memnode.cpp revision 0:a61af66fc99e
1/* 2 * Copyright 1997-2007 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25// Portions of code courtesy of Clifford Click 26 27// Optimization - Graph Style 28 29#include "incls/_precompiled.incl" 30#include "incls/_memnode.cpp.incl" 31 32//============================================================================= 33uint MemNode::size_of() const { return sizeof(*this); } 34 35const TypePtr *MemNode::adr_type() const { 36 Node* adr = in(Address); 37 const TypePtr* cross_check = NULL; 38 DEBUG_ONLY(cross_check = _adr_type); 39 return calculate_adr_type(adr->bottom_type(), cross_check); 40} 41 42#ifndef PRODUCT 43void MemNode::dump_spec(outputStream *st) const { 44 if (in(Address) == NULL) return; // node is dead 45#ifndef ASSERT 46 // fake the missing field 47 const TypePtr* _adr_type = NULL; 48 if (in(Address) != NULL) 49 _adr_type = in(Address)->bottom_type()->isa_ptr(); 50#endif 51 dump_adr_type(this, _adr_type, st); 52 53 Compile* C = Compile::current(); 54 if( C->alias_type(_adr_type)->is_volatile() ) 55 st->print(" Volatile!"); 56} 57 58void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 59 st->print(" @"); 60 if (adr_type == NULL) { 61 st->print("NULL"); 62 } else { 63 adr_type->dump_on(st); 64 Compile* C = Compile::current(); 65 Compile::AliasType* atp = NULL; 66 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 67 if (atp == NULL) 68 st->print(", idx=?\?;"); 69 else if (atp->index() == Compile::AliasIdxBot) 70 st->print(", idx=Bot;"); 71 else if (atp->index() == Compile::AliasIdxTop) 72 st->print(", idx=Top;"); 73 else if (atp->index() == Compile::AliasIdxRaw) 74 st->print(", idx=Raw;"); 75 else { 76 ciField* field = atp->field(); 77 if (field) { 78 st->print(", name="); 79 field->print_name_on(st); 80 } 81 st->print(", idx=%d;", atp->index()); 82 } 83 } 84} 85 86extern void print_alias_types(); 87 88#endif 89 90//--------------------------Ideal_common--------------------------------------- 91// Look for degenerate control and memory inputs. Bypass MergeMem inputs. 92// Unhook non-raw memories from complete (macro-expanded) initializations. 93Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 94 // If our control input is a dead region, kill all below the region 95 Node *ctl = in(MemNode::Control); 96 if (ctl && remove_dead_region(phase, can_reshape)) 97 return this; 98 99 // Ignore if memory is dead, or self-loop 100 Node *mem = in(MemNode::Memory); 101 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 102 assert( mem != this, "dead loop in MemNode::Ideal" ); 103 104 Node *address = in(MemNode::Address); 105 const Type *t_adr = phase->type( address ); 106 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 107 108 // Avoid independent memory operations 109 Node* old_mem = mem; 110 111 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 112 InitializeNode* init = mem->in(0)->as_Initialize(); 113 if (init->is_complete()) { // i.e., after macro expansion 114 const TypePtr* tp = t_adr->is_ptr(); 115 uint alias_idx = phase->C->get_alias_index(tp); 116 // Free this slice from the init. It was hooked, temporarily, 117 // by GraphKit::set_output_for_allocation. 118 if (alias_idx > Compile::AliasIdxRaw) { 119 mem = init->memory(alias_idx); 120 // ...but not with the raw-pointer slice. 121 } 122 } 123 } 124 125 if (mem->is_MergeMem()) { 126 MergeMemNode* mmem = mem->as_MergeMem(); 127 const TypePtr *tp = t_adr->is_ptr(); 128 uint alias_idx = phase->C->get_alias_index(tp); 129#ifdef ASSERT 130 { 131 // Check that current type is consistent with the alias index used during graph construction 132 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 133 const TypePtr *adr_t = adr_type(); 134 bool consistent = adr_t == NULL || adr_t->empty() || phase->C->must_alias(adr_t, alias_idx ); 135 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 136 if( !consistent && adr_t != NULL && !adr_t->empty() && 137 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 138 adr_t->isa_aryptr() && adr_t->offset() != Type::OffsetBot && 139 ( adr_t->offset() == arrayOopDesc::length_offset_in_bytes() || 140 adr_t->offset() == oopDesc::klass_offset_in_bytes() || 141 adr_t->offset() == oopDesc::mark_offset_in_bytes() ) ) { 142 // don't assert if it is dead code. 143 consistent = true; 144 } 145 if( !consistent ) { 146 tty->print("alias_idx==%d, adr_type()==", alias_idx); if( adr_t == NULL ) { tty->print("NULL"); } else { adr_t->dump(); } 147 tty->cr(); 148 print_alias_types(); 149 assert(consistent, "adr_type must match alias idx"); 150 } 151 } 152#endif 153 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 154 // means an array I have not precisely typed yet. Do not do any 155 // alias stuff with it any time soon. 156 const TypeInstPtr *tinst = tp->isa_instptr(); 157 if( tp->base() != Type::AnyPtr && 158 !(tinst && 159 tinst->klass()->is_java_lang_Object() && 160 tinst->offset() == Type::OffsetBot) ) { 161 // compress paths and change unreachable cycles to TOP 162 // If not, we can update the input infinitely along a MergeMem cycle 163 // Equivalent code in PhiNode::Ideal 164 Node* m = phase->transform(mmem); 165 // If tranformed to a MergeMem, get the desired slice 166 // Otherwise the returned node represents memory for every slice 167 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 168 // Update input if it is progress over what we have now 169 } 170 } 171 172 if (mem != old_mem) { 173 set_req(MemNode::Memory, mem); 174 return this; 175 } 176 177 // let the subclass continue analyzing... 178 return NULL; 179} 180 181// Helper function for proving some simple control dominations. 182// Attempt to prove that control input 'dom' dominates (or equals) 'sub'. 183// Already assumes that 'dom' is available at 'sub', and that 'sub' 184// is not a constant (dominated by the method's StartNode). 185// Used by MemNode::find_previous_store to prove that the 186// control input of a memory operation predates (dominates) 187// an allocation it wants to look past. 188bool MemNode::detect_dominating_control(Node* dom, Node* sub) { 189 if (dom == NULL) return false; 190 if (dom->is_Proj()) dom = dom->in(0); 191 if (dom->is_Start()) return true; // anything inside the method 192 if (dom->is_Root()) return true; // dom 'controls' a constant 193 int cnt = 20; // detect cycle or too much effort 194 while (sub != NULL) { // walk 'sub' up the chain to 'dom' 195 if (--cnt < 0) return false; // in a cycle or too complex 196 if (sub == dom) return true; 197 if (sub->is_Start()) return false; 198 if (sub->is_Root()) return false; 199 Node* up = sub->in(0); 200 if (sub == up && sub->is_Region()) { 201 for (uint i = 1; i < sub->req(); i++) { 202 Node* in = sub->in(i); 203 if (in != NULL && !in->is_top() && in != sub) { 204 up = in; break; // take any path on the way up to 'dom' 205 } 206 } 207 } 208 if (sub == up) return false; // some kind of tight cycle 209 sub = up; 210 } 211 return false; 212} 213 214//---------------------detect_ptr_independence--------------------------------- 215// Used by MemNode::find_previous_store to prove that two base 216// pointers are never equal. 217// The pointers are accompanied by their associated allocations, 218// if any, which have been previously discovered by the caller. 219bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 220 Node* p2, AllocateNode* a2, 221 PhaseTransform* phase) { 222 // Attempt to prove that these two pointers cannot be aliased. 223 // They may both manifestly be allocations, and they should differ. 224 // Or, if they are not both allocations, they can be distinct constants. 225 // Otherwise, one is an allocation and the other a pre-existing value. 226 if (a1 == NULL && a2 == NULL) { // neither an allocation 227 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 228 } else if (a1 != NULL && a2 != NULL) { // both allocations 229 return (a1 != a2); 230 } else if (a1 != NULL) { // one allocation a1 231 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 232 return detect_dominating_control(p2->in(0), a1->in(0)); 233 } else { //(a2 != NULL) // one allocation a2 234 return detect_dominating_control(p1->in(0), a2->in(0)); 235 } 236 return false; 237} 238 239 240// The logic for reordering loads and stores uses four steps: 241// (a) Walk carefully past stores and initializations which we 242// can prove are independent of this load. 243// (b) Observe that the next memory state makes an exact match 244// with self (load or store), and locate the relevant store. 245// (c) Ensure that, if we were to wire self directly to the store, 246// the optimizer would fold it up somehow. 247// (d) Do the rewiring, and return, depending on some other part of 248// the optimizer to fold up the load. 249// This routine handles steps (a) and (b). Steps (c) and (d) are 250// specific to loads and stores, so they are handled by the callers. 251// (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 252// 253Node* MemNode::find_previous_store(PhaseTransform* phase) { 254 Node* ctrl = in(MemNode::Control); 255 Node* adr = in(MemNode::Address); 256 intptr_t offset = 0; 257 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 258 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 259 260 if (offset == Type::OffsetBot) 261 return NULL; // cannot unalias unless there are precise offsets 262 263 intptr_t size_in_bytes = memory_size(); 264 265 Node* mem = in(MemNode::Memory); // start searching here... 266 267 int cnt = 50; // Cycle limiter 268 for (;;) { // While we can dance past unrelated stores... 269 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 270 271 if (mem->is_Store()) { 272 Node* st_adr = mem->in(MemNode::Address); 273 intptr_t st_offset = 0; 274 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 275 if (st_base == NULL) 276 break; // inscrutable pointer 277 if (st_offset != offset && st_offset != Type::OffsetBot) { 278 const int MAX_STORE = BytesPerLong; 279 if (st_offset >= offset + size_in_bytes || 280 st_offset <= offset - MAX_STORE || 281 st_offset <= offset - mem->as_Store()->memory_size()) { 282 // Success: The offsets are provably independent. 283 // (You may ask, why not just test st_offset != offset and be done? 284 // The answer is that stores of different sizes can co-exist 285 // in the same sequence of RawMem effects. We sometimes initialize 286 // a whole 'tile' of array elements with a single jint or jlong.) 287 mem = mem->in(MemNode::Memory); 288 continue; // (a) advance through independent store memory 289 } 290 } 291 if (st_base != base && 292 detect_ptr_independence(base, alloc, 293 st_base, 294 AllocateNode::Ideal_allocation(st_base, phase), 295 phase)) { 296 // Success: The bases are provably independent. 297 mem = mem->in(MemNode::Memory); 298 continue; // (a) advance through independent store memory 299 } 300 301 // (b) At this point, if the bases or offsets do not agree, we lose, 302 // since we have not managed to prove 'this' and 'mem' independent. 303 if (st_base == base && st_offset == offset) { 304 return mem; // let caller handle steps (c), (d) 305 } 306 307 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 308 InitializeNode* st_init = mem->in(0)->as_Initialize(); 309 AllocateNode* st_alloc = st_init->allocation(); 310 if (st_alloc == NULL) 311 break; // something degenerated 312 bool known_identical = false; 313 bool known_independent = false; 314 if (alloc == st_alloc) 315 known_identical = true; 316 else if (alloc != NULL) 317 known_independent = true; 318 else if (ctrl != NULL && 319 detect_dominating_control(ctrl, st_alloc->in(0))) 320 known_independent = true; 321 322 if (known_independent) { 323 // The bases are provably independent: Either they are 324 // manifestly distinct allocations, or else the control 325 // of this load dominates the store's allocation. 326 int alias_idx = phase->C->get_alias_index(adr_type()); 327 if (alias_idx == Compile::AliasIdxRaw) { 328 mem = st_alloc->in(TypeFunc::Memory); 329 } else { 330 mem = st_init->memory(alias_idx); 331 } 332 continue; // (a) advance through independent store memory 333 } 334 335 // (b) at this point, if we are not looking at a store initializing 336 // the same allocation we are loading from, we lose. 337 if (known_identical) { 338 // From caller, can_see_stored_value will consult find_captured_store. 339 return mem; // let caller handle steps (c), (d) 340 } 341 342 } 343 344 // Unless there is an explicit 'continue', we must bail out here, 345 // because 'mem' is an inscrutable memory state (e.g., a call). 346 break; 347 } 348 349 return NULL; // bail out 350} 351 352//----------------------calculate_adr_type------------------------------------- 353// Helper function. Notices when the given type of address hits top or bottom. 354// Also, asserts a cross-check of the type against the expected address type. 355const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 356 if (t == Type::TOP) return NULL; // does not touch memory any more? 357 #ifdef PRODUCT 358 cross_check = NULL; 359 #else 360 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 361 #endif 362 const TypePtr* tp = t->isa_ptr(); 363 if (tp == NULL) { 364 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 365 return TypePtr::BOTTOM; // touches lots of memory 366 } else { 367 #ifdef ASSERT 368 // %%%% [phh] We don't check the alias index if cross_check is 369 // TypeRawPtr::BOTTOM. Needs to be investigated. 370 if (cross_check != NULL && 371 cross_check != TypePtr::BOTTOM && 372 cross_check != TypeRawPtr::BOTTOM) { 373 // Recheck the alias index, to see if it has changed (due to a bug). 374 Compile* C = Compile::current(); 375 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 376 "must stay in the original alias category"); 377 // The type of the address must be contained in the adr_type, 378 // disregarding "null"-ness. 379 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 380 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 381 assert(cross_check->meet(tp_notnull) == cross_check, 382 "real address must not escape from expected memory type"); 383 } 384 #endif 385 return tp; 386 } 387} 388 389//------------------------adr_phi_is_loop_invariant---------------------------- 390// A helper function for Ideal_DU_postCCP to check if a Phi in a counted 391// loop is loop invariant. Make a quick traversal of Phi and associated 392// CastPP nodes, looking to see if they are a closed group within the loop. 393bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 394 // The idea is that the phi-nest must boil down to only CastPP nodes 395 // with the same data. This implies that any path into the loop already 396 // includes such a CastPP, and so the original cast, whatever its input, 397 // must be covered by an equivalent cast, with an earlier control input. 398 ResourceMark rm; 399 400 // The loop entry input of the phi should be the unique dominating 401 // node for every Phi/CastPP in the loop. 402 Unique_Node_List closure; 403 closure.push(adr_phi->in(LoopNode::EntryControl)); 404 405 // Add the phi node and the cast to the worklist. 406 Unique_Node_List worklist; 407 worklist.push(adr_phi); 408 if( cast != NULL ){ 409 if( !cast->is_ConstraintCast() ) return false; 410 worklist.push(cast); 411 } 412 413 // Begin recursive walk of phi nodes. 414 while( worklist.size() ){ 415 // Take a node off the worklist 416 Node *n = worklist.pop(); 417 if( !closure.member(n) ){ 418 // Add it to the closure. 419 closure.push(n); 420 // Make a sanity check to ensure we don't waste too much time here. 421 if( closure.size() > 20) return false; 422 // This node is OK if: 423 // - it is a cast of an identical value 424 // - or it is a phi node (then we add its inputs to the worklist) 425 // Otherwise, the node is not OK, and we presume the cast is not invariant 426 if( n->is_ConstraintCast() ){ 427 worklist.push(n->in(1)); 428 } else if( n->is_Phi() ) { 429 for( uint i = 1; i < n->req(); i++ ) { 430 worklist.push(n->in(i)); 431 } 432 } else { 433 return false; 434 } 435 } 436 } 437 438 // Quit when the worklist is empty, and we've found no offending nodes. 