memnode.cpp revision 168:7793bd37a336
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 32static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 33 34//============================================================================= 35uint MemNode::size_of() const { return sizeof(*this); } 36 37const TypePtr *MemNode::adr_type() const { 38 Node* adr = in(Address); 39 const TypePtr* cross_check = NULL; 40 DEBUG_ONLY(cross_check = _adr_type); 41 return calculate_adr_type(adr->bottom_type(), cross_check); 42} 43 44#ifndef PRODUCT 45void MemNode::dump_spec(outputStream *st) const { 46 if (in(Address) == NULL) return; // node is dead 47#ifndef ASSERT 48 // fake the missing field 49 const TypePtr* _adr_type = NULL; 50 if (in(Address) != NULL) 51 _adr_type = in(Address)->bottom_type()->isa_ptr(); 52#endif 53 dump_adr_type(this, _adr_type, st); 54 55 Compile* C = Compile::current(); 56 if( C->alias_type(_adr_type)->is_volatile() ) 57 st->print(" Volatile!"); 58} 59 60void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 61 st->print(" @"); 62 if (adr_type == NULL) { 63 st->print("NULL"); 64 } else { 65 adr_type->dump_on(st); 66 Compile* C = Compile::current(); 67 Compile::AliasType* atp = NULL; 68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 69 if (atp == NULL) 70 st->print(", idx=?\?;"); 71 else if (atp->index() == Compile::AliasIdxBot) 72 st->print(", idx=Bot;"); 73 else if (atp->index() == Compile::AliasIdxTop) 74 st->print(", idx=Top;"); 75 else if (atp->index() == Compile::AliasIdxRaw) 76 st->print(", idx=Raw;"); 77 else { 78 ciField* field = atp->field(); 79 if (field) { 80 st->print(", name="); 81 field->print_name_on(st); 82 } 83 st->print(", idx=%d;", atp->index()); 84 } 85 } 86} 87 88extern void print_alias_types(); 89 90#endif 91 92Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 93 const TypeOopPtr *tinst = t_adr->isa_oopptr(); 94 if (tinst == NULL || !tinst->is_instance_field()) 95 return mchain; // don't try to optimize non-instance types 96 uint instance_id = tinst->instance_id(); 97 Node *prev = NULL; 98 Node *result = mchain; 99 while (prev != result) { 100 prev = result; 101 // skip over a call which does not affect this memory slice 102 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 103 Node *proj_in = result->in(0); 104 if (proj_in->is_Call()) { 105 CallNode *call = proj_in->as_Call(); 106 if (!call->may_modify(t_adr, phase)) { 107 result = call->in(TypeFunc::Memory); 108 } 109 } else if (proj_in->is_Initialize()) { 110 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 111 // Stop if this is the initialization for the object instance which 112 // which contains this memory slice, otherwise skip over it. 113 if (alloc != NULL && alloc->_idx != instance_id) { 114 result = proj_in->in(TypeFunc::Memory); 115 } 116 } else if (proj_in->is_MemBar()) { 117 result = proj_in->in(TypeFunc::Memory); 118 } 119 } else if (result->is_MergeMem()) { 120 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 121 } 122 } 123 return result; 124} 125 126Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 127 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 128 bool is_instance = (t_oop != NULL) && t_oop->is_instance_field(); 129 PhaseIterGVN *igvn = phase->is_IterGVN(); 130 Node *result = mchain; 131 result = optimize_simple_memory_chain(result, t_adr, phase); 132 if (is_instance && igvn != NULL && result->is_Phi()) { 133 PhiNode *mphi = result->as_Phi(); 134 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 135 const TypePtr *t = mphi->adr_type(); 136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 137 t->isa_oopptr() && !t->is_oopptr()->is_instance() && 138 t->is_oopptr()->cast_to_instance(t_oop->instance_id()) == t_oop) { 139 // clone the Phi with our address type 140 result = mphi->split_out_instance(t_adr, igvn); 141 } else { 142 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 143 } 144 } 145 return result; 146} 147 148static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 149 uint alias_idx = phase->C->get_alias_index(tp); 150 Node *mem = mmem; 151#ifdef ASSERT 152 { 153 // Check that current type is consistent with the alias index used during graph construction 154 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 155 bool consistent = adr_check == NULL || adr_check->empty() || 156 phase->C->must_alias(adr_check, alias_idx ); 157 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 158 if( !consistent && adr_check != NULL && !adr_check->empty() && 159 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 160 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 161 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 162 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 163 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 164 // don't assert if it is dead code. 165 consistent = true; 166 } 167 if( !consistent ) { 168 st->print("alias_idx==%d, adr_check==", alias_idx); 169 if( adr_check == NULL ) { 170 st->print("NULL"); 171 } else { 172 adr_check->dump(); 173 } 174 st->cr(); 175 print_alias_types(); 176 assert(consistent, "adr_check must match alias idx"); 177 } 178 } 179#endif 180 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 181 // means an array I have not precisely typed yet. Do not do any 182 // alias stuff with it any time soon. 183 const TypeOopPtr *tinst = tp->isa_oopptr(); 184 if( tp->base() != Type::AnyPtr && 185 !(tinst && 186 tinst->klass()->is_java_lang_Object() && 187 tinst->offset() == Type::OffsetBot) ) { 188 // compress paths and change unreachable cycles to TOP 189 // If not, we can update the input infinitely along a MergeMem cycle 190 // Equivalent code in PhiNode::Ideal 191 Node* m = phase->transform(mmem); 192 // If tranformed to a MergeMem, get the desired slice 193 // Otherwise the returned node represents memory for every slice 194 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 195 // Update input if it is progress over what we have now 196 } 197 return mem; 198} 199 200//--------------------------Ideal_common--------------------------------------- 201// Look for degenerate control and memory inputs. Bypass MergeMem inputs. 202// Unhook non-raw memories from complete (macro-expanded) initializations. 203Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 204 // If our control input is a dead region, kill all below the region 205 Node *ctl = in(MemNode::Control); 206 if (ctl && remove_dead_region(phase, can_reshape)) 207 return this; 208 209 // Ignore if memory is dead, or self-loop 210 Node *mem = in(MemNode::Memory); 211 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 212 assert( mem != this, "dead loop in MemNode::Ideal" ); 213 214 Node *address = in(MemNode::Address); 215 const Type *t_adr = phase->type( address ); 216 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 217 218 // Avoid independent memory operations 219 Node* old_mem = mem; 220 221 // The code which unhooks non-raw memories from complete (macro-expanded) 222 // initializations was removed. After macro-expansion all stores catched 223 // by Initialize node became raw stores and there is no information 224 // which memory slices they modify. So it is unsafe to move any memory 225 // operation above these stores. Also in most cases hooked non-raw memories 226 // were already unhooked by using information from detect_ptr_independence() 227 // and find_previous_store(). 228 229 if (mem->is_MergeMem()) { 230 MergeMemNode* mmem = mem->as_MergeMem(); 231 const TypePtr *tp = t_adr->is_ptr(); 232 233 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 234 } 235 236 if (mem != old_mem) { 237 set_req(MemNode::Memory, mem); 238 return this; 239 } 240 241 // let the subclass continue analyzing... 242 return NULL; 243} 244 245// Helper function for proving some simple control dominations. 246// Attempt to prove that all control inputs of 'dom' dominate 'sub'. 247// Already assumes that 'dom' is available at 'sub', and that 'sub' 248// is not a constant (dominated by the method's StartNode). 249// Used by MemNode::find_previous_store to prove that the 250// control input of a memory operation predates (dominates) 251// an allocation it wants to look past. 252bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 253 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 254 return false; // Conservative answer for dead code 255 256 // Check 'dom'. 257 dom = dom->find_exact_control(dom); 258 if (dom == NULL || dom->is_top()) 259 return false; // Conservative answer for dead code 260 261 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 262 return true; 263 264 // 'dom' dominates 'sub' if its control edge and control edges 265 // of all its inputs dominate or equal to sub's control edge. 266 267 // Currently 'sub' is either Allocate, Initialize or Start nodes. 268 // Or Region for the check in LoadNode::Ideal(); 269 // 'sub' should have sub->in(0) != NULL. 270 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 271 sub->is_Region(), "expecting only these nodes"); 272 273 // Get control edge of 'sub'. 274 sub = sub->find_exact_control(sub->in(0)); 275 if (sub == NULL || sub->is_top()) 276 return false; // Conservative answer for dead code 277 278 assert(sub->is_CFG(), "expecting control"); 279 280 if (sub == dom) 281 return true; 282 283 if (sub->is_Start() || sub->is_Root()) 284 return false; 285 286 { 287 // Check all control edges of 'dom'. 288 289 ResourceMark rm; 290 Arena* arena = Thread::current()->resource_area(); 291 Node_List nlist(arena); 292 Unique_Node_List dom_list(arena); 293 294 dom_list.push(dom); 295 bool only_dominating_controls = false; 296 297 for (uint next = 0; next < dom_list.size(); next++) { 298 Node* n = dom_list.at(next); 299 if (!n->is_CFG() && n->pinned()) { 300 // Check only own control edge for pinned non-control nodes. 301 n = n->find_exact_control(n->in(0)); 302 if (n == NULL || n->is_top()) 303 return false; // Conservative answer for dead code 304 assert(n->is_CFG(), "expecting control"); 305 } 306 if (n->is_Con() || n->is_Start() || n->is_Root()) { 307 only_dominating_controls = true; 308 } else if (n->is_CFG()) { 309 if (n->dominates(sub, nlist)) 310 only_dominating_controls = true; 311 else 312 return false; 313 } else { 314 // First, own control edge. 315 Node* m = n->find_exact_control(n->in(0)); 316 if (m != NULL) { 317 if (m->is_top()) 318 return false; // Conservative answer for dead code 319 dom_list.push(m); 320 } 321 // Now, the rest of edges. 322 uint cnt = n->req(); 323 for (uint i = 1; i < cnt; i++) { 324 m = n->find_exact_control(n->in(i)); 325 if (m == NULL || m->is_top()) 326 continue; 327 dom_list.push(m); 328 } 329 } 330 } 331 return only_dominating_controls; 332 } 333} 334 335//---------------------detect_ptr_independence--------------------------------- 336// Used by MemNode::find_previous_store to prove that two base 337// pointers are never equal. 338// The pointers are accompanied by their associated allocations, 339// if any, which have been previously discovered by the caller. 340bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 341 Node* p2, AllocateNode* a2, 342 PhaseTransform* phase) { 343 // Attempt to prove that these two pointers cannot be aliased. 344 // They may both manifestly be allocations, and they should differ. 345 // Or, if they are not both allocations, they can be distinct constants. 346 // Otherwise, one is an allocation and the other a pre-existing value. 347 if (a1 == NULL && a2 == NULL) { // neither an allocation 348 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 349 } else if (a1 != NULL && a2 != NULL) { // both allocations 350 return (a1 != a2); 351 } else if (a1 != NULL) { // one allocation a1 352 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 353 return all_controls_dominate(p2, a1); 354 } else { //(a2 != NULL) // one allocation a2 355 return all_controls_dominate(p1, a2); 356 } 357 return false; 358} 359 360 361// The logic for reordering loads and stores uses four steps: 362// (a) Walk carefully past stores and initializations which we 363// can prove are independent of this load. 364// (b) Observe that the next memory state makes an exact match 365// with self (load or store), and locate the relevant store. 366// (c) Ensure that, if we were to wire self directly to the store, 367// the optimizer would fold it up somehow. 368// (d) Do the rewiring, and return, depending on some other part of 369// the optimizer to fold up the load. 370// This routine handles steps (a) and (b). Steps (c) and (d) are 371// specific to loads and stores, so they are handled by the callers. 372// (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 373// 374Node* MemNode::find_previous_store(PhaseTransform* phase) { 375 Node* ctrl = in(MemNode::Control); 376 Node* adr = in(MemNode::Address); 377 intptr_t offset = 0; 378 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 379 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 380 381 if (offset == Type::OffsetBot) 382 return NULL; // cannot unalias unless there are precise offsets 383 384 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 385 386 intptr_t size_in_bytes = memory_size(); 387 388 Node* mem = in(MemNode::Memory); // start searching here... 389 390 int cnt = 50; // Cycle limiter 391 for (;;) { // While we can dance past unrelated stores... 392 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 393 394 if (mem->is_Store()) { 395 Node* st_adr = mem->in(MemNode::Address); 396 intptr_t st_offset = 0; 397 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 398 if (st_base == NULL) 399 break; // inscrutable pointer 400 if (st_offset != offset && st_offset != Type::OffsetBot) { 401 const int MAX_STORE = BytesPerLong; 402 if (st_offset >= offset + size_in_bytes || 403 st_offset <= offset - MAX_STORE || 404 st_offset <= offset - mem->as_Store()->memory_size()) { 405 // Success: The offsets are provably independent. 406 // (You may ask, why not just test st_offset != offset and be done? 407 // The answer is that stores of different sizes can co-exist 408 // in the same sequence of RawMem effects. We sometimes initialize 409 // a whole 'tile' of array elements with a single jint or jlong.) 410 mem = mem->in(MemNode::Memory); 411 continue; // (a) advance through independent store memory 412 } 413 } 414 if (st_base != base && 415 detect_ptr_independence(base, alloc, 416 st_base, 417 AllocateNode::Ideal_allocation(st_base, phase), 418 phase)) { 419 // Success: The bases are provably independent. 420 mem = mem->in(MemNode::Memory); 421 continue; // (a) advance through independent store memory 422 } 423 424 // (b) At this point, if the bases or offsets do not agree, we lose, 425 // since we have not managed to prove 'this' and 'mem' independent. 426 if (st_base == base && st_offset == offset) { 427 return mem; // let caller handle steps (c), (d) 428 } 429 430 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 431 InitializeNode* st_init = mem->in(0)->as_Initialize(); 432 AllocateNode* st_alloc = st_init->allocation(); 433 if (st_alloc == NULL) 434 break; // something degenerated 435 bool known_identical = false; 436 bool known_independent = false; 437 if (alloc == st_alloc) 438 known_identical = true; 439 else if (alloc != NULL) 440 known_independent = true; 441 else if (all_controls_dominate(this, st_alloc)) 442 known_independent = true; 443 444 if (known_independent) { 445 // The bases are provably independent: Either they are 446 // manifestly distinct allocations, or else the control 447 // of this load dominates the store's allocation. 448 int alias_idx = phase->C->get_alias_index(adr_type()); 449 if (alias_idx == Compile::AliasIdxRaw) { 450 mem = st_alloc->in(TypeFunc::Memory); 451 } else { 452 mem = st_init->memory(alias_idx); 453 } 454 continue; // (a) advance through independent store memory 455 } 456 457 // (b) at this point, if we are not looking at a store initializing 458 // the same allocation we are loading from, we lose. 459 if (known_identical) { 460 // From caller, can_see_stored_value will consult find_captured_store. 461 return mem; // let caller handle steps (c), (d) 462 } 463 464 } else if (addr_t != NULL && addr_t->is_instance_field()) { 465 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 466 if (mem->is_Proj() && mem->in(0)->is_Call()) { 467 CallNode *call = mem->in(0)->as_Call(); 468 if (!call->may_modify(addr_t, phase)) { 469 mem = call->in(TypeFunc::Memory); 470 continue; // (a) advance through independent call memory 471 } 472 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 473 mem = mem->in(0)->in(TypeFunc::Memory); 474 continue; // (a) advance through independent MemBar memory 475 } else if (mem->is_MergeMem()) { 476 int alias_idx = phase->C->get_alias_index(adr_type()); 477 mem = mem->as_MergeMem()->memory_at(alias_idx); 478 continue; // (a) advance through independent MergeMem memory 479 } 480 } 481 482 // Unless there is an explicit 'continue', we must bail out here, 483 // because 'mem' is an inscrutable memory state (e.g., a call). 484 break; 485 } 486 487 return NULL; // bail out 488} 489 490//----------------------calculate_adr_type------------------------------------- 491// Helper function. Notices when the given type of address hits top or bottom. 492// Also, asserts a cross-check of the type against the expected address type. 493const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 494 if (t == Type::TOP) return NULL; // does not touch memory any more? 495 #ifdef PRODUCT 496 cross_check = NULL; 497 #else 498 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 499 #endif 500 const TypePtr* tp = t->isa_ptr(); 501 if (tp == NULL) { 502 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 503 return TypePtr::BOTTOM; // touches lots of memory 504 } else { 505 #ifdef ASSERT 506 // %%%% [phh] We don't check the alias index if cross_check is 507 // TypeRawPtr::BOTTOM. Needs to be investigated. 508 if (cross_check != NULL && 509 cross_check != TypePtr::BOTTOM && 510 cross_check != TypeRawPtr::BOTTOM) { 511 // Recheck the alias index, to see if it has changed (due to a bug). 512 Compile* C = Compile::current(); 513 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 514 "must stay in the original alias category"); 515 // The type of the address must be contained in the adr_type, 516 // disregarding "null"-ness. 517 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 518 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 519 assert(cross_check->meet(tp_notnull) == cross_check, 520 "real address must not escape from expected memory type"); 521 } 522 #endif 523 return tp; 524 } 525} 526 527//------------------------adr_phi_is_loop_invariant---------------------------- 528// A helper function for Ideal_DU_postCCP to check if a Phi in a counted 529// loop is loop invariant. Make a quick traversal of Phi and associated 530// CastPP nodes, looking to see if they are a closed group within the loop. 531bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 532 // The idea is that the phi-nest must boil down to only CastPP nodes 533 // with the same data. This implies that any path into the loop already 534 // includes such a CastPP, and so the original cast, whatever its input, 535 // must be covered by an equivalent cast, with an earlier control input. 