1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 // Portions of code courtesy of Clifford Click 26 27 // Optimization - Graph Style 28 29 #include "incls/_precompiled.incl" 30 #include "incls/_memnode.cpp.incl" 31 32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 33 34 //============================================================================= 35 uint MemNode::size_of() const { return sizeof(*this); } 36 37 const TypePtr *MemNode::adr_type() const { 38 Node* adr = in(Address); 39 const TypePtr* cross_check = NULL; 40 DEBUG_ONLY(cross_check = _adr_type); 41 return calculate_adr_type(adr->bottom_type(), cross_check); 42 } 43 44 #ifndef PRODUCT 45 void MemNode::dump_spec(outputStream *st) const { 46 if (in(Address) == NULL) return; // node is dead 47 #ifndef ASSERT 48 // fake the missing field 49 const TypePtr* _adr_type = NULL; 50 if (in(Address) != NULL) 51 _adr_type = in(Address)->bottom_type()->isa_ptr(); 52 #endif 53 dump_adr_type(this, _adr_type, st); 54 55 Compile* C = Compile::current(); 56 if( C->alias_type(_adr_type)->is_volatile() ) 57 st->print(" Volatile!"); 58 } 59 60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 61 st->print(" @"); 62 if (adr_type == NULL) { 63 st->print("NULL"); 64 } else { 65 adr_type->dump_on(st); 66 Compile* C = Compile::current(); 67 Compile::AliasType* atp = NULL; 68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 69 if (atp == NULL) 70 st->print(", idx=?\?;"); 71 else if (atp->index() == Compile::AliasIdxBot) 72 st->print(", idx=Bot;"); 73 else if (atp->index() == Compile::AliasIdxTop) 74 st->print(", idx=Top;"); 75 else if (atp->index() == Compile::AliasIdxRaw) 76 st->print(", idx=Raw;"); 77 else { 78 ciField* field = atp->field(); 79 if (field) { 80 st->print(", name="); 81 field->print_name_on(st); 82 } 83 st->print(", idx=%d;", atp->index()); 84 } 85 } 86 } 87 88 extern void print_alias_types(); 89 90 #endif 91 92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 93 const TypeOopPtr *tinst = t_adr->isa_oopptr(); 94 if (tinst == NULL || !tinst->is_known_instance_field()) 95 return mchain; // don't try to optimize non-instance types 96 uint instance_id = tinst->instance_id(); 97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 98 Node *prev = NULL; 99 Node *result = mchain; 100 while (prev != result) { 101 prev = result; 102 if (result == start_mem) 103 break; // hit one of our sentinels 104 // skip over a call which does not affect this memory slice 105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 106 Node *proj_in = result->in(0); 107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 108 break; // hit one of our sentinels 109 } else if (proj_in->is_Call()) { 110 CallNode *call = proj_in->as_Call(); 111 if (!call->may_modify(t_adr, phase)) { 112 result = call->in(TypeFunc::Memory); 113 } 114 } else if (proj_in->is_Initialize()) { 115 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 116 // Stop if this is the initialization for the object instance which 117 // which contains this memory slice, otherwise skip over it. 118 if (alloc != NULL && alloc->_idx != instance_id) { 119 result = proj_in->in(TypeFunc::Memory); 120 } 121 } else if (proj_in->is_MemBar()) { 122 result = proj_in->in(TypeFunc::Memory); 123 } else { 124 assert(false, "unexpected projection"); 125 } 126 } else if (result->Opcode() == Op_ClearArray) { 127 intptr_t offset; 128 AllocateNode* alloc = AllocateNode::Ideal_allocation(result->in(3), phase, offset); 129 // We can not bypass initialization of the instance we are looking for 130 // or when something is wrong. 131 if (alloc == NULL || alloc->_idx == (uint)tinst->instance_id()) 132 break; 133 // Otherwise skip it. 134 InitializeNode* init = alloc->initialization(); 135 if (init != NULL) 136 result = init->in(TypeFunc::Memory); 137 else 138 result = alloc->in(TypeFunc::Memory); 139 } else if (result->is_MergeMem()) { 140 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 141 } 142 } 143 return result; 144 } 145 146 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 147 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 148 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); 149 PhaseIterGVN *igvn = phase->is_IterGVN(); 150 Node *result = mchain; 151 result = optimize_simple_memory_chain(result, t_adr, phase); 152 if (is_instance && igvn != NULL && result->is_Phi()) { 153 PhiNode *mphi = result->as_Phi(); 154 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 155 const TypePtr *t = mphi->adr_type(); 156 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 157 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 158 t->is_oopptr()->cast_to_exactness(true) 159 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 160 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 161 // clone the Phi with our address type 162 result = mphi->split_out_instance(t_adr, igvn); 163 } else { 164 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 165 } 166 } 167 return result; 168 } 169 170 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 171 uint alias_idx = phase->C->get_alias_index(tp); 172 Node *mem = mmem; 173 #ifdef ASSERT 174 { 175 // Check that current type is consistent with the alias index used during graph construction 176 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 177 bool consistent = adr_check == NULL || adr_check->empty() || 178 phase->C->must_alias(adr_check, alias_idx ); 179 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 180 if( !consistent && adr_check != NULL && !adr_check->empty() && 181 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 182 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 183 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 184 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 185 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 186 // don't assert if it is dead code. 187 consistent = true; 188 } 189 if( !consistent ) { 190 st->print("alias_idx==%d, adr_check==", alias_idx); 191 if( adr_check == NULL ) { 192 st->print("NULL"); 193 } else { 194 adr_check->dump(); 195 } 196 st->cr(); 197 print_alias_types(); 198 assert(consistent, "adr_check must match alias idx"); 199 } 200 } 201 #endif 202 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 203 // means an array I have not precisely typed yet. Do not do any 204 // alias stuff with it any time soon. 205 const TypeOopPtr *tinst = tp->isa_oopptr(); 206 if( tp->base() != Type::AnyPtr && 207 !(tinst && 208 tinst->klass()->is_java_lang_Object() && 209 tinst->offset() == Type::OffsetBot) ) { 210 // compress paths and change unreachable cycles to TOP 211 // If not, we can update the input infinitely along a MergeMem cycle 212 // Equivalent code in PhiNode::Ideal 213 Node* m = phase->transform(mmem); 214 // If transformed to a MergeMem, get the desired slice 215 // Otherwise the returned node represents memory for every slice 216 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 217 // Update input if it is progress over what we have now 218 } 219 return mem; 220 } 221 222 //--------------------------Ideal_common--------------------------------------- 223 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 224 // Unhook non-raw memories from complete (macro-expanded) initializations. 225 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 226 // If our control input is a dead region, kill all below the region 227 Node *ctl = in(MemNode::Control); 228 if (ctl && remove_dead_region(phase, can_reshape)) 229 return this; 230 ctl = in(MemNode::Control); 231 // Don't bother trying to transform a dead node 232 if( ctl && ctl->is_top() ) return NodeSentinel; 233 234 PhaseIterGVN *igvn = phase->is_IterGVN(); 235 // Wait if control on the worklist. 236 if (ctl && can_reshape && igvn != NULL) { 237 Node* bol = NULL; 238 Node* cmp = NULL; 239 if (ctl->in(0)->is_If()) { 240 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 241 bol = ctl->in(0)->in(1); 242 if (bol->is_Bool()) 243 cmp = ctl->in(0)->in(1)->in(1); 244 } 245 if (igvn->_worklist.member(ctl) || 246 (bol != NULL && igvn->_worklist.member(bol)) || 247 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 248 // This control path may be dead. 249 // Delay this memory node transformation until the control is processed. 250 phase->is_IterGVN()->_worklist.push(this); 251 return NodeSentinel; // caller will return NULL 252 } 253 } 254 // Ignore if memory is dead, or self-loop 255 Node *mem = in(MemNode::Memory); 256 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 257 assert( mem != this, "dead loop in MemNode::Ideal" ); 258 259 Node *address = in(MemNode::Address); 260 const Type *t_adr = phase->type( address ); 261 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 262 263 if( can_reshape && igvn != NULL && 264 (igvn->_worklist.member(address) || phase->type(address) != adr_type()) ) { 265 // The address's base and type may change when the address is processed. 266 // Delay this mem node transformation until the address is processed. 267 phase->is_IterGVN()->_worklist.push(this); 268 return NodeSentinel; // caller will return NULL 269 } 270 271 // Do NOT remove or optimize the next lines: ensure a new alias index 272 // is allocated for an oop pointer type before Escape Analysis. 273 // Note: C++ will not remove it since the call has side effect. 274 if ( t_adr->isa_oopptr() ) { 275 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 276 } 277 278 #ifdef ASSERT 279 Node* base = NULL; 280 if (address->is_AddP()) 281 base = address->in(AddPNode::Base); 282 assert(base == NULL || t_adr->isa_rawptr() || 283 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); 284 #endif 285 286 // Avoid independent memory operations 287 Node* old_mem = mem; 288 289 // The code which unhooks non-raw memories from complete (macro-expanded) 290 // initializations was removed. After macro-expansion all stores catched 291 // by Initialize node became raw stores and there is no information 292 // which memory slices they modify. So it is unsafe to move any memory 293 // operation above these stores. Also in most cases hooked non-raw memories 294 // were already unhooked by using information from detect_ptr_independence() 295 // and find_previous_store(). 296 297 if (mem->is_MergeMem()) { 298 MergeMemNode* mmem = mem->as_MergeMem(); 299 const TypePtr *tp = t_adr->is_ptr(); 300 301 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 302 } 303 304 if (mem != old_mem) { 305 set_req(MemNode::Memory, mem); 306 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 307 return this; 308 } 309 310 // let the subclass continue analyzing... 311 return NULL; 312 } 313 314 // Helper function for proving some simple control dominations. 315 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 316 // Already assumes that 'dom' is available at 'sub', and that 'sub' 317 // is not a constant (dominated by the method's StartNode). 318 // Used by MemNode::find_previous_store to prove that the 319 // control input of a memory operation predates (dominates) 320 // an allocation it wants to look past. 321 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 322 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 323 return false; // Conservative answer for dead code 324 325 // Check 'dom'. Skip Proj and CatchProj nodes. 326 dom = dom->find_exact_control(dom); 327 if (dom == NULL || dom->is_top()) 328 return false; // Conservative answer for dead code 329 330 if (dom == sub) { 331 // For the case when, for example, 'sub' is Initialize and the original 332 // 'dom' is Proj node of the 'sub'. 333 return false; 334 } 335 336 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 337 return true; 338 339 // 'dom' dominates 'sub' if its control edge and control edges 340 // of all its inputs dominate or equal to sub's control edge. 341 342 // Currently 'sub' is either Allocate, Initialize or Start nodes. 343 // Or Region for the check in LoadNode::Ideal(); 344 // 'sub' should have sub->in(0) != NULL. 345 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 346 sub->is_Region(), "expecting only these nodes"); 347 348 // Get control edge of 'sub'. 349 Node* orig_sub = sub; 350 sub = sub->find_exact_control(sub->in(0)); 351 if (sub == NULL || sub->is_top()) 352 return false; // Conservative answer for dead code 353 354 assert(sub->is_CFG(), "expecting control"); 355 356 if (sub == dom) 357 return true; 358 359 if (sub->is_Start() || sub->is_Root()) 360 return false; 361 362 { 363 // Check all control edges of 'dom'. 364 365 ResourceMark rm; 366 Arena* arena = Thread::current()->resource_area(); 367 Node_List nlist(arena); 368 Unique_Node_List dom_list(arena); 369 370 dom_list.push(dom); 371 bool only_dominating_controls = false; 372 373 for (uint next = 0; next < dom_list.size(); next++) { 374 Node* n = dom_list.at(next); 375 if (n == orig_sub) 376 return false; // One of dom's inputs dominated by sub. 377 if (!n->is_CFG() && n->pinned()) { 378 // Check only own control edge for pinned non-control nodes. 379 n = n->find_exact_control(n->in(0)); 380 if (n == NULL || n->is_top()) 381 return false; // Conservative answer for dead code 382 assert(n->is_CFG(), "expecting control"); 383 dom_list.push(n); 384 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 385 only_dominating_controls = true; 386 } else if (n->is_CFG()) { 387 if (n->dominates(sub, nlist)) 388 only_dominating_controls = true; 389 else 390 return false; 391 } else { 392 // First, own control edge. 393 Node* m = n->find_exact_control(n->in(0)); 394 if (m != NULL) { 395 if (m->is_top()) 396 return false; // Conservative answer for dead code 397 dom_list.push(m); 398 } 399 // Now, the rest of edges. 400 uint cnt = n->req(); 401 for (uint i = 1; i < cnt; i++) { 402 m = n->find_exact_control(n->in(i)); 403 if (m == NULL || m->is_top()) 404 continue; 405 dom_list.push(m); 406 } 407 } 408 } 409 return only_dominating_controls; 410 } 411 } 412 413 //---------------------detect_ptr_independence--------------------------------- 414 // Used by MemNode::find_previous_store to prove that two base 415 // pointers are never equal. 416 // The pointers are accompanied by their associated allocations, 417 // if any, which have been previously discovered by the caller. 418 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 419 Node* p2, AllocateNode* a2, 420 PhaseTransform* phase) { 421 // Attempt to prove that these two pointers cannot be aliased. 422 // They may both manifestly be allocations, and they should differ. 423 // Or, if they are not both allocations, they can be distinct constants. 424 // Otherwise, one is an allocation and the other a pre-existing value. 425 if (a1 == NULL && a2 == NULL) { // neither an allocation 426 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 427 } else if (a1 != NULL && a2 != NULL) { // both allocations 428 return (a1 != a2); 429 } else if (a1 != NULL) { // one allocation a1 430 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 431 return all_controls_dominate(p2, a1); 432 } else { //(a2 != NULL) // one allocation a2 433 return all_controls_dominate(p1, a2); 434 } 435 return false; 436 } 437 438 439 // The logic for reordering loads and stores uses four steps: 440 // (a) Walk carefully past stores and initializations which we 441 // can prove are independent of this load. 442 // (b) Observe that the next memory state makes an exact match 443 // with self (load or store), and locate the relevant store. 444 // (c) Ensure that, if we were to wire self directly to the store, 445 // the optimizer would fold it up somehow. 446 // (d) Do the rewiring, and return, depending on some other part of 447 // the optimizer to fold up the load. 448 // This routine handles steps (a) and (b). Steps (c) and (d) are 449 // specific to loads and stores, so they are handled by the callers. 450 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 451 // 452 Node* MemNode::find_previous_store(PhaseTransform* phase) { 453 Node* ctrl = in(MemNode::Control); 454 Node* adr = in(MemNode::Address); 455 intptr_t offset = 0; 456 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 457 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 458 459 if (offset == Type::OffsetBot) 460 return NULL; // cannot unalias unless there are precise offsets 461 462 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 463 464 intptr_t size_in_bytes = memory_size(); 465 466 Node* mem = in(MemNode::Memory); // start searching here... 467 468 int cnt = 50; // Cycle limiter 469 for (;;) { // While we can dance past unrelated stores... 470 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 471 472 if (mem->is_Store()) { 473 Node* st_adr = mem->in(MemNode::Address); 474 intptr_t st_offset = 0; 475 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 476 if (st_base == NULL) 477 break; // inscrutable pointer 478 if (st_offset != offset && st_offset != Type::OffsetBot) { 479 const int MAX_STORE = BytesPerLong; 480 if (st_offset >= offset + size_in_bytes || 481 st_offset <= offset - MAX_STORE || 482 st_offset <= offset - mem->as_Store()->memory_size()) { 483 // Success: The offsets are provably independent. 484 // (You may ask, why not just test st_offset != offset and be done? 485 // The answer is that stores of different sizes can co-exist 486 // in the same sequence of RawMem effects. We sometimes initialize 487 // a whole 'tile' of array elements with a single jint or jlong.) 488 mem = mem->in(MemNode::Memory); 489 continue; // (a) advance through independent store memory 490 } 491 } 492 if (st_base != base && 493 detect_ptr_independence(base, alloc, 494 st_base, 495 AllocateNode::Ideal_allocation(st_base, phase), 496 phase)) { 497 // Success: The bases are provably independent. 498 mem = mem->in(MemNode::Memory); 499 continue; // (a) advance through independent store memory 500 } 501 502 // (b) At this point, if the bases or offsets do not agree, we lose, 503 // since we have not managed to prove 'this' and 'mem' independent. 504 if (st_base == base && st_offset == offset) { 505 return mem; // let caller handle steps (c), (d) 506 } 507 508 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 509 InitializeNode* st_init = mem->in(0)->as_Initialize(); 510 AllocateNode* st_alloc = st_init->allocation(); 511 if (st_alloc == NULL) 512 break; // something degenerated 513 bool known_identical = false; 514 bool known_independent = false; 515 if (alloc == st_alloc) 516 known_identical = true; 517 else if (alloc != NULL) 518 known_independent = true; 519 else if (all_controls_dominate(this, st_alloc)) 520 known_independent = true; 521 522 if (known_independent) { 523 // The bases are provably independent: Either they are 524 // manifestly distinct allocations, or else the control 525 // of this load dominates the store's allocation. 526 int alias_idx = phase->C->get_alias_index(adr_type()); 527 if (alias_idx == Compile::AliasIdxRaw) { 528 mem = st_alloc->in(TypeFunc::Memory); 529 } else { 530 mem = st_init->memory(alias_idx); 531 } 532 continue; // (a) advance through independent store memory 533 } 534 535 // (b) at this point, if we are not looking at a store initializing 536 // the same allocation we are loading from, we lose. 537 if (known_identical) { 538 // From caller, can_see_stored_value will consult find_captured_store. 539 return mem; // let caller handle steps (c), (d) 540 } 541 542 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 543 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 544 if (mem->is_Proj() && mem->in(0)->is_Call()) { 545 CallNode *call = mem->in(0)->as_Call(); 546 if (!call->may_modify(addr_t, phase)) { 547 mem = call->in(TypeFunc::Memory); 548 continue; // (a) advance through independent call memory 549 } 550 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 551 mem = mem->in(0)->in(TypeFunc::Memory); 552 continue; // (a) advance through independent MemBar memory 553 } else if (mem->Opcode() == Op_ClearArray) { 554 intptr_t offset; 555 AllocateNode* alloc = AllocateNode::Ideal_allocation(mem->in(3), phase, offset); 556 // We can not bypass initialization of the instance we are looking for 557 // or when something is wrong. 