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 // TypeOopPtr::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 *toop = tp->isa_oopptr(); 200 if( tp->base() != Type::AnyPtr && 201 !(toop && 202 toop->klass() != NULL && 203 toop->klass()->is_java_lang_Object() && 204 toop->offset() == Type::OffsetBot) ) { 205 // compress paths and change unreachable cycles to TOP 206 // If not, we can update the input infinitely along a MergeMem cycle 207 // Equivalent code in PhiNode::Ideal 208 Node* m = phase->transform(mmem); 209 // If transformed to a MergeMem, get the desired slice 210 // Otherwise the returned node represents memory for every slice 211 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 212 // Update input if it is progress over what we have now 213 } 214 return mem; 215 } 216 217 //--------------------------Ideal_common--------------------------------------- 218 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 219 // Unhook non-raw memories from complete (macro-expanded) initializations. 220 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 221 // If our control input is a dead region, kill all below the region 222 Node *ctl = in(MemNode::Control); 223 if (ctl && remove_dead_region(phase, can_reshape)) 224 return this; 225 ctl = in(MemNode::Control); 226 // Don't bother trying to transform a dead node 227 if( ctl && ctl->is_top() ) return NodeSentinel; 228 229 PhaseIterGVN *igvn = phase->is_IterGVN(); 230 // Wait if control on the worklist. 231 if (ctl && can_reshape && igvn != NULL) { 232 Node* bol = NULL; 233 Node* cmp = NULL; 234 if (ctl->in(0)->is_If()) { 235 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 236 bol = ctl->in(0)->in(1); 237 if (bol->is_Bool()) 238 cmp = ctl->in(0)->in(1)->in(1); 239 } 240 if (igvn->_worklist.member(ctl) || 241 (bol != NULL && igvn->_worklist.member(bol)) || 242 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 243 // This control path may be dead. 244 // Delay this memory node transformation until the control is processed. 245 phase->is_IterGVN()->_worklist.push(this); 246 return NodeSentinel; // caller will return NULL 247 } 248 } 249 // Ignore if memory is dead, or self-loop 250 Node *mem = in(MemNode::Memory); 251 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 252 assert( mem != this, "dead loop in MemNode::Ideal" ); 253 254 Node *address = in(MemNode::Address); 255 const Type *t_adr = phase->type( address ); 256 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 257 258 if( can_reshape && igvn != NULL && 259 (igvn->_worklist.member(address) || 260 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) { 261 // The address's base and type may change when the address is processed. 262 // Delay this mem node transformation until the address is processed. 263 phase->is_IterGVN()->_worklist.push(this); 264 return NodeSentinel; // caller will return NULL 265 } 266 267 // Do NOT remove or optimize the next lines: ensure a new alias index 268 // is allocated for an oop pointer type before Escape Analysis. 269 // Note: C++ will not remove it since the call has side effect. 270 if ( t_adr->isa_oopptr() ) { 271 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 272 } 273 274 #ifdef ASSERT 275 Node* base = NULL; 276 if (address->is_AddP()) 277 base = address->in(AddPNode::Base); 278 assert(base == NULL || t_adr->isa_rawptr() || 279 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); 280 #endif 281 282 // Avoid independent memory operations 283 Node* old_mem = mem; 284 285 // The code which unhooks non-raw memories from complete (macro-expanded) 286 // initializations was removed. After macro-expansion all stores catched 287 // by Initialize node became raw stores and there is no information 288 // which memory slices they modify. So it is unsafe to move any memory 289 // operation above these stores. Also in most cases hooked non-raw memories 290 // were already unhooked by using information from detect_ptr_independence() 291 // and find_previous_store(). 292 293 if (mem->is_MergeMem()) { 294 MergeMemNode* mmem = mem->as_MergeMem(); 295 const TypePtr *tp = t_adr->is_ptr(); 296 297 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 298 } 299 300 if (mem != old_mem) { 301 set_req(MemNode::Memory, mem); 302 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 303 return this; 304 } 305 306 // let the subclass continue analyzing... 307 return NULL; 308 } 309 310 // Helper function for proving some simple control dominations. 311 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 312 // Already assumes that 'dom' is available at 'sub', and that 'sub' 313 // is not a constant (dominated by the method's StartNode). 314 // Used by MemNode::find_previous_store to prove that the 315 // control input of a memory operation predates (dominates) 316 // an allocation it wants to look past. 317 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 318 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 319 return false; // Conservative answer for dead code 320 321 // Check 'dom'. Skip Proj and CatchProj nodes. 322 dom = dom->find_exact_control(dom); 323 if (dom == NULL || dom->is_top()) 324 return false; // Conservative answer for dead code 325 326 if (dom == sub) { 327 // For the case when, for example, 'sub' is Initialize and the original 328 // 'dom' is Proj node of the 'sub'. 329 return false; 330 } 331 332 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 333 return true; 334 335 // 'dom' dominates 'sub' if its control edge and control edges 336 // of all its inputs dominate or equal to sub's control edge. 337 338 // Currently 'sub' is either Allocate, Initialize or Start nodes. 339 // Or Region for the check in LoadNode::Ideal(); 340 // 'sub' should have sub->in(0) != NULL. 341 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 342 sub->is_Region(), "expecting only these nodes"); 343 344 // Get control edge of 'sub'. 345 Node* orig_sub = sub; 346 sub = sub->find_exact_control(sub->in(0)); 347 if (sub == NULL || sub->is_top()) 348 return false; // Conservative answer for dead code 349 350 assert(sub->is_CFG(), "expecting control"); 351 352 if (sub == dom) 353 return true; 354 355 if (sub->is_Start() || sub->is_Root()) 356 return false; 357 358 { 359 // Check all control edges of 'dom'. 360 361 ResourceMark rm; 362 Arena* arena = Thread::current()->resource_area(); 363 Node_List nlist(arena); 364 Unique_Node_List dom_list(arena); 365 366 dom_list.push(dom); 367 bool only_dominating_controls = false; 368 369 for (uint next = 0; next < dom_list.size(); next++) { 370 Node* n = dom_list.at(next); 371 if (n == orig_sub) 372 return false; // One of dom's inputs dominated by sub. 373 if (!n->is_CFG() && n->pinned()) { 374 // Check only own control edge for pinned non-control nodes. 375 n = n->find_exact_control(n->in(0)); 376 if (n == NULL || n->is_top()) 377 return false; // Conservative answer for dead code 378 assert(n->is_CFG(), "expecting control"); 379 dom_list.push(n); 380 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 381 only_dominating_controls = true; 382 } else if (n->is_CFG()) { 383 if (n->dominates(sub, nlist)) 384 only_dominating_controls = true; 385 else 386 return false; 387 } else { 388 // First, own control edge. 389 Node* m = n->find_exact_control(n->in(0)); 390 if (m != NULL) { 391 if (m->is_top()) 392 return false; // Conservative answer for dead code 393 dom_list.push(m); 394 } 395 // Now, the rest of edges. 396 uint cnt = n->req(); 397 for (uint i = 1; i < cnt; i++) { 398 m = n->find_exact_control(n->in(i)); 399 if (m == NULL || m->is_top()) 400 continue; 401 dom_list.push(m); 402 } 403 } 404 } 405 return only_dominating_controls; 406 } 407 } 408 409 //---------------------detect_ptr_independence--------------------------------- 410 // Used by MemNode::find_previous_store to prove that two base 411 // pointers are never equal. 412 // The pointers are accompanied by their associated allocations, 413 // if any, which have been previously discovered by the caller. 414 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 415 Node* p2, AllocateNode* a2, 416 PhaseTransform* phase) { 417 // Attempt to prove that these two pointers cannot be aliased. 418 // They may both manifestly be allocations, and they should differ. 419 // Or, if they are not both allocations, they can be distinct constants. 420 // Otherwise, one is an allocation and the other a pre-existing value. 421 if (a1 == NULL && a2 == NULL) { // neither an allocation 422 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 423 } else if (a1 != NULL && a2 != NULL) { // both allocations 424 return (a1 != a2); 425 } else if (a1 != NULL) { // one allocation a1 426 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 427 return all_controls_dominate(p2, a1); 428 } else { //(a2 != NULL) // one allocation a2 429 return all_controls_dominate(p1, a2); 430 } 431 return false; 432 } 433 434 435 // The logic for reordering loads and stores uses four steps: 436 // (a) Walk carefully past stores and initializations which we 437 // can prove are independent of this load. 438 // (b) Observe that the next memory state makes an exact match 439 // with self (load or store), and locate the relevant store. 440 // (c) Ensure that, if we were to wire self directly to the store, 441 // the optimizer would fold it up somehow. 442 // (d) Do the rewiring, and return, depending on some other part of 443 // the optimizer to fold up the load. 444 // This routine handles steps (a) and (b). Steps (c) and (d) are 445 // specific to loads and stores, so they are handled by the callers. 446 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 447 // 448 Node* MemNode::find_previous_store(PhaseTransform* phase) { 449 Node* ctrl = in(MemNode::Control); 450 Node* adr = in(MemNode::Address); 451 intptr_t offset = 0; 452 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 453 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 454 455 if (offset == Type::OffsetBot) 456 return NULL; // cannot unalias unless there are precise offsets 457 458 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 459 460 intptr_t size_in_bytes = memory_size(); 461 462 Node* mem = in(MemNode::Memory); // start searching here... 463 464 int cnt = 50; // Cycle limiter 465 for (;;) { // While we can dance past unrelated stores... 466 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 467 468 if (mem->is_Store()) { 469 Node* st_adr = mem->in(MemNode::Address); 470 intptr_t st_offset = 0; 471 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 472 if (st_base == NULL) 473 break; // inscrutable pointer 474 if (st_offset != offset && st_offset != Type::OffsetBot) { 475 const int MAX_STORE = BytesPerLong; 476 if (st_offset >= offset + size_in_bytes || 477 st_offset <= offset - MAX_STORE || 478 st_offset <= offset - mem->as_Store()->memory_size()) { 479 // Success: The offsets are provably independent. 480 // (You may ask, why not just test st_offset != offset and be done? 481 // The answer is that stores of different sizes can co-exist 482 // in the same sequence of RawMem effects. We sometimes initialize 483 // a whole 'tile' of array elements with a single jint or jlong.) 484 mem = mem->in(MemNode::Memory); 485 continue; // (a) advance through independent store memory 486 } 487 } 488 if (st_base != base && 489 detect_ptr_independence(base, alloc, 490 st_base, 491 AllocateNode::Ideal_allocation(st_base, phase), 492 phase)) { 493 // Success: The bases are provably independent. 494 mem = mem->in(MemNode::Memory); 495 continue; // (a) advance through independent store memory 496 } 497 498 // (b) At this point, if the bases or offsets do not agree, we lose, 499 // since we have not managed to prove 'this' and 'mem' independent. 500 if (st_base == base && st_offset == offset) { 501 return mem; // let caller handle steps (c), (d) 502 } 503 504 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 505 InitializeNode* st_init = mem->in(0)->as_Initialize(); 506 AllocateNode* st_alloc = st_init->allocation(); 507 if (st_alloc == NULL) 508 break; // something degenerated 509 bool known_identical = false; 510 bool known_independent = false; 511 if (alloc == st_alloc) 512 known_identical = true; 513 else if (alloc != NULL) 514 known_independent = true; 515 else if (all_controls_dominate(this, st_alloc)) 516 known_independent = true; 517 518 if (known_independent) { 519 // The bases are provably independent: Either they are 520 // manifestly distinct allocations, or else the control 521 // of this load dominates the store's allocation. 522 int alias_idx = phase->C->get_alias_index(adr_type()); 523 if (alias_idx == Compile::AliasIdxRaw) { 524 mem = st_alloc->in(TypeFunc::Memory); 525 } else { 526 mem = st_init->memory(alias_idx); 527 } 528 continue; // (a) advance through independent store memory 529 } 530 531 // (b) at this point, if we are not looking at a store initializing 532 // the same allocation we are loading from, we lose. 533 if (known_identical) { 534 // From caller, can_see_stored_value will consult find_captured_store. 535 return mem; // let caller handle steps (c), (d) 536 } 537 538 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 539 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 540 if (mem->is_Proj() && mem->in(0)->is_Call()) { 541 CallNode *call = mem->in(0)->as_Call(); 542 if (!call->may_modify(addr_t, phase)) { 543 mem = call->in(TypeFunc::Memory); 544 continue; // (a) advance through independent call memory 545 } 546 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 547 mem = mem->in(0)->in(TypeFunc::Memory); 548 continue; // (a) advance through independent MemBar memory 549 } else if (mem->is_ClearArray()) { 550 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 551 // (the call updated 'mem' value) 552 continue; // (a) advance through independent allocation memory 553 } else { 554 // Can not bypass initialization of the instance 555 // we are looking for. 556 return mem; 557 } 558 } else if (mem->is_MergeMem()) { 559 int alias_idx = phase->C->get_alias_index(adr_type()); 560 mem = mem->as_MergeMem()->memory_at(alias_idx); 561 continue; // (a) advance through independent MergeMem memory 562 } 563 } 564 565 // Unless there is an explicit 'continue', we must bail out here, 566 // because 'mem' is an inscrutable memory state (e.g., a call). 567 break; 568 } 569 570 return NULL; // bail out 571 } 572 573 //----------------------calculate_adr_type------------------------------------- 574 // Helper function. Notices when the given type of address hits top or bottom. 