536 ResourceMark rm; 537 538 // The loop entry input of the phi should be the unique dominating 539 // node for every Phi/CastPP in the loop. 540 Unique_Node_List closure; 541 closure.push(adr_phi->in(LoopNode::EntryControl)); 542 543 // Add the phi node and the cast to the worklist. 544 Unique_Node_List worklist; 545 worklist.push(adr_phi); 546 if( cast != NULL ){ 547 if( !cast->is_ConstraintCast() ) return false; 548 worklist.push(cast); 549 } 550 551 // Begin recursive walk of phi nodes. 552 while( worklist.size() ){ 553 // Take a node off the worklist 554 Node *n = worklist.pop(); 555 if( !closure.member(n) ){ 556 // Add it to the closure. 557 closure.push(n); 558 // Make a sanity check to ensure we don't waste too much time here. 559 if( closure.size() > 20) return false; 560 // This node is OK if: 561 // - it is a cast of an identical value 562 // - or it is a phi node (then we add its inputs to the worklist) 563 // Otherwise, the node is not OK, and we presume the cast is not invariant 564 if( n->is_ConstraintCast() ){ 565 worklist.push(n->in(1)); 566 } else if( n->is_Phi() ) { 567 for( uint i = 1; i < n->req(); i++ ) { 568 worklist.push(n->in(i)); 569 } 570 } else { 571 return false; 572 } 573 } 574 } 575 576 // Quit when the worklist is empty, and we've found no offending nodes. 577 return true; 578} 579 580//------------------------------Ideal_DU_postCCP------------------------------- 581// Find any cast-away of null-ness and keep its control. Null cast-aways are 582// going away in this pass and we need to make this memory op depend on the 583// gating null check. 584Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 585 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 586} 587 588// I tried to leave the CastPP's in. This makes the graph more accurate in 589// some sense; we get to keep around the knowledge that an oop is not-null 590// after some test. Alas, the CastPP's interfere with GVN (some values are 591// the regular oop, some are the CastPP of the oop, all merge at Phi's which 592// cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 593// some of the more trivial cases in the optimizer. Removing more useless 594// Phi's started allowing Loads to illegally float above null checks. I gave 595// up on this approach. CNC 10/20/2000 596// This static method may be called not from MemNode (EncodePNode calls it). 597// Only the control edge of the node 'n' might be updated. 598Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 599 Node *skipped_cast = NULL; 600 // Need a null check? Regular static accesses do not because they are 601 // from constant addresses. Array ops are gated by the range check (which 602 // always includes a NULL check). Just check field ops. 603 if( n->in(MemNode::Control) == NULL ) { 604 // Scan upwards for the highest location we can place this memory op. 605 while( true ) { 606 switch( adr->Opcode() ) { 607 608 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 609 adr = adr->in(AddPNode::Base); 610 continue; 611 612 case Op_DecodeN: // No change to NULL-ness, so peek thru 613 adr = adr->in(1); 614 continue; 615 616 case Op_CastPP: 617 // If the CastPP is useless, just peek on through it. 618 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 619 // Remember the cast that we've peeked though. If we peek 620 // through more than one, then we end up remembering the highest 621 // one, that is, if in a loop, the one closest to the top. 622 skipped_cast = adr; 623 adr = adr->in(1); 624 continue; 625 } 626 // CastPP is going away in this pass! We need this memory op to be 627 // control-dependent on the test that is guarding the CastPP. 628 ccp->hash_delete(n); 629 n->set_req(MemNode::Control, adr->in(0)); 630 ccp->hash_insert(n); 631 return n; 632 633 case Op_Phi: 634 // Attempt to float above a Phi to some dominating point. 635 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 636 // If we've already peeked through a Cast (which could have set the 637 // control), we can't float above a Phi, because the skipped Cast 638 // may not be loop invariant. 639 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 640 adr = adr->in(1); 641 continue; 642 } 643 } 644 645 // Intentional fallthrough! 646 647 // No obvious dominating point. The mem op is pinned below the Phi 648 // by the Phi itself. If the Phi goes away (no true value is merged) 649 // then the mem op can float, but not indefinitely. It must be pinned 650 // behind the controls leading to the Phi. 651 case Op_CheckCastPP: 652 // These usually stick around to change address type, however a 653 // useless one can be elided and we still need to pick up a control edge 654 if (adr->in(0) == NULL) { 655 // This CheckCastPP node has NO control and is likely useless. But we 656 // need check further up the ancestor chain for a control input to keep 657 // the node in place. 4959717. 658 skipped_cast = adr; 659 adr = adr->in(1); 660 continue; 661 } 662 ccp->hash_delete(n); 663 n->set_req(MemNode::Control, adr->in(0)); 664 ccp->hash_insert(n); 665 return n; 666 667 // List of "safe" opcodes; those that implicitly block the memory 668 // op below any null check. 669 case Op_CastX2P: // no null checks on native pointers 670 case Op_Parm: // 'this' pointer is not null 671 case Op_LoadP: // Loading from within a klass 672 case Op_LoadN: // Loading from within a klass 673 case Op_LoadKlass: // Loading from within a klass 674 case Op_LoadNKlass: // Loading from within a klass 675 case Op_ConP: // Loading from a klass 676 case Op_ConN: // Loading from a klass 677 case Op_CreateEx: // Sucking up the guts of an exception oop 678 case Op_Con: // Reading from TLS 679 case Op_CMoveP: // CMoveP is pinned 680 case Op_CMoveN: // CMoveN is pinned 681 break; // No progress 682 683 case Op_Proj: // Direct call to an allocation routine 684 case Op_SCMemProj: // Memory state from store conditional ops 685#ifdef ASSERT 686 { 687 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 688 const Node* call = adr->in(0); 689 if (call->is_CallJava()) { 690 const CallJavaNode* call_java = call->as_CallJava(); 691 const TypeTuple *r = call_java->tf()->range(); 692 assert(r->cnt() > TypeFunc::Parms, "must return value"); 693 const Type* ret_type = r->field_at(TypeFunc::Parms); 694 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 695 // We further presume that this is one of 696 // new_instance_Java, new_array_Java, or 697 // the like, but do not assert for this. 698 } else if (call->is_Allocate()) { 699 // similar case to new_instance_Java, etc. 700 } else if (!call->is_CallLeaf()) { 701 // Projections from fetch_oop (OSR) are allowed as well. 702 ShouldNotReachHere(); 703 } 704 } 705#endif 706 break; 707 default: 708 ShouldNotReachHere(); 709 } 710 break; 711 } 712 } 713 714 return NULL; // No progress 715} 716 717 718//============================================================================= 719uint LoadNode::size_of() const { return sizeof(*this); } 720uint LoadNode::cmp( const Node &n ) const 721{ return !Type::cmp( _type, ((LoadNode&)n)._type ); } 722const Type *LoadNode::bottom_type() const { return _type; } 723uint LoadNode::ideal_reg() const { 724 return Matcher::base2reg[_type->base()]; 725} 726 727#ifndef PRODUCT 728void LoadNode::dump_spec(outputStream *st) const { 729 MemNode::dump_spec(st); 730 if( !Verbose && !WizardMode ) { 731 // standard dump does this in Verbose and WizardMode 732 st->print(" #"); _type->dump_on(st); 733 } 734} 735#endif 736 737 738//----------------------------LoadNode::make----------------------------------- 739// Polymorphic factory method: 740Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 741 Compile* C = gvn.C; 742 743 // sanity check the alias category against the created node type 744 assert(!(adr_type->isa_oopptr() && 745 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 746 "use LoadKlassNode instead"); 747 assert(!(adr_type->isa_aryptr() && 748 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 749 "use LoadRangeNode instead"); 750 switch (bt) { 751 case T_BOOLEAN: 752 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() ); 753 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() ); 754 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() ); 755 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() ); 756 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() ); 757 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt ); 758 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt ); 759 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() ); 760 case T_OBJECT: 761#ifdef _LP64 762 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 763 const TypeNarrowOop* narrowtype; 764 if (rt->isa_narrowoop()) { 765 narrowtype = rt->is_narrowoop(); 766 } else { 767 narrowtype = rt->is_oopptr()->make_narrowoop(); 768 } 769 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, narrowtype)); 770 771 return DecodeNNode::decode(&gvn, load); 772 } else 773#endif 774 { 775 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 776 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 777 } 778 } 779 ShouldNotReachHere(); 780 return (LoadNode*)NULL; 781} 782 783LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 784 bool require_atomic = true; 785 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 786} 787 788 789 790 791//------------------------------hash------------------------------------------- 792uint LoadNode::hash() const { 793 // unroll addition of interesting fields 794 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 795} 796 797//---------------------------can_see_stored_value------------------------------ 798// This routine exists to make sure this set of tests is done the same 799// everywhere. We need to make a coordinated change: first LoadNode::Ideal 800// will change the graph shape in a way which makes memory alive twice at the 801// same time (uses the Oracle model of aliasing), then some 802// LoadXNode::Identity will fold things back to the equivalence-class model 803// of aliasing. 804Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 805 Node* ld_adr = in(MemNode::Address); 806 807 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 808 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 809 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 810 atp->field() != NULL && !atp->field()->is_volatile()) { 811 uint alias_idx = atp->index(); 812 bool final = atp->field()->is_final(); 813 Node* result = NULL; 814 Node* current = st; 815 // Skip through chains of MemBarNodes checking the MergeMems for 816 // new states for the slice of this load. Stop once any other 817 // kind of node is encountered. Loads from final memory can skip 818 // through any kind of MemBar but normal loads shouldn't skip 819 // through MemBarAcquire since the could allow them to move out of 820 // a synchronized region. 821 while (current->is_Proj()) { 822 int opc = current->in(0)->Opcode(); 823 if ((final && opc == Op_MemBarAcquire) || 824 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 825 Node* mem = current->in(0)->in(TypeFunc::Memory); 826 if (mem->is_MergeMem()) { 827 MergeMemNode* merge = mem->as_MergeMem(); 828 Node* new_st = merge->memory_at(alias_idx); 829 if (new_st == merge->base_memory()) { 830 // Keep searching 831 current = merge->base_memory(); 832 continue; 833 } 834 // Save the new memory state for the slice and fall through 835 // to exit. 836 result = new_st; 837 } 838 } 839 break; 840 } 841 if (result != NULL) { 842 st = result; 843 } 844 } 845 846 847 // Loop around twice in the case Load -> Initialize -> Store. 848 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 849 for (int trip = 0; trip <= 1; trip++) { 850 851 if (st->is_Store()) { 852 Node* st_adr = st->in(MemNode::Address); 853 if (!phase->eqv(st_adr, ld_adr)) { 854 // Try harder before giving up... Match raw and non-raw pointers. 855 intptr_t st_off = 0; 856 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 857 if (alloc == NULL) return NULL; 858 intptr_t ld_off = 0; 859 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 860 if (alloc != allo2) return NULL; 861 if (ld_off != st_off) return NULL; 862 // At this point we have proven something like this setup: 863 // A = Allocate(...) 864 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 865 // S = StoreQ(, AddP(, A.Parm , #Off), V) 866 // (Actually, we haven't yet proven the Q's are the same.) 867 // In other words, we are loading from a casted version of 868 // the same pointer-and-offset that we stored to. 869 // Thus, we are able to replace L by V. 870 } 871 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 872 if (store_Opcode() != st->Opcode()) 873 return NULL; 874 return st->in(MemNode::ValueIn); 875 } 876 877 intptr_t offset = 0; // scratch 878 879 // A load from a freshly-created object always returns zero. 880 // (This can happen after LoadNode::Ideal resets the load's memory input 881 // to find_captured_store, which returned InitializeNode::zero_memory.) 882 if (st->is_Proj() && st->in(0)->is_Allocate() && 883 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 884 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 885 // return a zero value for the load's basic type 886 // (This is one of the few places where a generic PhaseTransform 887 // can create new nodes. Think of it as lazily manifesting 888 // virtually pre-existing constants.) 889 return phase->zerocon(memory_type()); 890 } 891 892 // A load from an initialization barrier can match a captured store. 893 if (st->is_Proj() && st->in(0)->is_Initialize()) { 894 InitializeNode* init = st->in(0)->as_Initialize(); 895 AllocateNode* alloc = init->allocation(); 896 if (alloc != NULL && 897 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 898 // examine a captured store value 899 st = init->find_captured_store(offset, memory_size(), phase); 900 if (st != NULL) 901 continue; // take one more trip around 902 } 903 } 904 905 break; 906 } 907 908 return NULL; 909} 910 911//----------------------is_instance_field_load_with_local_phi------------------ 912bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 913 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 914 in(MemNode::Address)->is_AddP() ) { 915 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 916 // Only instances. 917 if( t_oop != NULL && t_oop->is_instance_field() && 918 t_oop->offset() != Type::OffsetBot && 919 t_oop->offset() != Type::OffsetTop) { 920 return true; 921 } 922 } 923 return false; 924} 925 926//------------------------------Identity--------------------------------------- 927// Loads are identity if previous store is to same address 928Node *LoadNode::Identity( PhaseTransform *phase ) { 929 // If the previous store-maker is the right kind of Store, and the store is 930 // to the same address, then we are equal to the value stored. 931 Node* mem = in(MemNode::Memory); 932 Node* value = can_see_stored_value(mem, phase); 933 if( value ) { 934 // byte, short & char stores truncate naturally. 935 // A load has to load the truncated value which requires 936 // some sort of masking operation and that requires an 937 // Ideal call instead of an Identity call. 938 if (memory_size() < BytesPerInt) { 939 // If the input to the store does not fit with the load's result type, 940 // it must be truncated via an Ideal call. 941 if (!phase->type(value)->higher_equal(phase->type(this))) 942 return this; 943 } 944 // (This works even when value is a Con, but LoadNode::Value 945 // usually runs first, producing the singleton type of the Con.) 946 return value; 947 } 948 949 // Search for an existing data phi which was generated before for the same 950 // instance's field to avoid infinite genertion of phis in a loop. 951 Node *region = mem->in(0); 952 if (is_instance_field_load_with_local_phi(region)) { 953 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 954 int this_index = phase->C->get_alias_index(addr_t); 955 int this_offset = addr_t->offset(); 956 int this_id = addr_t->is_oopptr()->instance_id(); 957 const Type* this_type = bottom_type(); 958 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 959 Node* phi = region->fast_out(i); 960 if (phi->is_Phi() && phi != mem && 961 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 962 return phi; 963 } 964 } 965 } 966 967 return this; 968} 969 970 971// Returns true if the AliasType refers to the field that holds the 972// cached box array. Currently only handles the IntegerCache case. 973static bool is_autobox_cache(Compile::AliasType* atp) { 974 if (atp != NULL && atp->field() != NULL) { 975 ciField* field = atp->field(); 976 ciSymbol* klass = field->holder()->name(); 977 if (field->name() == ciSymbol::cache_field_name() && 978 field->holder()->uses_default_loader() && 979 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 980 return true; 981 } 982 } 983 return false; 984} 985 986// Fetch the base value in the autobox array 987static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 988 if (atp != NULL && atp->field() != NULL) { 989 ciField* field = atp->field(); 990 ciSymbol* klass = field->holder()->name(); 991 if (field->name() == ciSymbol::cache_field_name() && 992 field->holder()->uses_default_loader() && 993 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 994 assert(field->is_constant(), "what?"); 995 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 996 // Fetch the box object at the base of the array and get its value 997 ciInstance* box = array->obj_at(0)->as_instance(); 998 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 999 if (ik->nof_nonstatic_fields() == 1) { 1000 // This should be true nonstatic_field_at requires calling 1001 // nof_nonstatic_fields so check it anyway 1002 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1003 cache_offset = c.as_int(); 1004 } 1005 return true; 1006 } 1007 } 1008 return false; 1009} 1010 1011// Returns true if the AliasType refers to the value field of an 1012// autobox object. Currently only handles Integer. 1013static bool is_autobox_object(Compile::AliasType* atp) { 1014 if (atp != NULL && atp->field() != NULL) { 1015 ciField* field = atp->field(); 1016 ciSymbol* klass = field->holder()->name(); 1017 if (field->name() == ciSymbol::value_name() && 1018 field->holder()->uses_default_loader() && 1019 klass == ciSymbol::java_lang_Integer()) { 1020 return true; 1021 } 1022 } 1023 return false; 1024} 1025 1026 1027// We're loading from an object which has autobox behaviour. 1028// If this object is result of a valueOf call we'll have a phi 1029// merging a newly allocated object and a load from the cache. 