memnode.cpp revision 579:0fbdb4381b99
1/* 2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25// Portions of code courtesy of Clifford Click 26 27// Optimization - Graph Style 28 29#include "incls/_precompiled.incl" 30#include "incls/_memnode.cpp.incl" 31 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_known_instance_field()) 95 return mchain; // don't try to optimize non-instance types 96 uint instance_id = tinst->instance_id(); 97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 98 Node *prev = NULL; 99 Node *result = mchain; 100 while (prev != result) { 101 prev = result; 102 if (result == start_mem) 103 break; // hit one of our sentinals 104 // skip over a call which does not affect this memory slice 105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 106 Node *proj_in = result->in(0); 107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 108 break; // hit one of our sentinals 109 } else if (proj_in->is_Call()) { 110 CallNode *call = proj_in->as_Call(); 111 if (!call->may_modify(t_adr, phase)) { 112 result = call->in(TypeFunc::Memory); 113 } 114 } else if (proj_in->is_Initialize()) { 115 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 116 // Stop if this is the initialization for the object instance which 117 // which contains this memory slice, otherwise skip over it. 118 if (alloc != NULL && alloc->_idx != instance_id) { 119 result = proj_in->in(TypeFunc::Memory); 120 } 121 } else if (proj_in->is_MemBar()) { 122 result = proj_in->in(TypeFunc::Memory); 123 } else { 124 assert(false, "unexpected projection"); 125 } 126 } else if (result->is_MergeMem()) { 127 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 128 } 129 } 130 return result; 131} 132 133Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 134 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 135 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); 136 PhaseIterGVN *igvn = phase->is_IterGVN(); 137 Node *result = mchain; 138 result = optimize_simple_memory_chain(result, t_adr, phase); 139 if (is_instance && igvn != NULL && result->is_Phi()) { 140 PhiNode *mphi = result->as_Phi(); 141 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 142 const TypePtr *t = mphi->adr_type(); 143 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 144 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 145 t->is_oopptr()->cast_to_exactness(true) 146 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 147 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 148 // clone the Phi with our address type 149 result = mphi->split_out_instance(t_adr, igvn); 150 } else { 151 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 152 } 153 } 154 return result; 155} 156 157static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 158 uint alias_idx = phase->C->get_alias_index(tp); 159 Node *mem = mmem; 160#ifdef ASSERT 161 { 162 // Check that current type is consistent with the alias index used during graph construction 163 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 164 bool consistent = adr_check == NULL || adr_check->empty() || 165 phase->C->must_alias(adr_check, alias_idx ); 166 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 167 if( !consistent && adr_check != NULL && !adr_check->empty() && 168 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 169 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 170 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 171 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 172 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 173 // don't assert if it is dead code. 174 consistent = true; 175 } 176 if( !consistent ) { 177 st->print("alias_idx==%d, adr_check==", alias_idx); 178 if( adr_check == NULL ) { 179 st->print("NULL"); 180 } else { 181 adr_check->dump(); 182 } 183 st->cr(); 184 print_alias_types(); 185 assert(consistent, "adr_check must match alias idx"); 186 } 187 } 188#endif 189 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 190 // means an array I have not precisely typed yet. Do not do any 191 // alias stuff with it any time soon. 192 const TypeOopPtr *tinst = tp->isa_oopptr(); 193 if( tp->base() != Type::AnyPtr && 194 !(tinst && 195 tinst->klass()->is_java_lang_Object() && 196 tinst->offset() == Type::OffsetBot) ) { 197 // compress paths and change unreachable cycles to TOP 198 // If not, we can update the input infinitely along a MergeMem cycle 199 // Equivalent code in PhiNode::Ideal 200 Node* m = phase->transform(mmem); 201 // If tranformed to a MergeMem, get the desired slice 202 // Otherwise the returned node represents memory for every slice 203 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 204 // Update input if it is progress over what we have now 205 } 206 return mem; 207} 208 209//--------------------------Ideal_common--------------------------------------- 210// Look for degenerate control and memory inputs. Bypass MergeMem inputs. 211// Unhook non-raw memories from complete (macro-expanded) initializations. 212Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 213 // If our control input is a dead region, kill all below the region 214 Node *ctl = in(MemNode::Control); 215 if (ctl && remove_dead_region(phase, can_reshape)) 216 return this; 217 ctl = in(MemNode::Control); 218 // Don't bother trying to transform a dead node 219 if( ctl && ctl->is_top() ) return NodeSentinel; 220 221 // Ignore if memory is dead, or self-loop 222 Node *mem = in(MemNode::Memory); 223 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 224 assert( mem != this, "dead loop in MemNode::Ideal" ); 225 226 Node *address = in(MemNode::Address); 227 const Type *t_adr = phase->type( address ); 228 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 229 230 PhaseIterGVN *igvn = phase->is_IterGVN(); 231 if( can_reshape && igvn != NULL && igvn->_worklist.member(address) ) { 232 // The address's base and type may change when the address is processed. 233 // Delay this mem node transformation until the address is processed. 234 phase->is_IterGVN()->_worklist.push(this); 235 return NodeSentinel; // caller will return NULL 236 } 237 238 // Avoid independent memory operations 239 Node* old_mem = mem; 240 241 // The code which unhooks non-raw memories from complete (macro-expanded) 242 // initializations was removed. After macro-expansion all stores catched 243 // by Initialize node became raw stores and there is no information 244 // which memory slices they modify. So it is unsafe to move any memory 245 // operation above these stores. Also in most cases hooked non-raw memories 246 // were already unhooked by using information from detect_ptr_independence() 247 // and find_previous_store(). 248 249 if (mem->is_MergeMem()) { 250 MergeMemNode* mmem = mem->as_MergeMem(); 251 const TypePtr *tp = t_adr->is_ptr(); 252 253 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 254 } 255 256 if (mem != old_mem) { 257 set_req(MemNode::Memory, mem); 258 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 259 return this; 260 } 261 262 // let the subclass continue analyzing... 263 return NULL; 264} 265 266// Helper function for proving some simple control dominations. 267// Attempt to prove that all control inputs of 'dom' dominate 'sub'. 268// Already assumes that 'dom' is available at 'sub', and that 'sub' 269// is not a constant (dominated by the method's StartNode). 270// Used by MemNode::find_previous_store to prove that the 271// control input of a memory operation predates (dominates) 272// an allocation it wants to look past. 273bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 274 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 275 return false; // Conservative answer for dead code 276 277 // Check 'dom'. Skip Proj and CatchProj nodes. 278 dom = dom->find_exact_control(dom); 279 if (dom == NULL || dom->is_top()) 280 return false; // Conservative answer for dead code 281 282 if (dom == sub) { 283 // For the case when, for example, 'sub' is Initialize and the original 284 // 'dom' is Proj node of the 'sub'. 285 return false; 286 } 287 288 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 289 return true; 290 291 // 'dom' dominates 'sub' if its control edge and control edges 292 // of all its inputs dominate or equal to sub's control edge. 293 294 // Currently 'sub' is either Allocate, Initialize or Start nodes. 295 // Or Region for the check in LoadNode::Ideal(); 296 // 'sub' should have sub->in(0) != NULL. 297 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 298 sub->is_Region(), "expecting only these nodes"); 299 300 // Get control edge of 'sub'. 301 Node* orig_sub = sub; 302 sub = sub->find_exact_control(sub->in(0)); 303 if (sub == NULL || sub->is_top()) 304 return false; // Conservative answer for dead code 305 306 assert(sub->is_CFG(), "expecting control"); 307 308 if (sub == dom) 309 return true; 310 311 if (sub->is_Start() || sub->is_Root()) 312 return false; 313 314 { 315 // Check all control edges of 'dom'. 316 317 ResourceMark rm; 318 Arena* arena = Thread::current()->resource_area(); 319 Node_List nlist(arena); 320 Unique_Node_List dom_list(arena); 321 322 dom_list.push(dom); 323 bool only_dominating_controls = false; 324 325 for (uint next = 0; next < dom_list.size(); next++) { 326 Node* n = dom_list.at(next); 327 if (n == orig_sub) 328 return false; // One of dom's inputs dominated by sub. 329 if (!n->is_CFG() && n->pinned()) { 330 // Check only own control edge for pinned non-control nodes. 331 n = n->find_exact_control(n->in(0)); 332 if (n == NULL || n->is_top()) 333 return false; // Conservative answer for dead code 334 assert(n->is_CFG(), "expecting control"); 335 dom_list.push(n); 336 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 337 only_dominating_controls = true; 338 } else if (n->is_CFG()) { 339 if (n->dominates(sub, nlist)) 340 only_dominating_controls = true; 341 else 342 return false; 343 } else { 344 // First, own control edge. 345 Node* m = n->find_exact_control(n->in(0)); 346 if (m != NULL) { 347 if (m->is_top()) 348 return false; // Conservative answer for dead code 349 dom_list.push(m); 350 } 351 // Now, the rest of edges. 352 uint cnt = n->req(); 353 for (uint i = 1; i < cnt; i++) { 354 m = n->find_exact_control(n->in(i)); 355 if (m == NULL || m->is_top()) 356 continue; 357 dom_list.push(m); 358 } 359 } 360 } 361 return only_dominating_controls; 362 } 363} 364 365//---------------------detect_ptr_independence--------------------------------- 366// Used by MemNode::find_previous_store to prove that two base 367// pointers are never equal. 368// The pointers are accompanied by their associated allocations, 369// if any, which have been previously discovered by the caller. 370bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 371 Node* p2, AllocateNode* a2, 372 PhaseTransform* phase) { 373 // Attempt to prove that these two pointers cannot be aliased. 374 // They may both manifestly be allocations, and they should differ. 375 // Or, if they are not both allocations, they can be distinct constants. 376 // Otherwise, one is an allocation and the other a pre-existing value. 377 if (a1 == NULL && a2 == NULL) { // neither an allocation 378 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 379 } else if (a1 != NULL && a2 != NULL) { // both allocations 380 return (a1 != a2); 381 } else if (a1 != NULL) { // one allocation a1 382 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 383 return all_controls_dominate(p2, a1); 384 } else { //(a2 != NULL) // one allocation a2 385 return all_controls_dominate(p1, a2); 386 } 387 return false; 388} 389 390 391// The logic for reordering loads and stores uses four steps: 392// (a) Walk carefully past stores and initializations which we 393// can prove are independent of this load. 394// (b) Observe that the next memory state makes an exact match 395// with self (load or store), and locate the relevant store. 396// (c) Ensure that, if we were to wire self directly to the store, 397// the optimizer would fold it up somehow. 398// (d) Do the rewiring, and return, depending on some other part of 399// the optimizer to fold up the load. 400// This routine handles steps (a) and (b). Steps (c) and (d) are 401// specific to loads and stores, so they are handled by the callers. 402// (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 403// 404Node* MemNode::find_previous_store(PhaseTransform* phase) { 405 Node* ctrl = in(MemNode::Control); 406 Node* adr = in(MemNode::Address); 407 intptr_t offset = 0; 408 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 409 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 410 411 if (offset == Type::OffsetBot) 412 return NULL; // cannot unalias unless there are precise offsets 413 414 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 415 416 intptr_t size_in_bytes = memory_size(); 417 418 Node* mem = in(MemNode::Memory); // start searching here... 419 420 int cnt = 50; // Cycle limiter 421 for (;;) { // While we can dance past unrelated stores... 422 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 423 424 if (mem->is_Store()) { 425 Node* st_adr = mem->in(MemNode::Address); 426 intptr_t st_offset = 0; 427 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 428 if (st_base == NULL) 429 break; // inscrutable pointer 430 if (st_offset != offset && st_offset != Type::OffsetBot) { 431 const int MAX_STORE = BytesPerLong; 432 if (st_offset >= offset + size_in_bytes || 433 st_offset <= offset - MAX_STORE || 434 st_offset <= offset - mem->as_Store()->memory_size()) { 435 // Success: The offsets are provably independent. 436 // (You may ask, why not just test st_offset != offset and be done? 437 // The answer is that stores of different sizes can co-exist 438 // in the same sequence of RawMem effects. We sometimes initialize 439 // a whole 'tile' of array elements with a single jint or jlong.) 440 mem = mem->in(MemNode::Memory); 441 continue; // (a) advance through independent store memory 442 } 443 } 444 if (st_base != base && 445 detect_ptr_independence(base, alloc, 446 st_base, 447 AllocateNode::Ideal_allocation(st_base, phase), 448 phase)) { 449 // Success: The bases are provably independent. 450 mem = mem->in(MemNode::Memory); 451 continue; // (a) advance through independent store memory 452 } 453 454 // (b) At this point, if the bases or offsets do not agree, we lose, 455 // since we have not managed to prove 'this' and 'mem' independent. 456 if (st_base == base && st_offset == offset) { 457 return mem; // let caller handle steps (c), (d) 458 } 459 460 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 461 InitializeNode* st_init = mem->in(0)->as_Initialize(); 462 AllocateNode* st_alloc = st_init->allocation(); 463 if (st_alloc == NULL) 464 break; // something degenerated 465 bool known_identical = false; 466 bool known_independent = false; 467 if (alloc == st_alloc) 468 known_identical = true; 469 else if (alloc != NULL) 470 known_independent = true; 471 else if (all_controls_dominate(this, st_alloc)) 472 known_independent = true; 473 474 if (known_independent) { 475 // The bases are provably independent: Either they are 476 // manifestly distinct allocations, or else the control 477 // of this load dominates the store's allocation. 478 int alias_idx = phase->C->get_alias_index(adr_type()); 479 if (alias_idx == Compile::AliasIdxRaw) { 480 mem = st_alloc->in(TypeFunc::Memory); 481 } else { 482 mem = st_init->memory(alias_idx); 483 } 484 continue; // (a) advance through independent store memory 485 } 486 487 // (b) at this point, if we are not looking at a store initializing 488 // the same allocation we are loading from, we lose. 489 if (known_identical) { 490 // From caller, can_see_stored_value will consult find_captured_store. 491 return mem; // let caller handle steps (c), (d) 492 } 493 494 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 495 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 496 if (mem->is_Proj() && mem->in(0)->is_Call()) { 497 CallNode *call = mem->in(0)->as_Call(); 498 if (!call->may_modify(addr_t, phase)) { 499 mem = call->in(TypeFunc::Memory); 500 continue; // (a) advance through independent call memory 501 } 502 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 503 mem = mem->in(0)->in(TypeFunc::Memory); 504 continue; // (a) advance through independent MemBar memory 505 } else if (mem->is_MergeMem()) { 506 int alias_idx = phase->C->get_alias_index(adr_type()); 507 mem = mem->as_MergeMem()->memory_at(alias_idx); 508 continue; // (a) advance through independent MergeMem memory 509 } 510 } 511 512 // Unless there is an explicit 'continue', we must bail out here, 513 // because 'mem' is an inscrutable memory state (e.g., a call). 514 break; 515 } 516 517 return NULL; // bail out 518} 519 520//----------------------calculate_adr_type------------------------------------- 521// Helper function. Notices when the given type of address hits top or bottom. 522// Also, asserts a cross-check of the type against the expected address type. 523const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 524 if (t == Type::TOP) return NULL; // does not touch memory any more? 525 #ifdef PRODUCT 526 cross_check = NULL; 527 #else 528 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 529 #endif 530 const TypePtr* tp = t->isa_ptr(); 531 if (tp == NULL) { 532 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 533 return TypePtr::BOTTOM; // touches lots of memory 534 } else { 535 #ifdef ASSERT 536 // %%%% [phh] We don't check the alias index if cross_check is 537 // TypeRawPtr::BOTTOM. Needs to be investigated. 538 if (cross_check != NULL && 539 cross_check != TypePtr::BOTTOM && 540 cross_check != TypeRawPtr::BOTTOM) { 541 // Recheck the alias index, to see if it has changed (due to a bug). 542 Compile* C = Compile::current(); 543 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 544 "must stay in the original alias category"); 545 // The type of the address must be contained in the adr_type, 546 // disregarding "null"-ness. 547 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 548 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 549 assert(cross_check->meet(tp_notnull) == cross_check, 550 "real address must not escape from expected memory type"); 551 } 552 #endif 553 return tp; 554 } 555} 556 557//------------------------adr_phi_is_loop_invariant---------------------------- 558// A helper function for Ideal_DU_postCCP to check if a Phi in a counted 559// loop is loop invariant. Make a quick traversal of Phi and associated 560// CastPP nodes, looking to see if they are a closed group within the loop. 561bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 562 // The idea is that the phi-nest must boil down to only CastPP nodes 563 // with the same data. This implies that any path into the loop already 564 // includes such a CastPP, and so the original cast, whatever its input, 565 // must be covered by an equivalent cast, with an earlier control input. 566 ResourceMark rm; 567 568 // The loop entry input of the phi should be the unique dominating 569 // node for every Phi/CastPP in the loop. 570 Unique_Node_List closure; 571 closure.push(adr_phi->in(LoopNode::EntryControl)); 572 573 // Add the phi node and the cast to the worklist. 