575 // Also, asserts a cross-check of the type against the expected address type. 576 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 577 if (t == Type::TOP) return NULL; // does not touch memory any more? 578 #ifdef PRODUCT 579 cross_check = NULL; 580 #else 581 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 582 #endif 583 const TypePtr* tp = t->isa_ptr(); 584 if (tp == NULL) { 585 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 586 return TypePtr::BOTTOM; // touches lots of memory 587 } else { 588 #ifdef ASSERT 589 // %%%% [phh] We don't check the alias index if cross_check is 590 // TypeRawPtr::BOTTOM. Needs to be investigated. 591 if (cross_check != NULL && 592 cross_check != TypePtr::BOTTOM && 593 cross_check != TypeRawPtr::BOTTOM) { 594 // Recheck the alias index, to see if it has changed (due to a bug). 595 Compile* C = Compile::current(); 596 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 597 "must stay in the original alias category"); 598 // The type of the address must be contained in the adr_type, 599 // disregarding "null"-ness. 600 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 601 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 602 assert(cross_check->meet(tp_notnull) == cross_check, 603 "real address must not escape from expected memory type"); 604 } 605 #endif 606 return tp; 607 } 608 } 609 610 //------------------------adr_phi_is_loop_invariant---------------------------- 611 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 612 // loop is loop invariant. Make a quick traversal of Phi and associated 613 // CastPP nodes, looking to see if they are a closed group within the loop. 614 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 615 // The idea is that the phi-nest must boil down to only CastPP nodes 616 // with the same data. This implies that any path into the loop already 617 // includes such a CastPP, and so the original cast, whatever its input, 618 // must be covered by an equivalent cast, with an earlier control input. 619 ResourceMark rm; 620 621 // The loop entry input of the phi should be the unique dominating 622 // node for every Phi/CastPP in the loop. 623 Unique_Node_List closure; 624 closure.push(adr_phi->in(LoopNode::EntryControl)); 625 626 // Add the phi node and the cast to the worklist. 627 Unique_Node_List worklist; 628 worklist.push(adr_phi); 629 if( cast != NULL ){ 630 if( !cast->is_ConstraintCast() ) return false; 631 worklist.push(cast); 632 } 633 634 // Begin recursive walk of phi nodes. 635 while( worklist.size() ){ 636 // Take a node off the worklist 637 Node *n = worklist.pop(); 638 if( !closure.member(n) ){ 639 // Add it to the closure. 640 closure.push(n); 641 // Make a sanity check to ensure we don't waste too much time here. 642 if( closure.size() > 20) return false; 643 // This node is OK if: 644 // - it is a cast of an identical value 645 // - or it is a phi node (then we add its inputs to the worklist) 646 // Otherwise, the node is not OK, and we presume the cast is not invariant 647 if( n->is_ConstraintCast() ){ 648 worklist.push(n->in(1)); 649 } else if( n->is_Phi() ) { 650 for( uint i = 1; i < n->req(); i++ ) { 651 worklist.push(n->in(i)); 652 } 653 } else { 654 return false; 655 } 656 } 657 } 658 659 // Quit when the worklist is empty, and we've found no offending nodes. 660 return true; 661 } 662 663 //------------------------------Ideal_DU_postCCP------------------------------- 664 // Find any cast-away of null-ness and keep its control. Null cast-aways are 665 // going away in this pass and we need to make this memory op depend on the 666 // gating null check. 667 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 668 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 669 } 670 671 // I tried to leave the CastPP's in. This makes the graph more accurate in 672 // some sense; we get to keep around the knowledge that an oop is not-null 673 // after some test. Alas, the CastPP's interfere with GVN (some values are 674 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 675 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 676 // some of the more trivial cases in the optimizer. Removing more useless 677 // Phi's started allowing Loads to illegally float above null checks. I gave 678 // up on this approach. CNC 10/20/2000 679 // This static method may be called not from MemNode (EncodePNode calls it). 680 // Only the control edge of the node 'n' might be updated. 681 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 682 Node *skipped_cast = NULL; 683 // Need a null check? Regular static accesses do not because they are 684 // from constant addresses. Array ops are gated by the range check (which 685 // always includes a NULL check). Just check field ops. 686 if( n->in(MemNode::Control) == NULL ) { 687 // Scan upwards for the highest location we can place this memory op. 688 while( true ) { 689 switch( adr->Opcode() ) { 690 691 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 692 adr = adr->in(AddPNode::Base); 693 continue; 694 695 case Op_DecodeN: // No change to NULL-ness, so peek thru 696 adr = adr->in(1); 697 continue; 698 699 case Op_CastPP: 700 // If the CastPP is useless, just peek on through it. 701 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 702 // Remember the cast that we've peeked though. If we peek 703 // through more than one, then we end up remembering the highest 704 // one, that is, if in a loop, the one closest to the top. 705 skipped_cast = adr; 706 adr = adr->in(1); 707 continue; 708 } 709 // CastPP is going away in this pass! We need this memory op to be 710 // control-dependent on the test that is guarding the CastPP. 711 ccp->hash_delete(n); 712 n->set_req(MemNode::Control, adr->in(0)); 713 ccp->hash_insert(n); 714 return n; 715 716 case Op_Phi: 717 // Attempt to float above a Phi to some dominating point. 718 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 719 // If we've already peeked through a Cast (which could have set the 720 // control), we can't float above a Phi, because the skipped Cast 721 // may not be loop invariant. 722 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 723 adr = adr->in(1); 724 continue; 725 } 726 } 727 728 // Intentional fallthrough! 729 730 // No obvious dominating point. The mem op is pinned below the Phi 731 // by the Phi itself. If the Phi goes away (no true value is merged) 732 // then the mem op can float, but not indefinitely. It must be pinned 733 // behind the controls leading to the Phi. 734 case Op_CheckCastPP: 735 // These usually stick around to change address type, however a 736 // useless one can be elided and we still need to pick up a control edge 737 if (adr->in(0) == NULL) { 738 // This CheckCastPP node has NO control and is likely useless. But we 739 // need check further up the ancestor chain for a control input to keep 740 // the node in place. 4959717. 741 skipped_cast = adr; 742 adr = adr->in(1); 743 continue; 744 } 745 ccp->hash_delete(n); 746 n->set_req(MemNode::Control, adr->in(0)); 747 ccp->hash_insert(n); 748 return n; 749 750 // List of "safe" opcodes; those that implicitly block the memory 751 // op below any null check. 752 case Op_CastX2P: // no null checks on native pointers 753 case Op_Parm: // 'this' pointer is not null 754 case Op_LoadP: // Loading from within a klass 755 case Op_LoadN: // Loading from within a klass 756 case Op_LoadKlass: // Loading from within a klass 757 case Op_LoadNKlass: // Loading from within a klass 758 case Op_ConP: // Loading from a klass 759 case Op_ConN: // Loading from a klass 760 case Op_CreateEx: // Sucking up the guts of an exception oop 761 case Op_Con: // Reading from TLS 762 case Op_CMoveP: // CMoveP is pinned 763 case Op_CMoveN: // CMoveN is pinned 764 break; // No progress 765 766 case Op_Proj: // Direct call to an allocation routine 767 case Op_SCMemProj: // Memory state from store conditional ops 768 #ifdef ASSERT 769 { 770 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 771 const Node* call = adr->in(0); 772 if (call->is_CallJava()) { 773 const CallJavaNode* call_java = call->as_CallJava(); 774 const TypeTuple *r = call_java->tf()->range(); 775 assert(r->cnt() > TypeFunc::Parms, "must return value"); 776 const Type* ret_type = r->field_at(TypeFunc::Parms); 777 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 778 // We further presume that this is one of 779 // new_instance_Java, new_array_Java, or 780 // the like, but do not assert for this. 781 } else if (call->is_Allocate()) { 782 // similar case to new_instance_Java, etc. 783 } else if (!call->is_CallLeaf()) { 784 // Projections from fetch_oop (OSR) are allowed as well. 785 ShouldNotReachHere(); 786 } 787 } 788 #endif 789 break; 790 default: 791 ShouldNotReachHere(); 792 } 793 break; 794 } 795 } 796 797 return NULL; // No progress 798 } 799 800 801 //============================================================================= 802 uint LoadNode::size_of() const { return sizeof(*this); } 803 uint LoadNode::cmp( const Node &n ) const 804 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 805 const Type *LoadNode::bottom_type() const { return _type; } 806 uint LoadNode::ideal_reg() const { 807 return Matcher::base2reg[_type->base()]; 808 } 809 810 #ifndef PRODUCT 811 void LoadNode::dump_spec(outputStream *st) const { 812 MemNode::dump_spec(st); 813 if( !Verbose && !WizardMode ) { 814 // standard dump does this in Verbose and WizardMode 815 st->print(" #"); _type->dump_on(st); 816 } 817 } 818 #endif 819 820 #ifdef ASSERT 821 //----------------------------is_immutable_value------------------------------- 822 // Helper function to allow a raw load without control edge for some cases 823 bool LoadNode::is_immutable_value(Node* adr) { 824 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 825 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 826 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 827 in_bytes(JavaThread::osthread_offset()))); 828 } 829 #endif 830 831 //----------------------------LoadNode::make----------------------------------- 832 // Polymorphic factory method: 833 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 834 Compile* C = gvn.C; 835 836 // sanity check the alias category against the created node type 837 assert(!(adr_type->isa_oopptr() && 838 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 839 "use LoadKlassNode instead"); 840 assert(!(adr_type->isa_aryptr() && 841 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 842 "use LoadRangeNode instead"); 843 // Check control edge of raw loads 844 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 845 // oop will be recorded in oop map if load crosses safepoint 846 rt->isa_oopptr() || is_immutable_value(adr), 847 "raw memory operations should have control edge"); 848 switch (bt) { 849 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 850 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 851 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 852 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 853 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 854 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 855 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 856 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 857 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 858 case T_OBJECT: 859 #ifdef _LP64 860 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 861 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 862 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 863 } else 864 #endif 865 { 866 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 867 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 868 } 869 } 870 ShouldNotReachHere(); 871 return (LoadNode*)NULL; 872 } 873 874 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 875 bool require_atomic = true; 876 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 877 } 878 879 880 881 882 //------------------------------hash------------------------------------------- 883 uint LoadNode::hash() const { 884 // unroll addition of interesting fields 885 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 886 } 887 888 //---------------------------can_see_stored_value------------------------------ 889 // This routine exists to make sure this set of tests is done the same 890 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 891 // will change the graph shape in a way which makes memory alive twice at the 892 // same time (uses the Oracle model of aliasing), then some 893 // LoadXNode::Identity will fold things back to the equivalence-class model 894 // of aliasing. 895 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 896 Node* ld_adr = in(MemNode::Address); 897 898 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 899 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 900 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 901 atp->field() != NULL && !atp->field()->is_volatile()) { 902 uint alias_idx = atp->index(); 903 bool final = atp->field()->is_final(); 904 Node* result = NULL; 905 Node* current = st; 906 // Skip through chains of MemBarNodes checking the MergeMems for 907 // new states for the slice of this load. Stop once any other 908 // kind of node is encountered. Loads from final memory can skip 909 // through any kind of MemBar but normal loads shouldn't skip 910 // through MemBarAcquire since the could allow them to move out of 911 // a synchronized region. 912 while (current->is_Proj()) { 913 int opc = current->in(0)->Opcode(); 914 if ((final && opc == Op_MemBarAcquire) || 915 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 916 Node* mem = current->in(0)->in(TypeFunc::Memory); 917 if (mem->is_MergeMem()) { 918 MergeMemNode* merge = mem->as_MergeMem(); 919 Node* new_st = merge->memory_at(alias_idx); 920 if (new_st == merge->base_memory()) { 921 // Keep searching 922 current = merge->base_memory(); 923 continue; 924 } 925 // Save the new memory state for the slice and fall through 926 // to exit. 927 result = new_st; 928 } 929 } 930 break; 931 } 932 if (result != NULL) { 933 st = result; 934 } 935 } 936 937 938 // Loop around twice in the case Load -> Initialize -> Store. 939 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 940 for (int trip = 0; trip <= 1; trip++) { 941 942 if (st->is_Store()) { 943 Node* st_adr = st->in(MemNode::Address); 944 if (!phase->eqv(st_adr, ld_adr)) { 945 // Try harder before giving up... Match raw and non-raw pointers. 946 intptr_t st_off = 0; 947 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 948 if (alloc == NULL) return NULL; 949 intptr_t ld_off = 0; 950 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 951 if (alloc != allo2) return NULL; 952 if (ld_off != st_off) return NULL; 953 // At this point we have proven something like this setup: 954 // A = Allocate(...) 