439 return true; 440} 441 442//------------------------------Ideal_DU_postCCP------------------------------- 443// Find any cast-away of null-ness and keep its control. Null cast-aways are 444// going away in this pass and we need to make this memory op depend on the 445// gating null check. 446 447// I tried to leave the CastPP's in. This makes the graph more accurate in 448// some sense; we get to keep around the knowledge that an oop is not-null 449// after some test. Alas, the CastPP's interfere with GVN (some values are 450// the regular oop, some are the CastPP of the oop, all merge at Phi's which 451// cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 452// some of the more trivial cases in the optimizer. Removing more useless 453// Phi's started allowing Loads to illegally float above null checks. I gave 454// up on this approach. CNC 10/20/2000 455Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 456 Node *ctr = in(MemNode::Control); 457 Node *mem = in(MemNode::Memory); 458 Node *adr = in(MemNode::Address); 459 Node *skipped_cast = NULL; 460 // Need a null check? Regular static accesses do not because they are 461 // from constant addresses. Array ops are gated by the range check (which 462 // always includes a NULL check). Just check field ops. 463 if( !ctr ) { 464 // Scan upwards for the highest location we can place this memory op. 465 while( true ) { 466 switch( adr->Opcode() ) { 467 468 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 469 adr = adr->in(AddPNode::Base); 470 continue; 471 472 case Op_CastPP: 473 // If the CastPP is useless, just peek on through it. 474 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 475 // Remember the cast that we've peeked though. If we peek 476 // through more than one, then we end up remembering the highest 477 // one, that is, if in a loop, the one closest to the top. 478 skipped_cast = adr; 479 adr = adr->in(1); 480 continue; 481 } 482 // CastPP is going away in this pass! We need this memory op to be 483 // control-dependent on the test that is guarding the CastPP. 484 ccp->hash_delete(this); 485 set_req(MemNode::Control, adr->in(0)); 486 ccp->hash_insert(this); 487 return this; 488 489 case Op_Phi: 490 // Attempt to float above a Phi to some dominating point. 491 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 492 // If we've already peeked through a Cast (which could have set the 493 // control), we can't float above a Phi, because the skipped Cast 494 // may not be loop invariant. 495 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 496 adr = adr->in(1); 497 continue; 498 } 499 } 500 501 // Intentional fallthrough! 502 503 // No obvious dominating point. The mem op is pinned below the Phi 504 // by the Phi itself. If the Phi goes away (no true value is merged) 505 // then the mem op can float, but not indefinitely. It must be pinned 506 // behind the controls leading to the Phi. 507 case Op_CheckCastPP: 508 // These usually stick around to change address type, however a 509 // useless one can be elided and we still need to pick up a control edge 510 if (adr->in(0) == NULL) { 511 // This CheckCastPP node has NO control and is likely useless. But we 512 // need check further up the ancestor chain for a control input to keep 513 // the node in place. 4959717. 514 skipped_cast = adr; 515 adr = adr->in(1); 516 continue; 517 } 518 ccp->hash_delete(this); 519 set_req(MemNode::Control, adr->in(0)); 520 ccp->hash_insert(this); 521 return this; 522 523 // List of "safe" opcodes; those that implicitly block the memory 524 // op below any null check. 525 case Op_CastX2P: // no null checks on native pointers 526 case Op_Parm: // 'this' pointer is not null 527 case Op_LoadP: // Loading from within a klass 528 case Op_LoadKlass: // Loading from within a klass 529 case Op_ConP: // Loading from a klass 530 case Op_CreateEx: // Sucking up the guts of an exception oop 531 case Op_Con: // Reading from TLS 532 case Op_CMoveP: // CMoveP is pinned 533 break; // No progress 534 535 case Op_Proj: // Direct call to an allocation routine 536 case Op_SCMemProj: // Memory state from store conditional ops 537#ifdef ASSERT 538 { 539 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 540 const Node* call = adr->in(0); 541 if (call->is_CallStaticJava()) { 542 const CallStaticJavaNode* call_java = call->as_CallStaticJava(); 543 assert(call_java && call_java->method() == NULL, "must be runtime call"); 544 // We further presume that this is one of 545 // new_instance_Java, new_array_Java, or 546 // the like, but do not assert for this. 547 } else if (call->is_Allocate()) { 548 // similar case to new_instance_Java, etc. 549 } else if (!call->is_CallLeaf()) { 550 // Projections from fetch_oop (OSR) are allowed as well. 551 ShouldNotReachHere(); 552 } 553 } 554#endif 555 break; 556 default: 557 ShouldNotReachHere(); 558 } 559 break; 560 } 561 } 562 563 return NULL; // No progress 564} 565 566 567//============================================================================= 568uint LoadNode::size_of() const { return sizeof(*this); } 569uint LoadNode::cmp( const Node &n ) const 570{ return !Type::cmp( _type, ((LoadNode&)n)._type ); } 571const Type *LoadNode::bottom_type() const { return _type; } 572uint LoadNode::ideal_reg() const { 573 return Matcher::base2reg[_type->base()]; 574} 575 576#ifndef PRODUCT 577void LoadNode::dump_spec(outputStream *st) const { 578 MemNode::dump_spec(st); 579 if( !Verbose && !WizardMode ) { 580 // standard dump does this in Verbose and WizardMode 581 st->print(" #"); _type->dump_on(st); 582 } 583} 584#endif 585 586 587//----------------------------LoadNode::make----------------------------------- 588// Polymorphic factory method: 589LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 590 // sanity check the alias category against the created node type 591 assert(!(adr_type->isa_oopptr() && 592 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 593 "use LoadKlassNode instead"); 594 assert(!(adr_type->isa_aryptr() && 595 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 596 "use LoadRangeNode instead"); 597 switch (bt) { 598 case T_BOOLEAN: 599 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() ); 600 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() ); 601 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() ); 602 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() ); 603 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() ); 604 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt ); 605 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt ); 606 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() ); 607 case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 608 } 609 ShouldNotReachHere(); 610 return (LoadNode*)NULL; 611} 612 613LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 614 bool require_atomic = true; 615 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 616} 617 618 619 620 621//------------------------------hash------------------------------------------- 622uint LoadNode::hash() const { 623 // unroll addition of interesting fields 624 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 625} 626 627//---------------------------can_see_stored_value------------------------------ 628// This routine exists to make sure this set of tests is done the same 629// everywhere. We need to make a coordinated change: first LoadNode::Ideal 630// will change the graph shape in a way which makes memory alive twice at the 631// same time (uses the Oracle model of aliasing), then some 632// LoadXNode::Identity will fold things back to the equivalence-class model 633// of aliasing. 634Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 635 Node* ld_adr = in(MemNode::Address); 636 637 // Loop around twice in the case Load -> Initialize -> Store. 638 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 639 for (int trip = 0; trip <= 1; trip++) { 640 641 if (st->is_Store()) { 642 Node* st_adr = st->in(MemNode::Address); 643 if (!phase->eqv(st_adr, ld_adr)) { 644 // Try harder before giving up... Match raw and non-raw pointers. 645 intptr_t st_off = 0; 646 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 647 if (alloc == NULL) return NULL; 648 intptr_t ld_off = 0; 649 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 650 if (alloc != allo2) return NULL; 651 if (ld_off != st_off) return NULL; 652 // At this point we have proven something like this setup: 653 // A = Allocate(...) 654 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 655 // S = StoreQ(, AddP(, A.Parm , #Off), V) 656 // (Actually, we haven't yet proven the Q's are the same.) 657 // In other words, we are loading from a casted version of 658 // the same pointer-and-offset that we stored to. 659 // Thus, we are able to replace L by V. 660 } 661 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 662 if (store_Opcode() != st->Opcode()) 663 return NULL; 664 return st->in(MemNode::ValueIn); 665 } 666 667 intptr_t offset = 0; // scratch 668 669 // A load from a freshly-created object always returns zero. 670 // (This can happen after LoadNode::Ideal resets the load's memory input 671 // to find_captured_store, which returned InitializeNode::zero_memory.) 672 if (st->is_Proj() && st->in(0)->is_Allocate() && 673 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 674 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 675 // return a zero value for the load's basic type 676 // (This is one of the few places where a generic PhaseTransform 677 // can create new nodes. Think of it as lazily manifesting 678 // virtually pre-existing constants.) 679 return phase->zerocon(memory_type()); 680 } 681 682 // A load from an initialization barrier can match a captured store. 683 if (st->is_Proj() && st->in(0)->is_Initialize()) { 684 InitializeNode* init = st->in(0)->as_Initialize(); 685 AllocateNode* alloc = init->allocation(); 686 if (alloc != NULL && 687 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 688 // examine a captured store value 689 st = init->find_captured_store(offset, memory_size(), phase); 690 if (st != NULL) 691 continue; // take one more trip around 692 } 693 } 694 695 break; 696 } 697 698 return NULL; 699} 700 701//------------------------------Identity--------------------------------------- 702// Loads are identity if previous store is to same address 703Node *LoadNode::Identity( PhaseTransform *phase ) { 704 // If the previous store-maker is the right kind of Store, and the store is 705 // to the same address, then we are equal to the value stored. 706 Node* mem = in(MemNode::Memory); 707 Node* value = can_see_stored_value(mem, phase); 708 if( value ) { 709 // byte, short & char stores truncate naturally. 710 // A load has to load the truncated value which requires 711 // some sort of masking operation and that requires an 712 // Ideal call instead of an Identity call. 713 if (memory_size() < BytesPerInt) { 714 // If the input to the store does not fit with the load's result type, 715 // it must be truncated via an Ideal call. 716 if (!phase->type(value)->higher_equal(phase->type(this))) 717 return this; 718 } 719 // (This works even when value is a Con, but LoadNode::Value 720 // usually runs first, producing the singleton type of the Con.) 721 return value; 722 } 723 return this; 724} 725 726//------------------------------Ideal------------------------------------------ 727// If the load is from Field memory and the pointer is non-null, we can 728// zero out the control input. 729// If the offset is constant and the base is an object allocation, 730// try to hook me up to the exact initializing store. 731Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 732 Node* p = MemNode::Ideal_common(phase, can_reshape); 733 if (p) return (p == NodeSentinel) ? NULL : p; 734 735 Node* ctrl = in(MemNode::Control); 736 Node* address = in(MemNode::Address); 737 738 // Skip up past a SafePoint control. Cannot do this for Stores because 739 // pointer stores & cardmarks must stay on the same side of a SafePoint. 740 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 741 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 742 ctrl = ctrl->in(0); 743 set_req(MemNode::Control,ctrl); 744 } 745 746 // Check for useless control edge in some common special cases 747 if (in(MemNode::Control) != NULL) { 748 intptr_t ignore = 0; 749 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 750 if (base != NULL 751 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 752 && detect_dominating_control(base->in(0), phase->C->start())) { 753 // A method-invariant, non-null address (constant or 'this' argument). 754 set_req(MemNode::Control, NULL); 755 } 756 } 757 758 // Check for prior store with a different base or offset; make Load 759 // independent. Skip through any number of them. Bail out if the stores 760 // are in an endless dead cycle and report no progress. This is a key 761 // transform for Reflection. However, if after skipping through the Stores 762 // we can't then fold up against a prior store do NOT do the transform as 763 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 764 // array memory alive twice: once for the hoisted Load and again after the 765 // bypassed Store. This situation only works if EVERYBODY who does 766 // anti-dependence work knows how to bypass. I.e. we need all 767 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 768 // the alias index stuff. So instead, peek through Stores and IFF we can 769 // fold up, do so. 770 Node* prev_mem = find_previous_store(phase); 771 // Steps (a), (b): Walk past independent stores to find an exact match. 772 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 773 // (c) See if we can fold up on the spot, but don't fold up here. 774 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or 775 // just return a prior value, which is done by Identity calls. 776 if (can_see_stored_value(prev_mem, phase)) { 777 // Make ready for step (d): 778 set_req(MemNode::Memory, prev_mem); 779 return this; 780 } 781 } 782 783 return NULL; // No further progress 784} 785 786// Helper to recognize certain Klass fields which are invariant across 787// some group of array types (e.g., int[] or all T[] where T < Object). 788const Type* 789LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 790 ciKlass* klass) const { 791 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 792 // The field is Klass::_modifier_flags. Return its (constant) value. 793 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 794 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 795 return TypeInt::make(klass->modifier_flags()); 796 } 797 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 798 // The field is Klass::_access_flags. Return its (constant) value. 799 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 800 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 801 return TypeInt::make(klass->access_flags()); 802 } 803 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 804 // The field is Klass::_layout_helper. Return its constant value if known. 805 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 806 return TypeInt::make(klass->layout_helper()); 807 } 808 809 // No match. 810 return NULL; 811} 812 813//------------------------------Value----------------------------------------- 814const Type *LoadNode::Value( PhaseTransform *phase ) const { 815 // Either input is TOP ==> the result is TOP 816 Node* mem = in(MemNode::Memory); 817 const Type *t1 = phase->type(mem); 818 if (t1 == Type::TOP) return Type::TOP; 819 Node* adr = in(MemNode::Address); 820 const TypePtr* tp = phase->type(adr)->isa_ptr(); 821 if (tp == NULL || tp->empty()) return Type::TOP; 822 int off = tp->offset(); 823 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 824 825 // Try to guess loaded type from pointer type 826 if (tp->base() == Type::AryPtr) { 827 const Type *t = tp->is_aryptr()->elem(); 828 // Don't do this for integer types. There is only potential profit if 829 // the element type t is lower than _type; that is, for int types, if _type is 830 // more restrictive than t. This only happens here if one is short and the other 831 // char (both 16 bits), and in those cases we've made an intentional decision 832 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 833 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 834 // 835 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 836 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 837 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 838 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 839 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 840 // In fact, that could have been the original type of p1, and p1 could have 841 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 842 // expression (LShiftL quux 3) independently optimized to the constant 8. 