574 Unique_Node_List worklist; 575 worklist.push(adr_phi); 576 if( cast != NULL ){ 577 if( !cast->is_ConstraintCast() ) return false; 578 worklist.push(cast); 579 } 580 581 // Begin recursive walk of phi nodes. 582 while( worklist.size() ){ 583 // Take a node off the worklist 584 Node *n = worklist.pop(); 585 if( !closure.member(n) ){ 586 // Add it to the closure. 587 closure.push(n); 588 // Make a sanity check to ensure we don't waste too much time here. 589 if( closure.size() > 20) return false; 590 // This node is OK if: 591 // - it is a cast of an identical value 592 // - or it is a phi node (then we add its inputs to the worklist) 593 // Otherwise, the node is not OK, and we presume the cast is not invariant 594 if( n->is_ConstraintCast() ){ 595 worklist.push(n->in(1)); 596 } else if( n->is_Phi() ) { 597 for( uint i = 1; i < n->req(); i++ ) { 598 worklist.push(n->in(i)); 599 } 600 } else { 601 return false; 602 } 603 } 604 } 605 606 // Quit when the worklist is empty, and we've found no offending nodes. 607 return true; 608} 609 610//------------------------------Ideal_DU_postCCP------------------------------- 611// Find any cast-away of null-ness and keep its control. Null cast-aways are 612// going away in this pass and we need to make this memory op depend on the 613// gating null check. 614Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 615 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 616} 617 618// I tried to leave the CastPP's in. This makes the graph more accurate in 619// some sense; we get to keep around the knowledge that an oop is not-null 620// after some test. Alas, the CastPP's interfere with GVN (some values are 621// the regular oop, some are the CastPP of the oop, all merge at Phi's which 622// cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 623// some of the more trivial cases in the optimizer. Removing more useless 624// Phi's started allowing Loads to illegally float above null checks. I gave 625// up on this approach. CNC 10/20/2000 626// This static method may be called not from MemNode (EncodePNode calls it). 627// Only the control edge of the node 'n' might be updated. 628Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 629 Node *skipped_cast = NULL; 630 // Need a null check? Regular static accesses do not because they are 631 // from constant addresses. Array ops are gated by the range check (which 632 // always includes a NULL check). Just check field ops. 633 if( n->in(MemNode::Control) == NULL ) { 634 // Scan upwards for the highest location we can place this memory op. 635 while( true ) { 636 switch( adr->Opcode() ) { 637 638 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 639 adr = adr->in(AddPNode::Base); 640 continue; 641 642 case Op_DecodeN: // No change to NULL-ness, so peek thru 643 adr = adr->in(1); 644 continue; 645 646 case Op_CastPP: 647 // If the CastPP is useless, just peek on through it. 648 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 649 // Remember the cast that we've peeked though. If we peek 650 // through more than one, then we end up remembering the highest 651 // one, that is, if in a loop, the one closest to the top. 652 skipped_cast = adr; 653 adr = adr->in(1); 654 continue; 655 } 656 // CastPP is going away in this pass! We need this memory op to be 657 // control-dependent on the test that is guarding the CastPP. 658 ccp->hash_delete(n); 659 n->set_req(MemNode::Control, adr->in(0)); 660 ccp->hash_insert(n); 661 return n; 662 663 case Op_Phi: 664 // Attempt to float above a Phi to some dominating point. 665 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 666 // If we've already peeked through a Cast (which could have set the 667 // control), we can't float above a Phi, because the skipped Cast 668 // may not be loop invariant. 669 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 670 adr = adr->in(1); 671 continue; 672 } 673 } 674 675 // Intentional fallthrough! 676 677 // No obvious dominating point. The mem op is pinned below the Phi 678 // by the Phi itself. If the Phi goes away (no true value is merged) 679 // then the mem op can float, but not indefinitely. It must be pinned 680 // behind the controls leading to the Phi. 681 case Op_CheckCastPP: 682 // These usually stick around to change address type, however a 683 // useless one can be elided and we still need to pick up a control edge 684 if (adr->in(0) == NULL) { 685 // This CheckCastPP node has NO control and is likely useless. But we 686 // need check further up the ancestor chain for a control input to keep 687 // the node in place. 4959717. 688 skipped_cast = adr; 689 adr = adr->in(1); 690 continue; 691 } 692 ccp->hash_delete(n); 693 n->set_req(MemNode::Control, adr->in(0)); 694 ccp->hash_insert(n); 695 return n; 696 697 // List of "safe" opcodes; those that implicitly block the memory 698 // op below any null check. 699 case Op_CastX2P: // no null checks on native pointers 700 case Op_Parm: // 'this' pointer is not null 701 case Op_LoadP: // Loading from within a klass 702 case Op_LoadN: // Loading from within a klass 703 case Op_LoadKlass: // Loading from within a klass 704 case Op_LoadNKlass: // Loading from within a klass 705 case Op_ConP: // Loading from a klass 706 case Op_ConN: // Loading from a klass 707 case Op_CreateEx: // Sucking up the guts of an exception oop 708 case Op_Con: // Reading from TLS 709 case Op_CMoveP: // CMoveP is pinned 710 case Op_CMoveN: // CMoveN is pinned 711 break; // No progress 712 713 case Op_Proj: // Direct call to an allocation routine 714 case Op_SCMemProj: // Memory state from store conditional ops 715#ifdef ASSERT 716 { 717 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 718 const Node* call = adr->in(0); 719 if (call->is_CallJava()) { 720 const CallJavaNode* call_java = call->as_CallJava(); 721 const TypeTuple *r = call_java->tf()->range(); 722 assert(r->cnt() > TypeFunc::Parms, "must return value"); 723 const Type* ret_type = r->field_at(TypeFunc::Parms); 724 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 725 // We further presume that this is one of 726 // new_instance_Java, new_array_Java, or 727 // the like, but do not assert for this. 728 } else if (call->is_Allocate()) { 729 // similar case to new_instance_Java, etc. 730 } else if (!call->is_CallLeaf()) { 731 // Projections from fetch_oop (OSR) are allowed as well. 732 ShouldNotReachHere(); 733 } 734 } 735#endif 736 break; 737 default: 738 ShouldNotReachHere(); 739 } 740 break; 741 } 742 } 743 744 return NULL; // No progress 745} 746 747 748//============================================================================= 749uint LoadNode::size_of() const { return sizeof(*this); } 750uint LoadNode::cmp( const Node &n ) const 751{ return !Type::cmp( _type, ((LoadNode&)n)._type ); } 752const Type *LoadNode::bottom_type() const { return _type; } 753uint LoadNode::ideal_reg() const { 754 return Matcher::base2reg[_type->base()]; 755} 756 757#ifndef PRODUCT 758void LoadNode::dump_spec(outputStream *st) const { 759 MemNode::dump_spec(st); 760 if( !Verbose && !WizardMode ) { 761 // standard dump does this in Verbose and WizardMode 762 st->print(" #"); _type->dump_on(st); 763 } 764} 765#endif 766 767 768//----------------------------LoadNode::make----------------------------------- 769// Polymorphic factory method: 770Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 771 Compile* C = gvn.C; 772 773 // sanity check the alias category against the created node type 774 assert(!(adr_type->isa_oopptr() && 775 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 776 "use LoadKlassNode instead"); 777 assert(!(adr_type->isa_aryptr() && 778 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 779 "use LoadRangeNode instead"); 780 switch (bt) { 781 case T_BOOLEAN: 782 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 783 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 784 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 785 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 786 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 787 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 788 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 789 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 790 case T_OBJECT: 791#ifdef _LP64 792 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 793 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 794 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 795 } else 796#endif 797 { 798 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 799 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 800 } 801 } 802 ShouldNotReachHere(); 803 return (LoadNode*)NULL; 804} 805 806LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 807 bool require_atomic = true; 808 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 809} 810 811 812 813 814//------------------------------hash------------------------------------------- 815uint LoadNode::hash() const { 816 // unroll addition of interesting fields 817 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 818} 819 820//---------------------------can_see_stored_value------------------------------ 821// This routine exists to make sure this set of tests is done the same 822// everywhere. We need to make a coordinated change: first LoadNode::Ideal 823// will change the graph shape in a way which makes memory alive twice at the 824// same time (uses the Oracle model of aliasing), then some 825// LoadXNode::Identity will fold things back to the equivalence-class model 826// of aliasing. 827Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 828 Node* ld_adr = in(MemNode::Address); 829 830 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 831 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 832 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 833 atp->field() != NULL && !atp->field()->is_volatile()) { 834 uint alias_idx = atp->index(); 835 bool final = atp->field()->is_final(); 836 Node* result = NULL; 837 Node* current = st; 838 // Skip through chains of MemBarNodes checking the MergeMems for 839 // new states for the slice of this load. Stop once any other 840 // kind of node is encountered. Loads from final memory can skip 841 // through any kind of MemBar but normal loads shouldn't skip 842 // through MemBarAcquire since the could allow them to move out of 843 // a synchronized region. 844 while (current->is_Proj()) { 845 int opc = current->in(0)->Opcode(); 846 if ((final && opc == Op_MemBarAcquire) || 847 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 848 Node* mem = current->in(0)->in(TypeFunc::Memory); 849 if (mem->is_MergeMem()) { 850 MergeMemNode* merge = mem->as_MergeMem(); 851 Node* new_st = merge->memory_at(alias_idx); 852 if (new_st == merge->base_memory()) { 853 // Keep searching 854 current = merge->base_memory(); 855 continue; 856 } 857 // Save the new memory state for the slice and fall through 858 // to exit. 859 result = new_st; 860 } 861 } 862 break; 863 } 864 if (result != NULL) { 865 st = result; 866 } 867 } 868 869 870 // Loop around twice in the case Load -> Initialize -> Store. 871 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 872 for (int trip = 0; trip <= 1; trip++) { 873 874 if (st->is_Store()) { 875 Node* st_adr = st->in(MemNode::Address); 876 if (!phase->eqv(st_adr, ld_adr)) { 877 // Try harder before giving up... Match raw and non-raw pointers. 878 intptr_t st_off = 0; 879 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 880 if (alloc == NULL) return NULL; 881 intptr_t ld_off = 0; 882 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 883 if (alloc != allo2) return NULL; 884 if (ld_off != st_off) return NULL; 885 // At this point we have proven something like this setup: 886 // A = Allocate(...) 887 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 888 // S = StoreQ(, AddP(, A.Parm , #Off), V) 889 // (Actually, we haven't yet proven the Q's are the same.) 890 // In other words, we are loading from a casted version of 891 // the same pointer-and-offset that we stored to. 892 // Thus, we are able to replace L by V. 893 } 894 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 895 if (store_Opcode() != st->Opcode()) 896 return NULL; 897 return st->in(MemNode::ValueIn); 898 } 899 900 intptr_t offset = 0; // scratch 901 902 // A load from a freshly-created object always returns zero. 903 // (This can happen after LoadNode::Ideal resets the load's memory input 904 // to find_captured_store, which returned InitializeNode::zero_memory.) 905 if (st->is_Proj() && st->in(0)->is_Allocate() && 906 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 907 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 908 // return a zero value for the load's basic type 909 // (This is one of the few places where a generic PhaseTransform 910 // can create new nodes. Think of it as lazily manifesting 911 // virtually pre-existing constants.) 912 return phase->zerocon(memory_type()); 913 } 914 915 // A load from an initialization barrier can match a captured store. 916 if (st->is_Proj() && st->in(0)->is_Initialize()) { 917 InitializeNode* init = st->in(0)->as_Initialize(); 918 AllocateNode* alloc = init->allocation(); 919 if (alloc != NULL && 920 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 921 // examine a captured store value 922 st = init->find_captured_store(offset, memory_size(), phase); 923 if (st != NULL) 924 continue; // take one more trip around 925 } 926 } 927 928 break; 929 } 930 931 return NULL; 932} 933 934//----------------------is_instance_field_load_with_local_phi------------------ 935bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 936 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 937 in(MemNode::Address)->is_AddP() ) { 938 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 939 // Only instances. 940 if( t_oop != NULL && t_oop->is_known_instance_field() && 941 t_oop->offset() != Type::OffsetBot && 942 t_oop->offset() != Type::OffsetTop) { 943 return true; 944 } 945 } 946 return false; 947} 948 949//------------------------------Identity--------------------------------------- 950// Loads are identity if previous store is to same address 951Node *LoadNode::Identity( PhaseTransform *phase ) { 952 // If the previous store-maker is the right kind of Store, and the store is 953 // to the same address, then we are equal to the value stored. 954 Node* mem = in(MemNode::Memory); 955 Node* value = can_see_stored_value(mem, phase); 956 if( value ) { 957 // byte, short & char stores truncate naturally. 958 // A load has to load the truncated value which requires 959 // some sort of masking operation and that requires an 960 // Ideal call instead of an Identity call. 961 if (memory_size() < BytesPerInt) { 962 // If the input to the store does not fit with the load's result type, 963 // it must be truncated via an Ideal call. 964 if (!phase->type(value)->higher_equal(phase->type(this))) 965 return this; 966 } 967 // (This works even when value is a Con, but LoadNode::Value 968 // usually runs first, producing the singleton type of the Con.) 969 return value; 970 } 971 972 // Search for an existing data phi which was generated before for the same 973 // instance's field to avoid infinite genertion of phis in a loop. 974 Node *region = mem->in(0); 975 if (is_instance_field_load_with_local_phi(region)) { 976 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 977 int this_index = phase->C->get_alias_index(addr_t); 978 int this_offset = addr_t->offset(); 979 int this_id = addr_t->is_oopptr()->instance_id(); 980 const Type* this_type = bottom_type(); 981 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 982 Node* phi = region->fast_out(i); 983 if (phi->is_Phi() && phi != mem && 984 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 985 return phi; 986 } 987 } 988 } 989 990 return this; 991} 992 993 994// Returns true if the AliasType refers to the field that holds the 995// cached box array. Currently only handles the IntegerCache case. 996static bool is_autobox_cache(Compile::AliasType* atp) { 997 if (atp != NULL && atp->field() != NULL) { 998 ciField* field = atp->field(); 999 ciSymbol* klass = field->holder()->name(); 1000 if (field->name() == ciSymbol::cache_field_name() && 1001 field->holder()->uses_default_loader() && 1002 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1003 return true; 1004 } 1005 } 1006 return false; 1007} 1008 1009// Fetch the base value in the autobox array 1010static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1011 if (atp != NULL && atp->field() != NULL) { 1012 ciField* field = atp->field(); 1013 ciSymbol* klass = field->holder()->name(); 1014 if (field->name() == ciSymbol::cache_field_name() && 1015 field->holder()->uses_default_loader() && 1016 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1017 assert(field->is_constant(), "what?"); 1018 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1019 // Fetch the box object at the base of the array and get its value 1020 ciInstance* box = array->obj_at(0)->as_instance(); 1021 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1022 if (ik->nof_nonstatic_fields() == 1) { 1023 // This should be true nonstatic_field_at requires calling 1024 // nof_nonstatic_fields so check it anyway 1025 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1026 cache_offset = c.as_int(); 1027 } 1028 return true; 1029 } 1030 } 1031 return false; 1032} 1033 1034// Returns true if the AliasType refers to the value field of an 1035// autobox object. Currently only handles Integer. 1036static bool is_autobox_object(Compile::AliasType* atp) { 1037 if (atp != NULL && atp->field() != NULL) { 1038 ciField* field = atp->field(); 1039 ciSymbol* klass = field->holder()->name(); 1040 if (field->name() == ciSymbol::value_name() && 1041 field->holder()->uses_default_loader() && 1042 klass == ciSymbol::java_lang_Integer()) { 1043 return true; 1044 } 1045 } 1046 return false; 1047} 1048 1049 1050// We're loading from an object which has autobox behaviour. 1051// If this object is result of a valueOf call we'll have a phi 1052// merging a newly allocated object and a load from the cache. 1053// We want to replace this load with the original incoming 1054// argument to the valueOf call. 1055Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1056 Node* base = in(Address)->in(AddPNode::Base); 1057 if (base->is_Phi() && base->req() == 3) { 1058 AllocateNode* allocation = NULL; 1059 int allocation_index = -1; 1060 int load_index = -1; 1061 for (uint i = 1; i < base->req(); i++) { 1062 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1063 if (allocation != NULL) { 1064 allocation_index = i; 1065 load_index = 3 - allocation_index; 1066 break; 1067 } 1068 } 1069 LoadNode* load = NULL; 1070 if (allocation != NULL && base->in(load_index)->is_Load()) { 1071 load = base->in(load_index)->as_Load(); 1072 } 1073 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1074 // Push the loads from the phi that comes from valueOf up 1075 // through it to allow elimination of the loads and the recovery 1076 // of the original value. 