955 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 956 // S = StoreQ(, AddP(, A.Parm , #Off), V) 957 // (Actually, we haven't yet proven the Q's are the same.) 958 // In other words, we are loading from a casted version of 959 // the same pointer-and-offset that we stored to. 960 // Thus, we are able to replace L by V. 961 } 962 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 963 if (store_Opcode() != st->Opcode()) 964 return NULL; 965 return st->in(MemNode::ValueIn); 966 } 967 968 intptr_t offset = 0; // scratch 969 970 // A load from a freshly-created object always returns zero. 971 // (This can happen after LoadNode::Ideal resets the load's memory input 972 // to find_captured_store, which returned InitializeNode::zero_memory.) 973 if (st->is_Proj() && st->in(0)->is_Allocate() && 974 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 975 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 976 // return a zero value for the load's basic type 977 // (This is one of the few places where a generic PhaseTransform 978 // can create new nodes. Think of it as lazily manifesting 979 // virtually pre-existing constants.) 980 return phase->zerocon(memory_type()); 981 } 982 983 // A load from an initialization barrier can match a captured store. 984 if (st->is_Proj() && st->in(0)->is_Initialize()) { 985 InitializeNode* init = st->in(0)->as_Initialize(); 986 AllocateNode* alloc = init->allocation(); 987 if (alloc != NULL && 988 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 989 // examine a captured store value 990 st = init->find_captured_store(offset, memory_size(), phase); 991 if (st != NULL) 992 continue; // take one more trip around 993 } 994 } 995 996 break; 997 } 998 999 return NULL; 1000 } 1001 1002 //----------------------is_instance_field_load_with_local_phi------------------ 1003 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1004 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 1005 in(MemNode::Address)->is_AddP() ) { 1006 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 1007 // Only instances. 1008 if( t_oop != NULL && t_oop->is_known_instance_field() && 1009 t_oop->offset() != Type::OffsetBot && 1010 t_oop->offset() != Type::OffsetTop) { 1011 return true; 1012 } 1013 } 1014 return false; 1015 } 1016 1017 //------------------------------Identity--------------------------------------- 1018 // Loads are identity if previous store is to same address 1019 Node *LoadNode::Identity( PhaseTransform *phase ) { 1020 // If the previous store-maker is the right kind of Store, and the store is 1021 // to the same address, then we are equal to the value stored. 1022 Node* mem = in(MemNode::Memory); 1023 Node* value = can_see_stored_value(mem, phase); 1024 if( value ) { 1025 // byte, short & char stores truncate naturally. 1026 // A load has to load the truncated value which requires 1027 // some sort of masking operation and that requires an 1028 // Ideal call instead of an Identity call. 1029 if (memory_size() < BytesPerInt) { 1030 // If the input to the store does not fit with the load's result type, 1031 // it must be truncated via an Ideal call. 1032 if (!phase->type(value)->higher_equal(phase->type(this))) 1033 return this; 1034 } 1035 // (This works even when value is a Con, but LoadNode::Value 1036 // usually runs first, producing the singleton type of the Con.) 1037 return value; 1038 } 1039 1040 // Search for an existing data phi which was generated before for the same 1041 // instance's field to avoid infinite generation of phis in a loop. 1042 Node *region = mem->in(0); 1043 if (is_instance_field_load_with_local_phi(region)) { 1044 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 1045 int this_index = phase->C->get_alias_index(addr_t); 1046 int this_offset = addr_t->offset(); 1047 int this_id = addr_t->is_oopptr()->instance_id(); 1048 const Type* this_type = bottom_type(); 1049 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1050 Node* phi = region->fast_out(i); 1051 if (phi->is_Phi() && phi != mem && 1052 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 1053 return phi; 1054 } 1055 } 1056 } 1057 1058 return this; 1059 } 1060 1061 1062 // Returns true if the AliasType refers to the field that holds the 1063 // cached box array. Currently only handles the IntegerCache case. 1064 static bool is_autobox_cache(Compile::AliasType* atp) { 1065 if (atp != NULL && atp->field() != NULL) { 1066 ciField* field = atp->field(); 1067 ciSymbol* klass = field->holder()->name(); 1068 if (field->name() == ciSymbol::cache_field_name() && 1069 field->holder()->uses_default_loader() && 1070 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1071 return true; 1072 } 1073 } 1074 return false; 1075 } 1076 1077 // Fetch the base value in the autobox array 1078 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1079 if (atp != NULL && atp->field() != NULL) { 1080 ciField* field = atp->field(); 1081 ciSymbol* klass = field->holder()->name(); 1082 if (field->name() == ciSymbol::cache_field_name() && 1083 field->holder()->uses_default_loader() && 1084 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1085 assert(field->is_constant(), "what?"); 1086 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1087 // Fetch the box object at the base of the array and get its value 1088 ciInstance* box = array->obj_at(0)->as_instance(); 1089 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1090 if (ik->nof_nonstatic_fields() == 1) { 1091 // This should be true nonstatic_field_at requires calling 1092 // nof_nonstatic_fields so check it anyway 1093 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1094 cache_offset = c.as_int(); 1095 } 1096 return true; 1097 } 1098 } 1099 return false; 1100 } 1101 1102 // Returns true if the AliasType refers to the value field of an 1103 // autobox object. Currently only handles Integer. 1104 static bool is_autobox_object(Compile::AliasType* atp) { 1105 if (atp != NULL && atp->field() != NULL) { 1106 ciField* field = atp->field(); 1107 ciSymbol* klass = field->holder()->name(); 1108 if (field->name() == ciSymbol::value_name() && 1109 field->holder()->uses_default_loader() && 1110 klass == ciSymbol::java_lang_Integer()) { 1111 return true; 1112 } 1113 } 1114 return false; 1115 } 1116 1117 1118 // We're loading from an object which has autobox behaviour. 1119 // If this object is result of a valueOf call we'll have a phi 1120 // merging a newly allocated object and a load from the cache. 1121 // We want to replace this load with the original incoming 1122 // argument to the valueOf call. 1123 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1124 Node* base = in(Address)->in(AddPNode::Base); 1125 if (base->is_Phi() && base->req() == 3) { 1126 AllocateNode* allocation = NULL; 1127 int allocation_index = -1; 1128 int load_index = -1; 1129 for (uint i = 1; i < base->req(); i++) { 1130 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1131 if (allocation != NULL) { 1132 allocation_index = i; 1133 load_index = 3 - allocation_index; 1134 break; 1135 } 1136 } 1137 bool has_load = ( allocation != NULL && 1138 (base->in(load_index)->is_Load() || 1139 base->in(load_index)->is_DecodeN() && 1140 base->in(load_index)->in(1)->is_Load()) ); 1141 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1142 // Push the loads from the phi that comes from valueOf up 1143 // through it to allow elimination of the loads and the recovery 1144 // of the original value. 1145 Node* mem_phi = in(Memory); 1146 Node* offset = in(Address)->in(AddPNode::Offset); 1147 Node* region = base->in(0); 1148 1149 Node* in1 = clone(); 1150 Node* in1_addr = in1->in(Address)->clone(); 1151 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1152 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1153 in1_addr->set_req(AddPNode::Offset, offset); 1154 in1->set_req(0, region->in(allocation_index)); 1155 in1->set_req(Address, in1_addr); 1156 in1->set_req(Memory, mem_phi->in(allocation_index)); 1157 1158 Node* in2 = clone(); 1159 Node* in2_addr = in2->in(Address)->clone(); 1160 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1161 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1162 in2_addr->set_req(AddPNode::Offset, offset); 1163 in2->set_req(0, region->in(load_index)); 1164 in2->set_req(Address, in2_addr); 1165 in2->set_req(Memory, mem_phi->in(load_index)); 1166 1167 in1_addr = phase->transform(in1_addr); 1168 in1 = phase->transform(in1); 1169 in2_addr = phase->transform(in2_addr); 1170 in2 = phase->transform(in2); 1171 1172 PhiNode* result = PhiNode::make_blank(region, this); 1173 result->set_req(allocation_index, in1); 1174 result->set_req(load_index, in2); 1175 return result; 1176 } 1177 } else if (base->is_Load() || 1178 base->is_DecodeN() && base->in(1)->is_Load()) { 1179 if (base->is_DecodeN()) { 1180 // Get LoadN node which loads cached Integer object 1181 base = base->in(1); 1182 } 1183 // Eliminate the load of Integer.value for integers from the cache 1184 // array by deriving the value from the index into the array. 1185 // Capture the offset of the load and then reverse the computation. 1186 Node* load_base = base->in(Address)->in(AddPNode::Base); 1187 if (load_base->is_DecodeN()) { 1188 // Get LoadN node which loads IntegerCache.cache field 1189 load_base = load_base->in(1); 1190 } 1191 if (load_base != NULL) { 1192 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1193 intptr_t cache_offset; 1194 int shift = -1; 1195 Node* cache = NULL; 1196 if (is_autobox_cache(atp)) { 1197 shift = exact_log2(type2aelembytes(T_OBJECT)); 1198 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1199 } 1200 if (cache != NULL && base->in(Address)->is_AddP()) { 1201 Node* elements[4]; 1202 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1203 int cache_low; 1204 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1205 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1206 // Add up all the offsets making of the address of the load 1207 Node* result = elements[0]; 1208 for (int i = 1; i < count; i++) { 1209 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1210 } 1211 // Remove the constant offset from the address and then 1212 // remove the scaling of the offset to recover the original index. 1213 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1214 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1215 // Peel the shift off directly but wrap it in a dummy node 1216 // since Ideal can't return existing nodes 1217 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1218 } else { 1219 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1220 } 1221 #ifdef _LP64 1222 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1223 #endif 1224 return result; 1225 } 1226 } 1227 } 1228 } 1229 return NULL; 1230 } 1231 1232 //------------------------------split_through_phi------------------------------ 1233 // Split instance field load through Phi. 1234 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1235 Node* mem = in(MemNode::Memory); 1236 Node* address = in(MemNode::Address); 1237 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1238 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1239 1240 assert(mem->is_Phi() && (t_oop != NULL) && 1241 t_oop->is_known_instance_field(), "invalide conditions"); 1242 1243 Node *region = mem->in(0); 1244 if (region == NULL) { 1245 return NULL; // Wait stable graph 1246 } 1247 uint cnt = mem->req(); 1248 for( uint i = 1; i < cnt; i++ ) { 1249 Node *in = mem->in(i); 1250 if( in == NULL ) { 1251 return NULL; // Wait stable graph 1252 } 1253 } 1254 // Check for loop invariant. 1255 if (cnt == 3) { 1256 for( uint i = 1; i < cnt; i++ ) { 1257 Node *in = mem->in(i); 1258 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1259 if (m == mem) { 1260 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1261 return this; 1262 } 1263 } 1264 } 1265 // Split through Phi (see original code in loopopts.cpp). 1266 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1267 1268 // Do nothing here if Identity will find a value 1269 // (to avoid infinite chain of value phis generation). 1270 if ( !phase->eqv(this, this->Identity(phase)) ) 1271 return NULL; 1272 1273 // Skip the split if the region dominates some control edge of the address. 1274 if (cnt == 3 && !MemNode::all_controls_dominate(address, region)) 1275 return NULL; 1276 1277 const Type* this_type = this->bottom_type(); 1278 int this_index = phase->C->get_alias_index(addr_t); 1279 int this_offset = addr_t->offset(); 1280 int this_iid = addr_t->is_oopptr()->instance_id(); 1281 int wins = 0; 1282 PhaseIterGVN *igvn = phase->is_IterGVN(); 1283 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1284 for( uint i = 1; i < region->req(); i++ ) { 1285 Node *x; 1286 Node* the_clone = NULL; 1287 if( region->in(i) == phase->C->top() ) { 1288 x = phase->C->top(); // Dead path? Use a dead data op 1289 } else { 1290 x = this->clone(); // Else clone up the data op 1291 the_clone = x; // Remember for possible deletion. 1292 // Alter data node to use pre-phi inputs 1293 if( this->in(0) == region ) { 1294 x->set_req( 0, region->in(i) ); 1295 } else { 1296 x->set_req( 0, NULL ); 1297 } 1298 for( uint j = 1; j < this->req(); j++ ) { 1299 Node *in = this->in(j); 1300 if( in->is_Phi() && in->in(0) == region ) 1301 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 1302 } 1303 } 1304 // Check for a 'win' on some paths 1305 const Type *t = x->Value(igvn); 1306 1307 bool singleton = t->singleton(); 1308 1309 // See comments in PhaseIdealLoop::split_thru_phi(). 1310 if( singleton && t == Type::TOP ) { 1311 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1312 } 1313 1314 if( singleton ) { 1315 wins++; 1316 x = igvn->makecon(t); 1317 } else { 1318 // We now call Identity to try to simplify the cloned node. 1319 // Note that some Identity methods call phase->type(this). 1320 // Make sure that the type array is big enough for 1321 // our new node, even though we may throw the node away. 1322 // (This tweaking with igvn only works because x is a new node.) 1323 igvn->set_type(x, t); 1324 // If x is a TypeNode, capture any more-precise type permanently into Node 1325 // otherwise it will be not updated during igvn->transform since 1326 // igvn->type(x) is set to x->Value() already. 1327 x->raise_bottom_type(t); 1328 Node *y = x->Identity(igvn); 1329 if( y != x ) { 1330 wins++; 1331 x = y; 1332 } else { 1333 y = igvn->hash_find(x); 1334 if( y ) { 1335 wins++; 1336 x = y; 1337 } else { 1338 // Else x is a new node we are keeping 1339 // We do not need register_new_node_with_optimizer 1340 // because set_type has already been called. 1341 igvn->_worklist.push(x); 1342 } 1343 } 1344 } 1345 if (x != the_clone && the_clone != NULL) 1346 igvn->remove_dead_node(the_clone); 1347 phi->set_req(i, x); 1348 } 1349 if( wins > 0 ) { 1350 // Record Phi 1351 igvn->register_new_node_with_optimizer(phi); 1352 return phi; 1353 } 1354 igvn->remove_dead_node(phi); 1355 return NULL; 1356 } 1357 1358 //------------------------------Ideal------------------------------------------ 1359 // If the load is from Field memory and the pointer is non-null, we can 1360 // zero out the control input. 