1030// We want to replace this load with the original incoming 1031// argument to the valueOf call. 1032Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1033 Node* base = in(Address)->in(AddPNode::Base); 1034 if (base->is_Phi() && base->req() == 3) { 1035 AllocateNode* allocation = NULL; 1036 int allocation_index = -1; 1037 int load_index = -1; 1038 for (uint i = 1; i < base->req(); i++) { 1039 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1040 if (allocation != NULL) { 1041 allocation_index = i; 1042 load_index = 3 - allocation_index; 1043 break; 1044 } 1045 } 1046 LoadNode* load = NULL; 1047 if (allocation != NULL && base->in(load_index)->is_Load()) { 1048 load = base->in(load_index)->as_Load(); 1049 } 1050 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1051 // Push the loads from the phi that comes from valueOf up 1052 // through it to allow elimination of the loads and the recovery 1053 // of the original value. 1054 Node* mem_phi = in(Memory); 1055 Node* offset = in(Address)->in(AddPNode::Offset); 1056 1057 Node* in1 = clone(); 1058 Node* in1_addr = in1->in(Address)->clone(); 1059 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1060 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1061 in1_addr->set_req(AddPNode::Offset, offset); 1062 in1->set_req(0, base->in(allocation_index)); 1063 in1->set_req(Address, in1_addr); 1064 in1->set_req(Memory, mem_phi->in(allocation_index)); 1065 1066 Node* in2 = clone(); 1067 Node* in2_addr = in2->in(Address)->clone(); 1068 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1069 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1070 in2_addr->set_req(AddPNode::Offset, offset); 1071 in2->set_req(0, base->in(load_index)); 1072 in2->set_req(Address, in2_addr); 1073 in2->set_req(Memory, mem_phi->in(load_index)); 1074 1075 in1_addr = phase->transform(in1_addr); 1076 in1 = phase->transform(in1); 1077 in2_addr = phase->transform(in2_addr); 1078 in2 = phase->transform(in2); 1079 1080 PhiNode* result = PhiNode::make_blank(base->in(0), this); 1081 result->set_req(allocation_index, in1); 1082 result->set_req(load_index, in2); 1083 return result; 1084 } 1085 } else if (base->is_Load()) { 1086 // Eliminate the load of Integer.value for integers from the cache 1087 // array by deriving the value from the index into the array. 1088 // Capture the offset of the load and then reverse the computation. 1089 Node* load_base = base->in(Address)->in(AddPNode::Base); 1090 if (load_base != NULL) { 1091 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1092 intptr_t cache_offset; 1093 int shift = -1; 1094 Node* cache = NULL; 1095 if (is_autobox_cache(atp)) { 1096 shift = exact_log2(type2aelembytes(T_OBJECT)); 1097 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1098 } 1099 if (cache != NULL && base->in(Address)->is_AddP()) { 1100 Node* elements[4]; 1101 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1102 int cache_low; 1103 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1104 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1105 // Add up all the offsets making of the address of the load 1106 Node* result = elements[0]; 1107 for (int i = 1; i < count; i++) { 1108 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1109 } 1110 // Remove the constant offset from the address and then 1111 // remove the scaling of the offset to recover the original index. 1112 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1113 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1114 // Peel the shift off directly but wrap it in a dummy node 1115 // since Ideal can't return existing nodes 1116 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1117 } else { 1118 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1119 } 1120#ifdef _LP64 1121 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1122#endif 1123 return result; 1124 } 1125 } 1126 } 1127 } 1128 return NULL; 1129} 1130 1131//------------------------------split_through_phi------------------------------ 1132// Split instance field load through Phi. 1133Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1134 Node* mem = in(MemNode::Memory); 1135 Node* address = in(MemNode::Address); 1136 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1137 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1138 1139 assert(mem->is_Phi() && (t_oop != NULL) && 1140 t_oop->is_instance_field(), "invalide conditions"); 1141 1142 Node *region = mem->in(0); 1143 if (region == NULL) { 1144 return NULL; // Wait stable graph 1145 } 1146 uint cnt = mem->req(); 1147 for( uint i = 1; i < cnt; i++ ) { 1148 Node *in = mem->in(i); 1149 if( in == NULL ) { 1150 return NULL; // Wait stable graph 1151 } 1152 } 1153 // Check for loop invariant. 1154 if (cnt == 3) { 1155 for( uint i = 1; i < cnt; i++ ) { 1156 Node *in = mem->in(i); 1157 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1158 if (m == mem) { 1159 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1160 return this; 1161 } 1162 } 1163 } 1164 // Split through Phi (see original code in loopopts.cpp). 1165 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1166 1167 // Do nothing here if Identity will find a value 1168 // (to avoid infinite chain of value phis generation). 1169 if ( !phase->eqv(this, this->Identity(phase)) ) 1170 return NULL; 1171 1172 // Skip the split if the region dominates some control edge of the address. 1173 if (cnt == 3 && !MemNode::all_controls_dominate(address, region)) 1174 return NULL; 1175 1176 const Type* this_type = this->bottom_type(); 1177 int this_index = phase->C->get_alias_index(addr_t); 1178 int this_offset = addr_t->offset(); 1179 int this_iid = addr_t->is_oopptr()->instance_id(); 1180 int wins = 0; 1181 PhaseIterGVN *igvn = phase->is_IterGVN(); 1182 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1183 for( uint i = 1; i < region->req(); i++ ) { 1184 Node *x; 1185 Node* the_clone = NULL; 1186 if( region->in(i) == phase->C->top() ) { 1187 x = phase->C->top(); // Dead path? Use a dead data op 1188 } else { 1189 x = this->clone(); // Else clone up the data op 1190 the_clone = x; // Remember for possible deletion. 1191 // Alter data node to use pre-phi inputs 1192 if( this->in(0) == region ) { 1193 x->set_req( 0, region->in(i) ); 1194 } else { 1195 x->set_req( 0, NULL ); 1196 } 1197 for( uint j = 1; j < this->req(); j++ ) { 1198 Node *in = this->in(j); 1199 if( in->is_Phi() && in->in(0) == region ) 1200 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 1201 } 1202 } 1203 // Check for a 'win' on some paths 1204 const Type *t = x->Value(igvn); 1205 1206 bool singleton = t->singleton(); 1207 1208 // See comments in PhaseIdealLoop::split_thru_phi(). 1209 if( singleton && t == Type::TOP ) { 1210 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1211 } 1212 1213 if( singleton ) { 1214 wins++; 1215 x = igvn->makecon(t); 1216 } else { 1217 // We now call Identity to try to simplify the cloned node. 1218 // Note that some Identity methods call phase->type(this). 1219 // Make sure that the type array is big enough for 1220 // our new node, even though we may throw the node away. 1221 // (This tweaking with igvn only works because x is a new node.) 1222 igvn->set_type(x, t); 1223 Node *y = x->Identity(igvn); 1224 if( y != x ) { 1225 wins++; 1226 x = y; 1227 } else { 1228 y = igvn->hash_find(x); 1229 if( y ) { 1230 wins++; 1231 x = y; 1232 } else { 1233 // Else x is a new node we are keeping 1234 // We do not need register_new_node_with_optimizer 1235 // because set_type has already been called. 1236 igvn->_worklist.push(x); 1237 } 1238 } 1239 } 1240 if (x != the_clone && the_clone != NULL) 1241 igvn->remove_dead_node(the_clone); 1242 phi->set_req(i, x); 1243 } 1244 if( wins > 0 ) { 1245 // Record Phi 1246 igvn->register_new_node_with_optimizer(phi); 1247 return phi; 1248 } 1249 igvn->remove_dead_node(phi); 1250 return NULL; 1251} 1252 1253//------------------------------Ideal------------------------------------------ 1254// If the load is from Field memory and the pointer is non-null, we can 1255// zero out the control input. 1256// If the offset is constant and the base is an object allocation, 1257// try to hook me up to the exact initializing store. 1258Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1259 Node* p = MemNode::Ideal_common(phase, can_reshape); 1260 if (p) return (p == NodeSentinel) ? NULL : p; 1261 1262 Node* ctrl = in(MemNode::Control); 1263 Node* address = in(MemNode::Address); 1264 1265 // Skip up past a SafePoint control. Cannot do this for Stores because 1266 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1267 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1268 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1269 ctrl = ctrl->in(0); 1270 set_req(MemNode::Control,ctrl); 1271 } 1272 1273 // Check for useless control edge in some common special cases 1274 if (in(MemNode::Control) != NULL) { 1275 intptr_t ignore = 0; 1276 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1277 if (base != NULL 1278 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1279 && all_controls_dominate(base, phase->C->start())) { 1280 // A method-invariant, non-null address (constant or 'this' argument). 1281 set_req(MemNode::Control, NULL); 1282 } 1283 } 1284 1285 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) { 1286 Node* base = in(Address)->in(AddPNode::Base); 1287 if (base != NULL) { 1288 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1289 if (is_autobox_object(atp)) { 1290 Node* result = eliminate_autobox(phase); 1291 if (result != NULL) return result; 1292 } 1293 } 1294 } 1295 1296 Node* mem = in(MemNode::Memory); 1297 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1298 1299 if (addr_t != NULL) { 1300 // try to optimize our memory input 1301 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1302 if (opt_mem != mem) { 1303 set_req(MemNode::Memory, opt_mem); 1304 return this; 1305 } 1306 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1307 if (can_reshape && opt_mem->is_Phi() && 1308 (t_oop != NULL) && t_oop->is_instance_field()) { 1309 // Split instance field load through Phi. 1310 Node* result = split_through_phi(phase); 1311 if (result != NULL) return result; 1312 } 1313 } 1314 1315 // Check for prior store with a different base or offset; make Load 1316 // independent. Skip through any number of them. Bail out if the stores 1317 // are in an endless dead cycle and report no progress. This is a key 1318 // transform for Reflection. However, if after skipping through the Stores 1319 // we can't then fold up against a prior store do NOT do the transform as 1320 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1321 // array memory alive twice: once for the hoisted Load and again after the 1322 // bypassed Store. This situation only works if EVERYBODY who does 1323 // anti-dependence work knows how to bypass. I.e. we need all 1324 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1325 // the alias index stuff. So instead, peek through Stores and IFF we can 1326 // fold up, do so. 1327 Node* prev_mem = find_previous_store(phase); 1328 // Steps (a), (b): Walk past independent stores to find an exact match. 1329 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1330 // (c) See if we can fold up on the spot, but don't fold up here. 1331 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or 1332 // just return a prior value, which is done by Identity calls. 1333 if (can_see_stored_value(prev_mem, phase)) { 1334 // Make ready for step (d): 1335 set_req(MemNode::Memory, prev_mem); 1336 return this; 1337 } 1338 } 1339 1340 return NULL; // No further progress 1341} 1342 1343// Helper to recognize certain Klass fields which are invariant across 1344// some group of array types (e.g., int[] or all T[] where T < Object). 1345const Type* 1346LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1347 ciKlass* klass) const { 1348 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1349 // The field is Klass::_modifier_flags. Return its (constant) value. 1350 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1351 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1352 return TypeInt::make(klass->modifier_flags()); 1353 } 1354 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1355 // The field is Klass::_access_flags. Return its (constant) value. 1356 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1357 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1358 return TypeInt::make(klass->access_flags()); 1359 } 1360 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1361 // The field is Klass::_layout_helper. Return its constant value if known. 1362 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1363 return TypeInt::make(klass->layout_helper()); 1364 } 1365 1366 // No match. 1367 return NULL; 1368} 1369 1370//------------------------------Value----------------------------------------- 1371const Type *LoadNode::Value( PhaseTransform *phase ) const { 1372 // Either input is TOP ==> the result is TOP 1373 Node* mem = in(MemNode::Memory); 1374 const Type *t1 = phase->type(mem); 1375 if (t1 == Type::TOP) return Type::TOP; 1376 Node* adr = in(MemNode::Address); 1377 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1378 if (tp == NULL || tp->empty()) return Type::TOP; 1379 int off = tp->offset(); 1380 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1381 1382 // Try to guess loaded type from pointer type 1383 if (tp->base() == Type::AryPtr) { 1384 const Type *t = tp->is_aryptr()->elem(); 1385 // Don't do this for integer types. There is only potential profit if 1386 // the element type t is lower than _type; that is, for int types, if _type is 1387 // more restrictive than t. This only happens here if one is short and the other 1388 // char (both 16 bits), and in those cases we've made an intentional decision 1389 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1390 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1391 // 1392 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1393 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1394 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1395 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1396 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1397 // In fact, that could have been the original type of p1, and p1 could have 1398 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1399 // expression (LShiftL quux 3) independently optimized to the constant 8. 1400 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1401 && Opcode() != Op_LoadKlass) { 1402 // t might actually be lower than _type, if _type is a unique 1403 // concrete subclass of abstract class t. 1404 // Make sure the reference is not into the header, by comparing 1405 // the offset against the offset of the start of the array's data. 1406 // Different array types begin at slightly different offsets (12 vs. 16). 1407 // We choose T_BYTE as an example base type that is least restrictive 1408 // as to alignment, which will therefore produce the smallest 1409 // possible base offset. 1410 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1411 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1412 const Type* jt = t->join(_type); 1413 // In any case, do not allow the join, per se, to empty out the type. 1414 if (jt->empty() && !t->empty()) { 1415 // This can happen if a interface-typed array narrows to a class type. 1416 jt = _type; 1417 } 1418 1419 if (EliminateAutoBox) { 1420 // The pointers in the autobox arrays are always non-null 1421 Node* base = in(Address)->in(AddPNode::Base); 1422 if (base != NULL) { 1423 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1424 if (is_autobox_cache(atp)) { 1425 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1426 } 1427 } 1428 } 1429 return jt; 1430 } 1431 } 1432 } else if (tp->base() == Type::InstPtr) { 1433 assert( off != Type::OffsetBot || 1434 // arrays can be cast to Objects 1435 tp->is_oopptr()->klass()->is_java_lang_Object() || 1436 // unsafe field access may not have a constant offset 1437 phase->C->has_unsafe_access(), 1438 "Field accesses must be precise" ); 1439 // For oop loads, we expect the _type to be precise 1440 } else if (tp->base() == Type::KlassPtr) { 1441 assert( off != Type::OffsetBot || 1442 // arrays can be cast to Objects 1443 tp->is_klassptr()->klass()->is_java_lang_Object() || 1444 // also allow array-loading from the primary supertype 1445 // array during subtype checks 1446 Opcode() == Op_LoadKlass, 1447 "Field accesses must be precise" ); 1448 // For klass/static loads, we expect the _type to be precise 1449 } 1450 1451 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1452 if (tkls != NULL && !StressReflectiveCode) { 1453 ciKlass* klass = tkls->klass(); 1454 if (klass->is_loaded() && tkls->klass_is_exact()) { 1455 // We are loading a field from a Klass metaobject whose identity 1456 // is known at compile time (the type is "exact" or "precise"). 1457 // Check for fields we know are maintained as constants by the VM. 1458 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1459 // The field is Klass::_super_check_offset. Return its (constant) value. 1460 // (Folds up type checking code.) 1461 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1462 return TypeInt::make(klass->super_check_offset()); 1463 } 1464 // Compute index into primary_supers array 1465 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1466 // Check for overflowing; use unsigned compare to handle the negative case. 1467 if( depth < ciKlass::primary_super_limit() ) { 1468 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1469 // (Folds up type checking code.) 1470 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1471 ciKlass *ss = klass->super_of_depth(depth); 1472 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1473 } 1474 const Type* aift = load_array_final_field(tkls, klass); 1475 if (aift != NULL) return aift; 1476 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1477 && klass->is_array_klass()) { 1478 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1479 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1480 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1481 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1482 } 1483 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1484 // The field is Klass::_java_mirror. Return its (constant) value. 1485 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1486 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1487 return TypeInstPtr::make(klass->java_mirror()); 1488 } 1489 } 1490 1491 // We can still check if we are loading from the primary_supers array at a 1492 // shallow enough depth. Even though the klass is not exact, entries less 1493 // than or equal to its super depth are correct. 