1077 Node* mem_phi = in(Memory); 1078 Node* offset = in(Address)->in(AddPNode::Offset); 1079 Node* region = base->in(0); 1080 1081 Node* in1 = clone(); 1082 Node* in1_addr = in1->in(Address)->clone(); 1083 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1084 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1085 in1_addr->set_req(AddPNode::Offset, offset); 1086 in1->set_req(0, region->in(allocation_index)); 1087 in1->set_req(Address, in1_addr); 1088 in1->set_req(Memory, mem_phi->in(allocation_index)); 1089 1090 Node* in2 = clone(); 1091 Node* in2_addr = in2->in(Address)->clone(); 1092 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1093 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1094 in2_addr->set_req(AddPNode::Offset, offset); 1095 in2->set_req(0, region->in(load_index)); 1096 in2->set_req(Address, in2_addr); 1097 in2->set_req(Memory, mem_phi->in(load_index)); 1098 1099 in1_addr = phase->transform(in1_addr); 1100 in1 = phase->transform(in1); 1101 in2_addr = phase->transform(in2_addr); 1102 in2 = phase->transform(in2); 1103 1104 PhiNode* result = PhiNode::make_blank(region, this); 1105 result->set_req(allocation_index, in1); 1106 result->set_req(load_index, in2); 1107 return result; 1108 } 1109 } else if (base->is_Load()) { 1110 // Eliminate the load of Integer.value for integers from the cache 1111 // array by deriving the value from the index into the array. 1112 // Capture the offset of the load and then reverse the computation. 1113 Node* load_base = base->in(Address)->in(AddPNode::Base); 1114 if (load_base != NULL) { 1115 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1116 intptr_t cache_offset; 1117 int shift = -1; 1118 Node* cache = NULL; 1119 if (is_autobox_cache(atp)) { 1120 shift = exact_log2(type2aelembytes(T_OBJECT)); 1121 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1122 } 1123 if (cache != NULL && base->in(Address)->is_AddP()) { 1124 Node* elements[4]; 1125 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1126 int cache_low; 1127 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1128 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1129 // Add up all the offsets making of the address of the load 1130 Node* result = elements[0]; 1131 for (int i = 1; i < count; i++) { 1132 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1133 } 1134 // Remove the constant offset from the address and then 1135 // remove the scaling of the offset to recover the original index. 1136 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1137 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1138 // Peel the shift off directly but wrap it in a dummy node 1139 // since Ideal can't return existing nodes 1140 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1141 } else { 1142 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1143 } 1144#ifdef _LP64 1145 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1146#endif 1147 return result; 1148 } 1149 } 1150 } 1151 } 1152 return NULL; 1153} 1154 1155//------------------------------split_through_phi------------------------------ 1156// Split instance field load through Phi. 1157Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1158 Node* mem = in(MemNode::Memory); 1159 Node* address = in(MemNode::Address); 1160 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1161 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1162 1163 assert(mem->is_Phi() && (t_oop != NULL) && 1164 t_oop->is_known_instance_field(), "invalide conditions"); 1165 1166 Node *region = mem->in(0); 1167 if (region == NULL) { 1168 return NULL; // Wait stable graph 1169 } 1170 uint cnt = mem->req(); 1171 for( uint i = 1; i < cnt; i++ ) { 1172 Node *in = mem->in(i); 1173 if( in == NULL ) { 1174 return NULL; // Wait stable graph 1175 } 1176 } 1177 // Check for loop invariant. 1178 if (cnt == 3) { 1179 for( uint i = 1; i < cnt; i++ ) { 1180 Node *in = mem->in(i); 1181 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1182 if (m == mem) { 1183 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1184 return this; 1185 } 1186 } 1187 } 1188 // Split through Phi (see original code in loopopts.cpp). 1189 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1190 1191 // Do nothing here if Identity will find a value 1192 // (to avoid infinite chain of value phis generation). 1193 if ( !phase->eqv(this, this->Identity(phase)) ) 1194 return NULL; 1195 1196 // Skip the split if the region dominates some control edge of the address. 1197 if (cnt == 3 && !MemNode::all_controls_dominate(address, region)) 1198 return NULL; 1199 1200 const Type* this_type = this->bottom_type(); 1201 int this_index = phase->C->get_alias_index(addr_t); 1202 int this_offset = addr_t->offset(); 1203 int this_iid = addr_t->is_oopptr()->instance_id(); 1204 int wins = 0; 1205 PhaseIterGVN *igvn = phase->is_IterGVN(); 1206 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1207 for( uint i = 1; i < region->req(); i++ ) { 1208 Node *x; 1209 Node* the_clone = NULL; 1210 if( region->in(i) == phase->C->top() ) { 1211 x = phase->C->top(); // Dead path? Use a dead data op 1212 } else { 1213 x = this->clone(); // Else clone up the data op 1214 the_clone = x; // Remember for possible deletion. 1215 // Alter data node to use pre-phi inputs 1216 if( this->in(0) == region ) { 1217 x->set_req( 0, region->in(i) ); 1218 } else { 1219 x->set_req( 0, NULL ); 1220 } 1221 for( uint j = 1; j < this->req(); j++ ) { 1222 Node *in = this->in(j); 1223 if( in->is_Phi() && in->in(0) == region ) 1224 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 1225 } 1226 } 1227 // Check for a 'win' on some paths 1228 const Type *t = x->Value(igvn); 1229 1230 bool singleton = t->singleton(); 1231 1232 // See comments in PhaseIdealLoop::split_thru_phi(). 1233 if( singleton && t == Type::TOP ) { 1234 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1235 } 1236 1237 if( singleton ) { 1238 wins++; 1239 x = igvn->makecon(t); 1240 } else { 1241 // We now call Identity to try to simplify the cloned node. 1242 // Note that some Identity methods call phase->type(this). 1243 // Make sure that the type array is big enough for 1244 // our new node, even though we may throw the node away. 1245 // (This tweaking with igvn only works because x is a new node.) 1246 igvn->set_type(x, t); 1247 // If x is a TypeNode, capture any more-precise type permanently into Node 1248 // othewise it will be not updated during igvn->transform since 1249 // igvn->type(x) is set to x->Value() already. 1250 x->raise_bottom_type(t); 1251 Node *y = x->Identity(igvn); 1252 if( y != x ) { 1253 wins++; 1254 x = y; 1255 } else { 1256 y = igvn->hash_find(x); 1257 if( y ) { 1258 wins++; 1259 x = y; 1260 } else { 1261 // Else x is a new node we are keeping 1262 // We do not need register_new_node_with_optimizer 1263 // because set_type has already been called. 1264 igvn->_worklist.push(x); 1265 } 1266 } 1267 } 1268 if (x != the_clone && the_clone != NULL) 1269 igvn->remove_dead_node(the_clone); 1270 phi->set_req(i, x); 1271 } 1272 if( wins > 0 ) { 1273 // Record Phi 1274 igvn->register_new_node_with_optimizer(phi); 1275 return phi; 1276 } 1277 igvn->remove_dead_node(phi); 1278 return NULL; 1279} 1280 1281//------------------------------Ideal------------------------------------------ 1282// If the load is from Field memory and the pointer is non-null, we can 1283// zero out the control input. 1284// If the offset is constant and the base is an object allocation, 1285// try to hook me up to the exact initializing store. 1286Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1287 Node* p = MemNode::Ideal_common(phase, can_reshape); 1288 if (p) return (p == NodeSentinel) ? NULL : p; 1289 1290 Node* ctrl = in(MemNode::Control); 1291 Node* address = in(MemNode::Address); 1292 1293 // Skip up past a SafePoint control. Cannot do this for Stores because 1294 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1295 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1296 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1297 ctrl = ctrl->in(0); 1298 set_req(MemNode::Control,ctrl); 1299 } 1300 1301 // Check for useless control edge in some common special cases 1302 if (in(MemNode::Control) != NULL) { 1303 intptr_t ignore = 0; 1304 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1305 if (base != NULL 1306 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1307 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw 1308 && all_controls_dominate(base, phase->C->start())) { 1309 // A method-invariant, non-null address (constant or 'this' argument). 1310 set_req(MemNode::Control, NULL); 1311 } 1312 } 1313 1314 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) { 1315 Node* base = in(Address)->in(AddPNode::Base); 1316 if (base != NULL) { 1317 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1318 if (is_autobox_object(atp)) { 1319 Node* result = eliminate_autobox(phase); 1320 if (result != NULL) return result; 1321 } 1322 } 1323 } 1324 1325 Node* mem = in(MemNode::Memory); 1326 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1327 1328 if (addr_t != NULL) { 1329 // try to optimize our memory input 1330 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1331 if (opt_mem != mem) { 1332 set_req(MemNode::Memory, opt_mem); 1333 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1334 return this; 1335 } 1336 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1337 if (can_reshape && opt_mem->is_Phi() && 1338 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1339 // Split instance field load through Phi. 1340 Node* result = split_through_phi(phase); 1341 if (result != NULL) return result; 1342 } 1343 } 1344 1345 // Check for prior store with a different base or offset; make Load 1346 // independent. Skip through any number of them. Bail out if the stores 1347 // are in an endless dead cycle and report no progress. This is a key 1348 // transform for Reflection. However, if after skipping through the Stores 1349 // we can't then fold up against a prior store do NOT do the transform as 1350 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1351 // array memory alive twice: once for the hoisted Load and again after the 1352 // bypassed Store. This situation only works if EVERYBODY who does 1353 // anti-dependence work knows how to bypass. I.e. we need all 1354 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1355 // the alias index stuff. So instead, peek through Stores and IFF we can 1356 // fold up, do so. 1357 Node* prev_mem = find_previous_store(phase); 1358 // Steps (a), (b): Walk past independent stores to find an exact match. 1359 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1360 // (c) See if we can fold up on the spot, but don't fold up here. 1361 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1362 // just return a prior value, which is done by Identity calls. 1363 if (can_see_stored_value(prev_mem, phase)) { 1364 // Make ready for step (d): 1365 set_req(MemNode::Memory, prev_mem); 1366 return this; 1367 } 1368 } 1369 1370 return NULL; // No further progress 1371} 1372 1373// Helper to recognize certain Klass fields which are invariant across 1374// some group of array types (e.g., int[] or all T[] where T < Object). 1375const Type* 1376LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1377 ciKlass* klass) const { 1378 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1379 // The field is Klass::_modifier_flags. Return its (constant) value. 1380 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1381 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1382 return TypeInt::make(klass->modifier_flags()); 1383 } 1384 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1385 // The field is Klass::_access_flags. Return its (constant) value. 1386 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1387 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1388 return TypeInt::make(klass->access_flags()); 1389 } 1390 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1391 // The field is Klass::_layout_helper. Return its constant value if known. 1392 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1393 return TypeInt::make(klass->layout_helper()); 1394 } 1395 1396 // No match. 1397 return NULL; 1398} 1399 1400//------------------------------Value----------------------------------------- 1401const Type *LoadNode::Value( PhaseTransform *phase ) const { 1402 // Either input is TOP ==> the result is TOP 1403 Node* mem = in(MemNode::Memory); 1404 const Type *t1 = phase->type(mem); 1405 if (t1 == Type::TOP) return Type::TOP; 1406 Node* adr = in(MemNode::Address); 1407 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1408 if (tp == NULL || tp->empty()) return Type::TOP; 1409 int off = tp->offset(); 1410 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1411 1412 // Try to guess loaded type from pointer type 1413 if (tp->base() == Type::AryPtr) { 1414 const Type *t = tp->is_aryptr()->elem(); 1415 // Don't do this for integer types. There is only potential profit if 1416 // the element type t is lower than _type; that is, for int types, if _type is 1417 // more restrictive than t. This only happens here if one is short and the other 1418 // char (both 16 bits), and in those cases we've made an intentional decision 1419 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1420 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1421 // 1422 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1423 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1424 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1425 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1426 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1427 // In fact, that could have been the original type of p1, and p1 could have 1428 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1429 // expression (LShiftL quux 3) independently optimized to the constant 8. 1430 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1431 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1432 // t might actually be lower than _type, if _type is a unique 1433 // concrete subclass of abstract class t. 1434 // Make sure the reference is not into the header, by comparing 1435 // the offset against the offset of the start of the array's data. 1436 // Different array types begin at slightly different offsets (12 vs. 16). 1437 // We choose T_BYTE as an example base type that is least restrictive 1438 // as to alignment, which will therefore produce the smallest 1439 // possible base offset. 1440 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1441 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1442 const Type* jt = t->join(_type); 1443 // In any case, do not allow the join, per se, to empty out the type. 1444 if (jt->empty() && !t->empty()) { 1445 // This can happen if a interface-typed array narrows to a class type. 1446 jt = _type; 1447 } 1448 1449 if (EliminateAutoBox) { 1450 // The pointers in the autobox arrays are always non-null 1451 Node* base = in(Address)->in(AddPNode::Base); 1452 if (base != NULL) { 1453 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1454 if (is_autobox_cache(atp)) { 1455 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1456 } 1457 } 1458 } 1459 return jt; 1460 } 1461 } 1462 } else if (tp->base() == Type::InstPtr) { 1463 assert( off != Type::OffsetBot || 1464 // arrays can be cast to Objects 1465 tp->is_oopptr()->klass()->is_java_lang_Object() || 1466 // unsafe field access may not have a constant offset 1467 phase->C->has_unsafe_access(), 1468 "Field accesses must be precise" ); 1469 // For oop loads, we expect the _type to be precise 1470 } else if (tp->base() == Type::KlassPtr) { 1471 assert( off != Type::OffsetBot || 1472 // arrays can be cast to Objects 1473 tp->is_klassptr()->klass()->is_java_lang_Object() || 1474 // also allow array-loading from the primary supertype 1475 // array during subtype checks 1476 Opcode() == Op_LoadKlass, 1477 "Field accesses must be precise" ); 1478 // For klass/static loads, we expect the _type to be precise 1479 } 1480 1481 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1482 if (tkls != NULL && !StressReflectiveCode) { 1483 ciKlass* klass = tkls->klass(); 1484 if (klass->is_loaded() && tkls->klass_is_exact()) { 1485 // We are loading a field from a Klass metaobject whose identity 1486 // is known at compile time (the type is "exact" or "precise"). 1487 // Check for fields we know are maintained as constants by the VM. 1488 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1489 // The field is Klass::_super_check_offset. Return its (constant) value. 1490 // (Folds up type checking code.) 1491 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1492 return TypeInt::make(klass->super_check_offset()); 1493 } 1494 // Compute index into primary_supers array 1495 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1496 // Check for overflowing; use unsigned compare to handle the negative case. 1497 if( depth < ciKlass::primary_super_limit() ) { 1498 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1499 // (Folds up type checking code.) 1500 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1501 ciKlass *ss = klass->super_of_depth(depth); 1502 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1503 } 1504 const Type* aift = load_array_final_field(tkls, klass); 1505 if (aift != NULL) return aift; 1506 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1507 && klass->is_array_klass()) { 1508 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1509 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1510 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1511 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1512 } 1513 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1514 // The field is Klass::_java_mirror. Return its (constant) value. 1515 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1516 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1517 return TypeInstPtr::make(klass->java_mirror()); 1518 } 1519 } 1520 1521 // We can still check if we are loading from the primary_supers array at a 1522 // shallow enough depth. Even though the klass is not exact, entries less 1523 // than or equal to its super depth are correct. 1524 if (klass->is_loaded() ) { 1525 ciType *inner = klass->klass(); 1526 while( inner->is_obj_array_klass() ) 1527 inner = inner->as_obj_array_klass()->base_element_type(); 1528 if( inner->is_instance_klass() && 1529 !inner->as_instance_klass()->flags().is_interface() ) { 1530 // Compute index into primary_supers array 1531 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1532 // Check for overflowing; use unsigned compare to handle the negative case. 1533 if( depth < ciKlass::primary_super_limit() && 1534 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1535 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1536 // (Folds up type checking code.) 1537 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1538 ciKlass *ss = klass->super_of_depth(depth); 1539 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1540 } 1541 } 1542 } 1543 1544 // If the type is enough to determine that the thing is not an array, 1545 // we can give the layout_helper a positive interval type. 1546 // This will help short-circuit some reflective code. 1547 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1548 && !klass->is_array_klass() // not directly typed as an array 1549 && !klass->is_interface() // specifically not Serializable & Cloneable 1550 && !klass->is_java_lang_Object() // not the supertype of all T[] 1551 ) { 1552 // Note: When interfaces are reliable, we can narrow the interface 1553 // test to (klass != Serializable && klass != Cloneable). 