843 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 844 && Opcode() != Op_LoadKlass) { 845 // t might actually be lower than _type, if _type is a unique 846 // concrete subclass of abstract class t. 847 // Make sure the reference is not into the header, by comparing 848 // the offset against the offset of the start of the array's data. 849 // Different array types begin at slightly different offsets (12 vs. 16). 850 // We choose T_BYTE as an example base type that is least restrictive 851 // as to alignment, which will therefore produce the smallest 852 // possible base offset. 853 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 854 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 855 const Type* jt = t->join(_type); 856 // In any case, do not allow the join, per se, to empty out the type. 857 if (jt->empty() && !t->empty()) { 858 // This can happen if a interface-typed array narrows to a class type. 859 jt = _type; 860 } 861 return jt; 862 } 863 } 864 } else if (tp->base() == Type::InstPtr) { 865 assert( off != Type::OffsetBot || 866 // arrays can be cast to Objects 867 tp->is_oopptr()->klass()->is_java_lang_Object() || 868 // unsafe field access may not have a constant offset 869 phase->C->has_unsafe_access(), 870 "Field accesses must be precise" ); 871 // For oop loads, we expect the _type to be precise 872 } else if (tp->base() == Type::KlassPtr) { 873 assert( off != Type::OffsetBot || 874 // arrays can be cast to Objects 875 tp->is_klassptr()->klass()->is_java_lang_Object() || 876 // also allow array-loading from the primary supertype 877 // array during subtype checks 878 Opcode() == Op_LoadKlass, 879 "Field accesses must be precise" ); 880 // For klass/static loads, we expect the _type to be precise 881 } 882 883 const TypeKlassPtr *tkls = tp->isa_klassptr(); 884 if (tkls != NULL && !StressReflectiveCode) { 885 ciKlass* klass = tkls->klass(); 886 if (klass->is_loaded() && tkls->klass_is_exact()) { 887 // We are loading a field from a Klass metaobject whose identity 888 // is known at compile time (the type is "exact" or "precise"). 889 // Check for fields we know are maintained as constants by the VM. 890 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 891 // The field is Klass::_super_check_offset. Return its (constant) value. 892 // (Folds up type checking code.) 893 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 894 return TypeInt::make(klass->super_check_offset()); 895 } 896 // Compute index into primary_supers array 897 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 898 // Check for overflowing; use unsigned compare to handle the negative case. 899 if( depth < ciKlass::primary_super_limit() ) { 900 // The field is an element of Klass::_primary_supers. Return its (constant) value. 901 // (Folds up type checking code.) 902 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 903 ciKlass *ss = klass->super_of_depth(depth); 904 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 905 } 906 const Type* aift = load_array_final_field(tkls, klass); 907 if (aift != NULL) return aift; 908 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 909 && klass->is_array_klass()) { 910 // The field is arrayKlass::_component_mirror. Return its (constant) value. 911 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 912 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 913 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 914 } 915 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 916 // The field is Klass::_java_mirror. Return its (constant) value. 917 // (Folds up the 2nd indirection in anObjConstant.getClass().) 918 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 919 return TypeInstPtr::make(klass->java_mirror()); 920 } 921 } 922 923 // We can still check if we are loading from the primary_supers array at a 924 // shallow enough depth. Even though the klass is not exact, entries less 925 // than or equal to its super depth are correct. 926 if (klass->is_loaded() ) { 927 ciType *inner = klass->klass(); 928 while( inner->is_obj_array_klass() ) 929 inner = inner->as_obj_array_klass()->base_element_type(); 930 if( inner->is_instance_klass() && 931 !inner->as_instance_klass()->flags().is_interface() ) { 932 // Compute index into primary_supers array 933 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 934 // Check for overflowing; use unsigned compare to handle the negative case. 935 if( depth < ciKlass::primary_super_limit() && 936 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 937 // The field is an element of Klass::_primary_supers. Return its (constant) value. 938 // (Folds up type checking code.) 939 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 940 ciKlass *ss = klass->super_of_depth(depth); 941 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 942 } 943 } 944 } 945 946 // If the type is enough to determine that the thing is not an array, 947 // we can give the layout_helper a positive interval type. 948 // This will help short-circuit some reflective code. 949 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 950 && !klass->is_array_klass() // not directly typed as an array 951 && !klass->is_interface() // specifically not Serializable & Cloneable 952 && !klass->is_java_lang_Object() // not the supertype of all T[] 953 ) { 954 // Note: When interfaces are reliable, we can narrow the interface 955 // test to (klass != Serializable && klass != Cloneable). 956 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 957 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 958 // The key property of this type is that it folds up tests 959 // for array-ness, since it proves that the layout_helper is positive. 960 // Thus, a generic value like the basic object layout helper works fine. 961 return TypeInt::make(min_size, max_jint, Type::WidenMin); 962 } 963 } 964 965 // If we are loading from a freshly-allocated object, produce a zero, 966 // if the load is provably beyond the header of the object. 967 // (Also allow a variable load from a fresh array to produce zero.) 968 if (ReduceFieldZeroing) { 969 Node* value = can_see_stored_value(mem,phase); 970 if (value != NULL && value->is_Con()) 971 return value->bottom_type(); 972 } 973 974 return _type; 975} 976 977//------------------------------match_edge------------------------------------- 978// Do we Match on this edge index or not? Match only the address. 979uint LoadNode::match_edge(uint idx) const { 980 return idx == MemNode::Address; 981} 982 983//--------------------------LoadBNode::Ideal-------------------------------------- 984// 985// If the previous store is to the same address as this load, 986// and the value stored was larger than a byte, replace this load 987// with the value stored truncated to a byte. If no truncation is 988// needed, the replacement is done in LoadNode::Identity(). 989// 990Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 991 Node* mem = in(MemNode::Memory); 992 Node* value = can_see_stored_value(mem,phase); 993 if( value && !phase->type(value)->higher_equal( _type ) ) { 994 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 995 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 996 } 997 // Identity call will handle the case where truncation is not needed. 998 return LoadNode::Ideal(phase, can_reshape); 999} 1000 1001//--------------------------LoadCNode::Ideal-------------------------------------- 1002// 1003// If the previous store is to the same address as this load, 1004// and the value stored was larger than a char, replace this load 1005// with the value stored truncated to a char. If no truncation is 1006// needed, the replacement is done in LoadNode::Identity(). 1007// 1008Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1009 Node* mem = in(MemNode::Memory); 1010 Node* value = can_see_stored_value(mem,phase); 1011 if( value && !phase->type(value)->higher_equal( _type ) ) 1012 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1013 // Identity call will handle the case where truncation is not needed. 1014 return LoadNode::Ideal(phase, can_reshape); 1015} 1016 1017//--------------------------LoadSNode::Ideal-------------------------------------- 1018// 1019// If the previous store is to the same address as this load, 1020// and the value stored was larger than a short, replace this load 1021// with the value stored truncated to a short. If no truncation is 1022// needed, the replacement is done in LoadNode::Identity(). 1023// 1024Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1025 Node* mem = in(MemNode::Memory); 1026 Node* value = can_see_stored_value(mem,phase); 1027 if( value && !phase->type(value)->higher_equal( _type ) ) { 1028 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1029 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1030 } 1031 // Identity call will handle the case where truncation is not needed. 1032 return LoadNode::Ideal(phase, can_reshape); 1033} 1034 1035//============================================================================= 1036//------------------------------Value------------------------------------------ 1037const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1038 // Either input is TOP ==> the result is TOP 1039 const Type *t1 = phase->type( in(MemNode::Memory) ); 1040 if (t1 == Type::TOP) return Type::TOP; 1041 Node *adr = in(MemNode::Address); 1042 const Type *t2 = phase->type( adr ); 1043 if (t2 == Type::TOP) return Type::TOP; 1044 const TypePtr *tp = t2->is_ptr(); 1045 if (TypePtr::above_centerline(tp->ptr()) || 1046 tp->ptr() == TypePtr::Null) return Type::TOP; 1047 1048 // Return a more precise klass, if possible 1049 const TypeInstPtr *tinst = tp->isa_instptr(); 1050 if (tinst != NULL) { 1051 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1052 int offset = tinst->offset(); 1053 if (ik == phase->C->env()->Class_klass() 1054 && (offset == java_lang_Class::klass_offset_in_bytes() || 1055 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1056 // We are loading a special hidden field from a Class mirror object, 1057 // the field which points to the VM's Klass metaobject. 1058 ciType* t = tinst->java_mirror_type(); 1059 // java_mirror_type returns non-null for compile-time Class constants. 1060 if (t != NULL) { 1061 // constant oop => constant klass 1062 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1063 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1064 } 1065 if (!t->is_klass()) { 1066 // a primitive Class (e.g., int.class) has NULL for a klass field 1067 return TypePtr::NULL_PTR; 1068 } 1069 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1070 return TypeKlassPtr::make(t->as_klass()); 1071 } 1072 // non-constant mirror, so we can't tell what's going on 1073 } 1074 if( !ik->is_loaded() ) 1075 return _type; // Bail out if not loaded 1076 if (offset == oopDesc::klass_offset_in_bytes()) { 1077 if (tinst->klass_is_exact()) { 1078 return TypeKlassPtr::make(ik); 1079 } 1080 // See if we can become precise: no subklasses and no interface 1081 // (Note: We need to support verified interfaces.) 1082 if (!ik->is_interface() && !ik->has_subklass()) { 1083 //assert(!UseExactTypes, "this code should be useless with exact types"); 1084 // Add a dependence; if any subclass added we need to recompile 1085 if (!ik->is_final()) { 1086 // %%% should use stronger assert_unique_concrete_subtype instead 1087 phase->C->dependencies()->assert_leaf_type(ik); 1088 } 1089 // Return precise klass 1090 return TypeKlassPtr::make(ik); 1091 } 1092 1093 // Return root of possible klass 1094 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1095 } 1096 } 1097 1098 // Check for loading klass from an array 1099 const TypeAryPtr *tary = tp->isa_aryptr(); 1100 if( tary != NULL ) { 1101 ciKlass *tary_klass = tary->klass(); 1102 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1103 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1104 if (tary->klass_is_exact()) { 1105 return TypeKlassPtr::make(tary_klass); 1106 } 1107 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1108 // If the klass is an object array, we defer the question to the 1109 // array component klass. 1110 if( ak->is_obj_array_klass() ) { 1111 assert( ak->is_loaded(), "" ); 1112 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1113 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1114 ciInstanceKlass* ik = base_k->as_instance_klass(); 1115 // See if we can become precise: no subklasses and no interface 1116 if (!ik->is_interface() && !ik->has_subklass()) { 1117 //assert(!UseExactTypes, "this code should be useless with exact types"); 1118 // Add a dependence; if any subclass added we need to recompile 1119 if (!ik->is_final()) { 1120 phase->C->dependencies()->assert_leaf_type(ik); 1121 } 1122 // Return precise array klass 1123 return TypeKlassPtr::make(ak); 1124 } 1125 } 1126 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1127 } else { // Found a type-array? 1128 //assert(!UseExactTypes, "this code should be useless with exact types"); 1129 assert( ak->is_type_array_klass(), "" ); 1130 return TypeKlassPtr::make(ak); // These are always precise 1131 } 1132 } 1133 } 1134 1135 // Check for loading klass from an array klass 1136 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1137 if (tkls != NULL && !StressReflectiveCode) { 1138 ciKlass* klass = tkls->klass(); 1139 if( !klass->is_loaded() ) 1140 return _type; // Bail out if not loaded 1141 if( klass->is_obj_array_klass() && 1142 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1143 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1144 // // Always returning precise element type is incorrect, 1145 // // e.g., element type could be object and array may contain strings 1146 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1147 1148 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1149 // according to the element type's subclassing. 1150 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1151 } 1152 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1153 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1154 ciKlass* sup = klass->as_instance_klass()->super(); 1155 // The field is Klass::_super. Return its (constant) value. 1156 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1157 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1158 } 1159 } 1160 1161 // Bailout case 1162 return LoadNode::Value(phase); 1163} 1164 1165//------------------------------Identity--------------------------------------- 1166// To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1167// Also feed through the klass in Allocate(...klass...)._klass. 1168Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1169 Node* x = LoadNode::Identity(phase); 1170 if (x != this) return x; 1171 1172 // Take apart the address into an oop and and offset. 1173 // Return 'this' if we cannot. 1174 Node* adr = in(MemNode::Address); 1175 intptr_t offset = 0; 1176 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1177 if (base == NULL) return this; 1178 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1179 if (toop == NULL) return this; 1180 1181 // We can fetch the klass directly through an AllocateNode. 1182 // This works even if the klass is not constant (clone or newArray). 1183 if (offset == oopDesc::klass_offset_in_bytes()) { 1184 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1185 if (allocated_klass != NULL) { 1186 return allocated_klass; 1187 } 1188 } 1189 1190 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1191 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1192 // See inline_native_Class_query for occurrences of these patterns. 