1 /* 2 * Copyright (c) 1997, 2009, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * 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->is_ClearArray()) { 127 if (!ClearArrayNode::step_through(&result, instance_id, phase)) { 128 // Can not bypass initialization of the instance 129 // we are looking for. 130 break; 131 } 132 // Otherwise skip it (the call updated 'result' value). 133 } else if (result->is_MergeMem()) { 134 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 135 } 136 } 137 return result; 138 } 139 140 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 141 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 142 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); 143 PhaseIterGVN *igvn = phase->is_IterGVN(); 144 Node *result = mchain; 145 result = optimize_simple_memory_chain(result, t_adr, phase); 146 if (is_instance && igvn != NULL && result->is_Phi()) { 147 PhiNode *mphi = result->as_Phi(); 148 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 149 const TypePtr *t = mphi->adr_type(); 150 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 151 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 152 t->is_oopptr()->cast_to_exactness(true) 153 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 154 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 155 // clone the Phi with our address type 156 result = mphi->split_out_instance(t_adr, igvn); 157 } else { 158 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 159 } 160 } 161 return result; 162 } 163 164 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 165 uint alias_idx = phase->C->get_alias_index(tp); 166 Node *mem = mmem; 167 #ifdef ASSERT 168 { 169 // Check that current type is consistent with the alias index used during graph construction 170 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 171 bool consistent = adr_check == NULL || adr_check->empty() || 172 phase->C->must_alias(adr_check, alias_idx ); 173 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 174 if( !consistent && adr_check != NULL && !adr_check->empty() && 175 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 176 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 177 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 178 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 179 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 180 // don't assert if it is dead code. 181 consistent = true; 182 } 183 if( !consistent ) { 184 st->print("alias_idx==%d, adr_check==", alias_idx); 185 if( adr_check == NULL ) { 186 st->print("NULL"); 187 } else { 188 adr_check->dump(); 189 } 190 st->cr(); 191 print_alias_types(); 192 assert(consistent, "adr_check must match alias idx"); 193 } 194 } 195 #endif 196 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 197 // means an array I have not precisely typed yet. Do not do any 198 // alias stuff with it any time soon. 199 const TypeOopPtr *tinst = tp->isa_oopptr(); 200 if( tp->base() != Type::AnyPtr && 201 !(tinst && 202 tinst->klass()->is_java_lang_Object() && 203 tinst->offset() == Type::OffsetBot) ) { 204 // compress paths and change unreachable cycles to TOP 205 // If not, we can update the input infinitely along a MergeMem cycle 206 // Equivalent code in PhiNode::Ideal 207 Node* m = phase->transform(mmem); 208 // If transformed to a MergeMem, get the desired slice 209 // Otherwise the returned node represents memory for every slice 210 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 211 // Update input if it is progress over what we have now 212 } 213 return mem; 214 } 215 216 //--------------------------Ideal_common--------------------------------------- 217 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 218 // Unhook non-raw memories from complete (macro-expanded) initializations. 219 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 220 // If our control input is a dead region, kill all below the region 221 Node *ctl = in(MemNode::Control); 222 if (ctl && remove_dead_region(phase, can_reshape)) 223 return this; 224 ctl = in(MemNode::Control); 225 // Don't bother trying to transform a dead node 226 if( ctl && ctl->is_top() ) return NodeSentinel; 227 228 PhaseIterGVN *igvn = phase->is_IterGVN(); 229 // Wait if control on the worklist. 230 if (ctl && can_reshape && igvn != NULL) { 231 Node* bol = NULL; 232 Node* cmp = NULL; 233 if (ctl->in(0)->is_If()) { 234 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 235 bol = ctl->in(0)->in(1); 236 if (bol->is_Bool()) 237 cmp = ctl->in(0)->in(1)->in(1); 238 } 239 if (igvn->_worklist.member(ctl) || 240 (bol != NULL && igvn->_worklist.member(bol)) || 241 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 242 // This control path may be dead. 243 // Delay this memory node transformation until the control is processed. 244 phase->is_IterGVN()->_worklist.push(this); 245 return NodeSentinel; // caller will return NULL 246 } 247 } 248 // Ignore if memory is dead, or self-loop 249 Node *mem = in(MemNode::Memory); 250 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 251 assert( mem != this, "dead loop in MemNode::Ideal" ); 252 253 Node *address = in(MemNode::Address); 254 const Type *t_adr = phase->type( address ); 255 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 256 257 if( can_reshape && igvn != NULL && 258 (igvn->_worklist.member(address) || phase->type(address) != adr_type()) ) { 259 // The address's base and type may change when the address is processed. 260 // Delay this mem node transformation until the address is processed. 261 phase->is_IterGVN()->_worklist.push(this); 262 return NodeSentinel; // caller will return NULL 263 } 264 265 // Do NOT remove or optimize the next lines: ensure a new alias index 266 // is allocated for an oop pointer type before Escape Analysis. 267 // Note: C++ will not remove it since the call has side effect. 268 if ( t_adr->isa_oopptr() ) { 269 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 270 } 271 272 #ifdef ASSERT 273 Node* base = NULL; 274 if (address->is_AddP()) 275 base = address->in(AddPNode::Base); 276 assert(base == NULL || t_adr->isa_rawptr() || 277 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); 278 #endif 279 280 // Avoid independent memory operations 281 Node* old_mem = mem; 282 283 // The code which unhooks non-raw memories from complete (macro-expanded) 284 // initializations was removed. After macro-expansion all stores catched 285 // by Initialize node became raw stores and there is no information 286 // which memory slices they modify. So it is unsafe to move any memory 287 // operation above these stores. Also in most cases hooked non-raw memories 288 // were already unhooked by using information from detect_ptr_independence() 289 // and find_previous_store(). 290 291 if (mem->is_MergeMem()) { 292 MergeMemNode* mmem = mem->as_MergeMem(); 293 const TypePtr *tp = t_adr->is_ptr(); 294 295 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 296 } 297 298 if (mem != old_mem) { 299 set_req(MemNode::Memory, mem); 300 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 301 return this; 302 } 303 304 // let the subclass continue analyzing... 305 return NULL; 306 } 307 308 // Helper function for proving some simple control dominations. 309 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 310 // Already assumes that 'dom' is available at 'sub', and that 'sub' 311 // is not a constant (dominated by the method's StartNode). 312 // Used by MemNode::find_previous_store to prove that the 313 // control input of a memory operation predates (dominates) 314 // an allocation it wants to look past. 315 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 316 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 317 return false; // Conservative answer for dead code 318 319 // Check 'dom'. Skip Proj and CatchProj nodes. 320 dom = dom->find_exact_control(dom); 321 if (dom == NULL || dom->is_top()) 322 return false; // Conservative answer for dead code 323 324 if (dom == sub) { 325 // For the case when, for example, 'sub' is Initialize and the original 326 // 'dom' is Proj node of the 'sub'. 327 return false; 328 } 329 330 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 331 return true; 332 333 // 'dom' dominates 'sub' if its control edge and control edges 334 // of all its inputs dominate or equal to sub's control edge. 335 336 // Currently 'sub' is either Allocate, Initialize or Start nodes. 337 // Or Region for the check in LoadNode::Ideal(); 338 // 'sub' should have sub->in(0) != NULL. 339 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 340 sub->is_Region(), "expecting only these nodes"); 341 342 // Get control edge of 'sub'. 343 Node* orig_sub = sub; 344 sub = sub->find_exact_control(sub->in(0)); 345 if (sub == NULL || sub->is_top()) 346 return false; // Conservative answer for dead code 347 348 assert(sub->is_CFG(), "expecting control"); 349 350 if (sub == dom) 351 return true; 352 353 if (sub->is_Start() || sub->is_Root()) 354 return false; 355 356 { 357 // Check all control edges of 'dom'. 358 359 ResourceMark rm; 360 Arena* arena = Thread::current()->resource_area(); 361 Node_List nlist(arena); 362 Unique_Node_List dom_list(arena); 363 364 dom_list.push(dom); 365 bool only_dominating_controls = false; 366 367 for (uint next = 0; next < dom_list.size(); next++) { 368 Node* n = dom_list.at(next); 369 if (n == orig_sub) 370 return false; // One of dom's inputs dominated by sub. 371 if (!n->is_CFG() && n->pinned()) { 372 // Check only own control edge for pinned non-control nodes. 373 n = n->find_exact_control(n->in(0)); 374 if (n == NULL || n->is_top()) 375 return false; // Conservative answer for dead code 376 assert(n->is_CFG(), "expecting control"); 377 dom_list.push(n); 378 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 379 only_dominating_controls = true; 380 } else if (n->is_CFG()) { 381 if (n->dominates(sub, nlist)) 382 only_dominating_controls = true; 383 else 384 return false; 385 } else { 386 // First, own control edge. 387 Node* m = n->find_exact_control(n->in(0)); 388 if (m != NULL) { 389 if (m->is_top()) 390 return false; // Conservative answer for dead code 391 dom_list.push(m); 392 } 393 // Now, the rest of edges. 394 uint cnt = n->req(); 395 for (uint i = 1; i < cnt; i++) { 396 m = n->find_exact_control(n->in(i)); 397 if (m == NULL || m->is_top()) 398 continue; 399 dom_list.push(m); 400 } 401 } 402 } 403 return only_dominating_controls; 404 } 405 } 406 407 //---------------------detect_ptr_independence--------------------------------- 408 // Used by MemNode::find_previous_store to prove that two base 409 // pointers are never equal. 410 // The pointers are accompanied by their associated allocations, 411 // if any, which have been previously discovered by the caller. 412 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 413 Node* p2, AllocateNode* a2, 414 PhaseTransform* phase) { 415 // Attempt to prove that these two pointers cannot be aliased. 416 // They may both manifestly be allocations, and they should differ. 417 // Or, if they are not both allocations, they can be distinct constants. 418 // Otherwise, one is an allocation and the other a pre-existing value. 419 if (a1 == NULL && a2 == NULL) { // neither an allocation 420 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 421 } else if (a1 != NULL && a2 != NULL) { // both allocations 422 return (a1 != a2); 423 } else if (a1 != NULL) { // one allocation a1 424 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 425 return all_controls_dominate(p2, a1); 426 } else { //(a2 != NULL) // one allocation a2 427 return all_controls_dominate(p1, a2); 428 } 429 return false; 430 } 431 432 433 // The logic for reordering loads and stores uses four steps: 434 // (a) Walk carefully past stores and initializations which we 435 // can prove are independent of this load. 436 // (b) Observe that the next memory state makes an exact match 437 // with self (load or store), and locate the relevant store. 438 // (c) Ensure that, if we were to wire self directly to the store, 439 // the optimizer would fold it up somehow. 440 // (d) Do the rewiring, and return, depending on some other part of 441 // the optimizer to fold up the load. 442 // This routine handles steps (a) and (b). Steps (c) and (d) are 443 // specific to loads and stores, so they are handled by the callers. 444 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 445 // 446 Node* MemNode::find_previous_store(PhaseTransform* phase) { 447 Node* ctrl = in(MemNode::Control); 448 Node* adr = in(MemNode::Address); 449 intptr_t offset = 0; 450 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 451 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 452 453 if (offset == Type::OffsetBot) 454 return NULL; // cannot unalias unless there are precise offsets 455 456 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 457 458 intptr_t size_in_bytes = memory_size(); 459 460 Node* mem = in(MemNode::Memory); // start searching here... 461 462 int cnt = 50; // Cycle limiter 463 for (;;) { // While we can dance past unrelated stores... 464 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 465 466 if (mem->is_Store()) { 467 Node* st_adr = mem->in(MemNode::Address); 468 intptr_t st_offset = 0; 469 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 470 if (st_base == NULL) 471 break; // inscrutable pointer 472 if (st_offset != offset && st_offset != Type::OffsetBot) { 473 const int MAX_STORE = BytesPerLong; 474 if (st_offset >= offset + size_in_bytes || 475 st_offset <= offset - MAX_STORE || 476 st_offset <= offset - mem->as_Store()->memory_size()) { 477 // Success: The offsets are provably independent. 478 // (You may ask, why not just test st_offset != offset and be done? 479 // The answer is that stores of different sizes can co-exist 480 // in the same sequence of RawMem effects. We sometimes initialize 481 // a whole 'tile' of array elements with a single jint or jlong.) 482 mem = mem->in(MemNode::Memory); 483 continue; // (a) advance through independent store memory 484 } 485 } 486 if (st_base != base && 487 detect_ptr_independence(base, alloc, 488 st_base, 489 AllocateNode::Ideal_allocation(st_base, phase), 490 phase)) { 491 // Success: The bases are provably independent. 492 mem = mem->in(MemNode::Memory); 493 continue; // (a) advance through independent store memory 494 } 495 496 // (b) At this point, if the bases or offsets do not agree, we lose, 497 // since we have not managed to prove 'this' and 'mem' independent. 498 if (st_base == base && st_offset == offset) { 499 return mem; // let caller handle steps (c), (d) 500 } 501 502 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 503 InitializeNode* st_init = mem->in(0)->as_Initialize(); 504 AllocateNode* st_alloc = st_init->allocation(); 505 if (st_alloc == NULL) 506 break; // something degenerated 507 bool known_identical = false; 508 bool known_independent = false; 509 if (alloc == st_alloc) 510 known_identical = true; 511 else if (alloc != NULL) 512 known_independent = true; 513 else if (all_controls_dominate(this, st_alloc)) 514 known_independent = true; 515 516 if (known_independent) { 517 // The bases are provably independent: Either they are 518 // manifestly distinct allocations, or else the control 519 // of this load dominates the store's allocation. 520 int alias_idx = phase->C->get_alias_index(adr_type()); 521 if (alias_idx == Compile::AliasIdxRaw) { 522 mem = st_alloc->in(TypeFunc::Memory); 523 } else { 524 mem = st_init->memory(alias_idx); 525 } 526 continue; // (a) advance through independent store memory 527 } 528 529 // (b) at this point, if we are not looking at a store initializing 530 // the same allocation we are loading from, we lose. 531 if (known_identical) { 532 // From caller, can_see_stored_value will consult find_captured_store. 533 return mem; // let caller handle steps (c), (d) 534 } 535 536 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 537 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 538 if (mem->is_Proj() && mem->in(0)->is_Call()) { 539 CallNode *call = mem->in(0)->as_Call(); 540 if (!call->may_modify(addr_t, phase)) { 541 mem = call->in(TypeFunc::Memory); 542 continue; // (a) advance through independent call memory 543 } 544 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 545 mem = mem->in(0)->in(TypeFunc::Memory); 546 continue; // (a) advance through independent MemBar memory 547 } else if (mem->is_ClearArray()) { 548 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 549 // (the call updated 'mem' value) 550 continue; // (a) advance through independent allocation memory 551 } else { 552 // Can not bypass initialization of the instance 553 // we are looking for. 554 return mem; 555 } 556 } else if (mem->is_MergeMem()) { 557 int alias_idx = phase->C->get_alias_index(adr_type()); 558 mem = mem->as_MergeMem()->memory_at(alias_idx); 559 continue; // (a) advance through independent MergeMem memory 560 } 561 } 562 563 // Unless there is an explicit 'continue', we must bail out here, 564 // because 'mem' is an inscrutable memory state (e.g., a call). 565 break; 566 } 567 568 return NULL; // bail out 569 } 570 571 //----------------------calculate_adr_type------------------------------------- 572 // Helper function. Notices when the given type of address hits top or bottom. 573 // Also, asserts a cross-check of the type against the expected address type. 574 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 575 if (t == Type::TOP) return NULL; // does not touch memory any more? 576 #ifdef PRODUCT 577 cross_check = NULL; 578 #else 579 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 580 #endif 581 const TypePtr* tp = t->isa_ptr(); 582 if (tp == NULL) { 583 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 584 return TypePtr::BOTTOM; // touches lots of memory 585 } else { 586 #ifdef ASSERT 587 // %%%% [phh] We don't check the alias index if cross_check is 588 // TypeRawPtr::BOTTOM. Needs to be investigated. 589 if (cross_check != NULL && 590 cross_check != TypePtr::BOTTOM && 591 cross_check != TypeRawPtr::BOTTOM) { 592 // Recheck the alias index, to see if it has changed (due to a bug). 593 Compile* C = Compile::current(); 594 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 595 "must stay in the original alias category"); 596 // The type of the address must be contained in the adr_type, 597 // disregarding "null"-ness. 598 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 599 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 600 assert(cross_check->meet(tp_notnull) == cross_check, 601 "real address must not escape from expected memory type"); 602 } 603 #endif 604 return tp; 605 } 606 } 607 608 //------------------------adr_phi_is_loop_invariant---------------------------- 609 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 610 // loop is loop invariant. Make a quick traversal of Phi and associated 611 // CastPP nodes, looking to see if they are a closed group within the loop. 612 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 613 // The idea is that the phi-nest must boil down to only CastPP nodes 614 // with the same data. This implies that any path into the loop already 615 // includes such a CastPP, and so the original cast, whatever its input, 616 // must be covered by an equivalent cast, with an earlier control input. 617 ResourceMark rm; 618 619 // The loop entry input of the phi should be the unique dominating 620 // node for every Phi/CastPP in the loop. 621 Unique_Node_List closure; 622 closure.push(adr_phi->in(LoopNode::EntryControl)); 623 624 // Add the phi node and the cast to the worklist. 625 Unique_Node_List worklist; 626 worklist.push(adr_phi); 627 if( cast != NULL ){ 628 if( !cast->is_ConstraintCast() ) return false; 629 worklist.push(cast); 630 } 631 632 // Begin recursive walk of phi nodes. 633 while( worklist.size() ){ 634 // Take a node off the worklist 635 Node *n = worklist.pop(); 636 if( !closure.member(n) ){ 637 // Add it to the closure. 