1554 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1555 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1556 // The key property of this type is that it folds up tests 1557 // for array-ness, since it proves that the layout_helper is positive. 1558 // Thus, a generic value like the basic object layout helper works fine. 1559 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1560 } 1561 } 1562 1563 // If we are loading from a freshly-allocated object, produce a zero, 1564 // if the load is provably beyond the header of the object. 1565 // (Also allow a variable load from a fresh array to produce zero.) 1566 if (ReduceFieldZeroing) { 1567 Node* value = can_see_stored_value(mem,phase); 1568 if (value != NULL && value->is_Con()) 1569 return value->bottom_type(); 1570 } 1571 1572 const TypeOopPtr *tinst = tp->isa_oopptr(); 1573 if (tinst != NULL && tinst->is_known_instance_field()) { 1574 // If we have an instance type and our memory input is the 1575 // programs's initial memory state, there is no matching store, 1576 // so just return a zero of the appropriate type 1577 Node *mem = in(MemNode::Memory); 1578 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1579 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1580 return Type::get_zero_type(_type->basic_type()); 1581 } 1582 } 1583 return _type; 1584} 1585 1586//------------------------------match_edge------------------------------------- 1587// Do we Match on this edge index or not? Match only the address. 1588uint LoadNode::match_edge(uint idx) const { 1589 return idx == MemNode::Address; 1590} 1591 1592//--------------------------LoadBNode::Ideal-------------------------------------- 1593// 1594// If the previous store is to the same address as this load, 1595// and the value stored was larger than a byte, replace this load 1596// with the value stored truncated to a byte. If no truncation is 1597// needed, the replacement is done in LoadNode::Identity(). 1598// 1599Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1600 Node* mem = in(MemNode::Memory); 1601 Node* value = can_see_stored_value(mem,phase); 1602 if( value && !phase->type(value)->higher_equal( _type ) ) { 1603 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1604 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1605 } 1606 // Identity call will handle the case where truncation is not needed. 1607 return LoadNode::Ideal(phase, can_reshape); 1608} 1609 1610//--------------------------LoadUSNode::Ideal------------------------------------- 1611// 1612// If the previous store is to the same address as this load, 1613// and the value stored was larger than a char, replace this load 1614// with the value stored truncated to a char. If no truncation is 1615// needed, the replacement is done in LoadNode::Identity(). 1616// 1617Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1618 Node* mem = in(MemNode::Memory); 1619 Node* value = can_see_stored_value(mem,phase); 1620 if( value && !phase->type(value)->higher_equal( _type ) ) 1621 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1622 // Identity call will handle the case where truncation is not needed. 1623 return LoadNode::Ideal(phase, can_reshape); 1624} 1625 1626//--------------------------LoadSNode::Ideal-------------------------------------- 1627// 1628// If the previous store is to the same address as this load, 1629// and the value stored was larger than a short, replace this load 1630// with the value stored truncated to a short. If no truncation is 1631// needed, the replacement is done in LoadNode::Identity(). 1632// 1633Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1634 Node* mem = in(MemNode::Memory); 1635 Node* value = can_see_stored_value(mem,phase); 1636 if( value && !phase->type(value)->higher_equal( _type ) ) { 1637 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1638 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1639 } 1640 // Identity call will handle the case where truncation is not needed. 1641 return LoadNode::Ideal(phase, can_reshape); 1642} 1643 1644//============================================================================= 1645//----------------------------LoadKlassNode::make------------------------------ 1646// Polymorphic factory method: 1647Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1648 Compile* C = gvn.C; 1649 Node *ctl = NULL; 1650 // sanity check the alias category against the created node type 1651 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1652 assert(adr_type != NULL, "expecting TypeOopPtr"); 1653#ifdef _LP64 1654 if (adr_type->is_ptr_to_narrowoop()) { 1655 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1656 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1657 } 1658#endif 1659 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1660 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1661} 1662 1663//------------------------------Value------------------------------------------ 1664const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1665 return klass_value_common(phase); 1666} 1667 1668const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1669 // Either input is TOP ==> the result is TOP 1670 const Type *t1 = phase->type( in(MemNode::Memory) ); 1671 if (t1 == Type::TOP) return Type::TOP; 1672 Node *adr = in(MemNode::Address); 1673 const Type *t2 = phase->type( adr ); 1674 if (t2 == Type::TOP) return Type::TOP; 1675 const TypePtr *tp = t2->is_ptr(); 1676 if (TypePtr::above_centerline(tp->ptr()) || 1677 tp->ptr() == TypePtr::Null) return Type::TOP; 1678 1679 // Return a more precise klass, if possible 1680 const TypeInstPtr *tinst = tp->isa_instptr(); 1681 if (tinst != NULL) { 1682 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1683 int offset = tinst->offset(); 1684 if (ik == phase->C->env()->Class_klass() 1685 && (offset == java_lang_Class::klass_offset_in_bytes() || 1686 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1687 // We are loading a special hidden field from a Class mirror object, 1688 // the field which points to the VM's Klass metaobject. 1689 ciType* t = tinst->java_mirror_type(); 1690 // java_mirror_type returns non-null for compile-time Class constants. 1691 if (t != NULL) { 1692 // constant oop => constant klass 1693 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1694 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1695 } 1696 if (!t->is_klass()) { 1697 // a primitive Class (e.g., int.class) has NULL for a klass field 1698 return TypePtr::NULL_PTR; 1699 } 1700 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1701 return TypeKlassPtr::make(t->as_klass()); 1702 } 1703 // non-constant mirror, so we can't tell what's going on 1704 } 1705 if( !ik->is_loaded() ) 1706 return _type; // Bail out if not loaded 1707 if (offset == oopDesc::klass_offset_in_bytes()) { 1708 if (tinst->klass_is_exact()) { 1709 return TypeKlassPtr::make(ik); 1710 } 1711 // See if we can become precise: no subklasses and no interface 1712 // (Note: We need to support verified interfaces.) 1713 if (!ik->is_interface() && !ik->has_subklass()) { 1714 //assert(!UseExactTypes, "this code should be useless with exact types"); 1715 // Add a dependence; if any subclass added we need to recompile 1716 if (!ik->is_final()) { 1717 // %%% should use stronger assert_unique_concrete_subtype instead 1718 phase->C->dependencies()->assert_leaf_type(ik); 1719 } 1720 // Return precise klass 1721 return TypeKlassPtr::make(ik); 1722 } 1723 1724 // Return root of possible klass 1725 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1726 } 1727 } 1728 1729 // Check for loading klass from an array 1730 const TypeAryPtr *tary = tp->isa_aryptr(); 1731 if( tary != NULL ) { 1732 ciKlass *tary_klass = tary->klass(); 1733 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1734 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1735 if (tary->klass_is_exact()) { 1736 return TypeKlassPtr::make(tary_klass); 1737 } 1738 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1739 // If the klass is an object array, we defer the question to the 1740 // array component klass. 1741 if( ak->is_obj_array_klass() ) { 1742 assert( ak->is_loaded(), "" ); 1743 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1744 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1745 ciInstanceKlass* ik = base_k->as_instance_klass(); 1746 // See if we can become precise: no subklasses and no interface 1747 if (!ik->is_interface() && !ik->has_subklass()) { 1748 //assert(!UseExactTypes, "this code should be useless with exact types"); 1749 // Add a dependence; if any subclass added we need to recompile 1750 if (!ik->is_final()) { 1751 phase->C->dependencies()->assert_leaf_type(ik); 1752 } 1753 // Return precise array klass 1754 return TypeKlassPtr::make(ak); 1755 } 1756 } 1757 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1758 } else { // Found a type-array? 1759 //assert(!UseExactTypes, "this code should be useless with exact types"); 1760 assert( ak->is_type_array_klass(), "" ); 1761 return TypeKlassPtr::make(ak); // These are always precise 1762 } 1763 } 1764 } 1765 1766 // Check for loading klass from an array klass 1767 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1768 if (tkls != NULL && !StressReflectiveCode) { 1769 ciKlass* klass = tkls->klass(); 1770 if( !klass->is_loaded() ) 1771 return _type; // Bail out if not loaded 1772 if( klass->is_obj_array_klass() && 1773 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1774 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1775 // // Always returning precise element type is incorrect, 1776 // // e.g., element type could be object and array may contain strings 1777 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1778 1779 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1780 // according to the element type's subclassing. 1781 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1782 } 1783 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1784 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1785 ciKlass* sup = klass->as_instance_klass()->super(); 1786 // The field is Klass::_super. Return its (constant) value. 1787 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1788 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1789 } 1790 } 1791 1792 // Bailout case 1793 return LoadNode::Value(phase); 1794} 1795 1796//------------------------------Identity--------------------------------------- 1797// To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1798// Also feed through the klass in Allocate(...klass...)._klass. 1799Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1800 return klass_identity_common(phase); 1801} 1802 1803Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 1804 Node* x = LoadNode::Identity(phase); 1805 if (x != this) return x; 1806 1807 // Take apart the address into an oop and and offset. 1808 // Return 'this' if we cannot. 1809 Node* adr = in(MemNode::Address); 1810 intptr_t offset = 0; 1811 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1812 if (base == NULL) return this; 1813 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1814 if (toop == NULL) return this; 1815 1816 // We can fetch the klass directly through an AllocateNode. 1817 // This works even if the klass is not constant (clone or newArray). 1818 if (offset == oopDesc::klass_offset_in_bytes()) { 1819 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1820 if (allocated_klass != NULL) { 1821 return allocated_klass; 1822 } 1823 } 1824 1825 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1826 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1827 // See inline_native_Class_query for occurrences of these patterns. 1828 // Java Example: x.getClass().isAssignableFrom(y) 1829 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1830 // 1831 // This improves reflective code, often making the Class 1832 // mirror go completely dead. (Current exception: Class 1833 // mirrors may appear in debug info, but we could clean them out by 1834 // introducing a new debug info operator for klassOop.java_mirror). 1835 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1836 && (offset == java_lang_Class::klass_offset_in_bytes() || 1837 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1838 // We are loading a special hidden field from a Class mirror, 1839 // the field which points to its Klass or arrayKlass metaobject. 1840 if (base->is_Load()) { 1841 Node* adr2 = base->in(MemNode::Address); 1842 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1843 if (tkls != NULL && !tkls->empty() 1844 && (tkls->klass()->is_instance_klass() || 1845 tkls->klass()->is_array_klass()) 1846 && adr2->is_AddP() 1847 ) { 1848 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1849 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1850 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1851 } 1852 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1853 return adr2->in(AddPNode::Base); 1854 } 1855 } 1856 } 1857 } 1858 1859 return this; 1860} 1861 1862 1863//------------------------------Value------------------------------------------ 1864const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 1865 const Type *t = klass_value_common(phase); 1866 if (t == Type::TOP) 1867 return t; 1868 1869 return t->make_narrowoop(); 1870} 1871 1872//------------------------------Identity--------------------------------------- 1873// To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 1874// Also feed through the klass in Allocate(...klass...)._klass. 1875Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 1876 Node *x = klass_identity_common(phase); 1877 1878 const Type *t = phase->type( x ); 1879 if( t == Type::TOP ) return x; 1880 if( t->isa_narrowoop()) return x; 1881 1882 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 1883} 1884 1885//------------------------------Value----------------------------------------- 1886const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1887 // Either input is TOP ==> the result is TOP 1888 const Type *t1 = phase->type( in(MemNode::Memory) ); 1889 if( t1 == Type::TOP ) return Type::TOP; 1890 Node *adr = in(MemNode::Address); 1891 const Type *t2 = phase->type( adr ); 1892 if( t2 == Type::TOP ) return Type::TOP; 1893 const TypePtr *tp = t2->is_ptr(); 1894 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 1895 const TypeAryPtr *tap = tp->isa_aryptr(); 1896 if( !tap ) return _type; 1897 return tap->size(); 1898} 1899 1900//-------------------------------Ideal--------------------------------------- 1901// Feed through the length in AllocateArray(...length...)._length. 1902Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1903 Node* p = MemNode::Ideal_common(phase, can_reshape); 1904 if (p) return (p == NodeSentinel) ? NULL : p; 1905 1906 // Take apart the address into an oop and and offset. 1907 // Return 'this' if we cannot. 1908 Node* adr = in(MemNode::Address); 1909 intptr_t offset = 0; 1910 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1911 if (base == NULL) return NULL; 1912 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1913 if (tary == NULL) return NULL; 1914 1915 // We can fetch the length directly through an AllocateArrayNode. 1916 // This works even if the length is not constant (clone or newArray). 1917 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1918 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 1919 if (alloc != NULL) { 1920 Node* allocated_length = alloc->Ideal_length(); 1921 Node* len = alloc->make_ideal_length(tary, phase); 1922 if (allocated_length != len) { 1923 // New CastII improves on this. 1924 return len; 1925 } 1926 } 1927 } 1928 1929 return NULL; 1930} 1931 1932//------------------------------Identity--------------------------------------- 1933// Feed through the length in AllocateArray(...length...)._length. 1934Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 1935 Node* x = LoadINode::Identity(phase); 1936 if (x != this) return x; 1937 1938 // Take apart the address into an oop and and offset. 1939 // Return 'this' if we cannot. 1940 Node* adr = in(MemNode::Address); 1941 intptr_t offset = 0; 1942 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1943 if (base == NULL) return this; 1944 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1945 if (tary == NULL) return this; 1946 1947 // We can fetch the length directly through an AllocateArrayNode. 1948 // This works even if the length is not constant (clone or newArray). 1949 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1950 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 1951 if (alloc != NULL) { 1952 Node* allocated_length = alloc->Ideal_length(); 1953 // Do not allow make_ideal_length to allocate a CastII node. 1954 Node* len = alloc->make_ideal_length(tary, phase, false); 1955 if (allocated_length == len) { 1956 // Return allocated_length only if it would not be improved by a CastII. 1957 return allocated_length; 1958 } 1959 } 1960 } 1961 1962 return this; 1963 1964} 1965 1966//============================================================================= 1967//---------------------------StoreNode::make----------------------------------- 1968// Polymorphic factory method: 1969StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 1970 Compile* C = gvn.C; 1971 1972 switch (bt) { 1973 case T_BOOLEAN: 1974 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 1975 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 1976 case T_CHAR: 1977 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 1978 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 1979 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 1980 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 1981 case T_ADDRESS: 1982 case T_OBJECT: 1983#ifdef _LP64 1984 if (adr->bottom_type()->is_ptr_to_narrowoop() || 1985 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 1986 adr->bottom_type()->isa_rawptr())) { 1987 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 1988 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 1989 } else 1990#endif 1991 { 1992 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 1993 } 1994 } 1995 ShouldNotReachHere(); 1996 return (StoreNode*)NULL; 1997} 1998 1999StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2000 bool require_atomic = true; 2001 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2002} 2003 2004 2005//--------------------------bottom_type---------------------------------------- 2006const Type *StoreNode::bottom_type() const { 2007 return Type::MEMORY; 2008} 2009 2010//------------------------------hash------------------------------------------- 2011uint StoreNode::hash() const { 2012 // unroll addition of interesting fields 2013 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2014 2015 // Since they are not commoned, do not hash them: 2016 return NO_HASH; 2017} 2018 2019//------------------------------Ideal------------------------------------------ 2020// Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2021// When a store immediately follows a relevant allocation/initialization, 2022// try to capture it into the initialization, or hoist it above. 