1193 // Java Example: x.getClass().isAssignableFrom(y) 1194 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1195 // 1196 // This improves reflective code, often making the Class 1197 // mirror go completely dead. (Current exception: Class 1198 // mirrors may appear in debug info, but we could clean them out by 1199 // introducing a new debug info operator for klassOop.java_mirror). 1200 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1201 && (offset == java_lang_Class::klass_offset_in_bytes() || 1202 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1203 // We are loading a special hidden field from a Class mirror, 1204 // the field which points to its Klass or arrayKlass metaobject. 1205 if (base->is_Load()) { 1206 Node* adr2 = base->in(MemNode::Address); 1207 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1208 if (tkls != NULL && !tkls->empty() 1209 && (tkls->klass()->is_instance_klass() || 1210 tkls->klass()->is_array_klass()) 1211 && adr2->is_AddP() 1212 ) { 1213 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1214 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1215 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1216 } 1217 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1218 return adr2->in(AddPNode::Base); 1219 } 1220 } 1221 } 1222 } 1223 1224 return this; 1225} 1226 1227//------------------------------Value----------------------------------------- 1228const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1229 // Either input is TOP ==> the result is TOP 1230 const Type *t1 = phase->type( in(MemNode::Memory) ); 1231 if( t1 == Type::TOP ) return Type::TOP; 1232 Node *adr = in(MemNode::Address); 1233 const Type *t2 = phase->type( adr ); 1234 if( t2 == Type::TOP ) return Type::TOP; 1235 const TypePtr *tp = t2->is_ptr(); 1236 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 1237 const TypeAryPtr *tap = tp->isa_aryptr(); 1238 if( !tap ) return _type; 1239 return tap->size(); 1240} 1241 1242//------------------------------Identity--------------------------------------- 1243// Feed through the length in AllocateArray(...length...)._length. 1244Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 1245 Node* x = LoadINode::Identity(phase); 1246 if (x != this) return x; 1247 1248 // Take apart the address into an oop and and offset. 1249 // Return 'this' if we cannot. 1250 Node* adr = in(MemNode::Address); 1251 intptr_t offset = 0; 1252 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1253 if (base == NULL) return this; 1254 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1255 if (tary == NULL) return this; 1256 1257 // We can fetch the length directly through an AllocateArrayNode. 1258 // This works even if the length is not constant (clone or newArray). 1259 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1260 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase); 1261 if (allocated_length != NULL) { 1262 return allocated_length; 1263 } 1264 } 1265 1266 return this; 1267 1268} 1269//============================================================================= 1270//---------------------------StoreNode::make----------------------------------- 1271// Polymorphic factory method: 1272StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 1273 switch (bt) { 1274 case T_BOOLEAN: 1275 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 1276 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 1277 case T_CHAR: 1278 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 1279 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 1280 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 1281 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 1282 case T_ADDRESS: 1283 case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 1284 } 1285 ShouldNotReachHere(); 1286 return (StoreNode*)NULL; 1287} 1288 1289StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 1290 bool require_atomic = true; 1291 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 1292} 1293 1294 1295//--------------------------bottom_type---------------------------------------- 1296const Type *StoreNode::bottom_type() const { 1297 return Type::MEMORY; 1298} 1299 1300//------------------------------hash------------------------------------------- 1301uint StoreNode::hash() const { 1302 // unroll addition of interesting fields 1303 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 1304 1305 // Since they are not commoned, do not hash them: 1306 return NO_HASH; 1307} 1308 1309//------------------------------Ideal------------------------------------------ 1310// Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 1311// When a store immediately follows a relevant allocation/initialization, 1312// try to capture it into the initialization, or hoist it above. 1313Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1314 Node* p = MemNode::Ideal_common(phase, can_reshape); 1315 if (p) return (p == NodeSentinel) ? NULL : p; 1316 1317 Node* mem = in(MemNode::Memory); 1318 Node* address = in(MemNode::Address); 1319 1320 // Back-to-back stores to same address? Fold em up. 1321 // Generally unsafe if I have intervening uses... 1322 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 1323 // Looking at a dead closed cycle of memory? 1324 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 1325 1326 assert(Opcode() == mem->Opcode() || 1327 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 1328 "no mismatched stores, except on raw memory"); 1329 1330 if (mem->outcnt() == 1 && // check for intervening uses 1331 mem->as_Store()->memory_size() <= this->memory_size()) { 1332 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 1333 // For example, 'mem' might be the final state at a conditional return. 1334 // Or, 'mem' might be used by some node which is live at the same time 1335 // 'this' is live, which might be unschedulable. So, require exactly 1336 // ONE user, the 'this' store, until such time as we clone 'mem' for 1337 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 1338 if (can_reshape) { // (%%% is this an anachronism?) 1339 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 1340 phase->is_IterGVN()); 1341 } else { 1342 // It's OK to do this in the parser, since DU info is always accurate, 1343 // and the parser always refers to nodes via SafePointNode maps. 1344 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 1345 } 1346 return this; 1347 } 1348 } 1349 1350 // Capture an unaliased, unconditional, simple store into an initializer. 1351 // Or, if it is independent of the allocation, hoist it above the allocation. 1352 if (ReduceFieldZeroing && /*can_reshape &&*/ 1353 mem->is_Proj() && mem->in(0)->is_Initialize()) { 1354 InitializeNode* init = mem->in(0)->as_Initialize(); 1355 intptr_t offset = init->can_capture_store(this, phase); 1356 if (offset > 0) { 1357 Node* moved = init->capture_store(this, offset, phase); 1358 // If the InitializeNode captured me, it made a raw copy of me, 1359 // and I need to disappear. 1360 if (moved != NULL) { 1361 // %%% hack to ensure that Ideal returns a new node: 1362 mem = MergeMemNode::make(phase->C, mem); 1363 return mem; // fold me away 1364 } 1365 } 1366 } 1367 1368 return NULL; // No further progress 1369} 1370 1371//------------------------------Value----------------------------------------- 1372const Type *StoreNode::Value( PhaseTransform *phase ) const { 1373 // Either input is TOP ==> the result is TOP 1374 const Type *t1 = phase->type( in(MemNode::Memory) ); 1375 if( t1 == Type::TOP ) return Type::TOP; 1376 const Type *t2 = phase->type( in(MemNode::Address) ); 1377 if( t2 == Type::TOP ) return Type::TOP; 1378 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 1379 if( t3 == Type::TOP ) return Type::TOP; 1380 return Type::MEMORY; 1381} 1382 1383//------------------------------Identity--------------------------------------- 1384// Remove redundant stores: 1385// Store(m, p, Load(m, p)) changes to m. 1386// Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 1387Node *StoreNode::Identity( PhaseTransform *phase ) { 1388 Node* mem = in(MemNode::Memory); 1389 Node* adr = in(MemNode::Address); 1390 Node* val = in(MemNode::ValueIn); 1391 1392 // Load then Store? Then the Store is useless 1393 if (val->is_Load() && 1394 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 1395 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 1396 val->as_Load()->store_Opcode() == Opcode()) { 1397 return mem; 1398 } 1399 1400 // Two stores in a row of the same value? 1401 if (mem->is_Store() && 1402 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 1403 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 1404 mem->Opcode() == Opcode()) { 1405 return mem; 1406 } 1407 1408 // Store of zero anywhere into a freshly-allocated object? 1409 // Then the store is useless. 1410 // (It must already have been captured by the InitializeNode.) 1411 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 1412 // a newly allocated object is already all-zeroes everywhere 1413 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 1414 return mem; 1415 } 1416 1417 // the store may also apply to zero-bits in an earlier object 1418 Node* prev_mem = find_previous_store(phase); 1419 // Steps (a), (b): Walk past independent stores to find an exact match. 1420 if (prev_mem != NULL) { 1421 Node* prev_val = can_see_stored_value(prev_mem, phase); 1422 if (prev_val != NULL && phase->eqv(prev_val, val)) { 1423 // prev_val and val might differ by a cast; it would be good 1424 // to keep the more informative of the two. 1425 return mem; 1426 } 1427 } 1428 } 1429 1430 return this; 1431} 1432 1433//------------------------------match_edge------------------------------------- 1434// Do we Match on this edge index or not? Match only memory & value 1435uint StoreNode::match_edge(uint idx) const { 1436 return idx == MemNode::Address || idx == MemNode::ValueIn; 1437} 1438 1439//------------------------------cmp-------------------------------------------- 1440// Do not common stores up together. They generally have to be split 1441// back up anyways, so do not bother. 1442uint StoreNode::cmp( const Node &n ) const { 1443 return (&n == this); // Always fail except on self 1444} 1445 1446//------------------------------Ideal_masked_input----------------------------- 1447// Check for a useless mask before a partial-word store 1448// (StoreB ... (AndI valIn conIa) ) 1449// If (conIa & mask == mask) this simplifies to 1450// (StoreB ... (valIn) ) 1451Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 1452 Node *val = in(MemNode::ValueIn); 1453 if( val->Opcode() == Op_AndI ) { 1454 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 1455 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 1456 set_req(MemNode::ValueIn, val->in(1)); 1457 return this; 1458 } 1459 } 1460 return NULL; 1461} 1462 1463 1464//------------------------------Ideal_sign_extended_input---------------------- 1465// Check for useless sign-extension before a partial-word store 1466// (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 1467// If (conIL == conIR && conIR <= num_bits) this simplifies to 1468// (StoreB ... (valIn) ) 1469Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 1470 Node *val = in(MemNode::ValueIn); 1471 if( val->Opcode() == Op_RShiftI ) { 1472 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 1473 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 1474 Node *shl = val->in(1); 1475 if( shl->Opcode() == Op_LShiftI ) { 1476 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 1477 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 1478 set_req(MemNode::ValueIn, shl->in(1)); 1479 return this; 1480 } 1481 } 1482 } 1483 } 1484 return NULL; 1485} 1486 1487//------------------------------value_never_loaded----------------------------------- 1488// Determine whether there are any possible loads of the value stored. 1489// For simplicity, we actually check if there are any loads from the 1490// address stored to, not just for loads of the value stored by this node. 1491// 1492bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 1493 Node *adr = in(Address); 1494 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 1495 if (adr_oop == NULL) 1496 return false; 1497 if (!adr_oop->is_instance()) 1498 return false; // if not a distinct instance, there may be aliases of the address 1499 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 1500 Node *use = adr->fast_out(i); 1501 int opc = use->Opcode(); 1502 if (use->is_Load() || use->is_LoadStore()) { 1503 return false; 1504 } 1505 } 1506 return true; 1507} 1508 1509//============================================================================= 1510//------------------------------Ideal------------------------------------------ 1511// If the store is from an AND mask that leaves the low bits untouched, then 1512// we can skip the AND operation. If the store is from a sign-extension 1513// (a left shift, then right shift) we can skip both. 1514Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1515 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 1516 if( progress != NULL ) return progress; 1517 1518 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 1519 if( progress != NULL ) return progress; 1520 1521 // Finally check the default case 1522 return StoreNode::Ideal(phase, can_reshape); 1523} 1524 1525//============================================================================= 1526//------------------------------Ideal------------------------------------------ 1527// If the store is from an AND mask that leaves the low bits untouched, then 1528// we can skip the AND operation 1529Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1530 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 1531 if( progress != NULL ) return progress; 1532 1533 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 1534 if( progress != NULL ) return progress; 1535 1536 // Finally check the default case 1537 return StoreNode::Ideal(phase, can_reshape); 1538} 1539 1540//============================================================================= 1541//------------------------------Identity--------------------------------------- 1542Node *StoreCMNode::Identity( PhaseTransform *phase ) { 1543 // No need to card mark when storing a null ptr 1544 Node* my_store = in(MemNode::OopStore); 1545 if (my_store->is_Store()) { 1546 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 1547 if( t1 == TypePtr::NULL_PTR ) { 1548 return in(MemNode::Memory); 1549 } 1550 } 1551 return this; 1552} 1553 1554//------------------------------Value----------------------------------------- 1555const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 1556 // If extra input is TOP ==> the result is TOP 1557 const Type *t1 = phase->type( in(MemNode::OopStore) ); 1558 if( t1 == Type::TOP ) return Type::TOP; 1559 1560 return StoreNode::Value( phase ); 1561} 1562 1563 1564//============================================================================= 1565//----------------------------------SCMemProjNode------------------------------ 1566const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 1567{ 1568 return bottom_type(); 1569} 1570 1571//============================================================================= 1572LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 1573 init_req(MemNode::Control, c ); 1574 init_req(MemNode::Memory , mem); 1575 init_req(MemNode::Address, adr); 1576 init_req(MemNode::ValueIn, val); 1577 init_req( ExpectedIn, ex ); 1578 init_class_id(Class_LoadStore); 1579 1580} 1581 1582//============================================================================= 1583//-------------------------------adr_type-------------------------------------- 1584// Do we Match on this edge index or not? Do not match memory 1585const TypePtr* ClearArrayNode::adr_type() const { 1586 Node *adr = in(3); 1587 return MemNode::calculate_adr_type(adr->bottom_type()); 1588} 1589 1590//------------------------------match_edge------------------------------------- 1591// Do we Match on this edge index or not? Do not match memory 1592uint ClearArrayNode::match_edge(uint idx) const { 1593 return idx > 1; 1594} 1595 1596//------------------------------Identity--------------------------------------- 1597// Clearing a zero length array does nothing 1598Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 1599 return phase->type(in(2))->higher_equal(TypeInt::ZERO) ? in(1) : this; 1600} 1601 1602//------------------------------Idealize--------------------------------------- 1603// Clearing a short array is faster with stores 1604Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1605 const int unit = BytesPerLong; 1606 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 1607 if (!t) return NULL; 1608 if (!t->is_con()) return NULL; 1609 intptr_t raw_count = t->get_con(); 1610 intptr_t size = raw_count; 1611 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 1612 // Clearing nothing uses the Identity call. 1613 // Negative clears are possible on dead ClearArrays 1614 // (see jck test stmt114.stmt11402.val). 