558 if (alloc == NULL || alloc->_idx == (uint)addr_t->instance_id()) 559 return mem; 560 // Otherwise skip it. 561 InitializeNode* init = alloc->initialization(); 562 if (init != NULL) 563 mem = init->in(TypeFunc::Memory); 564 else 565 mem = alloc->in(TypeFunc::Memory); 566 continue; // (a) advance through independent allocation memory 567 } else if (mem->is_MergeMem()) { 568 int alias_idx = phase->C->get_alias_index(adr_type()); 569 mem = mem->as_MergeMem()->memory_at(alias_idx); 570 continue; // (a) advance through independent MergeMem memory 571 } 572 } 573 574 // Unless there is an explicit 'continue', we must bail out here, 575 // because 'mem' is an inscrutable memory state (e.g., a call). 576 break; 577 } 578 579 return NULL; // bail out 580 } 581 582 //----------------------calculate_adr_type------------------------------------- 583 // Helper function. Notices when the given type of address hits top or bottom. 584 // Also, asserts a cross-check of the type against the expected address type. 585 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 586 if (t == Type::TOP) return NULL; // does not touch memory any more? 587 #ifdef PRODUCT 588 cross_check = NULL; 589 #else 590 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 591 #endif 592 const TypePtr* tp = t->isa_ptr(); 593 if (tp == NULL) { 594 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 595 return TypePtr::BOTTOM; // touches lots of memory 596 } else { 597 #ifdef ASSERT 598 // %%%% [phh] We don't check the alias index if cross_check is 599 // TypeRawPtr::BOTTOM. Needs to be investigated. 600 if (cross_check != NULL && 601 cross_check != TypePtr::BOTTOM && 602 cross_check != TypeRawPtr::BOTTOM) { 603 // Recheck the alias index, to see if it has changed (due to a bug). 604 Compile* C = Compile::current(); 605 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 606 "must stay in the original alias category"); 607 // The type of the address must be contained in the adr_type, 608 // disregarding "null"-ness. 609 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 610 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 611 assert(cross_check->meet(tp_notnull) == cross_check, 612 "real address must not escape from expected memory type"); 613 } 614 #endif 615 return tp; 616 } 617 } 618 619 //------------------------adr_phi_is_loop_invariant---------------------------- 620 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 621 // loop is loop invariant. Make a quick traversal of Phi and associated 622 // CastPP nodes, looking to see if they are a closed group within the loop. 623 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 624 // The idea is that the phi-nest must boil down to only CastPP nodes 625 // with the same data. This implies that any path into the loop already 626 // includes such a CastPP, and so the original cast, whatever its input, 627 // must be covered by an equivalent cast, with an earlier control input. 628 ResourceMark rm; 629 630 // The loop entry input of the phi should be the unique dominating 631 // node for every Phi/CastPP in the loop. 632 Unique_Node_List closure; 633 closure.push(adr_phi->in(LoopNode::EntryControl)); 634 635 // Add the phi node and the cast to the worklist. 636 Unique_Node_List worklist; 637 worklist.push(adr_phi); 638 if( cast != NULL ){ 639 if( !cast->is_ConstraintCast() ) return false; 640 worklist.push(cast); 641 } 642 643 // Begin recursive walk of phi nodes. 644 while( worklist.size() ){ 645 // Take a node off the worklist 646 Node *n = worklist.pop(); 647 if( !closure.member(n) ){ 648 // Add it to the closure. 649 closure.push(n); 650 // Make a sanity check to ensure we don't waste too much time here. 651 if( closure.size() > 20) return false; 652 // This node is OK if: 653 // - it is a cast of an identical value 654 // - or it is a phi node (then we add its inputs to the worklist) 655 // Otherwise, the node is not OK, and we presume the cast is not invariant 656 if( n->is_ConstraintCast() ){ 657 worklist.push(n->in(1)); 658 } else if( n->is_Phi() ) { 659 for( uint i = 1; i < n->req(); i++ ) { 660 worklist.push(n->in(i)); 661 } 662 } else { 663 return false; 664 } 665 } 666 } 667 668 // Quit when the worklist is empty, and we've found no offending nodes. 669 return true; 670 } 671 672 //------------------------------Ideal_DU_postCCP------------------------------- 673 // Find any cast-away of null-ness and keep its control. Null cast-aways are 674 // going away in this pass and we need to make this memory op depend on the 675 // gating null check. 676 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 677 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 678 } 679 680 // I tried to leave the CastPP's in. This makes the graph more accurate in 681 // some sense; we get to keep around the knowledge that an oop is not-null 682 // after some test. Alas, the CastPP's interfere with GVN (some values are 683 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 684 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 685 // some of the more trivial cases in the optimizer. Removing more useless 686 // Phi's started allowing Loads to illegally float above null checks. I gave 687 // up on this approach. CNC 10/20/2000 688 // This static method may be called not from MemNode (EncodePNode calls it). 689 // Only the control edge of the node 'n' might be updated. 690 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 691 Node *skipped_cast = NULL; 692 // Need a null check? Regular static accesses do not because they are 693 // from constant addresses. Array ops are gated by the range check (which 694 // always includes a NULL check). Just check field ops. 695 if( n->in(MemNode::Control) == NULL ) { 696 // Scan upwards for the highest location we can place this memory op. 697 while( true ) { 698 switch( adr->Opcode() ) { 699 700 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 701 adr = adr->in(AddPNode::Base); 702 continue; 703 704 case Op_DecodeN: // No change to NULL-ness, so peek thru 705 adr = adr->in(1); 706 continue; 707 708 case Op_CastPP: 709 // If the CastPP is useless, just peek on through it. 710 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 711 // Remember the cast that we've peeked though. If we peek 712 // through more than one, then we end up remembering the highest 713 // one, that is, if in a loop, the one closest to the top. 714 skipped_cast = adr; 715 adr = adr->in(1); 716 continue; 717 } 718 // CastPP is going away in this pass! We need this memory op to be 719 // control-dependent on the test that is guarding the CastPP. 720 ccp->hash_delete(n); 721 n->set_req(MemNode::Control, adr->in(0)); 722 ccp->hash_insert(n); 723 return n; 724 725 case Op_Phi: 726 // Attempt to float above a Phi to some dominating point. 727 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 728 // If we've already peeked through a Cast (which could have set the 729 // control), we can't float above a Phi, because the skipped Cast 730 // may not be loop invariant. 731 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 732 adr = adr->in(1); 733 continue; 734 } 735 } 736 737 // Intentional fallthrough! 738 739 // No obvious dominating point. The mem op is pinned below the Phi 740 // by the Phi itself. If the Phi goes away (no true value is merged) 741 // then the mem op can float, but not indefinitely. It must be pinned 742 // behind the controls leading to the Phi. 743 case Op_CheckCastPP: 744 // These usually stick around to change address type, however a 745 // useless one can be elided and we still need to pick up a control edge 746 if (adr->in(0) == NULL) { 747 // This CheckCastPP node has NO control and is likely useless. But we 748 // need check further up the ancestor chain for a control input to keep 749 // the node in place. 4959717. 750 skipped_cast = adr; 751 adr = adr->in(1); 752 continue; 753 } 754 ccp->hash_delete(n); 755 n->set_req(MemNode::Control, adr->in(0)); 756 ccp->hash_insert(n); 757 return n; 758 759 // List of "safe" opcodes; those that implicitly block the memory 760 // op below any null check. 761 case Op_CastX2P: // no null checks on native pointers 762 case Op_Parm: // 'this' pointer is not null 763 case Op_LoadP: // Loading from within a klass 764 case Op_LoadN: // Loading from within a klass 765 case Op_LoadKlass: // Loading from within a klass 766 case Op_LoadNKlass: // Loading from within a klass 767 case Op_ConP: // Loading from a klass 768 case Op_ConN: // Loading from a klass 769 case Op_CreateEx: // Sucking up the guts of an exception oop 770 case Op_Con: // Reading from TLS 771 case Op_CMoveP: // CMoveP is pinned 772 case Op_CMoveN: // CMoveN is pinned 773 break; // No progress 774 775 case Op_Proj: // Direct call to an allocation routine 776 case Op_SCMemProj: // Memory state from store conditional ops 777 #ifdef ASSERT 778 { 779 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 780 const Node* call = adr->in(0); 781 if (call->is_CallJava()) { 782 const CallJavaNode* call_java = call->as_CallJava(); 783 const TypeTuple *r = call_java->tf()->range(); 784 assert(r->cnt() > TypeFunc::Parms, "must return value"); 785 const Type* ret_type = r->field_at(TypeFunc::Parms); 786 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 787 // We further presume that this is one of 788 // new_instance_Java, new_array_Java, or 789 // the like, but do not assert for this. 790 } else if (call->is_Allocate()) { 791 // similar case to new_instance_Java, etc. 792 } else if (!call->is_CallLeaf()) { 793 // Projections from fetch_oop (OSR) are allowed as well. 794 ShouldNotReachHere(); 795 } 796 } 797 #endif 798 break; 799 default: 800 ShouldNotReachHere(); 801 } 802 break; 803 } 804 } 805 806 return NULL; // No progress 807 } 808 809 810 //============================================================================= 811 uint LoadNode::size_of() const { return sizeof(*this); } 812 uint LoadNode::cmp( const Node &n ) const 813 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 814 const Type *LoadNode::bottom_type() const { return _type; } 815 uint LoadNode::ideal_reg() const { 816 return Matcher::base2reg[_type->base()]; 817 } 818 819 #ifndef PRODUCT 820 void LoadNode::dump_spec(outputStream *st) const { 821 MemNode::dump_spec(st); 822 if( !Verbose && !WizardMode ) { 823 // standard dump does this in Verbose and WizardMode 824 st->print(" #"); _type->dump_on(st); 825 } 826 } 827 #endif 828 829 830 //----------------------------LoadNode::make----------------------------------- 831 // Polymorphic factory method: 832 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 833 Compile* C = gvn.C; 834 835 // sanity check the alias category against the created node type 836 assert(!(adr_type->isa_oopptr() && 837 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 838 "use LoadKlassNode instead"); 839 assert(!(adr_type->isa_aryptr() && 840 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 841 "use LoadRangeNode instead"); 842 switch (bt) { 843 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 844 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 845 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 846 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 847 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 848 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 849 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 850 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 851 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 852 case T_OBJECT: 853 #ifdef _LP64 854 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 855 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 856 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 857 } else 858 #endif 859 { 860 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 861 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 862 } 863 } 864 ShouldNotReachHere(); 865 return (LoadNode*)NULL; 866 } 867 868 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 869 bool require_atomic = true; 870 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 871 } 872 873 874 875 876 //------------------------------hash------------------------------------------- 877 uint LoadNode::hash() const { 878 // unroll addition of interesting fields 879 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 880 } 881 882 //---------------------------can_see_stored_value------------------------------ 883 // This routine exists to make sure this set of tests is done the same 884 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 885 // will change the graph shape in a way which makes memory alive twice at the 886 // same time (uses the Oracle model of aliasing), then some 887 // LoadXNode::Identity will fold things back to the equivalence-class model 888 // of aliasing. 889 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 890 Node* ld_adr = in(MemNode::Address); 891 892 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 893 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 894 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 895 atp->field() != NULL && !atp->field()->is_volatile()) { 896 uint alias_idx = atp->index(); 897 bool final = atp->field()->is_final(); 898 Node* result = NULL; 899 Node* current = st; 900 // Skip through chains of MemBarNodes checking the MergeMems for 901 // new states for the slice of this load. Stop once any other 902 // kind of node is encountered. Loads from final memory can skip 903 // through any kind of MemBar but normal loads shouldn't skip 904 // through MemBarAcquire since the could allow them to move out of 905 // a synchronized region. 906 while (current->is_Proj()) { 907 int opc = current->in(0)->Opcode(); 908 if ((final && opc == Op_MemBarAcquire) || 909 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 910 Node* mem = current->in(0)->in(TypeFunc::Memory); 911 if (mem->is_MergeMem()) { 912 MergeMemNode* merge = mem->as_MergeMem(); 913 Node* new_st = merge->memory_at(alias_idx); 914 if (new_st == merge->base_memory()) { 915 // Keep searching 916 current = merge->base_memory(); 917 continue; 918 } 919 // Save the new memory state for the slice and fall through 920 // to exit. 921 result = new_st; 922 } 923 } 924 break; 925 } 926 if (result != NULL) { 927 st = result; 928 } 929 } 930 931 932 // Loop around twice in the case Load -> Initialize -> Store. 933 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 934 for (int trip = 0; trip <= 1; trip++) { 935 936 if (st->is_Store()) { 937 Node* st_adr = st->in(MemNode::Address); 938 if (!phase->eqv(st_adr, ld_adr)) { 939 // Try harder before giving up... Match raw and non-raw pointers. 940 intptr_t st_off = 0; 941 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 942 if (alloc == NULL) return NULL; 943 intptr_t ld_off = 0; 944 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 945 if (alloc != allo2) return NULL; 946 if (ld_off != st_off) return NULL; 947 // At this point we have proven something like this setup: 948 // A = Allocate(...) 949 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 950 // S = StoreQ(, AddP(, A.Parm , #Off), V) 951 // (Actually, we haven't yet proven the Q's are the same.) 952 // In other words, we are loading from a casted version of 953 // the same pointer-and-offset that we stored to. 954 // Thus, we are able to replace L by V. 955 } 956 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 957 if (store_Opcode() != st->Opcode()) 958 return NULL; 959 return st->in(MemNode::ValueIn); 960 } 961 962 intptr_t offset = 0; // scratch 963 964 // A load from a freshly-created object always returns zero. 965 // (This can happen after LoadNode::Ideal resets the load's memory input 966 // to find_captured_store, which returned InitializeNode::zero_memory.) 967 if (st->is_Proj() && st->in(0)->is_Allocate() && 968 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 969 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 970 // return a zero value for the load's basic type 971 // (This is one of the few places where a generic PhaseTransform 972 // can create new nodes. Think of it as lazily manifesting 973 // virtually pre-existing constants.) 974 return phase->zerocon(memory_type()); 975 } 976 977 // A load from an initialization barrier can match a captured store. 978 if (st->is_Proj() && st->in(0)->is_Initialize()) { 979 InitializeNode* init = st->in(0)->as_Initialize(); 980 AllocateNode* alloc = init->allocation(); 981 if (alloc != NULL && 982 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 983 // examine a captured store value 984 st = init->find_captured_store(offset, memory_size(), phase); 985 if (st != NULL) 986 continue; // take one more trip around 987 } 988 } 989 990 break; 991 } 992 993 return NULL; 994 } 995 996 //----------------------is_instance_field_load_with_local_phi------------------ 997 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 998 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 999 in(MemNode::Address)->is_AddP() ) { 1000 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 1001 // Only instances. 1002 if( t_oop != NULL && t_oop->is_known_instance_field() && 1003 t_oop->offset() != Type::OffsetBot && 1004 t_oop->offset() != Type::OffsetTop) { 1005 return true; 1006 } 1007 } 1008 return false; 1009 } 1010 1011 //------------------------------Identity--------------------------------------- 1012 // Loads are identity if previous store is to same address 1013 Node *LoadNode::Identity( PhaseTransform *phase ) { 1014 // If the previous store-maker is the right kind of Store, and the store is 1015 // to the same address, then we are equal to the value stored. 1016 Node* mem = in(MemNode::Memory); 1017 Node* value = can_see_stored_value(mem, phase); 1018 if( value ) { 1019 // byte, short & char stores truncate naturally. 1020 // A load has to load the truncated value which requires 1021 // some sort of masking operation and that requires an 1022 // Ideal call instead of an Identity call. 1023 if (memory_size() < BytesPerInt) { 1024 // If the input to the store does not fit with the load's result type, 1025 // it must be truncated via an Ideal call. 1026 if (!phase->type(value)->higher_equal(phase->type(this))) 1027 return this; 1028 } 1029 // (This works even when value is a Con, but LoadNode::Value 1030 // usually runs first, producing the singleton type of the Con.) 1031 return value; 1032 } 1033 1034 // Search for an existing data phi which was generated before for the same 1035 // instance's field to avoid infinite generation of phis in a loop. 1036 Node *region = mem->in(0); 1037 if (is_instance_field_load_with_local_phi(region)) { 1038 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 1039 int this_index = phase->C->get_alias_index(addr_t); 1040 int this_offset = addr_t->offset(); 1041 int this_id = addr_t->is_oopptr()->instance_id(); 1042 const Type* this_type = bottom_type(); 1043 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1044 Node* phi = region->fast_out(i); 1045 if (phi->is_Phi() && phi != mem && 1046 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 1047 return phi; 1048 } 1049 } 1050 } 1051 1052 return this; 1053 } 1054 1055 1056 // Returns true if the AliasType refers to the field that holds the 1057 // cached box array. Currently only handles the IntegerCache case. 1058 static bool is_autobox_cache(Compile::AliasType* atp) { 1059 if (atp != NULL && atp->field() != NULL) { 1060 ciField* field = atp->field(); 1061 ciSymbol* klass = field->holder()->name(); 1062 if (field->name() == ciSymbol::cache_field_name() && 1063 field->holder()->uses_default_loader() && 1064 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1065 return true; 1066 } 1067 } 1068 return false; 1069 } 1070 1071 // Fetch the base value in the autobox array 1072 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1073 if (atp != NULL && atp->field() != NULL) { 1074 ciField* field = atp->field(); 1075 ciSymbol* klass = field->holder()->name(); 1076 if (field->name() == ciSymbol::cache_field_name() && 1077 field->holder()->uses_default_loader() && 1078 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1079 assert(field->is_constant(), "what?"); 1080 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1081 // Fetch the box object at the base of the array and get its value 1082 ciInstance* box = array->obj_at(0)->as_instance(); 1083 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1084 if (ik->nof_nonstatic_fields() == 1) { 1085 // This should be true nonstatic_field_at requires calling 1086 // nof_nonstatic_fields so check it anyway 1087 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1088 cache_offset = c.as_int(); 1089 } 1090 return true; 1091 } 1092 } 1093 return false; 1094 } 1095 1096 // Returns true if the AliasType refers to the value field of an 1097 // autobox object. Currently only handles Integer. 