2023Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2024 Node* p = MemNode::Ideal_common(phase, can_reshape); 2025 if (p) return (p == NodeSentinel) ? NULL : p; 2026 2027 Node* mem = in(MemNode::Memory); 2028 Node* address = in(MemNode::Address); 2029 2030 // Back-to-back stores to same address? Fold em up. 2031 // Generally unsafe if I have intervening uses... 2032 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 2033 // Looking at a dead closed cycle of memory? 2034 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2035 2036 assert(Opcode() == mem->Opcode() || 2037 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2038 "no mismatched stores, except on raw memory"); 2039 2040 if (mem->outcnt() == 1 && // check for intervening uses 2041 mem->as_Store()->memory_size() <= this->memory_size()) { 2042 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2043 // For example, 'mem' might be the final state at a conditional return. 2044 // Or, 'mem' might be used by some node which is live at the same time 2045 // 'this' is live, which might be unschedulable. So, require exactly 2046 // ONE user, the 'this' store, until such time as we clone 'mem' for 2047 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2048 if (can_reshape) { // (%%% is this an anachronism?) 2049 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2050 phase->is_IterGVN()); 2051 } else { 2052 // It's OK to do this in the parser, since DU info is always accurate, 2053 // and the parser always refers to nodes via SafePointNode maps. 2054 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2055 } 2056 return this; 2057 } 2058 } 2059 2060 // Capture an unaliased, unconditional, simple store into an initializer. 2061 // Or, if it is independent of the allocation, hoist it above the allocation. 2062 if (ReduceFieldZeroing && /*can_reshape &&*/ 2063 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2064 InitializeNode* init = mem->in(0)->as_Initialize(); 2065 intptr_t offset = init->can_capture_store(this, phase); 2066 if (offset > 0) { 2067 Node* moved = init->capture_store(this, offset, phase); 2068 // If the InitializeNode captured me, it made a raw copy of me, 2069 // and I need to disappear. 2070 if (moved != NULL) { 2071 // %%% hack to ensure that Ideal returns a new node: 2072 mem = MergeMemNode::make(phase->C, mem); 2073 return mem; // fold me away 2074 } 2075 } 2076 } 2077 2078 return NULL; // No further progress 2079} 2080 2081//------------------------------Value----------------------------------------- 2082const Type *StoreNode::Value( PhaseTransform *phase ) const { 2083 // Either input is TOP ==> the result is TOP 2084 const Type *t1 = phase->type( in(MemNode::Memory) ); 2085 if( t1 == Type::TOP ) return Type::TOP; 2086 const Type *t2 = phase->type( in(MemNode::Address) ); 2087 if( t2 == Type::TOP ) return Type::TOP; 2088 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2089 if( t3 == Type::TOP ) return Type::TOP; 2090 return Type::MEMORY; 2091} 2092 2093//------------------------------Identity--------------------------------------- 2094// Remove redundant stores: 2095// Store(m, p, Load(m, p)) changes to m. 2096// Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2097Node *StoreNode::Identity( PhaseTransform *phase ) { 2098 Node* mem = in(MemNode::Memory); 2099 Node* adr = in(MemNode::Address); 2100 Node* val = in(MemNode::ValueIn); 2101 2102 // Load then Store? Then the Store is useless 2103 if (val->is_Load() && 2104 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 2105 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 2106 val->as_Load()->store_Opcode() == Opcode()) { 2107 return mem; 2108 } 2109 2110 // Two stores in a row of the same value? 2111 if (mem->is_Store() && 2112 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 2113 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 2114 mem->Opcode() == Opcode()) { 2115 return mem; 2116 } 2117 2118 // Store of zero anywhere into a freshly-allocated object? 2119 // Then the store is useless. 2120 // (It must already have been captured by the InitializeNode.) 2121 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2122 // a newly allocated object is already all-zeroes everywhere 2123 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2124 return mem; 2125 } 2126 2127 // the store may also apply to zero-bits in an earlier object 2128 Node* prev_mem = find_previous_store(phase); 2129 // Steps (a), (b): Walk past independent stores to find an exact match. 2130 if (prev_mem != NULL) { 2131 Node* prev_val = can_see_stored_value(prev_mem, phase); 2132 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2133 // prev_val and val might differ by a cast; it would be good 2134 // to keep the more informative of the two. 2135 return mem; 2136 } 2137 } 2138 } 2139 2140 return this; 2141} 2142 2143//------------------------------match_edge------------------------------------- 2144// Do we Match on this edge index or not? Match only memory & value 2145uint StoreNode::match_edge(uint idx) const { 2146 return idx == MemNode::Address || idx == MemNode::ValueIn; 2147} 2148 2149//------------------------------cmp-------------------------------------------- 2150// Do not common stores up together. They generally have to be split 2151// back up anyways, so do not bother. 2152uint StoreNode::cmp( const Node &n ) const { 2153 return (&n == this); // Always fail except on self 2154} 2155 2156//------------------------------Ideal_masked_input----------------------------- 2157// Check for a useless mask before a partial-word store 2158// (StoreB ... (AndI valIn conIa) ) 2159// If (conIa & mask == mask) this simplifies to 2160// (StoreB ... (valIn) ) 2161Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2162 Node *val = in(MemNode::ValueIn); 2163 if( val->Opcode() == Op_AndI ) { 2164 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2165 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2166 set_req(MemNode::ValueIn, val->in(1)); 2167 return this; 2168 } 2169 } 2170 return NULL; 2171} 2172 2173 2174//------------------------------Ideal_sign_extended_input---------------------- 2175// Check for useless sign-extension before a partial-word store 2176// (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2177// If (conIL == conIR && conIR <= num_bits) this simplifies to 2178// (StoreB ... (valIn) ) 2179Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2180 Node *val = in(MemNode::ValueIn); 2181 if( val->Opcode() == Op_RShiftI ) { 2182 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2183 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2184 Node *shl = val->in(1); 2185 if( shl->Opcode() == Op_LShiftI ) { 2186 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2187 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2188 set_req(MemNode::ValueIn, shl->in(1)); 2189 return this; 2190 } 2191 } 2192 } 2193 } 2194 return NULL; 2195} 2196 2197//------------------------------value_never_loaded----------------------------------- 2198// Determine whether there are any possible loads of the value stored. 2199// For simplicity, we actually check if there are any loads from the 2200// address stored to, not just for loads of the value stored by this node. 2201// 2202bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2203 Node *adr = in(Address); 2204 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2205 if (adr_oop == NULL) 2206 return false; 2207 if (!adr_oop->is_known_instance_field()) 2208 return false; // if not a distinct instance, there may be aliases of the address 2209 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2210 Node *use = adr->fast_out(i); 2211 int opc = use->Opcode(); 2212 if (use->is_Load() || use->is_LoadStore()) { 2213 return false; 2214 } 2215 } 2216 return true; 2217} 2218 2219//============================================================================= 2220//------------------------------Ideal------------------------------------------ 2221// If the store is from an AND mask that leaves the low bits untouched, then 2222// we can skip the AND operation. If the store is from a sign-extension 2223// (a left shift, then right shift) we can skip both. 2224Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2225 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2226 if( progress != NULL ) return progress; 2227 2228 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2229 if( progress != NULL ) return progress; 2230 2231 // Finally check the default case 2232 return StoreNode::Ideal(phase, can_reshape); 2233} 2234 2235//============================================================================= 2236//------------------------------Ideal------------------------------------------ 2237// If the store is from an AND mask that leaves the low bits untouched, then 2238// we can skip the AND operation 2239Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2240 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2241 if( progress != NULL ) return progress; 2242 2243 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2244 if( progress != NULL ) return progress; 2245 2246 // Finally check the default case 2247 return StoreNode::Ideal(phase, can_reshape); 2248} 2249 2250//============================================================================= 2251//------------------------------Identity--------------------------------------- 2252Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2253 // No need to card mark when storing a null ptr 2254 Node* my_store = in(MemNode::OopStore); 2255 if (my_store->is_Store()) { 2256 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2257 if( t1 == TypePtr::NULL_PTR ) { 2258 return in(MemNode::Memory); 2259 } 2260 } 2261 return this; 2262} 2263 2264//------------------------------Value----------------------------------------- 2265const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2266 // Either input is TOP ==> the result is TOP 2267 const Type *t = phase->type( in(MemNode::Memory) ); 2268 if( t == Type::TOP ) return Type::TOP; 2269 t = phase->type( in(MemNode::Address) ); 2270 if( t == Type::TOP ) return Type::TOP; 2271 t = phase->type( in(MemNode::ValueIn) ); 2272 if( t == Type::TOP ) return Type::TOP; 2273 // If extra input is TOP ==> the result is TOP 2274 t = phase->type( in(MemNode::OopStore) ); 2275 if( t == Type::TOP ) return Type::TOP; 2276 2277 return StoreNode::Value( phase ); 2278} 2279 2280 2281//============================================================================= 2282//----------------------------------SCMemProjNode------------------------------ 2283const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2284{ 2285 return bottom_type(); 2286} 2287 2288//============================================================================= 2289LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2290 init_req(MemNode::Control, c ); 2291 init_req(MemNode::Memory , mem); 2292 init_req(MemNode::Address, adr); 2293 init_req(MemNode::ValueIn, val); 2294 init_req( ExpectedIn, ex ); 2295 init_class_id(Class_LoadStore); 2296 2297} 2298 2299//============================================================================= 2300//-------------------------------adr_type-------------------------------------- 2301// Do we Match on this edge index or not? Do not match memory 2302const TypePtr* ClearArrayNode::adr_type() const { 2303 Node *adr = in(3); 2304 return MemNode::calculate_adr_type(adr->bottom_type()); 2305} 2306 2307//------------------------------match_edge------------------------------------- 2308// Do we Match on this edge index or not? Do not match memory 2309uint ClearArrayNode::match_edge(uint idx) const { 2310 return idx > 1; 2311} 2312 2313//------------------------------Identity--------------------------------------- 2314// Clearing a zero length array does nothing 2315Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2316 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2317} 2318 2319//------------------------------Idealize--------------------------------------- 2320// Clearing a short array is faster with stores 2321Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2322 const int unit = BytesPerLong; 2323 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2324 if (!t) return NULL; 2325 if (!t->is_con()) return NULL; 2326 intptr_t raw_count = t->get_con(); 2327 intptr_t size = raw_count; 2328 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2329 // Clearing nothing uses the Identity call. 2330 // Negative clears are possible on dead ClearArrays 2331 // (see jck test stmt114.stmt11402.val). 2332 if (size <= 0 || size % unit != 0) return NULL; 2333 intptr_t count = size / unit; 2334 // Length too long; use fast hardware clear 2335 if (size > Matcher::init_array_short_size) return NULL; 2336 Node *mem = in(1); 2337 if( phase->type(mem)==Type::TOP ) return NULL; 2338 Node *adr = in(3); 2339 const Type* at = phase->type(adr); 2340 if( at==Type::TOP ) return NULL; 2341 const TypePtr* atp = at->isa_ptr(); 2342 // adjust atp to be the correct array element address type 2343 if (atp == NULL) atp = TypePtr::BOTTOM; 2344 else atp = atp->add_offset(Type::OffsetBot); 2345 // Get base for derived pointer purposes 2346 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2347 Node *base = adr->in(1); 2348 2349 Node *zero = phase->makecon(TypeLong::ZERO); 2350 Node *off = phase->MakeConX(BytesPerLong); 2351 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2352 count--; 2353 while( count-- ) { 2354 mem = phase->transform(mem); 2355 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2356 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2357 } 2358 return mem; 2359} 2360 2361//----------------------------clear_memory------------------------------------- 2362// Generate code to initialize object storage to zero. 2363Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2364 intptr_t start_offset, 2365 Node* end_offset, 2366 PhaseGVN* phase) { 2367 Compile* C = phase->C; 2368 intptr_t offset = start_offset; 2369 2370 int unit = BytesPerLong; 2371 if ((offset % unit) != 0) { 2372 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2373 adr = phase->transform(adr); 2374 const TypePtr* atp = TypeRawPtr::BOTTOM; 2375 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2376 mem = phase->transform(mem); 2377 offset += BytesPerInt; 2378 } 2379 assert((offset % unit) == 0, ""); 2380 2381 // Initialize the remaining stuff, if any, with a ClearArray. 2382 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2383} 2384 2385Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2386 Node* start_offset, 2387 Node* end_offset, 2388 PhaseGVN* phase) { 2389 if (start_offset == end_offset) { 2390 // nothing to do 2391 return mem; 2392 } 2393 2394 Compile* C = phase->C; 2395 int unit = BytesPerLong; 2396 Node* zbase = start_offset; 2397 Node* zend = end_offset; 2398 2399 // Scale to the unit required by the CPU: 2400 if (!Matcher::init_array_count_is_in_bytes) { 2401 Node* shift = phase->intcon(exact_log2(unit)); 2402 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2403 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2404 } 2405 2406 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2407 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2408 2409 // Bulk clear double-words 2410 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2411 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2412 return phase->transform(mem); 2413} 2414 2415Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2416 intptr_t start_offset, 2417 intptr_t end_offset, 2418 PhaseGVN* phase) { 2419 if (start_offset == end_offset) { 2420 // nothing to do 2421 return mem; 2422 } 2423 2424 Compile* C = phase->C; 2425 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2426 intptr_t done_offset = end_offset; 2427 if ((done_offset % BytesPerLong) != 0) { 2428 done_offset -= BytesPerInt; 2429 } 2430 if (done_offset > start_offset) { 2431 mem = clear_memory(ctl, mem, dest, 2432 start_offset, phase->MakeConX(done_offset), phase); 2433 } 2434 if (done_offset < end_offset) { // emit the final 32-bit store 2435 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2436 adr = phase->transform(adr); 2437 const TypePtr* atp = TypeRawPtr::BOTTOM; 2438 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2439 mem = phase->transform(mem); 2440 done_offset += BytesPerInt; 2441 } 2442 assert(done_offset == end_offset, ""); 2443 return mem; 2444} 2445 2446//============================================================================= 2447// Do we match on this edge? No memory edges 2448uint StrCompNode::match_edge(uint idx) const { 2449 return idx == 5 || idx == 6; 2450} 2451 2452//------------------------------Ideal------------------------------------------ 2453// Return a node which is more "ideal" than the current node. Strip out 2454// control copies 2455Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2456 return remove_dead_region(phase, can_reshape) ? this : NULL; 2457} 2458 2459//------------------------------Ideal------------------------------------------ 2460// Return a node which is more "ideal" than the current node. Strip out 2461// control copies 2462Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2463 return remove_dead_region(phase, can_reshape) ? this : NULL; 2464} 2465 2466 2467//============================================================================= 2468MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2469 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2470 _adr_type(C->get_adr_type(alias_idx)) 2471{ 2472 init_class_id(Class_MemBar); 2473 Node* top = C->top(); 2474 init_req(TypeFunc::I_O,top); 2475 init_req(TypeFunc::FramePtr,top); 2476 init_req(TypeFunc::ReturnAdr,top); 2477 if (precedent != NULL) 2478 init_req(TypeFunc::Parms, precedent); 2479} 2480 2481//------------------------------cmp-------------------------------------------- 2482uint MemBarNode::hash() const { return NO_HASH; } 2483uint MemBarNode::cmp( const Node &n ) const { 2484 return (&n == this); // Always fail except on self 2485} 2486 2487//------------------------------make------------------------------------------- 2488MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2489 int len = Precedent + (pn == NULL? 0: 1); 2490 switch (opcode) { 2491 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2492 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2493 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2494 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2495 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2496 default: ShouldNotReachHere(); return NULL; 2497 } 2498} 2499 2500//------------------------------Ideal------------------------------------------ 2501// Return a node which is more "ideal" than the current node. Strip out 2502// control copies 2503Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2504 return remove_dead_region(phase, can_reshape) ? this : NULL; 2505} 2506 2507//------------------------------Value------------------------------------------ 2508const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2509 if( !in(0) ) return Type::TOP; 2510 if( phase->type(in(0)) == Type::TOP ) 2511 return Type::TOP; 2512 return TypeTuple::MEMBAR; 2513} 2514 2515//------------------------------match------------------------------------------ 2516// Construct projections for memory. 