638 closure.push(n); 639 // Make a sanity check to ensure we don't waste too much time here. 640 if( closure.size() > 20) return false; 641 // This node is OK if: 642 // - it is a cast of an identical value 643 // - or it is a phi node (then we add its inputs to the worklist) 644 // Otherwise, the node is not OK, and we presume the cast is not invariant 645 if( n->is_ConstraintCast() ){ 646 worklist.push(n->in(1)); 647 } else if( n->is_Phi() ) { 648 for( uint i = 1; i < n->req(); i++ ) { 649 worklist.push(n->in(i)); 650 } 651 } else { 652 return false; 653 } 654 } 655 } 656 657 // Quit when the worklist is empty, and we've found no offending nodes. 658 return true; 659 } 660 661 //------------------------------Ideal_DU_postCCP------------------------------- 662 // Find any cast-away of null-ness and keep its control. Null cast-aways are 663 // going away in this pass and we need to make this memory op depend on the 664 // gating null check. 665 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 666 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 667 } 668 669 // I tried to leave the CastPP's in. This makes the graph more accurate in 670 // some sense; we get to keep around the knowledge that an oop is not-null 671 // after some test. Alas, the CastPP's interfere with GVN (some values are 672 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 673 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 674 // some of the more trivial cases in the optimizer. Removing more useless 675 // Phi's started allowing Loads to illegally float above null checks. I gave 676 // up on this approach. CNC 10/20/2000 677 // This static method may be called not from MemNode (EncodePNode calls it). 678 // Only the control edge of the node 'n' might be updated. 679 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 680 Node *skipped_cast = NULL; 681 // Need a null check? Regular static accesses do not because they are 682 // from constant addresses. Array ops are gated by the range check (which 683 // always includes a NULL check). Just check field ops. 684 if( n->in(MemNode::Control) == NULL ) { 685 // Scan upwards for the highest location we can place this memory op. 686 while( true ) { 687 switch( adr->Opcode() ) { 688 689 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 690 adr = adr->in(AddPNode::Base); 691 continue; 692 693 case Op_DecodeN: // No change to NULL-ness, so peek thru 694 adr = adr->in(1); 695 continue; 696 697 case Op_CastPP: 698 // If the CastPP is useless, just peek on through it. 699 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 700 // Remember the cast that we've peeked though. If we peek 701 // through more than one, then we end up remembering the highest 702 // one, that is, if in a loop, the one closest to the top. 703 skipped_cast = adr; 704 adr = adr->in(1); 705 continue; 706 } 707 // CastPP is going away in this pass! We need this memory op to be 708 // control-dependent on the test that is guarding the CastPP. 709 ccp->hash_delete(n); 710 n->set_req(MemNode::Control, adr->in(0)); 711 ccp->hash_insert(n); 712 return n; 713 714 case Op_Phi: 715 // Attempt to float above a Phi to some dominating point. 716 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 717 // If we've already peeked through a Cast (which could have set the 718 // control), we can't float above a Phi, because the skipped Cast 719 // may not be loop invariant. 720 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 721 adr = adr->in(1); 722 continue; 723 } 724 } 725 726 // Intentional fallthrough! 727 728 // No obvious dominating point. The mem op is pinned below the Phi 729 // by the Phi itself. If the Phi goes away (no true value is merged) 730 // then the mem op can float, but not indefinitely. It must be pinned 731 // behind the controls leading to the Phi. 732 case Op_CheckCastPP: 733 // These usually stick around to change address type, however a 734 // useless one can be elided and we still need to pick up a control edge 735 if (adr->in(0) == NULL) { 736 // This CheckCastPP node has NO control and is likely useless. But we 737 // need check further up the ancestor chain for a control input to keep 738 // the node in place. 4959717. 739 skipped_cast = adr; 740 adr = adr->in(1); 741 continue; 742 } 743 ccp->hash_delete(n); 744 n->set_req(MemNode::Control, adr->in(0)); 745 ccp->hash_insert(n); 746 return n; 747 748 // List of "safe" opcodes; those that implicitly block the memory 749 // op below any null check. 750 case Op_CastX2P: // no null checks on native pointers 751 case Op_Parm: // 'this' pointer is not null 752 case Op_LoadP: // Loading from within a klass 753 case Op_LoadN: // Loading from within a klass 754 case Op_LoadKlass: // Loading from within a klass 755 case Op_LoadNKlass: // Loading from within a klass 756 case Op_ConP: // Loading from a klass 757 case Op_ConN: // Loading from a klass 758 case Op_CreateEx: // Sucking up the guts of an exception oop 759 case Op_Con: // Reading from TLS 760 case Op_CMoveP: // CMoveP is pinned 761 case Op_CMoveN: // CMoveN is pinned 762 break; // No progress 763 764 case Op_Proj: // Direct call to an allocation routine 765 case Op_SCMemProj: // Memory state from store conditional ops 766 #ifdef ASSERT 767 { 768 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 769 const Node* call = adr->in(0); 770 if (call->is_CallJava()) { 771 const CallJavaNode* call_java = call->as_CallJava(); 772 const TypeTuple *r = call_java->tf()->range(); 773 assert(r->cnt() > TypeFunc::Parms, "must return value"); 774 const Type* ret_type = r->field_at(TypeFunc::Parms); 775 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 776 // We further presume that this is one of 777 // new_instance_Java, new_array_Java, or 778 // the like, but do not assert for this. 779 } else if (call->is_Allocate()) { 780 // similar case to new_instance_Java, etc. 781 } else if (!call->is_CallLeaf()) { 782 // Projections from fetch_oop (OSR) are allowed as well. 783 ShouldNotReachHere(); 784 } 785 } 786 #endif 787 break; 788 default: 789 ShouldNotReachHere(); 790 } 791 break; 792 } 793 } 794 795 return NULL; // No progress 796 } 797 798 799 //============================================================================= 800 uint LoadNode::size_of() const { return sizeof(*this); } 801 uint LoadNode::cmp( const Node &n ) const 802 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 803 const Type *LoadNode::bottom_type() const { return _type; } 804 uint LoadNode::ideal_reg() const { 805 return Matcher::base2reg[_type->base()]; 806 } 807 808 #ifndef PRODUCT 809 void LoadNode::dump_spec(outputStream *st) const { 810 MemNode::dump_spec(st); 811 if( !Verbose && !WizardMode ) { 812 // standard dump does this in Verbose and WizardMode 813 st->print(" #"); _type->dump_on(st); 814 } 815 } 816 #endif 817 818 #ifdef ASSERT 819 //----------------------------is_immutable_value------------------------------- 820 // Helper function to allow a raw load without control edge for some cases 821 bool LoadNode::is_immutable_value(Node* adr) { 822 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 823 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 824 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 825 in_bytes(JavaThread::osthread_offset()))); 826 } 827 #endif 828 829 //----------------------------LoadNode::make----------------------------------- 830 // Polymorphic factory method: 831 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 832 Compile* C = gvn.C; 833 834 // sanity check the alias category against the created node type 835 assert(!(adr_type->isa_oopptr() && 836 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 837 "use LoadKlassNode instead"); 838 assert(!(adr_type->isa_aryptr() && 839 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 840 "use LoadRangeNode instead"); 841 // Check control edge of raw loads 842 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 843 // oop will be recorded in oop map if load crosses safepoint 844 rt->isa_oopptr() || is_immutable_value(adr), 845 "raw memory operations should have control edge"); 846 switch (bt) { 847 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 848 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 849 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 850 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 851 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 852 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 853 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 854 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 855 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 856 case T_OBJECT: 857 #ifdef _LP64 858 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 859 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 860 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 861 } else 862 #endif 863 { 864 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 865 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 866 } 867 } 868 ShouldNotReachHere(); 869 return (LoadNode*)NULL; 870 } 871 872 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 873 bool require_atomic = true; 874 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 875 } 876 877 878 879 880 //------------------------------hash------------------------------------------- 881 uint LoadNode::hash() const { 882 // unroll addition of interesting fields 883 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 884 } 885 886 //---------------------------can_see_stored_value------------------------------ 887 // This routine exists to make sure this set of tests is done the same 888 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 889 // will change the graph shape in a way which makes memory alive twice at the 890 // same time (uses the Oracle model of aliasing), then some 891 // LoadXNode::Identity will fold things back to the equivalence-class model 892 // of aliasing. 893 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 894 Node* ld_adr = in(MemNode::Address); 895 896 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 897 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 898 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 899 atp->field() != NULL && !atp->field()->is_volatile()) { 900 uint alias_idx = atp->index(); 901 bool final = atp->field()->is_final(); 902 Node* result = NULL; 903 Node* current = st; 904 // Skip through chains of MemBarNodes checking the MergeMems for 905 // new states for the slice of this load. Stop once any other 906 // kind of node is encountered. Loads from final memory can skip 907 // through any kind of MemBar but normal loads shouldn't skip 908 // through MemBarAcquire since the could allow them to move out of 909 // a synchronized region. 910 while (current->is_Proj()) { 911 int opc = current->in(0)->Opcode(); 912 if ((final && opc == Op_MemBarAcquire) || 913 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 914 Node* mem = current->in(0)->in(TypeFunc::Memory); 915 if (mem->is_MergeMem()) { 916 MergeMemNode* merge = mem->as_MergeMem(); 917 Node* new_st = merge->memory_at(alias_idx); 918 if (new_st == merge->base_memory()) { 919 // Keep searching 920 current = merge->base_memory(); 921 continue; 922 } 923 // Save the new memory state for the slice and fall through 924 // to exit. 925 result = new_st; 926 } 927 } 928 break; 929 } 930 if (result != NULL) { 931 st = result; 932 } 933 } 934 935 936 // Loop around twice in the case Load -> Initialize -> Store. 937 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 938 for (int trip = 0; trip <= 1; trip++) { 939 940 if (st->is_Store()) { 941 Node* st_adr = st->in(MemNode::Address); 942 if (!phase->eqv(st_adr, ld_adr)) { 943 // Try harder before giving up... Match raw and non-raw pointers. 944 intptr_t st_off = 0; 945 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 946 if (alloc == NULL) return NULL; 947 intptr_t ld_off = 0; 948 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 949 if (alloc != allo2) return NULL; 950 if (ld_off != st_off) return NULL; 951 // At this point we have proven something like this setup: 952 // A = Allocate(...) 953 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 954 // S = StoreQ(, AddP(, A.Parm , #Off), V) 955 // (Actually, we haven't yet proven the Q's are the same.) 956 // In other words, we are loading from a casted version of 957 // the same pointer-and-offset that we stored to. 958 // Thus, we are able to replace L by V. 959 } 960 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 961 if (store_Opcode() != st->Opcode()) 962 return NULL; 963 return st->in(MemNode::ValueIn); 964 } 965 966 intptr_t offset = 0; // scratch 967 968 // A load from a freshly-created object always returns zero. 969 // (This can happen after LoadNode::Ideal resets the load's memory input 970 // to find_captured_store, which returned InitializeNode::zero_memory.) 971 if (st->is_Proj() && st->in(0)->is_Allocate() && 972 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 973 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 974 // return a zero value for the load's basic type 975 // (This is one of the few places where a generic PhaseTransform 976 // can create new nodes. Think of it as lazily manifesting 977 // virtually pre-existing constants.) 978 return phase->zerocon(memory_type()); 979 } 980 981 // A load from an initialization barrier can match a captured store. 982 if (st->is_Proj() && st->in(0)->is_Initialize()) { 983 InitializeNode* init = st->in(0)->as_Initialize(); 984 AllocateNode* alloc = init->allocation(); 985 if (alloc != NULL && 986 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 987 // examine a captured store value 988 st = init->find_captured_store(offset, memory_size(), phase); 989 if (st != NULL) 990 continue; // take one more trip around 991 } 992 } 993 994 break; 995 } 996 997 return NULL; 998 } 999 1000 //----------------------is_instance_field_load_with_local_phi------------------ 1001 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1002 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 1003 in(MemNode::Address)->is_AddP() ) { 1004 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 1005 // Only instances. 1006 if( t_oop != NULL && t_oop->is_known_instance_field() && 1007 t_oop->offset() != Type::OffsetBot && 1008 t_oop->offset() != Type::OffsetTop) { 1009 return true; 1010 } 1011 } 1012 return false; 1013 } 1014 1015 //------------------------------Identity--------------------------------------- 1016 // Loads are identity if previous store is to same address 1017 Node *LoadNode::Identity( PhaseTransform *phase ) { 1018 // If the previous store-maker is the right kind of Store, and the store is 1019 // to the same address, then we are equal to the value stored. 1020 Node* mem = in(MemNode::Memory); 1021 Node* value = can_see_stored_value(mem, phase); 1022 if( value ) { 1023 // byte, short & char stores truncate naturally. 1024 // A load has to load the truncated value which requires 1025 // some sort of masking operation and that requires an 1026 // Ideal call instead of an Identity call. 1027 if (memory_size() < BytesPerInt) { 1028 // If the input to the store does not fit with the load's result type, 1029 // it must be truncated via an Ideal call. 1030 if (!phase->type(value)->higher_equal(phase->type(this))) 1031 return this; 1032 } 1033 // (This works even when value is a Con, but LoadNode::Value 1034 // usually runs first, producing the singleton type of the Con.) 1035 return value; 1036 } 1037 1038 // Search for an existing data phi which was generated before for the same 1039 // instance's field to avoid infinite generation of phis in a loop. 1040 Node *region = mem->in(0); 1041 if (is_instance_field_load_with_local_phi(region)) { 1042 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 1043 int this_index = phase->C->get_alias_index(addr_t); 1044 int this_offset = addr_t->offset(); 1045 int this_id = addr_t->is_oopptr()->instance_id(); 1046 const Type* this_type = bottom_type(); 1047 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1048 Node* phi = region->fast_out(i); 1049 if (phi->is_Phi() && phi != mem && 1050 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 1051 return phi; 1052 } 1053 } 1054 } 1055 1056 return this; 1057 } 1058 1059 1060 // Returns true if the AliasType refers to the field that holds the 1061 // cached box array. Currently only handles the IntegerCache case. 1062 static bool is_autobox_cache(Compile::AliasType* atp) { 1063 if (atp != NULL && atp->field() != NULL) { 1064 ciField* field = atp->field(); 1065 ciSymbol* klass = field->holder()->name(); 1066 if (field->name() == ciSymbol::cache_field_name() && 1067 field->holder()->uses_default_loader() && 1068 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1069 return true; 1070 } 1071 } 1072 return false; 1073 } 1074 1075 // Fetch the base value in the autobox array 1076 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1077 if (atp != NULL && atp->field() != NULL) { 1078 ciField* field = atp->field(); 1079 ciSymbol* klass = field->holder()->name(); 1080 if (field->name() == ciSymbol::cache_field_name() && 1081 field->holder()->uses_default_loader() && 1082 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1083 assert(field->is_constant(), "what?"); 1084 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1085 // Fetch the box object at the base of the array and get its value 1086 ciInstance* box = array->obj_at(0)->as_instance(); 1087 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1088 if (ik->nof_nonstatic_fields() == 1) { 1089 // This should be true nonstatic_field_at requires calling 1090 // nof_nonstatic_fields so check it anyway 1091 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1092 cache_offset = c.as_int(); 1093 } 1094 return true; 1095 } 1096 } 1097 return false; 1098 } 1099 1100 // Returns true if the AliasType refers to the value field of an 1101 // autobox object. Currently only handles Integer. 1102 static bool is_autobox_object(Compile::AliasType* atp) { 1103 if (atp != NULL && atp->field() != NULL) { 1104 ciField* field = atp->field(); 1105 ciSymbol* klass = field->holder()->name(); 1106 if (field->name() == ciSymbol::value_name() && 1107 field->holder()->uses_default_loader() && 1108 klass == ciSymbol::java_lang_Integer()) { 1109 return true; 1110 } 1111 } 1112 return false; 1113 } 1114 1115 1116 // We're loading from an object which has autobox behaviour. 1117 // If this object is result of a valueOf call we'll have a phi 1118 // merging a newly allocated object and a load from the cache. 1119 // We want to replace this load with the original incoming 1120 // argument to the valueOf call. 1121 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1122 Node* base = in(Address)->in(AddPNode::Base); 1123 if (base->is_Phi() && base->req() == 3) { 1124 AllocateNode* allocation = NULL; 1125 int allocation_index = -1; 1126 int load_index = -1; 1127 for (uint i = 1; i < base->req(); i++) { 1128 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1129 if (allocation != NULL) { 1130 allocation_index = i; 1131 load_index = 3 - allocation_index; 1132 break; 1133 } 1134 } 1135 bool has_load = ( allocation != NULL && 1136 (base->in(load_index)->is_Load() || 1137 base->in(load_index)->is_DecodeN() && 1138 base->in(load_index)->in(1)->is_Load()) ); 1139 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1140 // Push the loads from the phi that comes from valueOf up 1141 // through it to allow elimination of the loads and the recovery 1142 // of the original value. 1143 Node* mem_phi = in(Memory); 1144 Node* offset = in(Address)->in(AddPNode::Offset); 1145 Node* region = base->in(0); 1146 1147 Node* in1 = clone(); 1148 Node* in1_addr = in1->in(Address)->clone(); 1149 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1150 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1151 in1_addr->set_req(AddPNode::Offset, offset); 1152 in1->set_req(0, region->in(allocation_index)); 1153 in1->set_req(Address, in1_addr); 1154 in1->set_req(Memory, mem_phi->in(allocation_index)); 1155 1156 Node* in2 = clone(); 1157 Node* in2_addr = in2->in(Address)->clone(); 1158 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1159 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1160 in2_addr->set_req(AddPNode::Offset, offset); 1161 in2->set_req(0, region->in(load_index)); 1162 in2->set_req(Address, in2_addr); 1163 in2->set_req(Memory, mem_phi->in(load_index)); 1164 1165 in1_addr = phase->transform(in1_addr); 1166 in1 = phase->transform(in1); 1167 in2_addr = phase->transform(in2_addr); 1168 in2 = phase->transform(in2); 1169 1170 PhiNode* result = PhiNode::make_blank(region, this); 1171 result->set_req(allocation_index, in1); 1172 result->set_req(load_index, in2); 1173 return result; 1174 } 1175 } else if (base->is_Load() || 1176 base->is_DecodeN() && base->in(1)->is_Load()) { 1177 if (base->is_DecodeN()) { 1178 // Get LoadN node which loads cached Integer object 1179 base = base->in(1); 1180 } 1181 // Eliminate the load of Integer.value for integers from the cache 1182 // array by deriving the value from the index into the array. 