1098 static bool is_autobox_object(Compile::AliasType* atp) { 1099 if (atp != NULL && atp->field() != NULL) { 1100 ciField* field = atp->field(); 1101 ciSymbol* klass = field->holder()->name(); 1102 if (field->name() == ciSymbol::value_name() && 1103 field->holder()->uses_default_loader() && 1104 klass == ciSymbol::java_lang_Integer()) { 1105 return true; 1106 } 1107 } 1108 return false; 1109 } 1110 1111 1112 // We're loading from an object which has autobox behaviour. 1113 // If this object is result of a valueOf call we'll have a phi 1114 // merging a newly allocated object and a load from the cache. 1115 // We want to replace this load with the original incoming 1116 // argument to the valueOf call. 1117 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1118 Node* base = in(Address)->in(AddPNode::Base); 1119 if (base->is_Phi() && base->req() == 3) { 1120 AllocateNode* allocation = NULL; 1121 int allocation_index = -1; 1122 int load_index = -1; 1123 for (uint i = 1; i < base->req(); i++) { 1124 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1125 if (allocation != NULL) { 1126 allocation_index = i; 1127 load_index = 3 - allocation_index; 1128 break; 1129 } 1130 } 1131 bool has_load = ( allocation != NULL && 1132 (base->in(load_index)->is_Load() || 1133 base->in(load_index)->is_DecodeN() && 1134 base->in(load_index)->in(1)->is_Load()) ); 1135 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1136 // Push the loads from the phi that comes from valueOf up 1137 // through it to allow elimination of the loads and the recovery 1138 // of the original value. 1139 Node* mem_phi = in(Memory); 1140 Node* offset = in(Address)->in(AddPNode::Offset); 1141 Node* region = base->in(0); 1142 1143 Node* in1 = clone(); 1144 Node* in1_addr = in1->in(Address)->clone(); 1145 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1146 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1147 in1_addr->set_req(AddPNode::Offset, offset); 1148 in1->set_req(0, region->in(allocation_index)); 1149 in1->set_req(Address, in1_addr); 1150 in1->set_req(Memory, mem_phi->in(allocation_index)); 1151 1152 Node* in2 = clone(); 1153 Node* in2_addr = in2->in(Address)->clone(); 1154 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1155 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1156 in2_addr->set_req(AddPNode::Offset, offset); 1157 in2->set_req(0, region->in(load_index)); 1158 in2->set_req(Address, in2_addr); 1159 in2->set_req(Memory, mem_phi->in(load_index)); 1160 1161 in1_addr = phase->transform(in1_addr); 1162 in1 = phase->transform(in1); 1163 in2_addr = phase->transform(in2_addr); 1164 in2 = phase->transform(in2); 1165 1166 PhiNode* result = PhiNode::make_blank(region, this); 1167 result->set_req(allocation_index, in1); 1168 result->set_req(load_index, in2); 1169 return result; 1170 } 1171 } else if (base->is_Load() || 1172 base->is_DecodeN() && base->in(1)->is_Load()) { 1173 if (base->is_DecodeN()) { 1174 // Get LoadN node which loads cached Integer object 1175 base = base->in(1); 1176 } 1177 // Eliminate the load of Integer.value for integers from the cache 1178 // array by deriving the value from the index into the array. 1179 // Capture the offset of the load and then reverse the computation. 1180 Node* load_base = base->in(Address)->in(AddPNode::Base); 1181 if (load_base->is_DecodeN()) { 1182 // Get LoadN node which loads IntegerCache.cache field 1183 load_base = load_base->in(1); 1184 } 1185 if (load_base != NULL) { 1186 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1187 intptr_t cache_offset; 1188 int shift = -1; 1189 Node* cache = NULL; 1190 if (is_autobox_cache(atp)) { 1191 shift = exact_log2(type2aelembytes(T_OBJECT)); 1192 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1193 } 1194 if (cache != NULL && base->in(Address)->is_AddP()) { 1195 Node* elements[4]; 1196 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1197 int cache_low; 1198 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1199 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1200 // Add up all the offsets making of the address of the load 1201 Node* result = elements[0]; 1202 for (int i = 1; i < count; i++) { 1203 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1204 } 1205 // Remove the constant offset from the address and then 1206 // remove the scaling of the offset to recover the original index. 1207 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1208 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1209 // Peel the shift off directly but wrap it in a dummy node 1210 // since Ideal can't return existing nodes 1211 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1212 } else { 1213 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1214 } 1215 #ifdef _LP64 1216 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1217 #endif 1218 return result; 1219 } 1220 } 1221 } 1222 } 1223 return NULL; 1224 } 1225 1226 //------------------------------split_through_phi------------------------------ 1227 // Split instance field load through Phi. 1228 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1229 Node* mem = in(MemNode::Memory); 1230 Node* address = in(MemNode::Address); 1231 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1232 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1233 1234 assert(mem->is_Phi() && (t_oop != NULL) && 1235 t_oop->is_known_instance_field(), "invalide conditions"); 1236 1237 Node *region = mem->in(0); 1238 if (region == NULL) { 1239 return NULL; // Wait stable graph 1240 } 1241 uint cnt = mem->req(); 1242 for( uint i = 1; i < cnt; i++ ) { 1243 Node *in = mem->in(i); 1244 if( in == NULL ) { 1245 return NULL; // Wait stable graph 1246 } 1247 } 1248 // Check for loop invariant. 1249 if (cnt == 3) { 1250 for( uint i = 1; i < cnt; i++ ) { 1251 Node *in = mem->in(i); 1252 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1253 if (m == mem) { 1254 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1255 return this; 1256 } 1257 } 1258 } 1259 // Split through Phi (see original code in loopopts.cpp). 1260 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1261 1262 // Do nothing here if Identity will find a value 1263 // (to avoid infinite chain of value phis generation). 1264 if ( !phase->eqv(this, this->Identity(phase)) ) 1265 return NULL; 1266 1267 // Skip the split if the region dominates some control edge of the address. 1268 if (cnt == 3 && !MemNode::all_controls_dominate(address, region)) 1269 return NULL; 1270 1271 const Type* this_type = this->bottom_type(); 1272 int this_index = phase->C->get_alias_index(addr_t); 1273 int this_offset = addr_t->offset(); 1274 int this_iid = addr_t->is_oopptr()->instance_id(); 1275 int wins = 0; 1276 PhaseIterGVN *igvn = phase->is_IterGVN(); 1277 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1278 for( uint i = 1; i < region->req(); i++ ) { 1279 Node *x; 1280 Node* the_clone = NULL; 1281 if( region->in(i) == phase->C->top() ) { 1282 x = phase->C->top(); // Dead path? Use a dead data op 1283 } else { 1284 x = this->clone(); // Else clone up the data op 1285 the_clone = x; // Remember for possible deletion. 1286 // Alter data node to use pre-phi inputs 1287 if( this->in(0) == region ) { 1288 x->set_req( 0, region->in(i) ); 1289 } else { 1290 x->set_req( 0, NULL ); 1291 } 1292 for( uint j = 1; j < this->req(); j++ ) { 1293 Node *in = this->in(j); 1294 if( in->is_Phi() && in->in(0) == region ) 1295 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 1296 } 1297 } 1298 // Check for a 'win' on some paths 1299 const Type *t = x->Value(igvn); 1300 1301 bool singleton = t->singleton(); 1302 1303 // See comments in PhaseIdealLoop::split_thru_phi(). 1304 if( singleton && t == Type::TOP ) { 1305 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1306 } 1307 1308 if( singleton ) { 1309 wins++; 1310 x = igvn->makecon(t); 1311 } else { 1312 // We now call Identity to try to simplify the cloned node. 1313 // Note that some Identity methods call phase->type(this). 1314 // Make sure that the type array is big enough for 1315 // our new node, even though we may throw the node away. 1316 // (This tweaking with igvn only works because x is a new node.) 1317 igvn->set_type(x, t); 1318 // If x is a TypeNode, capture any more-precise type permanently into Node 1319 // otherwise it will be not updated during igvn->transform since 1320 // igvn->type(x) is set to x->Value() already. 1321 x->raise_bottom_type(t); 1322 Node *y = x->Identity(igvn); 1323 if( y != x ) { 1324 wins++; 1325 x = y; 1326 } else { 1327 y = igvn->hash_find(x); 1328 if( y ) { 1329 wins++; 1330 x = y; 1331 } else { 1332 // Else x is a new node we are keeping 1333 // We do not need register_new_node_with_optimizer 1334 // because set_type has already been called. 1335 igvn->_worklist.push(x); 1336 } 1337 } 1338 } 1339 if (x != the_clone && the_clone != NULL) 1340 igvn->remove_dead_node(the_clone); 1341 phi->set_req(i, x); 1342 } 1343 if( wins > 0 ) { 1344 // Record Phi 1345 igvn->register_new_node_with_optimizer(phi); 1346 return phi; 1347 } 1348 igvn->remove_dead_node(phi); 1349 return NULL; 1350 } 1351 1352 //------------------------------Ideal------------------------------------------ 1353 // If the load is from Field memory and the pointer is non-null, we can 1354 // zero out the control input. 1355 // If the offset is constant and the base is an object allocation, 1356 // try to hook me up to the exact initializing store. 1357 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1358 Node* p = MemNode::Ideal_common(phase, can_reshape); 1359 if (p) return (p == NodeSentinel) ? NULL : p; 1360 1361 Node* ctrl = in(MemNode::Control); 1362 Node* address = in(MemNode::Address); 1363 1364 // Skip up past a SafePoint control. Cannot do this for Stores because 1365 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1366 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1367 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1368 ctrl = ctrl->in(0); 1369 set_req(MemNode::Control,ctrl); 1370 } 1371 1372 intptr_t ignore = 0; 1373 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1374 if (base != NULL 1375 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1376 // Check for useless control edge in some common special cases 1377 if (in(MemNode::Control) != NULL 1378 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1379 && all_controls_dominate(base, phase->C->start())) { 1380 // A method-invariant, non-null address (constant or 'this' argument). 1381 set_req(MemNode::Control, NULL); 1382 } 1383 1384 if (EliminateAutoBox && can_reshape) { 1385 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); 1386 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1387 if (is_autobox_object(atp)) { 1388 Node* result = eliminate_autobox(phase); 1389 if (result != NULL) return result; 1390 } 1391 } 1392 } 1393 1394 Node* mem = in(MemNode::Memory); 1395 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1396 1397 if (addr_t != NULL) { 1398 // try to optimize our memory input 1399 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1400 if (opt_mem != mem) { 1401 set_req(MemNode::Memory, opt_mem); 1402 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1403 return this; 1404 } 1405 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1406 if (can_reshape && opt_mem->is_Phi() && 1407 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1408 // Split instance field load through Phi. 1409 Node* result = split_through_phi(phase); 1410 if (result != NULL) return result; 1411 } 1412 } 1413 1414 // Check for prior store with a different base or offset; make Load 1415 // independent. Skip through any number of them. Bail out if the stores 1416 // are in an endless dead cycle and report no progress. This is a key 1417 // transform for Reflection. However, if after skipping through the Stores 1418 // we can't then fold up against a prior store do NOT do the transform as 1419 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1420 // array memory alive twice: once for the hoisted Load and again after the 1421 // bypassed Store. This situation only works if EVERYBODY who does 1422 // anti-dependence work knows how to bypass. I.e. we need all 1423 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1424 // the alias index stuff. So instead, peek through Stores and IFF we can 1425 // fold up, do so. 1426 Node* prev_mem = find_previous_store(phase); 1427 // Steps (a), (b): Walk past independent stores to find an exact match. 1428 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1429 // (c) See if we can fold up on the spot, but don't fold up here. 1430 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1431 // just return a prior value, which is done by Identity calls. 1432 if (can_see_stored_value(prev_mem, phase)) { 1433 // Make ready for step (d): 1434 set_req(MemNode::Memory, prev_mem); 1435 return this; 1436 } 1437 } 1438 1439 return NULL; // No further progress 1440 } 1441 1442 // Helper to recognize certain Klass fields which are invariant across 1443 // some group of array types (e.g., int[] or all T[] where T < Object). 1444 const Type* 1445 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1446 ciKlass* klass) const { 1447 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1448 // The field is Klass::_modifier_flags. Return its (constant) value. 1449 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1450 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1451 return TypeInt::make(klass->modifier_flags()); 1452 } 1453 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1454 // The field is Klass::_access_flags. Return its (constant) value. 1455 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1456 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1457 return TypeInt::make(klass->access_flags()); 1458 } 1459 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1460 // The field is Klass::_layout_helper. Return its constant value if known. 1461 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1462 return TypeInt::make(klass->layout_helper()); 1463 } 1464 1465 // No match. 1466 return NULL; 1467 } 1468 1469 //------------------------------Value----------------------------------------- 1470 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1471 // Either input is TOP ==> the result is TOP 1472 Node* mem = in(MemNode::Memory); 1473 const Type *t1 = phase->type(mem); 1474 if (t1 == Type::TOP) return Type::TOP; 1475 Node* adr = in(MemNode::Address); 1476 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1477 if (tp == NULL || tp->empty()) return Type::TOP; 1478 int off = tp->offset(); 1479 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1480 1481 // Try to guess loaded type from pointer type 1482 if (tp->base() == Type::AryPtr) { 1483 const Type *t = tp->is_aryptr()->elem(); 1484 // Don't do this for integer types. There is only potential profit if 1485 // the element type t is lower than _type; that is, for int types, if _type is 1486 // more restrictive than t. This only happens here if one is short and the other 1487 // char (both 16 bits), and in those cases we've made an intentional decision 1488 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1489 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1490 // 1491 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1492 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1493 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1494 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1495 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1496 // In fact, that could have been the original type of p1, and p1 could have 1497 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1498 // expression (LShiftL quux 3) independently optimized to the constant 8. 1499 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1500 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1501 // t might actually be lower than _type, if _type is a unique 1502 // concrete subclass of abstract class t. 1503 // Make sure the reference is not into the header, by comparing 1504 // the offset against the offset of the start of the array's data. 1505 // Different array types begin at slightly different offsets (12 vs. 16). 1506 // We choose T_BYTE as an example base type that is least restrictive 1507 // as to alignment, which will therefore produce the smallest 1508 // possible base offset. 1509 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1510 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1511 const Type* jt = t->join(_type); 1512 // In any case, do not allow the join, per se, to empty out the type. 1513 if (jt->empty() && !t->empty()) { 1514 // This can happen if a interface-typed array narrows to a class type. 1515 jt = _type; 1516 } 1517 1518 if (EliminateAutoBox && adr->is_AddP()) { 1519 // The pointers in the autobox arrays are always non-null 1520 Node* base = adr->in(AddPNode::Base); 1521 if (base != NULL && 1522 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { 1523 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1524 if (is_autobox_cache(atp)) { 1525 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1526 } 1527 } 1528 } 1529 return jt; 1530 } 1531 } 1532 } else if (tp->base() == Type::InstPtr) { 1533 assert( off != Type::OffsetBot || 1534 // arrays can be cast to Objects 1535 tp->is_oopptr()->klass()->is_java_lang_Object() || 1536 // unsafe field access may not have a constant offset 1537 phase->C->has_unsafe_access(), 1538 "Field accesses must be precise" ); 1539 // For oop loads, we expect the _type to be precise 1540 } else if (tp->base() == Type::KlassPtr) { 1541 assert( off != Type::OffsetBot || 1542 // arrays can be cast to Objects 1543 tp->is_klassptr()->klass()->is_java_lang_Object() || 1544 // also allow array-loading from the primary supertype 1545 // array during subtype checks 1546 Opcode() == Op_LoadKlass, 1547 "Field accesses must be precise" ); 1548 // For klass/static loads, we expect the _type to be precise 1549 } 1550 1551 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1552 if (tkls != NULL && !StressReflectiveCode) { 1553 ciKlass* klass = tkls->klass(); 1554 if (klass->is_loaded() && tkls->klass_is_exact()) { 1555 // We are loading a field from a Klass metaobject whose identity 1556 // is known at compile time (the type is "exact" or "precise"). 1557 // Check for fields we know are maintained as constants by the VM. 1558 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1559 // The field is Klass::_super_check_offset. Return its (constant) value. 1560 // (Folds up type checking code.) 1561 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1562 return TypeInt::make(klass->super_check_offset()); 1563 } 1564 // Compute index into primary_supers array 1565 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1566 // Check for overflowing; use unsigned compare to handle the negative case. 1567 if( depth < ciKlass::primary_super_limit() ) { 1568 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1569 // (Folds up type checking code.) 1570 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1571 ciKlass *ss = klass->super_of_depth(depth); 1572 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1573 } 1574 const Type* aift = load_array_final_field(tkls, klass); 1575 if (aift != NULL) return aift; 1576 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1577 && klass->is_array_klass()) { 1578 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1579 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1580 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1581 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1582 } 1583 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1584 // The field is Klass::_java_mirror. Return its (constant) value. 