1615 if (size <= 0 || size % unit != 0) return NULL; 1616 intptr_t count = size / unit; 1617 // Length too long; use fast hardware clear 1618 if (size > Matcher::init_array_short_size) return NULL; 1619 Node *mem = in(1); 1620 if( phase->type(mem)==Type::TOP ) return NULL; 1621 Node *adr = in(3); 1622 const Type* at = phase->type(adr); 1623 if( at==Type::TOP ) return NULL; 1624 const TypePtr* atp = at->isa_ptr(); 1625 // adjust atp to be the correct array element address type 1626 if (atp == NULL) atp = TypePtr::BOTTOM; 1627 else atp = atp->add_offset(Type::OffsetBot); 1628 // Get base for derived pointer purposes 1629 if( adr->Opcode() != Op_AddP ) Unimplemented(); 1630 Node *base = adr->in(1); 1631 1632 Node *zero = phase->makecon(TypeLong::ZERO); 1633 Node *off = phase->MakeConX(BytesPerLong); 1634 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 1635 count--; 1636 while( count-- ) { 1637 mem = phase->transform(mem); 1638 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 1639 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 1640 } 1641 return mem; 1642} 1643 1644//----------------------------clear_memory------------------------------------- 1645// Generate code to initialize object storage to zero. 1646Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 1647 intptr_t start_offset, 1648 Node* end_offset, 1649 PhaseGVN* phase) { 1650 Compile* C = phase->C; 1651 intptr_t offset = start_offset; 1652 1653 int unit = BytesPerLong; 1654 if ((offset % unit) != 0) { 1655 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 1656 adr = phase->transform(adr); 1657 const TypePtr* atp = TypeRawPtr::BOTTOM; 1658 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 1659 mem = phase->transform(mem); 1660 offset += BytesPerInt; 1661 } 1662 assert((offset % unit) == 0, ""); 1663 1664 // Initialize the remaining stuff, if any, with a ClearArray. 1665 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 1666} 1667 1668Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 1669 Node* start_offset, 1670 Node* end_offset, 1671 PhaseGVN* phase) { 1672 Compile* C = phase->C; 1673 int unit = BytesPerLong; 1674 Node* zbase = start_offset; 1675 Node* zend = end_offset; 1676 1677 // Scale to the unit required by the CPU: 1678 if (!Matcher::init_array_count_is_in_bytes) { 1679 Node* shift = phase->intcon(exact_log2(unit)); 1680 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 1681 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 1682 } 1683 1684 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 1685 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 1686 1687 // Bulk clear double-words 1688 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 1689 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 1690 return phase->transform(mem); 1691} 1692 1693Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 1694 intptr_t start_offset, 1695 intptr_t end_offset, 1696 PhaseGVN* phase) { 1697 Compile* C = phase->C; 1698 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 1699 intptr_t done_offset = end_offset; 1700 if ((done_offset % BytesPerLong) != 0) { 1701 done_offset -= BytesPerInt; 1702 } 1703 if (done_offset > start_offset) { 1704 mem = clear_memory(ctl, mem, dest, 1705 start_offset, phase->MakeConX(done_offset), phase); 1706 } 1707 if (done_offset < end_offset) { // emit the final 32-bit store 1708 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 1709 adr = phase->transform(adr); 1710 const TypePtr* atp = TypeRawPtr::BOTTOM; 1711 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 1712 mem = phase->transform(mem); 1713 done_offset += BytesPerInt; 1714 } 1715 assert(done_offset == end_offset, ""); 1716 return mem; 1717} 1718 1719//============================================================================= 1720// Do we match on this edge? No memory edges 1721uint StrCompNode::match_edge(uint idx) const { 1722 return idx == 5 || idx == 6; 1723} 1724 1725//------------------------------Ideal------------------------------------------ 1726// Return a node which is more "ideal" than the current node. Strip out 1727// control copies 1728Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1729 return remove_dead_region(phase, can_reshape) ? this : NULL; 1730} 1731 1732 1733//============================================================================= 1734MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 1735 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 1736 _adr_type(C->get_adr_type(alias_idx)) 1737{ 1738 init_class_id(Class_MemBar); 1739 Node* top = C->top(); 1740 init_req(TypeFunc::I_O,top); 1741 init_req(TypeFunc::FramePtr,top); 1742 init_req(TypeFunc::ReturnAdr,top); 1743 if (precedent != NULL) 1744 init_req(TypeFunc::Parms, precedent); 1745} 1746 1747//------------------------------cmp-------------------------------------------- 1748uint MemBarNode::hash() const { return NO_HASH; } 1749uint MemBarNode::cmp( const Node &n ) const { 1750 return (&n == this); // Always fail except on self 1751} 1752 1753//------------------------------make------------------------------------------- 1754MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 1755 int len = Precedent + (pn == NULL? 0: 1); 1756 switch (opcode) { 1757 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 1758 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 1759 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 1760 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 1761 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 1762 default: ShouldNotReachHere(); return NULL; 1763 } 1764} 1765 1766//------------------------------Ideal------------------------------------------ 1767// Return a node which is more "ideal" than the current node. Strip out 1768// control copies 1769Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1770 if (remove_dead_region(phase, can_reshape)) return this; 1771 return NULL; 1772} 1773 1774//------------------------------Value------------------------------------------ 1775const Type *MemBarNode::Value( PhaseTransform *phase ) const { 1776 if( !in(0) ) return Type::TOP; 1777 if( phase->type(in(0)) == Type::TOP ) 1778 return Type::TOP; 1779 return TypeTuple::MEMBAR; 1780} 1781 1782//------------------------------match------------------------------------------ 1783// Construct projections for memory. 1784Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 1785 switch (proj->_con) { 1786 case TypeFunc::Control: 1787 case TypeFunc::Memory: 1788 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 1789 } 1790 ShouldNotReachHere(); 1791 return NULL; 1792} 1793 1794//===========================InitializeNode==================================== 1795// SUMMARY: 1796// This node acts as a memory barrier on raw memory, after some raw stores. 1797// The 'cooked' oop value feeds from the Initialize, not the Allocation. 1798// The Initialize can 'capture' suitably constrained stores as raw inits. 1799// It can coalesce related raw stores into larger units (called 'tiles'). 1800// It can avoid zeroing new storage for memory units which have raw inits. 1801// At macro-expansion, it is marked 'complete', and does not optimize further. 1802// 1803// EXAMPLE: 1804// The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 1805// ctl = incoming control; mem* = incoming memory 1806// (Note: A star * on a memory edge denotes I/O and other standard edges.) 1807// First allocate uninitialized memory and fill in the header: 1808// alloc = (Allocate ctl mem* 16 #short[].klass ...) 1809// ctl := alloc.Control; mem* := alloc.Memory* 1810// rawmem = alloc.Memory; rawoop = alloc.RawAddress 1811// Then initialize to zero the non-header parts of the raw memory block: 1812// init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 1813// ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 1814// After the initialize node executes, the object is ready for service: 1815// oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 1816// Suppose its body is immediately initialized as {1,2}: 1817// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 1818// store2 = (StoreC init.Control store1 (+ oop 14) 2) 1819// mem.SLICE(#short[*]) := store2 1820// 1821// DETAILS: 1822// An InitializeNode collects and isolates object initialization after 1823// an AllocateNode and before the next possible safepoint. As a 1824// memory barrier (MemBarNode), it keeps critical stores from drifting 1825// down past any safepoint or any publication of the allocation. 1826// Before this barrier, a newly-allocated object may have uninitialized bits. 1827// After this barrier, it may be treated as a real oop, and GC is allowed. 1828// 1829// The semantics of the InitializeNode include an implicit zeroing of 1830// the new object from object header to the end of the object. 1831// (The object header and end are determined by the AllocateNode.) 1832// 1833// Certain stores may be added as direct inputs to the InitializeNode. 1834// These stores must update raw memory, and they must be to addresses 1835// derived from the raw address produced by AllocateNode, and with 1836// a constant offset. They must be ordered by increasing offset. 1837// The first one is at in(RawStores), the last at in(req()-1). 1838// Unlike most memory operations, they are not linked in a chain, 1839// but are displayed in parallel as users of the rawmem output of 1840// the allocation. 1841// 1842// (See comments in InitializeNode::capture_store, which continue 1843// the example given above.) 1844// 1845// When the associated Allocate is macro-expanded, the InitializeNode 1846// may be rewritten to optimize collected stores. A ClearArrayNode 1847// may also be created at that point to represent any required zeroing. 1848// The InitializeNode is then marked 'complete', prohibiting further 1849// capturing of nearby memory operations. 1850// 1851// During macro-expansion, all captured initializations which store 1852// constant values of 32 bits or smaller are coalesced (if advantagous) 1853// into larger 'tiles' 32 or 64 bits. This allows an object to be 1854// initialized in fewer memory operations. Memory words which are 1855// covered by neither tiles nor non-constant stores are pre-zeroed 1856// by explicit stores of zero. (The code shape happens to do all 1857// zeroing first, then all other stores, with both sequences occurring 1858// in order of ascending offsets.) 1859// 1860// Alternatively, code may be inserted between an AllocateNode and its 1861// InitializeNode, to perform arbitrary initialization of the new object. 1862// E.g., the object copying intrinsics insert complex data transfers here. 1863// The initialization must then be marked as 'complete' disable the 1864// built-in zeroing semantics and the collection of initializing stores. 1865// 1866// While an InitializeNode is incomplete, reads from the memory state 1867// produced by it are optimizable if they match the control edge and 1868// new oop address associated with the allocation/initialization. 1869// They return a stored value (if the offset matches) or else zero. 1870// A write to the memory state, if it matches control and address, 1871// and if it is to a constant offset, may be 'captured' by the 1872// InitializeNode. It is cloned as a raw memory operation and rewired 1873// inside the initialization, to the raw oop produced by the allocation. 1874// Operations on addresses which are provably distinct (e.g., to 1875// other AllocateNodes) are allowed to bypass the initialization. 1876// 1877// The effect of all this is to consolidate object initialization 1878// (both arrays and non-arrays, both piecewise and bulk) into a 1879// single location, where it can be optimized as a unit. 1880// 1881// Only stores with an offset less than TrackedInitializationLimit words 1882// will be considered for capture by an InitializeNode. This puts a 1883// reasonable limit on the complexity of optimized initializations. 1884 1885//---------------------------InitializeNode------------------------------------ 1886InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 1887 : _is_complete(false), 1888 MemBarNode(C, adr_type, rawoop) 1889{ 1890 init_class_id(Class_Initialize); 1891 1892 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 1893 assert(in(RawAddress) == rawoop, "proper init"); 1894 // Note: allocation() can be NULL, for secondary initialization barriers 1895} 1896 1897// Since this node is not matched, it will be processed by the 1898// register allocator. Declare that there are no constraints 1899// on the allocation of the RawAddress edge. 1900const RegMask &InitializeNode::in_RegMask(uint idx) const { 1901 // This edge should be set to top, by the set_complete. But be conservative. 1902 if (idx == InitializeNode::RawAddress) 1903 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 1904 return RegMask::Empty; 1905} 1906 1907Node* InitializeNode::memory(uint alias_idx) { 1908 Node* mem = in(Memory); 1909 if (mem->is_MergeMem()) { 1910 return mem->as_MergeMem()->memory_at(alias_idx); 1911 } else { 1912 // incoming raw memory is not split 1913 return mem; 1914 } 1915} 1916 1917bool InitializeNode::is_non_zero() { 1918 if (is_complete()) return false; 1919 remove_extra_zeroes(); 1920 return (req() > RawStores); 1921} 1922 1923void InitializeNode::set_complete(PhaseGVN* phase) { 1924 assert(!is_complete(), "caller responsibility"); 1925 _is_complete = true; 1926 1927 // After this node is complete, it contains a bunch of 1928 // raw-memory initializations. There is no need for 1929 // it to have anything to do with non-raw memory effects. 1930 // Therefore, tell all non-raw users to re-optimize themselves, 1931 // after skipping the memory effects of this initialization. 1932 PhaseIterGVN* igvn = phase->is_IterGVN(); 1933 if (igvn) igvn->add_users_to_worklist(this); 1934} 1935 1936// convenience function 1937// return false if the init contains any stores already 1938bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 1939 InitializeNode* init = initialization(); 1940 if (init == NULL || init->is_complete()) return false; 1941 init->remove_extra_zeroes(); 1942 // for now, if this allocation has already collected any inits, bail: 1943 if (init->is_non_zero()) return false; 1944 init->set_complete(phase); 1945 return true; 1946} 1947 1948void InitializeNode::remove_extra_zeroes() { 1949 if (req() == RawStores) return; 1950 Node* zmem = zero_memory(); 1951 uint fill = RawStores; 1952 for (uint i = fill; i < req(); i++) { 1953 Node* n = in(i); 1954 if (n->is_top() || n == zmem) continue; // skip 1955 if (fill < i) set_req(fill, n); // compact 1956 ++fill; 1957 } 1958 // delete any empty spaces created: 1959 while (fill < req()) { 1960 del_req(fill); 1961 } 1962} 1963 1964// Helper for remembering which stores go with which offsets. 1965intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 1966 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 1967 intptr_t offset = -1; 1968 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 1969 phase, offset); 1970 if (base == NULL) return -1; // something is dead, 1971 if (offset < 0) return -1; // dead, dead 1972 return offset; 1973} 1974 1975// Helper for proving that an initialization expression is 1976// "simple enough" to be folded into an object initialization. 