1183 // Capture the offset of the load and then reverse the computation. 1184 Node* load_base = base->in(Address)->in(AddPNode::Base); 1185 if (load_base->is_DecodeN()) { 1186 // Get LoadN node which loads IntegerCache.cache field 1187 load_base = load_base->in(1); 1188 } 1189 if (load_base != NULL) { 1190 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1191 intptr_t cache_offset; 1192 int shift = -1; 1193 Node* cache = NULL; 1194 if (is_autobox_cache(atp)) { 1195 shift = exact_log2(type2aelembytes(T_OBJECT)); 1196 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1197 } 1198 if (cache != NULL && base->in(Address)->is_AddP()) { 1199 Node* elements[4]; 1200 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1201 int cache_low; 1202 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1203 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1204 // Add up all the offsets making of the address of the load 1205 Node* result = elements[0]; 1206 for (int i = 1; i < count; i++) { 1207 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1208 } 1209 // Remove the constant offset from the address and then 1210 // remove the scaling of the offset to recover the original index. 1211 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1212 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1213 // Peel the shift off directly but wrap it in a dummy node 1214 // since Ideal can't return existing nodes 1215 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1216 } else { 1217 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1218 } 1219 #ifdef _LP64 1220 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1221 #endif 1222 return result; 1223 } 1224 } 1225 } 1226 } 1227 return NULL; 1228 } 1229 1230 //------------------------------split_through_phi------------------------------ 1231 // Split instance field load through Phi. 1232 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1233 Node* mem = in(MemNode::Memory); 1234 Node* address = in(MemNode::Address); 1235 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1236 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1237 1238 assert(mem->is_Phi() && (t_oop != NULL) && 1239 t_oop->is_known_instance_field(), "invalide conditions"); 1240 1241 Node *region = mem->in(0); 1242 if (region == NULL) { 1243 return NULL; // Wait stable graph 1244 } 1245 uint cnt = mem->req(); 1246 for( uint i = 1; i < cnt; i++ ) { 1247 Node *in = mem->in(i); 1248 if( in == NULL ) { 1249 return NULL; // Wait stable graph 1250 } 1251 } 1252 // Check for loop invariant. 1253 if (cnt == 3) { 1254 for( uint i = 1; i < cnt; i++ ) { 1255 Node *in = mem->in(i); 1256 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1257 if (m == mem) { 1258 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1259 return this; 1260 } 1261 } 1262 } 1263 // Split through Phi (see original code in loopopts.cpp). 1264 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1265 1266 // Do nothing here if Identity will find a value 1267 // (to avoid infinite chain of value phis generation). 1268 if ( !phase->eqv(this, this->Identity(phase)) ) 1269 return NULL; 1270 1271 // Skip the split if the region dominates some control edge of the address. 1272 if (cnt == 3 && !MemNode::all_controls_dominate(address, region)) 1273 return NULL; 1274 1275 const Type* this_type = this->bottom_type(); 1276 int this_index = phase->C->get_alias_index(addr_t); 1277 int this_offset = addr_t->offset(); 1278 int this_iid = addr_t->is_oopptr()->instance_id(); 1279 int wins = 0; 1280 PhaseIterGVN *igvn = phase->is_IterGVN(); 1281 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1282 for( uint i = 1; i < region->req(); i++ ) { 1283 Node *x; 1284 Node* the_clone = NULL; 1285 if( region->in(i) == phase->C->top() ) { 1286 x = phase->C->top(); // Dead path? Use a dead data op 1287 } else { 1288 x = this->clone(); // Else clone up the data op 1289 the_clone = x; // Remember for possible deletion. 1290 // Alter data node to use pre-phi inputs 1291 if( this->in(0) == region ) { 1292 x->set_req( 0, region->in(i) ); 1293 } else { 1294 x->set_req( 0, NULL ); 1295 } 1296 for( uint j = 1; j < this->req(); j++ ) { 1297 Node *in = this->in(j); 1298 if( in->is_Phi() && in->in(0) == region ) 1299 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 1300 } 1301 } 1302 // Check for a 'win' on some paths 1303 const Type *t = x->Value(igvn); 1304 1305 bool singleton = t->singleton(); 1306 1307 // See comments in PhaseIdealLoop::split_thru_phi(). 1308 if( singleton && t == Type::TOP ) { 1309 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1310 } 1311 1312 if( singleton ) { 1313 wins++; 1314 x = igvn->makecon(t); 1315 } else { 1316 // We now call Identity to try to simplify the cloned node. 1317 // Note that some Identity methods call phase->type(this). 1318 // Make sure that the type array is big enough for 1319 // our new node, even though we may throw the node away. 1320 // (This tweaking with igvn only works because x is a new node.) 1321 igvn->set_type(x, t); 1322 // If x is a TypeNode, capture any more-precise type permanently into Node 1323 // otherwise it will be not updated during igvn->transform since 1324 // igvn->type(x) is set to x->Value() already. 1325 x->raise_bottom_type(t); 1326 Node *y = x->Identity(igvn); 1327 if( y != x ) { 1328 wins++; 1329 x = y; 1330 } else { 1331 y = igvn->hash_find(x); 1332 if( y ) { 1333 wins++; 1334 x = y; 1335 } else { 1336 // Else x is a new node we are keeping 1337 // We do not need register_new_node_with_optimizer 1338 // because set_type has already been called. 1339 igvn->_worklist.push(x); 1340 } 1341 } 1342 } 1343 if (x != the_clone && the_clone != NULL) 1344 igvn->remove_dead_node(the_clone); 1345 phi->set_req(i, x); 1346 } 1347 if( wins > 0 ) { 1348 // Record Phi 1349 igvn->register_new_node_with_optimizer(phi); 1350 return phi; 1351 } 1352 igvn->remove_dead_node(phi); 1353 return NULL; 1354 } 1355 1356 //------------------------------Ideal------------------------------------------ 1357 // If the load is from Field memory and the pointer is non-null, we can 1358 // zero out the control input. 1359 // If the offset is constant and the base is an object allocation, 1360 // try to hook me up to the exact initializing store. 1361 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1362 Node* p = MemNode::Ideal_common(phase, can_reshape); 1363 if (p) return (p == NodeSentinel) ? NULL : p; 1364 1365 Node* ctrl = in(MemNode::Control); 1366 Node* address = in(MemNode::Address); 1367 1368 // Skip up past a SafePoint control. Cannot do this for Stores because 1369 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1370 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1371 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1372 ctrl = ctrl->in(0); 1373 set_req(MemNode::Control,ctrl); 1374 } 1375 1376 intptr_t ignore = 0; 1377 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1378 if (base != NULL 1379 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1380 // Check for useless control edge in some common special cases 1381 if (in(MemNode::Control) != NULL 1382 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1383 && all_controls_dominate(base, phase->C->start())) { 1384 // A method-invariant, non-null address (constant or 'this' argument). 1385 set_req(MemNode::Control, NULL); 1386 } 1387 1388 if (EliminateAutoBox && can_reshape) { 1389 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); 1390 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1391 if (is_autobox_object(atp)) { 1392 Node* result = eliminate_autobox(phase); 1393 if (result != NULL) return result; 1394 } 1395 } 1396 } 1397 1398 Node* mem = in(MemNode::Memory); 1399 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1400 1401 if (addr_t != NULL) { 1402 // try to optimize our memory input 1403 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1404 if (opt_mem != mem) { 1405 set_req(MemNode::Memory, opt_mem); 1406 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1407 return this; 1408 } 1409 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1410 if (can_reshape && opt_mem->is_Phi() && 1411 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1412 // Split instance field load through Phi. 1413 Node* result = split_through_phi(phase); 1414 if (result != NULL) return result; 1415 } 1416 } 1417 1418 // Check for prior store with a different base or offset; make Load 1419 // independent. Skip through any number of them. Bail out if the stores 1420 // are in an endless dead cycle and report no progress. This is a key 1421 // transform for Reflection. However, if after skipping through the Stores 1422 // we can't then fold up against a prior store do NOT do the transform as 1423 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1424 // array memory alive twice: once for the hoisted Load and again after the 1425 // bypassed Store. This situation only works if EVERYBODY who does 1426 // anti-dependence work knows how to bypass. I.e. we need all 1427 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1428 // the alias index stuff. So instead, peek through Stores and IFF we can 1429 // fold up, do so. 1430 Node* prev_mem = find_previous_store(phase); 1431 // Steps (a), (b): Walk past independent stores to find an exact match. 1432 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1433 // (c) See if we can fold up on the spot, but don't fold up here. 1434 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1435 // just return a prior value, which is done by Identity calls. 1436 if (can_see_stored_value(prev_mem, phase)) { 1437 // Make ready for step (d): 1438 set_req(MemNode::Memory, prev_mem); 1439 return this; 1440 } 1441 } 1442 1443 return NULL; // No further progress 1444 } 1445 1446 // Helper to recognize certain Klass fields which are invariant across 1447 // some group of array types (e.g., int[] or all T[] where T < Object). 1448 const Type* 1449 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1450 ciKlass* klass) const { 1451 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1452 // The field is Klass::_modifier_flags. Return its (constant) value. 1453 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1454 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1455 return TypeInt::make(klass->modifier_flags()); 1456 } 1457 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1458 // The field is Klass::_access_flags. Return its (constant) value. 1459 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1460 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1461 return TypeInt::make(klass->access_flags()); 1462 } 1463 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1464 // The field is Klass::_layout_helper. Return its constant value if known. 1465 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1466 return TypeInt::make(klass->layout_helper()); 1467 } 1468 1469 // No match. 1470 return NULL; 1471 } 1472 1473 //------------------------------Value----------------------------------------- 1474 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1475 // Either input is TOP ==> the result is TOP 1476 Node* mem = in(MemNode::Memory); 1477 const Type *t1 = phase->type(mem); 1478 if (t1 == Type::TOP) return Type::TOP; 1479 Node* adr = in(MemNode::Address); 1480 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1481 if (tp == NULL || tp->empty()) return Type::TOP; 1482 int off = tp->offset(); 1483 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1484 1485 // Try to guess loaded type from pointer type 1486 if (tp->base() == Type::AryPtr) { 1487 const Type *t = tp->is_aryptr()->elem(); 1488 // Don't do this for integer types. There is only potential profit if 1489 // the element type t is lower than _type; that is, for int types, if _type is 1490 // more restrictive than t. This only happens here if one is short and the other 1491 // char (both 16 bits), and in those cases we've made an intentional decision 1492 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1493 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1494 // 1495 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1496 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1497 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1498 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1499 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1500 // In fact, that could have been the original type of p1, and p1 could have 1501 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1502 // expression (LShiftL quux 3) independently optimized to the constant 8. 1503 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1504 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1505 // t might actually be lower than _type, if _type is a unique 1506 // concrete subclass of abstract class t. 1507 // Make sure the reference is not into the header, by comparing 1508 // the offset against the offset of the start of the array's data. 1509 // Different array types begin at slightly different offsets (12 vs. 16). 1510 // We choose T_BYTE as an example base type that is least restrictive 1511 // as to alignment, which will therefore produce the smallest 1512 // possible base offset. 1513 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1514 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1515 const Type* jt = t->join(_type); 1516 // In any case, do not allow the join, per se, to empty out the type. 1517 if (jt->empty() && !t->empty()) { 1518 // This can happen if a interface-typed array narrows to a class type. 1519 jt = _type; 1520 } 1521 1522 if (EliminateAutoBox && adr->is_AddP()) { 1523 // The pointers in the autobox arrays are always non-null 1524 Node* base = adr->in(AddPNode::Base); 1525 if (base != NULL && 1526 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { 1527 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1528 if (is_autobox_cache(atp)) { 1529 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1530 } 1531 } 1532 } 1533 return jt; 1534 } 1535 } 1536 } else if (tp->base() == Type::InstPtr) { 1537 const TypeInstPtr* tinst = tp->is_instptr(); 1538 ciKlass* klass = tinst->klass(); 1539 assert( off != Type::OffsetBot || 1540 // arrays can be cast to Objects 1541 tp->is_oopptr()->klass()->is_java_lang_Object() || 1542 // unsafe field access may not have a constant offset 1543 phase->C->has_unsafe_access(), 1544 "Field accesses must be precise" ); 1545 // For oop loads, we expect the _type to be precise 1546 if (OptimizeStringConcat && klass == phase->C->env()->String_klass() && 1547 adr->is_AddP() && off != Type::OffsetBot) { 1548 // For constant Strings treat the fields as compile time constants. 1549 Node* base = adr->in(AddPNode::Base); 1550 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1551 if (t != NULL && t->singleton()) { 1552 ciObject* string = t->const_oop(); 1553 ciConstant constant = string->as_instance()->field_value_by_offset(off); 1554 if (constant.basic_type() == T_INT) { 1555 return TypeInt::make(constant.as_int()); 1556 } else if (constant.basic_type() == T_ARRAY) { 1557 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1558 return TypeNarrowOop::make_from_constant(constant.as_object()); 1559 } else { 1560 return TypeOopPtr::make_from_constant(constant.as_object()); 1561 } 1562 } 1563 } 1564 } 1565 } else if (tp->base() == Type::KlassPtr) { 1566 assert( off != Type::OffsetBot || 1567 // arrays can be cast to Objects 1568 tp->is_klassptr()->klass()->is_java_lang_Object() || 1569 // also allow array-loading from the primary supertype 1570 // array during subtype checks 1571 Opcode() == Op_LoadKlass, 1572 "Field accesses must be precise" ); 1573 // For klass/static loads, we expect the _type to be precise 1574 } 1575 1576 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1577 if (tkls != NULL && !StressReflectiveCode) { 1578 ciKlass* klass = tkls->klass(); 1579 if (klass->is_loaded() && tkls->klass_is_exact()) { 1580 // We are loading a field from a Klass metaobject whose identity 1581 // is known at compile time (the type is "exact" or "precise"). 1582 // Check for fields we know are maintained as constants by the VM. 1583 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1584 // The field is Klass::_super_check_offset. Return its (constant) value. 1585 // (Folds up type checking code.) 1586 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1587 return TypeInt::make(klass->super_check_offset()); 1588 } 1589 // Compute index into primary_supers array 1590 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1591 // Check for overflowing; use unsigned compare to handle the negative case. 1592 if( depth < ciKlass::primary_super_limit() ) { 1593 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1594 // (Folds up type checking code.) 1595 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1596 ciKlass *ss = klass->super_of_depth(depth); 1597 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1598 } 1599 const Type* aift = load_array_final_field(tkls, klass); 1600 if (aift != NULL) return aift; 1601 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1602 && klass->is_array_klass()) { 1603 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1604 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1605 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1606 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1607 } 1608 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1609 // The field is Klass::_java_mirror. Return its (constant) value. 1610 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1611 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1612 return TypeInstPtr::make(klass->java_mirror()); 1613 } 1614 } 1615 1616 // We can still check if we are loading from the primary_supers array at a 1617 // shallow enough depth. Even though the klass is not exact, entries less 1618 // than or equal to its super depth are correct. 1619 if (klass->is_loaded() ) { 1620 ciType *inner = klass->klass(); 1621 while( inner->is_obj_array_klass() ) 1622 inner = inner->as_obj_array_klass()->base_element_type(); 1623 if( inner->is_instance_klass() && 1624 !inner->as_instance_klass()->flags().is_interface() ) { 1625 // Compute index into primary_supers array 1626 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1627 // Check for overflowing; use unsigned compare to handle the negative case. 1628 if( depth < ciKlass::primary_super_limit() && 1629 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1630 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1631 // (Folds up type checking code.) 1632 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1633 ciKlass *ss = klass->super_of_depth(depth); 1634 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1635 } 1636 } 1637 } 1638 1639 // If the type is enough to determine that the thing is not an array, 1640 // we can give the layout_helper a positive interval type. 1641 // This will help short-circuit some reflective code. 1642 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1643 && !klass->is_array_klass() // not directly typed as an array 1644 && !klass->is_interface() // specifically not Serializable & Cloneable 1645 && !klass->is_java_lang_Object() // not the supertype of all T[] 1646 ) { 1647 // Note: When interfaces are reliable, we can narrow the interface 1648 // test to (klass != Serializable && klass != Cloneable). 