1361 // If the offset is constant and the base is an object allocation, 1362 // try to hook me up to the exact initializing store. 1363 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1364 Node* p = MemNode::Ideal_common(phase, can_reshape); 1365 if (p) return (p == NodeSentinel) ? NULL : p; 1366 1367 Node* ctrl = in(MemNode::Control); 1368 Node* address = in(MemNode::Address); 1369 1370 // Skip up past a SafePoint control. Cannot do this for Stores because 1371 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1372 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1373 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1374 ctrl = ctrl->in(0); 1375 set_req(MemNode::Control,ctrl); 1376 } 1377 1378 intptr_t ignore = 0; 1379 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1380 if (base != NULL 1381 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1382 // Check for useless control edge in some common special cases 1383 if (in(MemNode::Control) != NULL 1384 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1385 && all_controls_dominate(base, phase->C->start())) { 1386 // A method-invariant, non-null address (constant or 'this' argument). 1387 set_req(MemNode::Control, NULL); 1388 } 1389 1390 if (EliminateAutoBox && can_reshape) { 1391 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); 1392 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1393 if (is_autobox_object(atp)) { 1394 Node* result = eliminate_autobox(phase); 1395 if (result != NULL) return result; 1396 } 1397 } 1398 } 1399 1400 Node* mem = in(MemNode::Memory); 1401 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1402 1403 if (addr_t != NULL) { 1404 // try to optimize our memory input 1405 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1406 if (opt_mem != mem) { 1407 set_req(MemNode::Memory, opt_mem); 1408 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1409 return this; 1410 } 1411 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1412 if (can_reshape && opt_mem->is_Phi() && 1413 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1414 // Split instance field load through Phi. 1415 Node* result = split_through_phi(phase); 1416 if (result != NULL) return result; 1417 } 1418 } 1419 1420 // Check for prior store with a different base or offset; make Load 1421 // independent. Skip through any number of them. Bail out if the stores 1422 // are in an endless dead cycle and report no progress. This is a key 1423 // transform for Reflection. However, if after skipping through the Stores 1424 // we can't then fold up against a prior store do NOT do the transform as 1425 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1426 // array memory alive twice: once for the hoisted Load and again after the 1427 // bypassed Store. This situation only works if EVERYBODY who does 1428 // anti-dependence work knows how to bypass. I.e. we need all 1429 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1430 // the alias index stuff. So instead, peek through Stores and IFF we can 1431 // fold up, do so. 1432 Node* prev_mem = find_previous_store(phase); 1433 // Steps (a), (b): Walk past independent stores to find an exact match. 1434 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1435 // (c) See if we can fold up on the spot, but don't fold up here. 1436 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1437 // just return a prior value, which is done by Identity calls. 1438 if (can_see_stored_value(prev_mem, phase)) { 1439 // Make ready for step (d): 1440 set_req(MemNode::Memory, prev_mem); 1441 return this; 1442 } 1443 } 1444 1445 return NULL; // No further progress 1446 } 1447 1448 // Helper to recognize certain Klass fields which are invariant across 1449 // some group of array types (e.g., int[] or all T[] where T < Object). 1450 const Type* 1451 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1452 ciKlass* klass) const { 1453 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1454 // The field is Klass::_modifier_flags. Return its (constant) value. 1455 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1456 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1457 return TypeInt::make(klass->modifier_flags()); 1458 } 1459 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1460 // The field is Klass::_access_flags. Return its (constant) value. 1461 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1462 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1463 return TypeInt::make(klass->access_flags()); 1464 } 1465 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1466 // The field is Klass::_layout_helper. Return its constant value if known. 1467 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1468 return TypeInt::make(klass->layout_helper()); 1469 } 1470 1471 // No match. 1472 return NULL; 1473 } 1474 1475 //------------------------------Value----------------------------------------- 1476 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1477 // Either input is TOP ==> the result is TOP 1478 Node* mem = in(MemNode::Memory); 1479 const Type *t1 = phase->type(mem); 1480 if (t1 == Type::TOP) return Type::TOP; 1481 Node* adr = in(MemNode::Address); 1482 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1483 if (tp == NULL || tp->empty()) return Type::TOP; 1484 int off = tp->offset(); 1485 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1486 1487 // Try to guess loaded type from pointer type 1488 if (tp->base() == Type::AryPtr) { 1489 const Type *t = tp->is_aryptr()->elem(); 1490 // Don't do this for integer types. There is only potential profit if 1491 // the element type t is lower than _type; that is, for int types, if _type is 1492 // more restrictive than t. This only happens here if one is short and the other 1493 // char (both 16 bits), and in those cases we've made an intentional decision 1494 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1495 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1496 // 1497 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1498 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1499 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1500 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1501 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1502 // In fact, that could have been the original type of p1, and p1 could have 1503 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1504 // expression (LShiftL quux 3) independently optimized to the constant 8. 1505 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1506 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1507 // t might actually be lower than _type, if _type is a unique 1508 // concrete subclass of abstract class t. 1509 // Make sure the reference is not into the header, by comparing 1510 // the offset against the offset of the start of the array's data. 1511 // Different array types begin at slightly different offsets (12 vs. 16). 1512 // We choose T_BYTE as an example base type that is least restrictive 1513 // as to alignment, which will therefore produce the smallest 1514 // possible base offset. 1515 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1516 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1517 const Type* jt = t->join(_type); 1518 // In any case, do not allow the join, per se, to empty out the type. 1519 if (jt->empty() && !t->empty()) { 1520 // This can happen if a interface-typed array narrows to a class type. 1521 jt = _type; 1522 } 1523 1524 if (EliminateAutoBox && adr->is_AddP()) { 1525 // The pointers in the autobox arrays are always non-null 1526 Node* base = adr->in(AddPNode::Base); 1527 if (base != NULL && 1528 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { 1529 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1530 if (is_autobox_cache(atp)) { 1531 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1532 } 1533 } 1534 } 1535 return jt; 1536 } 1537 } 1538 } else if (tp->base() == Type::InstPtr) { 1539 const TypeInstPtr* tinst = tp->is_instptr(); 1540 ciKlass* klass = tinst->klass(); 1541 assert( off != Type::OffsetBot || 1542 // arrays can be cast to Objects 1543 tp->is_oopptr()->klass()->is_java_lang_Object() || 1544 // unsafe field access may not have a constant offset 1545 phase->C->has_unsafe_access(), 1546 "Field accesses must be precise" ); 1547 // For oop loads, we expect the _type to be precise 1548 if (OptimizeStringConcat && klass == phase->C->env()->String_klass() && 1549 adr->is_AddP() && off != Type::OffsetBot) { 1550 // For constant Strings treat the fields as compile time constants. 1551 Node* base = adr->in(AddPNode::Base); 1552 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1553 if (t != NULL && t->singleton()) { 1554 ciObject* string = t->const_oop(); 1555 ciConstant constant = string->as_instance()->field_value_by_offset(off); 1556 if (constant.basic_type() == T_INT) { 1557 return TypeInt::make(constant.as_int()); 1558 } else if (constant.basic_type() == T_ARRAY) { 1559 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1560 return TypeNarrowOop::make_from_constant(constant.as_object()); 1561 } else { 1562 return TypeOopPtr::make_from_constant(constant.as_object()); 1563 } 1564 } 1565 } 1566 } 1567 } else if (tp->base() == Type::KlassPtr) { 1568 assert( off != Type::OffsetBot || 1569 // arrays can be cast to Objects 1570 tp->is_klassptr()->klass()->is_java_lang_Object() || 1571 // also allow array-loading from the primary supertype 1572 // array during subtype checks 1573 Opcode() == Op_LoadKlass, 1574 "Field accesses must be precise" ); 1575 // For klass/static loads, we expect the _type to be precise 1576 } 1577 1578 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1579 if (tkls != NULL && !StressReflectiveCode) { 1580 ciKlass* klass = tkls->klass(); 1581 if (klass->is_loaded() && tkls->klass_is_exact()) { 1582 // We are loading a field from a Klass metaobject whose identity 1583 // is known at compile time (the type is "exact" or "precise"). 1584 // Check for fields we know are maintained as constants by the VM. 1585 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1586 // The field is Klass::_super_check_offset. Return its (constant) value. 1587 // (Folds up type checking code.) 1588 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1589 return TypeInt::make(klass->super_check_offset()); 1590 } 1591 // Compute index into primary_supers array 1592 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1593 // Check for overflowing; use unsigned compare to handle the negative case. 1594 if( depth < ciKlass::primary_super_limit() ) { 1595 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1596 // (Folds up type checking code.) 1597 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1598 ciKlass *ss = klass->super_of_depth(depth); 1599 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1600 } 1601 const Type* aift = load_array_final_field(tkls, klass); 1602 if (aift != NULL) return aift; 1603 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1604 && klass->is_array_klass()) { 1605 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1606 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1607 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1608 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1609 } 1610 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1611 // The field is Klass::_java_mirror. Return its (constant) value. 1612 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1613 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1614 return TypeInstPtr::make(klass->java_mirror()); 1615 } 1616 } 1617 1618 // We can still check if we are loading from the primary_supers array at a 1619 // shallow enough depth. Even though the klass is not exact, entries less 1620 // than or equal to its super depth are correct. 1621 if (klass->is_loaded() ) { 1622 ciType *inner = klass->klass(); 1623 while( inner->is_obj_array_klass() ) 1624 inner = inner->as_obj_array_klass()->base_element_type(); 1625 if( inner->is_instance_klass() && 1626 !inner->as_instance_klass()->flags().is_interface() ) { 1627 // Compute index into primary_supers array 1628 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1629 // Check for overflowing; use unsigned compare to handle the negative case. 1630 if( depth < ciKlass::primary_super_limit() && 1631 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1632 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1633 // (Folds up type checking code.) 1634 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1635 ciKlass *ss = klass->super_of_depth(depth); 1636 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1637 } 1638 } 1639 } 1640 1641 // If the type is enough to determine that the thing is not an array, 1642 // we can give the layout_helper a positive interval type. 1643 // This will help short-circuit some reflective code. 1644 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1645 && !klass->is_array_klass() // not directly typed as an array 1646 && !klass->is_interface() // specifically not Serializable & Cloneable 1647 && !klass->is_java_lang_Object() // not the supertype of all T[] 1648 ) { 1649 // Note: When interfaces are reliable, we can narrow the interface 1650 // test to (klass != Serializable && klass != Cloneable). 1651 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1652 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1653 // The key property of this type is that it folds up tests 1654 // for array-ness, since it proves that the layout_helper is positive. 1655 // Thus, a generic value like the basic object layout helper works fine. 1656 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1657 } 1658 } 1659 1660 // If we are loading from a freshly-allocated object, produce a zero, 1661 // if the load is provably beyond the header of the object. 1662 // (Also allow a variable load from a fresh array to produce zero.) 1663 if (ReduceFieldZeroing) { 1664 Node* value = can_see_stored_value(mem,phase); 1665 if (value != NULL && value->is_Con()) 1666 return value->bottom_type(); 1667 } 1668 1669 const TypeOopPtr *tinst = tp->isa_oopptr(); 1670 if (tinst != NULL && tinst->is_known_instance_field()) { 1671 // If we have an instance type and our memory input is the 1672 // programs's initial memory state, there is no matching store, 1673 // so just return a zero of the appropriate type 1674 Node *mem = in(MemNode::Memory); 1675 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1676 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1677 return Type::get_zero_type(_type->basic_type()); 1678 } 1679 } 1680 return _type; 1681 } 1682 1683 //------------------------------match_edge------------------------------------- 1684 // Do we Match on this edge index or not? Match only the address. 