1494 if (klass->is_loaded() ) { 1495 ciType *inner = klass->klass(); 1496 while( inner->is_obj_array_klass() ) 1497 inner = inner->as_obj_array_klass()->base_element_type(); 1498 if( inner->is_instance_klass() && 1499 !inner->as_instance_klass()->flags().is_interface() ) { 1500 // Compute index into primary_supers array 1501 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1502 // Check for overflowing; use unsigned compare to handle the negative case. 1503 if( depth < ciKlass::primary_super_limit() && 1504 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1505 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1506 // (Folds up type checking code.) 1507 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1508 ciKlass *ss = klass->super_of_depth(depth); 1509 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1510 } 1511 } 1512 } 1513 1514 // If the type is enough to determine that the thing is not an array, 1515 // we can give the layout_helper a positive interval type. 1516 // This will help short-circuit some reflective code. 1517 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1518 && !klass->is_array_klass() // not directly typed as an array 1519 && !klass->is_interface() // specifically not Serializable & Cloneable 1520 && !klass->is_java_lang_Object() // not the supertype of all T[] 1521 ) { 1522 // Note: When interfaces are reliable, we can narrow the interface 1523 // test to (klass != Serializable && klass != Cloneable). 1524 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1525 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1526 // The key property of this type is that it folds up tests 1527 // for array-ness, since it proves that the layout_helper is positive. 1528 // Thus, a generic value like the basic object layout helper works fine. 1529 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1530 } 1531 } 1532 1533 // If we are loading from a freshly-allocated object, produce a zero, 1534 // if the load is provably beyond the header of the object. 1535 // (Also allow a variable load from a fresh array to produce zero.) 1536 if (ReduceFieldZeroing) { 1537 Node* value = can_see_stored_value(mem,phase); 1538 if (value != NULL && value->is_Con()) 1539 return value->bottom_type(); 1540 } 1541 1542 const TypeOopPtr *tinst = tp->isa_oopptr(); 1543 if (tinst != NULL && tinst->is_instance_field()) { 1544 // If we have an instance type and our memory input is the 1545 // programs's initial memory state, there is no matching store, 1546 // so just return a zero of the appropriate type 1547 Node *mem = in(MemNode::Memory); 1548 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1549 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1550 return Type::get_zero_type(_type->basic_type()); 1551 } 1552 } 1553 return _type; 1554} 1555 1556//------------------------------match_edge------------------------------------- 1557// Do we Match on this edge index or not? Match only the address. 1558uint LoadNode::match_edge(uint idx) const { 1559 return idx == MemNode::Address; 1560} 1561 1562//--------------------------LoadBNode::Ideal-------------------------------------- 1563// 1564// If the previous store is to the same address as this load, 1565// and the value stored was larger than a byte, replace this load 1566// with the value stored truncated to a byte. If no truncation is 1567// needed, the replacement is done in LoadNode::Identity(). 1568// 1569Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1570 Node* mem = in(MemNode::Memory); 1571 Node* value = can_see_stored_value(mem,phase); 1572 if( value && !phase->type(value)->higher_equal( _type ) ) { 1573 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1574 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1575 } 1576 // Identity call will handle the case where truncation is not needed. 1577 return LoadNode::Ideal(phase, can_reshape); 1578} 1579 1580//--------------------------LoadCNode::Ideal-------------------------------------- 1581// 1582// If the previous store is to the same address as this load, 1583// and the value stored was larger than a char, replace this load 1584// with the value stored truncated to a char. If no truncation is 1585// needed, the replacement is done in LoadNode::Identity(). 1586// 1587Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1588 Node* mem = in(MemNode::Memory); 1589 Node* value = can_see_stored_value(mem,phase); 1590 if( value && !phase->type(value)->higher_equal( _type ) ) 1591 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1592 // Identity call will handle the case where truncation is not needed. 1593 return LoadNode::Ideal(phase, can_reshape); 1594} 1595 1596//--------------------------LoadSNode::Ideal-------------------------------------- 1597// 1598// If the previous store is to the same address as this load, 1599// and the value stored was larger than a short, replace this load 1600// with the value stored truncated to a short. If no truncation is 1601// needed, the replacement is done in LoadNode::Identity(). 1602// 1603Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1604 Node* mem = in(MemNode::Memory); 1605 Node* value = can_see_stored_value(mem,phase); 1606 if( value && !phase->type(value)->higher_equal( _type ) ) { 1607 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1608 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1609 } 1610 // Identity call will handle the case where truncation is not needed. 1611 return LoadNode::Ideal(phase, can_reshape); 1612} 1613 1614//============================================================================= 1615//----------------------------LoadKlassNode::make------------------------------ 1616// Polymorphic factory method: 1617Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1618 Compile* C = gvn.C; 1619 Node *ctl = NULL; 1620 // sanity check the alias category against the created node type 1621 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1622 assert(adr_type != NULL, "expecting TypeOopPtr"); 1623#ifdef _LP64 1624 if (adr_type->is_ptr_to_narrowoop()) { 1625 const TypeNarrowOop* narrowtype = tk->is_oopptr()->make_narrowoop(); 1626 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, narrowtype)); 1627 return DecodeNNode::decode(&gvn, load_klass); 1628 } 1629#endif 1630 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1631 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1632} 1633 1634//------------------------------Value------------------------------------------ 1635const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1636 return klass_value_common(phase); 1637} 1638 1639const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1640 // Either input is TOP ==> the result is TOP 1641 const Type *t1 = phase->type( in(MemNode::Memory) ); 1642 if (t1 == Type::TOP) return Type::TOP; 1643 Node *adr = in(MemNode::Address); 1644 const Type *t2 = phase->type( adr ); 1645 if (t2 == Type::TOP) return Type::TOP; 1646 const TypePtr *tp = t2->is_ptr(); 1647 if (TypePtr::above_centerline(tp->ptr()) || 1648 tp->ptr() == TypePtr::Null) return Type::TOP; 1649 1650 // Return a more precise klass, if possible 1651 const TypeInstPtr *tinst = tp->isa_instptr(); 1652 if (tinst != NULL) { 1653 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1654 int offset = tinst->offset(); 1655 if (ik == phase->C->env()->Class_klass() 1656 && (offset == java_lang_Class::klass_offset_in_bytes() || 1657 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1658 // We are loading a special hidden field from a Class mirror object, 1659 // the field which points to the VM's Klass metaobject. 1660 ciType* t = tinst->java_mirror_type(); 1661 // java_mirror_type returns non-null for compile-time Class constants. 1662 if (t != NULL) { 1663 // constant oop => constant klass 1664 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1665 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1666 } 1667 if (!t->is_klass()) { 1668 // a primitive Class (e.g., int.class) has NULL for a klass field 1669 return TypePtr::NULL_PTR; 1670 } 1671 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1672 return TypeKlassPtr::make(t->as_klass()); 1673 } 1674 // non-constant mirror, so we can't tell what's going on 1675 } 1676 if( !ik->is_loaded() ) 1677 return _type; // Bail out if not loaded 1678 if (offset == oopDesc::klass_offset_in_bytes()) { 1679 if (tinst->klass_is_exact()) { 1680 return TypeKlassPtr::make(ik); 1681 } 1682 // See if we can become precise: no subklasses and no interface 1683 // (Note: We need to support verified interfaces.) 1684 if (!ik->is_interface() && !ik->has_subklass()) { 1685 //assert(!UseExactTypes, "this code should be useless with exact types"); 1686 // Add a dependence; if any subclass added we need to recompile 1687 if (!ik->is_final()) { 1688 // %%% should use stronger assert_unique_concrete_subtype instead 1689 phase->C->dependencies()->assert_leaf_type(ik); 1690 } 1691 // Return precise klass 1692 return TypeKlassPtr::make(ik); 1693 } 1694 1695 // Return root of possible klass 1696 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1697 } 1698 } 1699 1700 // Check for loading klass from an array 1701 const TypeAryPtr *tary = tp->isa_aryptr(); 1702 if( tary != NULL ) { 1703 ciKlass *tary_klass = tary->klass(); 1704 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1705 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1706 if (tary->klass_is_exact()) { 1707 return TypeKlassPtr::make(tary_klass); 1708 } 1709 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1710 // If the klass is an object array, we defer the question to the 1711 // array component klass. 1712 if( ak->is_obj_array_klass() ) { 1713 assert( ak->is_loaded(), "" ); 1714 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1715 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1716 ciInstanceKlass* ik = base_k->as_instance_klass(); 1717 // See if we can become precise: no subklasses and no interface 1718 if (!ik->is_interface() && !ik->has_subklass()) { 1719 //assert(!UseExactTypes, "this code should be useless with exact types"); 1720 // Add a dependence; if any subclass added we need to recompile 1721 if (!ik->is_final()) { 1722 phase->C->dependencies()->assert_leaf_type(ik); 1723 } 1724 // Return precise array klass 1725 return TypeKlassPtr::make(ak); 1726 } 1727 } 1728 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1729 } else { // Found a type-array? 1730 //assert(!UseExactTypes, "this code should be useless with exact types"); 1731 assert( ak->is_type_array_klass(), "" ); 1732 return TypeKlassPtr::make(ak); // These are always precise 1733 } 1734 } 1735 } 1736 1737 // Check for loading klass from an array klass 1738 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1739 if (tkls != NULL && !StressReflectiveCode) { 1740 ciKlass* klass = tkls->klass(); 1741 if( !klass->is_loaded() ) 1742 return _type; // Bail out if not loaded 1743 if( klass->is_obj_array_klass() && 1744 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1745 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1746 // // Always returning precise element type is incorrect, 1747 // // e.g., element type could be object and array may contain strings 1748 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1749 1750 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1751 // according to the element type's subclassing. 1752 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1753 } 1754 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1755 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1756 ciKlass* sup = klass->as_instance_klass()->super(); 1757 // The field is Klass::_super. Return its (constant) value. 1758 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1759 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1760 } 1761 } 1762 1763 // Bailout case 1764 return LoadNode::Value(phase); 1765} 1766 1767//------------------------------Identity--------------------------------------- 1768// To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1769// Also feed through the klass in Allocate(...klass...)._klass. 1770Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1771 return klass_identity_common(phase); 1772} 1773 1774Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 1775 Node* x = LoadNode::Identity(phase); 1776 if (x != this) return x; 1777 1778 // Take apart the address into an oop and and offset. 1779 // Return 'this' if we cannot. 1780 Node* adr = in(MemNode::Address); 1781 intptr_t offset = 0; 1782 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1783 if (base == NULL) return this; 1784 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1785 if (toop == NULL) return this; 1786 1787 // We can fetch the klass directly through an AllocateNode. 1788 // This works even if the klass is not constant (clone or newArray). 1789 if (offset == oopDesc::klass_offset_in_bytes()) { 1790 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1791 if (allocated_klass != NULL) { 1792 return allocated_klass; 1793 } 1794 } 1795 1796 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1797 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1798 // See inline_native_Class_query for occurrences of these patterns. 1799 // Java Example: x.getClass().isAssignableFrom(y) 1800 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1801 // 1802 // This improves reflective code, often making the Class 1803 // mirror go completely dead. (Current exception: Class 1804 // mirrors may appear in debug info, but we could clean them out by 1805 // introducing a new debug info operator for klassOop.java_mirror). 1806 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1807 && (offset == java_lang_Class::klass_offset_in_bytes() || 1808 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1809 // We are loading a special hidden field from a Class mirror, 1810 // the field which points to its Klass or arrayKlass metaobject. 1811 if (base->is_Load()) { 1812 Node* adr2 = base->in(MemNode::Address); 1813 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1814 if (tkls != NULL && !tkls->empty() 1815 && (tkls->klass()->is_instance_klass() || 1816 tkls->klass()->is_array_klass()) 1817 && adr2->is_AddP() 1818 ) { 1819 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1820 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1821 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1822 } 1823 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1824 return adr2->in(AddPNode::Base); 1825 } 1826 } 1827 } 1828 } 1829 1830 return this; 1831} 1832 1833 1834//------------------------------Value------------------------------------------ 1835const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 1836 const Type *t = klass_value_common(phase); 1837 1838 if (t == TypePtr::NULL_PTR) { 1839 return TypeNarrowOop::NULL_PTR; 1840 } 1841 if (t != Type::TOP && !t->isa_narrowoop()) { 1842 assert(t->is_oopptr(), "sanity"); 1843 t = t->is_oopptr()->make_narrowoop(); 1844 } 1845 return t; 1846} 1847 1848//------------------------------Identity--------------------------------------- 1849// To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 1850// Also feed through the klass in Allocate(...klass...)._klass. 1851Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 1852 Node *x = klass_identity_common(phase); 1853 1854 const Type *t = phase->type( x ); 1855 if( t == Type::TOP ) return x; 1856 if( t->isa_narrowoop()) return x; 1857 1858 return EncodePNode::encode(phase, x); 1859} 1860 1861//------------------------------Value----------------------------------------- 1862const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1863 // Either input is TOP ==> the result is TOP 1864 const Type *t1 = phase->type( in(MemNode::Memory) ); 1865 if( t1 == Type::TOP ) return Type::TOP; 1866 Node *adr = in(MemNode::Address); 1867 const Type *t2 = phase->type( adr ); 1868 if( t2 == Type::TOP ) return Type::TOP; 1869 const TypePtr *tp = t2->is_ptr(); 1870 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 1871 const TypeAryPtr *tap = tp->isa_aryptr(); 1872 if( !tap ) return _type; 1873 return tap->size(); 1874} 1875 1876//------------------------------Identity--------------------------------------- 1877// Feed through the length in AllocateArray(...length...)._length. 1878Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 1879 Node* x = LoadINode::Identity(phase); 1880 if (x != this) return x; 1881 1882 // Take apart the address into an oop and and offset. 1883 // Return 'this' if we cannot. 1884 Node* adr = in(MemNode::Address); 1885 intptr_t offset = 0; 1886 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1887 if (base == NULL) return this; 1888 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1889 if (tary == NULL) return this; 1890 1891 // We can fetch the length directly through an AllocateArrayNode. 1892 // This works even if the length is not constant (clone or newArray). 1893 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1894 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase); 1895 if (allocated_length != NULL) { 1896 return allocated_length; 1897 } 1898 } 1899 1900 return this; 1901 1902} 1903//============================================================================= 1904//---------------------------StoreNode::make----------------------------------- 1905// Polymorphic factory method: 1906StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 1907 Compile* C = gvn.C; 1908 1909 switch (bt) { 1910 case T_BOOLEAN: 1911 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 1912 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 1913 case T_CHAR: 1914 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 1915 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 1916 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 1917 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 1918 case T_ADDRESS: 1919 case T_OBJECT: 1920#ifdef _LP64 1921 if (adr->bottom_type()->is_ptr_to_narrowoop() || 1922 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 1923 adr->bottom_type()->isa_rawptr())) { 1924 const TypePtr* type = val->bottom_type()->is_ptr(); 1925 Node* cp = EncodePNode::encode(&gvn, val); 1926 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp); 1927 } else 1928#endif 1929 { 1930 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 1931 } 1932 } 1933 ShouldNotReachHere(); 1934 return (StoreNode*)NULL; 1935} 1936 1937StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 1938 bool require_atomic = true; 1939 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 1940} 1941 1942 1943//--------------------------bottom_type---------------------------------------- 1944const Type *StoreNode::bottom_type() const { 1945 return Type::MEMORY; 1946} 1947 1948//------------------------------hash------------------------------------------- 1949uint StoreNode::hash() const { 1950 // unroll addition of interesting fields 1951 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 1952 1953 // Since they are not commoned, do not hash them: 1954 return NO_HASH; 1955} 1956 1957//------------------------------Ideal------------------------------------------ 1958// Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 1959// When a store immediately follows a relevant allocation/initialization, 1960// try to capture it into the initialization, or hoist it above. 1961Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1962 Node* p = MemNode::Ideal_common(phase, can_reshape); 1963 if (p) return (p == NodeSentinel) ? NULL : p; 1964 1965 Node* mem = in(MemNode::Memory); 1966 Node* address = in(MemNode::Address); 1967 1968 // Back-to-back stores to same address? Fold em up. 1969 // Generally unsafe if I have intervening uses... 1970 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 1971 // Looking at a dead closed cycle of memory? 