2517Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2518 switch (proj->_con) { 2519 case TypeFunc::Control: 2520 case TypeFunc::Memory: 2521 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2522 } 2523 ShouldNotReachHere(); 2524 return NULL; 2525} 2526 2527//===========================InitializeNode==================================== 2528// SUMMARY: 2529// This node acts as a memory barrier on raw memory, after some raw stores. 2530// The 'cooked' oop value feeds from the Initialize, not the Allocation. 2531// The Initialize can 'capture' suitably constrained stores as raw inits. 2532// It can coalesce related raw stores into larger units (called 'tiles'). 2533// It can avoid zeroing new storage for memory units which have raw inits. 2534// At macro-expansion, it is marked 'complete', and does not optimize further. 2535// 2536// EXAMPLE: 2537// The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2538// ctl = incoming control; mem* = incoming memory 2539// (Note: A star * on a memory edge denotes I/O and other standard edges.) 2540// First allocate uninitialized memory and fill in the header: 2541// alloc = (Allocate ctl mem* 16 #short[].klass ...) 2542// ctl := alloc.Control; mem* := alloc.Memory* 2543// rawmem = alloc.Memory; rawoop = alloc.RawAddress 2544// Then initialize to zero the non-header parts of the raw memory block: 2545// init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2546// ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2547// After the initialize node executes, the object is ready for service: 2548// oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2549// Suppose its body is immediately initialized as {1,2}: 2550// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2551// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2552// mem.SLICE(#short[*]) := store2 2553// 2554// DETAILS: 2555// An InitializeNode collects and isolates object initialization after 2556// an AllocateNode and before the next possible safepoint. As a 2557// memory barrier (MemBarNode), it keeps critical stores from drifting 2558// down past any safepoint or any publication of the allocation. 2559// Before this barrier, a newly-allocated object may have uninitialized bits. 2560// After this barrier, it may be treated as a real oop, and GC is allowed. 2561// 2562// The semantics of the InitializeNode include an implicit zeroing of 2563// the new object from object header to the end of the object. 2564// (The object header and end are determined by the AllocateNode.) 2565// 2566// Certain stores may be added as direct inputs to the InitializeNode. 2567// These stores must update raw memory, and they must be to addresses 2568// derived from the raw address produced by AllocateNode, and with 2569// a constant offset. They must be ordered by increasing offset. 2570// The first one is at in(RawStores), the last at in(req()-1). 2571// Unlike most memory operations, they are not linked in a chain, 2572// but are displayed in parallel as users of the rawmem output of 2573// the allocation. 2574// 2575// (See comments in InitializeNode::capture_store, which continue 2576// the example given above.) 2577// 2578// When the associated Allocate is macro-expanded, the InitializeNode 2579// may be rewritten to optimize collected stores. A ClearArrayNode 2580// may also be created at that point to represent any required zeroing. 2581// The InitializeNode is then marked 'complete', prohibiting further 2582// capturing of nearby memory operations. 2583// 2584// During macro-expansion, all captured initializations which store 2585// constant values of 32 bits or smaller are coalesced (if advantagous) 2586// into larger 'tiles' 32 or 64 bits. This allows an object to be 2587// initialized in fewer memory operations. Memory words which are 2588// covered by neither tiles nor non-constant stores are pre-zeroed 2589// by explicit stores of zero. (The code shape happens to do all 2590// zeroing first, then all other stores, with both sequences occurring 2591// in order of ascending offsets.) 2592// 2593// Alternatively, code may be inserted between an AllocateNode and its 2594// InitializeNode, to perform arbitrary initialization of the new object. 2595// E.g., the object copying intrinsics insert complex data transfers here. 2596// The initialization must then be marked as 'complete' disable the 2597// built-in zeroing semantics and the collection of initializing stores. 2598// 2599// While an InitializeNode is incomplete, reads from the memory state 2600// produced by it are optimizable if they match the control edge and 2601// new oop address associated with the allocation/initialization. 2602// They return a stored value (if the offset matches) or else zero. 2603// A write to the memory state, if it matches control and address, 2604// and if it is to a constant offset, may be 'captured' by the 2605// InitializeNode. It is cloned as a raw memory operation and rewired 2606// inside the initialization, to the raw oop produced by the allocation. 2607// Operations on addresses which are provably distinct (e.g., to 2608// other AllocateNodes) are allowed to bypass the initialization. 2609// 2610// The effect of all this is to consolidate object initialization 2611// (both arrays and non-arrays, both piecewise and bulk) into a 2612// single location, where it can be optimized as a unit. 2613// 2614// Only stores with an offset less than TrackedInitializationLimit words 2615// will be considered for capture by an InitializeNode. This puts a 2616// reasonable limit on the complexity of optimized initializations. 2617 2618//---------------------------InitializeNode------------------------------------ 2619InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2620 : _is_complete(false), 2621 MemBarNode(C, adr_type, rawoop) 2622{ 2623 init_class_id(Class_Initialize); 2624 2625 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2626 assert(in(RawAddress) == rawoop, "proper init"); 2627 // Note: allocation() can be NULL, for secondary initialization barriers 2628} 2629 2630// Since this node is not matched, it will be processed by the 2631// register allocator. Declare that there are no constraints 2632// on the allocation of the RawAddress edge. 2633const RegMask &InitializeNode::in_RegMask(uint idx) const { 2634 // This edge should be set to top, by the set_complete. But be conservative. 2635 if (idx == InitializeNode::RawAddress) 2636 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2637 return RegMask::Empty; 2638} 2639 2640Node* InitializeNode::memory(uint alias_idx) { 2641 Node* mem = in(Memory); 2642 if (mem->is_MergeMem()) { 2643 return mem->as_MergeMem()->memory_at(alias_idx); 2644 } else { 2645 // incoming raw memory is not split 2646 return mem; 2647 } 2648} 2649 2650bool InitializeNode::is_non_zero() { 2651 if (is_complete()) return false; 2652 remove_extra_zeroes(); 2653 return (req() > RawStores); 2654} 2655 2656void InitializeNode::set_complete(PhaseGVN* phase) { 2657 assert(!is_complete(), "caller responsibility"); 2658 _is_complete = true; 2659 2660 // After this node is complete, it contains a bunch of 2661 // raw-memory initializations. There is no need for 2662 // it to have anything to do with non-raw memory effects. 2663 // Therefore, tell all non-raw users to re-optimize themselves, 2664 // after skipping the memory effects of this initialization. 2665 PhaseIterGVN* igvn = phase->is_IterGVN(); 2666 if (igvn) igvn->add_users_to_worklist(this); 2667} 2668 2669// convenience function 2670// return false if the init contains any stores already 2671bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2672 InitializeNode* init = initialization(); 2673 if (init == NULL || init->is_complete()) return false; 2674 init->remove_extra_zeroes(); 2675 // for now, if this allocation has already collected any inits, bail: 2676 if (init->is_non_zero()) return false; 2677 init->set_complete(phase); 2678 return true; 2679} 2680 2681void InitializeNode::remove_extra_zeroes() { 2682 if (req() == RawStores) return; 2683 Node* zmem = zero_memory(); 2684 uint fill = RawStores; 2685 for (uint i = fill; i < req(); i++) { 2686 Node* n = in(i); 2687 if (n->is_top() || n == zmem) continue; // skip 2688 if (fill < i) set_req(fill, n); // compact 2689 ++fill; 2690 } 2691 // delete any empty spaces created: 2692 while (fill < req()) { 2693 del_req(fill); 2694 } 2695} 2696 2697// Helper for remembering which stores go with which offsets. 2698intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2699 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2700 intptr_t offset = -1; 2701 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2702 phase, offset); 2703 if (base == NULL) return -1; // something is dead, 2704 if (offset < 0) return -1; // dead, dead 2705 return offset; 2706} 2707 2708// Helper for proving that an initialization expression is 2709// "simple enough" to be folded into an object initialization. 2710// Attempts to prove that a store's initial value 'n' can be captured 2711// within the initialization without creating a vicious cycle, such as: 2712// { Foo p = new Foo(); p.next = p; } 2713// True for constants and parameters and small combinations thereof. 2714bool InitializeNode::detect_init_independence(Node* n, 2715 bool st_is_pinned, 2716 int& count) { 2717 if (n == NULL) return true; // (can this really happen?) 2718 if (n->is_Proj()) n = n->in(0); 2719 if (n == this) return false; // found a cycle 2720 if (n->is_Con()) return true; 2721 if (n->is_Start()) return true; // params, etc., are OK 2722 if (n->is_Root()) return true; // even better 2723 2724 Node* ctl = n->in(0); 2725 if (ctl != NULL && !ctl->is_top()) { 2726 if (ctl->is_Proj()) ctl = ctl->in(0); 2727 if (ctl == this) return false; 2728 2729 // If we already know that the enclosing memory op is pinned right after 2730 // the init, then any control flow that the store has picked up 2731 // must have preceded the init, or else be equal to the init. 2732 // Even after loop optimizations (which might change control edges) 2733 // a store is never pinned *before* the availability of its inputs. 2734 if (!MemNode::all_controls_dominate(n, this)) 2735 return false; // failed to prove a good control 2736 2737 } 2738 2739 // Check data edges for possible dependencies on 'this'. 2740 if ((count += 1) > 20) return false; // complexity limit 2741 for (uint i = 1; i < n->req(); i++) { 2742 Node* m = n->in(i); 2743 if (m == NULL || m == n || m->is_top()) continue; 2744 uint first_i = n->find_edge(m); 2745 if (i != first_i) continue; // process duplicate edge just once 2746 if (!detect_init_independence(m, st_is_pinned, count)) { 2747 return false; 2748 } 2749 } 2750 2751 return true; 2752} 2753 2754// Here are all the checks a Store must pass before it can be moved into 2755// an initialization. Returns zero if a check fails. 2756// On success, returns the (constant) offset to which the store applies, 2757// within the initialized memory. 2758intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2759 const int FAIL = 0; 2760 if (st->req() != MemNode::ValueIn + 1) 2761 return FAIL; // an inscrutable StoreNode (card mark?) 2762 Node* ctl = st->in(MemNode::Control); 2763 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2764 return FAIL; // must be unconditional after the initialization 2765 Node* mem = st->in(MemNode::Memory); 2766 if (!(mem->is_Proj() && mem->in(0) == this)) 2767 return FAIL; // must not be preceded by other stores 2768 Node* adr = st->in(MemNode::Address); 2769 intptr_t offset; 2770 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2771 if (alloc == NULL) 2772 return FAIL; // inscrutable address 2773 if (alloc != allocation()) 2774 return FAIL; // wrong allocation! (store needs to float up) 2775 Node* val = st->in(MemNode::ValueIn); 2776 int complexity_count = 0; 2777 if (!detect_init_independence(val, true, complexity_count)) 2778 return FAIL; // stored value must be 'simple enough' 2779 2780 return offset; // success 2781} 2782 2783// Find the captured store in(i) which corresponds to the range 2784// [start..start+size) in the initialized object. 2785// If there is one, return its index i. If there isn't, return the 2786// negative of the index where it should be inserted. 2787// Return 0 if the queried range overlaps an initialization boundary 2788// or if dead code is encountered. 2789// If size_in_bytes is zero, do not bother with overlap checks. 2790int InitializeNode::captured_store_insertion_point(intptr_t start, 2791 int size_in_bytes, 2792 PhaseTransform* phase) { 2793 const int FAIL = 0, MAX_STORE = BytesPerLong; 2794 2795 if (is_complete()) 2796 return FAIL; // arraycopy got here first; punt 2797 2798 assert(allocation() != NULL, "must be present"); 2799 2800 // no negatives, no header fields: 2801 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 2802 2803 // after a certain size, we bail out on tracking all the stores: 2804 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2805 if (start >= ti_limit) return FAIL; 2806 2807 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 2808 if (i >= limit) return -(int)i; // not found; here is where to put it 2809 2810 Node* st = in(i); 2811 intptr_t st_off = get_store_offset(st, phase); 2812 if (st_off < 0) { 2813 if (st != zero_memory()) { 2814 return FAIL; // bail out if there is dead garbage 2815 } 2816 } else if (st_off > start) { 2817 // ...we are done, since stores are ordered 2818 if (st_off < start + size_in_bytes) { 2819 return FAIL; // the next store overlaps 2820 } 2821 return -(int)i; // not found; here is where to put it 2822 } else if (st_off < start) { 2823 if (size_in_bytes != 0 && 2824 start < st_off + MAX_STORE && 2825 start < st_off + st->as_Store()->memory_size()) { 2826 return FAIL; // the previous store overlaps 2827 } 2828 } else { 2829 if (size_in_bytes != 0 && 2830 st->as_Store()->memory_size() != size_in_bytes) { 2831 return FAIL; // mismatched store size 2832 } 2833 return i; 2834 } 2835 2836 ++i; 2837 } 2838} 2839 2840// Look for a captured store which initializes at the offset 'start' 2841// with the given size. If there is no such store, and no other 2842// initialization interferes, then return zero_memory (the memory 2843// projection of the AllocateNode). 2844Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 2845 PhaseTransform* phase) { 2846 assert(stores_are_sane(phase), ""); 2847 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2848 if (i == 0) { 2849 return NULL; // something is dead 2850 } else if (i < 0) { 2851 return zero_memory(); // just primordial zero bits here 2852 } else { 2853 Node* st = in(i); // here is the store at this position 2854 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 2855 return st; 2856 } 2857} 2858 2859// Create, as a raw pointer, an address within my new object at 'offset'. 2860Node* InitializeNode::make_raw_address(intptr_t offset, 2861 PhaseTransform* phase) { 2862 Node* addr = in(RawAddress); 2863 if (offset != 0) { 2864 Compile* C = phase->C; 2865 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 2866 phase->MakeConX(offset)) ); 2867 } 2868 return addr; 2869} 2870 2871// Clone the given store, converting it into a raw store 2872// initializing a field or element of my new object. 2873// Caller is responsible for retiring the original store, 2874// with subsume_node or the like. 2875// 2876// From the example above InitializeNode::InitializeNode, 2877// here are the old stores to be captured: 2878// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2879// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2880// 2881// Here is the changed code; note the extra edges on init: 2882// alloc = (Allocate ...) 2883// rawoop = alloc.RawAddress 2884// rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 2885// rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 2886// init = (Initialize alloc.Control alloc.Memory rawoop 2887// rawstore1 rawstore2) 2888// 2889Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 2890 PhaseTransform* phase) { 2891 assert(stores_are_sane(phase), ""); 2892 2893 if (start < 0) return NULL; 2894 assert(can_capture_store(st, phase) == start, "sanity"); 2895 2896 Compile* C = phase->C; 2897 int size_in_bytes = st->memory_size(); 2898 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2899 if (i == 0) return NULL; // bail out 2900 Node* prev_mem = NULL; // raw memory for the captured store 2901 if (i > 0) { 2902 prev_mem = in(i); // there is a pre-existing store under this one 2903 set_req(i, C->top()); // temporarily disconnect it 2904 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 2905 } else { 2906 i = -i; // no pre-existing store 2907 prev_mem = zero_memory(); // a slice of the newly allocated object 2908 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 2909 set_req(--i, C->top()); // reuse this edge; it has been folded away 2910 else 2911 ins_req(i, C->top()); // build a new edge 2912 } 2913 Node* new_st = st->clone(); 2914 new_st->set_req(MemNode::Control, in(Control)); 2915 new_st->set_req(MemNode::Memory, prev_mem); 2916 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 2917 new_st = phase->transform(new_st); 2918 2919 // At this point, new_st might have swallowed a pre-existing store 2920 // at the same offset, or perhaps new_st might have disappeared, 2921 // if it redundantly stored the same value (or zero to fresh memory). 2922 2923 // In any case, wire it in: 2924 set_req(i, new_st); 2925 2926 // The caller may now kill the old guy. 2927 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 2928 assert(check_st == new_st || check_st == NULL, "must be findable"); 2929 assert(!is_complete(), ""); 2930 return new_st; 2931} 2932 2933static bool store_constant(jlong* tiles, int num_tiles, 2934 intptr_t st_off, int st_size, 2935 jlong con) { 2936 if ((st_off & (st_size-1)) != 0) 2937 return false; // strange store offset (assume size==2**N) 2938 address addr = (address)tiles + st_off; 2939 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 2940 switch (st_size) { 2941 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 2942 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 2943 case sizeof(jint): *(jint*) addr = (jint) con; break; 2944 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 2945 default: return false; // strange store size (detect size!=2**N here) 2946 } 2947 return true; // return success to caller 2948} 2949 2950// Coalesce subword constants into int constants and possibly 2951// into long constants. The goal, if the CPU permits, 2952// is to initialize the object with a small number of 64-bit tiles. 2953// Also, convert floating-point constants to bit patterns. 2954// Non-constants are not relevant to this pass. 2955// 2956// In terms of the running example on InitializeNode::InitializeNode 2957// and InitializeNode::capture_store, here is the transformation 2958// of rawstore1 and rawstore2 into rawstore12: 2959// alloc = (Allocate ...) 