1649 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1650 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1651 // The key property of this type is that it folds up tests 1652 // for array-ness, since it proves that the layout_helper is positive. 1653 // Thus, a generic value like the basic object layout helper works fine. 1654 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1655 } 1656 } 1657 1658 // If we are loading from a freshly-allocated object, produce a zero, 1659 // if the load is provably beyond the header of the object. 1660 // (Also allow a variable load from a fresh array to produce zero.) 1661 if (ReduceFieldZeroing) { 1662 Node* value = can_see_stored_value(mem,phase); 1663 if (value != NULL && value->is_Con()) 1664 return value->bottom_type(); 1665 } 1666 1667 const TypeOopPtr *tinst = tp->isa_oopptr(); 1668 if (tinst != NULL && tinst->is_known_instance_field()) { 1669 // If we have an instance type and our memory input is the 1670 // programs's initial memory state, there is no matching store, 1671 // so just return a zero of the appropriate type 1672 Node *mem = in(MemNode::Memory); 1673 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1674 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1675 return Type::get_zero_type(_type->basic_type()); 1676 } 1677 } 1678 return _type; 1679 } 1680 1681 //------------------------------match_edge------------------------------------- 1682 // Do we Match on this edge index or not? Match only the address. 1683 uint LoadNode::match_edge(uint idx) const { 1684 return idx == MemNode::Address; 1685 } 1686 1687 //--------------------------LoadBNode::Ideal-------------------------------------- 1688 // 1689 // If the previous store is to the same address as this load, 1690 // and the value stored was larger than a byte, replace this load 1691 // with the value stored truncated to a byte. If no truncation is 1692 // needed, the replacement is done in LoadNode::Identity(). 1693 // 1694 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1695 Node* mem = in(MemNode::Memory); 1696 Node* value = can_see_stored_value(mem,phase); 1697 if( value && !phase->type(value)->higher_equal( _type ) ) { 1698 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1699 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1700 } 1701 // Identity call will handle the case where truncation is not needed. 1702 return LoadNode::Ideal(phase, can_reshape); 1703 } 1704 1705 //--------------------------LoadUBNode::Ideal------------------------------------- 1706 // 1707 // If the previous store is to the same address as this load, 1708 // and the value stored was larger than a byte, replace this load 1709 // with the value stored truncated to a byte. If no truncation is 1710 // needed, the replacement is done in LoadNode::Identity(). 1711 // 1712 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1713 Node* mem = in(MemNode::Memory); 1714 Node* value = can_see_stored_value(mem, phase); 1715 if (value && !phase->type(value)->higher_equal(_type)) 1716 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); 1717 // Identity call will handle the case where truncation is not needed. 1718 return LoadNode::Ideal(phase, can_reshape); 1719 } 1720 1721 //--------------------------LoadUSNode::Ideal------------------------------------- 1722 // 1723 // If the previous store is to the same address as this load, 1724 // and the value stored was larger than a char, replace this load 1725 // with the value stored truncated to a char. If no truncation is 1726 // needed, the replacement is done in LoadNode::Identity(). 1727 // 1728 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1729 Node* mem = in(MemNode::Memory); 1730 Node* value = can_see_stored_value(mem,phase); 1731 if( value && !phase->type(value)->higher_equal( _type ) ) 1732 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1733 // Identity call will handle the case where truncation is not needed. 1734 return LoadNode::Ideal(phase, can_reshape); 1735 } 1736 1737 //--------------------------LoadSNode::Ideal-------------------------------------- 1738 // 1739 // If the previous store is to the same address as this load, 1740 // and the value stored was larger than a short, replace this load 1741 // with the value stored truncated to a short. If no truncation is 1742 // needed, the replacement is done in LoadNode::Identity(). 1743 // 1744 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1745 Node* mem = in(MemNode::Memory); 1746 Node* value = can_see_stored_value(mem,phase); 1747 if( value && !phase->type(value)->higher_equal( _type ) ) { 1748 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1749 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1750 } 1751 // Identity call will handle the case where truncation is not needed. 1752 return LoadNode::Ideal(phase, can_reshape); 1753 } 1754 1755 //============================================================================= 1756 //----------------------------LoadKlassNode::make------------------------------ 1757 // Polymorphic factory method: 1758 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1759 Compile* C = gvn.C; 1760 Node *ctl = NULL; 1761 // sanity check the alias category against the created node type 1762 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1763 assert(adr_type != NULL, "expecting TypeOopPtr"); 1764 #ifdef _LP64 1765 if (adr_type->is_ptr_to_narrowoop()) { 1766 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1767 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1768 } 1769 #endif 1770 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1771 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1772 } 1773 1774 //------------------------------Value------------------------------------------ 1775 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1776 return klass_value_common(phase); 1777 } 1778 1779 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1780 // Either input is TOP ==> the result is TOP 1781 const Type *t1 = phase->type( in(MemNode::Memory) ); 1782 if (t1 == Type::TOP) return Type::TOP; 1783 Node *adr = in(MemNode::Address); 1784 const Type *t2 = phase->type( adr ); 1785 if (t2 == Type::TOP) return Type::TOP; 1786 const TypePtr *tp = t2->is_ptr(); 1787 if (TypePtr::above_centerline(tp->ptr()) || 1788 tp->ptr() == TypePtr::Null) return Type::TOP; 1789 1790 // Return a more precise klass, if possible 1791 const TypeInstPtr *tinst = tp->isa_instptr(); 1792 if (tinst != NULL) { 1793 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1794 int offset = tinst->offset(); 1795 if (ik == phase->C->env()->Class_klass() 1796 && (offset == java_lang_Class::klass_offset_in_bytes() || 1797 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1798 // We are loading a special hidden field from a Class mirror object, 1799 // the field which points to the VM's Klass metaobject. 1800 ciType* t = tinst->java_mirror_type(); 1801 // java_mirror_type returns non-null for compile-time Class constants. 1802 if (t != NULL) { 1803 // constant oop => constant klass 1804 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1805 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1806 } 1807 if (!t->is_klass()) { 1808 // a primitive Class (e.g., int.class) has NULL for a klass field 1809 return TypePtr::NULL_PTR; 1810 } 1811 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1812 return TypeKlassPtr::make(t->as_klass()); 1813 } 1814 // non-constant mirror, so we can't tell what's going on 1815 } 1816 if( !ik->is_loaded() ) 1817 return _type; // Bail out if not loaded 1818 if (offset == oopDesc::klass_offset_in_bytes()) { 1819 if (tinst->klass_is_exact()) { 1820 return TypeKlassPtr::make(ik); 1821 } 1822 // See if we can become precise: no subklasses and no interface 1823 // (Note: We need to support verified interfaces.) 1824 if (!ik->is_interface() && !ik->has_subklass()) { 1825 //assert(!UseExactTypes, "this code should be useless with exact types"); 1826 // Add a dependence; if any subclass added we need to recompile 1827 if (!ik->is_final()) { 1828 // %%% should use stronger assert_unique_concrete_subtype instead 1829 phase->C->dependencies()->assert_leaf_type(ik); 1830 } 1831 // Return precise klass 1832 return TypeKlassPtr::make(ik); 1833 } 1834 1835 // Return root of possible klass 1836 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1837 } 1838 } 1839 1840 // Check for loading klass from an array 1841 const TypeAryPtr *tary = tp->isa_aryptr(); 1842 if( tary != NULL ) { 1843 ciKlass *tary_klass = tary->klass(); 1844 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1845 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1846 if (tary->klass_is_exact()) { 1847 return TypeKlassPtr::make(tary_klass); 1848 } 1849 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1850 // If the klass is an object array, we defer the question to the 1851 // array component klass. 1852 if( ak->is_obj_array_klass() ) { 1853 assert( ak->is_loaded(), "" ); 1854 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1855 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1856 ciInstanceKlass* ik = base_k->as_instance_klass(); 1857 // See if we can become precise: no subklasses and no interface 1858 if (!ik->is_interface() && !ik->has_subklass()) { 1859 //assert(!UseExactTypes, "this code should be useless with exact types"); 1860 // Add a dependence; if any subclass added we need to recompile 1861 if (!ik->is_final()) { 1862 phase->C->dependencies()->assert_leaf_type(ik); 1863 } 1864 // Return precise array klass 1865 return TypeKlassPtr::make(ak); 1866 } 1867 } 1868 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1869 } else { // Found a type-array? 1870 //assert(!UseExactTypes, "this code should be useless with exact types"); 1871 assert( ak->is_type_array_klass(), "" ); 1872 return TypeKlassPtr::make(ak); // These are always precise 1873 } 1874 } 1875 } 1876 1877 // Check for loading klass from an array klass 1878 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1879 if (tkls != NULL && !StressReflectiveCode) { 1880 ciKlass* klass = tkls->klass(); 1881 if( !klass->is_loaded() ) 1882 return _type; // Bail out if not loaded 1883 if( klass->is_obj_array_klass() && 1884 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1885 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1886 // // Always returning precise element type is incorrect, 1887 // // e.g., element type could be object and array may contain strings 1888 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1889 1890 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1891 // according to the element type's subclassing. 1892 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1893 } 1894 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1895 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1896 ciKlass* sup = klass->as_instance_klass()->super(); 1897 // The field is Klass::_super. Return its (constant) value. 1898 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1899 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1900 } 1901 } 1902 1903 // Bailout case 1904 return LoadNode::Value(phase); 1905 } 1906 1907 //------------------------------Identity--------------------------------------- 1908 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1909 // Also feed through the klass in Allocate(...klass...)._klass. 1910 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1911 return klass_identity_common(phase); 1912 } 1913 1914 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 1915 Node* x = LoadNode::Identity(phase); 1916 if (x != this) return x; 1917 1918 // Take apart the address into an oop and and offset. 1919 // Return 'this' if we cannot. 1920 Node* adr = in(MemNode::Address); 1921 intptr_t offset = 0; 1922 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1923 if (base == NULL) return this; 1924 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1925 if (toop == NULL) return this; 1926 1927 // We can fetch the klass directly through an AllocateNode. 1928 // This works even if the klass is not constant (clone or newArray). 1929 if (offset == oopDesc::klass_offset_in_bytes()) { 1930 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1931 if (allocated_klass != NULL) { 1932 return allocated_klass; 1933 } 1934 } 1935 1936 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1937 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1938 // See inline_native_Class_query for occurrences of these patterns. 1939 // Java Example: x.getClass().isAssignableFrom(y) 1940 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1941 // 1942 // This improves reflective code, often making the Class 1943 // mirror go completely dead. (Current exception: Class 1944 // mirrors may appear in debug info, but we could clean them out by 1945 // introducing a new debug info operator for klassOop.java_mirror). 1946 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1947 && (offset == java_lang_Class::klass_offset_in_bytes() || 1948 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1949 // We are loading a special hidden field from a Class mirror, 1950 // the field which points to its Klass or arrayKlass metaobject. 1951 if (base->is_Load()) { 1952 Node* adr2 = base->in(MemNode::Address); 1953 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1954 if (tkls != NULL && !tkls->empty() 1955 && (tkls->klass()->is_instance_klass() || 1956 tkls->klass()->is_array_klass()) 1957 && adr2->is_AddP() 1958 ) { 1959 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1960 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1961 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1962 } 1963 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1964 return adr2->in(AddPNode::Base); 1965 } 1966 } 1967 } 1968 } 1969 1970 return this; 1971 } 1972 1973 1974 //------------------------------Value------------------------------------------ 1975 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 1976 const Type *t = klass_value_common(phase); 1977 if (t == Type::TOP) 1978 return t; 1979 1980 return t->make_narrowoop(); 1981 } 1982 1983 //------------------------------Identity--------------------------------------- 1984 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 1985 // Also feed through the klass in Allocate(...klass...)._klass. 1986 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 1987 Node *x = klass_identity_common(phase); 1988 1989 const Type *t = phase->type( x ); 1990 if( t == Type::TOP ) return x; 1991 if( t->isa_narrowoop()) return x; 1992 1993 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 1994 } 1995 1996 //------------------------------Value----------------------------------------- 1997 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1998 // Either input is TOP ==> the result is TOP 1999 const Type *t1 = phase->type( in(MemNode::Memory) ); 2000 if( t1 == Type::TOP ) return Type::TOP; 2001 Node *adr = in(MemNode::Address); 2002 const Type *t2 = phase->type( adr ); 2003 if( t2 == Type::TOP ) return Type::TOP; 2004 const TypePtr *tp = t2->is_ptr(); 2005 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2006 const TypeAryPtr *tap = tp->isa_aryptr(); 2007 if( !tap ) return _type; 2008 return tap->size(); 2009 } 2010 2011 //-------------------------------Ideal--------------------------------------- 2012 // Feed through the length in AllocateArray(...length...)._length. 2013 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2014 Node* p = MemNode::Ideal_common(phase, can_reshape); 2015 if (p) return (p == NodeSentinel) ? NULL : p; 2016 2017 // Take apart the address into an oop and and offset. 2018 // Return 'this' if we cannot. 2019 Node* adr = in(MemNode::Address); 2020 intptr_t offset = 0; 2021 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2022 if (base == NULL) return NULL; 2023 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2024 if (tary == NULL) return NULL; 2025 2026 // We can fetch the length directly through an AllocateArrayNode. 2027 // This works even if the length is not constant (clone or newArray). 2028 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2029 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2030 if (alloc != NULL) { 2031 Node* allocated_length = alloc->Ideal_length(); 2032 Node* len = alloc->make_ideal_length(tary, phase); 2033 if (allocated_length != len) { 2034 // New CastII improves on this. 2035 return len; 2036 } 2037 } 2038 } 2039 2040 return NULL; 2041 } 2042 2043 //------------------------------Identity--------------------------------------- 2044 // Feed through the length in AllocateArray(...length...)._length. 2045 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2046 Node* x = LoadINode::Identity(phase); 2047 if (x != this) return x; 2048 2049 // Take apart the address into an oop and and offset. 2050 // Return 'this' if we cannot. 2051 Node* adr = in(MemNode::Address); 2052 intptr_t offset = 0; 2053 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2054 if (base == NULL) return this; 2055 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2056 if (tary == NULL) return this; 2057 2058 // We can fetch the length directly through an AllocateArrayNode. 2059 // This works even if the length is not constant (clone or newArray). 2060 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2061 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2062 if (alloc != NULL) { 2063 Node* allocated_length = alloc->Ideal_length(); 2064 // Do not allow make_ideal_length to allocate a CastII node. 2065 Node* len = alloc->make_ideal_length(tary, phase, false); 2066 if (allocated_length == len) { 2067 // Return allocated_length only if it would not be improved by a CastII. 2068 return allocated_length; 2069 } 2070 } 2071 } 2072 2073 return this; 2074 2075 } 2076 2077 //============================================================================= 2078 //---------------------------StoreNode::make----------------------------------- 2079 // Polymorphic factory method: 2080 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 2081 Compile* C = gvn.C; 2082 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2083 ctl != NULL, "raw memory operations should have control edge"); 2084 2085 switch (bt) { 2086 case T_BOOLEAN: 2087 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 2088 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 2089 case T_CHAR: 2090 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 2091 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 2092 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 2093 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 2094 case T_ADDRESS: 2095 case T_OBJECT: 2096 #ifdef _LP64 2097 if (adr->bottom_type()->is_ptr_to_narrowoop() || 2098 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 2099 adr->bottom_type()->isa_rawptr())) { 2100 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2101 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 2102 } else 2103 #endif 2104 { 2105 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 2106 } 2107 } 2108 ShouldNotReachHere(); 2109 return (StoreNode*)NULL; 2110 } 2111 2112 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2113 bool require_atomic = true; 2114 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2115 } 2116 2117 2118 //--------------------------bottom_type---------------------------------------- 2119 const Type *StoreNode::bottom_type() const { 2120 return Type::MEMORY; 2121 } 2122 2123 //------------------------------hash------------------------------------------- 2124 uint StoreNode::hash() const { 2125 // unroll addition of interesting fields 2126 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2127 2128 // Since they are not commoned, do not hash them: 2129 return NO_HASH; 2130 } 2131 2132 //------------------------------Ideal------------------------------------------ 2133 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2134 // When a store immediately follows a relevant allocation/initialization, 2135 // try to capture it into the initialization, or hoist it above. 