1972 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 1973 1974 assert(Opcode() == mem->Opcode() || 1975 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 1976 "no mismatched stores, except on raw memory"); 1977 1978 if (mem->outcnt() == 1 && // check for intervening uses 1979 mem->as_Store()->memory_size() <= this->memory_size()) { 1980 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 1981 // For example, 'mem' might be the final state at a conditional return. 1982 // Or, 'mem' might be used by some node which is live at the same time 1983 // 'this' is live, which might be unschedulable. So, require exactly 1984 // ONE user, the 'this' store, until such time as we clone 'mem' for 1985 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 1986 if (can_reshape) { // (%%% is this an anachronism?) 1987 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 1988 phase->is_IterGVN()); 1989 } else { 1990 // It's OK to do this in the parser, since DU info is always accurate, 1991 // and the parser always refers to nodes via SafePointNode maps. 1992 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 1993 } 1994 return this; 1995 } 1996 } 1997 1998 // Capture an unaliased, unconditional, simple store into an initializer. 1999 // Or, if it is independent of the allocation, hoist it above the allocation. 2000 if (ReduceFieldZeroing && /*can_reshape &&*/ 2001 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2002 InitializeNode* init = mem->in(0)->as_Initialize(); 2003 intptr_t offset = init->can_capture_store(this, phase); 2004 if (offset > 0) { 2005 Node* moved = init->capture_store(this, offset, phase); 2006 // If the InitializeNode captured me, it made a raw copy of me, 2007 // and I need to disappear. 2008 if (moved != NULL) { 2009 // %%% hack to ensure that Ideal returns a new node: 2010 mem = MergeMemNode::make(phase->C, mem); 2011 return mem; // fold me away 2012 } 2013 } 2014 } 2015 2016 return NULL; // No further progress 2017} 2018 2019//------------------------------Value----------------------------------------- 2020const Type *StoreNode::Value( PhaseTransform *phase ) const { 2021 // Either input is TOP ==> the result is TOP 2022 const Type *t1 = phase->type( in(MemNode::Memory) ); 2023 if( t1 == Type::TOP ) return Type::TOP; 2024 const Type *t2 = phase->type( in(MemNode::Address) ); 2025 if( t2 == Type::TOP ) return Type::TOP; 2026 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2027 if( t3 == Type::TOP ) return Type::TOP; 2028 return Type::MEMORY; 2029} 2030 2031//------------------------------Identity--------------------------------------- 2032// Remove redundant stores: 2033// Store(m, p, Load(m, p)) changes to m. 2034// Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2035Node *StoreNode::Identity( PhaseTransform *phase ) { 2036 Node* mem = in(MemNode::Memory); 2037 Node* adr = in(MemNode::Address); 2038 Node* val = in(MemNode::ValueIn); 2039 2040 // Load then Store? Then the Store is useless 2041 if (val->is_Load() && 2042 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 2043 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 2044 val->as_Load()->store_Opcode() == Opcode()) { 2045 return mem; 2046 } 2047 2048 // Two stores in a row of the same value? 2049 if (mem->is_Store() && 2050 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 2051 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 2052 mem->Opcode() == Opcode()) { 2053 return mem; 2054 } 2055 2056 // Store of zero anywhere into a freshly-allocated object? 2057 // Then the store is useless. 2058 // (It must already have been captured by the InitializeNode.) 2059 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2060 // a newly allocated object is already all-zeroes everywhere 2061 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2062 return mem; 2063 } 2064 2065 // the store may also apply to zero-bits in an earlier object 2066 Node* prev_mem = find_previous_store(phase); 2067 // Steps (a), (b): Walk past independent stores to find an exact match. 2068 if (prev_mem != NULL) { 2069 Node* prev_val = can_see_stored_value(prev_mem, phase); 2070 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2071 // prev_val and val might differ by a cast; it would be good 2072 // to keep the more informative of the two. 2073 return mem; 2074 } 2075 } 2076 } 2077 2078 return this; 2079} 2080 2081//------------------------------match_edge------------------------------------- 2082// Do we Match on this edge index or not? Match only memory & value 2083uint StoreNode::match_edge(uint idx) const { 2084 return idx == MemNode::Address || idx == MemNode::ValueIn; 2085} 2086 2087//------------------------------cmp-------------------------------------------- 2088// Do not common stores up together. They generally have to be split 2089// back up anyways, so do not bother. 2090uint StoreNode::cmp( const Node &n ) const { 2091 return (&n == this); // Always fail except on self 2092} 2093 2094//------------------------------Ideal_masked_input----------------------------- 2095// Check for a useless mask before a partial-word store 2096// (StoreB ... (AndI valIn conIa) ) 2097// If (conIa & mask == mask) this simplifies to 2098// (StoreB ... (valIn) ) 2099Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2100 Node *val = in(MemNode::ValueIn); 2101 if( val->Opcode() == Op_AndI ) { 2102 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2103 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2104 set_req(MemNode::ValueIn, val->in(1)); 2105 return this; 2106 } 2107 } 2108 return NULL; 2109} 2110 2111 2112//------------------------------Ideal_sign_extended_input---------------------- 2113// Check for useless sign-extension before a partial-word store 2114// (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2115// If (conIL == conIR && conIR <= num_bits) this simplifies to 2116// (StoreB ... (valIn) ) 2117Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2118 Node *val = in(MemNode::ValueIn); 2119 if( val->Opcode() == Op_RShiftI ) { 2120 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2121 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2122 Node *shl = val->in(1); 2123 if( shl->Opcode() == Op_LShiftI ) { 2124 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2125 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2126 set_req(MemNode::ValueIn, shl->in(1)); 2127 return this; 2128 } 2129 } 2130 } 2131 } 2132 return NULL; 2133} 2134 2135//------------------------------value_never_loaded----------------------------------- 2136// Determine whether there are any possible loads of the value stored. 2137// For simplicity, we actually check if there are any loads from the 2138// address stored to, not just for loads of the value stored by this node. 2139// 2140bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2141 Node *adr = in(Address); 2142 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2143 if (adr_oop == NULL) 2144 return false; 2145 if (!adr_oop->is_instance_field()) 2146 return false; // if not a distinct instance, there may be aliases of the address 2147 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2148 Node *use = adr->fast_out(i); 2149 int opc = use->Opcode(); 2150 if (use->is_Load() || use->is_LoadStore()) { 2151 return false; 2152 } 2153 } 2154 return true; 2155} 2156 2157//============================================================================= 2158//------------------------------Ideal------------------------------------------ 2159// If the store is from an AND mask that leaves the low bits untouched, then 2160// we can skip the AND operation. If the store is from a sign-extension 2161// (a left shift, then right shift) we can skip both. 2162Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2163 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2164 if( progress != NULL ) return progress; 2165 2166 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2167 if( progress != NULL ) return progress; 2168 2169 // Finally check the default case 2170 return StoreNode::Ideal(phase, can_reshape); 2171} 2172 2173//============================================================================= 2174//------------------------------Ideal------------------------------------------ 2175// If the store is from an AND mask that leaves the low bits untouched, then 2176// we can skip the AND operation 2177Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2178 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2179 if( progress != NULL ) return progress; 2180 2181 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2182 if( progress != NULL ) return progress; 2183 2184 // Finally check the default case 2185 return StoreNode::Ideal(phase, can_reshape); 2186} 2187 2188//============================================================================= 2189//------------------------------Identity--------------------------------------- 2190Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2191 // No need to card mark when storing a null ptr 2192 Node* my_store = in(MemNode::OopStore); 2193 if (my_store->is_Store()) { 2194 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2195 if( t1 == TypePtr::NULL_PTR ) { 2196 return in(MemNode::Memory); 2197 } 2198 } 2199 return this; 2200} 2201 2202//------------------------------Value----------------------------------------- 2203const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2204 // Either input is TOP ==> the result is TOP 2205 const Type *t = phase->type( in(MemNode::Memory) ); 2206 if( t == Type::TOP ) return Type::TOP; 2207 t = phase->type( in(MemNode::Address) ); 2208 if( t == Type::TOP ) return Type::TOP; 2209 t = phase->type( in(MemNode::ValueIn) ); 2210 if( t == Type::TOP ) return Type::TOP; 2211 // If extra input is TOP ==> the result is TOP 2212 t = phase->type( in(MemNode::OopStore) ); 2213 if( t == Type::TOP ) return Type::TOP; 2214 2215 return StoreNode::Value( phase ); 2216} 2217 2218 2219//============================================================================= 2220//----------------------------------SCMemProjNode------------------------------ 2221const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2222{ 2223 return bottom_type(); 2224} 2225 2226//============================================================================= 2227LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2228 init_req(MemNode::Control, c ); 2229 init_req(MemNode::Memory , mem); 2230 init_req(MemNode::Address, adr); 2231 init_req(MemNode::ValueIn, val); 2232 init_req( ExpectedIn, ex ); 2233 init_class_id(Class_LoadStore); 2234 2235} 2236 2237//============================================================================= 2238//-------------------------------adr_type-------------------------------------- 2239// Do we Match on this edge index or not? Do not match memory 2240const TypePtr* ClearArrayNode::adr_type() const { 2241 Node *adr = in(3); 2242 return MemNode::calculate_adr_type(adr->bottom_type()); 2243} 2244 2245//------------------------------match_edge------------------------------------- 2246// Do we Match on this edge index or not? Do not match memory 2247uint ClearArrayNode::match_edge(uint idx) const { 2248 return idx > 1; 2249} 2250 2251//------------------------------Identity--------------------------------------- 2252// Clearing a zero length array does nothing 2253Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2254 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2255} 2256 2257//------------------------------Idealize--------------------------------------- 2258// Clearing a short array is faster with stores 2259Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2260 const int unit = BytesPerLong; 2261 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2262 if (!t) return NULL; 2263 if (!t->is_con()) return NULL; 2264 intptr_t raw_count = t->get_con(); 2265 intptr_t size = raw_count; 2266 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2267 // Clearing nothing uses the Identity call. 2268 // Negative clears are possible on dead ClearArrays 2269 // (see jck test stmt114.stmt11402.val). 2270 if (size <= 0 || size % unit != 0) return NULL; 2271 intptr_t count = size / unit; 2272 // Length too long; use fast hardware clear 2273 if (size > Matcher::init_array_short_size) return NULL; 2274 Node *mem = in(1); 2275 if( phase->type(mem)==Type::TOP ) return NULL; 2276 Node *adr = in(3); 2277 const Type* at = phase->type(adr); 2278 if( at==Type::TOP ) return NULL; 2279 const TypePtr* atp = at->isa_ptr(); 2280 // adjust atp to be the correct array element address type 2281 if (atp == NULL) atp = TypePtr::BOTTOM; 2282 else atp = atp->add_offset(Type::OffsetBot); 2283 // Get base for derived pointer purposes 2284 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2285 Node *base = adr->in(1); 2286 2287 Node *zero = phase->makecon(TypeLong::ZERO); 2288 Node *off = phase->MakeConX(BytesPerLong); 2289 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2290 count--; 2291 while( count-- ) { 2292 mem = phase->transform(mem); 2293 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2294 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2295 } 2296 return mem; 2297} 2298 2299//----------------------------clear_memory------------------------------------- 2300// Generate code to initialize object storage to zero. 2301Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2302 intptr_t start_offset, 2303 Node* end_offset, 2304 PhaseGVN* phase) { 2305 Compile* C = phase->C; 2306 intptr_t offset = start_offset; 2307 2308 int unit = BytesPerLong; 2309 if ((offset % unit) != 0) { 2310 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2311 adr = phase->transform(adr); 2312 const TypePtr* atp = TypeRawPtr::BOTTOM; 2313 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2314 mem = phase->transform(mem); 2315 offset += BytesPerInt; 2316 } 2317 assert((offset % unit) == 0, ""); 2318 2319 // Initialize the remaining stuff, if any, with a ClearArray. 2320 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2321} 2322 2323Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2324 Node* start_offset, 2325 Node* end_offset, 2326 PhaseGVN* phase) { 2327 if (start_offset == end_offset) { 2328 // nothing to do 2329 return mem; 2330 } 2331 2332 Compile* C = phase->C; 2333 int unit = BytesPerLong; 2334 Node* zbase = start_offset; 2335 Node* zend = end_offset; 2336 2337 // Scale to the unit required by the CPU: 2338 if (!Matcher::init_array_count_is_in_bytes) { 2339 Node* shift = phase->intcon(exact_log2(unit)); 2340 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2341 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2342 } 2343 2344 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2345 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2346 2347 // Bulk clear double-words 2348 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2349 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2350 return phase->transform(mem); 2351} 2352 2353Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2354 intptr_t start_offset, 2355 intptr_t end_offset, 2356 PhaseGVN* phase) { 2357 if (start_offset == end_offset) { 2358 // nothing to do 2359 return mem; 2360 } 2361 2362 Compile* C = phase->C; 2363 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2364 intptr_t done_offset = end_offset; 2365 if ((done_offset % BytesPerLong) != 0) { 2366 done_offset -= BytesPerInt; 2367 } 2368 if (done_offset > start_offset) { 2369 mem = clear_memory(ctl, mem, dest, 2370 start_offset, phase->MakeConX(done_offset), phase); 2371 } 2372 if (done_offset < end_offset) { // emit the final 32-bit store 2373 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2374 adr = phase->transform(adr); 2375 const TypePtr* atp = TypeRawPtr::BOTTOM; 2376 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2377 mem = phase->transform(mem); 2378 done_offset += BytesPerInt; 2379 } 2380 assert(done_offset == end_offset, ""); 2381 return mem; 2382} 2383 2384//============================================================================= 2385// Do we match on this edge? No memory edges 2386uint StrCompNode::match_edge(uint idx) const { 2387 return idx == 5 || idx == 6; 2388} 2389 2390//------------------------------Ideal------------------------------------------ 2391// Return a node which is more "ideal" than the current node. Strip out 2392// control copies 2393Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2394 return remove_dead_region(phase, can_reshape) ? this : NULL; 2395} 2396 2397 2398//============================================================================= 2399MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2400 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2401 _adr_type(C->get_adr_type(alias_idx)) 2402{ 2403 init_class_id(Class_MemBar); 2404 Node* top = C->top(); 2405 init_req(TypeFunc::I_O,top); 2406 init_req(TypeFunc::FramePtr,top); 2407 init_req(TypeFunc::ReturnAdr,top); 2408 if (precedent != NULL) 2409 init_req(TypeFunc::Parms, precedent); 2410} 2411 2412//------------------------------cmp-------------------------------------------- 2413uint MemBarNode::hash() const { return NO_HASH; } 2414uint MemBarNode::cmp( const Node &n ) const { 2415 return (&n == this); // Always fail except on self 2416} 2417 2418//------------------------------make------------------------------------------- 2419MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2420 int len = Precedent + (pn == NULL? 0: 1); 2421 switch (opcode) { 2422 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2423 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2424 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2425 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2426 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2427 default: ShouldNotReachHere(); return NULL; 2428 } 2429} 2430 2431//------------------------------Ideal------------------------------------------ 2432// Return a node which is more "ideal" than the current node. Strip out 2433// control copies 2434Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2435 if (remove_dead_region(phase, can_reshape)) return this; 2436 return NULL; 2437} 2438 2439//------------------------------Value------------------------------------------ 2440const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2441 if( !in(0) ) return Type::TOP; 2442 if( phase->type(in(0)) == Type::TOP ) 2443 return Type::TOP; 2444 return TypeTuple::MEMBAR; 2445} 2446 2447//------------------------------match------------------------------------------ 2448// Construct projections for memory. 2449Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2450 switch (proj->_con) { 2451 case TypeFunc::Control: 2452 case TypeFunc::Memory: 2453 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2454 } 2455 ShouldNotReachHere(); 2456 return NULL; 2457} 2458 2459//===========================InitializeNode==================================== 2460// SUMMARY: 2461// This node acts as a memory barrier on raw memory, after some raw stores. 