1685 uint LoadNode::match_edge(uint idx) const { 1686 return idx == MemNode::Address; 1687 } 1688 1689 //--------------------------LoadBNode::Ideal-------------------------------------- 1690 // 1691 // If the previous store is to the same address as this load, 1692 // and the value stored was larger than a byte, replace this load 1693 // with the value stored truncated to a byte. If no truncation is 1694 // needed, the replacement is done in LoadNode::Identity(). 1695 // 1696 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1697 Node* mem = in(MemNode::Memory); 1698 Node* value = can_see_stored_value(mem,phase); 1699 if( value && !phase->type(value)->higher_equal( _type ) ) { 1700 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1701 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1702 } 1703 // Identity call will handle the case where truncation is not needed. 1704 return LoadNode::Ideal(phase, can_reshape); 1705 } 1706 1707 //--------------------------LoadUBNode::Ideal------------------------------------- 1708 // 1709 // If the previous store is to the same address as this load, 1710 // and the value stored was larger than a byte, replace this load 1711 // with the value stored truncated to a byte. If no truncation is 1712 // needed, the replacement is done in LoadNode::Identity(). 1713 // 1714 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1715 Node* mem = in(MemNode::Memory); 1716 Node* value = can_see_stored_value(mem, phase); 1717 if (value && !phase->type(value)->higher_equal(_type)) 1718 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); 1719 // Identity call will handle the case where truncation is not needed. 1720 return LoadNode::Ideal(phase, can_reshape); 1721 } 1722 1723 //--------------------------LoadUSNode::Ideal------------------------------------- 1724 // 1725 // If the previous store is to the same address as this load, 1726 // and the value stored was larger than a char, replace this load 1727 // with the value stored truncated to a char. If no truncation is 1728 // needed, the replacement is done in LoadNode::Identity(). 1729 // 1730 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1731 Node* mem = in(MemNode::Memory); 1732 Node* value = can_see_stored_value(mem,phase); 1733 if( value && !phase->type(value)->higher_equal( _type ) ) 1734 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1735 // Identity call will handle the case where truncation is not needed. 1736 return LoadNode::Ideal(phase, can_reshape); 1737 } 1738 1739 //--------------------------LoadSNode::Ideal-------------------------------------- 1740 // 1741 // If the previous store is to the same address as this load, 1742 // and the value stored was larger than a short, replace this load 1743 // with the value stored truncated to a short. If no truncation is 1744 // needed, the replacement is done in LoadNode::Identity(). 1745 // 1746 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1747 Node* mem = in(MemNode::Memory); 1748 Node* value = can_see_stored_value(mem,phase); 1749 if( value && !phase->type(value)->higher_equal( _type ) ) { 1750 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1751 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1752 } 1753 // Identity call will handle the case where truncation is not needed. 1754 return LoadNode::Ideal(phase, can_reshape); 1755 } 1756 1757 //============================================================================= 1758 //----------------------------LoadKlassNode::make------------------------------ 1759 // Polymorphic factory method: 1760 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1761 Compile* C = gvn.C; 1762 Node *ctl = NULL; 1763 // sanity check the alias category against the created node type 1764 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1765 assert(adr_type != NULL, "expecting TypeOopPtr"); 1766 #ifdef _LP64 1767 if (adr_type->is_ptr_to_narrowoop()) { 1768 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1769 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1770 } 1771 #endif 1772 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1773 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1774 } 1775 1776 //------------------------------Value------------------------------------------ 1777 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1778 return klass_value_common(phase); 1779 } 1780 1781 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1782 // Either input is TOP ==> the result is TOP 1783 const Type *t1 = phase->type( in(MemNode::Memory) ); 1784 if (t1 == Type::TOP) return Type::TOP; 1785 Node *adr = in(MemNode::Address); 1786 const Type *t2 = phase->type( adr ); 1787 if (t2 == Type::TOP) return Type::TOP; 1788 const TypePtr *tp = t2->is_ptr(); 1789 if (TypePtr::above_centerline(tp->ptr()) || 1790 tp->ptr() == TypePtr::Null) return Type::TOP; 1791 1792 // Return a more precise klass, if possible 1793 const TypeInstPtr *tinst = tp->isa_instptr(); 1794 if (tinst != NULL) { 1795 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1796 int offset = tinst->offset(); 1797 if (ik == phase->C->env()->Class_klass() 1798 && (offset == java_lang_Class::klass_offset_in_bytes() || 1799 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1800 // We are loading a special hidden field from a Class mirror object, 1801 // the field which points to the VM's Klass metaobject. 1802 ciType* t = tinst->java_mirror_type(); 1803 // java_mirror_type returns non-null for compile-time Class constants. 1804 if (t != NULL) { 1805 // constant oop => constant klass 1806 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1807 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1808 } 1809 if (!t->is_klass()) { 1810 // a primitive Class (e.g., int.class) has NULL for a klass field 1811 return TypePtr::NULL_PTR; 1812 } 1813 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1814 return TypeKlassPtr::make(t->as_klass()); 1815 } 1816 // non-constant mirror, so we can't tell what's going on 1817 } 1818 if( !ik->is_loaded() ) 1819 return _type; // Bail out if not loaded 1820 if (offset == oopDesc::klass_offset_in_bytes()) { 1821 if (tinst->klass_is_exact()) { 1822 return TypeKlassPtr::make(ik); 1823 } 1824 // See if we can become precise: no subklasses and no interface 1825 // (Note: We need to support verified interfaces.) 1826 if (!ik->is_interface() && !ik->has_subklass()) { 1827 //assert(!UseExactTypes, "this code should be useless with exact types"); 1828 // Add a dependence; if any subclass added we need to recompile 1829 if (!ik->is_final()) { 1830 // %%% should use stronger assert_unique_concrete_subtype instead 1831 phase->C->dependencies()->assert_leaf_type(ik); 1832 } 1833 // Return precise klass 1834 return TypeKlassPtr::make(ik); 1835 } 1836 1837 // Return root of possible klass 1838 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1839 } 1840 } 1841 1842 // Check for loading klass from an array 1843 const TypeAryPtr *tary = tp->isa_aryptr(); 1844 if( tary != NULL ) { 1845 ciKlass *tary_klass = tary->klass(); 1846 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1847 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1848 if (tary->klass_is_exact()) { 1849 return TypeKlassPtr::make(tary_klass); 1850 } 1851 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1852 // If the klass is an object array, we defer the question to the 1853 // array component klass. 1854 if( ak->is_obj_array_klass() ) { 1855 assert( ak->is_loaded(), "" ); 1856 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1857 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1858 ciInstanceKlass* ik = base_k->as_instance_klass(); 1859 // See if we can become precise: no subklasses and no interface 1860 if (!ik->is_interface() && !ik->has_subklass()) { 1861 //assert(!UseExactTypes, "this code should be useless with exact types"); 1862 // Add a dependence; if any subclass added we need to recompile 1863 if (!ik->is_final()) { 1864 phase->C->dependencies()->assert_leaf_type(ik); 1865 } 1866 // Return precise array klass 1867 return TypeKlassPtr::make(ak); 1868 } 1869 } 1870 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1871 } else { // Found a type-array? 1872 //assert(!UseExactTypes, "this code should be useless with exact types"); 1873 assert( ak->is_type_array_klass(), "" ); 1874 return TypeKlassPtr::make(ak); // These are always precise 1875 } 1876 } 1877 } 1878 1879 // Check for loading klass from an array klass 1880 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1881 if (tkls != NULL && !StressReflectiveCode) { 1882 ciKlass* klass = tkls->klass(); 1883 if( !klass->is_loaded() ) 1884 return _type; // Bail out if not loaded 1885 if( klass->is_obj_array_klass() && 1886 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1887 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1888 // // Always returning precise element type is incorrect, 1889 // // e.g., element type could be object and array may contain strings 1890 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1891 1892 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1893 // according to the element type's subclassing. 1894 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1895 } 1896 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1897 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1898 ciKlass* sup = klass->as_instance_klass()->super(); 1899 // The field is Klass::_super. Return its (constant) value. 1900 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1901 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1902 } 1903 } 1904 1905 // Bailout case 1906 return LoadNode::Value(phase); 1907 } 1908 1909 //------------------------------Identity--------------------------------------- 1910 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1911 // Also feed through the klass in Allocate(...klass...)._klass. 1912 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1913 return klass_identity_common(phase); 1914 } 1915 1916 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 1917 Node* x = LoadNode::Identity(phase); 1918 if (x != this) return x; 1919 1920 // Take apart the address into an oop and and offset. 1921 // Return 'this' if we cannot. 1922 Node* adr = in(MemNode::Address); 1923 intptr_t offset = 0; 1924 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1925 if (base == NULL) return this; 1926 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1927 if (toop == NULL) return this; 1928 1929 // We can fetch the klass directly through an AllocateNode. 1930 // This works even if the klass is not constant (clone or newArray). 1931 if (offset == oopDesc::klass_offset_in_bytes()) { 1932 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1933 if (allocated_klass != NULL) { 1934 return allocated_klass; 1935 } 1936 } 1937 1938 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1939 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1940 // See inline_native_Class_query for occurrences of these patterns. 1941 // Java Example: x.getClass().isAssignableFrom(y) 1942 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1943 // 1944 // This improves reflective code, often making the Class 1945 // mirror go completely dead. (Current exception: Class 1946 // mirrors may appear in debug info, but we could clean them out by 1947 // introducing a new debug info operator for klassOop.java_mirror). 1948 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1949 && (offset == java_lang_Class::klass_offset_in_bytes() || 1950 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1951 // We are loading a special hidden field from a Class mirror, 1952 // the field which points to its Klass or arrayKlass metaobject. 1953 if (base->is_Load()) { 1954 Node* adr2 = base->in(MemNode::Address); 1955 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1956 if (tkls != NULL && !tkls->empty() 1957 && (tkls->klass()->is_instance_klass() || 1958 tkls->klass()->is_array_klass()) 1959 && adr2->is_AddP() 1960 ) { 1961 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1962 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1963 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1964 } 1965 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1966 return adr2->in(AddPNode::Base); 1967 } 1968 } 1969 } 1970 } 1971 1972 return this; 1973 } 1974 1975 1976 //------------------------------Value------------------------------------------ 1977 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 1978 const Type *t = klass_value_common(phase); 1979 if (t == Type::TOP) 1980 return t; 1981 1982 return t->make_narrowoop(); 1983 } 1984 1985 //------------------------------Identity--------------------------------------- 1986 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 1987 // Also feed through the klass in Allocate(...klass...)._klass. 1988 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 1989 Node *x = klass_identity_common(phase); 1990 1991 const Type *t = phase->type( x ); 1992 if( t == Type::TOP ) return x; 1993 if( t->isa_narrowoop()) return x; 1994 1995 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 1996 } 1997 1998 //------------------------------Value----------------------------------------- 1999 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2000 // Either input is TOP ==> the result is TOP 2001 const Type *t1 = phase->type( in(MemNode::Memory) ); 2002 if( t1 == Type::TOP ) return Type::TOP; 2003 Node *adr = in(MemNode::Address); 2004 const Type *t2 = phase->type( adr ); 2005 if( t2 == Type::TOP ) return Type::TOP; 2006 const TypePtr *tp = t2->is_ptr(); 2007 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2008 const TypeAryPtr *tap = tp->isa_aryptr(); 2009 if( !tap ) return _type; 2010 return tap->size(); 2011 } 2012 2013 //-------------------------------Ideal--------------------------------------- 2014 // Feed through the length in AllocateArray(...length...)._length. 2015 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2016 Node* p = MemNode::Ideal_common(phase, can_reshape); 2017 if (p) return (p == NodeSentinel) ? NULL : p; 2018 2019 // Take apart the address into an oop and and offset. 2020 // Return 'this' if we cannot. 2021 Node* adr = in(MemNode::Address); 2022 intptr_t offset = 0; 2023 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2024 if (base == NULL) return NULL; 2025 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2026 if (tary == NULL) return NULL; 2027 2028 // We can fetch the length directly through an AllocateArrayNode. 2029 // This works even if the length is not constant (clone or newArray). 2030 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2031 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2032 if (alloc != NULL) { 2033 Node* allocated_length = alloc->Ideal_length(); 2034 Node* len = alloc->make_ideal_length(tary, phase); 2035 if (allocated_length != len) { 2036 // New CastII improves on this. 2037 return len; 2038 } 2039 } 2040 } 2041 2042 return NULL; 2043 } 2044 2045 //------------------------------Identity--------------------------------------- 2046 // Feed through the length in AllocateArray(...length...)