2136 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2137 Node* p = MemNode::Ideal_common(phase, can_reshape); 2138 if (p) return (p == NodeSentinel) ? NULL : p; 2139 2140 Node* mem = in(MemNode::Memory); 2141 Node* address = in(MemNode::Address); 2142 2143 // Back-to-back stores to same address? Fold em up. 2144 // Generally unsafe if I have intervening uses... 2145 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 2146 // Looking at a dead closed cycle of memory? 2147 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2148 2149 assert(Opcode() == mem->Opcode() || 2150 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2151 "no mismatched stores, except on raw memory"); 2152 2153 if (mem->outcnt() == 1 && // check for intervening uses 2154 mem->as_Store()->memory_size() <= this->memory_size()) { 2155 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2156 // For example, 'mem' might be the final state at a conditional return. 2157 // Or, 'mem' might be used by some node which is live at the same time 2158 // 'this' is live, which might be unschedulable. So, require exactly 2159 // ONE user, the 'this' store, until such time as we clone 'mem' for 2160 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2161 if (can_reshape) { // (%%% is this an anachronism?) 2162 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2163 phase->is_IterGVN()); 2164 } else { 2165 // It's OK to do this in the parser, since DU info is always accurate, 2166 // and the parser always refers to nodes via SafePointNode maps. 2167 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2168 } 2169 return this; 2170 } 2171 } 2172 2173 // Capture an unaliased, unconditional, simple store into an initializer. 2174 // Or, if it is independent of the allocation, hoist it above the allocation. 2175 if (ReduceFieldZeroing && /*can_reshape &&*/ 2176 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2177 InitializeNode* init = mem->in(0)->as_Initialize(); 2178 intptr_t offset = init->can_capture_store(this, phase); 2179 if (offset > 0) { 2180 Node* moved = init->capture_store(this, offset, phase); 2181 // If the InitializeNode captured me, it made a raw copy of me, 2182 // and I need to disappear. 2183 if (moved != NULL) { 2184 // %%% hack to ensure that Ideal returns a new node: 2185 mem = MergeMemNode::make(phase->C, mem); 2186 return mem; // fold me away 2187 } 2188 } 2189 } 2190 2191 return NULL; // No further progress 2192 } 2193 2194 //------------------------------Value----------------------------------------- 2195 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2196 // Either input is TOP ==> the result is TOP 2197 const Type *t1 = phase->type( in(MemNode::Memory) ); 2198 if( t1 == Type::TOP ) return Type::TOP; 2199 const Type *t2 = phase->type( in(MemNode::Address) ); 2200 if( t2 == Type::TOP ) return Type::TOP; 2201 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2202 if( t3 == Type::TOP ) return Type::TOP; 2203 return Type::MEMORY; 2204 } 2205 2206 //------------------------------Identity--------------------------------------- 2207 // Remove redundant stores: 2208 // Store(m, p, Load(m, p)) changes to m. 2209 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2210 Node *StoreNode::Identity( PhaseTransform *phase ) { 2211 Node* mem = in(MemNode::Memory); 2212 Node* adr = in(MemNode::Address); 2213 Node* val = in(MemNode::ValueIn); 2214 2215 // Load then Store? Then the Store is useless 2216 if (val->is_Load() && 2217 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 2218 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 2219 val->as_Load()->store_Opcode() == Opcode()) { 2220 return mem; 2221 } 2222 2223 // Two stores in a row of the same value? 2224 if (mem->is_Store() && 2225 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 2226 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 2227 mem->Opcode() == Opcode()) { 2228 return mem; 2229 } 2230 2231 // Store of zero anywhere into a freshly-allocated object? 2232 // Then the store is useless. 2233 // (It must already have been captured by the InitializeNode.) 2234 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2235 // a newly allocated object is already all-zeroes everywhere 2236 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2237 return mem; 2238 } 2239 2240 // the store may also apply to zero-bits in an earlier object 2241 Node* prev_mem = find_previous_store(phase); 2242 // Steps (a), (b): Walk past independent stores to find an exact match. 2243 if (prev_mem != NULL) { 2244 Node* prev_val = can_see_stored_value(prev_mem, phase); 2245 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2246 // prev_val and val might differ by a cast; it would be good 2247 // to keep the more informative of the two. 2248 return mem; 2249 } 2250 } 2251 } 2252 2253 return this; 2254 } 2255 2256 //------------------------------match_edge------------------------------------- 2257 // Do we Match on this edge index or not? Match only memory & value 2258 uint StoreNode::match_edge(uint idx) const { 2259 return idx == MemNode::Address || idx == MemNode::ValueIn; 2260 } 2261 2262 //------------------------------cmp-------------------------------------------- 2263 // Do not common stores up together. They generally have to be split 2264 // back up anyways, so do not bother. 2265 uint StoreNode::cmp( const Node &n ) const { 2266 return (&n == this); // Always fail except on self 2267 } 2268 2269 //------------------------------Ideal_masked_input----------------------------- 2270 // Check for a useless mask before a partial-word store 2271 // (StoreB ... (AndI valIn conIa) ) 2272 // If (conIa & mask == mask) this simplifies to 2273 // (StoreB ... (valIn) ) 2274 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2275 Node *val = in(MemNode::ValueIn); 2276 if( val->Opcode() == Op_AndI ) { 2277 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2278 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2279 set_req(MemNode::ValueIn, val->in(1)); 2280 return this; 2281 } 2282 } 2283 return NULL; 2284 } 2285 2286 2287 //------------------------------Ideal_sign_extended_input---------------------- 2288 // Check for useless sign-extension before a partial-word store 2289 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2290 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2291 // (StoreB ... (valIn) ) 2292 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2293 Node *val = in(MemNode::ValueIn); 2294 if( val->Opcode() == Op_RShiftI ) { 2295 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2296 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2297 Node *shl = val->in(1); 2298 if( shl->Opcode() == Op_LShiftI ) { 2299 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2300 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2301 set_req(MemNode::ValueIn, shl->in(1)); 2302 return this; 2303 } 2304 } 2305 } 2306 } 2307 return NULL; 2308 } 2309 2310 //------------------------------value_never_loaded----------------------------------- 2311 // Determine whether there are any possible loads of the value stored. 2312 // For simplicity, we actually check if there are any loads from the 2313 // address stored to, not just for loads of the value stored by this node. 2314 // 2315 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2316 Node *adr = in(Address); 2317 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2318 if (adr_oop == NULL) 2319 return false; 2320 if (!adr_oop->is_known_instance_field()) 2321 return false; // if not a distinct instance, there may be aliases of the address 2322 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2323 Node *use = adr->fast_out(i); 2324 int opc = use->Opcode(); 2325 if (use->is_Load() || use->is_LoadStore()) { 2326 return false; 2327 } 2328 } 2329 return true; 2330 } 2331 2332 //============================================================================= 2333 //------------------------------Ideal------------------------------------------ 2334 // If the store is from an AND mask that leaves the low bits untouched, then 2335 // we can skip the AND operation. If the store is from a sign-extension 2336 // (a left shift, then right shift) we can skip both. 2337 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2338 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2339 if( progress != NULL ) return progress; 2340 2341 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2342 if( progress != NULL ) return progress; 2343 2344 // Finally check the default case 2345 return StoreNode::Ideal(phase, can_reshape); 2346 } 2347 2348 //============================================================================= 2349 //------------------------------Ideal------------------------------------------ 2350 // If the store is from an AND mask that leaves the low bits untouched, then 2351 // we can skip the AND operation 2352 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2353 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2354 if( progress != NULL ) return progress; 2355 2356 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2357 if( progress != NULL ) return progress; 2358 2359 // Finally check the default case 2360 return StoreNode::Ideal(phase, can_reshape); 2361 } 2362 2363 //============================================================================= 2364 //------------------------------Identity--------------------------------------- 2365 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2366 // No need to card mark when storing a null ptr 2367 Node* my_store = in(MemNode::OopStore); 2368 if (my_store->is_Store()) { 2369 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2370 if( t1 == TypePtr::NULL_PTR ) { 2371 return in(MemNode::Memory); 2372 } 2373 } 2374 return this; 2375 } 2376 2377 //============================================================================= 2378 //------------------------------Ideal--------------------------------------- 2379 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2380 Node* progress = StoreNode::Ideal(phase, can_reshape); 2381 if (progress != NULL) return progress; 2382 2383 Node* my_store = in(MemNode::OopStore); 2384 if (my_store->is_MergeMem()) { 2385 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2386 set_req(MemNode::OopStore, mem); 2387 return this; 2388 } 2389 2390 return NULL; 2391 } 2392 2393 //------------------------------Value----------------------------------------- 2394 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2395 // Either input is TOP ==> the result is TOP 2396 const Type *t = phase->type( in(MemNode::Memory) ); 2397 if( t == Type::TOP ) return Type::TOP; 2398 t = phase->type( in(MemNode::Address) ); 2399 if( t == Type::TOP ) return Type::TOP; 2400 t = phase->type( in(MemNode::ValueIn) ); 2401 if( t == Type::TOP ) return Type::TOP; 2402 // If extra input is TOP ==> the result is TOP 2403 t = phase->type( in(MemNode::OopStore) ); 2404 if( t == Type::TOP ) return Type::TOP; 2405 2406 return StoreNode::Value( phase ); 2407 } 2408 2409 2410 //============================================================================= 2411 //----------------------------------SCMemProjNode------------------------------ 2412 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2413 { 2414 return bottom_type(); 2415 } 2416 2417 //============================================================================= 2418 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2419 init_req(MemNode::Control, c ); 2420 init_req(MemNode::Memory , mem); 2421 init_req(MemNode::Address, adr); 2422 init_req(MemNode::ValueIn, val); 2423 init_req( ExpectedIn, ex ); 2424 init_class_id(Class_LoadStore); 2425 2426 } 2427 2428 //============================================================================= 2429 //-------------------------------adr_type-------------------------------------- 2430 // Do we Match on this edge index or not? Do not match memory 2431 const TypePtr* ClearArrayNode::adr_type() const { 2432 Node *adr = in(3); 2433 return MemNode::calculate_adr_type(adr->bottom_type()); 2434 } 2435 2436 //------------------------------match_edge------------------------------------- 2437 // Do we Match on this edge index or not? Do not match memory 2438 uint ClearArrayNode::match_edge(uint idx) const { 2439 return idx > 1; 2440 } 2441 2442 //------------------------------Identity--------------------------------------- 2443 // Clearing a zero length array does nothing 2444 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2445 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2446 } 2447 2448 //------------------------------Idealize--------------------------------------- 2449 // Clearing a short array is faster with stores 2450 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2451 const int unit = BytesPerLong; 2452 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2453 if (!t) return NULL; 2454 if (!t->is_con()) return NULL; 2455 intptr_t raw_count = t->get_con(); 2456 intptr_t size = raw_count; 2457 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2458 // Clearing nothing uses the Identity call. 2459 // Negative clears are possible on dead ClearArrays 2460 // (see jck test stmt114.stmt11402.val). 2461 if (size <= 0 || size % unit != 0) return NULL; 2462 intptr_t count = size / unit; 2463 // Length too long; use fast hardware clear 2464 if (size > Matcher::init_array_short_size) return NULL; 2465 Node *mem = in(1); 2466 if( phase->type(mem)==Type::TOP ) return NULL; 2467 Node *adr = in(3); 2468 const Type* at = phase->type(adr); 2469 if( at==Type::TOP ) return NULL; 2470 const TypePtr* atp = at->isa_ptr(); 2471 // adjust atp to be the correct array element address type 2472 if (atp == NULL) atp = TypePtr::BOTTOM; 2473 else atp = atp->add_offset(Type::OffsetBot); 2474 // Get base for derived pointer purposes 2475 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2476 Node *base = adr->in(1); 2477 2478 Node *zero = phase->makecon(TypeLong::ZERO); 2479 Node *off = phase->MakeConX(BytesPerLong); 2480 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2481 count--; 2482 while( count-- ) { 2483 mem = phase->transform(mem); 2484 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2485 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2486 } 2487 return mem; 2488 } 2489 2490 //----------------------------step_through---------------------------------- 2491 // Return allocation input memory edge if it is different instance 2492 // or itself if it is the one we are looking for. 2493 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2494 Node* n = *np; 2495 assert(n->is_ClearArray(), "sanity"); 2496 intptr_t offset; 2497 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2498 // This method is called only before Allocate nodes are expanded during 2499 // macro nodes expansion. Before that ClearArray nodes are only generated 2500 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2501 assert(alloc != NULL, "should have allocation"); 2502 if (alloc->_idx == instance_id) { 2503 // Can not bypass initialization of the instance we are looking for. 2504 return false; 2505 } 2506 // Otherwise skip it. 2507 InitializeNode* init = alloc->initialization(); 2508 if (init != NULL) 2509 *np = init->in(TypeFunc::Memory); 2510 else 2511 *np = alloc->in(TypeFunc::Memory); 2512 return true; 2513 } 2514 2515 //----------------------------clear_memory------------------------------------- 2516 // Generate code to initialize object storage to zero. 2517 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2518 intptr_t start_offset, 2519 Node* end_offset, 2520 PhaseGVN* phase) { 2521 Compile* C = phase->C; 2522 intptr_t offset = start_offset; 2523 2524 int unit = BytesPerLong; 2525 if ((offset % unit) != 0) { 2526 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2527 adr = phase->transform(adr); 2528 const TypePtr* atp = TypeRawPtr::BOTTOM; 2529 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2530 mem = phase->transform(mem); 2531 offset += BytesPerInt; 2532 } 2533 assert((offset % unit) == 0, ""); 2534 2535 // Initialize the remaining stuff, if any, with a ClearArray. 2536 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2537 } 2538 2539 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2540 Node* start_offset, 2541 Node* end_offset, 2542 PhaseGVN* phase) { 2543 if (start_offset == end_offset) { 2544 // nothing to do 2545 return mem; 2546 } 2547 2548 Compile* C = phase->C; 2549 int unit = BytesPerLong; 2550 Node* zbase = start_offset; 2551 Node* zend = end_offset; 2552 2553 // Scale to the unit required by the CPU: 2554 if (!Matcher::init_array_count_is_in_bytes) { 2555 Node* shift = phase->intcon(exact_log2(unit)); 2556 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2557 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2558 } 2559 2560 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2561 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2562 2563 // Bulk clear double-words 2564 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2565 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2566 return phase->transform(mem); 2567 } 2568 2569 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2570 intptr_t start_offset, 2571 intptr_t end_offset, 2572 PhaseGVN* phase) { 2573 if (start_offset == end_offset) { 2574 // nothing to do 2575 return mem; 2576 } 2577 2578 Compile* C = phase->C; 2579 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2580 intptr_t done_offset = end_offset; 2581 if ((done_offset % BytesPerLong) != 0) { 2582 done_offset -= BytesPerInt; 2583 } 2584 if (done_offset > start_offset) { 2585 mem = clear_memory(ctl, mem, dest, 2586 start_offset, phase->MakeConX(done_offset), phase); 2587 } 2588 if (done_offset < end_offset) { // emit the final 32-bit store 2589 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2590 adr = phase->transform(adr); 2591 const TypePtr* atp = TypeRawPtr::BOTTOM; 2592 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2593 mem = phase->transform(mem); 2594 done_offset += BytesPerInt; 2595 } 2596 assert(done_offset == end_offset, ""); 2597 return mem; 2598 } 2599 2600 //============================================================================= 2601 // Do we match on this edge? No memory edges 2602 uint StrCompNode::match_edge(uint idx) const { 2603 return idx == 2 || idx == 3; // StrComp (Binary str1 cnt1) (Binary str2 cnt2) 2604 } 2605 2606 //------------------------------Ideal------------------------------------------ 2607 // Return a node which is more "ideal" than the current node. Strip out 2608 // control copies 2609 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2610 return remove_dead_region(phase, can_reshape) ? this : NULL; 2611 } 2612 2613 //============================================================================= 2614 // Do we match on this edge? No memory edges 2615 uint StrEqualsNode::match_edge(uint idx) const { 2616 return idx == 2 || idx == 3; // StrEquals (Binary str1 str2) cnt 2617 } 2618 2619 //------------------------------Ideal------------------------------------------ 2620 // Return a node which is more "ideal" than the current node. Strip out 2621 // control copies 2622 Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2623 return remove_dead_region(phase, can_reshape) ? this : NULL; 2624 } 2625 2626 //============================================================================= 2627 // Do we match on this edge? No memory edges 2628 uint StrIndexOfNode::match_edge(uint idx) const { 2629 return idx == 2 || idx == 3; // StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2) 2630 } 2631 2632 //------------------------------Ideal------------------------------------------ 2633 // Return a node which is more "ideal" than the current node. Strip out 2634 // control copies 2635 Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2636 return remove_dead_region(phase, can_reshape) ? this : NULL; 2637 } 2638 2639 //============================================================================= 2640 // Do we match on this edge? No memory edges 2641 uint AryEqNode::match_edge(uint idx) const { 2642 return idx == 2 || idx == 3; // StrEquals ary1 ary2 2643 } 2644 //------------------------------Ideal------------------------------------------ 2645 // Return a node which is more "ideal" than the current node. Strip out 2646 // control copies 2647 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2648 return remove_dead_region(phase, can_reshape) ? this : NULL; 2649 } 2650 2651 //============================================================================= 2652 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2653 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2654 _adr_type(C->get_adr_type(alias_idx)) 2655 { 2656 init_class_id(Class_MemBar); 2657 Node* top = C->top(); 2658 init_req(TypeFunc::I_O,top); 2659 init_req(TypeFunc::FramePtr,top); 2660 init_req(TypeFunc::ReturnAdr,top); 2661 if (precedent != NULL) 2662 init_req(TypeFunc::Parms, precedent); 2663 } 2664 2665 //------------------------------cmp-------------------------------------------- 2666 uint MemBarNode::hash() const { return NO_HASH; } 2667 uint MemBarNode::cmp( const Node &n ) const { 2668 return (&n == this); // Always fail except on self 2669 } 2670 2671 //------------------------------make------------------------------------------- 2672 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2673 int len = Precedent + (pn == NULL? 0: 1); 2674 switch (opcode) { 2675 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2676 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2677 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2678 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2679 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2680 default: ShouldNotReachHere(); return NULL; 2681 } 2682 } 2683 2684 //------------------------------Ideal------------------------------------------ 2685 // Return a node which is more "ideal" than the current node. Strip out 2686 // control copies 2687 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2688 if (remove_dead_region(phase, can_reshape)) return this; 2689 2690 // Eliminate volatile MemBars for scalar replaced objects. 