2960// rawoop = alloc.RawAddress 2961// tile12 = 0x00010002 2962// rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 2963// init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 2964// 2965void 2966InitializeNode::coalesce_subword_stores(intptr_t header_size, 2967 Node* size_in_bytes, 2968 PhaseGVN* phase) { 2969 Compile* C = phase->C; 2970 2971 assert(stores_are_sane(phase), ""); 2972 // Note: After this pass, they are not completely sane, 2973 // since there may be some overlaps. 2974 2975 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 2976 2977 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2978 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 2979 size_limit = MIN2(size_limit, ti_limit); 2980 size_limit = align_size_up(size_limit, BytesPerLong); 2981 int num_tiles = size_limit / BytesPerLong; 2982 2983 // allocate space for the tile map: 2984 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 2985 jlong tiles_buf[small_len]; 2986 Node* nodes_buf[small_len]; 2987 jlong inits_buf[small_len]; 2988 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 2989 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2990 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 2991 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 2992 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 2993 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2994 // tiles: exact bitwise model of all primitive constants 2995 // nodes: last constant-storing node subsumed into the tiles model 2996 // inits: which bytes (in each tile) are touched by any initializations 2997 2998 //// Pass A: Fill in the tile model with any relevant stores. 2999 3000 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3001 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3002 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3003 Node* zmem = zero_memory(); // initially zero memory state 3004 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3005 Node* st = in(i); 3006 intptr_t st_off = get_store_offset(st, phase); 3007 3008 // Figure out the store's offset and constant value: 3009 if (st_off < header_size) continue; //skip (ignore header) 3010 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3011 int st_size = st->as_Store()->memory_size(); 3012 if (st_off + st_size > size_limit) break; 3013 3014 // Record which bytes are touched, whether by constant or not. 3015 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3016 continue; // skip (strange store size) 3017 3018 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3019 if (!val->singleton()) continue; //skip (non-con store) 3020 BasicType type = val->basic_type(); 3021 3022 jlong con = 0; 3023 switch (type) { 3024 case T_INT: con = val->is_int()->get_con(); break; 3025 case T_LONG: con = val->is_long()->get_con(); break; 3026 case T_FLOAT: con = jint_cast(val->getf()); break; 3027 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3028 default: continue; //skip (odd store type) 3029 } 3030 3031 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3032 st->Opcode() == Op_StoreL) { 3033 continue; // This StoreL is already optimal. 3034 } 3035 3036 // Store down the constant. 3037 store_constant(tiles, num_tiles, st_off, st_size, con); 3038 3039 intptr_t j = st_off >> LogBytesPerLong; 3040 3041 if (type == T_INT && st_size == BytesPerInt 3042 && (st_off & BytesPerInt) == BytesPerInt) { 3043 jlong lcon = tiles[j]; 3044 if (!Matcher::isSimpleConstant64(lcon) && 3045 st->Opcode() == Op_StoreI) { 3046 // This StoreI is already optimal by itself. 3047 jint* intcon = (jint*) &tiles[j]; 3048 intcon[1] = 0; // undo the store_constant() 3049 3050 // If the previous store is also optimal by itself, back up and 3051 // undo the action of the previous loop iteration... if we can. 3052 // But if we can't, just let the previous half take care of itself. 3053 st = nodes[j]; 3054 st_off -= BytesPerInt; 3055 con = intcon[0]; 3056 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3057 assert(st_off >= header_size, "still ignoring header"); 3058 assert(get_store_offset(st, phase) == st_off, "must be"); 3059 assert(in(i-1) == zmem, "must be"); 3060 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3061 assert(con == tcon->is_int()->get_con(), "must be"); 3062 // Undo the effects of the previous loop trip, which swallowed st: 3063 intcon[0] = 0; // undo store_constant() 3064 set_req(i-1, st); // undo set_req(i, zmem) 3065 nodes[j] = NULL; // undo nodes[j] = st 3066 --old_subword; // undo ++old_subword 3067 } 3068 continue; // This StoreI is already optimal. 3069 } 3070 } 3071 3072 // This store is not needed. 3073 set_req(i, zmem); 3074 nodes[j] = st; // record for the moment 3075 if (st_size < BytesPerLong) // something has changed 3076 ++old_subword; // includes int/float, but who's counting... 3077 else ++old_long; 3078 } 3079 3080 if ((old_subword + old_long) == 0) 3081 return; // nothing more to do 3082 3083 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3084 // Be sure to insert them before overlapping non-constant stores. 3085 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3086 for (int j = 0; j < num_tiles; j++) { 3087 jlong con = tiles[j]; 3088 jlong init = inits[j]; 3089 if (con == 0) continue; 3090 jint con0, con1; // split the constant, address-wise 3091 jint init0, init1; // split the init map, address-wise 3092 { union { jlong con; jint intcon[2]; } u; 3093 u.con = con; 3094 con0 = u.intcon[0]; 3095 con1 = u.intcon[1]; 3096 u.con = init; 3097 init0 = u.intcon[0]; 3098 init1 = u.intcon[1]; 3099 } 3100 3101 Node* old = nodes[j]; 3102 assert(old != NULL, "need the prior store"); 3103 intptr_t offset = (j * BytesPerLong); 3104 3105 bool split = !Matcher::isSimpleConstant64(con); 3106 3107 if (offset < header_size) { 3108 assert(offset + BytesPerInt >= header_size, "second int counts"); 3109 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3110 split = true; // only the second word counts 3111 // Example: int a[] = { 42 ... } 3112 } else if (con0 == 0 && init0 == -1) { 3113 split = true; // first word is covered by full inits 3114 // Example: int a[] = { ... foo(), 42 ... } 3115 } else if (con1 == 0 && init1 == -1) { 3116 split = true; // second word is covered by full inits 3117 // Example: int a[] = { ... 42, foo() ... } 3118 } 3119 3120 // Here's a case where init0 is neither 0 nor -1: 3121 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3122 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3123 // In this case the tile is not split; it is (jlong)42. 3124 // The big tile is stored down, and then the foo() value is inserted. 3125 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3126 3127 Node* ctl = old->in(MemNode::Control); 3128 Node* adr = make_raw_address(offset, phase); 3129 const TypePtr* atp = TypeRawPtr::BOTTOM; 3130 3131 // One or two coalesced stores to plop down. 3132 Node* st[2]; 3133 intptr_t off[2]; 3134 int nst = 0; 3135 if (!split) { 3136 ++new_long; 3137 off[nst] = offset; 3138 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3139 phase->longcon(con), T_LONG); 3140 } else { 3141 // Omit either if it is a zero. 3142 if (con0 != 0) { 3143 ++new_int; 3144 off[nst] = offset; 3145 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3146 phase->intcon(con0), T_INT); 3147 } 3148 if (con1 != 0) { 3149 ++new_int; 3150 offset += BytesPerInt; 3151 adr = make_raw_address(offset, phase); 3152 off[nst] = offset; 3153 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3154 phase->intcon(con1), T_INT); 3155 } 3156 } 3157 3158 // Insert second store first, then the first before the second. 3159 // Insert each one just before any overlapping non-constant stores. 3160 while (nst > 0) { 3161 Node* st1 = st[--nst]; 3162 C->copy_node_notes_to(st1, old); 3163 st1 = phase->transform(st1); 3164 offset = off[nst]; 3165 assert(offset >= header_size, "do not smash header"); 3166 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3167 guarantee(ins_idx != 0, "must re-insert constant store"); 3168 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3169 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3170 set_req(--ins_idx, st1); 3171 else 3172 ins_req(ins_idx, st1); 3173 } 3174 } 3175 3176 if (PrintCompilation && WizardMode) 3177 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3178 old_subword, old_long, new_int, new_long); 3179 if (C->log() != NULL) 3180 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3181 old_subword, old_long, new_int, new_long); 3182 3183 // Clean up any remaining occurrences of zmem: 3184 remove_extra_zeroes(); 3185} 3186 3187// Explore forward from in(start) to find the first fully initialized 3188// word, and return its offset. Skip groups of subword stores which 3189// together initialize full words. If in(start) is itself part of a 3190// fully initialized word, return the offset of in(start). If there 3191// are no following full-word stores, or if something is fishy, return 3192// a negative value. 3193intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3194 int int_map = 0; 3195 intptr_t int_map_off = 0; 3196 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3197 3198 for (uint i = start, limit = req(); i < limit; i++) { 3199 Node* st = in(i); 3200 3201 intptr_t st_off = get_store_offset(st, phase); 3202 if (st_off < 0) break; // return conservative answer 3203 3204 int st_size = st->as_Store()->memory_size(); 3205 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3206 return st_off; // we found a complete word init 3207 } 3208 3209 // update the map: 3210 3211 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3212 if (this_int_off != int_map_off) { 3213 // reset the map: 3214 int_map = 0; 3215 int_map_off = this_int_off; 3216 } 3217 3218 int subword_off = st_off - this_int_off; 3219 int_map |= right_n_bits(st_size) << subword_off; 3220 if ((int_map & FULL_MAP) == FULL_MAP) { 3221 return this_int_off; // we found a complete word init 3222 } 3223 3224 // Did this store hit or cross the word boundary? 3225 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3226 if (next_int_off == this_int_off + BytesPerInt) { 3227 // We passed the current int, without fully initializing it. 3228 int_map_off = next_int_off; 3229 int_map >>= BytesPerInt; 3230 } else if (next_int_off > this_int_off + BytesPerInt) { 3231 // We passed the current and next int. 3232 return this_int_off + BytesPerInt; 3233 } 3234 } 3235 3236 return -1; 3237} 3238 3239 3240// Called when the associated AllocateNode is expanded into CFG. 3241// At this point, we may perform additional optimizations. 3242// Linearize the stores by ascending offset, to make memory 3243// activity as coherent as possible. 3244Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3245 intptr_t header_size, 3246 Node* size_in_bytes, 3247 PhaseGVN* phase) { 3248 assert(!is_complete(), "not already complete"); 3249 assert(stores_are_sane(phase), ""); 3250 assert(allocation() != NULL, "must be present"); 3251 3252 remove_extra_zeroes(); 3253 3254 if (ReduceFieldZeroing || ReduceBulkZeroing) 3255 // reduce instruction count for common initialization patterns 3256 coalesce_subword_stores(header_size, size_in_bytes, phase); 3257 3258 Node* zmem = zero_memory(); // initially zero memory state 3259 Node* inits = zmem; // accumulating a linearized chain of inits 3260 #ifdef ASSERT 3261 intptr_t first_offset = allocation()->minimum_header_size(); 3262 intptr_t last_init_off = first_offset; // previous init offset 3263 intptr_t last_init_end = first_offset; // previous init offset+size 3264 intptr_t last_tile_end = first_offset; // previous tile offset+size 3265 #endif 3266 intptr_t zeroes_done = header_size; 3267 3268 bool do_zeroing = true; // we might give up if inits are very sparse 3269 int big_init_gaps = 0; // how many large gaps have we seen? 3270 3271 if (ZeroTLAB) do_zeroing = false; 3272 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3273 3274 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3275 Node* st = in(i); 3276 intptr_t st_off = get_store_offset(st, phase); 3277 if (st_off < 0) 3278 break; // unknown junk in the inits 3279 if (st->in(MemNode::Memory) != zmem) 3280 break; // complicated store chains somehow in list 3281 3282 int st_size = st->as_Store()->memory_size(); 3283 intptr_t next_init_off = st_off + st_size; 3284 3285 if (do_zeroing && zeroes_done < next_init_off) { 3286 // See if this store needs a zero before it or under it. 3287 intptr_t zeroes_needed = st_off; 3288 3289 if (st_size < BytesPerInt) { 3290 // Look for subword stores which only partially initialize words. 3291 // If we find some, we must lay down some word-level zeroes first, 3292 // underneath the subword stores. 3293 // 3294 // Examples: 3295 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3296 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3297 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3298 // 3299 // Note: coalesce_subword_stores may have already done this, 3300 // if it was prompted by constant non-zero subword initializers. 3301 // But this case can still arise with non-constant stores. 3302 3303 intptr_t next_full_store = find_next_fullword_store(i, phase); 3304 3305 // In the examples above: 3306 // in(i) p q r s x y z 3307 // st_off 12 13 14 15 12 13 14 3308 // st_size 1 1 1 1 1 1 1 3309 // next_full_s. 12 16 16 16 16 16 16 3310 // z's_done 12 16 16 16 12 16 12 3311 // z's_needed 12 16 16 16 16 16 16 3312 // zsize 0 0 0 0 4 0 4 3313 if (next_full_store < 0) { 3314 // Conservative tack: Zero to end of current word. 3315 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3316 } else { 3317 // Zero to beginning of next fully initialized word. 3318 // Or, don't zero at all, if we are already in that word. 3319 assert(next_full_store >= zeroes_needed, "must go forward"); 3320 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3321 zeroes_needed = next_full_store; 3322 } 3323 } 3324 3325 if (zeroes_needed > zeroes_done) { 3326 intptr_t zsize = zeroes_needed - zeroes_done; 3327 // Do some incremental zeroing on rawmem, in parallel with inits. 3328 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3329 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3330 zeroes_done, zeroes_needed, 3331 phase); 3332 zeroes_done = zeroes_needed; 3333 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3334 do_zeroing = false; // leave the hole, next time 3335 } 3336 } 3337 3338 // Collect the store and move on: 3339 st->set_req(MemNode::Memory, inits); 3340 inits = st; // put it on the linearized chain 3341 set_req(i, zmem); // unhook from previous position 3342 3343 if (zeroes_done == st_off) 3344 zeroes_done = next_init_off; 3345 3346 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3347 3348 #ifdef ASSERT 3349 // Various order invariants. Weaker than stores_are_sane because 3350 // a large constant tile can be filled in by smaller non-constant stores. 3351 assert(st_off >= last_init_off, "inits do not reverse"); 3352 last_init_off = st_off; 3353 const Type* val = NULL; 3354 if (st_size >= BytesPerInt && 3355 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3356 (int)val->basic_type() < (int)T_OBJECT) { 3357 assert(st_off >= last_tile_end, "tiles do not overlap"); 3358 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3359 last_tile_end = MAX2(last_tile_end, next_init_off); 3360 } else { 3361 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3362 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3363 assert(st_off >= last_init_end, "inits do not overlap"); 3364 last_init_end = next_init_off; // it's a non-tile 3365 } 3366 #endif //ASSERT 3367 } 3368 3369 remove_extra_zeroes(); // clear out all the zmems left over 3370 add_req(inits); 3371 3372 if (!ZeroTLAB) { 3373 // If anything remains to be zeroed, zero it all now. 3374 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3375 // if it is the last unused 4 bytes of an instance, forget about it 3376 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3377 if (zeroes_done + BytesPerLong >= size_limit) { 3378 assert(allocation() != NULL, ""); 3379 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3380 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3381 if (zeroes_done == k->layout_helper()) 3382 zeroes_done = size_limit; 3383 } 3384 if (zeroes_done < size_limit) { 3385 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3386 zeroes_done, size_in_bytes, phase); 3387 } 3388 } 3389 3390 set_complete(phase); 3391 return rawmem; 3392} 3393 3394 3395#ifdef ASSERT 3396bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3397 if (is_complete()) 3398 return true; // stores could be anything at this point 3399 assert(allocation() != NULL, "must be present"); 3400 intptr_t last_off = allocation()->minimum_header_size(); 3401 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3402 Node* st = in(i); 3403 intptr_t st_off = get_store_offset(st, phase); 3404 if (st_off < 0) continue; // ignore dead garbage 3405 if (last_off > st_off) { 3406 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3407 this->dump(2); 3408 assert(false, "ascending store offsets"); 3409 return false; 3410 } 3411 last_off = st_off + st->as_Store()->memory_size(); 3412 } 3413 return true; 3414} 3415#endif //ASSERT 3416 3417 3418 3419 3420//============================MergeMemNode===================================== 3421// 3422// SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3423// contributing store or call operations. Each contributor provides the memory 3424// state for a particular "alias type" (see Compile::alias_type). For example, 3425// if a MergeMem has an input X for alias category #6, then any memory reference 3426// to alias category #6 may use X as its memory state input, as an exact equivalent 3427// to using the MergeMem as a whole. 3428// Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3429// 3430// (Here, the <N> notation gives the index of the relevant adr_type.) 3431// 3432// In one special case (and more cases in the future), alias categories overlap. 3433// The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3434// states. Therefore, if a MergeMem has only one contributing input W for Bot, 3435// it is exactly equivalent to that state W: 3436// MergeMem(<Bot>: W) <==> W 3437// 3438// Usually, the merge has more than one input. In that case, where inputs 3439// overlap (i.e., one is Bot), the narrower alias type determines the memory 3440// state for that type, and the wider alias type (Bot) fills in everywhere else: 3441// Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3442// Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3443// 3444// A merge can take a "wide" memory state as one of its narrow inputs. 3445// This simply means that the merge observes out only the relevant parts of 3446// the wide input. That is, wide memory states arriving at narrow merge inputs 3447// are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3448// 3449// These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3450// and that memory slices "leak through": 3451// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3452// 3453// But, in such a cascade, repeated memory slices can "block the leak": 3454// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3455// 3456// In the last example, Y is not part of the combined memory state of the 3457// outermost MergeMem. The system must, of course, prevent unschedulable 3458// memory states from arising, so you can be sure that the state Y is somehow 3459// a precursor to state Y'. 