._length. 2047 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2048 Node* x = LoadINode::Identity(phase); 2049 if (x != this) return x; 2050 2051 // Take apart the address into an oop and and offset. 2052 // Return 'this' if we cannot. 2053 Node* adr = in(MemNode::Address); 2054 intptr_t offset = 0; 2055 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2056 if (base == NULL) return this; 2057 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2058 if (tary == NULL) return this; 2059 2060 // We can fetch the length directly through an AllocateArrayNode. 2061 // This works even if the length is not constant (clone or newArray). 2062 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2063 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2064 if (alloc != NULL) { 2065 Node* allocated_length = alloc->Ideal_length(); 2066 // Do not allow make_ideal_length to allocate a CastII node. 2067 Node* len = alloc->make_ideal_length(tary, phase, false); 2068 if (allocated_length == len) { 2069 // Return allocated_length only if it would not be improved by a CastII. 2070 return allocated_length; 2071 } 2072 } 2073 } 2074 2075 return this; 2076 2077 } 2078 2079 //============================================================================= 2080 //---------------------------StoreNode::make----------------------------------- 2081 // Polymorphic factory method: 2082 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 2083 Compile* C = gvn.C; 2084 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2085 ctl != NULL, "raw memory operations should have control edge"); 2086 2087 switch (bt) { 2088 case T_BOOLEAN: 2089 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 2090 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 2091 case T_CHAR: 2092 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 2093 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 2094 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 2095 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 2096 case T_ADDRESS: 2097 case T_OBJECT: 2098 #ifdef _LP64 2099 if (adr->bottom_type()->is_ptr_to_narrowoop() || 2100 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 2101 adr->bottom_type()->isa_rawptr())) { 2102 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2103 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 2104 } else 2105 #endif 2106 { 2107 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 2108 } 2109 } 2110 ShouldNotReachHere(); 2111 return (StoreNode*)NULL; 2112 } 2113 2114 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2115 bool require_atomic = true; 2116 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2117 } 2118 2119 2120 //--------------------------bottom_type---------------------------------------- 2121 const Type *StoreNode::bottom_type() const { 2122 return Type::MEMORY; 2123 } 2124 2125 //------------------------------hash------------------------------------------- 2126 uint StoreNode::hash() const { 2127 // unroll addition of interesting fields 2128 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2129 2130 // Since they are not commoned, do not hash them: 2131 return NO_HASH; 2132 } 2133 2134 //------------------------------Ideal------------------------------------------ 2135 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2136 // When a store immediately follows a relevant allocation/initialization, 2137 // try to capture it into the initialization, or hoist it above. 2138 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2139 Node* p = MemNode::Ideal_common(phase, can_reshape); 2140 if (p) return (p == NodeSentinel) ? NULL : p; 2141 2142 Node* mem = in(MemNode::Memory); 2143 Node* address = in(MemNode::Address); 2144 2145 // Back-to-back stores to same address? Fold em up. 2146 // Generally unsafe if I have intervening uses... 2147 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 2148 // Looking at a dead closed cycle of memory? 2149 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2150 2151 assert(Opcode() == mem->Opcode() || 2152 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2153 "no mismatched stores, except on raw memory"); 2154 2155 if (mem->outcnt() == 1 && // check for intervening uses 2156 mem->as_Store()->memory_size() <= this->memory_size()) { 2157 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2158 // For example, 'mem' might be the final state at a conditional return. 2159 // Or, 'mem' might be used by some node which is live at the same time 2160 // 'this' is live, which might be unschedulable. So, require exactly 2161 // ONE user, the 'this' store, until such time as we clone 'mem' for 2162 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2163 if (can_reshape) { // (%%% is this an anachronism?) 2164 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2165 phase->is_IterGVN()); 2166 } else { 2167 // It's OK to do this in the parser, since DU info is always accurate, 2168 // and the parser always refers to nodes via SafePointNode maps. 2169 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2170 } 2171 return this; 2172 } 2173 } 2174 2175 // Capture an unaliased, unconditional, simple store into an initializer. 2176 // Or, if it is independent of the allocation, hoist it above the allocation. 2177 if (ReduceFieldZeroing && /*can_reshape &&*/ 2178 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2179 InitializeNode* init = mem->in(0)->as_Initialize(); 2180 intptr_t offset = init->can_capture_store(this, phase); 2181 if (offset > 0) { 2182 Node* moved = init->capture_store(this, offset, phase); 2183 // If the InitializeNode captured me, it made a raw copy of me, 2184 // and I need to disappear. 2185 if (moved != NULL) { 2186 // %%% hack to ensure that Ideal returns a new node: 2187 mem = MergeMemNode::make(phase->C, mem); 2188 return mem; // fold me away 2189 } 2190 } 2191 } 2192 2193 return NULL; // No further progress 2194 } 2195 2196 //------------------------------Value----------------------------------------- 2197 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2198 // Either input is TOP ==> the result is TOP 2199 const Type *t1 = phase->type( in(MemNode::Memory) ); 2200 if( t1 == Type::TOP ) return Type::TOP; 2201 const Type *t2 = phase->type( in(MemNode::Address) ); 2202 if( t2 == Type::TOP ) return Type::TOP; 2203 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2204 if( t3 == Type::TOP ) return Type::TOP; 2205 return Type::MEMORY; 2206 } 2207 2208 //------------------------------Identity--------------------------------------- 2209 // Remove redundant stores: 2210 // Store(m, p, Load(m, p)) changes to m. 2211 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2212 Node *StoreNode::Identity( PhaseTransform *phase ) { 2213 Node* mem = in(MemNode::Memory); 2214 Node* adr = in(MemNode::Address); 2215 Node* val = in(MemNode::ValueIn); 2216 2217 // Load then Store? Then the Store is useless 2218 if (val->is_Load() && 2219 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 2220 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 2221 val->as_Load()->store_Opcode() == Opcode()) { 2222 return mem; 2223 } 2224 2225 // Two stores in a row of the same value? 2226 if (mem->is_Store() && 2227 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 2228 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 2229 mem->Opcode() == Opcode()) { 2230 return mem; 2231 } 2232 2233 // Store of zero anywhere into a freshly-allocated object? 2234 // Then the store is useless. 2235 // (It must already have been captured by the InitializeNode.) 2236 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2237 // a newly allocated object is already all-zeroes everywhere 2238 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2239 return mem; 2240 } 2241 2242 // the store may also apply to zero-bits in an earlier object 2243 Node* prev_mem = find_previous_store(phase); 2244 // Steps (a), (b): Walk past independent stores to find an exact match. 2245 if (prev_mem != NULL) { 2246 Node* prev_val = can_see_stored_value(prev_mem, phase); 2247 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2248 // prev_val and val might differ by a cast; it would be good 2249 // to keep the more informative of the two. 2250 return mem; 2251 } 2252 } 2253 } 2254 2255 return this; 2256 } 2257 2258 //------------------------------match_edge------------------------------------- 2259 // Do we Match on this edge index or not? Match only memory & value 2260 uint StoreNode::match_edge(uint idx) const { 2261 return idx == MemNode::Address || idx == MemNode::ValueIn; 2262 } 2263 2264 //------------------------------cmp-------------------------------------------- 2265 // Do not common stores up together. They generally have to be split 2266 // back up anyways, so do not bother. 2267 uint StoreNode::cmp( const Node &n ) const { 2268 return (&n == this); // Always fail except on self 2269 } 2270 2271 //------------------------------Ideal_masked_input----------------------------- 2272 // Check for a useless mask before a partial-word store 2273 // (StoreB ... (AndI valIn conIa) ) 2274 // If (conIa & mask == mask) this simplifies to 2275 // (StoreB ... (valIn) ) 2276 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2277 Node *val = in(MemNode::ValueIn); 2278 if( val->Opcode() == Op_AndI ) { 2279 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2280 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2281 set_req(MemNode::ValueIn, val->in(1)); 2282 return this; 2283 } 2284 } 2285 return NULL; 2286 } 2287 2288 2289 //------------------------------Ideal_sign_extended_input---------------------- 2290 // Check for useless sign-extension before a partial-word store 2291 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2292 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2293 // (StoreB ... (valIn) ) 2294 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2295 Node *val = in(MemNode::ValueIn); 2296 if( val->Opcode() == Op_RShiftI ) { 2297 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2298 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2299 Node *shl = val->in(1); 2300 if( shl->Opcode() == Op_LShiftI ) { 2301 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2302 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2303 set_req(MemNode::ValueIn, shl->in(1)); 2304 return this; 2305 } 2306 } 2307 } 2308 } 2309 return NULL; 2310 } 2311 2312 //------------------------------value_never_loaded----------------------------------- 2313 // Determine whether there are any possible loads of the value stored. 2314 // For simplicity, we actually check if there are any loads from the 2315 // address stored to, not just for loads of the value stored by this node. 2316 // 2317 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2318 Node *adr = in(Address); 2319 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2320 if (adr_oop == NULL) 2321 return false; 2322 if (!adr_oop->is_known_instance_field()) 2323 return false; // if not a distinct instance, there may be aliases of the address 2324 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2325 Node *use = adr->fast_out(i); 2326 int opc = use->Opcode(); 2327 if (use->is_Load() || use->is_LoadStore()) { 2328 return false; 2329 } 2330 } 2331 return true; 2332 } 2333 2334 //============================================================================= 2335 //------------------------------Ideal------------------------------------------ 2336 // If the store is from an AND mask that leaves the low bits untouched, then 2337 // we can skip the AND operation. If the store is from a sign-extension 2338 // (a left shift, then right shift) we can skip both. 2339 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2340 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2341 if( progress != NULL ) return progress; 2342 2343 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2344 if( progress != NULL ) return progress; 2345 2346 // Finally check the default case 2347 return StoreNode::Ideal(phase, can_reshape); 2348 } 2349 2350 //============================================================================= 2351 //------------------------------Ideal------------------------------------------ 2352 // If the store is from an AND mask that leaves the low bits untouched, then 2353 // we can skip the AND operation 2354 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2355 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2356 if( progress != NULL ) return progress; 2357 2358 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2359 if( progress != NULL ) return progress; 2360 2361 // Finally check the default case 2362 return StoreNode::Ideal(phase, can_reshape); 2363 } 2364 2365 //============================================================================= 2366 //------------------------------Identity--------------------------------------- 2367 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2368 // No need to card mark when storing a null ptr 2369 Node* my_store = in(MemNode::OopStore); 2370 if (my_store->is_Store()) { 2371 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2372 if( t1 == TypePtr::NULL_PTR ) { 2373 return in(MemNode::Memory); 2374 } 2375 } 2376 return this; 2377 } 2378 2379 //============================================================================= 2380 //------------------------------Ideal--------------------------------------- 2381 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2382 Node* progress = StoreNode::Ideal(phase, can_reshape); 2383 if (progress != NULL) return progress; 2384 2385 Node* my_store = in(MemNode::OopStore); 2386 if (my_store->is_MergeMem()) { 2387 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2388 set_req(MemNode::OopStore, mem); 2389 return this; 2390 } 2391 2392 return NULL; 2393 } 2394 2395 //------------------------------Value----------------------------------------- 2396 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2397 // Either input is TOP ==> the result is TOP 2398 const Type *t = phase->type( in(MemNode::Memory) ); 2399 if( t == Type::TOP ) return Type::TOP; 2400 t = phase->type( in(MemNode::Address) ); 2401 if( t == Type::TOP ) return Type::TOP; 2402 t = phase->type( in(MemNode::ValueIn) ); 2403 if( t == Type::TOP ) return Type::TOP; 2404 // If extra input is TOP ==> the result is TOP 2405 t = phase->type( in(MemNode::OopStore) ); 2406 if( t == Type::TOP ) return Type::TOP; 2407 2408 return StoreNode::Value( phase ); 2409 } 2410 2411 2412 //============================================================================= 2413 //----------------------------------SCMemProjNode------------------------------ 2414 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2415 { 2416 return bottom_type(); 2417 } 2418 2419 //============================================================================= 2420 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2421 init_req(MemNode::Control, c ); 2422 init_req(MemNode::Memory , mem); 2423 init_req(MemNode::Address, adr); 2424 init_req(MemNode::ValueIn, val); 2425 init_req( ExpectedIn, ex ); 2426 init_class_id(Class_LoadStore); 2427 2428 } 2429 2430 //============================================================================= 2431 //-------------------------------adr_type-------------------------------------- 2432 // Do we Match on this edge index or not? Do not match memory 2433 const TypePtr* ClearArrayNode::adr_type() const { 2434 Node *adr = in(3); 2435 return MemNode::calculate_adr_type(adr->bottom_type()); 2436 } 2437 2438 //------------------------------match_edge------------------------------------- 2439 // Do we Match on this edge index or not? Do not match memory 2440 uint ClearArrayNode::match_edge(uint idx) const { 2441 return idx > 1; 2442 } 2443 2444 //------------------------------Identity--------------------------------------- 2445 // Clearing a zero length array does nothing 2446 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2447 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2448 } 2449 2450 //------------------------------Idealize--------------------------------------- 2451 // Clearing a short array is faster with stores 2452 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2453 const int unit = BytesPerLong; 2454 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2455 if (!t) return NULL; 2456 if (!t->is_con()) return NULL; 2457 intptr_t raw_count = t->get_con(); 2458 intptr_t size = raw_count; 2459 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2460 // Clearing nothing uses the Identity call. 2461 // Negative clears are possible on dead ClearArrays 2462 // (see jck test stmt114.stmt11402.val). 