2462// The 'cooked' oop value feeds from the Initialize, not the Allocation. 2463// The Initialize can 'capture' suitably constrained stores as raw inits. 2464// It can coalesce related raw stores into larger units (called 'tiles'). 2465// It can avoid zeroing new storage for memory units which have raw inits. 2466// At macro-expansion, it is marked 'complete', and does not optimize further. 2467// 2468// EXAMPLE: 2469// The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2470// ctl = incoming control; mem* = incoming memory 2471// (Note: A star * on a memory edge denotes I/O and other standard edges.) 2472// First allocate uninitialized memory and fill in the header: 2473// alloc = (Allocate ctl mem* 16 #short[].klass ...) 2474// ctl := alloc.Control; mem* := alloc.Memory* 2475// rawmem = alloc.Memory; rawoop = alloc.RawAddress 2476// Then initialize to zero the non-header parts of the raw memory block: 2477// init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2478// ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2479// After the initialize node executes, the object is ready for service: 2480// oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2481// Suppose its body is immediately initialized as {1,2}: 2482// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2483// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2484// mem.SLICE(#short[*]) := store2 2485// 2486// DETAILS: 2487// An InitializeNode collects and isolates object initialization after 2488// an AllocateNode and before the next possible safepoint. As a 2489// memory barrier (MemBarNode), it keeps critical stores from drifting 2490// down past any safepoint or any publication of the allocation. 2491// Before this barrier, a newly-allocated object may have uninitialized bits. 2492// After this barrier, it may be treated as a real oop, and GC is allowed. 2493// 2494// The semantics of the InitializeNode include an implicit zeroing of 2495// the new object from object header to the end of the object. 2496// (The object header and end are determined by the AllocateNode.) 2497// 2498// Certain stores may be added as direct inputs to the InitializeNode. 2499// These stores must update raw memory, and they must be to addresses 2500// derived from the raw address produced by AllocateNode, and with 2501// a constant offset. They must be ordered by increasing offset. 2502// The first one is at in(RawStores), the last at in(req()-1). 2503// Unlike most memory operations, they are not linked in a chain, 2504// but are displayed in parallel as users of the rawmem output of 2505// the allocation. 2506// 2507// (See comments in InitializeNode::capture_store, which continue 2508// the example given above.) 2509// 2510// When the associated Allocate is macro-expanded, the InitializeNode 2511// may be rewritten to optimize collected stores. A ClearArrayNode 2512// may also be created at that point to represent any required zeroing. 2513// The InitializeNode is then marked 'complete', prohibiting further 2514// capturing of nearby memory operations. 2515// 2516// During macro-expansion, all captured initializations which store 2517// constant values of 32 bits or smaller are coalesced (if advantagous) 2518// into larger 'tiles' 32 or 64 bits. This allows an object to be 2519// initialized in fewer memory operations. Memory words which are 2520// covered by neither tiles nor non-constant stores are pre-zeroed 2521// by explicit stores of zero. (The code shape happens to do all 2522// zeroing first, then all other stores, with both sequences occurring 2523// in order of ascending offsets.) 2524// 2525// Alternatively, code may be inserted between an AllocateNode and its 2526// InitializeNode, to perform arbitrary initialization of the new object. 2527// E.g., the object copying intrinsics insert complex data transfers here. 2528// The initialization must then be marked as 'complete' disable the 2529// built-in zeroing semantics and the collection of initializing stores. 2530// 2531// While an InitializeNode is incomplete, reads from the memory state 2532// produced by it are optimizable if they match the control edge and 2533// new oop address associated with the allocation/initialization. 2534// They return a stored value (if the offset matches) or else zero. 2535// A write to the memory state, if it matches control and address, 2536// and if it is to a constant offset, may be 'captured' by the 2537// InitializeNode. It is cloned as a raw memory operation and rewired 2538// inside the initialization, to the raw oop produced by the allocation. 2539// Operations on addresses which are provably distinct (e.g., to 2540// other AllocateNodes) are allowed to bypass the initialization. 2541// 2542// The effect of all this is to consolidate object initialization 2543// (both arrays and non-arrays, both piecewise and bulk) into a 2544// single location, where it can be optimized as a unit. 2545// 2546// Only stores with an offset less than TrackedInitializationLimit words 2547// will be considered for capture by an InitializeNode. This puts a 2548// reasonable limit on the complexity of optimized initializations. 2549 2550//---------------------------InitializeNode------------------------------------ 2551InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2552 : _is_complete(false), 2553 MemBarNode(C, adr_type, rawoop) 2554{ 2555 init_class_id(Class_Initialize); 2556 2557 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2558 assert(in(RawAddress) == rawoop, "proper init"); 2559 // Note: allocation() can be NULL, for secondary initialization barriers 2560} 2561 2562// Since this node is not matched, it will be processed by the 2563// register allocator. Declare that there are no constraints 2564// on the allocation of the RawAddress edge. 2565const RegMask &InitializeNode::in_RegMask(uint idx) const { 2566 // This edge should be set to top, by the set_complete. But be conservative. 2567 if (idx == InitializeNode::RawAddress) 2568 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2569 return RegMask::Empty; 2570} 2571 2572Node* InitializeNode::memory(uint alias_idx) { 2573 Node* mem = in(Memory); 2574 if (mem->is_MergeMem()) { 2575 return mem->as_MergeMem()->memory_at(alias_idx); 2576 } else { 2577 // incoming raw memory is not split 2578 return mem; 2579 } 2580} 2581 2582bool InitializeNode::is_non_zero() { 2583 if (is_complete()) return false; 2584 remove_extra_zeroes(); 2585 return (req() > RawStores); 2586} 2587 2588void InitializeNode::set_complete(PhaseGVN* phase) { 2589 assert(!is_complete(), "caller responsibility"); 2590 _is_complete = true; 2591 2592 // After this node is complete, it contains a bunch of 2593 // raw-memory initializations. There is no need for 2594 // it to have anything to do with non-raw memory effects. 2595 // Therefore, tell all non-raw users to re-optimize themselves, 2596 // after skipping the memory effects of this initialization. 2597 PhaseIterGVN* igvn = phase->is_IterGVN(); 2598 if (igvn) igvn->add_users_to_worklist(this); 2599} 2600 2601// convenience function 2602// return false if the init contains any stores already 2603bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2604 InitializeNode* init = initialization(); 2605 if (init == NULL || init->is_complete()) return false; 2606 init->remove_extra_zeroes(); 2607 // for now, if this allocation has already collected any inits, bail: 2608 if (init->is_non_zero()) return false; 2609 init->set_complete(phase); 2610 return true; 2611} 2612 2613void InitializeNode::remove_extra_zeroes() { 2614 if (req() == RawStores) return; 2615 Node* zmem = zero_memory(); 2616 uint fill = RawStores; 2617 for (uint i = fill; i < req(); i++) { 2618 Node* n = in(i); 2619 if (n->is_top() || n == zmem) continue; // skip 2620 if (fill < i) set_req(fill, n); // compact 2621 ++fill; 2622 } 2623 // delete any empty spaces created: 2624 while (fill < req()) { 2625 del_req(fill); 2626 } 2627} 2628 2629// Helper for remembering which stores go with which offsets. 2630intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2631 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2632 intptr_t offset = -1; 2633 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2634 phase, offset); 2635 if (base == NULL) return -1; // something is dead, 2636 if (offset < 0) return -1; // dead, dead 2637 return offset; 2638} 2639 2640// Helper for proving that an initialization expression is 2641// "simple enough" to be folded into an object initialization. 2642// Attempts to prove that a store's initial value 'n' can be captured 2643// within the initialization without creating a vicious cycle, such as: 2644// { Foo p = new Foo(); p.next = p; } 2645// True for constants and parameters and small combinations thereof. 2646bool InitializeNode::detect_init_independence(Node* n, 2647 bool st_is_pinned, 2648 int& count) { 2649 if (n == NULL) return true; // (can this really happen?) 2650 if (n->is_Proj()) n = n->in(0); 2651 if (n == this) return false; // found a cycle 2652 if (n->is_Con()) return true; 2653 if (n->is_Start()) return true; // params, etc., are OK 2654 if (n->is_Root()) return true; // even better 2655 2656 Node* ctl = n->in(0); 2657 if (ctl != NULL && !ctl->is_top()) { 2658 if (ctl->is_Proj()) ctl = ctl->in(0); 2659 if (ctl == this) return false; 2660 2661 // If we already know that the enclosing memory op is pinned right after 2662 // the init, then any control flow that the store has picked up 2663 // must have preceded the init, or else be equal to the init. 2664 // Even after loop optimizations (which might change control edges) 2665 // a store is never pinned *before* the availability of its inputs. 2666 if (!MemNode::all_controls_dominate(n, this)) 2667 return false; // failed to prove a good control 2668 2669 } 2670 2671 // Check data edges for possible dependencies on 'this'. 2672 if ((count += 1) > 20) return false; // complexity limit 2673 for (uint i = 1; i < n->req(); i++) { 2674 Node* m = n->in(i); 2675 if (m == NULL || m == n || m->is_top()) continue; 2676 uint first_i = n->find_edge(m); 2677 if (i != first_i) continue; // process duplicate edge just once 2678 if (!detect_init_independence(m, st_is_pinned, count)) { 2679 return false; 2680 } 2681 } 2682 2683 return true; 2684} 2685 2686// Here are all the checks a Store must pass before it can be moved into 2687// an initialization. Returns zero if a check fails. 2688// On success, returns the (constant) offset to which the store applies, 2689// within the initialized memory. 2690intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2691 const int FAIL = 0; 2692 if (st->req() != MemNode::ValueIn + 1) 2693 return FAIL; // an inscrutable StoreNode (card mark?) 2694 Node* ctl = st->in(MemNode::Control); 2695 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2696 return FAIL; // must be unconditional after the initialization 2697 Node* mem = st->in(MemNode::Memory); 2698 if (!(mem->is_Proj() && mem->in(0) == this)) 2699 return FAIL; // must not be preceded by other stores 2700 Node* adr = st->in(MemNode::Address); 2701 intptr_t offset; 2702 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2703 if (alloc == NULL) 2704 return FAIL; // inscrutable address 2705 if (alloc != allocation()) 2706 return FAIL; // wrong allocation! (store needs to float up) 2707 Node* val = st->in(MemNode::ValueIn); 2708 int complexity_count = 0; 2709 if (!detect_init_independence(val, true, complexity_count)) 2710 return FAIL; // stored value must be 'simple enough' 2711 2712 return offset; // success 2713} 2714 2715// Find the captured store in(i) which corresponds to the range 2716// [start..start+size) in the initialized object. 2717// If there is one, return its index i. If there isn't, return the 2718// negative of the index where it should be inserted. 2719// Return 0 if the queried range overlaps an initialization boundary 2720// or if dead code is encountered. 2721// If size_in_bytes is zero, do not bother with overlap checks. 2722int InitializeNode::captured_store_insertion_point(intptr_t start, 2723 int size_in_bytes, 2724 PhaseTransform* phase) { 2725 const int FAIL = 0, MAX_STORE = BytesPerLong; 2726 2727 if (is_complete()) 2728 return FAIL; // arraycopy got here first; punt 2729 2730 assert(allocation() != NULL, "must be present"); 2731 2732 // no negatives, no header fields: 2733 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 2734 2735 // after a certain size, we bail out on tracking all the stores: 2736 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2737 if (start >= ti_limit) return FAIL; 2738 2739 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 2740 if (i >= limit) return -(int)i; // not found; here is where to put it 2741 2742 Node* st = in(i); 2743 intptr_t st_off = get_store_offset(st, phase); 2744 if (st_off < 0) { 2745 if (st != zero_memory()) { 2746 return FAIL; // bail out if there is dead garbage 2747 } 2748 } else if (st_off > start) { 2749 // ...we are done, since stores are ordered 2750 if (st_off < start + size_in_bytes) { 2751 return FAIL; // the next store overlaps 2752 } 2753 return -(int)i; // not found; here is where to put it 2754 } else if (st_off < start) { 2755 if (size_in_bytes != 0 && 2756 start < st_off + MAX_STORE && 2757 start < st_off + st->as_Store()->memory_size()) { 2758 return FAIL; // the previous store overlaps 2759 } 2760 } else { 2761 if (size_in_bytes != 0 && 2762 st->as_Store()->memory_size() != size_in_bytes) { 2763 return FAIL; // mismatched store size 2764 } 2765 return i; 2766 } 2767 2768 ++i; 2769 } 2770} 2771 2772// Look for a captured store which initializes at the offset 'start' 2773// with the given size. If there is no such store, and no other 2774// initialization interferes, then return zero_memory (the memory 2775// projection of the AllocateNode). 2776Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 2777 PhaseTransform* phase) { 2778 assert(stores_are_sane(phase), ""); 2779 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2780 if (i == 0) { 2781 return NULL; // something is dead 2782 } else if (i < 0) { 2783 return zero_memory(); // just primordial zero bits here 2784 } else { 2785 Node* st = in(i); // here is the store at this position 2786 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 2787 return st; 2788 } 2789} 2790 2791// Create, as a raw pointer, an address within my new object at 'offset'. 2792Node* InitializeNode::make_raw_address(intptr_t offset, 2793 PhaseTransform* phase) { 2794 Node* addr = in(RawAddress); 2795 if (offset != 0) { 2796 Compile* C = phase->C; 2797 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 2798 phase->MakeConX(offset)) ); 2799 } 2800 return addr; 2801} 2802 2803// Clone the given store, converting it into a raw store 2804// initializing a field or element of my new object. 2805// Caller is responsible for retiring the original store, 2806// with subsume_node or the like. 2807// 2808// From the example above InitializeNode::InitializeNode, 2809// here are the old stores to be captured: 2810// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2811// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2812// 2813// Here is the changed code; note the extra edges on init: 2814// alloc = (Allocate ...) 2815// rawoop = alloc.RawAddress 2816// rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 2817// rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 2818// init = (Initialize alloc.Control alloc.Memory rawoop 2819// rawstore1 rawstore2) 2820// 2821Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 2822 PhaseTransform* phase) { 2823 assert(stores_are_sane(phase), ""); 2824 2825 if (start < 0) return NULL; 2826 assert(can_capture_store(st, phase) == start, "sanity"); 2827 2828 Compile* C = phase->C; 2829 int size_in_bytes = st->memory_size(); 2830 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2831 if (i == 0) return NULL; // bail out 2832 Node* prev_mem = NULL; // raw memory for the captured store 2833 if (i > 0) { 2834 prev_mem = in(i); // there is a pre-existing store under this one 2835 set_req(i, C->top()); // temporarily disconnect it 2836 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 2837 } else { 2838 i = -i; // no pre-existing store 2839 prev_mem = zero_memory(); // a slice of the newly allocated object 2840 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 2841 set_req(--i, C->top()); // reuse this edge; it has been folded away 2842 else 2843 ins_req(i, C->top()); // build a new edge 2844 } 2845 Node* new_st = st->clone(); 2846 new_st->set_req(MemNode::Control, in(Control)); 2847 new_st->set_req(MemNode::Memory, prev_mem); 2848 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 2849 new_st = phase->transform(new_st); 2850 2851 // At this point, new_st might have swallowed a pre-existing store 2852 // at the same offset, or perhaps new_st might have disappeared, 2853 // if it redundantly stored the same value (or zero to fresh memory). 2854 2855 // In any case, wire it in: 2856 set_req(i, new_st); 2857 2858 // The caller may now kill the old guy. 2859 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 2860 assert(check_st == new_st || check_st == NULL, "must be findable"); 2861 assert(!is_complete(), ""); 2862 return new_st; 2863} 2864 2865static bool store_constant(jlong* tiles, int num_tiles, 2866 intptr_t st_off, int st_size, 2867 jlong con) { 2868 if ((st_off & (st_size-1)) != 0) 2869 return false; // strange store offset (assume size==2**N) 2870 address addr = (address)tiles + st_off; 2871 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 2872 switch (st_size) { 2873 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 2874 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 2875 case sizeof(jint): *(jint*) addr = (jint) con; break; 2876 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 2877 default: return false; // strange store size (detect size!=2**N here) 2878 } 2879 return true; // return success to caller 2880} 2881 2882// Coalesce subword constants into int constants and possibly 2883// into long constants. The goal, if the CPU permits, 2884// is to initialize the object with a small number of 64-bit tiles. 2885// Also, convert floating-point constants to bit patterns. 2886// Non-constants are not relevant to this pass. 2887// 2888// In terms of the running example on InitializeNode::InitializeNode 2889// and InitializeNode::capture_store, here is the transformation 2890// of rawstore1 and rawstore2 into rawstore12: 2891// alloc = (Allocate ...) 2892// rawoop = alloc.RawAddress 2893// tile12 = 0x00010002 2894// rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 2895// init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 2896// 2897void 2898InitializeNode::coalesce_subword_stores(intptr_t header_size, 2899 Node* size_in_bytes, 2900 PhaseGVN* phase) { 2901 Compile* C = phase->C; 2902 2903 assert(stores_are_sane(phase), ""); 2904 // Note: After this pass, they are not completely sane, 2905 // since there may be some overlaps. 