1977// Attempts to prove that a store's initial value 'n' can be captured 1978// within the initialization without creating a vicious cycle, such as: 1979// { Foo p = new Foo(); p.next = p; } 1980// True for constants and parameters and small combinations thereof. 1981bool InitializeNode::detect_init_independence(Node* n, 1982 bool st_is_pinned, 1983 int& count) { 1984 if (n == NULL) return true; // (can this really happen?) 1985 if (n->is_Proj()) n = n->in(0); 1986 if (n == this) return false; // found a cycle 1987 if (n->is_Con()) return true; 1988 if (n->is_Start()) return true; // params, etc., are OK 1989 if (n->is_Root()) return true; // even better 1990 1991 Node* ctl = n->in(0); 1992 if (ctl != NULL && !ctl->is_top()) { 1993 if (ctl->is_Proj()) ctl = ctl->in(0); 1994 if (ctl == this) return false; 1995 1996 // If we already know that the enclosing memory op is pinned right after 1997 // the init, then any control flow that the store has picked up 1998 // must have preceded the init, or else be equal to the init. 1999 // Even after loop optimizations (which might change control edges) 2000 // a store is never pinned *before* the availability of its inputs. 2001 if (!MemNode::detect_dominating_control(ctl, this->in(0))) 2002 return false; // failed to prove a good control 2003 2004 } 2005 2006 // Check data edges for possible dependencies on 'this'. 2007 if ((count += 1) > 20) return false; // complexity limit 2008 for (uint i = 1; i < n->req(); i++) { 2009 Node* m = n->in(i); 2010 if (m == NULL || m == n || m->is_top()) continue; 2011 uint first_i = n->find_edge(m); 2012 if (i != first_i) continue; // process duplicate edge just once 2013 if (!detect_init_independence(m, st_is_pinned, count)) { 2014 return false; 2015 } 2016 } 2017 2018 return true; 2019} 2020 2021// Here are all the checks a Store must pass before it can be moved into 2022// an initialization. Returns zero if a check fails. 2023// On success, returns the (constant) offset to which the store applies, 2024// within the initialized memory. 2025intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2026 const int FAIL = 0; 2027 if (st->req() != MemNode::ValueIn + 1) 2028 return FAIL; // an inscrutable StoreNode (card mark?) 2029 Node* ctl = st->in(MemNode::Control); 2030 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2031 return FAIL; // must be unconditional after the initialization 2032 Node* mem = st->in(MemNode::Memory); 2033 if (!(mem->is_Proj() && mem->in(0) == this)) 2034 return FAIL; // must not be preceded by other stores 2035 Node* adr = st->in(MemNode::Address); 2036 intptr_t offset; 2037 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2038 if (alloc == NULL) 2039 return FAIL; // inscrutable address 2040 if (alloc != allocation()) 2041 return FAIL; // wrong allocation! (store needs to float up) 2042 Node* val = st->in(MemNode::ValueIn); 2043 int complexity_count = 0; 2044 if (!detect_init_independence(val, true, complexity_count)) 2045 return FAIL; // stored value must be 'simple enough' 2046 2047 return offset; // success 2048} 2049 2050// Find the captured store in(i) which corresponds to the range 2051// [start..start+size) in the initialized object. 2052// If there is one, return its index i. If there isn't, return the 2053// negative of the index where it should be inserted. 2054// Return 0 if the queried range overlaps an initialization boundary 2055// or if dead code is encountered. 2056// If size_in_bytes is zero, do not bother with overlap checks. 2057int InitializeNode::captured_store_insertion_point(intptr_t start, 2058 int size_in_bytes, 2059 PhaseTransform* phase) { 2060 const int FAIL = 0, MAX_STORE = BytesPerLong; 2061 2062 if (is_complete()) 2063 return FAIL; // arraycopy got here first; punt 2064 2065 assert(allocation() != NULL, "must be present"); 2066 2067 // no negatives, no header fields: 2068 if (start < (intptr_t) sizeof(oopDesc)) return FAIL; 2069 if (start < (intptr_t) sizeof(arrayOopDesc) && 2070 start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 2071 2072 // after a certain size, we bail out on tracking all the stores: 2073 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2074 if (start >= ti_limit) return FAIL; 2075 2076 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 2077 if (i >= limit) return -(int)i; // not found; here is where to put it 2078 2079 Node* st = in(i); 2080 intptr_t st_off = get_store_offset(st, phase); 2081 if (st_off < 0) { 2082 if (st != zero_memory()) { 2083 return FAIL; // bail out if there is dead garbage 2084 } 2085 } else if (st_off > start) { 2086 // ...we are done, since stores are ordered 2087 if (st_off < start + size_in_bytes) { 2088 return FAIL; // the next store overlaps 2089 } 2090 return -(int)i; // not found; here is where to put it 2091 } else if (st_off < start) { 2092 if (size_in_bytes != 0 && 2093 start < st_off + MAX_STORE && 2094 start < st_off + st->as_Store()->memory_size()) { 2095 return FAIL; // the previous store overlaps 2096 } 2097 } else { 2098 if (size_in_bytes != 0 && 2099 st->as_Store()->memory_size() != size_in_bytes) { 2100 return FAIL; // mismatched store size 2101 } 2102 return i; 2103 } 2104 2105 ++i; 2106 } 2107} 2108 2109// Look for a captured store which initializes at the offset 'start' 2110// with the given size. If there is no such store, and no other 2111// initialization interferes, then return zero_memory (the memory 2112// projection of the AllocateNode). 2113Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 2114 PhaseTransform* phase) { 2115 assert(stores_are_sane(phase), ""); 2116 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2117 if (i == 0) { 2118 return NULL; // something is dead 2119 } else if (i < 0) { 2120 return zero_memory(); // just primordial zero bits here 2121 } else { 2122 Node* st = in(i); // here is the store at this position 2123 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 2124 return st; 2125 } 2126} 2127 2128// Create, as a raw pointer, an address within my new object at 'offset'. 2129Node* InitializeNode::make_raw_address(intptr_t offset, 2130 PhaseTransform* phase) { 2131 Node* addr = in(RawAddress); 2132 if (offset != 0) { 2133 Compile* C = phase->C; 2134 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 2135 phase->MakeConX(offset)) ); 2136 } 2137 return addr; 2138} 2139 2140// Clone the given store, converting it into a raw store 2141// initializing a field or element of my new object. 2142// Caller is responsible for retiring the original store, 2143// with subsume_node or the like. 2144// 2145// From the example above InitializeNode::InitializeNode, 2146// here are the old stores to be captured: 2147// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2148// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2149// 2150// Here is the changed code; note the extra edges on init: 2151// alloc = (Allocate ...) 2152// rawoop = alloc.RawAddress 2153// rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 2154// rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 2155// init = (Initialize alloc.Control alloc.Memory rawoop 2156// rawstore1 rawstore2) 2157// 2158Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 2159 PhaseTransform* phase) { 2160 assert(stores_are_sane(phase), ""); 2161 2162 if (start < 0) return NULL; 2163 assert(can_capture_store(st, phase) == start, "sanity"); 2164 2165 Compile* C = phase->C; 2166 int size_in_bytes = st->memory_size(); 2167 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2168 if (i == 0) return NULL; // bail out 2169 Node* prev_mem = NULL; // raw memory for the captured store 2170 if (i > 0) { 2171 prev_mem = in(i); // there is a pre-existing store under this one 2172 set_req(i, C->top()); // temporarily disconnect it 2173 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 2174 } else { 2175 i = -i; // no pre-existing store 2176 prev_mem = zero_memory(); // a slice of the newly allocated object 2177 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 2178 set_req(--i, C->top()); // reuse this edge; it has been folded away 2179 else 2180 ins_req(i, C->top()); // build a new edge 2181 } 2182 Node* new_st = st->clone(); 2183 new_st->set_req(MemNode::Control, in(Control)); 2184 new_st->set_req(MemNode::Memory, prev_mem); 2185 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 2186 new_st = phase->transform(new_st); 2187 2188 // At this point, new_st might have swallowed a pre-existing store 2189 // at the same offset, or perhaps new_st might have disappeared, 2190 // if it redundantly stored the same value (or zero to fresh memory). 2191 2192 // In any case, wire it in: 2193 set_req(i, new_st); 2194 2195 // The caller may now kill the old guy. 2196 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 2197 assert(check_st == new_st || check_st == NULL, "must be findable"); 2198 assert(!is_complete(), ""); 2199 return new_st; 2200} 2201 2202static bool store_constant(jlong* tiles, int num_tiles, 2203 intptr_t st_off, int st_size, 2204 jlong con) { 2205 if ((st_off & (st_size-1)) != 0) 2206 return false; // strange store offset (assume size==2**N) 2207 address addr = (address)tiles + st_off; 2208 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 2209 switch (st_size) { 2210 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 2211 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 2212 case sizeof(jint): *(jint*) addr = (jint) con; break; 2213 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 2214 default: return false; // strange store size (detect size!=2**N here) 2215 } 2216 return true; // return success to caller 2217} 2218 2219// Coalesce subword constants into int constants and possibly 2220// into long constants. The goal, if the CPU permits, 2221// is to initialize the object with a small number of 64-bit tiles. 2222// Also, convert floating-point constants to bit patterns. 2223// Non-constants are not relevant to this pass. 2224// 2225// In terms of the running example on InitializeNode::InitializeNode 2226// and InitializeNode::capture_store, here is the transformation 2227// of rawstore1 and rawstore2 into rawstore12: 2228// alloc = (Allocate ...) 2229// rawoop = alloc.RawAddress 2230// tile12 = 0x00010002 2231// rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 2232// init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 2233// 2234void 2235InitializeNode::coalesce_subword_stores(intptr_t header_size, 2236 Node* size_in_bytes, 2237 PhaseGVN* phase) { 2238 Compile* C = phase->C; 2239 2240 assert(stores_are_sane(phase), ""); 2241 // Note: After this pass, they are not completely sane, 2242 // since there may be some overlaps. 2243 2244 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 2245 2246 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2247 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 2248 size_limit = MIN2(size_limit, ti_limit); 2249 size_limit = align_size_up(size_limit, BytesPerLong); 2250 int num_tiles = size_limit / BytesPerLong; 2251 2252 // allocate space for the tile map: 2253 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 2254 jlong tiles_buf[small_len]; 2255 Node* nodes_buf[small_len]; 2256 jlong inits_buf[small_len]; 2257 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 2258 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2259 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 2260 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 2261 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 2262 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2263 // tiles: exact bitwise model of all primitive constants 2264 // nodes: last constant-storing node subsumed into the tiles model 2265 // inits: which bytes (in each tile) are touched by any initializations 2266 2267 //// Pass A: Fill in the tile model with any relevant stores. 2268 2269 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 2270 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 2271 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 2272 Node* zmem = zero_memory(); // initially zero memory state 2273 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 2274 Node* st = in(i); 2275 intptr_t st_off = get_store_offset(st, phase); 2276 2277 // Figure out the store's offset and constant value: 2278 if (st_off < header_size) continue; //skip (ignore header) 2279 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 2280 int st_size = st->as_Store()->memory_size(); 2281 if (st_off + st_size > size_limit) break; 2282 2283 // Record which bytes are touched, whether by constant or not. 2284 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 2285 continue; // skip (strange store size) 2286 2287 const Type* val = phase->type(st->in(MemNode::ValueIn)); 2288 if (!val->singleton()) continue; //skip (non-con store) 2289 BasicType type = val->basic_type(); 2290 2291 jlong con = 0; 2292 switch (type) { 2293 case T_INT: con = val->is_int()->get_con(); break; 2294 case T_LONG: con = val->is_long()->get_con(); break; 2295 case T_FLOAT: con = jint_cast(val->getf()); break; 2296 case T_DOUBLE: con = jlong_cast(val->getd()); break; 2297 default: continue; //skip (odd store type) 2298 } 2299 2300 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 2301 st->Opcode() == Op_StoreL) { 2302 continue; // This StoreL is already optimal. 2303 } 2304 2305 // Store down the constant. 2306 store_constant(tiles, num_tiles, st_off, st_size, con); 2307 2308 intptr_t j = st_off >> LogBytesPerLong; 2309 2310 if (type == T_INT && st_size == BytesPerInt 2311 && (st_off & BytesPerInt) == BytesPerInt) { 2312 jlong lcon = tiles[j]; 2313 if (!Matcher::isSimpleConstant64(lcon) && 2314 st->Opcode() == Op_StoreI) { 2315 // This StoreI is already optimal by itself. 2316 jint* intcon = (jint*) &tiles[j]; 2317 intcon[1] = 0; // undo the store_constant() 2318 2319 // If the previous store is also optimal by itself, back up and 2320 // undo the action of the previous loop iteration... if we can. 2321 // But if we can't, just let the previous half take care of itself. 2322 st = nodes[j]; 2323 st_off -= BytesPerInt; 2324 con = intcon[0]; 2325 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 2326 assert(st_off >= header_size, "still ignoring header"); 2327 assert(get_store_offset(st, phase) == st_off, "must be"); 2328 assert(in(i-1) == zmem, "must be"); 2329 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 2330 assert(con == tcon->is_int()->get_con(), "must be"); 2331 // Undo the effects of the previous loop trip, which swallowed st: 2332 intcon[0] = 0; // undo store_constant() 2333 set_req(i-1, st); // undo set_req(i, zmem) 2334 nodes[j] = NULL; // undo nodes[j] = st 2335 --old_subword; // undo ++old_subword 2336 } 2337 continue; // This StoreI is already optimal. 2338 } 2339 } 2340 2341 // This store is not needed. 2342 set_req(i, zmem); 2343 nodes[j] = st; // record for the moment 2344 if (st_size < BytesPerLong) // something has changed 2345 ++old_subword; // includes int/float, but who's counting... 2346 else ++old_long; 2347 } 2348 2349 if ((old_subword + old_long) == 0) 2350 return; // nothing more to do 2351 2352 //// Pass B: Convert any non-zero tiles into optimal constant stores. 2353 // Be sure to insert them before overlapping non-constant stores. 2354 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 2355 for (int j = 0; j < num_tiles; j++) { 2356 jlong con = tiles[j]; 2357 jlong init = inits[j]; 2358 if (con == 0) continue; 2359 jint con0, con1; // split the constant, address-wise 2360 jint init0, init1; // split the init map, address-wise 2361 { union { jlong con; jint intcon[2]; } u; 2362 u.con = con; 2363 con0 = u.intcon[0]; 2364 con1 = u.intcon[1]; 2365 u.con = init; 2366 init0 = u.intcon[0]; 2367 init1 = u.intcon[1]; 2368 } 2369 2370 Node* old = nodes[j]; 2371 assert(old != NULL, "need the prior store"); 2372 intptr_t offset = (j * BytesPerLong); 2373 2374 bool split = !Matcher::isSimpleConstant64(con); 2375 2376 if (offset < header_size) { 2377 assert(offset + BytesPerInt >= header_size, "second int counts"); 2378 assert(*(jint*)&tiles[j] == 0, "junk in header"); 2379 split = true; // only the second word counts 2380 // Example: int a[] = { 42 ... } 2381 } else if (con0 == 0 && init0 == -1) { 2382 split = true; // first word is covered by full inits 2383 // Example: int a[] = { ... foo(), 42 ... } 2384 } else if (con1 == 0 && init1 == -1) { 2385 split = true; // second word is covered by full inits 2386 // Example: int a[] = { ... 