1585 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1586 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1587 return TypeInstPtr::make(klass->java_mirror()); 1588 } 1589 } 1590 1591 // We can still check if we are loading from the primary_supers array at a 1592 // shallow enough depth. Even though the klass is not exact, entries less 1593 // than or equal to its super depth are correct. 1594 if (klass->is_loaded() ) { 1595 ciType *inner = klass->klass(); 1596 while( inner->is_obj_array_klass() ) 1597 inner = inner->as_obj_array_klass()->base_element_type(); 1598 if( inner->is_instance_klass() && 1599 !inner->as_instance_klass()->flags().is_interface() ) { 1600 // Compute index into primary_supers array 1601 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1602 // Check for overflowing; use unsigned compare to handle the negative case. 1603 if( depth < ciKlass::primary_super_limit() && 1604 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1605 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1606 // (Folds up type checking code.) 1607 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1608 ciKlass *ss = klass->super_of_depth(depth); 1609 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1610 } 1611 } 1612 } 1613 1614 // If the type is enough to determine that the thing is not an array, 1615 // we can give the layout_helper a positive interval type. 1616 // This will help short-circuit some reflective code. 1617 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1618 && !klass->is_array_klass() // not directly typed as an array 1619 && !klass->is_interface() // specifically not Serializable & Cloneable 1620 && !klass->is_java_lang_Object() // not the supertype of all T[] 1621 ) { 1622 // Note: When interfaces are reliable, we can narrow the interface 1623 // test to (klass != Serializable && klass != Cloneable). 1624 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1625 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1626 // The key property of this type is that it folds up tests 1627 // for array-ness, since it proves that the layout_helper is positive. 1628 // Thus, a generic value like the basic object layout helper works fine. 1629 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1630 } 1631 } 1632 1633 // If we are loading from a freshly-allocated object, produce a zero, 1634 // if the load is provably beyond the header of the object. 1635 // (Also allow a variable load from a fresh array to produce zero.) 1636 if (ReduceFieldZeroing) { 1637 Node* value = can_see_stored_value(mem,phase); 1638 if (value != NULL && value->is_Con()) 1639 return value->bottom_type(); 1640 } 1641 1642 const TypeOopPtr *tinst = tp->isa_oopptr(); 1643 if (tinst != NULL && tinst->is_known_instance_field()) { 1644 // If we have an instance type and our memory input is the 1645 // programs's initial memory state, there is no matching store, 1646 // so just return a zero of the appropriate type 1647 Node *mem = in(MemNode::Memory); 1648 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1649 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1650 return Type::get_zero_type(_type->basic_type()); 1651 } 1652 } 1653 return _type; 1654 } 1655 1656 //------------------------------match_edge------------------------------------- 1657 // Do we Match on this edge index or not? Match only the address. 1658 uint LoadNode::match_edge(uint idx) const { 1659 return idx == MemNode::Address; 1660 } 1661 1662 //--------------------------LoadBNode::Ideal-------------------------------------- 1663 // 1664 // If the previous store is to the same address as this load, 1665 // and the value stored was larger than a byte, replace this load 1666 // with the value stored truncated to a byte. If no truncation is 1667 // needed, the replacement is done in LoadNode::Identity(). 1668 // 1669 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1670 Node* mem = in(MemNode::Memory); 1671 Node* value = can_see_stored_value(mem,phase); 1672 if( value && !phase->type(value)->higher_equal( _type ) ) { 1673 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1674 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1675 } 1676 // Identity call will handle the case where truncation is not needed. 1677 return LoadNode::Ideal(phase, can_reshape); 1678 } 1679 1680 //--------------------------LoadUBNode::Ideal------------------------------------- 1681 // 1682 // If the previous store is to the same address as this load, 1683 // and the value stored was larger than a byte, replace this load 1684 // with the value stored truncated to a byte. If no truncation is 1685 // needed, the replacement is done in LoadNode::Identity(). 1686 // 1687 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1688 Node* mem = in(MemNode::Memory); 1689 Node* value = can_see_stored_value(mem, phase); 1690 if (value && !phase->type(value)->higher_equal(_type)) 1691 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); 1692 // Identity call will handle the case where truncation is not needed. 1693 return LoadNode::Ideal(phase, can_reshape); 1694 } 1695 1696 //--------------------------LoadUSNode::Ideal------------------------------------- 1697 // 1698 // If the previous store is to the same address as this load, 1699 // and the value stored was larger than a char, replace this load 1700 // with the value stored truncated to a char. If no truncation is 1701 // needed, the replacement is done in LoadNode::Identity(). 1702 // 1703 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1704 Node* mem = in(MemNode::Memory); 1705 Node* value = can_see_stored_value(mem,phase); 1706 if( value && !phase->type(value)->higher_equal( _type ) ) 1707 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1708 // Identity call will handle the case where truncation is not needed. 1709 return LoadNode::Ideal(phase, can_reshape); 1710 } 1711 1712 //--------------------------LoadSNode::Ideal-------------------------------------- 1713 // 1714 // If the previous store is to the same address as this load, 1715 // and the value stored was larger than a short, replace this load 1716 // with the value stored truncated to a short. If no truncation is 1717 // needed, the replacement is done in LoadNode::Identity(). 1718 // 1719 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1720 Node* mem = in(MemNode::Memory); 1721 Node* value = can_see_stored_value(mem,phase); 1722 if( value && !phase->type(value)->higher_equal( _type ) ) { 1723 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1724 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1725 } 1726 // Identity call will handle the case where truncation is not needed. 1727 return LoadNode::Ideal(phase, can_reshape); 1728 } 1729 1730 //============================================================================= 1731 //----------------------------LoadKlassNode::make------------------------------ 1732 // Polymorphic factory method: 1733 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1734 Compile* C = gvn.C; 1735 Node *ctl = NULL; 1736 // sanity check the alias category against the created node type 1737 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1738 assert(adr_type != NULL, "expecting TypeOopPtr"); 1739 #ifdef _LP64 1740 if (adr_type->is_ptr_to_narrowoop()) { 1741 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1742 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1743 } 1744 #endif 1745 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1746 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1747 } 1748 1749 //------------------------------Value------------------------------------------ 1750 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1751 return klass_value_common(phase); 1752 } 1753 1754 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1755 // Either input is TOP ==> the result is TOP 1756 const Type *t1 = phase->type( in(MemNode::Memory) ); 1757 if (t1 == Type::TOP) return Type::TOP; 1758 Node *adr = in(MemNode::Address); 1759 const Type *t2 = phase->type( adr ); 1760 if (t2 == Type::TOP) return Type::TOP; 1761 const TypePtr *tp = t2->is_ptr(); 1762 if (TypePtr::above_centerline(tp->ptr()) || 1763 tp->ptr() == TypePtr::Null) return Type::TOP; 1764 1765 // Return a more precise klass, if possible 1766 const TypeInstPtr *tinst = tp->isa_instptr(); 1767 if (tinst != NULL) { 1768 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1769 int offset = tinst->offset(); 1770 if (ik == phase->C->env()->Class_klass() 1771 && (offset == java_lang_Class::klass_offset_in_bytes() || 1772 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1773 // We are loading a special hidden field from a Class mirror object, 1774 // the field which points to the VM's Klass metaobject. 1775 ciType* t = tinst->java_mirror_type(); 1776 // java_mirror_type returns non-null for compile-time Class constants. 1777 if (t != NULL) { 1778 // constant oop => constant klass 1779 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1780 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1781 } 1782 if (!t->is_klass()) { 1783 // a primitive Class (e.g., int.class) has NULL for a klass field 1784 return TypePtr::NULL_PTR; 1785 } 1786 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1787 return TypeKlassPtr::make(t->as_klass()); 1788 } 1789 // non-constant mirror, so we can't tell what's going on 1790 } 1791 if( !ik->is_loaded() ) 1792 return _type; // Bail out if not loaded 1793 if (offset == oopDesc::klass_offset_in_bytes()) { 1794 if (tinst->klass_is_exact()) { 1795 return TypeKlassPtr::make(ik); 1796 } 1797 // See if we can become precise: no subklasses and no interface 1798 // (Note: We need to support verified interfaces.) 1799 if (!ik->is_interface() && !ik->has_subklass()) { 1800 //assert(!UseExactTypes, "this code should be useless with exact types"); 1801 // Add a dependence; if any subclass added we need to recompile 1802 if (!ik->is_final()) { 1803 // %%% should use stronger assert_unique_concrete_subtype instead 1804 phase->C->dependencies()->assert_leaf_type(ik); 1805 } 1806 // Return precise klass 1807 return TypeKlassPtr::make(ik); 1808 } 1809 1810 // Return root of possible klass 1811 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1812 } 1813 } 1814 1815 // Check for loading klass from an array 1816 const TypeAryPtr *tary = tp->isa_aryptr(); 1817 if( tary != NULL ) { 1818 ciKlass *tary_klass = tary->klass(); 1819 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1820 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1821 if (tary->klass_is_exact()) { 1822 return TypeKlassPtr::make(tary_klass); 1823 } 1824 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1825 // If the klass is an object array, we defer the question to the 1826 // array component klass. 1827 if( ak->is_obj_array_klass() ) { 1828 assert( ak->is_loaded(), "" ); 1829 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1830 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1831 ciInstanceKlass* ik = base_k->as_instance_klass(); 1832 // See if we can become precise: no subklasses and no interface 1833 if (!ik->is_interface() && !ik->has_subklass()) { 1834 //assert(!UseExactTypes, "this code should be useless with exact types"); 1835 // Add a dependence; if any subclass added we need to recompile 1836 if (!ik->is_final()) { 1837 phase->C->dependencies()->assert_leaf_type(ik); 1838 } 1839 // Return precise array klass 1840 return TypeKlassPtr::make(ak); 1841 } 1842 } 1843 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1844 } else { // Found a type-array? 1845 //assert(!UseExactTypes, "this code should be useless with exact types"); 1846 assert( ak->is_type_array_klass(), "" ); 1847 return TypeKlassPtr::make(ak); // These are always precise 1848 } 1849 } 1850 } 1851 1852 // Check for loading klass from an array klass 1853 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1854 if (tkls != NULL && !StressReflectiveCode) { 1855 ciKlass* klass = tkls->klass(); 1856 if( !klass->is_loaded() ) 1857 return _type; // Bail out if not loaded 1858 if( klass->is_obj_array_klass() && 1859 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1860 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1861 // // Always returning precise element type is incorrect, 1862 // // e.g., element type could be object and array may contain strings 1863 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1864 1865 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1866 // according to the element type's subclassing. 1867 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1868 } 1869 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1870 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1871 ciKlass* sup = klass->as_instance_klass()->super(); 1872 // The field is Klass::_super. Return its (constant) value. 1873 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1874 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1875 } 1876 } 1877 1878 // Bailout case 1879 return LoadNode::Value(phase); 1880 } 1881 1882 //------------------------------Identity--------------------------------------- 1883 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1884 // Also feed through the klass in Allocate(...klass...)._klass. 1885 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1886 return klass_identity_common(phase); 1887 } 1888 1889 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 1890 Node* x = LoadNode::Identity(phase); 1891 if (x != this) return x; 1892 1893 // Take apart the address into an oop and and offset. 1894 // Return 'this' if we cannot. 1895 Node* adr = in(MemNode::Address); 1896 intptr_t offset = 0; 1897 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1898 if (base == NULL) return this; 1899 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1900 if (toop == NULL) return this; 1901 1902 // We can fetch the klass directly through an AllocateNode. 1903 // This works even if the klass is not constant (clone or newArray). 1904 if (offset == oopDesc::klass_offset_in_bytes()) { 1905 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1906 if (allocated_klass != NULL) { 1907 return allocated_klass; 1908 } 1909 } 1910 1911 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1912 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1913 // See inline_native_Class_query for occurrences of these patterns. 1914 // Java Example: x.getClass().isAssignableFrom(y) 1915 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1916 // 1917 // This improves reflective code, often making the Class 1918 // mirror go completely dead. (Current exception: Class 1919 // mirrors may appear in debug info, but we could clean them out by 1920 // introducing a new debug info operator for klassOop.java_mirror). 1921 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1922 && (offset == java_lang_Class::klass_offset_in_bytes() || 1923 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1924 // We are loading a special hidden field from a Class mirror, 1925 // the field which points to its Klass or arrayKlass metaobject. 1926 if (base->is_Load()) { 1927 Node* adr2 = base->in(MemNode::Address); 1928 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1929 if (tkls != NULL && !tkls->empty() 1930 && (tkls->klass()->is_instance_klass() || 1931 tkls->klass()->is_array_klass()) 1932 && adr2->is_AddP() 1933 ) { 1934 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1935 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1936 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1937 } 1938 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1939 return adr2->in(AddPNode::Base); 1940 } 1941 } 1942 } 1943 } 1944 1945 return this; 1946 } 1947 1948 1949 //------------------------------Value------------------------------------------ 1950 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 1951 const Type *t = klass_value_common(phase); 1952 if (t == Type::TOP) 1953 return t; 1954 1955 return t->make_narrowoop(); 1956 } 1957 1958 //------------------------------Identity--------------------------------------- 1959 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 1960 // Also feed through the klass in Allocate(...klass...)._klass. 1961 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 1962 Node *x = klass_identity_common(phase); 1963 1964 const Type *t = phase->type( x ); 1965 if( t == Type::TOP ) return x; 1966 if( t->isa_narrowoop()) return x; 1967 1968 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 1969 } 1970 1971 //------------------------------Value----------------------------------------- 1972 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1973 // Either input is TOP ==> the result is TOP 1974 const Type *t1 = phase->type( in(MemNode::Memory) ); 1975 if( t1 == Type::TOP ) return Type::TOP; 1976 Node *adr = in(MemNode::Address); 1977 const Type *t2 = phase->type( adr ); 1978 if( t2 == Type::TOP ) return Type::TOP; 1979 const TypePtr *tp = t2->is_ptr(); 1980 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 1981 const TypeAryPtr *tap = tp->isa_aryptr(); 1982 if( !tap ) return _type; 1983 return tap->size(); 1984 } 1985 1986 //-------------------------------Ideal--------------------------------------- 1987 // Feed through the length in AllocateArray(...length...)._length. 1988 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1989 Node* p = MemNode::Ideal_common(phase, can_reshape); 1990 if (p) return (p == NodeSentinel) ? NULL : p; 1991 1992 // Take apart the address into an oop and and offset. 1993 // Return 'this' if we cannot. 1994 Node* adr = in(MemNode::Address); 1995 intptr_t offset = 0; 1996 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1997 if (base == NULL) return NULL; 1998 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1999 if (tary == NULL) return NULL; 2000 2001 // We can fetch the length directly through an AllocateArrayNode. 2002 // This works even if the length is not constant (clone or newArray). 2003 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2004 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2005 if (alloc != NULL) { 2006 Node* allocated_length = alloc->Ideal_length(); 2007 Node* len = alloc->make_ideal_length(tary, phase); 2008 if (allocated_length != len) { 2009 // New CastII improves on this. 2010 return len; 2011 } 2012 } 2013 } 2014 2015 return NULL; 2016 } 2017 2018 //------------------------------Identity--------------------------------------- 2019 // Feed through the length in AllocateArray(...length...)._length. 2020 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2021 Node* x = LoadINode::Identity(phase); 2022 if (x != this) return x; 2023 2024 // Take apart the address into an oop and and offset. 2025 // Return 'this' if we cannot. 2026 Node* adr = in(MemNode::Address); 2027 intptr_t offset = 0; 2028 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2029 if (base == NULL) return this; 2030 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2031 if (tary == NULL) return this; 2032 2033 // We can fetch the length directly through an AllocateArrayNode. 2034 // This works even if the length is not constant (clone or newArray). 2035 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2036 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2037 if (alloc != NULL) { 2038 Node* allocated_length = alloc->Ideal_length(); 2039 // Do not allow make_ideal_length to allocate a CastII node. 2040 Node* len = alloc->make_ideal_length(tary, phase, false); 2041 if (allocated_length == len) { 2042 // Return allocated_length only if it would not be improved by a CastII. 