2691 if (can_reshape && req() == (Precedent+1) && 2692 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { 2693 // Volatile field loads and stores. 2694 Node* my_mem = in(MemBarNode::Precedent); 2695 if (my_mem != NULL && my_mem->is_Mem()) { 2696 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2697 // Check for scalar replaced object reference. 2698 if( t_oop != NULL && t_oop->is_known_instance_field() && 2699 t_oop->offset() != Type::OffsetBot && 2700 t_oop->offset() != Type::OffsetTop) { 2701 // Replace MemBar projections by its inputs. 2702 PhaseIterGVN* igvn = phase->is_IterGVN(); 2703 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2704 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2705 // Must return either the original node (now dead) or a new node 2706 // (Do not return a top here, since that would break the uniqueness of top.) 2707 return new (phase->C, 1) ConINode(TypeInt::ZERO); 2708 } 2709 } 2710 } 2711 return NULL; 2712 } 2713 2714 //------------------------------Value------------------------------------------ 2715 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2716 if( !in(0) ) return Type::TOP; 2717 if( phase->type(in(0)) == Type::TOP ) 2718 return Type::TOP; 2719 return TypeTuple::MEMBAR; 2720 } 2721 2722 //------------------------------match------------------------------------------ 2723 // Construct projections for memory. 2724 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2725 switch (proj->_con) { 2726 case TypeFunc::Control: 2727 case TypeFunc::Memory: 2728 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2729 } 2730 ShouldNotReachHere(); 2731 return NULL; 2732 } 2733 2734 //===========================InitializeNode==================================== 2735 // SUMMARY: 2736 // This node acts as a memory barrier on raw memory, after some raw stores. 2737 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2738 // The Initialize can 'capture' suitably constrained stores as raw inits. 2739 // It can coalesce related raw stores into larger units (called 'tiles'). 2740 // It can avoid zeroing new storage for memory units which have raw inits. 2741 // At macro-expansion, it is marked 'complete', and does not optimize further. 2742 // 2743 // EXAMPLE: 2744 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2745 // ctl = incoming control; mem* = incoming memory 2746 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2747 // First allocate uninitialized memory and fill in the header: 2748 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2749 // ctl := alloc.Control; mem* := alloc.Memory* 2750 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2751 // Then initialize to zero the non-header parts of the raw memory block: 2752 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2753 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2754 // After the initialize node executes, the object is ready for service: 2755 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2756 // Suppose its body is immediately initialized as {1,2}: 2757 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2758 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2759 // mem.SLICE(#short[*]) := store2 2760 // 2761 // DETAILS: 2762 // An InitializeNode collects and isolates object initialization after 2763 // an AllocateNode and before the next possible safepoint. As a 2764 // memory barrier (MemBarNode), it keeps critical stores from drifting 2765 // down past any safepoint or any publication of the allocation. 2766 // Before this barrier, a newly-allocated object may have uninitialized bits. 2767 // After this barrier, it may be treated as a real oop, and GC is allowed. 2768 // 2769 // The semantics of the InitializeNode include an implicit zeroing of 2770 // the new object from object header to the end of the object. 2771 // (The object header and end are determined by the AllocateNode.) 2772 // 2773 // Certain stores may be added as direct inputs to the InitializeNode. 2774 // These stores must update raw memory, and they must be to addresses 2775 // derived from the raw address produced by AllocateNode, and with 2776 // a constant offset. They must be ordered by increasing offset. 2777 // The first one is at in(RawStores), the last at in(req()-1). 2778 // Unlike most memory operations, they are not linked in a chain, 2779 // but are displayed in parallel as users of the rawmem output of 2780 // the allocation. 2781 // 2782 // (See comments in InitializeNode::capture_store, which continue 2783 // the example given above.) 2784 // 2785 // When the associated Allocate is macro-expanded, the InitializeNode 2786 // may be rewritten to optimize collected stores. A ClearArrayNode 2787 // may also be created at that point to represent any required zeroing. 2788 // The InitializeNode is then marked 'complete', prohibiting further 2789 // capturing of nearby memory operations. 2790 // 2791 // During macro-expansion, all captured initializations which store 2792 // constant values of 32 bits or smaller are coalesced (if advantageous) 2793 // into larger 'tiles' 32 or 64 bits. This allows an object to be 2794 // initialized in fewer memory operations. Memory words which are 2795 // covered by neither tiles nor non-constant stores are pre-zeroed 2796 // by explicit stores of zero. (The code shape happens to do all 2797 // zeroing first, then all other stores, with both sequences occurring 2798 // in order of ascending offsets.) 2799 // 2800 // Alternatively, code may be inserted between an AllocateNode and its 2801 // InitializeNode, to perform arbitrary initialization of the new object. 2802 // E.g., the object copying intrinsics insert complex data transfers here. 2803 // The initialization must then be marked as 'complete' disable the 2804 // built-in zeroing semantics and the collection of initializing stores. 2805 // 2806 // While an InitializeNode is incomplete, reads from the memory state 2807 // produced by it are optimizable if they match the control edge and 2808 // new oop address associated with the allocation/initialization. 2809 // They return a stored value (if the offset matches) or else zero. 2810 // A write to the memory state, if it matches control and address, 2811 // and if it is to a constant offset, may be 'captured' by the 2812 // InitializeNode. It is cloned as a raw memory operation and rewired 2813 // inside the initialization, to the raw oop produced by the allocation. 2814 // Operations on addresses which are provably distinct (e.g., to 2815 // other AllocateNodes) are allowed to bypass the initialization. 2816 // 2817 // The effect of all this is to consolidate object initialization 2818 // (both arrays and non-arrays, both piecewise and bulk) into a 2819 // single location, where it can be optimized as a unit. 2820 // 2821 // Only stores with an offset less than TrackedInitializationLimit words 2822 // will be considered for capture by an InitializeNode. This puts a 2823 // reasonable limit on the complexity of optimized initializations. 2824 2825 //---------------------------InitializeNode------------------------------------ 2826 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2827 : _is_complete(false), 2828 MemBarNode(C, adr_type, rawoop) 2829 { 2830 init_class_id(Class_Initialize); 2831 2832 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2833 assert(in(RawAddress) == rawoop, "proper init"); 2834 // Note: allocation() can be NULL, for secondary initialization barriers 2835 } 2836 2837 // Since this node is not matched, it will be processed by the 2838 // register allocator. Declare that there are no constraints 2839 // on the allocation of the RawAddress edge. 2840 const RegMask &InitializeNode::in_RegMask(uint idx) const { 2841 // This edge should be set to top, by the set_complete. But be conservative. 2842 if (idx == InitializeNode::RawAddress) 2843 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2844 return RegMask::Empty; 2845 } 2846 2847 Node* InitializeNode::memory(uint alias_idx) { 2848 Node* mem = in(Memory); 2849 if (mem->is_MergeMem()) { 2850 return mem->as_MergeMem()->memory_at(alias_idx); 2851 } else { 2852 // incoming raw memory is not split 2853 return mem; 2854 } 2855 } 2856 2857 bool InitializeNode::is_non_zero() { 2858 if (is_complete()) return false; 2859 remove_extra_zeroes(); 2860 return (req() > RawStores); 2861 } 2862 2863 void InitializeNode::set_complete(PhaseGVN* phase) { 2864 assert(!is_complete(), "caller responsibility"); 2865 _is_complete = true; 2866 2867 // After this node is complete, it contains a bunch of 2868 // raw-memory initializations. There is no need for 2869 // it to have anything to do with non-raw memory effects. 2870 // Therefore, tell all non-raw users to re-optimize themselves, 2871 // after skipping the memory effects of this initialization. 2872 PhaseIterGVN* igvn = phase->is_IterGVN(); 2873 if (igvn) igvn->add_users_to_worklist(this); 2874 } 2875 2876 // convenience function 2877 // return false if the init contains any stores already 2878 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2879 InitializeNode* init = initialization(); 2880 if (init == NULL || init->is_complete()) return false; 2881 init->remove_extra_zeroes(); 2882 // for now, if this allocation has already collected any inits, bail: 2883 if (init->is_non_zero()) return false; 2884 init->set_complete(phase); 2885 return true; 2886 } 2887 2888 void InitializeNode::remove_extra_zeroes() { 2889 if (req() == RawStores) return; 2890 Node* zmem = zero_memory(); 2891 uint fill = RawStores; 2892 for (uint i = fill; i < req(); i++) { 2893 Node* n = in(i); 2894 if (n->is_top() || n == zmem) continue; // skip 2895 if (fill < i) set_req(fill, n); // compact 2896 ++fill; 2897 } 2898 // delete any empty spaces created: 2899 while (fill < req()) { 2900 del_req(fill); 2901 } 2902 } 2903 2904 // Helper for remembering which stores go with which offsets. 2905 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2906 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2907 intptr_t offset = -1; 2908 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2909 phase, offset); 2910 if (base == NULL) return -1; // something is dead, 2911 if (offset < 0) return -1; // dead, dead 2912 return offset; 2913 } 2914 2915 // Helper for proving that an initialization expression is 2916 // "simple enough" to be folded into an object initialization. 2917 // Attempts to prove that a store's initial value 'n' can be captured 2918 // within the initialization without creating a vicious cycle, such as: 2919 // { Foo p = new Foo(); p.next = p; } 2920 // True for constants and parameters and small combinations thereof. 2921 bool InitializeNode::detect_init_independence(Node* n, 2922 bool st_is_pinned, 2923 int& count) { 2924 if (n == NULL) return true; // (can this really happen?) 2925 if (n->is_Proj()) n = n->in(0); 2926 if (n == this) return false; // found a cycle 2927 if (n->is_Con()) return true; 2928 if (n->is_Start()) return true; // params, etc., are OK 2929 if (n->is_Root()) return true; // even better 2930 2931 Node* ctl = n->in(0); 2932 if (ctl != NULL && !ctl->is_top()) { 2933 if (ctl->is_Proj()) ctl = ctl->in(0); 2934 if (ctl == this) return false; 2935 2936 // If we already know that the enclosing memory op is pinned right after 2937 // the init, then any control flow that the store has picked up 2938 // must have preceded the init, or else be equal to the init. 2939 // Even after loop optimizations (which might change control edges) 2940 // a store is never pinned *before* the availability of its inputs. 2941 if (!MemNode::all_controls_dominate(n, this)) 2942 return false; // failed to prove a good control 2943 2944 } 2945 2946 // Check data edges for possible dependencies on 'this'. 2947 if ((count += 1) > 20) return false; // complexity limit 2948 for (uint i = 1; i < n->req(); i++) { 2949 Node* m = n->in(i); 2950 if (m == NULL || m == n || m->is_top()) continue; 2951 uint first_i = n->find_edge(m); 2952 if (i != first_i) continue; // process duplicate edge just once 2953 if (!detect_init_independence(m, st_is_pinned, count)) { 2954 return false; 2955 } 2956 } 2957 2958 return true; 2959 } 2960 2961 // Here are all the checks a Store must pass before it can be moved into 2962 // an initialization. Returns zero if a check fails. 2963 // On success, returns the (constant) offset to which the store applies, 2964 // within the initialized memory. 2965 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2966 const int FAIL = 0; 2967 if (st->req() != MemNode::ValueIn + 1) 2968 return FAIL; // an inscrutable StoreNode (card mark?) 2969 Node* ctl = st->in(MemNode::Control); 2970 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2971 return FAIL; // must be unconditional after the initialization 2972 Node* mem = st->in(MemNode::Memory); 2973 if (!(mem->is_Proj() && mem->in(0) == this)) 2974 return FAIL; // must not be preceded by other stores 2975 Node* adr = st->in(MemNode::Address); 2976 intptr_t offset; 2977 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2978 if (alloc == NULL) 2979 return FAIL; // inscrutable address 2980 if (alloc != allocation()) 2981 return FAIL; // wrong allocation! (store needs to float up) 2982 Node* val = st->in(MemNode::ValueIn); 2983 int complexity_count = 0; 2984 if (!detect_init_independence(val, true, complexity_count)) 2985 return FAIL; // stored value must be 'simple enough' 2986 2987 return offset; // success 2988 } 2989 2990 // Find the captured store in(i) which corresponds to the range 2991 // [start..start+size) in the initialized object. 2992 // If there is one, return its index i. If there isn't, return the 2993 // negative of the index where it should be inserted. 2994 // Return 0 if the queried range overlaps an initialization boundary 2995 // or if dead code is encountered. 2996 // If size_in_bytes is zero, do not bother with overlap checks. 2997 int InitializeNode::captured_store_insertion_point(intptr_t start, 2998 int size_in_bytes, 2999 PhaseTransform* phase) { 3000 const int FAIL = 0, MAX_STORE = BytesPerLong; 3001 3002 if (is_complete()) 3003 return FAIL; // arraycopy got here first; punt 3004 3005 assert(allocation() != NULL, "must be present"); 3006 3007 // no negatives, no header fields: 3008 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3009 3010 // after a certain size, we bail out on tracking all the stores: 3011 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3012 if (start >= ti_limit) return FAIL; 3013 3014 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3015 if (i >= limit) return -(int)i; // not found; here is where to put it 3016 3017 Node* st = in(i); 3018 intptr_t st_off = get_store_offset(st, phase); 3019 if (st_off < 0) { 3020 if (st != zero_memory()) { 3021 return FAIL; // bail out if there is dead garbage 3022 } 3023 } else if (st_off > start) { 3024 // ...we are done, since stores are ordered 3025 if (st_off < start + size_in_bytes) { 3026 return FAIL; // the next store overlaps 3027 } 3028 return -(int)i; // not found; here is where to put it 3029 } else if (st_off < start) { 3030 if (size_in_bytes != 0 && 3031 start < st_off + MAX_STORE && 3032 start < st_off + st->as_Store()->memory_size()) { 3033 return FAIL; // the previous store overlaps 3034 } 3035 } else { 3036 if (size_in_bytes != 0 && 3037 st->as_Store()->memory_size() != size_in_bytes) { 3038 return FAIL; // mismatched store size 3039 } 3040 return i; 3041 } 3042 3043 ++i; 3044 } 3045 } 3046 3047 // Look for a captured store which initializes at the offset 'start' 3048 // with the given size. If there is no such store, and no other 3049 // initialization interferes, then return zero_memory (the memory 3050 // projection of the AllocateNode). 3051 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3052 PhaseTransform* phase) { 3053 assert(stores_are_sane(phase), ""); 3054 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3055 if (i == 0) { 3056 return NULL; // something is dead 3057 } else if (i < 0) { 3058 return zero_memory(); // just primordial zero bits here 3059 } else { 3060 Node* st = in(i); // here is the store at this position 3061 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3062 return st; 3063 } 3064 } 3065 3066 // Create, as a raw pointer, an address within my new object at 'offset'. 3067 Node* InitializeNode::make_raw_address(intptr_t offset, 3068 PhaseTransform* phase) { 3069 Node* addr = in(RawAddress); 3070 if (offset != 0) { 3071 Compile* C = phase->C; 3072 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 3073 phase->MakeConX(offset)) ); 3074 } 3075 return addr; 3076 } 3077 3078 // Clone the given store, converting it into a raw store 3079 // initializing a field or element of my new object. 3080 // Caller is responsible for retiring the original store, 3081 // with subsume_node or the like. 3082 // 3083 // From the example above InitializeNode::InitializeNode, 3084 // here are the old stores to be captured: 3085 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3086 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3087 // 3088 // Here is the changed code; note the extra edges on init: 3089 // alloc = (Allocate ...) 3090 // rawoop = alloc.RawAddress 3091 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3092 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3093 // init = (Initialize alloc.Control alloc.Memory rawoop 3094 // rawstore1 rawstore2) 3095 // 3096 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3097 PhaseTransform* phase) { 3098 assert(stores_are_sane(phase), ""); 3099 3100 if (start < 0) return NULL; 3101 assert(can_capture_store(st, phase) == start, "sanity"); 3102 3103 Compile* C = phase->C; 3104 int size_in_bytes = st->memory_size(); 3105 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3106 if (i == 0) return NULL; // bail out 3107 Node* prev_mem = NULL; // raw memory for the captured store 3108 if (i > 0) { 3109 prev_mem = in(i); // there is a pre-existing store under this one 3110 set_req(i, C->top()); // temporarily disconnect it 3111 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3112 } else { 3113 i = -i; // no pre-existing store 3114 prev_mem = zero_memory(); // a slice of the newly allocated object 3115 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3116 set_req(--i, C->top()); // reuse this edge; it has been folded away 3117 else 3118 ins_req(i, C->top()); // build a new edge 3119 } 3120 Node* new_st = st->clone(); 3121 new_st->set_req(MemNode::Control, in(Control)); 3122 new_st->set_req(MemNode::Memory, prev_mem); 3123 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3124 new_st = phase->transform(new_st); 3125 3126 // At this point, new_st might have swallowed a pre-existing store 3127 // at the same offset, or perhaps new_st might have disappeared, 3128 // if it redundantly stored the same value (or zero to fresh memory). 