322 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 3323 bool expected_wide_mem = false; 3324 if (n == m->base_memory()) { 3325 expected_wide_mem = true; 3326 } else if (alias_idx == Compile::AliasIdxRaw || 3327 n == m->memory_at(Compile::AliasIdxRaw)) { 3328 expected_wide_mem = true; 3329 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 3330 // memory can "leak through" calls on channels that 3331 // are write-once. Allow this also. 3332 expected_wide_mem = true; 3333 } 3334 assert(expected_wide_mem, "expected narrow slice replacement"); 3335 } 3336 } 3337 #else // !ASSERT 3338 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 3339 #endif 3340 3341 3342 //-----------------------------memory_at--------------------------------------- 3343 Node* MergeMemNode::memory_at(uint alias_idx) const { 3344 assert(alias_idx >= Compile::AliasIdxRaw || 3345 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 3346 "must avoid base_memory and AliasIdxTop"); 3347 3348 // Otherwise, it is a narrow slice. 3349 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 3350 Compile *C = Compile::current(); 3351 if (is_empty_memory(n)) { 3352 // the array is sparse; empty slots are the "top" node 3353 n = base_memory(); 3354 assert(Node::in_dump() 3355 || n == NULL || n->bottom_type() == Type::TOP 3356 || n->adr_type() == TypePtr::BOTTOM 3357 || n->adr_type() == TypeRawPtr::BOTTOM 3358 || Compile::current()->AliasLevel() == 0, 3359 "must be a wide memory"); 3360 // AliasLevel == 0 if we are organizing the memory states manually. 3361 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 3362 } else { 3363 // make sure the stored slice is sane 3364 #ifdef ASSERT 3365 if (is_error_reported() || Node::in_dump()) { 3366 } else if (might_be_same(n, base_memory())) { 3367 // Give it a pass: It is a mostly harmless repetition of the base. 3368 // This can arise normally from node subsumption during optimization. 3369 } else { 3370 verify_memory_slice(this, alias_idx, n); 3371 } 3372 #endif 3373 } 3374 return n; 3375 } 3376 3377 //---------------------------set_memory_at------------------------------------- 3378 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 3379 verify_memory_slice(this, alias_idx, n); 3380 Node* empty_mem = empty_memory(); 3381 if (n == base_memory()) n = empty_mem; // collapse default 3382 uint need_req = alias_idx+1; 3383 if (req() < need_req) { 3384 if (n == empty_mem) return; // already the default, so do not grow me 3385 // grow the sparse array 3386 do { 3387 add_req(empty_mem); 3388 } while (req() < need_req); 3389 } 3390 set_req( alias_idx, n ); 3391 } 3392 3393 3394 3395 //--------------------------iteration_setup------------------------------------ 3396 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 3397 if (other != NULL) { 3398 grow_to_match(other); 3399 // invariant: the finite support of mm2 is within mm->req() 3400 #ifdef ASSERT 3401 for (uint i = req(); i < other->req(); i++) { 3402 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 3403 } 3404 #endif 3405 } 3406 // Replace spurious copies of base_memory by top. 3407 Node* base_mem = base_memory(); 3408 if (base_mem != NULL && !base_mem->is_top()) { 3409 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 3410 if (in(i) == base_mem) 3411 set_req(i, empty_memory()); 3412 } 3413 } 3414 } 3415 3416 //---------------------------grow_to_match------------------------------------- 3417 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 3418 Node* empty_mem = empty_memory(); 3419 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 3420 // look for the finite support of the other memory 3421 for (uint i = other->req(); --i >= req(); ) { 3422 if (other->in(i) != empty_mem) { 3423 uint new_len = i+1; 3424 while (req() < new_len) add_req(empty_mem); 3425 break; 3426 } 3427 } 3428 } 3429 3430 //---------------------------verify_sparse------------------------------------- 3431 #ifndef PRODUCT 3432 bool MergeMemNode::verify_sparse() const { 3433 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 3434 Node* base_mem = base_memory(); 3435 // The following can happen in degenerate cases, since empty==top. 3436 if (is_empty_memory(base_mem)) return true; 3437 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3438 assert(in(i) != NULL, "sane slice"); 3439 if (in(i) == base_mem) return false; // should have been the sentinel value! 3440 } 3441 return true; 3442 } 3443 3444 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 3445 Node* n; 3446 n = mm->in(idx); 3447 if (mem == n) return true; // might be empty_memory() 3448 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 3449 if (mem == n) return true; 3450 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 3451 if (mem == n) return true; 3452 if (n == NULL) break; 3453 } 3454 return false; 3455 } 3456 #endif // !PRODUCT