2043 return allocated_length; 2044 } 2045 } 2046 } 2047 2048 return this; 2049 2050 } 2051 2052 //============================================================================= 2053 //---------------------------StoreNode::make----------------------------------- 2054 // Polymorphic factory method: 2055 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 2056 Compile* C = gvn.C; 2057 2058 switch (bt) { 2059 case T_BOOLEAN: 2060 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 2061 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 2062 case T_CHAR: 2063 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 2064 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 2065 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 2066 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 2067 case T_ADDRESS: 2068 case T_OBJECT: 2069 #ifdef _LP64 2070 if (adr->bottom_type()->is_ptr_to_narrowoop() || 2071 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 2072 adr->bottom_type()->isa_rawptr())) { 2073 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2074 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 2075 } else 2076 #endif 2077 { 2078 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 2079 } 2080 } 2081 ShouldNotReachHere(); 2082 return (StoreNode*)NULL; 2083 } 2084 2085 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2086 bool require_atomic = true; 2087 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2088 } 2089 2090 2091 //--------------------------bottom_type---------------------------------------- 2092 const Type *StoreNode::bottom_type() const { 2093 return Type::MEMORY; 2094 } 2095 2096 //------------------------------hash------------------------------------------- 2097 uint StoreNode::hash() const { 2098 // unroll addition of interesting fields 2099 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2100 2101 // Since they are not commoned, do not hash them: 2102 return NO_HASH; 2103 } 2104 2105 //------------------------------Ideal------------------------------------------ 2106 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2107 // When a store immediately follows a relevant allocation/initialization, 2108 // try to capture it into the initialization, or hoist it above. 2109 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2110 Node* p = MemNode::Ideal_common(phase, can_reshape); 2111 if (p) return (p == NodeSentinel) ? NULL : p; 2112 2113 Node* mem = in(MemNode::Memory); 2114 Node* address = in(MemNode::Address); 2115 2116 // Back-to-back stores to same address? Fold em up. 2117 // Generally unsafe if I have intervening uses... 2118 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 2119 // Looking at a dead closed cycle of memory? 2120 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2121 2122 assert(Opcode() == mem->Opcode() || 2123 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2124 "no mismatched stores, except on raw memory"); 2125 2126 if (mem->outcnt() == 1 && // check for intervening uses 2127 mem->as_Store()->memory_size() <= this->memory_size()) { 2128 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2129 // For example, 'mem' might be the final state at a conditional return. 2130 // Or, 'mem' might be used by some node which is live at the same time 2131 // 'this' is live, which might be unschedulable. So, require exactly 2132 // ONE user, the 'this' store, until such time as we clone 'mem' for 2133 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2134 if (can_reshape) { // (%%% is this an anachronism?) 2135 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2136 phase->is_IterGVN()); 2137 } else { 2138 // It's OK to do this in the parser, since DU info is always accurate, 2139 // and the parser always refers to nodes via SafePointNode maps. 2140 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2141 } 2142 return this; 2143 } 2144 } 2145 2146 // Capture an unaliased, unconditional, simple store into an initializer. 2147 // Or, if it is independent of the allocation, hoist it above the allocation. 2148 if (ReduceFieldZeroing && /*can_reshape &&*/ 2149 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2150 InitializeNode* init = mem->in(0)->as_Initialize(); 2151 intptr_t offset = init->can_capture_store(this, phase); 2152 if (offset > 0) { 2153 Node* moved = init->capture_store(this, offset, phase); 2154 // If the InitializeNode captured me, it made a raw copy of me, 2155 // and I need to disappear. 2156 if (moved != NULL) { 2157 // %%% hack to ensure that Ideal returns a new node: 2158 mem = MergeMemNode::make(phase->C, mem); 2159 return mem; // fold me away 2160 } 2161 } 2162 } 2163 2164 return NULL; // No further progress 2165 } 2166 2167 //------------------------------Value----------------------------------------- 2168 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2169 // Either input is TOP ==> the result is TOP 2170 const Type *t1 = phase->type( in(MemNode::Memory) ); 2171 if( t1 == Type::TOP ) return Type::TOP; 2172 const Type *t2 = phase->type( in(MemNode::Address) ); 2173 if( t2 == Type::TOP ) return Type::TOP; 2174 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2175 if( t3 == Type::TOP ) return Type::TOP; 2176 return Type::MEMORY; 2177 } 2178 2179 //------------------------------Identity--------------------------------------- 2180 // Remove redundant stores: 2181 // Store(m, p, Load(m, p)) changes to m. 2182 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2183 Node *StoreNode::Identity( PhaseTransform *phase ) { 2184 Node* mem = in(MemNode::Memory); 2185 Node* adr = in(MemNode::Address); 2186 Node* val = in(MemNode::ValueIn); 2187 2188 // Load then Store? Then the Store is useless 2189 if (val->is_Load() && 2190 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 2191 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 2192 val->as_Load()->store_Opcode() == Opcode()) { 2193 return mem; 2194 } 2195 2196 // Two stores in a row of the same value? 2197 if (mem->is_Store() && 2198 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 2199 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 2200 mem->Opcode() == Opcode()) { 2201 return mem; 2202 } 2203 2204 // Store of zero anywhere into a freshly-allocated object? 2205 // Then the store is useless. 2206 // (It must already have been captured by the InitializeNode.) 2207 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2208 // a newly allocated object is already all-zeroes everywhere 2209 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2210 return mem; 2211 } 2212 2213 // the store may also apply to zero-bits in an earlier object 2214 Node* prev_mem = find_previous_store(phase); 2215 // Steps (a), (b): Walk past independent stores to find an exact match. 2216 if (prev_mem != NULL) { 2217 Node* prev_val = can_see_stored_value(prev_mem, phase); 2218 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2219 // prev_val and val might differ by a cast; it would be good 2220 // to keep the more informative of the two. 2221 return mem; 2222 } 2223 } 2224 } 2225 2226 return this; 2227 } 2228 2229 //------------------------------match_edge------------------------------------- 2230 // Do we Match on this edge index or not? Match only memory & value 2231 uint StoreNode::match_edge(uint idx) const { 2232 return idx == MemNode::Address || idx == MemNode::ValueIn; 2233 } 2234 2235 //------------------------------cmp-------------------------------------------- 2236 // Do not common stores up together. They generally have to be split 2237 // back up anyways, so do not bother. 2238 uint StoreNode::cmp( const Node &n ) const { 2239 return (&n == this); // Always fail except on self 2240 } 2241 2242 //------------------------------Ideal_masked_input----------------------------- 2243 // Check for a useless mask before a partial-word store 2244 // (StoreB ... (AndI valIn conIa) ) 2245 // If (conIa & mask == mask) this simplifies to 2246 // (StoreB ... (valIn) ) 2247 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2248 Node *val = in(MemNode::ValueIn); 2249 if( val->Opcode() == Op_AndI ) { 2250 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2251 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2252 set_req(MemNode::ValueIn, val->in(1)); 2253 return this; 2254 } 2255 } 2256 return NULL; 2257 } 2258 2259 2260 //------------------------------Ideal_sign_extended_input---------------------- 2261 // Check for useless sign-extension before a partial-word store 2262 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2263 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2264 // (StoreB ... (valIn) ) 2265 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2266 Node *val = in(MemNode::ValueIn); 2267 if( val->Opcode() == Op_RShiftI ) { 2268 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2269 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2270 Node *shl = val->in(1); 2271 if( shl->Opcode() == Op_LShiftI ) { 2272 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2273 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2274 set_req(MemNode::ValueIn, shl->in(1)); 2275 return this; 2276 } 2277 } 2278 } 2279 } 2280 return NULL; 2281 } 2282 2283 //------------------------------value_never_loaded----------------------------------- 2284 // Determine whether there are any possible loads of the value stored. 2285 // For simplicity, we actually check if there are any loads from the 2286 // address stored to, not just for loads of the value stored by this node. 2287 // 2288 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2289 Node *adr = in(Address); 2290 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2291 if (adr_oop == NULL) 2292 return false; 2293 if (!adr_oop->is_known_instance_field()) 2294 return false; // if not a distinct instance, there may be aliases of the address 2295 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2296 Node *use = adr->fast_out(i); 2297 int opc = use->Opcode(); 2298 if (use->is_Load() || use->is_LoadStore()) { 2299 return false; 2300 } 2301 } 2302 return true; 2303 } 2304 2305 //============================================================================= 2306 //------------------------------Ideal------------------------------------------ 2307 // If the store is from an AND mask that leaves the low bits untouched, then 2308 // we can skip the AND operation. If the store is from a sign-extension 2309 // (a left shift, then right shift) we can skip both. 2310 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2311 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2312 if( progress != NULL ) return progress; 2313 2314 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2315 if( progress != NULL ) return progress; 2316 2317 // Finally check the default case 2318 return StoreNode::Ideal(phase, can_reshape); 2319 } 2320 2321 //============================================================================= 2322 //------------------------------Ideal------------------------------------------ 2323 // If the store is from an AND mask that leaves the low bits untouched, then 2324 // we can skip the AND operation 2325 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2326 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2327 if( progress != NULL ) return progress; 2328 2329 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2330 if( progress != NULL ) return progress; 2331 2332 // Finally check the default case 2333 return StoreNode::Ideal(phase, can_reshape); 2334 } 2335 2336 //============================================================================= 2337 //------------------------------Identity--------------------------------------- 2338 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2339 // No need to card mark when storing a null ptr 2340 Node* my_store = in(MemNode::OopStore); 2341 if (my_store->is_Store()) { 2342 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2343 if( t1 == TypePtr::NULL_PTR ) { 2344 return in(MemNode::Memory); 2345 } 2346 } 2347 return this; 2348 } 2349 2350 //============================================================================= 2351 //------------------------------Ideal--------------------------------------- 2352 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2353 Node* progress = StoreNode::Ideal(phase, can_reshape); 2354 if (progress != NULL) return progress; 2355 2356 Node* my_store = in(MemNode::OopStore); 2357 if (my_store->is_MergeMem()) { 2358 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2359 set_req(MemNode::OopStore, mem); 2360 return this; 2361 } 2362 2363 return NULL; 2364 } 2365 2366 //------------------------------Value----------------------------------------- 2367 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2368 // Either input is TOP ==> the result is TOP 2369 const Type *t = phase->type( in(MemNode::Memory) ); 2370 if( t == Type::TOP ) return Type::TOP; 2371 t = phase->type( in(MemNode::Address) ); 2372 if( t == Type::TOP ) return Type::TOP; 2373 t = phase->type( in(MemNode::ValueIn) ); 2374 if( t == Type::TOP ) return Type::TOP; 2375 // If extra input is TOP ==> the result is TOP 2376 t = phase->type( in(MemNode::OopStore) ); 2377 if( t == Type::TOP ) return Type::TOP; 2378 2379 return StoreNode::Value( phase ); 2380 } 2381 2382 2383 //============================================================================= 2384 //----------------------------------SCMemProjNode------------------------------ 2385 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2386 { 2387 return bottom_type(); 2388 } 2389 2390 //============================================================================= 2391 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2392 init_req(MemNode::Control, c ); 2393 init_req(MemNode::Memory , mem); 2394 init_req(MemNode::Address, adr); 2395 init_req(MemNode::ValueIn, val); 2396 init_req( ExpectedIn, ex ); 2397 init_class_id(Class_LoadStore); 2398 2399 } 2400 2401 //============================================================================= 2402 //-------------------------------adr_type-------------------------------------- 2403 // Do we Match on this edge index or not? Do not match memory 2404 const TypePtr* ClearArrayNode::adr_type() const { 2405 Node *adr = in(3); 2406 return MemNode::calculate_adr_type(adr->bottom_type()); 2407 } 2408 2409 //------------------------------match_edge------------------------------------- 2410 // Do we Match on this edge index or not? Do not match memory 2411 uint ClearArrayNode::match_edge(uint idx) const { 2412 return idx > 1; 2413 } 2414 2415 //------------------------------Identity--------------------------------------- 2416 // Clearing a zero length array does nothing 2417 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2418 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2419 } 2420 2421 //------------------------------Idealize--------------------------------------- 2422 // Clearing a short array is faster with stores 2423 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2424 const int unit = BytesPerLong; 2425 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2426 if (!t) return NULL; 2427 if (!t->is_con()) return NULL; 2428 intptr_t raw_count = t->get_con(); 2429 intptr_t size = raw_count; 2430 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2431 // Clearing nothing uses the Identity call. 2432 // Negative clears are possible on dead ClearArrays 2433 // (see jck test stmt114.stmt11402.val). 2434 if (size <= 0 || size % unit != 0) return NULL; 2435 intptr_t count = size / unit; 2436 // Length too long; use fast hardware clear 2437 if (size > Matcher::init_array_short_size) return NULL; 2438 Node *mem = in(1); 2439 if( phase->type(mem)==Type::TOP ) return NULL; 2440 Node *adr = in(3); 2441 const Type* at = phase->type(adr); 2442 if( at==Type::TOP ) return NULL; 2443 const TypePtr* atp = at->isa_ptr(); 2444 // adjust atp to be the correct array element address type 2445 if (atp == NULL) atp = TypePtr::BOTTOM; 2446 else atp = atp->add_offset(Type::OffsetBot); 2447 // Get base for derived pointer purposes 2448 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2449 Node *base = adr->in(1); 2450 2451 Node *zero = phase->makecon(TypeLong::ZERO); 2452 Node *off = phase->MakeConX(BytesPerLong); 2453 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2454 count--; 2455 while( count-- ) { 2456 mem = phase->transform(mem); 2457 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2458 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2459 } 2460 return mem; 2461 } 2462 2463 //----------------------------clear_memory------------------------------------- 2464 // Generate code to initialize object storage to zero. 2465 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2466 intptr_t start_offset, 2467 Node* end_offset, 2468 PhaseGVN* phase) { 2469 Compile* C = phase->C; 2470 intptr_t offset = start_offset; 2471 2472 int unit = BytesPerLong; 2473 if ((offset % unit) != 0) { 2474 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2475 adr = phase->transform(adr); 2476 const TypePtr* atp = TypeRawPtr::BOTTOM; 2477 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2478 mem = phase->transform(mem); 2479 offset += BytesPerInt; 2480 } 2481 assert((offset % unit) == 0, ""); 2482 2483 // Initialize the remaining stuff, if any, with a ClearArray. 2484 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2485 } 2486 2487 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2488 Node* start_offset, 2489 Node* end_offset, 2490 PhaseGVN* phase) { 2491 if (start_offset == end_offset) { 2492 // nothing to do 2493 return mem; 2494 } 2495 2496 Compile* C = phase->C; 2497 int unit = BytesPerLong; 2498 Node* zbase = start_offset; 2499 Node* zend = end_offset; 2500 2501 // Scale to the unit required by the CPU: 2502 if (!Matcher::init_array_count_is_in_bytes) { 2503 Node* shift = phase->intcon(exact_log2(unit)); 2504 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2505 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2506 } 2507 2508 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2509 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2510 2511 // Bulk clear double-words 2512 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2513 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2514 return phase->transform(mem); 2515 } 2516 2517 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2518 intptr_t start_offset, 2519 intptr_t end_offset, 2520 PhaseGVN* phase) { 2521 if (start_offset == end_offset) { 2522 // nothing to do 2523 return mem; 2524 } 2525 2526 Compile* C = phase->C; 2527 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2528 intptr_t done_offset = end_offset; 2529 if ((done_offset % BytesPerLong) != 0) { 2530 done_offset -= BytesPerInt; 2531 } 2532 if (done_offset > start_offset) { 2533 mem = clear_memory(ctl, mem, dest, 2534 start_offset, phase->MakeConX(done_offset), phase); 2535 } 2536 if (done_offset < end_offset) { // emit the final 32-bit store 2537 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2538 adr = phase->transform(adr); 2539 const TypePtr* atp = TypeRawPtr::BOTTOM; 2540 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2541 mem = phase->transform(mem); 2542 done_offset += BytesPerInt; 2543 } 2544 assert(done_offset == end_offset, ""); 2545 return mem; 2546 } 2547 2548 //============================================================================= 2549 // Do we match on this edge? No memory edges 2550 uint StrCompNode::match_edge(uint idx) const { 2551 return idx == 2 || idx == 3; // StrComp (Binary str1 cnt1) (Binary str2 cnt2) 2552 } 2553 2554 //------------------------------Ideal------------------------------------------ 2555 // Return a node which is more "ideal" than the current node. Strip out 2556 // control copies 2557 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2558 return remove_dead_region(phase, can_reshape) ? this : NULL; 2559 } 2560 2561 //============================================================================= 2562 // Do we match on this edge? No memory edges 2563 uint StrEqualsNode::match_edge(uint idx) const { 2564 return idx == 2 || idx == 3; // StrEquals (Binary str1 str2) cnt 2565 } 2566 2567 //------------------------------Ideal------------------------------------------ 2568 // Return a node which is more "ideal" than the current node. Strip out 2569 // control copies 2570 Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2571 return remove_dead_region(phase, can_reshape) ? this : NULL; 2572 } 2573 2574 //============================================================================= 2575 // Do we match on this edge? No memory edges 2576 uint StrIndexOfNode::match_edge(uint idx) const { 2577 return idx == 2 || idx == 3; // StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2) 2578 } 2579 2580 //------------------------------Ideal------------------------------------------ 2581 // Return a node which is more "ideal" than the current node. Strip out 2582 // control copies 2583 Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2584 return remove_dead_region(phase, can_reshape) ? this : NULL; 2585 } 2586 2587 //============================================================================= 2588 // Do we match on this edge? No memory edges 2589 uint AryEqNode::match_edge(uint idx) const { 2590 return idx == 2 || idx == 3; // StrEquals ary1 ary2 2591 } 2592 //------------------------------Ideal------------------------------------------ 2593 // Return a node which is more "ideal" than the current node. Strip out 2594 // control copies 2595 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2596 return remove_dead_region(phase, can_reshape) ? this : NULL; 2597 } 2598 2599 //============================================================================= 2600 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2601 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2602 _adr_type(C->get_adr_type(alias_idx)) 2603 { 2604 init_class_id(Class_MemBar); 2605 Node* top = C->top(); 2606 init_req(TypeFunc::I_O,top); 2607 init_req(TypeFunc::FramePtr,top); 2608 init_req(TypeFunc::ReturnAdr,top); 2609 if (precedent != NULL) 2610 init_req(TypeFunc::Parms, precedent); 2611 } 2612 2613 //------------------------------cmp-------------------------------------------- 2614 uint MemBarNode::hash() const { return NO_HASH; } 2615 uint MemBarNode::cmp( const Node &n ) const { 2616 return (&n == this); // Always fail except on self 2617 } 2618 2619 //------------------------------make------------------------------------------- 2620 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2621 int len = Precedent + (pn == NULL? 0: 1); 2622 switch (opcode) { 2623 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2624 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2625 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2626 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2627 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2628 default: ShouldNotReachHere(); return NULL; 2629 } 2630 } 2631 2632 //------------------------------Ideal------------------------------------------ 2633 // Return a node which is more "ideal" than the current node. Strip out 2634 // control copies 2635 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2636 if (remove_dead_region(phase, can_reshape)) return this; 2637 2638 // Eliminate volatile MemBars for scalar replaced objects. 