3129 3130 // In any case, wire it in: 3131 set_req(i, new_st); 3132 3133 // The caller may now kill the old guy. 3134 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3135 assert(check_st == new_st || check_st == NULL, "must be findable"); 3136 assert(!is_complete(), ""); 3137 return new_st; 3138 } 3139 3140 static bool store_constant(jlong* tiles, int num_tiles, 3141 intptr_t st_off, int st_size, 3142 jlong con) { 3143 if ((st_off & (st_size-1)) != 0) 3144 return false; // strange store offset (assume size==2**N) 3145 address addr = (address)tiles + st_off; 3146 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3147 switch (st_size) { 3148 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3149 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3150 case sizeof(jint): *(jint*) addr = (jint) con; break; 3151 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3152 default: return false; // strange store size (detect size!=2**N here) 3153 } 3154 return true; // return success to caller 3155 } 3156 3157 // Coalesce subword constants into int constants and possibly 3158 // into long constants. The goal, if the CPU permits, 3159 // is to initialize the object with a small number of 64-bit tiles. 3160 // Also, convert floating-point constants to bit patterns. 3161 // Non-constants are not relevant to this pass. 3162 // 3163 // In terms of the running example on InitializeNode::InitializeNode 3164 // and InitializeNode::capture_store, here is the transformation 3165 // of rawstore1 and rawstore2 into rawstore12: 3166 // alloc = (Allocate ...) 3167 // rawoop = alloc.RawAddress 3168 // tile12 = 0x00010002 3169 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3170 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3171 // 3172 void 3173 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3174 Node* size_in_bytes, 3175 PhaseGVN* phase) { 3176 Compile* C = phase->C; 3177 3178 assert(stores_are_sane(phase), ""); 3179 // Note: After this pass, they are not completely sane, 3180 // since there may be some overlaps. 3181 3182 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3183 3184 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3185 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3186 size_limit = MIN2(size_limit, ti_limit); 3187 size_limit = align_size_up(size_limit, BytesPerLong); 3188 int num_tiles = size_limit / BytesPerLong; 3189 3190 // allocate space for the tile map: 3191 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3192 jlong tiles_buf[small_len]; 3193 Node* nodes_buf[small_len]; 3194 jlong inits_buf[small_len]; 3195 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3196 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3197 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3198 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3199 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3200 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3201 // tiles: exact bitwise model of all primitive constants 3202 // nodes: last constant-storing node subsumed into the tiles model 3203 // inits: which bytes (in each tile) are touched by any initializations 3204 3205 //// Pass A: Fill in the tile model with any relevant stores. 3206 3207 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3208 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3209 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3210 Node* zmem = zero_memory(); // initially zero memory state 3211 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3212 Node* st = in(i); 3213 intptr_t st_off = get_store_offset(st, phase); 3214 3215 // Figure out the store's offset and constant value: 3216 if (st_off < header_size) continue; //skip (ignore header) 3217 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3218 int st_size = st->as_Store()->memory_size(); 3219 if (st_off + st_size > size_limit) break; 3220 3221 // Record which bytes are touched, whether by constant or not. 3222 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3223 continue; // skip (strange store size) 3224 3225 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3226 if (!val->singleton()) continue; //skip (non-con store) 3227 BasicType type = val->basic_type(); 3228 3229 jlong con = 0; 3230 switch (type) { 3231 case T_INT: con = val->is_int()->get_con(); break; 3232 case T_LONG: con = val->is_long()->get_con(); break; 3233 case T_FLOAT: con = jint_cast(val->getf()); break; 3234 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3235 default: continue; //skip (odd store type) 3236 } 3237 3238 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3239 st->Opcode() == Op_StoreL) { 3240 continue; // This StoreL is already optimal. 3241 } 3242 3243 // Store down the constant. 3244 store_constant(tiles, num_tiles, st_off, st_size, con); 3245 3246 intptr_t j = st_off >> LogBytesPerLong; 3247 3248 if (type == T_INT && st_size == BytesPerInt 3249 && (st_off & BytesPerInt) == BytesPerInt) { 3250 jlong lcon = tiles[j]; 3251 if (!Matcher::isSimpleConstant64(lcon) && 3252 st->Opcode() == Op_StoreI) { 3253 // This StoreI is already optimal by itself. 3254 jint* intcon = (jint*) &tiles[j]; 3255 intcon[1] = 0; // undo the store_constant() 3256 3257 // If the previous store is also optimal by itself, back up and 3258 // undo the action of the previous loop iteration... if we can. 3259 // But if we can't, just let the previous half take care of itself. 3260 st = nodes[j]; 3261 st_off -= BytesPerInt; 3262 con = intcon[0]; 3263 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3264 assert(st_off >= header_size, "still ignoring header"); 3265 assert(get_store_offset(st, phase) == st_off, "must be"); 3266 assert(in(i-1) == zmem, "must be"); 3267 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3268 assert(con == tcon->is_int()->get_con(), "must be"); 3269 // Undo the effects of the previous loop trip, which swallowed st: 3270 intcon[0] = 0; // undo store_constant() 3271 set_req(i-1, st); // undo set_req(i, zmem) 3272 nodes[j] = NULL; // undo nodes[j] = st 3273 --old_subword; // undo ++old_subword 3274 } 3275 continue; // This StoreI is already optimal. 3276 } 3277 } 3278 3279 // This store is not needed. 3280 set_req(i, zmem); 3281 nodes[j] = st; // record for the moment 3282 if (st_size < BytesPerLong) // something has changed 3283 ++old_subword; // includes int/float, but who's counting... 3284 else ++old_long; 3285 } 3286 3287 if ((old_subword + old_long) == 0) 3288 return; // nothing more to do 3289 3290 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3291 // Be sure to insert them before overlapping non-constant stores. 3292 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3293 for (int j = 0; j < num_tiles; j++) { 3294 jlong con = tiles[j]; 3295 jlong init = inits[j]; 3296 if (con == 0) continue; 3297 jint con0, con1; // split the constant, address-wise 3298 jint init0, init1; // split the init map, address-wise 3299 { union { jlong con; jint intcon[2]; } u; 3300 u.con = con; 3301 con0 = u.intcon[0]; 3302 con1 = u.intcon[1]; 3303 u.con = init; 3304 init0 = u.intcon[0]; 3305 init1 = u.intcon[1]; 3306 } 3307 3308 Node* old = nodes[j]; 3309 assert(old != NULL, "need the prior store"); 3310 intptr_t offset = (j * BytesPerLong); 3311 3312 bool split = !Matcher::isSimpleConstant64(con); 3313 3314 if (offset < header_size) { 3315 assert(offset + BytesPerInt >= header_size, "second int counts"); 3316 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3317 split = true; // only the second word counts 3318 // Example: int a[] = { 42 ... } 3319 } else if (con0 == 0 && init0 == -1) { 3320 split = true; // first word is covered by full inits 3321 // Example: int a[] = { ... foo(), 42 ... } 3322 } else if (con1 == 0 && init1 == -1) { 3323 split = true; // second word is covered by full inits 3324 // Example: int a[] = { ... 42, foo() ... } 3325 } 3326 3327 // Here's a case where init0 is neither 0 nor -1: 3328 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3329 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3330 // In this case the tile is not split; it is (jlong)42. 3331 // The big tile is stored down, and then the foo() value is inserted. 3332 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3333 3334 Node* ctl = old->in(MemNode::Control); 3335 Node* adr = make_raw_address(offset, phase); 3336 const TypePtr* atp = TypeRawPtr::BOTTOM; 3337 3338 // One or two coalesced stores to plop down. 3339 Node* st[2]; 3340 intptr_t off[2]; 3341 int nst = 0; 3342 if (!split) { 3343 ++new_long; 3344 off[nst] = offset; 3345 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3346 phase->longcon(con), T_LONG); 3347 } else { 3348 // Omit either if it is a zero. 3349 if (con0 != 0) { 3350 ++new_int; 3351 off[nst] = offset; 3352 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3353 phase->intcon(con0), T_INT); 3354 } 3355 if (con1 != 0) { 3356 ++new_int; 3357 offset += BytesPerInt; 3358 adr = make_raw_address(offset, phase); 3359 off[nst] = offset; 3360 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3361 phase->intcon(con1), T_INT); 3362 } 3363 } 3364 3365 // Insert second store first, then the first before the second. 3366 // Insert each one just before any overlapping non-constant stores. 3367 while (nst > 0) { 3368 Node* st1 = st[--nst]; 3369 C->copy_node_notes_to(st1, old); 3370 st1 = phase->transform(st1); 3371 offset = off[nst]; 3372 assert(offset >= header_size, "do not smash header"); 3373 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3374 guarantee(ins_idx != 0, "must re-insert constant store"); 3375 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3376 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3377 set_req(--ins_idx, st1); 3378 else 3379 ins_req(ins_idx, st1); 3380 } 3381 } 3382 3383 if (PrintCompilation && WizardMode) 3384 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3385 old_subword, old_long, new_int, new_long); 3386 if (C->log() != NULL) 3387 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3388 old_subword, old_long, new_int, new_long); 3389 3390 // Clean up any remaining occurrences of zmem: 3391 remove_extra_zeroes(); 3392 } 3393 3394 // Explore forward from in(start) to find the first fully initialized 3395 // word, and return its offset. Skip groups of subword stores which 3396 // together initialize full words. If in(start) is itself part of a 3397 // fully initialized word, return the offset of in(start). If there 3398 // are no following full-word stores, or if something is fishy, return 3399 // a negative value. 3400 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3401 int int_map = 0; 3402 intptr_t int_map_off = 0; 3403 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3404 3405 for (uint i = start, limit = req(); i < limit; i++) { 3406 Node* st = in(i); 3407 3408 intptr_t st_off = get_store_offset(st, phase); 3409 if (st_off < 0) break; // return conservative answer 3410 3411 int st_size = st->as_Store()->memory_size(); 3412 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3413 return st_off; // we found a complete word init 3414 } 3415 3416 // update the map: 3417 3418 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3419 if (this_int_off != int_map_off) { 3420 // reset the map: 3421 int_map = 0; 3422 int_map_off = this_int_off; 3423 } 3424 3425 int subword_off = st_off - this_int_off; 3426 int_map |= right_n_bits(st_size) << subword_off; 3427 if ((int_map & FULL_MAP) == FULL_MAP) { 3428 return this_int_off; // we found a complete word init 3429 } 3430 3431 // Did this store hit or cross the word boundary? 3432 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3433 if (next_int_off == this_int_off + BytesPerInt) { 3434 // We passed the current int, without fully initializing it. 3435 int_map_off = next_int_off; 3436 int_map >>= BytesPerInt; 3437 } else if (next_int_off > this_int_off + BytesPerInt) { 3438 // We passed the current and next int. 3439 return this_int_off + BytesPerInt; 3440 } 3441 } 3442 3443 return -1; 3444 } 3445 3446 3447 // Called when the associated AllocateNode is expanded into CFG. 3448 // At this point, we may perform additional optimizations. 3449 // Linearize the stores by ascending offset, to make memory 3450 // activity as coherent as possible. 3451 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3452 intptr_t header_size, 3453 Node* size_in_bytes, 3454 PhaseGVN* phase) { 3455 assert(!is_complete(), "not already complete"); 3456 assert(stores_are_sane(phase), ""); 3457 assert(allocation() != NULL, "must be present"); 3458 3459 remove_extra_zeroes(); 3460 3461 if (ReduceFieldZeroing || ReduceBulkZeroing) 3462 // reduce instruction count for common initialization patterns 3463 coalesce_subword_stores(header_size, size_in_bytes, phase); 3464 3465 Node* zmem = zero_memory(); // initially zero memory state 3466 Node* inits = zmem; // accumulating a linearized chain of inits 3467 #ifdef ASSERT 3468 intptr_t first_offset = allocation()->minimum_header_size(); 3469 intptr_t last_init_off = first_offset; // previous init offset 3470 intptr_t last_init_end = first_offset; // previous init offset+size 3471 intptr_t last_tile_end = first_offset; // previous tile offset+size 3472 #endif 3473 intptr_t zeroes_done = header_size; 3474 3475 bool do_zeroing = true; // we might give up if inits are very sparse 3476 int big_init_gaps = 0; // how many large gaps have we seen? 3477 3478 if (ZeroTLAB) do_zeroing = false; 3479 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3480 3481 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3482 Node* st = in(i); 3483 intptr_t st_off = get_store_offset(st, phase); 3484 if (st_off < 0) 3485 break; // unknown junk in the inits 3486 if (st->in(MemNode::Memory) != zmem) 3487 break; // complicated store chains somehow in list 3488 3489 int st_size = st->as_Store()->memory_size(); 3490 intptr_t next_init_off = st_off + st_size; 3491 3492 if (do_zeroing && zeroes_done < next_init_off) { 3493 // See if this store needs a zero before it or under it. 3494 intptr_t zeroes_needed = st_off; 3495 3496 if (st_size < BytesPerInt) { 3497 // Look for subword stores which only partially initialize words. 3498 // If we find some, we must lay down some word-level zeroes first, 3499 // underneath the subword stores. 3500 // 3501 // Examples: 3502 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3503 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3504 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3505 // 3506 // Note: coalesce_subword_stores may have already done this, 3507 // if it was prompted by constant non-zero subword initializers. 3508 // But this case can still arise with non-constant stores. 3509 3510 intptr_t next_full_store = find_next_fullword_store(i, phase); 3511 3512 // In the examples above: 3513 // in(i) p q r s x y z 3514 // st_off 12 13 14 15 12 13 14 3515 // st_size 1 1 1 1 1 1 1 3516 // next_full_s. 12 16 16 16 16 16 16 3517 // z's_done 12 16 16 16 12 16 12 3518 // z's_needed 12 16 16 16 16 16 16 3519 // zsize 0 0 0 0 4 0 4 3520 if (next_full_store < 0) { 3521 // Conservative tack: Zero to end of current word. 3522 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3523 } else { 3524 // Zero to beginning of next fully initialized word. 3525 // Or, don't zero at all, if we are already in that word. 3526 assert(next_full_store >= zeroes_needed, "must go forward"); 3527 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3528 zeroes_needed = next_full_store; 3529 } 3530 } 3531 3532 if (zeroes_needed > zeroes_done) { 3533 intptr_t zsize = zeroes_needed - zeroes_done; 3534 // Do some incremental zeroing on rawmem, in parallel with inits. 3535 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3536 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3537 zeroes_done, zeroes_needed, 3538 phase); 3539 zeroes_done = zeroes_needed; 3540 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3541 do_zeroing = false; // leave the hole, next time 3542 } 3543 } 3544 3545 // Collect the store and move on: 3546 st->set_req(MemNode::Memory, inits); 3547 inits = st; // put it on the linearized chain 3548 set_req(i, zmem); // unhook from previous position 3549 3550 if (zeroes_done == st_off) 3551 zeroes_done = next_init_off; 3552 3553 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3554 3555 #ifdef ASSERT 3556 // Various order invariants. Weaker than stores_are_sane because 3557 // a large constant tile can be filled in by smaller non-constant stores. 3558 assert(st_off >= last_init_off, "inits do not reverse"); 3559 last_init_off = st_off; 3560 const Type* val = NULL; 3561 if (st_size >= BytesPerInt && 3562 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3563 (int)val->basic_type() < (int)T_OBJECT) { 3564 assert(st_off >= last_tile_end, "tiles do not overlap"); 3565 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3566 last_tile_end = MAX2(last_tile_end, next_init_off); 3567 } else { 3568 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3569 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3570 assert(st_off >= last_init_end, "inits do not overlap"); 3571 last_init_end = next_init_off; // it's a non-tile 3572 } 3573 #endif //ASSERT 3574 } 3575 3576 remove_extra_zeroes(); // clear out all the zmems left over 3577 add_req(inits); 3578 3579 if (!ZeroTLAB) { 3580 // If anything remains to be zeroed, zero it all now. 3581 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3582 // if it is the last unused 4 bytes of an instance, forget about it 3583 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3584 if (zeroes_done + BytesPerLong >= size_limit) { 3585 assert(allocation() != NULL, ""); 3586 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3587 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3588 if (zeroes_done == k->layout_helper()) 3589 zeroes_done = size_limit; 3590 } 3591 if (zeroes_done < size_limit) { 3592 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3593 zeroes_done, size_in_bytes, phase); 3594 } 3595 } 3596 3597 set_complete(phase); 3598 return rawmem; 3599 } 3600 3601 3602 #ifdef ASSERT 3603 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3604 if (is_complete()) 3605 return true; // stores could be anything at this point 3606 assert(allocation() != NULL, "must be present"); 3607 intptr_t last_off = allocation()->minimum_header_size(); 3608 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3609 Node* st = in(i); 3610 intptr_t st_off = get_store_offset(st, phase); 3611 if (st_off < 0) continue; // ignore dead garbage 3612 if (last_off > st_off) { 3613 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3614 this->dump(2); 3615 assert(false, "ascending store offsets"); 3616 return false; 3617 } 3618 last_off = st_off + st->as_Store()->memory_size(); 3619 } 3620 return true; 3621 } 3622 #endif //ASSERT 3623 3624 3625 3626 3627 //============================MergeMemNode===================================== 3628 // 3629 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3630 // contributing store or call operations. Each contributor provides the memory 3631 // state for a particular "alias type" (see Compile::alias_type). For example, 3632 // if a MergeMem has an input X for alias category #6, then any memory reference 3633 // to alias category #6 may use X as its memory state input, as an exact equivalent 3634 // to using the MergeMem as a whole. 3635 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3636 // 3637 // (Here, the <N> notation gives the index of the relevant adr_type.) 3638 // 3639 // In one special case (and more cases in the future), alias categories overlap. 