2906 2907 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 2908 2909 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2910 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 2911 size_limit = MIN2(size_limit, ti_limit); 2912 size_limit = align_size_up(size_limit, BytesPerLong); 2913 int num_tiles = size_limit / BytesPerLong; 2914 2915 // allocate space for the tile map: 2916 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 2917 jlong tiles_buf[small_len]; 2918 Node* nodes_buf[small_len]; 2919 jlong inits_buf[small_len]; 2920 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 2921 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2922 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 2923 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 2924 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 2925 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2926 // tiles: exact bitwise model of all primitive constants 2927 // nodes: last constant-storing node subsumed into the tiles model 2928 // inits: which bytes (in each tile) are touched by any initializations 2929 2930 //// Pass A: Fill in the tile model with any relevant stores. 2931 2932 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 2933 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 2934 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 2935 Node* zmem = zero_memory(); // initially zero memory state 2936 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 2937 Node* st = in(i); 2938 intptr_t st_off = get_store_offset(st, phase); 2939 2940 // Figure out the store's offset and constant value: 2941 if (st_off < header_size) continue; //skip (ignore header) 2942 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 2943 int st_size = st->as_Store()->memory_size(); 2944 if (st_off + st_size > size_limit) break; 2945 2946 // Record which bytes are touched, whether by constant or not. 2947 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 2948 continue; // skip (strange store size) 2949 2950 const Type* val = phase->type(st->in(MemNode::ValueIn)); 2951 if (!val->singleton()) continue; //skip (non-con store) 2952 BasicType type = val->basic_type(); 2953 2954 jlong con = 0; 2955 switch (type) { 2956 case T_INT: con = val->is_int()->get_con(); break; 2957 case T_LONG: con = val->is_long()->get_con(); break; 2958 case T_FLOAT: con = jint_cast(val->getf()); break; 2959 case T_DOUBLE: con = jlong_cast(val->getd()); break; 2960 default: continue; //skip (odd store type) 2961 } 2962 2963 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 2964 st->Opcode() == Op_StoreL) { 2965 continue; // This StoreL is already optimal. 2966 } 2967 2968 // Store down the constant. 2969 store_constant(tiles, num_tiles, st_off, st_size, con); 2970 2971 intptr_t j = st_off >> LogBytesPerLong; 2972 2973 if (type == T_INT && st_size == BytesPerInt 2974 && (st_off & BytesPerInt) == BytesPerInt) { 2975 jlong lcon = tiles[j]; 2976 if (!Matcher::isSimpleConstant64(lcon) && 2977 st->Opcode() == Op_StoreI) { 2978 // This StoreI is already optimal by itself. 2979 jint* intcon = (jint*) &tiles[j]; 2980 intcon[1] = 0; // undo the store_constant() 2981 2982 // If the previous store is also optimal by itself, back up and 2983 // undo the action of the previous loop iteration... if we can. 2984 // But if we can't, just let the previous half take care of itself. 2985 st = nodes[j]; 2986 st_off -= BytesPerInt; 2987 con = intcon[0]; 2988 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 2989 assert(st_off >= header_size, "still ignoring header"); 2990 assert(get_store_offset(st, phase) == st_off, "must be"); 2991 assert(in(i-1) == zmem, "must be"); 2992 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 2993 assert(con == tcon->is_int()->get_con(), "must be"); 2994 // Undo the effects of the previous loop trip, which swallowed st: 2995 intcon[0] = 0; // undo store_constant() 2996 set_req(i-1, st); // undo set_req(i, zmem) 2997 nodes[j] = NULL; // undo nodes[j] = st 2998 --old_subword; // undo ++old_subword 2999 } 3000 continue; // This StoreI is already optimal. 3001 } 3002 } 3003 3004 // This store is not needed. 3005 set_req(i, zmem); 3006 nodes[j] = st; // record for the moment 3007 if (st_size < BytesPerLong) // something has changed 3008 ++old_subword; // includes int/float, but who's counting... 3009 else ++old_long; 3010 } 3011 3012 if ((old_subword + old_long) == 0) 3013 return; // nothing more to do 3014 3015 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3016 // Be sure to insert them before overlapping non-constant stores. 3017 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3018 for (int j = 0; j < num_tiles; j++) { 3019 jlong con = tiles[j]; 3020 jlong init = inits[j]; 3021 if (con == 0) continue; 3022 jint con0, con1; // split the constant, address-wise 3023 jint init0, init1; // split the init map, address-wise 3024 { union { jlong con; jint intcon[2]; } u; 3025 u.con = con; 3026 con0 = u.intcon[0]; 3027 con1 = u.intcon[1]; 3028 u.con = init; 3029 init0 = u.intcon[0]; 3030 init1 = u.intcon[1]; 3031 } 3032 3033 Node* old = nodes[j]; 3034 assert(old != NULL, "need the prior store"); 3035 intptr_t offset = (j * BytesPerLong); 3036 3037 bool split = !Matcher::isSimpleConstant64(con); 3038 3039 if (offset < header_size) { 3040 assert(offset + BytesPerInt >= header_size, "second int counts"); 3041 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3042 split = true; // only the second word counts 3043 // Example: int a[] = { 42 ... } 3044 } else if (con0 == 0 && init0 == -1) { 3045 split = true; // first word is covered by full inits 3046 // Example: int a[] = { ... foo(), 42 ... } 3047 } else if (con1 == 0 && init1 == -1) { 3048 split = true; // second word is covered by full inits 3049 // Example: int a[] = { ... 42, foo() ... } 3050 } 3051 3052 // Here's a case where init0 is neither 0 nor -1: 3053 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3054 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3055 // In this case the tile is not split; it is (jlong)42. 3056 // The big tile is stored down, and then the foo() value is inserted. 3057 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3058 3059 Node* ctl = old->in(MemNode::Control); 3060 Node* adr = make_raw_address(offset, phase); 3061 const TypePtr* atp = TypeRawPtr::BOTTOM; 3062 3063 // One or two coalesced stores to plop down. 3064 Node* st[2]; 3065 intptr_t off[2]; 3066 int nst = 0; 3067 if (!split) { 3068 ++new_long; 3069 off[nst] = offset; 3070 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3071 phase->longcon(con), T_LONG); 3072 } else { 3073 // Omit either if it is a zero. 3074 if (con0 != 0) { 3075 ++new_int; 3076 off[nst] = offset; 3077 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3078 phase->intcon(con0), T_INT); 3079 } 3080 if (con1 != 0) { 3081 ++new_int; 3082 offset += BytesPerInt; 3083 adr = make_raw_address(offset, phase); 3084 off[nst] = offset; 3085 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3086 phase->intcon(con1), T_INT); 3087 } 3088 } 3089 3090 // Insert second store first, then the first before the second. 3091 // Insert each one just before any overlapping non-constant stores. 3092 while (nst > 0) { 3093 Node* st1 = st[--nst]; 3094 C->copy_node_notes_to(st1, old); 3095 st1 = phase->transform(st1); 3096 offset = off[nst]; 3097 assert(offset >= header_size, "do not smash header"); 3098 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3099 guarantee(ins_idx != 0, "must re-insert constant store"); 3100 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3101 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3102 set_req(--ins_idx, st1); 3103 else 3104 ins_req(ins_idx, st1); 3105 } 3106 } 3107 3108 if (PrintCompilation && WizardMode) 3109 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3110 old_subword, old_long, new_int, new_long); 3111 if (C->log() != NULL) 3112 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3113 old_subword, old_long, new_int, new_long); 3114 3115 // Clean up any remaining occurrences of zmem: 3116 remove_extra_zeroes(); 3117} 3118 3119// Explore forward from in(start) to find the first fully initialized 3120// word, and return its offset. Skip groups of subword stores which 3121// together initialize full words. If in(start) is itself part of a 3122// fully initialized word, return the offset of in(start). If there 3123// are no following full-word stores, or if something is fishy, return 3124// a negative value. 3125intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3126 int int_map = 0; 3127 intptr_t int_map_off = 0; 3128 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3129 3130 for (uint i = start, limit = req(); i < limit; i++) { 3131 Node* st = in(i); 3132 3133 intptr_t st_off = get_store_offset(st, phase); 3134 if (st_off < 0) break; // return conservative answer 3135 3136 int st_size = st->as_Store()->memory_size(); 3137 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3138 return st_off; // we found a complete word init 3139 } 3140 3141 // update the map: 3142 3143 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3144 if (this_int_off != int_map_off) { 3145 // reset the map: 3146 int_map = 0; 3147 int_map_off = this_int_off; 3148 } 3149 3150 int subword_off = st_off - this_int_off; 3151 int_map |= right_n_bits(st_size) << subword_off; 3152 if ((int_map & FULL_MAP) == FULL_MAP) { 3153 return this_int_off; // we found a complete word init 3154 } 3155 3156 // Did this store hit or cross the word boundary? 3157 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3158 if (next_int_off == this_int_off + BytesPerInt) { 3159 // We passed the current int, without fully initializing it. 3160 int_map_off = next_int_off; 3161 int_map >>= BytesPerInt; 3162 } else if (next_int_off > this_int_off + BytesPerInt) { 3163 // We passed the current and next int. 3164 return this_int_off + BytesPerInt; 3165 } 3166 } 3167 3168 return -1; 3169} 3170 3171 3172// Called when the associated AllocateNode is expanded into CFG. 3173// At this point, we may perform additional optimizations. 3174// Linearize the stores by ascending offset, to make memory 3175// activity as coherent as possible. 3176Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3177 intptr_t header_size, 3178 Node* size_in_bytes, 3179 PhaseGVN* phase) { 3180 assert(!is_complete(), "not already complete"); 3181 assert(stores_are_sane(phase), ""); 3182 assert(allocation() != NULL, "must be present"); 3183 3184 remove_extra_zeroes(); 3185 3186 if (ReduceFieldZeroing || ReduceBulkZeroing) 3187 // reduce instruction count for common initialization patterns 3188 coalesce_subword_stores(header_size, size_in_bytes, phase); 3189 3190 Node* zmem = zero_memory(); // initially zero memory state 3191 Node* inits = zmem; // accumulating a linearized chain of inits 3192 #ifdef ASSERT 3193 intptr_t first_offset = allocation()->minimum_header_size(); 3194 intptr_t last_init_off = first_offset; // previous init offset 3195 intptr_t last_init_end = first_offset; // previous init offset+size 3196 intptr_t last_tile_end = first_offset; // previous tile offset+size 3197 #endif 3198 intptr_t zeroes_done = header_size; 3199 3200 bool do_zeroing = true; // we might give up if inits are very sparse 3201 int big_init_gaps = 0; // how many large gaps have we seen? 3202 3203 if (ZeroTLAB) do_zeroing = false; 3204 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3205 3206 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3207 Node* st = in(i); 3208 intptr_t st_off = get_store_offset(st, phase); 3209 if (st_off < 0) 3210 break; // unknown junk in the inits 3211 if (st->in(MemNode::Memory) != zmem) 3212 break; // complicated store chains somehow in list 3213 3214 int st_size = st->as_Store()->memory_size(); 3215 intptr_t next_init_off = st_off + st_size; 3216 3217 if (do_zeroing && zeroes_done < next_init_off) { 3218 // See if this store needs a zero before it or under it. 3219 intptr_t zeroes_needed = st_off; 3220 3221 if (st_size < BytesPerInt) { 3222 // Look for subword stores which only partially initialize words. 3223 // If we find some, we must lay down some word-level zeroes first, 3224 // underneath the subword stores. 3225 // 3226 // Examples: 3227 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3228 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3229 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3230 // 3231 // Note: coalesce_subword_stores may have already done this, 3232 // if it was prompted by constant non-zero subword initializers. 3233 // But this case can still arise with non-constant stores. 3234 3235 intptr_t next_full_store = find_next_fullword_store(i, phase); 3236 3237 // In the examples above: 3238 // in(i) p q r s x y z 3239 // st_off 12 13 14 15 12 13 14 3240 // st_size 1 1 1 1 1 1 1 3241 // next_full_s. 12 16 16 16 16 16 16 3242 // z's_done 12 16 16 16 12 16 12 3243 // z's_needed 12 16 16 16 16 16 16 3244 // zsize 0 0 0 0 4 0 4 3245 if (next_full_store < 0) { 3246 // Conservative tack: Zero to end of current word. 3247 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3248 } else { 3249 // Zero to beginning of next fully initialized word. 3250 // Or, don't zero at all, if we are already in that word. 3251 assert(next_full_store >= zeroes_needed, "must go forward"); 3252 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3253 zeroes_needed = next_full_store; 3254 } 3255 } 3256 3257 if (zeroes_needed > zeroes_done) { 3258 intptr_t zsize = zeroes_needed - zeroes_done; 3259 // Do some incremental zeroing on rawmem, in parallel with inits. 3260 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3261 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3262 zeroes_done, zeroes_needed, 3263 phase); 3264 zeroes_done = zeroes_needed; 3265 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3266 do_zeroing = false; // leave the hole, next time 3267 } 3268 } 3269 3270 // Collect the store and move on: 3271 st->set_req(MemNode::Memory, inits); 3272 inits = st; // put it on the linearized chain 3273 set_req(i, zmem); // unhook from previous position 3274 3275 if (zeroes_done == st_off) 3276 zeroes_done = next_init_off; 3277 3278 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3279 3280 #ifdef ASSERT 3281 // Various order invariants. Weaker than stores_are_sane because 3282 // a large constant tile can be filled in by smaller non-constant stores. 3283 assert(st_off >= last_init_off, "inits do not reverse"); 3284 last_init_off = st_off; 3285 const Type* val = NULL; 3286 if (st_size >= BytesPerInt && 3287 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3288 (int)val->basic_type() < (int)T_OBJECT) { 3289 assert(st_off >= last_tile_end, "tiles do not overlap"); 3290 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3291 last_tile_end = MAX2(last_tile_end, next_init_off); 3292 } else { 3293 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3294 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3295 assert(st_off >= last_init_end, "inits do not overlap"); 3296 last_init_end = next_init_off; // it's a non-tile 3297 } 3298 #endif //ASSERT 3299 } 3300 3301 remove_extra_zeroes(); // clear out all the zmems left over 3302 add_req(inits); 3303 3304 if (!ZeroTLAB) { 3305 // If anything remains to be zeroed, zero it all now. 3306 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3307 // if it is the last unused 4 bytes of an instance, forget about it 3308 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3309 if (zeroes_done + BytesPerLong >= size_limit) { 3310 assert(allocation() != NULL, ""); 3311 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3312 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3313 if (zeroes_done == k->layout_helper()) 3314 zeroes_done = size_limit; 3315 } 3316 if (zeroes_done < size_limit) { 3317 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3318 zeroes_done, size_in_bytes, phase); 3319 } 3320 } 3321 3322 set_complete(phase); 3323 return rawmem; 3324} 3325 3326 3327#ifdef ASSERT 3328bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3329 if (is_complete()) 3330 return true; // stores could be anything at this point 3331 assert(allocation() != NULL, "must be present"); 3332 intptr_t last_off = allocation()->minimum_header_size(); 3333 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3334 Node* st = in(i); 3335 intptr_t st_off = get_store_offset(st, phase); 3336 if (st_off < 0) continue; // ignore dead garbage 3337 if (last_off > st_off) { 3338 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3339 this->dump(2); 3340 assert(false, "ascending store offsets"); 3341 return false; 3342 } 3343 last_off = st_off + st->as_Store()->memory_size(); 3344 } 3345 return true; 3346} 3347#endif //ASSERT 3348 3349 3350 3351 3352//============================MergeMemNode===================================== 3353// 3354// SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3355// contributing store or call operations. Each contributor provides the memory 3356// state for a particular "alias type" (see Compile::alias_type). For example, 3357// if a MergeMem has an input X for alias category #6, then any memory reference 3358// to alias category #6 may use X as its memory state input, as an exact equivalent 3359// to using the MergeMem as a whole. 3360// Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3361// 3362// (Here, the <N> notation gives the index of the relevant adr_type.) 3363// 3364// In one special case (and more cases in the future), alias categories overlap. 3365// The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3366// states. Therefore, if a MergeMem has only one contributing input W for Bot, 3367// it is exactly equivalent to that state W: 3368// MergeMem(<Bot>: W) <==> W 3369// 3370// Usually, the merge has more than one input. In that case, where inputs 3371// overlap (i.e., one is Bot), the narrower alias type determines the memory 3372// state for that type, and the wider alias type (Bot) fills in everywhere else: 3373// Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3374// Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3375// 3376// A merge can take a "wide" memory state as one of its narrow inputs. 3377// This simply means that the merge observes out only the relevant parts of 3378// the wide input. That is, wide memory states arriving at narrow merge inputs 3379// are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3380// 3381// These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3382// and that memory slices "leak through": 3383// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3384// 3385// But, in such a cascade, repeated memory slices can "block the leak": 3386// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3387// 3388// In the last example, Y is not part of the combined memory state of the 3389// outermost MergeMem. The system must, of course, prevent unschedulable 3390// memory states from arising, so you can be sure that the state Y is somehow 3391// a precursor to state Y'. 