2463 if (size <= 0 || size % unit != 0) return NULL; 2464 intptr_t count = size / unit; 2465 // Length too long; use fast hardware clear 2466 if (size > Matcher::init_array_short_size) return NULL; 2467 Node *mem = in(1); 2468 if( phase->type(mem)==Type::TOP ) return NULL; 2469 Node *adr = in(3); 2470 const Type* at = phase->type(adr); 2471 if( at==Type::TOP ) return NULL; 2472 const TypePtr* atp = at->isa_ptr(); 2473 // adjust atp to be the correct array element address type 2474 if (atp == NULL) atp = TypePtr::BOTTOM; 2475 else atp = atp->add_offset(Type::OffsetBot); 2476 // Get base for derived pointer purposes 2477 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2478 Node *base = adr->in(1); 2479 2480 Node *zero = phase->makecon(TypeLong::ZERO); 2481 Node *off = phase->MakeConX(BytesPerLong); 2482 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2483 count--; 2484 while( count-- ) { 2485 mem = phase->transform(mem); 2486 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2487 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2488 } 2489 return mem; 2490 } 2491 2492 //----------------------------step_through---------------------------------- 2493 // Return allocation input memory edge if it is different instance 2494 // or itself if it is the one we are looking for. 2495 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2496 Node* n = *np; 2497 assert(n->is_ClearArray(), "sanity"); 2498 intptr_t offset; 2499 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2500 // This method is called only before Allocate nodes are expanded during 2501 // macro nodes expansion. Before that ClearArray nodes are only generated 2502 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2503 assert(alloc != NULL, "should have allocation"); 2504 if (alloc->_idx == instance_id) { 2505 // Can not bypass initialization of the instance we are looking for. 2506 return false; 2507 } 2508 // Otherwise skip it. 2509 InitializeNode* init = alloc->initialization(); 2510 if (init != NULL) 2511 *np = init->in(TypeFunc::Memory); 2512 else 2513 *np = alloc->in(TypeFunc::Memory); 2514 return true; 2515 } 2516 2517 //----------------------------clear_memory------------------------------------- 2518 // Generate code to initialize object storage to zero. 2519 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2520 intptr_t start_offset, 2521 Node* end_offset, 2522 PhaseGVN* phase) { 2523 Compile* C = phase->C; 2524 intptr_t offset = start_offset; 2525 2526 int unit = BytesPerLong; 2527 if ((offset % unit) != 0) { 2528 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2529 adr = phase->transform(adr); 2530 const TypePtr* atp = TypeRawPtr::BOTTOM; 2531 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2532 mem = phase->transform(mem); 2533 offset += BytesPerInt; 2534 } 2535 assert((offset % unit) == 0, ""); 2536 2537 // Initialize the remaining stuff, if any, with a ClearArray. 2538 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2539 } 2540 2541 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2542 Node* start_offset, 2543 Node* end_offset, 2544 PhaseGVN* phase) { 2545 if (start_offset == end_offset) { 2546 // nothing to do 2547 return mem; 2548 } 2549 2550 Compile* C = phase->C; 2551 int unit = BytesPerLong; 2552 Node* zbase = start_offset; 2553 Node* zend = end_offset; 2554 2555 // Scale to the unit required by the CPU: 2556 if (!Matcher::init_array_count_is_in_bytes) { 2557 Node* shift = phase->intcon(exact_log2(unit)); 2558 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2559 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2560 } 2561 2562 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2563 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2564 2565 // Bulk clear double-words 2566 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2567 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2568 return phase->transform(mem); 2569 } 2570 2571 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2572 intptr_t start_offset, 2573 intptr_t end_offset, 2574 PhaseGVN* phase) { 2575 if (start_offset == end_offset) { 2576 // nothing to do 2577 return mem; 2578 } 2579 2580 Compile* C = phase->C; 2581 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2582 intptr_t done_offset = end_offset; 2583 if ((done_offset % BytesPerLong) != 0) { 2584 done_offset -= BytesPerInt; 2585 } 2586 if (done_offset > start_offset) { 2587 mem = clear_memory(ctl, mem, dest, 2588 start_offset, phase->MakeConX(done_offset), phase); 2589 } 2590 if (done_offset < end_offset) { // emit the final 32-bit store 2591 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2592 adr = phase->transform(adr); 2593 const TypePtr* atp = TypeRawPtr::BOTTOM; 2594 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2595 mem = phase->transform(mem); 2596 done_offset += BytesPerInt; 2597 } 2598 assert(done_offset == end_offset, ""); 2599 return mem; 2600 } 2601 2602 //============================================================================= 2603 // Do we match on this edge? No memory edges 2604 uint StrCompNode::match_edge(uint idx) const { 2605 return idx == 2 || idx == 3; // StrComp (Binary str1 cnt1) (Binary str2 cnt2) 2606 } 2607 2608 //------------------------------Ideal------------------------------------------ 2609 // Return a node which is more "ideal" than the current node. Strip out 2610 // control copies 2611 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2612 return remove_dead_region(phase, can_reshape) ? this : NULL; 2613 } 2614 2615 //============================================================================= 2616 // Do we match on this edge? No memory edges 2617 uint StrEqualsNode::match_edge(uint idx) const { 2618 return idx == 2 || idx == 3; // StrEquals (Binary str1 str2) cnt 2619 } 2620 2621 //------------------------------Ideal------------------------------------------ 2622 // Return a node which is more "ideal" than the current node. Strip out 2623 // control copies 2624 Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2625 return remove_dead_region(phase, can_reshape) ? this : NULL; 2626 } 2627 2628 //============================================================================= 2629 // Do we match on this edge? No memory edges 2630 uint StrIndexOfNode::match_edge(uint idx) const { 2631 return idx == 2 || idx == 3; // StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2) 2632 } 2633 2634 //------------------------------Ideal------------------------------------------ 2635 // Return a node which is more "ideal" than the current node. Strip out 2636 // control copies 2637 Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2638 return remove_dead_region(phase, can_reshape) ? this : NULL; 2639 } 2640 2641 //============================================================================= 2642 // Do we match on this edge? No memory edges 2643 uint AryEqNode::match_edge(uint idx) const { 2644 return idx == 2 || idx == 3; // StrEquals ary1 ary2 2645 } 2646 //------------------------------Ideal------------------------------------------ 2647 // Return a node which is more "ideal" than the current node. Strip out 2648 // control copies 2649 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2650 return remove_dead_region(phase, can_reshape) ? this : NULL; 2651 } 2652 2653 //============================================================================= 2654 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2655 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2656 _adr_type(C->get_adr_type(alias_idx)) 2657 { 2658 init_class_id(Class_MemBar); 2659 Node* top = C->top(); 2660 init_req(TypeFunc::I_O,top); 2661 init_req(TypeFunc::FramePtr,top); 2662 init_req(TypeFunc::ReturnAdr,top); 2663 if (precedent != NULL) 2664 init_req(TypeFunc::Parms, precedent); 2665 } 2666 2667 //------------------------------cmp-------------------------------------------- 2668 uint MemBarNode::hash() const { return NO_HASH; } 2669 uint MemBarNode::cmp( const Node &n ) const { 2670 return (&n == this); // Always fail except on self 2671 } 2672 2673 //------------------------------make------------------------------------------- 2674 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2675 int len = Precedent + (pn == NULL? 0: 1); 2676 switch (opcode) { 2677 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2678 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2679 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2680 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2681 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2682 default: ShouldNotReachHere(); return NULL; 2683 } 2684 } 2685 2686 //------------------------------Ideal------------------------------------------ 2687 // Return a node which is more "ideal" than the current node. Strip out 2688 // control copies 2689 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2690 if (remove_dead_region(phase, can_reshape)) return this; 2691 2692 // Eliminate volatile MemBars for scalar replaced objects. 2693 if (can_reshape && req() == (Precedent+1) && 2694 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { 2695 // Volatile field loads and stores. 2696 Node* my_mem = in(MemBarNode::Precedent); 2697 if (my_mem != NULL && my_mem->is_Mem()) { 2698 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2699 // Check for scalar replaced object reference. 2700 if( t_oop != NULL && t_oop->is_known_instance_field() && 2701 t_oop->offset() != Type::OffsetBot && 2702 t_oop->offset() != Type::OffsetTop) { 2703 // Replace MemBar projections by its inputs. 2704 PhaseIterGVN* igvn = phase->is_IterGVN(); 2705 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2706 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2707 // Must return either the original node (now dead) or a new node 2708 // (Do not return a top here, since that would break the uniqueness of top.) 2709 return new (phase->C, 1) ConINode(TypeInt::ZERO); 2710 } 2711 } 2712 } 2713 return NULL; 2714 } 2715 2716 //------------------------------Value------------------------------------------ 2717 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2718 if( !in(0) ) return Type::TOP; 2719 if( phase->type(in(0)) == Type::TOP ) 2720 return Type::TOP; 2721 return TypeTuple::MEMBAR; 2722 } 2723 2724 //------------------------------match------------------------------------------ 2725 // Construct projections for memory. 2726 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2727 switch (proj->_con) { 2728 case TypeFunc::Control: 2729 case TypeFunc::Memory: 2730 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2731 } 2732 ShouldNotReachHere(); 2733 return NULL; 2734 } 2735 2736 //===========================InitializeNode==================================== 2737 // SUMMARY: 2738 // This node acts as a memory barrier on raw memory, after some raw stores. 2739 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2740 // The Initialize can 'capture' suitably constrained stores as raw inits. 2741 // It can coalesce related raw stores into larger units (called 'tiles'). 2742 // It can avoid zeroing new storage for memory units which have raw inits. 2743 // At macro-expansion, it is marked 'complete', and does not optimize further. 2744 // 2745 // EXAMPLE: 2746 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2747 // ctl = incoming control; mem* = incoming memory 2748 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2749 // First allocate uninitialized memory and fill in the header: 2750 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2751 // ctl := alloc.Control; mem* := alloc.Memory* 2752 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2753 // Then initialize to zero the non-header parts of the raw memory block: 2754 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2755 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2756 // After the initialize node executes, the object is ready for service: 2757 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2758 // Suppose its body is immediately initialized as {1,2}: 2759 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2760 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2761 // mem.SLICE(#short[*]) := store2 2762 // 2763 // DETAILS: 2764 // An InitializeNode collects and isolates object initialization after 2765 // an AllocateNode and before the next possible safepoint. As a 2766 // memory barrier (MemBarNode), it keeps critical stores from drifting 2767 // down past any safepoint or any publication of the allocation. 2768 // Before this barrier, a newly-allocated object may have uninitialized bits. 2769 // After this barrier, it may be treated as a real oop, and GC is allowed. 2770 // 2771 // The semantics of the InitializeNode include an implicit zeroing of 2772 // the new object from object header to the end of the object. 