3460// 3461// 3462// REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3463// of each MergeMemNode array are exactly the numerical alias indexes, including 3464// but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3465// Compile::alias_type (and kin) produce and manage these indexes. 3466// 3467// By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3468// (Note that this provides quick access to the top node inside MergeMem methods, 3469// without the need to reach out via TLS to Compile::current.) 3470// 3471// As a consequence of what was just described, a MergeMem that represents a full 3472// memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3473// containing all alias categories. 3474// 3475// MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3476// 3477// All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3478// a memory state for the alias type <N>, or else the top node, meaning that 3479// there is no particular input for that alias type. Note that the length of 3480// a MergeMem is variable, and may be extended at any time to accommodate new 3481// memory states at larger alias indexes. When merges grow, they are of course 3482// filled with "top" in the unused in() positions. 3483// 3484// This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3485// (Top was chosen because it works smoothly with passes like GCM.) 3486// 3487// For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3488// the type of random VM bits like TLS references.) Since it is always the 3489// first non-Bot memory slice, some low-level loops use it to initialize an 3490// index variable: for (i = AliasIdxRaw; i < req(); i++). 3491// 3492// 3493// ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3494// the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3495// the memory state for alias type <N>, or (if there is no particular slice at <N>, 3496// it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3497// or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3498// MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3499// 3500// %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3501// really that different from the other memory inputs. An abbreviation called 3502// "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3503// 3504// 3505// PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3506// partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3507// that "emerges though" the base memory will be marked as excluding the alias types 3508// of the other (narrow-memory) copies which "emerged through" the narrow edges: 3509// 3510// Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3511// ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3512// 3513// This strange "subtraction" effect is necessary to ensure IGVN convergence. 3514// (It is currently unimplemented.) As you can see, the resulting merge is 3515// actually a disjoint union of memory states, rather than an overlay. 3516// 3517 3518//------------------------------MergeMemNode----------------------------------- 3519Node* MergeMemNode::make_empty_memory() { 3520 Node* empty_memory = (Node*) Compile::current()->top(); 3521 assert(empty_memory->is_top(), "correct sentinel identity"); 3522 return empty_memory; 3523} 3524 3525MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3526 init_class_id(Class_MergeMem); 3527 // all inputs are nullified in Node::Node(int) 3528 // set_input(0, NULL); // no control input 3529 3530 // Initialize the edges uniformly to top, for starters. 3531 Node* empty_mem = make_empty_memory(); 3532 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3533 init_req(i,empty_mem); 3534 } 3535 assert(empty_memory() == empty_mem, ""); 3536 3537 if( new_base != NULL && new_base->is_MergeMem() ) { 3538 MergeMemNode* mdef = new_base->as_MergeMem(); 3539 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3540 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3541 mms.set_memory(mms.memory2()); 3542 } 3543 assert(base_memory() == mdef->base_memory(), ""); 3544 } else { 3545 set_base_memory(new_base); 3546 } 3547} 3548 3549// Make a new, untransformed MergeMem with the same base as 'mem'. 3550// If mem is itself a MergeMem, populate the result with the same edges. 3551MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3552 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3553} 3554 3555//------------------------------cmp-------------------------------------------- 3556uint MergeMemNode::hash() const { return NO_HASH; } 3557uint MergeMemNode::cmp( const Node &n ) const { 3558 return (&n == this); // Always fail except on self 3559} 3560 3561//------------------------------Identity--------------------------------------- 3562Node* MergeMemNode::Identity(PhaseTransform *phase) { 3563 // Identity if this merge point does not record any interesting memory 3564 // disambiguations. 3565 Node* base_mem = base_memory(); 3566 Node* empty_mem = empty_memory(); 3567 if (base_mem != empty_mem) { // Memory path is not dead? 3568 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3569 Node* mem = in(i); 3570 if (mem != empty_mem && mem != base_mem) { 3571 return this; // Many memory splits; no change 3572 } 3573 } 3574 } 3575 return base_mem; // No memory splits; ID on the one true input 3576} 3577 3578//------------------------------Ideal------------------------------------------ 3579// This method is invoked recursively on chains of MergeMem nodes 3580Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3581 // Remove chain'd MergeMems 3582 // 3583 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3584 // relative to the "in(Bot)". Since we are patching both at the same time, 3585 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3586 // but rewrite each "in(i)" relative to the new "in(Bot)". 3587 Node *progress = NULL; 3588 3589 3590 Node* old_base = base_memory(); 3591 Node* empty_mem = empty_memory(); 3592 if (old_base == empty_mem) 3593 return NULL; // Dead memory path. 3594 3595 MergeMemNode* old_mbase; 3596 if (old_base != NULL && old_base->is_MergeMem()) 3597 old_mbase = old_base->as_MergeMem(); 3598 else 3599 old_mbase = NULL; 3600 Node* new_base = old_base; 3601 3602 // simplify stacked MergeMems in base memory 3603 if (old_mbase) new_base = old_mbase->base_memory(); 3604 3605 // the base memory might contribute new slices beyond my req() 3606 if (old_mbase) grow_to_match(old_mbase); 3607 3608 // Look carefully at the base node if it is a phi. 3609 PhiNode* phi_base; 3610 if (new_base != NULL && new_base->is_Phi()) 3611 phi_base = new_base->as_Phi(); 3612 else 3613 phi_base = NULL; 3614 3615 Node* phi_reg = NULL; 3616 uint phi_len = (uint)-1; 3617 if (phi_base != NULL && !phi_base->is_copy()) { 3618 // do not examine phi if degraded to a copy 3619 phi_reg = phi_base->region(); 3620 phi_len = phi_base->req(); 3621 // see if the phi is unfinished 3622 for (uint i = 1; i < phi_len; i++) { 3623 if (phi_base->in(i) == NULL) { 3624 // incomplete phi; do not look at it yet! 3625 phi_reg = NULL; 3626 phi_len = (uint)-1; 3627 break; 3628 } 3629 } 3630 } 3631 3632 // Note: We do not call verify_sparse on entry, because inputs 3633 // can normalize to the base_memory via subsume_node or similar 3634 // mechanisms. This method repairs that damage. 3635 3636 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3637 3638 // Look at each slice. 3639 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3640 Node* old_in = in(i); 3641 // calculate the old memory value 3642 Node* old_mem = old_in; 3643 if (old_mem == empty_mem) old_mem = old_base; 3644 assert(old_mem == memory_at(i), ""); 3645 3646 // maybe update (reslice) the old memory value 3647 3648 // simplify stacked MergeMems 3649 Node* new_mem = old_mem; 3650 MergeMemNode* old_mmem; 3651 if (old_mem != NULL && old_mem->is_MergeMem()) 3652 old_mmem = old_mem->as_MergeMem(); 3653 else 3654 old_mmem = NULL; 3655 if (old_mmem == this) { 3656 // This can happen if loops break up and safepoints disappear. 3657 // A merge of BotPtr (default) with a RawPtr memory derived from a 3658 // safepoint can be rewritten to a merge of the same BotPtr with 3659 // the BotPtr phi coming into the loop. If that phi disappears 3660 // also, we can end up with a self-loop of the mergemem. 3661 // In general, if loops degenerate and memory effects disappear, 3662 // a mergemem can be left looking at itself. This simply means 3663 // that the mergemem's default should be used, since there is 3664 // no longer any apparent effect on this slice. 3665 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3666 // from start. Update the input to TOP. 3667 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3668 } 3669 else if (old_mmem != NULL) { 3670 new_mem = old_mmem->memory_at(i); 3671 } 3672 // else preceeding memory was not a MergeMem 3673 3674 // replace equivalent phis (unfortunately, they do not GVN together) 3675 if (new_mem != NULL && new_mem != new_base && 3676 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3677 if (new_mem->is_Phi()) { 3678 PhiNode* phi_mem = new_mem->as_Phi(); 3679 for (uint i = 1; i < phi_len; i++) { 3680 if (phi_base->in(i) != phi_mem->in(i)) { 3681 phi_mem = NULL; 3682 break; 3683 } 3684 } 3685 if (phi_mem != NULL) { 3686 // equivalent phi nodes; revert to the def 3687 new_mem = new_base; 3688 } 3689 } 3690 } 3691 3692 // maybe store down a new value 3693 Node* new_in = new_mem; 3694 if (new_in == new_base) new_in = empty_mem; 3695 3696 if (new_in != old_in) { 3697 // Warning: Do not combine this "if" with the previous "if" 3698 // A memory slice might have be be rewritten even if it is semantically 3699 // unchanged, if the base_memory value has changed. 3700 set_req(i, new_in); 3701 progress = this; // Report progress 3702 } 3703 } 3704 3705 if (new_base != old_base) { 3706 set_req(Compile::AliasIdxBot, new_base); 3707 // Don't use set_base_memory(new_base), because we need to update du. 3708 assert(base_memory() == new_base, ""); 3709 progress = this; 3710 } 3711 3712 if( base_memory() == this ) { 3713 // a self cycle indicates this memory path is dead 3714 set_req(Compile::AliasIdxBot, empty_mem); 3715 } 3716 3717 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3718 // Recursion must occur after the self cycle check above 3719 if( base_memory()->is_MergeMem() ) { 3720 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3721 Node *m = phase->transform(new_mbase); // Rollup any cycles 3722 if( m != NULL && (m->is_top() || 3723 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3724 // propagate rollup of dead cycle to self 3725 set_req(Compile::AliasIdxBot, empty_mem); 3726 } 3727 } 3728 3729 if( base_memory() == empty_mem ) { 3730 progress = this; 3731 // Cut inputs during Parse phase only. 3732 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3733 if( !can_reshape ) { 3734 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3735 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3736 } 3737 } 3738 } 3739 3740 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3741 // Check if PhiNode::Ideal's "Split phis through memory merges" 3742 // transform should be attempted. Look for this->phi->this cycle. 3743 uint merge_width = req(); 3744 if (merge_width > Compile::AliasIdxRaw) { 3745 PhiNode* phi = base_memory()->as_Phi(); 3746 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3747 if (phi->in(i) == this) { 3748 phase->is_IterGVN()->_worklist.push(phi); 3749 break; 3750 } 3751 } 3752 } 3753 } 3754 3755 assert(progress || verify_sparse(), "please, no dups of base"); 3756 return progress; 3757} 3758 3759//-------------------------set_base_memory------------------------------------- 3760void MergeMemNode::set_base_memory(Node *new_base) { 3761 Node* empty_mem = empty_memory(); 3762 set_req(Compile::AliasIdxBot, new_base); 3763 assert(memory_at(req()) == new_base, "must set default memory"); 3764 // Clear out other occurrences of new_base: 3765 if (new_base != empty_mem) { 3766 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3767 if (in(i) == new_base) set_req(i, empty_mem); 3768 } 3769 } 3770} 3771 3772//------------------------------out_RegMask------------------------------------ 3773const RegMask &MergeMemNode::out_RegMask() const { 3774 return RegMask::Empty; 3775} 3776 3777//------------------------------dump_spec-------------------------------------- 3778#ifndef PRODUCT 3779void MergeMemNode::dump_spec(outputStream *st) const { 3780 st->print(" {"); 3781 Node* base_mem = base_memory(); 3782 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3783 Node* mem = memory_at(i); 3784 if (mem == base_mem) { st->print(" -"); continue; } 3785 st->print( " N%d:", mem->_idx ); 3786 Compile::current()->get_adr_type(i)->dump_on(st); 3787 } 3788 st->print(" }"); 3789} 3790#endif // !PRODUCT 3791 3792 3793#ifdef ASSERT 3794static bool might_be_same(Node* a, Node* b) { 3795 if (a == b) return true; 3796 if (!(a->is_Phi() || b->is_Phi())) return false; 3797 // phis shift around during optimization 3798 return true; // pretty stupid... 3799} 3800 3801// verify a narrow slice (either incoming or outgoing) 3802static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 3803 if (!VerifyAliases) return; // don't bother to verify unless requested 3804 if (is_error_reported()) return; // muzzle asserts when debugging an error 3805 if (Node::in_dump()) return; // muzzle asserts when printing 3806 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 3807 assert(n != NULL, ""); 3808 // Elide intervening MergeMem's 3809 while (n->is_MergeMem()) { 3810 n = n->as_MergeMem()->memory_at(alias_idx); 3811 } 3812 Compile* C = Compile::current(); 3813 const TypePtr* n_adr_type = n->adr_type(); 3814 if (n == m->empty_memory()) { 3815 // Implicit copy of base_memory() 3816 } else if (n_adr_type != TypePtr::BOTTOM) { 3817 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 3818 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 3819 } else { 3820 // A few places like make_runtime_call "know" that VM calls are narrow, 3821 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 3822 bool expected_wide_mem = false; 3823 if (n == m->base_memory()) { 3824 expected_wide_mem = true; 3825 } else if (alias_idx == Compile::AliasIdxRaw || 3826 n == m->memory_at(Compile::AliasIdxRaw)) { 3827 expected_wide_mem = true; 3828 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 3829 // memory can "leak through" calls on channels that 3830 // are write-once. Allow this also. 3831 expected_wide_mem = true; 3832 } 3833 assert(expected_wide_mem, "expected narrow slice replacement"); 3834 } 3835} 3836#else // !ASSERT 3837#define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 3838#endif 3839 3840 3841//-----------------------------memory_at--------------------------------------- 3842Node* MergeMemNode::memory_at(uint alias_idx) const { 3843 assert(alias_idx >= Compile::AliasIdxRaw || 3844 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 3845 "must avoid base_memory and AliasIdxTop"); 3846 3847 // Otherwise, it is a narrow slice. 3848 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 3849 Compile *C = Compile::current(); 3850 if (is_empty_memory(n)) { 3851 // the array is sparse; empty slots are the "top" node 3852 n = base_memory(); 3853 assert(Node::in_dump() 3854 || n == NULL || n->bottom_type() == Type::TOP 3855 || n->adr_type() == TypePtr::BOTTOM 3856 || n->adr_type() == TypeRawPtr::BOTTOM 3857 || Compile::current()->AliasLevel() == 0, 3858 "must be a wide memory"); 3859 // AliasLevel == 0 if we are organizing the memory states manually. 3860 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 3861 } else { 3862 // make sure the stored slice is sane 3863 #ifdef ASSERT 3864 if (is_error_reported() || Node::in_dump()) { 3865 } else if (might_be_same(n, base_memory())) { 3866 // Give it a pass: It is a mostly harmless repetition of the base. 3867 // This can arise normally from node subsumption during optimization. 3868 } else { 3869 verify_memory_slice(this, alias_idx, n); 3870 } 3871 #endif 3872 } 3873 return n; 3874} 3875 3876//---------------------------set_memory_at------------------------------------- 3877void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 3878 verify_memory_slice(this, alias_idx, n); 3879 Node* empty_mem = empty_memory(); 3880 if (n == base_memory()) n = empty_mem; // collapse default 3881 uint need_req = alias_idx+1; 3882 if (req() < need_req) { 3883 if (n == empty_mem) return; // already the default, so do not grow me 3884 // grow the sparse array 3885 do { 3886 add_req(empty_mem); 3887 } while (req() < need_req); 3888 } 3889 set_req( alias_idx, n ); 3890} 3891 3892 3893 3894//--------------------------iteration_setup------------------------------------ 3895void MergeMemNode::iteration_setup(const MergeMemNode* other) { 3896 if (other != NULL) { 3897 grow_to_match(other); 3898 // invariant: the finite support of mm2 is within mm->req() 3899 #ifdef ASSERT 3900 for (uint i = req(); i < other->req(); i++) { 3901 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 3902 } 3903 #endif 3904 } 3905 // Replace spurious copies of base_memory by top. 3906 Node* base_mem = base_memory(); 3907 if (base_mem != NULL && !base_mem->is_top()) { 3908 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 3909 if (in(i) == base_mem) 3910 set_req(i, empty_memory()); 3911 } 3912 } 3913} 3914 3915//---------------------------grow_to_match------------------------------------- 3916void MergeMemNode::grow_to_match(const MergeMemNode* other) { 3917 Node* empty_mem = empty_memory(); 3918 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 3919 // look for the finite support of the other memory 3920 for (uint i = other->req(); --i >= req(); ) { 3921 if (other->in(i) != empty_mem) { 3922 uint new_len = i+1; 3923 while (req() < new_len) add_req(empty_mem); 3924 break; 3925 } 3926 } 3927} 3928 3929//---------------------------verify_sparse------------------------------------- 3930#ifndef PRODUCT 3931bool MergeMemNode::verify_sparse() const { 3932 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 3933 Node* base_mem = base_memory(); 3934 // The following can happen in degenerate cases, since empty==top. 3935 if (is_empty_memory(base_mem)) return true; 3936 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3937 assert(in(i) != NULL, "sane slice"); 3938 if (in(i) == base_mem) return false; // should have been the sentinel value! 3939 } 3940 return true; 3941} 3942 3943bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 3944 Node* n; 3945 n = mm->in(idx); 3946 if (mem == n) return true; // might be empty_memory() 3947 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 3948 if (mem == n) return true; 3949 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 3950 if (mem == n) return true; 3951 if (n == NULL) break; 3952 } 3953 return false; 3954} 3955#endif // !PRODUCT 3956