3392// 3393// 3394// REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3395// of each MergeMemNode array are exactly the numerical alias indexes, including 3396// but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3397// Compile::alias_type (and kin) produce and manage these indexes. 3398// 3399// By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3400// (Note that this provides quick access to the top node inside MergeMem methods, 3401// without the need to reach out via TLS to Compile::current.) 3402// 3403// As a consequence of what was just described, a MergeMem that represents a full 3404// memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3405// containing all alias categories. 3406// 3407// MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3408// 3409// All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3410// a memory state for the alias type <N>, or else the top node, meaning that 3411// there is no particular input for that alias type. Note that the length of 3412// a MergeMem is variable, and may be extended at any time to accommodate new 3413// memory states at larger alias indexes. When merges grow, they are of course 3414// filled with "top" in the unused in() positions. 3415// 3416// This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3417// (Top was chosen because it works smoothly with passes like GCM.) 3418// 3419// For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3420// the type of random VM bits like TLS references.) Since it is always the 3421// first non-Bot memory slice, some low-level loops use it to initialize an 3422// index variable: for (i = AliasIdxRaw; i < req(); i++). 3423// 3424// 3425// ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3426// the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3427// the memory state for alias type <N>, or (if there is no particular slice at <N>, 3428// it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3429// or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3430// MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3431// 3432// %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3433// really that different from the other memory inputs. An abbreviation called 3434// "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3435// 3436// 3437// PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3438// partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3439// that "emerges though" the base memory will be marked as excluding the alias types 3440// of the other (narrow-memory) copies which "emerged through" the narrow edges: 3441// 3442// Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3443// ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3444// 3445// This strange "subtraction" effect is necessary to ensure IGVN convergence. 3446// (It is currently unimplemented.) As you can see, the resulting merge is 3447// actually a disjoint union of memory states, rather than an overlay. 3448// 3449 3450//------------------------------MergeMemNode----------------------------------- 3451Node* MergeMemNode::make_empty_memory() { 3452 Node* empty_memory = (Node*) Compile::current()->top(); 3453 assert(empty_memory->is_top(), "correct sentinel identity"); 3454 return empty_memory; 3455} 3456 3457MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3458 init_class_id(Class_MergeMem); 3459 // all inputs are nullified in Node::Node(int) 3460 // set_input(0, NULL); // no control input 3461 3462 // Initialize the edges uniformly to top, for starters. 3463 Node* empty_mem = make_empty_memory(); 3464 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3465 init_req(i,empty_mem); 3466 } 3467 assert(empty_memory() == empty_mem, ""); 3468 3469 if( new_base != NULL && new_base->is_MergeMem() ) { 3470 MergeMemNode* mdef = new_base->as_MergeMem(); 3471 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3472 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3473 mms.set_memory(mms.memory2()); 3474 } 3475 assert(base_memory() == mdef->base_memory(), ""); 3476 } else { 3477 set_base_memory(new_base); 3478 } 3479} 3480 3481// Make a new, untransformed MergeMem with the same base as 'mem'. 3482// If mem is itself a MergeMem, populate the result with the same edges. 3483MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3484 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3485} 3486 3487//------------------------------cmp-------------------------------------------- 3488uint MergeMemNode::hash() const { return NO_HASH; } 3489uint MergeMemNode::cmp( const Node &n ) const { 3490 return (&n == this); // Always fail except on self 3491} 3492 3493//------------------------------Identity--------------------------------------- 3494Node* MergeMemNode::Identity(PhaseTransform *phase) { 3495 // Identity if this merge point does not record any interesting memory 3496 // disambiguations. 3497 Node* base_mem = base_memory(); 3498 Node* empty_mem = empty_memory(); 3499 if (base_mem != empty_mem) { // Memory path is not dead? 3500 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3501 Node* mem = in(i); 3502 if (mem != empty_mem && mem != base_mem) { 3503 return this; // Many memory splits; no change 3504 } 3505 } 3506 } 3507 return base_mem; // No memory splits; ID on the one true input 3508} 3509 3510//------------------------------Ideal------------------------------------------ 3511// This method is invoked recursively on chains of MergeMem nodes 3512Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3513 // Remove chain'd MergeMems 3514 // 3515 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3516 // relative to the "in(Bot)". Since we are patching both at the same time, 3517 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3518 // but rewrite each "in(i)" relative to the new "in(Bot)". 3519 Node *progress = NULL; 3520 3521 3522 Node* old_base = base_memory(); 3523 Node* empty_mem = empty_memory(); 3524 if (old_base == empty_mem) 3525 return NULL; // Dead memory path. 3526 3527 MergeMemNode* old_mbase; 3528 if (old_base != NULL && old_base->is_MergeMem()) 3529 old_mbase = old_base->as_MergeMem(); 3530 else 3531 old_mbase = NULL; 3532 Node* new_base = old_base; 3533 3534 // simplify stacked MergeMems in base memory 3535 if (old_mbase) new_base = old_mbase->base_memory(); 3536 3537 // the base memory might contribute new slices beyond my req() 3538 if (old_mbase) grow_to_match(old_mbase); 3539 3540 // Look carefully at the base node if it is a phi. 3541 PhiNode* phi_base; 3542 if (new_base != NULL && new_base->is_Phi()) 3543 phi_base = new_base->as_Phi(); 3544 else 3545 phi_base = NULL; 3546 3547 Node* phi_reg = NULL; 3548 uint phi_len = (uint)-1; 3549 if (phi_base != NULL && !phi_base->is_copy()) { 3550 // do not examine phi if degraded to a copy 3551 phi_reg = phi_base->region(); 3552 phi_len = phi_base->req(); 3553 // see if the phi is unfinished 3554 for (uint i = 1; i < phi_len; i++) { 3555 if (phi_base->in(i) == NULL) { 3556 // incomplete phi; do not look at it yet! 3557 phi_reg = NULL; 3558 phi_len = (uint)-1; 3559 break; 3560 } 3561 } 3562 } 3563 3564 // Note: We do not call verify_sparse on entry, because inputs 3565 // can normalize to the base_memory via subsume_node or similar 3566 // mechanisms. This method repairs that damage. 3567 3568 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3569 3570 // Look at each slice. 3571 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3572 Node* old_in = in(i); 3573 // calculate the old memory value 3574 Node* old_mem = old_in; 3575 if (old_mem == empty_mem) old_mem = old_base; 3576 assert(old_mem == memory_at(i), ""); 3577 3578 // maybe update (reslice) the old memory value 3579 3580 // simplify stacked MergeMems 3581 Node* new_mem = old_mem; 3582 MergeMemNode* old_mmem; 3583 if (old_mem != NULL && old_mem->is_MergeMem()) 3584 old_mmem = old_mem->as_MergeMem(); 3585 else 3586 old_mmem = NULL; 3587 if (old_mmem == this) { 3588 // This can happen if loops break up and safepoints disappear. 3589 // A merge of BotPtr (default) with a RawPtr memory derived from a 3590 // safepoint can be rewritten to a merge of the same BotPtr with 3591 // the BotPtr phi coming into the loop. If that phi disappears 3592 // also, we can end up with a self-loop of the mergemem. 3593 // In general, if loops degenerate and memory effects disappear, 3594 // a mergemem can be left looking at itself. This simply means 3595 // that the mergemem's default should be used, since there is 3596 // no longer any apparent effect on this slice. 3597 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3598 // from start. Update the input to TOP. 3599 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3600 } 3601 else if (old_mmem != NULL) { 3602 new_mem = old_mmem->memory_at(i); 3603 } 3604 // else preceeding memory was not a MergeMem 3605 3606 // replace equivalent phis (unfortunately, they do not GVN together) 3607 if (new_mem != NULL && new_mem != new_base && 3608 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3609 if (new_mem->is_Phi()) { 3610 PhiNode* phi_mem = new_mem->as_Phi(); 3611 for (uint i = 1; i < phi_len; i++) { 3612 if (phi_base->in(i) != phi_mem->in(i)) { 3613 phi_mem = NULL; 3614 break; 3615 } 3616 } 3617 if (phi_mem != NULL) { 3618 // equivalent phi nodes; revert to the def 3619 new_mem = new_base; 3620 } 3621 } 3622 } 3623 3624 // maybe store down a new value 3625 Node* new_in = new_mem; 3626 if (new_in == new_base) new_in = empty_mem; 3627 3628 if (new_in != old_in) { 3629 // Warning: Do not combine this "if" with the previous "if" 3630 // A memory slice might have be be rewritten even if it is semantically 3631 // unchanged, if the base_memory value has changed. 3632 set_req(i, new_in); 3633 progress = this; // Report progress 3634 } 3635 } 3636 3637 if (new_base != old_base) { 3638 set_req(Compile::AliasIdxBot, new_base); 3639 // Don't use set_base_memory(new_base), because we need to update du. 3640 assert(base_memory() == new_base, ""); 3641 progress = this; 3642 } 3643 3644 if( base_memory() == this ) { 3645 // a self cycle indicates this memory path is dead 3646 set_req(Compile::AliasIdxBot, empty_mem); 3647 } 3648 3649 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3650 // Recursion must occur after the self cycle check above 3651 if( base_memory()->is_MergeMem() ) { 3652 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3653 Node *m = phase->transform(new_mbase); // Rollup any cycles 3654 if( m != NULL && (m->is_top() || 3655 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3656 // propagate rollup of dead cycle to self 3657 set_req(Compile::AliasIdxBot, empty_mem); 3658 } 3659 } 3660 3661 if( base_memory() == empty_mem ) { 3662 progress = this; 3663 // Cut inputs during Parse phase only. 3664 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3665 if( !can_reshape ) { 3666 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3667 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3668 } 3669 } 3670 } 3671 3672 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3673 // Check if PhiNode::Ideal's "Split phis through memory merges" 3674 // transform should be attempted. Look for this->phi->this cycle. 3675 uint merge_width = req(); 3676 if (merge_width > Compile::AliasIdxRaw) { 3677 PhiNode* phi = base_memory()->as_Phi(); 3678 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3679 if (phi->in(i) == this) { 3680 phase->is_IterGVN()->_worklist.push(phi); 3681 break; 3682 } 3683 } 3684 } 3685 } 3686 3687 assert(progress || verify_sparse(), "please, no dups of base"); 3688 return progress; 3689} 3690 3691//-------------------------set_base_memory------------------------------------- 3692void MergeMemNode::set_base_memory(Node *new_base) { 3693 Node* empty_mem = empty_memory(); 3694 set_req(Compile::AliasIdxBot, new_base); 3695 assert(memory_at(req()) == new_base, "must set default memory"); 3696 // Clear out other occurrences of new_base: 3697 if (new_base != empty_mem) { 3698 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3699 if (in(i) == new_base) set_req(i, empty_mem); 3700 } 3701 } 3702} 3703 3704//------------------------------out_RegMask------------------------------------ 3705const RegMask &MergeMemNode::out_RegMask() const { 3706 return RegMask::Empty; 3707} 3708 3709//------------------------------dump_spec-------------------------------------- 3710#ifndef PRODUCT 3711void MergeMemNode::dump_spec(outputStream *st) const { 3712 st->print(" {"); 3713 Node* base_mem = base_memory(); 3714 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3715 Node* mem = memory_at(i); 3716 if (mem == base_mem) { st->print(" -"); continue; } 3717 st->print( " N%d:", mem->_idx ); 3718 Compile::current()->get_adr_type(i)->dump_on(st); 3719 } 3720 st->print(" }"); 3721} 3722#endif // !PRODUCT 3723 3724 3725#ifdef ASSERT 3726static bool might_be_same(Node* a, Node* b) { 3727 if (a == b) return true; 3728 if (!(a->is_Phi() || b->is_Phi())) return false; 3729 // phis shift around during optimization 3730 return true; // pretty stupid... 3731} 3732 3733// verify a narrow slice (either incoming or outgoing) 3734static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 3735 if (!VerifyAliases) return; // don't bother to verify unless requested 3736 if (is_error_reported()) return; // muzzle asserts when debugging an error 3737 if (Node::in_dump()) return; // muzzle asserts when printing 3738 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 3739 assert(n != NULL, ""); 3740 // Elide intervening MergeMem's 3741 while (n->is_MergeMem()) { 3742 n = n->as_MergeMem()->memory_at(alias_idx); 3743 } 3744 Compile* C = Compile::current(); 3745 const TypePtr* n_adr_type = n->adr_type(); 3746 if (n == m->empty_memory()) { 3747 // Implicit copy of base_memory() 3748 } else if (n_adr_type != TypePtr::BOTTOM) { 3749 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 3750 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 3751 } else { 3752 // A few places like make_runtime_call "know" that VM calls are narrow, 3753 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 3754 bool expected_wide_mem = false; 3755 if (n == m->base_memory()) { 3756 expected_wide_mem = true; 3757 } else if (alias_idx == Compile::AliasIdxRaw || 3758 n == m->memory_at(Compile::AliasIdxRaw)) { 3759 expected_wide_mem = true; 3760 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 3761 // memory can "leak through" calls on channels that 3762 // are write-once. Allow this also. 3763 expected_wide_mem = true; 3764 } 3765 assert(expected_wide_mem, "expected narrow slice replacement"); 3766 } 3767} 3768#else // !ASSERT 3769#define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 3770#endif 3771 3772 3773//-----------------------------memory_at--------------------------------------- 3774Node* MergeMemNode::memory_at(uint alias_idx) const { 3775 assert(alias_idx >= Compile::AliasIdxRaw || 3776 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 3777 "must avoid base_memory and AliasIdxTop"); 3778 3779 // Otherwise, it is a narrow slice. 3780 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 3781 Compile *C = Compile::current(); 3782 if (is_empty_memory(n)) { 3783 // the array is sparse; empty slots are the "top" node 3784 n = base_memory(); 3785 assert(Node::in_dump() 3786 || n == NULL || n->bottom_type() == Type::TOP 3787 || n->adr_type() == TypePtr::BOTTOM 3788 || n->adr_type() == TypeRawPtr::BOTTOM 3789 || Compile::current()->AliasLevel() == 0, 3790 "must be a wide memory"); 3791 // AliasLevel == 0 if we are organizing the memory states manually. 3792 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 3793 } else { 3794 // make sure the stored slice is sane 3795 #ifdef ASSERT 3796 if (is_error_reported() || Node::in_dump()) { 3797 } else if (might_be_same(n, base_memory())) { 3798 // Give it a pass: It is a mostly harmless repetition of the base. 3799 // This can arise normally from node subsumption during optimization. 3800 } else { 3801 verify_memory_slice(this, alias_idx, n); 3802 } 3803 #endif 3804 } 3805 return n; 3806} 3807 3808//---------------------------set_memory_at------------------------------------- 3809void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 3810 verify_memory_slice(this, alias_idx, n); 3811 Node* empty_mem = empty_memory(); 3812 if (n == base_memory()) n = empty_mem; // collapse default 3813 uint need_req = alias_idx+1; 3814 if (req() < need_req) { 3815 if (n == empty_mem) return; // already the default, so do not grow me 3816 // grow the sparse array 3817 do { 3818 add_req(empty_mem); 3819 } while (req() < need_req); 3820 } 3821 set_req( alias_idx, n ); 3822} 3823 3824 3825 3826//--------------------------iteration_setup------------------------------------ 3827void MergeMemNode::iteration_setup(const MergeMemNode* other) { 3828 if (other != NULL) { 3829 grow_to_match(other); 3830 // invariant: the finite support of mm2 is within mm->req() 3831 #ifdef ASSERT 3832 for (uint i = req(); i < other->req(); i++) { 3833 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 3834 } 3835 #endif 3836 } 3837 // Replace spurious copies of base_memory by top. 3838 Node* base_mem = base_memory(); 3839 if (base_mem != NULL && !base_mem->is_top()) { 3840 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 3841 if (in(i) == base_mem) 3842 set_req(i, empty_memory()); 3843 } 3844 } 3845} 3846 3847//---------------------------grow_to_match------------------------------------- 3848void MergeMemNode::grow_to_match(const MergeMemNode* other) { 3849 Node* empty_mem = empty_memory(); 3850 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 3851 // look for the finite support of the other memory 3852 for (uint i = other->req(); --i >= req(); ) { 3853 if (other->in(i) != empty_mem) { 3854 uint new_len = i+1; 3855 while (req() < new_len) add_req(empty_mem); 3856 break; 3857 } 3858 } 3859} 3860 3861//---------------------------verify_sparse------------------------------------- 3862#ifndef PRODUCT 3863bool MergeMemNode::verify_sparse() const { 3864 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 3865 Node* base_mem = base_memory(); 3866 // The following can happen in degenerate cases, since empty==top. 3867 if (is_empty_memory(base_mem)) return true; 3868 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3869 assert(in(i) != NULL, "sane slice"); 3870 if (in(i) == base_mem) return false; // should have been the sentinel value! 3871 } 3872 return true; 3873} 3874 3875bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 3876 Node* n; 3877 n = mm->in(idx); 3878 if (mem == n) return true; // might be empty_memory() 3879 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 3880 if (mem == n) return true; 3881 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 3882 if (mem == n) return true; 3883 if (n == NULL) break; 3884 } 3885 return false; 3886} 3887#endif // !PRODUCT 3888