3640 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3641 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3642 // it is exactly equivalent to that state W: 3643 // MergeMem(<Bot>: W) <==> W 3644 // 3645 // Usually, the merge has more than one input. In that case, where inputs 3646 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3647 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3648 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3649 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3650 // 3651 // A merge can take a "wide" memory state as one of its narrow inputs. 3652 // This simply means that the merge observes out only the relevant parts of 3653 // the wide input. That is, wide memory states arriving at narrow merge inputs 3654 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3655 // 3656 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3657 // and that memory slices "leak through": 3658 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3659 // 3660 // But, in such a cascade, repeated memory slices can "block the leak": 3661 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3662 // 3663 // In the last example, Y is not part of the combined memory state of the 3664 // outermost MergeMem. The system must, of course, prevent unschedulable 3665 // memory states from arising, so you can be sure that the state Y is somehow 3666 // a precursor to state Y'. 3667 // 3668 // 3669 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3670 // of each MergeMemNode array are exactly the numerical alias indexes, including 3671 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3672 // Compile::alias_type (and kin) produce and manage these indexes. 3673 // 3674 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3675 // (Note that this provides quick access to the top node inside MergeMem methods, 3676 // without the need to reach out via TLS to Compile::current.) 3677 // 3678 // As a consequence of what was just described, a MergeMem that represents a full 3679 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3680 // containing all alias categories. 3681 // 3682 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3683 // 3684 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3685 // a memory state for the alias type <N>, or else the top node, meaning that 3686 // there is no particular input for that alias type. Note that the length of 3687 // a MergeMem is variable, and may be extended at any time to accommodate new 3688 // memory states at larger alias indexes. When merges grow, they are of course 3689 // filled with "top" in the unused in() positions. 3690 // 3691 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3692 // (Top was chosen because it works smoothly with passes like GCM.) 3693 // 3694 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3695 // the type of random VM bits like TLS references.) Since it is always the 3696 // first non-Bot memory slice, some low-level loops use it to initialize an 3697 // index variable: for (i = AliasIdxRaw; i < req(); i++). 3698 // 3699 // 3700 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3701 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3702 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 3703 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3704 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3705 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3706 // 3707 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3708 // really that different from the other memory inputs. An abbreviation called 3709 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3710 // 3711 // 3712 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3713 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3714 // that "emerges though" the base memory will be marked as excluding the alias types 3715 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 3716 // 3717 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3718 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3719 // 3720 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 3721 // (It is currently unimplemented.) As you can see, the resulting merge is 3722 // actually a disjoint union of memory states, rather than an overlay. 3723 // 3724 3725 //------------------------------MergeMemNode----------------------------------- 3726 Node* MergeMemNode::make_empty_memory() { 3727 Node* empty_memory = (Node*) Compile::current()->top(); 3728 assert(empty_memory->is_top(), "correct sentinel identity"); 3729 return empty_memory; 3730 } 3731 3732 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3733 init_class_id(Class_MergeMem); 3734 // all inputs are nullified in Node::Node(int) 3735 // set_input(0, NULL); // no control input 3736 3737 // Initialize the edges uniformly to top, for starters. 3738 Node* empty_mem = make_empty_memory(); 3739 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3740 init_req(i,empty_mem); 3741 } 3742 assert(empty_memory() == empty_mem, ""); 3743 3744 if( new_base != NULL && new_base->is_MergeMem() ) { 3745 MergeMemNode* mdef = new_base->as_MergeMem(); 3746 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3747 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3748 mms.set_memory(mms.memory2()); 3749 } 3750 assert(base_memory() == mdef->base_memory(), ""); 3751 } else { 3752 set_base_memory(new_base); 3753 } 3754 } 3755 3756 // Make a new, untransformed MergeMem with the same base as 'mem'. 3757 // If mem is itself a MergeMem, populate the result with the same edges. 3758 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3759 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3760 } 3761 3762 //------------------------------cmp-------------------------------------------- 3763 uint MergeMemNode::hash() const { return NO_HASH; } 3764 uint MergeMemNode::cmp( const Node &n ) const { 3765 return (&n == this); // Always fail except on self 3766 } 3767 3768 //------------------------------Identity--------------------------------------- 3769 Node* MergeMemNode::Identity(PhaseTransform *phase) { 3770 // Identity if this merge point does not record any interesting memory 3771 // disambiguations. 3772 Node* base_mem = base_memory(); 3773 Node* empty_mem = empty_memory(); 3774 if (base_mem != empty_mem) { // Memory path is not dead? 3775 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3776 Node* mem = in(i); 3777 if (mem != empty_mem && mem != base_mem) { 3778 return this; // Many memory splits; no change 3779 } 3780 } 3781 } 3782 return base_mem; // No memory splits; ID on the one true input 3783 } 3784 3785 //------------------------------Ideal------------------------------------------ 3786 // This method is invoked recursively on chains of MergeMem nodes 3787 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3788 // Remove chain'd MergeMems 3789 // 3790 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3791 // relative to the "in(Bot)". Since we are patching both at the same time, 3792 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3793 // but rewrite each "in(i)" relative to the new "in(Bot)". 3794 Node *progress = NULL; 3795 3796 3797 Node* old_base = base_memory(); 3798 Node* empty_mem = empty_memory(); 3799 if (old_base == empty_mem) 3800 return NULL; // Dead memory path. 3801 3802 MergeMemNode* old_mbase; 3803 if (old_base != NULL && old_base->is_MergeMem()) 3804 old_mbase = old_base->as_MergeMem(); 3805 else 3806 old_mbase = NULL; 3807 Node* new_base = old_base; 3808 3809 // simplify stacked MergeMems in base memory 3810 if (old_mbase) new_base = old_mbase->base_memory(); 3811 3812 // the base memory might contribute new slices beyond my req() 3813 if (old_mbase) grow_to_match(old_mbase); 3814 3815 // Look carefully at the base node if it is a phi. 3816 PhiNode* phi_base; 3817 if (new_base != NULL && new_base->is_Phi()) 3818 phi_base = new_base->as_Phi(); 3819 else 3820 phi_base = NULL; 3821 3822 Node* phi_reg = NULL; 3823 uint phi_len = (uint)-1; 3824 if (phi_base != NULL && !phi_base->is_copy()) { 3825 // do not examine phi if degraded to a copy 3826 phi_reg = phi_base->region(); 3827 phi_len = phi_base->req(); 3828 // see if the phi is unfinished 3829 for (uint i = 1; i < phi_len; i++) { 3830 if (phi_base->in(i) == NULL) { 3831 // incomplete phi; do not look at it yet! 3832 phi_reg = NULL; 3833 phi_len = (uint)-1; 3834 break; 3835 } 3836 } 3837 } 3838 3839 // Note: We do not call verify_sparse on entry, because inputs 3840 // can normalize to the base_memory via subsume_node or similar 3841 // mechanisms. This method repairs that damage. 3842 3843 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3844 3845 // Look at each slice. 3846 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3847 Node* old_in = in(i); 3848 // calculate the old memory value 3849 Node* old_mem = old_in; 3850 if (old_mem == empty_mem) old_mem = old_base; 3851 assert(old_mem == memory_at(i), ""); 3852 3853 // maybe update (reslice) the old memory value 3854 3855 // simplify stacked MergeMems 3856 Node* new_mem = old_mem; 3857 MergeMemNode* old_mmem; 3858 if (old_mem != NULL && old_mem->is_MergeMem()) 3859 old_mmem = old_mem->as_MergeMem(); 3860 else 3861 old_mmem = NULL; 3862 if (old_mmem == this) { 3863 // This can happen if loops break up and safepoints disappear. 3864 // A merge of BotPtr (default) with a RawPtr memory derived from a 3865 // safepoint can be rewritten to a merge of the same BotPtr with 3866 // the BotPtr phi coming into the loop. If that phi disappears 3867 // also, we can end up with a self-loop of the mergemem. 3868 // In general, if loops degenerate and memory effects disappear, 3869 // a mergemem can be left looking at itself. This simply means 3870 // that the mergemem's default should be used, since there is 3871 // no longer any apparent effect on this slice. 3872 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3873 // from start. Update the input to TOP. 3874 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3875 } 3876 else if (old_mmem != NULL) { 3877 new_mem = old_mmem->memory_at(i); 3878 } 3879 // else preceding memory was not a MergeMem 3880 3881 // replace equivalent phis (unfortunately, they do not GVN together) 3882 if (new_mem != NULL && new_mem != new_base && 3883 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3884 if (new_mem->is_Phi()) { 3885 PhiNode* phi_mem = new_mem->as_Phi(); 3886 for (uint i = 1; i < phi_len; i++) { 3887 if (phi_base->in(i) != phi_mem->in(i)) { 3888 phi_mem = NULL; 3889 break; 3890 } 3891 } 3892 if (phi_mem != NULL) { 3893 // equivalent phi nodes; revert to the def 3894 new_mem = new_base; 3895 } 3896 } 3897 } 3898 3899 // maybe store down a new value 3900 Node* new_in = new_mem; 3901 if (new_in == new_base) new_in = empty_mem; 3902 3903 if (new_in != old_in) { 3904 // Warning: Do not combine this "if" with the previous "if" 3905 // A memory slice might have be be rewritten even if it is semantically 3906 // unchanged, if the base_memory value has changed. 3907 set_req(i, new_in); 3908 progress = this; // Report progress 3909 } 3910 } 3911 3912 if (new_base != old_base) { 3913 set_req(Compile::AliasIdxBot, new_base); 3914 // Don't use set_base_memory(new_base), because we need to update du. 3915 assert(base_memory() == new_base, ""); 3916 progress = this; 3917 } 3918 3919 if( base_memory() == this ) { 3920 // a self cycle indicates this memory path is dead 3921 set_req(Compile::AliasIdxBot, empty_mem); 3922 } 3923 3924 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3925 // Recursion must occur after the self cycle check above 3926 if( base_memory()->is_MergeMem() ) { 3927 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3928 Node *m = phase->transform(new_mbase); // Rollup any cycles 3929 if( m != NULL && (m->is_top() || 3930 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3931 // propagate rollup of dead cycle to self 3932 set_req(Compile::AliasIdxBot, empty_mem); 3933 } 3934 } 3935 3936 if( base_memory() == empty_mem ) { 3937 progress = this; 3938 // Cut inputs during Parse phase only. 3939 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3940 if( !can_reshape ) { 3941 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3942 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3943 } 3944 } 3945 } 3946 3947 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3948 // Check if PhiNode::Ideal's "Split phis through memory merges" 3949 // transform should be attempted. Look for this->phi->this cycle. 3950 uint merge_width = req(); 3951 if (merge_width > Compile::AliasIdxRaw) { 3952 PhiNode* phi = base_memory()->as_Phi(); 3953 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3954 if (phi->in(i) == this) { 3955 phase->is_IterGVN()->_worklist.push(phi); 3956 break; 3957 } 3958 } 3959 } 3960 } 3961 3962 assert(progress || verify_sparse(), "please, no dups of base"); 3963 return progress; 3964 } 3965 3966 //-------------------------set_base_memory------------------------------------- 3967 void MergeMemNode::set_base_memory(Node *new_base) { 3968 Node* empty_mem = empty_memory(); 3969 set_req(Compile::AliasIdxBot, new_base); 3970 assert(memory_at(req()) == new_base, "must set default memory"); 3971 // Clear out other occurrences of new_base: 3972 if (new_base != empty_mem) { 3973 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3974 if (in(i) == new_base) set_req(i, empty_mem); 3975 } 3976 } 3977 } 3978 3979 //------------------------------out_RegMask------------------------------------ 3980 const RegMask &MergeMemNode::out_RegMask() const { 3981 return RegMask::Empty; 3982 } 3983 3984 //------------------------------dump_spec-------------------------------------- 3985 #ifndef PRODUCT 3986 void MergeMemNode::dump_spec(outputStream *st) const { 3987 st->print(" {"); 3988 Node* base_mem = base_memory(); 3989 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3990 Node* mem = memory_at(i); 3991 if (mem == base_mem) { st->print(" -"); continue; } 3992 st->print( " N%d:", mem->_idx ); 3993 Compile::current()->get_adr_type(i)->dump_on(st); 3994 } 3995 st->print(" }"); 3996 } 3997 #endif // !PRODUCT 3998 3999 4000 #ifdef ASSERT 4001 static bool might_be_same(Node* a, Node* b) { 4002 if (a == b) return true; 4003 if (!(a->is_Phi() || b->is_Phi())) return false; 4004 // phis shift around during optimization 4005 return true; // pretty stupid... 4006 } 4007 4008 // verify a narrow slice (either incoming or outgoing) 4009 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4010 if (!VerifyAliases) return; // don't bother to verify unless requested 4011 if (is_error_reported()) return; // muzzle asserts when debugging an error 4012 if (Node::in_dump()) return; // muzzle asserts when printing 4013 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4014 assert(n != NULL, ""); 4015 // Elide intervening MergeMem's 4016 while (n->is_MergeMem()) { 4017 n = n->as_MergeMem()->memory_at(alias_idx); 4018 } 4019 Compile* C = Compile::current(); 4020 const TypePtr* n_adr_type = n->adr_type(); 4021 if (n == m->empty_memory()) { 4022 // Implicit copy of base_memory() 4023 } else if (n_adr_type != TypePtr::BOTTOM) { 4024 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4025 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4026 } else { 4027 // A few places like make_runtime_call "know" that VM calls are narrow, 4028 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4029 bool expected_wide_mem = false; 4030 if (n == m->base_memory()) { 4031 expected_wide_mem = true; 4032 } else if (alias_idx == Compile::AliasIdxRaw || 4033 n == m->memory_at(Compile::AliasIdxRaw)) { 4034 expected_wide_mem = true; 4035 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4036 // memory can "leak through" calls on channels that 4037 // are write-once. Allow this also. 4038 expected_wide_mem = true; 4039 } 4040 assert(expected_wide_mem, "expected narrow slice replacement"); 4041 } 4042 } 4043 #else // !ASSERT 4044 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 4045 #endif 4046 4047 4048 //-----------------------------memory_at--------------------------------------- 4049 Node* MergeMemNode::memory_at(uint alias_idx) const { 4050 assert(alias_idx >= Compile::AliasIdxRaw || 4051 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4052 "must avoid base_memory and AliasIdxTop"); 4053 4054 // Otherwise, it is a narrow slice. 4055 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4056 Compile *C = Compile::current(); 4057 if (is_empty_memory(n)) { 4058 // the array is sparse; empty slots are the "top" node 4059 n = base_memory(); 4060 assert(Node::in_dump() 4061 || n == NULL || n->bottom_type() == Type::TOP 4062 || n->adr_type() == TypePtr::BOTTOM 4063 || n->adr_type() == TypeRawPtr::BOTTOM 4064 || Compile::current()->AliasLevel() == 0, 4065 "must be a wide memory"); 4066 // AliasLevel == 0 if we are organizing the memory states manually. 4067 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4068 } else { 4069 // make sure the stored slice is sane 4070 #ifdef ASSERT 4071 if (is_error_reported() || Node::in_dump()) { 4072 } else if (might_be_same(n, base_memory())) { 4073 // Give it a pass: It is a mostly harmless repetition of the base. 4074 // This can arise normally from node subsumption during optimization. 4075 } else { 4076 verify_memory_slice(this, alias_idx, n); 4077 } 4078 #endif 4079 } 4080 return n; 4081 } 4082 4083 //---------------------------set_memory_at------------------------------------- 4084 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4085 verify_memory_slice(this, alias_idx, n); 4086 Node* empty_mem = empty_memory(); 4087 if (n == base_memory()) n = empty_mem; // collapse default 4088 uint need_req = alias_idx+1; 4089 if (req() < need_req) { 4090 if (n == empty_mem) return; // already the default, so do not grow me 4091 // grow the sparse array 4092 do { 4093 add_req(empty_mem); 4094 } while (req() < need_req); 4095 } 4096 set_req( alias_idx, n ); 4097 } 4098 4099 4100 4101 //--------------------------iteration_setup------------------------------------ 4102 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4103 if (other != NULL) { 4104 grow_to_match(other); 4105 // invariant: the finite support of mm2 is within mm->req() 4106 #ifdef ASSERT 4107 for (uint i = req(); i < other->req(); i++) { 4108 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4109 } 4110 #endif 4111 } 4112 // Replace spurious copies of base_memory by top. 4113 Node* base_mem = base_memory(); 4114 if (base_mem != NULL && !base_mem->is_top()) { 4115 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4116 if (in(i) == base_mem) 4117 set_req(i, empty_memory()); 4118 } 4119 } 4120 } 4121 4122 //---------------------------grow_to_match------------------------------------- 4123 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4124 Node* empty_mem = empty_memory(); 4125 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4126 // look for the finite support of the other memory 4127 for (uint i = other->req(); --i >= req(); ) { 4128 if (other->in(i) != empty_mem) { 4129 uint new_len = i+1; 4130 while (req() < new_len) add_req(empty_mem); 4131 break; 4132 } 4133 } 4134 } 4135 4136 //---------------------------verify_sparse------------------------------------- 4137 #ifndef PRODUCT 4138 bool MergeMemNode::verify_sparse() const { 4139 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4140 Node* base_mem = base_memory(); 4141 // The following can happen in degenerate cases, since empty==top. 4142 if (is_empty_memory(base_mem)) return true; 4143 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4144 assert(in(i) != NULL, "sane slice"); 4145 if (in(i) == base_mem) return false; // should have been the sentinel value! 4146 } 4147 return true; 4148 } 4149 4150 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4151 Node* n; 4152 n = mm->in(idx); 4153 if (mem == n) return true; // might be empty_memory() 4154 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4155 if (mem == n) return true; 4156 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4157 if (mem == n) return true; 4158 if (n == NULL) break; 4159 } 4160 return false; 4161 } 4162 #endif // !PRODUCT