2639 if (can_reshape && req() == (Precedent+1) && 2640 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { 2641 // Volatile field loads and stores. 2642 Node* my_mem = in(MemBarNode::Precedent); 2643 if (my_mem != NULL && my_mem->is_Mem()) { 2644 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2645 // Check for scalar replaced object reference. 2646 if( t_oop != NULL && t_oop->is_known_instance_field() && 2647 t_oop->offset() != Type::OffsetBot && 2648 t_oop->offset() != Type::OffsetTop) { 2649 // Replace MemBar projections by its inputs. 2650 PhaseIterGVN* igvn = phase->is_IterGVN(); 2651 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2652 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2653 // Must return either the original node (now dead) or a new node 2654 // (Do not return a top here, since that would break the uniqueness of top.) 2655 return new (phase->C, 1) ConINode(TypeInt::ZERO); 2656 } 2657 } 2658 } 2659 return NULL; 2660 } 2661 2662 //------------------------------Value------------------------------------------ 2663 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2664 if( !in(0) ) return Type::TOP; 2665 if( phase->type(in(0)) == Type::TOP ) 2666 return Type::TOP; 2667 return TypeTuple::MEMBAR; 2668 } 2669 2670 //------------------------------match------------------------------------------ 2671 // Construct projections for memory. 2672 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2673 switch (proj->_con) { 2674 case TypeFunc::Control: 2675 case TypeFunc::Memory: 2676 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2677 } 2678 ShouldNotReachHere(); 2679 return NULL; 2680 } 2681 2682 //===========================InitializeNode==================================== 2683 // SUMMARY: 2684 // This node acts as a memory barrier on raw memory, after some raw stores. 2685 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2686 // The Initialize can 'capture' suitably constrained stores as raw inits. 2687 // It can coalesce related raw stores into larger units (called 'tiles'). 2688 // It can avoid zeroing new storage for memory units which have raw inits. 2689 // At macro-expansion, it is marked 'complete', and does not optimize further. 2690 // 2691 // EXAMPLE: 2692 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2693 // ctl = incoming control; mem* = incoming memory 2694 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2695 // First allocate uninitialized memory and fill in the header: 2696 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2697 // ctl := alloc.Control; mem* := alloc.Memory* 2698 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2699 // Then initialize to zero the non-header parts of the raw memory block: 2700 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2701 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2702 // After the initialize node executes, the object is ready for service: 2703 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2704 // Suppose its body is immediately initialized as {1,2}: 2705 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2706 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2707 // mem.SLICE(#short[*]) := store2 2708 // 2709 // DETAILS: 2710 // An InitializeNode collects and isolates object initialization after 2711 // an AllocateNode and before the next possible safepoint. As a 2712 // memory barrier (MemBarNode), it keeps critical stores from drifting 2713 // down past any safepoint or any publication of the allocation. 2714 // Before this barrier, a newly-allocated object may have uninitialized bits. 2715 // After this barrier, it may be treated as a real oop, and GC is allowed. 2716 // 2717 // The semantics of the InitializeNode include an implicit zeroing of 2718 // the new object from object header to the end of the object. 2719 // (The object header and end are determined by the AllocateNode.) 2720 // 2721 // Certain stores may be added as direct inputs to the InitializeNode. 2722 // These stores must update raw memory, and they must be to addresses 2723 // derived from the raw address produced by AllocateNode, and with 2724 // a constant offset. They must be ordered by increasing offset. 2725 // The first one is at in(RawStores), the last at in(req()-1). 2726 // Unlike most memory operations, they are not linked in a chain, 2727 // but are displayed in parallel as users of the rawmem output of 2728 // the allocation. 2729 // 2730 // (See comments in InitializeNode::capture_store, which continue 2731 // the example given above.) 2732 // 2733 // When the associated Allocate is macro-expanded, the InitializeNode 2734 // may be rewritten to optimize collected stores. A ClearArrayNode 2735 // may also be created at that point to represent any required zeroing. 2736 // The InitializeNode is then marked 'complete', prohibiting further 2737 // capturing of nearby memory operations. 2738 // 2739 // During macro-expansion, all captured initializations which store 2740 // constant values of 32 bits or smaller are coalesced (if advantageous) 2741 // into larger 'tiles' 32 or 64 bits. This allows an object to be 2742 // initialized in fewer memory operations. Memory words which are 2743 // covered by neither tiles nor non-constant stores are pre-zeroed 2744 // by explicit stores of zero. (The code shape happens to do all 2745 // zeroing first, then all other stores, with both sequences occurring 2746 // in order of ascending offsets.) 2747 // 2748 // Alternatively, code may be inserted between an AllocateNode and its 2749 // InitializeNode, to perform arbitrary initialization of the new object. 2750 // E.g., the object copying intrinsics insert complex data transfers here. 2751 // The initialization must then be marked as 'complete' disable the 2752 // built-in zeroing semantics and the collection of initializing stores. 2753 // 2754 // While an InitializeNode is incomplete, reads from the memory state 2755 // produced by it are optimizable if they match the control edge and 2756 // new oop address associated with the allocation/initialization. 2757 // They return a stored value (if the offset matches) or else zero. 2758 // A write to the memory state, if it matches control and address, 2759 // and if it is to a constant offset, may be 'captured' by the 2760 // InitializeNode. It is cloned as a raw memory operation and rewired 2761 // inside the initialization, to the raw oop produced by the allocation. 2762 // Operations on addresses which are provably distinct (e.g., to 2763 // other AllocateNodes) are allowed to bypass the initialization. 2764 // 2765 // The effect of all this is to consolidate object initialization 2766 // (both arrays and non-arrays, both piecewise and bulk) into a 2767 // single location, where it can be optimized as a unit. 2768 // 2769 // Only stores with an offset less than TrackedInitializationLimit words 2770 // will be considered for capture by an InitializeNode. This puts a 2771 // reasonable limit on the complexity of optimized initializations. 2772 2773 //---------------------------InitializeNode------------------------------------ 2774 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2775 : _is_complete(false), 2776 MemBarNode(C, adr_type, rawoop) 2777 { 2778 init_class_id(Class_Initialize); 2779 2780 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2781 assert(in(RawAddress) == rawoop, "proper init"); 2782 // Note: allocation() can be NULL, for secondary initialization barriers 2783 } 2784 2785 // Since this node is not matched, it will be processed by the 2786 // register allocator. Declare that there are no constraints 2787 // on the allocation of the RawAddress edge. 2788 const RegMask &InitializeNode::in_RegMask(uint idx) const { 2789 // This edge should be set to top, by the set_complete. But be conservative. 2790 if (idx == InitializeNode::RawAddress) 2791 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2792 return RegMask::Empty; 2793 } 2794 2795 Node* InitializeNode::memory(uint alias_idx) { 2796 Node* mem = in(Memory); 2797 if (mem->is_MergeMem()) { 2798 return mem->as_MergeMem()->memory_at(alias_idx); 2799 } else { 2800 // incoming raw memory is not split 2801 return mem; 2802 } 2803 } 2804 2805 bool InitializeNode::is_non_zero() { 2806 if (is_complete()) return false; 2807 remove_extra_zeroes(); 2808 return (req() > RawStores); 2809 } 2810 2811 void InitializeNode::set_complete(PhaseGVN* phase) { 2812 assert(!is_complete(), "caller responsibility"); 2813 _is_complete = true; 2814 2815 // After this node is complete, it contains a bunch of 2816 // raw-memory initializations. There is no need for 2817 // it to have anything to do with non-raw memory effects. 2818 // Therefore, tell all non-raw users to re-optimize themselves, 2819 // after skipping the memory effects of this initialization. 2820 PhaseIterGVN* igvn = phase->is_IterGVN(); 2821 if (igvn) igvn->add_users_to_worklist(this); 2822 } 2823 2824 // convenience function 2825 // return false if the init contains any stores already 2826 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2827 InitializeNode* init = initialization(); 2828 if (init == NULL || init->is_complete()) return false; 2829 init->remove_extra_zeroes(); 2830 // for now, if this allocation has already collected any inits, bail: 2831 if (init->is_non_zero()) return false; 2832 init->set_complete(phase); 2833 return true; 2834 } 2835 2836 void InitializeNode::remove_extra_zeroes() { 2837 if (req() == RawStores) return; 2838 Node* zmem = zero_memory(); 2839 uint fill = RawStores; 2840 for (uint i = fill; i < req(); i++) { 2841 Node* n = in(i); 2842 if (n->is_top() || n == zmem) continue; // skip 2843 if (fill < i) set_req(fill, n); // compact 2844 ++fill; 2845 } 2846 // delete any empty spaces created: 2847 while (fill < req()) { 2848 del_req(fill); 2849 } 2850 } 2851 2852 // Helper for remembering which stores go with which offsets. 2853 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2854 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2855 intptr_t offset = -1; 2856 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2857 phase, offset); 2858 if (base == NULL) return -1; // something is dead, 2859 if (offset < 0) return -1; // dead, dead 2860 return offset; 2861 } 2862 2863 // Helper for proving that an initialization expression is 2864 // "simple enough" to be folded into an object initialization. 2865 // Attempts to prove that a store's initial value 'n' can be captured 2866 // within the initialization without creating a vicious cycle, such as: 2867 // { Foo p = new Foo(); p.next = p; } 2868 // True for constants and parameters and small combinations thereof. 2869 bool InitializeNode::detect_init_independence(Node* n, 2870 bool st_is_pinned, 2871 int& count) { 2872 if (n == NULL) return true; // (can this really happen?) 2873 if (n->is_Proj()) n = n->in(0); 2874 if (n == this) return false; // found a cycle 2875 if (n->is_Con()) return true; 2876 if (n->is_Start()) return true; // params, etc., are OK 2877 if (n->is_Root()) return true; // even better 2878 2879 Node* ctl = n->in(0); 2880 if (ctl != NULL && !ctl->is_top()) { 2881 if (ctl->is_Proj()) ctl = ctl->in(0); 2882 if (ctl == this) return false; 2883 2884 // If we already know that the enclosing memory op is pinned right after 2885 // the init, then any control flow that the store has picked up 2886 // must have preceded the init, or else be equal to the init. 2887 // Even after loop optimizations (which might change control edges) 2888 // a store is never pinned *before* the availability of its inputs. 2889 if (!MemNode::all_controls_dominate(n, this)) 2890 return false; // failed to prove a good control 2891 2892 } 2893 2894 // Check data edges for possible dependencies on 'this'. 2895 if ((count += 1) > 20) return false; // complexity limit 2896 for (uint i = 1; i < n->req(); i++) { 2897 Node* m = n->in(i); 2898 if (m == NULL || m == n || m->is_top()) continue; 2899 uint first_i = n->find_edge(m); 2900 if (i != first_i) continue; // process duplicate edge just once 2901 if (!detect_init_independence(m, st_is_pinned, count)) { 2902 return false; 2903 } 2904 } 2905 2906 return true; 2907 } 2908 2909 // Here are all the checks a Store must pass before it can be moved into 2910 // an initialization. Returns zero if a check fails. 2911 // On success, returns the (constant) offset to which the store applies, 2912 // within the initialized memory. 2913 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2914 const int FAIL = 0; 2915 if (st->req() != MemNode::ValueIn + 1) 2916 return FAIL; // an inscrutable StoreNode (card mark?) 2917 Node* ctl = st->in(MemNode::Control); 2918 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2919 return FAIL; // must be unconditional after the initialization 2920 Node* mem = st->in(MemNode::Memory); 2921 if (!(mem->is_Proj() && mem->in(0) == this)) 2922 return FAIL; // must not be preceded by other stores 2923 Node* adr = st->in(MemNode::Address); 2924 intptr_t offset; 2925 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2926 if (alloc == NULL) 2927 return FAIL; // inscrutable address 2928 if (alloc != allocation()) 2929 return FAIL; // wrong allocation! (store needs to float up) 2930 Node* val = st->in(MemNode::ValueIn); 2931 int complexity_count = 0; 2932 if (!detect_init_independence(val, true, complexity_count)) 2933 return FAIL; // stored value must be 'simple enough' 2934 2935 return offset; // success 2936 } 2937 2938 // Find the captured store in(i) which corresponds to the range 2939 // [start..start+size) in the initialized object. 2940 // If there is one, return its index i. If there isn't, return the 2941 // negative of the index where it should be inserted. 2942 // Return 0 if the queried range overlaps an initialization boundary 2943 // or if dead code is encountered. 2944 // If size_in_bytes is zero, do not bother with overlap checks. 2945 int InitializeNode::captured_store_insertion_point(intptr_t start, 2946 int size_in_bytes, 2947 PhaseTransform* phase) { 2948 const int FAIL = 0, MAX_STORE = BytesPerLong; 2949 2950 if (is_complete()) 2951 return FAIL; // arraycopy got here first; punt 2952 2953 assert(allocation() != NULL, "must be present"); 2954 2955 // no negatives, no header fields: 2956 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 2957 2958 // after a certain size, we bail out on tracking all the stores: 2959 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2960 if (start >= ti_limit) return FAIL; 2961 2962 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 2963 if (i >= limit) return -(int)i; // not found; here is where to put it 2964 2965 Node* st = in(i); 2966 intptr_t st_off = get_store_offset(st, phase); 2967 if (st_off < 0) { 2968 if (st != zero_memory()) { 2969 return FAIL; // bail out if there is dead garbage 2970 } 2971 } else if (st_off > start) { 2972 // ...we are done, since stores are ordered 2973 if (st_off < start + size_in_bytes) { 2974 return FAIL; // the next store overlaps 2975 } 2976 return -(int)i; // not found; here is where to put it 2977 } else if (st_off < start) { 2978 if (size_in_bytes != 0 && 2979 start < st_off + MAX_STORE && 2980 start < st_off + st->as_Store()->memory_size()) { 2981 return FAIL; // the previous store overlaps 2982 } 2983 } else { 2984 if (size_in_bytes != 0 && 2985 st->as_Store()->memory_size() != size_in_bytes) { 2986 return FAIL; // mismatched store size 2987 } 2988 return i; 2989 } 2990 2991 ++i; 2992 } 2993 } 2994 2995 // Look for a captured store which initializes at the offset 'start' 2996 // with the given size. If there is no such store, and no other 2997 // initialization interferes, then return zero_memory (the memory 2998 // projection of the AllocateNode). 2999 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3000 PhaseTransform* phase) { 3001 assert(stores_are_sane(phase), ""); 3002 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3003 if (i == 0) { 3004 return NULL; // something is dead 3005 } else if (i < 0) { 3006 return zero_memory(); // just primordial zero bits here 3007 } else { 3008 Node* st = in(i); // here is the store at this position 3009 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3010 return st; 3011 } 3012 } 3013 3014 // Create, as a raw pointer, an address within my new object at 'offset'. 3015 Node* InitializeNode::make_raw_address(intptr_t offset, 3016 PhaseTransform* phase) { 3017 Node* addr = in(RawAddress); 3018 if (offset != 0) { 3019 Compile* C = phase->C; 3020 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 3021 phase->MakeConX(offset)) ); 3022 } 3023 return addr; 3024 } 3025 3026 // Clone the given store, converting it into a raw store 3027 // initializing a field or element of my new object. 3028 // Caller is responsible for retiring the original store, 3029 // with subsume_node or the like. 3030 // 3031 // From the example above InitializeNode::InitializeNode, 3032 // here are the old stores to be captured: 3033 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3034 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3035 // 3036 // Here is the changed code; note the extra edges on init: 3037 // alloc = (Allocate ...) 3038 // rawoop = alloc.RawAddress 3039 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3040 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3041 // init = (Initialize alloc.Control alloc.Memory rawoop 3042 // rawstore1 rawstore2) 3043 // 3044 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3045 PhaseTransform* phase) { 3046 assert(stores_are_sane(phase), ""); 3047 3048 if (start < 0) return NULL; 3049 assert(can_capture_store(st, phase) == start, "sanity"); 3050 3051 Compile* C = phase->C; 3052 int size_in_bytes = st->memory_size(); 3053 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3054 if (i == 0) return NULL; // bail out 3055 Node* prev_mem = NULL; // raw memory for the captured store 3056 if (i > 0) { 3057 prev_mem = in(i); // there is a pre-existing store under this one 3058 set_req(i, C->top()); // temporarily disconnect it 3059 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3060 } else { 3061 i = -i; // no pre-existing store 3062 prev_mem = zero_memory(); // a slice of the newly allocated object 3063 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3064 set_req(--i, C->top()); // reuse this edge; it has been folded away 3065 else 3066 ins_req(i, C->top()); // build a new edge 3067 } 3068 Node* new_st = st->clone(); 3069 new_st->set_req(MemNode::Control, in(Control)); 3070 new_st->set_req(MemNode::Memory, prev_mem); 3071 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3072 new_st = phase->transform(new_st); 3073 3074 // At this point, new_st might have swallowed a pre-existing store 3075 // at the same offset, or perhaps new_st might have disappeared, 3076 // if it redundantly stored the same value (or zero to fresh memory). 