42, foo() ... } 2387 } 2388 2389 // Here's a case where init0 is neither 0 nor -1: 2390 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 2391 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 2392 // In this case the tile is not split; it is (jlong)42. 2393 // The big tile is stored down, and then the foo() value is inserted. 2394 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 2395 2396 Node* ctl = old->in(MemNode::Control); 2397 Node* adr = make_raw_address(offset, phase); 2398 const TypePtr* atp = TypeRawPtr::BOTTOM; 2399 2400 // One or two coalesced stores to plop down. 2401 Node* st[2]; 2402 intptr_t off[2]; 2403 int nst = 0; 2404 if (!split) { 2405 ++new_long; 2406 off[nst] = offset; 2407 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, 2408 phase->longcon(con), T_LONG); 2409 } else { 2410 // Omit either if it is a zero. 2411 if (con0 != 0) { 2412 ++new_int; 2413 off[nst] = offset; 2414 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, 2415 phase->intcon(con0), T_INT); 2416 } 2417 if (con1 != 0) { 2418 ++new_int; 2419 offset += BytesPerInt; 2420 adr = make_raw_address(offset, phase); 2421 off[nst] = offset; 2422 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, 2423 phase->intcon(con1), T_INT); 2424 } 2425 } 2426 2427 // Insert second store first, then the first before the second. 2428 // Insert each one just before any overlapping non-constant stores. 2429 while (nst > 0) { 2430 Node* st1 = st[--nst]; 2431 C->copy_node_notes_to(st1, old); 2432 st1 = phase->transform(st1); 2433 offset = off[nst]; 2434 assert(offset >= header_size, "do not smash header"); 2435 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 2436 guarantee(ins_idx != 0, "must re-insert constant store"); 2437 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 2438 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 2439 set_req(--ins_idx, st1); 2440 else 2441 ins_req(ins_idx, st1); 2442 } 2443 } 2444 2445 if (PrintCompilation && WizardMode) 2446 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 2447 old_subword, old_long, new_int, new_long); 2448 if (C->log() != NULL) 2449 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 2450 old_subword, old_long, new_int, new_long); 2451 2452 // Clean up any remaining occurrences of zmem: 2453 remove_extra_zeroes(); 2454} 2455 2456// Explore forward from in(start) to find the first fully initialized 2457// word, and return its offset. Skip groups of subword stores which 2458// together initialize full words. If in(start) is itself part of a 2459// fully initialized word, return the offset of in(start). If there 2460// are no following full-word stores, or if something is fishy, return 2461// a negative value. 2462intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 2463 int int_map = 0; 2464 intptr_t int_map_off = 0; 2465 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 2466 2467 for (uint i = start, limit = req(); i < limit; i++) { 2468 Node* st = in(i); 2469 2470 intptr_t st_off = get_store_offset(st, phase); 2471 if (st_off < 0) break; // return conservative answer 2472 2473 int st_size = st->as_Store()->memory_size(); 2474 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 2475 return st_off; // we found a complete word init 2476 } 2477 2478 // update the map: 2479 2480 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 2481 if (this_int_off != int_map_off) { 2482 // reset the map: 2483 int_map = 0; 2484 int_map_off = this_int_off; 2485 } 2486 2487 int subword_off = st_off - this_int_off; 2488 int_map |= right_n_bits(st_size) << subword_off; 2489 if ((int_map & FULL_MAP) == FULL_MAP) { 2490 return this_int_off; // we found a complete word init 2491 } 2492 2493 // Did this store hit or cross the word boundary? 2494 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 2495 if (next_int_off == this_int_off + BytesPerInt) { 2496 // We passed the current int, without fully initializing it. 2497 int_map_off = next_int_off; 2498 int_map >>= BytesPerInt; 2499 } else if (next_int_off > this_int_off + BytesPerInt) { 2500 // We passed the current and next int. 2501 return this_int_off + BytesPerInt; 2502 } 2503 } 2504 2505 return -1; 2506} 2507 2508 2509// Called when the associated AllocateNode is expanded into CFG. 2510// At this point, we may perform additional optimizations. 2511// Linearize the stores by ascending offset, to make memory 2512// activity as coherent as possible. 2513Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 2514 intptr_t header_size, 2515 Node* size_in_bytes, 2516 PhaseGVN* phase) { 2517 assert(!is_complete(), "not already complete"); 2518 assert(stores_are_sane(phase), ""); 2519 assert(allocation() != NULL, "must be present"); 2520 2521 remove_extra_zeroes(); 2522 2523 if (ReduceFieldZeroing || ReduceBulkZeroing) 2524 // reduce instruction count for common initialization patterns 2525 coalesce_subword_stores(header_size, size_in_bytes, phase); 2526 2527 Node* zmem = zero_memory(); // initially zero memory state 2528 Node* inits = zmem; // accumulating a linearized chain of inits 2529 #ifdef ASSERT 2530 intptr_t last_init_off = sizeof(oopDesc); // previous init offset 2531 intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size 2532 intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size 2533 #endif 2534 intptr_t zeroes_done = header_size; 2535 2536 bool do_zeroing = true; // we might give up if inits are very sparse 2537 int big_init_gaps = 0; // how many large gaps have we seen? 2538 2539 if (ZeroTLAB) do_zeroing = false; 2540 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 2541 2542 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 2543 Node* st = in(i); 2544 intptr_t st_off = get_store_offset(st, phase); 2545 if (st_off < 0) 2546 break; // unknown junk in the inits 2547 if (st->in(MemNode::Memory) != zmem) 2548 break; // complicated store chains somehow in list 2549 2550 int st_size = st->as_Store()->memory_size(); 2551 intptr_t next_init_off = st_off + st_size; 2552 2553 if (do_zeroing && zeroes_done < next_init_off) { 2554 // See if this store needs a zero before it or under it. 2555 intptr_t zeroes_needed = st_off; 2556 2557 if (st_size < BytesPerInt) { 2558 // Look for subword stores which only partially initialize words. 2559 // If we find some, we must lay down some word-level zeroes first, 2560 // underneath the subword stores. 2561 // 2562 // Examples: 2563 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 2564 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 2565 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 2566 // 2567 // Note: coalesce_subword_stores may have already done this, 2568 // if it was prompted by constant non-zero subword initializers. 2569 // But this case can still arise with non-constant stores. 2570 2571 intptr_t next_full_store = find_next_fullword_store(i, phase); 2572 2573 // In the examples above: 2574 // in(i) p q r s x y z 2575 // st_off 12 13 14 15 12 13 14 2576 // st_size 1 1 1 1 1 1 1 2577 // next_full_s. 12 16 16 16 16 16 16 2578 // z's_done 12 16 16 16 12 16 12 2579 // z's_needed 12 16 16 16 16 16 16 2580 // zsize 0 0 0 0 4 0 4 2581 if (next_full_store < 0) { 2582 // Conservative tack: Zero to end of current word. 2583 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 2584 } else { 2585 // Zero to beginning of next fully initialized word. 2586 // Or, don't zero at all, if we are already in that word. 2587 assert(next_full_store >= zeroes_needed, "must go forward"); 2588 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 2589 zeroes_needed = next_full_store; 2590 } 2591 } 2592 2593 if (zeroes_needed > zeroes_done) { 2594 intptr_t zsize = zeroes_needed - zeroes_done; 2595 // Do some incremental zeroing on rawmem, in parallel with inits. 2596 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 2597 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 2598 zeroes_done, zeroes_needed, 2599 phase); 2600 zeroes_done = zeroes_needed; 2601 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 2602 do_zeroing = false; // leave the hole, next time 2603 } 2604 } 2605 2606 // Collect the store and move on: 2607 st->set_req(MemNode::Memory, inits); 2608 inits = st; // put it on the linearized chain 2609 set_req(i, zmem); // unhook from previous position 2610 2611 if (zeroes_done == st_off) 2612 zeroes_done = next_init_off; 2613 2614 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 2615 2616 #ifdef ASSERT 2617 // Various order invariants. Weaker than stores_are_sane because 2618 // a large constant tile can be filled in by smaller non-constant stores. 2619 assert(st_off >= last_init_off, "inits do not reverse"); 2620 last_init_off = st_off; 2621 const Type* val = NULL; 2622 if (st_size >= BytesPerInt && 2623 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 2624 (int)val->basic_type() < (int)T_OBJECT) { 2625 assert(st_off >= last_tile_end, "tiles do not overlap"); 2626 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 2627 last_tile_end = MAX2(last_tile_end, next_init_off); 2628 } else { 2629 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 2630 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 2631 assert(st_off >= last_init_end, "inits do not overlap"); 2632 last_init_end = next_init_off; // it's a non-tile 2633 } 2634 #endif //ASSERT 2635 } 2636 2637 remove_extra_zeroes(); // clear out all the zmems left over 2638 add_req(inits); 2639 2640 if (!ZeroTLAB) { 2641 // If anything remains to be zeroed, zero it all now. 2642 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 2643 // if it is the last unused 4 bytes of an instance, forget about it 2644 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 2645 if (zeroes_done + BytesPerLong >= size_limit) { 2646 assert(allocation() != NULL, ""); 2647 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 2648 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 2649 if (zeroes_done == k->layout_helper()) 2650 zeroes_done = size_limit; 2651 } 2652 if (zeroes_done < size_limit) { 2653 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 2654 zeroes_done, size_in_bytes, phase); 2655 } 2656 } 2657 2658 set_complete(phase); 2659 return rawmem; 2660} 2661 2662 2663#ifdef ASSERT 2664bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 2665 if (is_complete()) 2666 return true; // stores could be anything at this point 2667 intptr_t last_off = sizeof(oopDesc); 2668 for (uint i = InitializeNode::RawStores; i < req(); i++) { 2669 Node* st = in(i); 2670 intptr_t st_off = get_store_offset(st, phase); 2671 if (st_off < 0) continue; // ignore dead garbage 2672 if (last_off > st_off) { 2673 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 2674 this->dump(2); 2675 assert(false, "ascending store offsets"); 2676 return false; 2677 } 2678 last_off = st_off + st->as_Store()->memory_size(); 2679 } 2680 return true; 2681} 2682#endif //ASSERT 2683 2684 2685 2686 2687//============================MergeMemNode===================================== 2688// 2689// SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 2690// contributing store or call operations. Each contributor provides the memory 2691// state for a particular "alias type" (see Compile::alias_type). For example, 2692// if a MergeMem has an input X for alias category #6, then any memory reference 2693// to alias category #6 may use X as its memory state input, as an exact equivalent 2694// to using the MergeMem as a whole. 2695// Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 2696// 2697// (Here, the <N> notation gives the index of the relevant adr_type.) 2698// 2699// In one special case (and more cases in the future), alias categories overlap. 2700// The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 2701// states. Therefore, if a MergeMem has only one contributing input W for Bot, 2702// it is exactly equivalent to that state W: 2703// MergeMem(<Bot>: W) <==> W 2704// 2705// Usually, the merge has more than one input. In that case, where inputs 2706// overlap (i.e., one is Bot), the narrower alias type determines the memory 2707// state for that type, and the wider alias type (Bot) fills in everywhere else: 2708// Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 2709// Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 2710// 2711// A merge can take a "wide" memory state as one of its narrow inputs. 2712// This simply means that the merge observes out only the relevant parts of 2713// the wide input. That is, wide memory states arriving at narrow merge inputs 2714// are implicitly "filtered" or "sliced" as necessary. (This is rare.) 2715// 2716// These rules imply that MergeMem nodes may cascade (via their <Bot> links), 2717// and that memory slices "leak through": 2718// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 2719// 2720// But, in such a cascade, repeated memory slices can "block the leak": 2721// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 2722// 2723// In the last example, Y is not part of the combined memory state of the 2724// outermost MergeMem. The system must, of course, prevent unschedulable 2725// memory states from arising, so you can be sure that the state Y is somehow 2726// a precursor to state Y'. 2727// 2728// 2729// REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 2730// of each MergeMemNode array are exactly the numerical alias indexes, including 2731// but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 2732// Compile::alias_type (and kin) produce and manage these indexes. 2733// 2734// By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 2735// (Note that this provides quick access to the top node inside MergeMem methods, 2736// without the need to reach out via TLS to Compile::current.) 2737// 2738// As a consequence of what was just described, a MergeMem that represents a full 2739// memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 2740// containing all alias categories. 2741// 2742// MergeMem nodes never (?) have control inputs, so in(0) is NULL. 2743// 2744// All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 2745// a memory state for the alias type <N>, or else the top node, meaning that 2746// there is no particular input for that alias type. Note that the length of 2747// a MergeMem is variable, and may be extended at any time to accommodate new 2748// memory states at larger alias indexes. When merges grow, they are of course 2749// filled with "top" in the unused in() positions. 2750// 2751// This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 2752// (Top was chosen because it works smoothly with passes like GCM.) 2753// 2754// For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 2755// the type of random VM bits like TLS references.) Since it is always the 2756// first non-Bot memory slice, some low-level loops use it to initialize an 2757// index variable: for (i = AliasIdxRaw; i < req(); i++). 2758// 2759// 2760// ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 2761// the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 2762// the memory state for alias type <N>, or (if there is no particular slice at <N>, 2763// it returns the base memory. To prevent bugs, memory_at does not accept <Top> 2764// or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 2765// MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 2766// 2767// %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 2768// really that different from the other memory inputs. An abbreviation called 2769// "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 2770// 2771// 2772// PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 2773// partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 2774// that "emerges though" the base memory will be marked as excluding the alias types 2775// of the other (narrow-memory) copies which "emerged through" the narrow edges: 2776// 2777// Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 2778// ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 2779// 2780// This strange "subtraction" effect is necessary to ensure IGVN convergence. 