3077 3078 // In any case, wire it in: 3079 set_req(i, new_st); 3080 3081 // The caller may now kill the old guy. 3082 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3083 assert(check_st == new_st || check_st == NULL, "must be findable"); 3084 assert(!is_complete(), ""); 3085 return new_st; 3086 } 3087 3088 static bool store_constant(jlong* tiles, int num_tiles, 3089 intptr_t st_off, int st_size, 3090 jlong con) { 3091 if ((st_off & (st_size-1)) != 0) 3092 return false; // strange store offset (assume size==2**N) 3093 address addr = (address)tiles + st_off; 3094 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3095 switch (st_size) { 3096 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3097 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3098 case sizeof(jint): *(jint*) addr = (jint) con; break; 3099 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3100 default: return false; // strange store size (detect size!=2**N here) 3101 } 3102 return true; // return success to caller 3103 } 3104 3105 // Coalesce subword constants into int constants and possibly 3106 // into long constants. The goal, if the CPU permits, 3107 // is to initialize the object with a small number of 64-bit tiles. 3108 // Also, convert floating-point constants to bit patterns. 3109 // Non-constants are not relevant to this pass. 3110 // 3111 // In terms of the running example on InitializeNode::InitializeNode 3112 // and InitializeNode::capture_store, here is the transformation 3113 // of rawstore1 and rawstore2 into rawstore12: 3114 // alloc = (Allocate ...) 3115 // rawoop = alloc.RawAddress 3116 // tile12 = 0x00010002 3117 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3118 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3119 // 3120 void 3121 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3122 Node* size_in_bytes, 3123 PhaseGVN* phase) { 3124 Compile* C = phase->C; 3125 3126 assert(stores_are_sane(phase), ""); 3127 // Note: After this pass, they are not completely sane, 3128 // since there may be some overlaps. 3129 3130 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3131 3132 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3133 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3134 size_limit = MIN2(size_limit, ti_limit); 3135 size_limit = align_size_up(size_limit, BytesPerLong); 3136 int num_tiles = size_limit / BytesPerLong; 3137 3138 // allocate space for the tile map: 3139 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3140 jlong tiles_buf[small_len]; 3141 Node* nodes_buf[small_len]; 3142 jlong inits_buf[small_len]; 3143 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3144 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3145 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3146 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3147 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3148 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3149 // tiles: exact bitwise model of all primitive constants 3150 // nodes: last constant-storing node subsumed into the tiles model 3151 // inits: which bytes (in each tile) are touched by any initializations 3152 3153 //// Pass A: Fill in the tile model with any relevant stores. 3154 3155 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3156 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3157 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3158 Node* zmem = zero_memory(); // initially zero memory state 3159 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3160 Node* st = in(i); 3161 intptr_t st_off = get_store_offset(st, phase); 3162 3163 // Figure out the store's offset and constant value: 3164 if (st_off < header_size) continue; //skip (ignore header) 3165 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3166 int st_size = st->as_Store()->memory_size(); 3167 if (st_off + st_size > size_limit) break; 3168 3169 // Record which bytes are touched, whether by constant or not. 3170 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3171 continue; // skip (strange store size) 3172 3173 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3174 if (!val->singleton()) continue; //skip (non-con store) 3175 BasicType type = val->basic_type(); 3176 3177 jlong con = 0; 3178 switch (type) { 3179 case T_INT: con = val->is_int()->get_con(); break; 3180 case T_LONG: con = val->is_long()->get_con(); break; 3181 case T_FLOAT: con = jint_cast(val->getf()); break; 3182 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3183 default: continue; //skip (odd store type) 3184 } 3185 3186 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3187 st->Opcode() == Op_StoreL) { 3188 continue; // This StoreL is already optimal. 3189 } 3190 3191 // Store down the constant. 3192 store_constant(tiles, num_tiles, st_off, st_size, con); 3193 3194 intptr_t j = st_off >> LogBytesPerLong; 3195 3196 if (type == T_INT && st_size == BytesPerInt 3197 && (st_off & BytesPerInt) == BytesPerInt) { 3198 jlong lcon = tiles[j]; 3199 if (!Matcher::isSimpleConstant64(lcon) && 3200 st->Opcode() == Op_StoreI) { 3201 // This StoreI is already optimal by itself. 3202 jint* intcon = (jint*) &tiles[j]; 3203 intcon[1] = 0; // undo the store_constant() 3204 3205 // If the previous store is also optimal by itself, back up and 3206 // undo the action of the previous loop iteration... if we can. 3207 // But if we can't, just let the previous half take care of itself. 3208 st = nodes[j]; 3209 st_off -= BytesPerInt; 3210 con = intcon[0]; 3211 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3212 assert(st_off >= header_size, "still ignoring header"); 3213 assert(get_store_offset(st, phase) == st_off, "must be"); 3214 assert(in(i-1) == zmem, "must be"); 3215 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3216 assert(con == tcon->is_int()->get_con(), "must be"); 3217 // Undo the effects of the previous loop trip, which swallowed st: 3218 intcon[0] = 0; // undo store_constant() 3219 set_req(i-1, st); // undo set_req(i, zmem) 3220 nodes[j] = NULL; // undo nodes[j] = st 3221 --old_subword; // undo ++old_subword 3222 } 3223 continue; // This StoreI is already optimal. 3224 } 3225 } 3226 3227 // This store is not needed. 3228 set_req(i, zmem); 3229 nodes[j] = st; // record for the moment 3230 if (st_size < BytesPerLong) // something has changed 3231 ++old_subword; // includes int/float, but who's counting... 3232 else ++old_long; 3233 } 3234 3235 if ((old_subword + old_long) == 0) 3236 return; // nothing more to do 3237 3238 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3239 // Be sure to insert them before overlapping non-constant stores. 3240 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3241 for (int j = 0; j < num_tiles; j++) { 3242 jlong con = tiles[j]; 3243 jlong init = inits[j]; 3244 if (con == 0) continue; 3245 jint con0, con1; // split the constant, address-wise 3246 jint init0, init1; // split the init map, address-wise 3247 { union { jlong con; jint intcon[2]; } u; 3248 u.con = con; 3249 con0 = u.intcon[0]; 3250 con1 = u.intcon[1]; 3251 u.con = init; 3252 init0 = u.intcon[0]; 3253 init1 = u.intcon[1]; 3254 } 3255 3256 Node* old = nodes[j]; 3257 assert(old != NULL, "need the prior store"); 3258 intptr_t offset = (j * BytesPerLong); 3259 3260 bool split = !Matcher::isSimpleConstant64(con); 3261 3262 if (offset < header_size) { 3263 assert(offset + BytesPerInt >= header_size, "second int counts"); 3264 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3265 split = true; // only the second word counts 3266 // Example: int a[] = { 42 ... } 3267 } else if (con0 == 0 && init0 == -1) { 3268 split = true; // first word is covered by full inits 3269 // Example: int a[] = { ... foo(), 42 ... } 3270 } else if (con1 == 0 && init1 == -1) { 3271 split = true; // second word is covered by full inits 3272 // Example: int a[] = { ... 42, foo() ... } 3273 } 3274 3275 // Here's a case where init0 is neither 0 nor -1: 3276 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3277 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3278 // In this case the tile is not split; it is (jlong)42. 3279 // The big tile is stored down, and then the foo() value is inserted. 3280 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3281 3282 Node* ctl = old->in(MemNode::Control); 3283 Node* adr = make_raw_address(offset, phase); 3284 const TypePtr* atp = TypeRawPtr::BOTTOM; 3285 3286 // One or two coalesced stores to plop down. 3287 Node* st[2]; 3288 intptr_t off[2]; 3289 int nst = 0; 3290 if (!split) { 3291 ++new_long; 3292 off[nst] = offset; 3293 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3294 phase->longcon(con), T_LONG); 3295 } else { 3296 // Omit either if it is a zero. 3297 if (con0 != 0) { 3298 ++new_int; 3299 off[nst] = offset; 3300 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3301 phase->intcon(con0), T_INT); 3302 } 3303 if (con1 != 0) { 3304 ++new_int; 3305 offset += BytesPerInt; 3306 adr = make_raw_address(offset, phase); 3307 off[nst] = offset; 3308 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3309 phase->intcon(con1), T_INT); 3310 } 3311 } 3312 3313 // Insert second store first, then the first before the second. 3314 // Insert each one just before any overlapping non-constant stores. 3315 while (nst > 0) { 3316 Node* st1 = st[--nst]; 3317 C->copy_node_notes_to(st1, old); 3318 st1 = phase->transform(st1); 3319 offset = off[nst]; 3320 assert(offset >= header_size, "do not smash header"); 3321 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3322 guarantee(ins_idx != 0, "must re-insert constant store"); 3323 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3324 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3325 set_req(--ins_idx, st1); 3326 else 3327 ins_req(ins_idx, st1); 3328 } 3329 } 3330 3331 if (PrintCompilation && WizardMode) 3332 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3333 old_subword, old_long, new_int, new_long); 3334 if (C->log() != NULL) 3335 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3336 old_subword, old_long, new_int, new_long); 3337 3338 // Clean up any remaining occurrences of zmem: 3339 remove_extra_zeroes(); 3340 } 3341 3342 // Explore forward from in(start) to find the first fully initialized 3343 // word, and return its offset. Skip groups of subword stores which 3344 // together initialize full words. If in(start) is itself part of a 3345 // fully initialized word, return the offset of in(start). If there 3346 // are no following full-word stores, or if something is fishy, return 3347 // a negative value. 3348 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3349 int int_map = 0; 3350 intptr_t int_map_off = 0; 3351 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3352 3353 for (uint i = start, limit = req(); i < limit; i++) { 3354 Node* st = in(i); 3355 3356 intptr_t st_off = get_store_offset(st, phase); 3357 if (st_off < 0) break; // return conservative answer 3358 3359 int st_size = st->as_Store()->memory_size(); 3360 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3361 return st_off; // we found a complete word init 3362 } 3363 3364 // update the map: 3365 3366 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3367 if (this_int_off != int_map_off) { 3368 // reset the map: 3369 int_map = 0; 3370 int_map_off = this_int_off; 3371 } 3372 3373 int subword_off = st_off - this_int_off; 3374 int_map |= right_n_bits(st_size) << subword_off; 3375 if ((int_map & FULL_MAP) == FULL_MAP) { 3376 return this_int_off; // we found a complete word init 3377 } 3378 3379 // Did this store hit or cross the word boundary? 3380 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3381 if (next_int_off == this_int_off + BytesPerInt) { 3382 // We passed the current int, without fully initializing it. 3383 int_map_off = next_int_off; 3384 int_map >>= BytesPerInt; 3385 } else if (next_int_off > this_int_off + BytesPerInt) { 3386 // We passed the current and next int. 3387 return this_int_off + BytesPerInt; 3388 } 3389 } 3390 3391 return -1; 3392 } 3393 3394 3395 // Called when the associated AllocateNode is expanded into CFG. 3396 // At this point, we may perform additional optimizations. 3397 // Linearize the stores by ascending offset, to make memory 3398 // activity as coherent as possible. 3399 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3400 intptr_t header_size, 3401 Node* size_in_bytes, 3402 PhaseGVN* phase) { 3403 assert(!is_complete(), "not already complete"); 3404 assert(stores_are_sane(phase), ""); 3405 assert(allocation() != NULL, "must be present"); 3406 3407 remove_extra_zeroes(); 3408 3409 if (ReduceFieldZeroing || ReduceBulkZeroing) 3410 // reduce instruction count for common initialization patterns 3411 coalesce_subword_stores(header_size, size_in_bytes, phase); 3412 3413 Node* zmem = zero_memory(); // initially zero memory state 3414 Node* inits = zmem; // accumulating a linearized chain of inits 3415 #ifdef ASSERT 3416 intptr_t first_offset = allocation()->minimum_header_size(); 3417 intptr_t last_init_off = first_offset; // previous init offset 3418 intptr_t last_init_end = first_offset; // previous init offset+size 3419 intptr_t last_tile_end = first_offset; // previous tile offset+size 3420 #endif 3421 intptr_t zeroes_done = header_size; 3422 3423 bool do_zeroing = true; // we might give up if inits are very sparse 3424 int big_init_gaps = 0; // how many large gaps have we seen? 3425 3426 if (ZeroTLAB) do_zeroing = false; 3427 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3428 3429 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3430 Node* st = in(i); 3431 intptr_t st_off = get_store_offset(st, phase); 3432 if (st_off < 0) 3433 break; // unknown junk in the inits 3434 if (st->in(MemNode::Memory) != zmem) 3435 break; // complicated store chains somehow in list 3436 3437 int st_size = st->as_Store()->memory_size(); 3438 intptr_t next_init_off = st_off + st_size; 3439 3440 if (do_zeroing && zeroes_done < next_init_off) { 3441 // See if this store needs a zero before it or under it. 3442 intptr_t zeroes_needed = st_off; 3443 3444 if (st_size < BytesPerInt) { 3445 // Look for subword stores which only partially initialize words. 3446 // If we find some, we must lay down some word-level zeroes first, 3447 // underneath the subword stores. 3448 // 3449 // Examples: 3450 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3451 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3452 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3453 // 3454 // Note: coalesce_subword_stores may have already done this, 3455 // if it was prompted by constant non-zero subword initializers. 3456 // But this case can still arise with non-constant stores. 3457 3458 intptr_t next_full_store = find_next_fullword_store(i, phase); 3459 3460 // In the examples above: 3461 // in(i) p q r s x y z 3462 // st_off 12 13 14 15 12 13 14 3463 // st_size 1 1 1 1 1 1 1 3464 // next_full_s. 12 16 16 16 16 16 16 3465 // z's_done 12 16 16 16 12 16 12 3466 // z's_needed 12 16 16 16 16 16 16 3467 // zsize 0 0 0 0 4 0 4 3468 if (next_full_store < 0) { 3469 // Conservative tack: Zero to end of current word. 3470 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3471 } else { 3472 // Zero to beginning of next fully initialized word. 3473 // Or, don't zero at all, if we are already in that word. 3474 assert(next_full_store >= zeroes_needed, "must go forward"); 3475 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3476 zeroes_needed = next_full_store; 3477 } 3478 } 3479 3480 if (zeroes_needed > zeroes_done) { 3481 intptr_t zsize = zeroes_needed - zeroes_done; 3482 // Do some incremental zeroing on rawmem, in parallel with inits. 3483 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3484 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3485 zeroes_done, zeroes_needed, 3486 phase); 3487 zeroes_done = zeroes_needed; 3488 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3489 do_zeroing = false; // leave the hole, next time 3490 } 3491 } 3492 3493 // Collect the store and move on: 3494 st->set_req(MemNode::Memory, inits); 3495 inits = st; // put it on the linearized chain 3496 set_req(i, zmem); // unhook from previous position 3497 3498 if (zeroes_done == st_off) 3499 zeroes_done = next_init_off; 3500 3501 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3502 3503 #ifdef ASSERT 3504 // Various order invariants. Weaker than stores_are_sane because 3505 // a large constant tile can be filled in by smaller non-constant stores. 3506 assert(st_off >= last_init_off, "inits do not reverse"); 3507 last_init_off = st_off; 3508 const Type* val = NULL; 3509 if (st_size >= BytesPerInt && 3510 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3511 (int)val->basic_type() < (int)T_OBJECT) { 3512 assert(st_off >= last_tile_end, "tiles do not overlap"); 3513 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3514 last_tile_end = MAX2(last_tile_end, next_init_off); 3515 } else { 3516 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3517 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3518 assert(st_off >= last_init_end, "inits do not overlap"); 3519 last_init_end = next_init_off; // it's a non-tile 3520 } 3521 #endif //ASSERT 3522 } 3523 3524 remove_extra_zeroes(); // clear out all the zmems left over 3525 add_req(inits); 3526 3527 if (!ZeroTLAB) { 3528 // If anything remains to be zeroed, zero it all now. 3529 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3530 // if it is the last unused 4 bytes of an instance, forget about it 3531 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3532 if (zeroes_done + BytesPerLong >= size_limit) { 3533 assert(allocation() != NULL, ""); 3534 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3535 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3536 if (zeroes_done == k->layout_helper()) 3537 zeroes_done = size_limit; 3538 } 3539 if (zeroes_done < size_limit) { 3540 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3541 zeroes_done, size_in_bytes, phase); 3542 } 3543 } 3544 3545 set_complete(phase); 3546 return rawmem; 3547 } 3548 3549 3550 #ifdef ASSERT 3551 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3552 if (is_complete()) 3553 return true; // stores could be anything at this point 3554 assert(allocation() != NULL, "must be present"); 3555 intptr_t last_off = allocation()->minimum_header_size(); 3556 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3557 Node* st = in(i); 3558 intptr_t st_off = get_store_offset(st, phase); 3559 if (st_off < 0) continue; // ignore dead garbage 3560 if (last_off > st_off) { 3561 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3562 this->dump(2); 3563 assert(false, "ascending store offsets"); 3564 return false; 3565 } 3566 last_off = st_off + st->as_Store()->memory_size(); 3567 } 3568 return true; 3569 } 3570 #endif //ASSERT 3571 3572 3573 3574 3575 //============================MergeMemNode===================================== 3576 // 3577 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3578 // contributing store or call operations. Each contributor provides the memory 3579 // state for a particular "alias type" (see Compile::alias_type). For example, 3580 // if a MergeMem has an input X for alias category #6, then any memory reference 3581 // to alias category #6 may use X as its memory state input, as an exact equivalent 3582 // to using the MergeMem as a whole. 3583 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3584 // 3585 // (Here, the <N> notation gives the index of the relevant adr_type.) 3586 // 3587 // In one special case (and more cases in the future), alias categories overlap. 