2781// (It is currently unimplemented.) As you can see, the resulting merge is 2782// actually a disjoint union of memory states, rather than an overlay. 2783// 2784 2785//------------------------------MergeMemNode----------------------------------- 2786Node* MergeMemNode::make_empty_memory() { 2787 Node* empty_memory = (Node*) Compile::current()->top(); 2788 assert(empty_memory->is_top(), "correct sentinel identity"); 2789 return empty_memory; 2790} 2791 2792MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 2793 init_class_id(Class_MergeMem); 2794 // all inputs are nullified in Node::Node(int) 2795 // set_input(0, NULL); // no control input 2796 2797 // Initialize the edges uniformly to top, for starters. 2798 Node* empty_mem = make_empty_memory(); 2799 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 2800 init_req(i,empty_mem); 2801 } 2802 assert(empty_memory() == empty_mem, ""); 2803 2804 if( new_base != NULL && new_base->is_MergeMem() ) { 2805 MergeMemNode* mdef = new_base->as_MergeMem(); 2806 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 2807 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 2808 mms.set_memory(mms.memory2()); 2809 } 2810 assert(base_memory() == mdef->base_memory(), ""); 2811 } else { 2812 set_base_memory(new_base); 2813 } 2814} 2815 2816// Make a new, untransformed MergeMem with the same base as 'mem'. 2817// If mem is itself a MergeMem, populate the result with the same edges. 2818MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 2819 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 2820} 2821 2822//------------------------------cmp-------------------------------------------- 2823uint MergeMemNode::hash() const { return NO_HASH; } 2824uint MergeMemNode::cmp( const Node &n ) const { 2825 return (&n == this); // Always fail except on self 2826} 2827 2828//------------------------------Identity--------------------------------------- 2829Node* MergeMemNode::Identity(PhaseTransform *phase) { 2830 // Identity if this merge point does not record any interesting memory 2831 // disambiguations. 2832 Node* base_mem = base_memory(); 2833 Node* empty_mem = empty_memory(); 2834 if (base_mem != empty_mem) { // Memory path is not dead? 2835 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 2836 Node* mem = in(i); 2837 if (mem != empty_mem && mem != base_mem) { 2838 return this; // Many memory splits; no change 2839 } 2840 } 2841 } 2842 return base_mem; // No memory splits; ID on the one true input 2843} 2844 2845//------------------------------Ideal------------------------------------------ 2846// This method is invoked recursively on chains of MergeMem nodes 2847Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2848 // Remove chain'd MergeMems 2849 // 2850 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 2851 // relative to the "in(Bot)". Since we are patching both at the same time, 2852 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 2853 // but rewrite each "in(i)" relative to the new "in(Bot)". 2854 Node *progress = NULL; 2855 2856 2857 Node* old_base = base_memory(); 2858 Node* empty_mem = empty_memory(); 2859 if (old_base == empty_mem) 2860 return NULL; // Dead memory path. 2861 2862 MergeMemNode* old_mbase; 2863 if (old_base != NULL && old_base->is_MergeMem()) 2864 old_mbase = old_base->as_MergeMem(); 2865 else 2866 old_mbase = NULL; 2867 Node* new_base = old_base; 2868 2869 // simplify stacked MergeMems in base memory 2870 if (old_mbase) new_base = old_mbase->base_memory(); 2871 2872 // the base memory might contribute new slices beyond my req() 2873 if (old_mbase) grow_to_match(old_mbase); 2874 2875 // Look carefully at the base node if it is a phi. 2876 PhiNode* phi_base; 2877 if (new_base != NULL && new_base->is_Phi()) 2878 phi_base = new_base->as_Phi(); 2879 else 2880 phi_base = NULL; 2881 2882 Node* phi_reg = NULL; 2883 uint phi_len = (uint)-1; 2884 if (phi_base != NULL && !phi_base->is_copy()) { 2885 // do not examine phi if degraded to a copy 2886 phi_reg = phi_base->region(); 2887 phi_len = phi_base->req(); 2888 // see if the phi is unfinished 2889 for (uint i = 1; i < phi_len; i++) { 2890 if (phi_base->in(i) == NULL) { 2891 // incomplete phi; do not look at it yet! 2892 phi_reg = NULL; 2893 phi_len = (uint)-1; 2894 break; 2895 } 2896 } 2897 } 2898 2899 // Note: We do not call verify_sparse on entry, because inputs 2900 // can normalize to the base_memory via subsume_node or similar 2901 // mechanisms. This method repairs that damage. 2902 2903 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 2904 2905 // Look at each slice. 2906 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 2907 Node* old_in = in(i); 2908 // calculate the old memory value 2909 Node* old_mem = old_in; 2910 if (old_mem == empty_mem) old_mem = old_base; 2911 assert(old_mem == memory_at(i), ""); 2912 2913 // maybe update (reslice) the old memory value 2914 2915 // simplify stacked MergeMems 2916 Node* new_mem = old_mem; 2917 MergeMemNode* old_mmem; 2918 if (old_mem != NULL && old_mem->is_MergeMem()) 2919 old_mmem = old_mem->as_MergeMem(); 2920 else 2921 old_mmem = NULL; 2922 if (old_mmem == this) { 2923 // This can happen if loops break up and safepoints disappear. 2924 // A merge of BotPtr (default) with a RawPtr memory derived from a 2925 // safepoint can be rewritten to a merge of the same BotPtr with 2926 // the BotPtr phi coming into the loop. If that phi disappears 2927 // also, we can end up with a self-loop of the mergemem. 2928 // In general, if loops degenerate and memory effects disappear, 2929 // a mergemem can be left looking at itself. This simply means 2930 // that the mergemem's default should be used, since there is 2931 // no longer any apparent effect on this slice. 2932 // Note: If a memory slice is a MergeMem cycle, it is unreachable 2933 // from start. Update the input to TOP. 2934 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 2935 } 2936 else if (old_mmem != NULL) { 2937 new_mem = old_mmem->memory_at(i); 2938 } 2939 // else preceeding memory was not a MergeMem 2940 2941 // replace equivalent phis (unfortunately, they do not GVN together) 2942 if (new_mem != NULL && new_mem != new_base && 2943 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 2944 if (new_mem->is_Phi()) { 2945 PhiNode* phi_mem = new_mem->as_Phi(); 2946 for (uint i = 1; i < phi_len; i++) { 2947 if (phi_base->in(i) != phi_mem->in(i)) { 2948 phi_mem = NULL; 2949 break; 2950 } 2951 } 2952 if (phi_mem != NULL) { 2953 // equivalent phi nodes; revert to the def 2954 new_mem = new_base; 2955 } 2956 } 2957 } 2958 2959 // maybe store down a new value 2960 Node* new_in = new_mem; 2961 if (new_in == new_base) new_in = empty_mem; 2962 2963 if (new_in != old_in) { 2964 // Warning: Do not combine this "if" with the previous "if" 2965 // A memory slice might have be be rewritten even if it is semantically 2966 // unchanged, if the base_memory value has changed. 2967 set_req(i, new_in); 2968 progress = this; // Report progress 2969 } 2970 } 2971 2972 if (new_base != old_base) { 2973 set_req(Compile::AliasIdxBot, new_base); 2974 // Don't use set_base_memory(new_base), because we need to update du. 2975 assert(base_memory() == new_base, ""); 2976 progress = this; 2977 } 2978 2979 if( base_memory() == this ) { 2980 // a self cycle indicates this memory path is dead 2981 set_req(Compile::AliasIdxBot, empty_mem); 2982 } 2983 2984 // Resolve external cycles by calling Ideal on a MergeMem base_memory 2985 // Recursion must occur after the self cycle check above 2986 if( base_memory()->is_MergeMem() ) { 2987 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 2988 Node *m = phase->transform(new_mbase); // Rollup any cycles 2989 if( m != NULL && (m->is_top() || 2990 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 2991 // propagate rollup of dead cycle to self 2992 set_req(Compile::AliasIdxBot, empty_mem); 2993 } 2994 } 2995 2996 if( base_memory() == empty_mem ) { 2997 progress = this; 2998 // Cut inputs during Parse phase only. 2999 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3000 if( !can_reshape ) { 3001 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3002 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3003 } 3004 } 3005 } 3006 3007 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3008 // Check if PhiNode::Ideal's "Split phis through memory merges" 3009 // transform should be attempted. Look for this->phi->this cycle. 3010 uint merge_width = req(); 3011 if (merge_width > Compile::AliasIdxRaw) { 3012 PhiNode* phi = base_memory()->as_Phi(); 3013 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3014 if (phi->in(i) == this) { 3015 phase->is_IterGVN()->_worklist.push(phi); 3016 break; 3017 } 3018 } 3019 } 3020 } 3021 3022 assert(verify_sparse(), "please, no dups of base"); 3023 return progress; 3024} 3025 3026//-------------------------set_base_memory------------------------------------- 3027void MergeMemNode::set_base_memory(Node *new_base) { 3028 Node* empty_mem = empty_memory(); 3029 set_req(Compile::AliasIdxBot, new_base); 3030 assert(memory_at(req()) == new_base, "must set default memory"); 3031 // Clear out other occurrences of new_base: 3032 if (new_base != empty_mem) { 3033 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3034 if (in(i) == new_base) set_req(i, empty_mem); 3035 } 3036 } 3037} 3038 3039//------------------------------out_RegMask------------------------------------ 3040const RegMask &MergeMemNode::out_RegMask() const { 3041 return RegMask::Empty; 3042} 3043 3044//------------------------------dump_spec-------------------------------------- 3045#ifndef PRODUCT 3046void MergeMemNode::dump_spec(outputStream *st) const { 3047 st->print(" {"); 3048 Node* base_mem = base_memory(); 3049 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3050 Node* mem = memory_at(i); 3051 if (mem == base_mem) { st->print(" -"); continue; } 3052 st->print( " N%d:", mem->_idx ); 3053 Compile::current()->get_adr_type(i)->dump_on(st); 3054 } 3055 st->print(" }"); 3056} 3057#endif // !PRODUCT 3058 3059 3060#ifdef ASSERT 3061static bool might_be_same(Node* a, Node* b) { 3062 if (a == b) return true; 3063 if (!(a->is_Phi() || b->is_Phi())) return false; 3064 // phis shift around during optimization 3065 return true; // pretty stupid... 3066} 3067 3068// verify a narrow slice (either incoming or outgoing) 3069static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 3070 if (!VerifyAliases) return; // don't bother to verify unless requested 3071 if (is_error_reported()) return; // muzzle asserts when debugging an error 3072 if (Node::in_dump()) return; // muzzle asserts when printing 3073 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 3074 assert(n != NULL, ""); 3075 // Elide intervening MergeMem's 3076 while (n->is_MergeMem()) { 3077 n = n->as_MergeMem()->memory_at(alias_idx); 3078 } 3079 Compile* C = Compile::current(); 3080 const TypePtr* n_adr_type = n->adr_type(); 3081 if (n == m->empty_memory()) { 3082 // Implicit copy of base_memory() 3083 } else if (n_adr_type != TypePtr::BOTTOM) { 3084 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 3085 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 3086 } else { 3087 // A few places like make_runtime_call "know" that VM calls are narrow, 3088 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 3089 bool expected_wide_mem = false; 3090 if (n == m->base_memory()) { 3091 expected_wide_mem = true; 3092 } else if (alias_idx == Compile::AliasIdxRaw || 3093 n == m->memory_at(Compile::AliasIdxRaw)) { 3094 expected_wide_mem = true; 3095 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 3096 // memory can "leak through" calls on channels that 3097 // are write-once. Allow this also. 3098 expected_wide_mem = true; 3099 } 3100 assert(expected_wide_mem, "expected narrow slice replacement"); 3101 } 3102} 3103#else // !ASSERT 3104#define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 3105#endif 3106 3107 3108//-----------------------------memory_at--------------------------------------- 3109Node* MergeMemNode::memory_at(uint alias_idx) const { 3110 assert(alias_idx >= Compile::AliasIdxRaw || 3111 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 3112 "must avoid base_memory and AliasIdxTop"); 3113 3114 // Otherwise, it is a narrow slice. 3115 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 3116 Compile *C = Compile::current(); 3117 if (is_empty_memory(n)) { 3118 // the array is sparse; empty slots are the "top" node 3119 n = base_memory(); 3120 assert(Node::in_dump() 3121 || n == NULL || n->bottom_type() == Type::TOP 3122 || n->adr_type() == TypePtr::BOTTOM 3123 || n->adr_type() == TypeRawPtr::BOTTOM 3124 || Compile::current()->AliasLevel() == 0, 3125 "must be a wide memory"); 3126 // AliasLevel == 0 if we are organizing the memory states manually. 3127 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 3128 } else { 3129 // make sure the stored slice is sane 3130 #ifdef ASSERT 3131 if (is_error_reported() || Node::in_dump()) { 3132 } else if (might_be_same(n, base_memory())) { 3133 // Give it a pass: It is a mostly harmless repetition of the base. 3134 // This can arise normally from node subsumption during optimization. 3135 } else { 3136 verify_memory_slice(this, alias_idx, n); 3137 } 3138 #endif 3139 } 3140 return n; 3141} 3142 3143//---------------------------set_memory_at------------------------------------- 3144void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 3145 verify_memory_slice(this, alias_idx, n); 3146 Node* empty_mem = empty_memory(); 3147 if (n == base_memory()) n = empty_mem; // collapse default 3148 uint need_req = alias_idx+1; 3149 if (req() < need_req) { 3150 if (n == empty_mem) return; // already the default, so do not grow me 3151 // grow the sparse array 3152 do { 3153 add_req(empty_mem); 3154 } while (req() < need_req); 3155 } 3156 set_req( alias_idx, n ); 3157} 3158 3159 3160 3161//--------------------------iteration_setup------------------------------------ 3162void MergeMemNode::iteration_setup(const MergeMemNode* other) { 3163 if (other != NULL) { 3164 grow_to_match(other); 3165 // invariant: the finite support of mm2 is within mm->req() 3166 #ifdef ASSERT 3167 for (uint i = req(); i < other->req(); i++) { 3168 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 3169 } 3170 #endif 3171 } 3172 // Replace spurious copies of base_memory by top. 3173 Node* base_mem = base_memory(); 3174 if (base_mem != NULL && !base_mem->is_top()) { 3175 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 3176 if (in(i) == base_mem) 3177 set_req(i, empty_memory()); 3178 } 3179 } 3180} 3181 3182//---------------------------grow_to_match------------------------------------- 3183void MergeMemNode::grow_to_match(const MergeMemNode* other) { 3184 Node* empty_mem = empty_memory(); 3185 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 3186 // look for the finite support of the other memory 3187 for (uint i = other->req(); --i >= req(); ) { 3188 if (other->in(i) != empty_mem) { 3189 uint new_len = i+1; 3190 while (req() < new_len) add_req(empty_mem); 3191 break; 3192 } 3193 } 3194} 3195 3196//---------------------------verify_sparse------------------------------------- 3197#ifndef PRODUCT 3198bool MergeMemNode::verify_sparse() const { 3199 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 3200 Node* base_mem = base_memory(); 3201 // The following can happen in degenerate cases, since empty==top. 3202 if (is_empty_memory(base_mem)) return true; 3203 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3204 assert(in(i) != NULL, "sane slice"); 3205 if (in(i) == base_mem) return false; // should have been the sentinel value! 3206 } 3207 return true; 3208} 3209 3210bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 3211 Node* n; 3212 n = mm->in(idx); 3213 if (mem == n) return true; // might be empty_memory() 3214 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 3215 if (mem == n) return true; 3216 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 3217 if (mem == n) return true; 3218 if (n == NULL) break; 3219 } 3220 return false; 3221} 3222#endif // !PRODUCT 3223