3588 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3589 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3590 // it is exactly equivalent to that state W: 3591 // MergeMem(<Bot>: W) <==> W 3592 // 3593 // Usually, the merge has more than one input. In that case, where inputs 3594 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3595 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3596 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3597 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3598 // 3599 // A merge can take a "wide" memory state as one of its narrow inputs. 3600 // This simply means that the merge observes out only the relevant parts of 3601 // the wide input. That is, wide memory states arriving at narrow merge inputs 3602 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3603 // 3604 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3605 // and that memory slices "leak through": 3606 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3607 // 3608 // But, in such a cascade, repeated memory slices can "block the leak": 3609 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3610 // 3611 // In the last example, Y is not part of the combined memory state of the 3612 // outermost MergeMem. The system must, of course, prevent unschedulable 3613 // memory states from arising, so you can be sure that the state Y is somehow 3614 // a precursor to state Y'. 3615 // 3616 // 3617 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3618 // of each MergeMemNode array are exactly the numerical alias indexes, including 3619 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3620 // Compile::alias_type (and kin) produce and manage these indexes. 3621 // 3622 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3623 // (Note that this provides quick access to the top node inside MergeMem methods, 3624 // without the need to reach out via TLS to Compile::current.) 3625 // 3626 // As a consequence of what was just described, a MergeMem that represents a full 3627 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3628 // containing all alias categories. 3629 // 3630 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3631 // 3632 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3633 // a memory state for the alias type <N>, or else the top node, meaning that 3634 // there is no particular input for that alias type. Note that the length of 3635 // a MergeMem is variable, and may be extended at any time to accommodate new 3636 // memory states at larger alias indexes. When merges grow, they are of course 3637 // filled with "top" in the unused in() positions. 3638 // 3639 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3640 // (Top was chosen because it works smoothly with passes like GCM.) 3641 // 3642 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3643 // the type of random VM bits like TLS references.) Since it is always the 3644 // first non-Bot memory slice, some low-level loops use it to initialize an 3645 // index variable: for (i = AliasIdxRaw; i < req(); i++). 3646 // 3647 // 3648 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3649 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3650 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 3651 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3652 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3653 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3654 // 3655 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3656 // really that different from the other memory inputs. An abbreviation called 3657 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3658 // 3659 // 3660 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3661 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3662 // that "emerges though" the base memory will be marked as excluding the alias types 3663 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 3664 // 3665 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3666 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3667 // 3668 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 3669 // (It is currently unimplemented.) As you can see, the resulting merge is 3670 // actually a disjoint union of memory states, rather than an overlay. 3671 // 3672 3673 //------------------------------MergeMemNode----------------------------------- 3674 Node* MergeMemNode::make_empty_memory() { 3675 Node* empty_memory = (Node*) Compile::current()->top(); 3676 assert(empty_memory->is_top(), "correct sentinel identity"); 3677 return empty_memory; 3678 } 3679 3680 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3681 init_class_id(Class_MergeMem); 3682 // all inputs are nullified in Node::Node(int) 3683 // set_input(0, NULL); // no control input 3684 3685 // Initialize the edges uniformly to top, for starters. 3686 Node* empty_mem = make_empty_memory(); 3687 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3688 init_req(i,empty_mem); 3689 } 3690 assert(empty_memory() == empty_mem, ""); 3691 3692 if( new_base != NULL && new_base->is_MergeMem() ) { 3693 MergeMemNode* mdef = new_base->as_MergeMem(); 3694 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3695 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3696 mms.set_memory(mms.memory2()); 3697 } 3698 assert(base_memory() == mdef->base_memory(), ""); 3699 } else { 3700 set_base_memory(new_base); 3701 } 3702 } 3703 3704 // Make a new, untransformed MergeMem with the same base as 'mem'. 3705 // If mem is itself a MergeMem, populate the result with the same edges. 3706 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3707 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3708 } 3709 3710 //------------------------------cmp-------------------------------------------- 3711 uint MergeMemNode::hash() const { return NO_HASH; } 3712 uint MergeMemNode::cmp( const Node &n ) const { 3713 return (&n == this); // Always fail except on self 3714 } 3715 3716 //------------------------------Identity--------------------------------------- 3717 Node* MergeMemNode::Identity(PhaseTransform *phase) { 3718 // Identity if this merge point does not record any interesting memory 3719 // disambiguations. 3720 Node* base_mem = base_memory(); 3721 Node* empty_mem = empty_memory(); 3722 if (base_mem != empty_mem) { // Memory path is not dead? 3723 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3724 Node* mem = in(i); 3725 if (mem != empty_mem && mem != base_mem) { 3726 return this; // Many memory splits; no change 3727 } 3728 } 3729 } 3730 return base_mem; // No memory splits; ID on the one true input 3731 } 3732 3733 //------------------------------Ideal------------------------------------------ 3734 // This method is invoked recursively on chains of MergeMem nodes 3735 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3736 // Remove chain'd MergeMems 3737 // 3738 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3739 // relative to the "in(Bot)". Since we are patching both at the same time, 3740 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3741 // but rewrite each "in(i)" relative to the new "in(Bot)". 3742 Node *progress = NULL; 3743 3744 3745 Node* old_base = base_memory(); 3746 Node* empty_mem = empty_memory(); 3747 if (old_base == empty_mem) 3748 return NULL; // Dead memory path. 3749 3750 MergeMemNode* old_mbase; 3751 if (old_base != NULL && old_base->is_MergeMem()) 3752 old_mbase = old_base->as_MergeMem(); 3753 else 3754 old_mbase = NULL; 3755 Node* new_base = old_base; 3756 3757 // simplify stacked MergeMems in base memory 3758 if (old_mbase) new_base = old_mbase->base_memory(); 3759 3760 // the base memory might contribute new slices beyond my req() 3761 if (old_mbase) grow_to_match(old_mbase); 3762 3763 // Look carefully at the base node if it is a phi. 3764 PhiNode* phi_base; 3765 if (new_base != NULL && new_base->is_Phi()) 3766 phi_base = new_base->as_Phi(); 3767 else 3768 phi_base = NULL; 3769 3770 Node* phi_reg = NULL; 3771 uint phi_len = (uint)-1; 3772 if (phi_base != NULL && !phi_base->is_copy()) { 3773 // do not examine phi if degraded to a copy 3774 phi_reg = phi_base->region(); 3775 phi_len = phi_base->req(); 3776 // see if the phi is unfinished 3777 for (uint i = 1; i < phi_len; i++) { 3778 if (phi_base->in(i) == NULL) { 3779 // incomplete phi; do not look at it yet! 3780 phi_reg = NULL; 3781 phi_len = (uint)-1; 3782 break; 3783 } 3784 } 3785 } 3786 3787 // Note: We do not call verify_sparse on entry, because inputs 3788 // can normalize to the base_memory via subsume_node or similar 3789 // mechanisms. This method repairs that damage. 3790 3791 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3792 3793 // Look at each slice. 3794 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3795 Node* old_in = in(i); 3796 // calculate the old memory value 3797 Node* old_mem = old_in; 3798 if (old_mem == empty_mem) old_mem = old_base; 3799 assert(old_mem == memory_at(i), ""); 3800 3801 // maybe update (reslice) the old memory value 3802 3803 // simplify stacked MergeMems 3804 Node* new_mem = old_mem; 3805 MergeMemNode* old_mmem; 3806 if (old_mem != NULL && old_mem->is_MergeMem()) 3807 old_mmem = old_mem->as_MergeMem(); 3808 else 3809 old_mmem = NULL; 3810 if (old_mmem == this) { 3811 // This can happen if loops break up and safepoints disappear. 3812 // A merge of BotPtr (default) with a RawPtr memory derived from a 3813 // safepoint can be rewritten to a merge of the same BotPtr with 3814 // the BotPtr phi coming into the loop. If that phi disappears 3815 // also, we can end up with a self-loop of the mergemem. 3816 // In general, if loops degenerate and memory effects disappear, 3817 // a mergemem can be left looking at itself. This simply means 3818 // that the mergemem's default should be used, since there is 3819 // no longer any apparent effect on this slice. 3820 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3821 // from start. Update the input to TOP. 3822 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3823 } 3824 else if (old_mmem != NULL) { 3825 new_mem = old_mmem->memory_at(i); 3826 } 3827 // else preceding memory was not a MergeMem 3828 3829 // replace equivalent phis (unfortunately, they do not GVN together) 3830 if (new_mem != NULL && new_mem != new_base && 3831 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3832 if (new_mem->is_Phi()) { 3833 PhiNode* phi_mem = new_mem->as_Phi(); 3834 for (uint i = 1; i < phi_len; i++) { 3835 if (phi_base->in(i) != phi_mem->in(i)) { 3836 phi_mem = NULL; 3837 break; 3838 } 3839 } 3840 if (phi_mem != NULL) { 3841 // equivalent phi nodes; revert to the def 3842 new_mem = new_base; 3843 } 3844 } 3845 } 3846 3847 // maybe store down a new value 3848 Node* new_in = new_mem; 3849 if (new_in == new_base) new_in = empty_mem; 3850 3851 if (new_in != old_in) { 3852 // Warning: Do not combine this "if" with the previous "if" 3853 // A memory slice might have be be rewritten even if it is semantically 3854 // unchanged, if the base_memory value has changed. 3855 set_req(i, new_in); 3856 progress = this; // Report progress 3857 } 3858 } 3859 3860 if (new_base != old_base) { 3861 set_req(Compile::AliasIdxBot, new_base); 3862 // Don't use set_base_memory(new_base), because we need to update du. 3863 assert(base_memory() == new_base, ""); 3864 progress = this; 3865 } 3866 3867 if( base_memory() == this ) { 3868 // a self cycle indicates this memory path is dead 3869 set_req(Compile::AliasIdxBot, empty_mem); 3870 } 3871 3872 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3873 // Recursion must occur after the self cycle check above 3874 if( base_memory()->is_MergeMem() ) { 3875 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3876 Node *m = phase->transform(new_mbase); // Rollup any cycles 3877 if( m != NULL && (m->is_top() || 3878 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3879 // propagate rollup of dead cycle to self 3880 set_req(Compile::AliasIdxBot, empty_mem); 3881 } 3882 } 3883 3884 if( base_memory() == empty_mem ) { 3885 progress = this; 3886 // Cut inputs during Parse phase only. 3887 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3888 if( !can_reshape ) { 3889 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3890 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3891 } 3892 } 3893 } 3894 3895 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3896 // Check if PhiNode::Ideal's "Split phis through memory merges" 3897 // transform should be attempted. Look for this->phi->this cycle. 3898 uint merge_width = req(); 3899 if (merge_width > Compile::AliasIdxRaw) { 3900 PhiNode* phi = base_memory()->as_Phi(); 3901 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3902 if (phi->in(i) == this) { 3903 phase->is_IterGVN()->_worklist.push(phi); 3904 break; 3905 } 3906 } 3907 } 3908 } 3909 3910 assert(progress || verify_sparse(), "please, no dups of base"); 3911 return progress; 3912 } 3913 3914 //-------------------------set_base_memory------------------------------------- 3915 void MergeMemNode::set_base_memory(Node *new_base) { 3916 Node* empty_mem = empty_memory(); 3917 set_req(Compile::AliasIdxBot, new_base); 3918 assert(memory_at(req()) == new_base, "must set default memory"); 3919 // Clear out other occurrences of new_base: 3920 if (new_base != empty_mem) { 3921 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3922 if (in(i) == new_base) set_req(i, empty_mem); 3923 } 3924 } 3925 } 3926 3927 //------------------------------out_RegMask------------------------------------ 3928 const RegMask &MergeMemNode::out_RegMask() const { 3929 return RegMask::Empty; 3930 } 3931 3932 //------------------------------dump_spec-------------------------------------- 3933 #ifndef PRODUCT 3934 void MergeMemNode::dump_spec(outputStream *st) const { 3935 st->print(" {"); 3936 Node* base_mem = base_memory(); 3937 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3938 Node* mem = memory_at(i); 3939 if (mem == base_mem) { st->print(" -"); continue; } 3940 st->print( " N%d:", mem->_idx ); 3941 Compile::current()->get_adr_type(i)->dump_on(st); 3942 } 3943 st->print(" }"); 3944 } 3945 #endif // !PRODUCT 3946 3947 3948 #ifdef ASSERT 3949 static bool might_be_same(Node* a, Node* b) { 3950 if (a == b) return true; 3951 if (!(a->is_Phi() || b->is_Phi())) return false; 3952 // phis shift around during optimization 3953 return true; // pretty stupid... 3954 } 3955 3956 // verify a narrow slice (either incoming or outgoing) 3957 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 3958 if (!VerifyAliases) return; // don't bother to verify unless requested 3959 if (is_error_reported()) return; // muzzle asserts when debugging an error 3960 if (Node::in_dump()) return; // muzzle asserts when printing 3961 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 3962 assert(n != NULL, ""); 3963 // Elide intervening MergeMem's 3964 while (n->is_MergeMem()) { 3965 n = n->as_MergeMem()->memory_at(alias_idx); 3966 } 3967 Compile* C = Compile::current(); 3968 const TypePtr* n_adr_type = n->adr_type(); 3969 if (n == m->empty_memory()) { 3970 // Implicit copy of base_memory() 3971 } else if (n_adr_type != TypePtr::BOTTOM) { 3972 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 3973 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 3974 } else { 3975 // A few places like make_runtime_call "know" that VM calls are narrow, 3976 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 3977 bool expected_wide_mem = false; 3978 if (n == m->base_memory()) { 3979 expected_wide_mem = true; 3980 } else if (alias_idx == Compile::AliasIdxRaw || 3981 n == m->memory_at(Compile::AliasIdxRaw)) { 3982 expected_wide_mem = true; 3983 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 3984 // memory can "leak through" calls on channels that 3985 // are write-once. Allow this also. 3986 expected_wide_mem = true; 3987 } 3988 assert(expected_wide_mem, "expected narrow slice replacement"); 3989 } 3990 } 3991 #else // !ASSERT 3992 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 3993 #endif 3994 3995 3996 //-----------------------------memory_at--------------------------------------- 3997 Node* MergeMemNode::memory_at(uint alias_idx) const { 3998 assert(alias_idx >= Compile::AliasIdxRaw || 3999 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4000 "must avoid base_memory and AliasIdxTop"); 4001 4002 // Otherwise, it is a narrow slice. 4003 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4004 Compile *C = Compile::current(); 4005 if (is_empty_memory(n)) { 4006 // the array is sparse; empty slots are the "top" node 4007 n = base_memory(); 4008 assert(Node::in_dump() 4009 || n == NULL || n->bottom_type() == Type::TOP 4010 || n->adr_type() == TypePtr::BOTTOM 4011 || n->adr_type() == TypeRawPtr::BOTTOM 4012 || Compile::current()->AliasLevel() == 0, 4013 "must be a wide memory"); 4014 // AliasLevel == 0 if we are organizing the memory states manually. 4015 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4016 } else { 4017 // make sure the stored slice is sane 4018 #ifdef ASSERT 4019 if (is_error_reported() || Node::in_dump()) { 4020 } else if (might_be_same(n, base_memory())) { 4021 // Give it a pass: It is a mostly harmless repetition of the base. 4022 // This can arise normally from node subsumption during optimization. 4023 } else { 4024 verify_memory_slice(this, alias_idx, n); 4025 } 4026 #endif 4027 } 4028 return n; 4029 } 4030 4031 //---------------------------set_memory_at------------------------------------- 4032 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4033 verify_memory_slice(this, alias_idx, n); 4034 Node* empty_mem = empty_memory(); 4035 if (n == base_memory()) n = empty_mem; // collapse default 4036 uint need_req = alias_idx+1; 4037 if (req() < need_req) { 4038 if (n == empty_mem) return; // already the default, so do not grow me 4039 // grow the sparse array 4040 do { 4041 add_req(empty_mem); 4042 } while (req() < need_req); 4043 } 4044 set_req( alias_idx, n ); 4045 } 4046 4047 4048 4049 //--------------------------iteration_setup------------------------------------ 4050 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4051 if (other != NULL) { 4052 grow_to_match(other); 4053 // invariant: the finite support of mm2 is within mm->req() 4054 #ifdef ASSERT 4055 for (uint i = req(); i < other->req(); i++) { 4056 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4057 } 4058 #endif 4059 } 4060 // Replace spurious copies of base_memory by top. 4061 Node* base_mem = base_memory(); 4062 if (base_mem != NULL && !base_mem->is_top()) { 4063 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4064 if (in(i) == base_mem) 4065 set_req(i, empty_memory()); 4066 } 4067 } 4068 } 4069 4070 //---------------------------grow_to_match------------------------------------- 4071 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4072 Node* empty_mem = empty_memory(); 4073 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4074 // look for the finite support of the other memory 4075 for (uint i = other->req(); --i >= req(); ) { 4076 if (other->in(i) != empty_mem) { 4077 uint new_len = i+1; 4078 while (req() < new_len) add_req(empty_mem); 4079 break; 4080 } 4081 } 4082 } 4083 4084 //---------------------------verify_sparse------------------------------------- 4085 #ifndef PRODUCT 4086 bool MergeMemNode::verify_sparse() const { 4087 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4088 Node* base_mem = base_memory(); 4089 // The following can happen in degenerate cases, since empty==top. 4090 if (is_empty_memory(base_mem)) return true; 4091 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4092 assert(in(i) != NULL, "sane slice"); 4093 if (in(i) == base_mem) return false; // should have been the sentinel value! 4094 } 4095 return true; 4096 } 4097 4098 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4099 Node* n; 4100 n = mm->in(idx); 4101 if (mem == n) return true; // might be empty_memory() 4102 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4103 if (mem == n) return true; 4104 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4105 if (mem == n) return true; 4106 if (n == NULL) break; 4107 } 4108 return false; 4109 } 4110 #endif // !PRODUCT