2773 // (The object header and end are determined by the AllocateNode.) 2774 // 2775 // Certain stores may be added as direct inputs to the InitializeNode. 2776 // These stores must update raw memory, and they must be to addresses 2777 // derived from the raw address produced by AllocateNode, and with 2778 // a constant offset. They must be ordered by increasing offset. 2779 // The first one is at in(RawStores), the last at in(req()-1). 2780 // Unlike most memory operations, they are not linked in a chain, 2781 // but are displayed in parallel as users of the rawmem output of 2782 // the allocation. 2783 // 2784 // (See comments in InitializeNode::capture_store, which continue 2785 // the example given above.) 2786 // 2787 // When the associated Allocate is macro-expanded, the InitializeNode 2788 // may be rewritten to optimize collected stores. A ClearArrayNode 2789 // may also be created at that point to represent any required zeroing. 2790 // The InitializeNode is then marked 'complete', prohibiting further 2791 // capturing of nearby memory operations. 2792 // 2793 // During macro-expansion, all captured initializations which store 2794 // constant values of 32 bits or smaller are coalesced (if advantageous) 2795 // into larger 'tiles' 32 or 64 bits. This allows an object to be 2796 // initialized in fewer memory operations. Memory words which are 2797 // covered by neither tiles nor non-constant stores are pre-zeroed 2798 // by explicit stores of zero. (The code shape happens to do all 2799 // zeroing first, then all other stores, with both sequences occurring 2800 // in order of ascending offsets.) 2801 // 2802 // Alternatively, code may be inserted between an AllocateNode and its 2803 // InitializeNode, to perform arbitrary initialization of the new object. 2804 // E.g., the object copying intrinsics insert complex data transfers here. 2805 // The initialization must then be marked as 'complete' disable the 2806 // built-in zeroing semantics and the collection of initializing stores. 2807 // 2808 // While an InitializeNode is incomplete, reads from the memory state 2809 // produced by it are optimizable if they match the control edge and 2810 // new oop address associated with the allocation/initialization. 2811 // They return a stored value (if the offset matches) or else zero. 2812 // A write to the memory state, if it matches control and address, 2813 // and if it is to a constant offset, may be 'captured' by the 2814 // InitializeNode. It is cloned as a raw memory operation and rewired 2815 // inside the initialization, to the raw oop produced by the allocation. 2816 // Operations on addresses which are provably distinct (e.g., to 2817 // other AllocateNodes) are allowed to bypass the initialization. 2818 // 2819 // The effect of all this is to consolidate object initialization 2820 // (both arrays and non-arrays, both piecewise and bulk) into a 2821 // single location, where it can be optimized as a unit. 2822 // 2823 // Only stores with an offset less than TrackedInitializationLimit words 2824 // will be considered for capture by an InitializeNode. This puts a 2825 // reasonable limit on the complexity of optimized initializations. 2826 2827 //---------------------------InitializeNode------------------------------------ 2828 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2829 : _is_complete(false), 2830 MemBarNode(C, adr_type, rawoop) 2831 { 2832 init_class_id(Class_Initialize); 2833 2834 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2835 assert(in(RawAddress) == rawoop, "proper init"); 2836 // Note: allocation() can be NULL, for secondary initialization barriers 2837 } 2838 2839 // Since this node is not matched, it will be processed by the 2840 // register allocator. Declare that there are no constraints 2841 // on the allocation of the RawAddress edge. 2842 const RegMask &InitializeNode::in_RegMask(uint idx) const { 2843 // This edge should be set to top, by the set_complete. But be conservative. 2844 if (idx == InitializeNode::RawAddress) 2845 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2846 return RegMask::Empty; 2847 } 2848 2849 Node* InitializeNode::memory(uint alias_idx) { 2850 Node* mem = in(Memory); 2851 if (mem->is_MergeMem()) { 2852 return mem->as_MergeMem()->memory_at(alias_idx); 2853 } else { 2854 // incoming raw memory is not split 2855 return mem; 2856 } 2857 } 2858 2859 bool InitializeNode::is_non_zero() { 2860 if (is_complete()) return false; 2861 remove_extra_zeroes(); 2862 return (req() > RawStores); 2863 } 2864 2865 void InitializeNode::set_complete(PhaseGVN* phase) { 2866 assert(!is_complete(), "caller responsibility"); 2867 _is_complete = true; 2868 2869 // After this node is complete, it contains a bunch of 2870 // raw-memory initializations. There is no need for 2871 // it to have anything to do with non-raw memory effects. 2872 // Therefore, tell all non-raw users to re-optimize themselves, 2873 // after skipping the memory effects of this initialization. 2874 PhaseIterGVN* igvn = phase->is_IterGVN(); 2875 if (igvn) igvn->add_users_to_worklist(this); 2876 } 2877 2878 // convenience function 2879 // return false if the init contains any stores already 2880 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2881 InitializeNode* init = initialization(); 2882 if (init == NULL || init->is_complete()) return false; 2883 init->remove_extra_zeroes(); 2884 // for now, if this allocation has already collected any inits, bail: 2885 if (init->is_non_zero()) return false; 2886 init->set_complete(phase); 2887 return true; 2888 } 2889 2890 void InitializeNode::remove_extra_zeroes() { 2891 if (req() == RawStores) return; 2892 Node* zmem = zero_memory(); 2893 uint fill = RawStores; 2894 for (uint i = fill; i < req(); i++) { 2895 Node* n = in(i); 2896 if (n->is_top() || n == zmem) continue; // skip 2897 if (fill < i) set_req(fill, n); // compact 2898 ++fill; 2899 } 2900 // delete any empty spaces created: 2901 while (fill < req()) { 2902 del_req(fill); 2903 } 2904 } 2905 2906 // Helper for remembering which stores go with which offsets. 2907 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2908 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2909 intptr_t offset = -1; 2910 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2911 phase, offset); 2912 if (base == NULL) return -1; // something is dead, 2913 if (offset < 0) return -1; // dead, dead 2914 return offset; 2915 } 2916 2917 // Helper for proving that an initialization expression is 2918 // "simple enough" to be folded into an object initialization. 2919 // Attempts to prove that a store's initial value 'n' can be captured 2920 // within the initialization without creating a vicious cycle, such as: 2921 // { Foo p = new Foo(); p.next = p; } 2922 // True for constants and parameters and small combinations thereof. 2923 bool InitializeNode::detect_init_independence(Node* n, 2924 bool st_is_pinned, 2925 int& count) { 2926 if (n == NULL) return true; // (can this really happen?) 2927 if (n->is_Proj()) n = n->in(0); 2928 if (n == this) return false; // found a cycle 2929 if (n->is_Con()) return true; 2930 if (n->is_Start()) return true; // params, etc., are OK 2931 if (n->is_Root()) return true; // even better 2932 2933 Node* ctl = n->in(0); 2934 if (ctl != NULL && !ctl->is_top()) { 2935 if (ctl->is_Proj()) ctl = ctl->in(0); 2936 if (ctl == this) return false; 2937 2938 // If we already know that the enclosing memory op is pinned right after 2939 // the init, then any control flow that the store has picked up 2940 // must have preceded the init, or else be equal to the init. 2941 // Even after loop optimizations (which might change control edges) 2942 // a store is never pinned *before* the availability of its inputs. 2943 if (!MemNode::all_controls_dominate(n, this)) 2944 return false; // failed to prove a good control 2945 2946 } 2947 2948 // Check data edges for possible dependencies on 'this'. 2949 if ((count += 1) > 20) return false; // complexity limit 2950 for (uint i = 1; i < n->req(); i++) { 2951 Node* m = n->in(i); 2952 if (m == NULL || m == n || m->is_top()) continue; 2953 uint first_i = n->find_edge(m); 2954 if (i != first_i) continue; // process duplicate edge just once 2955 if (!detect_init_independence(m, st_is_pinned, count)) { 2956 return false; 2957 } 2958 } 2959 2960 return true; 2961 } 2962 2963 // Here are all the checks a Store must pass before it can be moved into 2964 // an initialization. Returns zero if a check fails. 2965 // On success, returns the (constant) offset to which the store applies, 2966 // within the initialized memory. 2967 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2968 const int FAIL = 0; 2969 if (st->req() != MemNode::ValueIn + 1) 2970 return FAIL; // an inscrutable StoreNode (card mark?) 2971 Node* ctl = st->in(MemNode::Control); 2972 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2973 return FAIL; // must be unconditional after the initialization 2974 Node* mem = st->in(MemNode::Memory); 2975 if (!(mem->is_Proj() && mem->in(0) == this)) 2976 return FAIL; // must not be preceded by other stores 2977 Node* adr = st->in(MemNode::Address); 2978 intptr_t offset; 2979 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2980 if (alloc == NULL) 2981 return FAIL; // inscrutable address 2982 if (alloc != allocation()) 2983 return FAIL; // wrong allocation! (store needs to float up) 2984 Node* val = st->in(MemNode::ValueIn); 2985 int complexity_count = 0; 2986 if (!detect_init_independence(val, true, complexity_count)) 2987 return FAIL; // stored value must be 'simple enough' 2988 2989 return offset; // success 2990 } 2991 2992 // Find the captured store in(i) which corresponds to the range 2993 // [start..start+size) in the initialized object. 2994 // If there is one, return its index i. If there isn't, return the 2995 // negative of the index where it should be inserted. 2996 // Return 0 if the queried range overlaps an initialization boundary 2997 // or if dead code is encountered. 2998 // If size_in_bytes is zero, do not bother with overlap checks. 2999 int InitializeNode::captured_store_insertion_point(intptr_t start, 3000 int size_in_bytes, 3001 PhaseTransform* phase) { 3002 const int FAIL = 0, MAX_STORE = BytesPerLong; 3003 3004 if (is_complete()) 3005 return FAIL; // arraycopy got here first; punt 3006 3007 assert(allocation() != NULL, "must be present"); 3008 3009 // no negatives, no header fields: 3010 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3011 3012 // after a certain size, we bail out on tracking all the stores: 3013 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3014 if (start >= ti_limit) return FAIL; 3015 3016 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3017 if (i >= limit) return -(int)i; // not found; here is where to put it 3018 3019 Node* st = in(i); 3020 intptr_t st_off = get_store_offset(st, phase); 3021 if (st_off < 0) { 3022 if (st != zero_memory()) { 3023 return FAIL; // bail out if there is dead garbage 3024 } 3025 } else if (st_off > start) { 3026 // ...we are done, since stores are ordered 3027 if (st_off < start + size_in_bytes) { 3028 return FAIL; // the next store overlaps 3029 } 3030 return -(int)i; // not found; here is where to put it 3031 } else if (st_off < start) { 3032 if (size_in_bytes != 0 && 3033 start < st_off + MAX_STORE && 3034 start < st_off + st->as_Store()->memory_size()) { 3035 return FAIL; // the previous store overlaps 3036 } 3037 } else { 3038 if (size_in_bytes != 0 && 3039 st->as_Store()->memory_size() != size_in_bytes) { 3040 return FAIL; // mismatched store size 3041 } 3042 return i; 3043 } 3044 3045 ++i; 3046 } 3047 } 3048 3049 // Look for a captured store which initializes at the offset 'start' 3050 // with the given size. If there is no such store, and no other 3051 // initialization interferes, then return zero_memory (the memory 3052 // projection of the AllocateNode). 3053 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3054 PhaseTransform* phase) { 3055 assert(stores_are_sane(phase), ""); 3056 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3057 if (i == 0) { 3058 return NULL; // something is dead 3059 } else if (i < 0) { 3060 return zero_memory(); // just primordial zero bits here 3061 } else { 3062 Node* st = in(i); // here is the store at this position 3063 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3064 return st; 3065 } 3066 } 3067 3068 // Create, as a raw pointer, an address within my new object at 'offset'. 3069 Node* InitializeNode::make_raw_address(intptr_t offset, 3070 PhaseTransform* phase) { 3071 Node* addr = in(RawAddress); 3072 if (offset != 0) { 3073 Compile* C = phase->C; 3074 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 3075 phase->MakeConX(offset)) ); 3076 } 3077 return addr; 3078 } 3079 3080 // Clone the given store, converting it into a raw store 3081 // initializing a field or element of my new object. 3082 // Caller is responsible for retiring the original store, 3083 // with subsume_node or the like. 3084 // 3085 // From the example above InitializeNode::InitializeNode, 3086 // here are the old stores to be captured: 3087 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3088 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3089 // 3090 // Here is the changed code; note the extra edges on init: 3091 // alloc = (Allocate ...) 3092 // rawoop = alloc.RawAddress 3093 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3094 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3095 // init = (Initialize alloc.Control alloc.Memory rawoop 3096 // rawstore1 rawstore2) 3097 // 3098 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3099 PhaseTransform* phase) { 3100 assert(stores_are_sane(phase), ""); 3101 3102 if (start < 0) return NULL; 3103 assert(can_capture_store(st, phase) == start, "sanity"); 3104 3105 Compile* C = phase->C; 3106 int size_in_bytes = st->memory_size(); 3107 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3108 if (i == 0) return NULL; // bail out 3109 Node* prev_mem = NULL; // raw memory for the captured store 3110 if (i > 0) { 3111 prev_mem = in(i); // there is a pre-existing store under this one 3112 set_req(i, C->top()); // temporarily disconnect it 3113 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3114 } else { 3115 i = -i; // no pre-existing store 3116 prev_mem = zero_memory(); // a slice of the newly allocated object 3117 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3118 set_req(--i, C->top()); // reuse this edge; it has been folded away 3119 else 3120 ins_req(i, C->top()); // build a new edge 3121 } 3122 Node* new_st = st->clone(); 3123 new_st->set_req(MemNode::Control, in(Control)); 3124 new_st->set_req(MemNode::Memory, prev_mem); 3125 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3126 new_st = phase->transform(new_st); 3127 3128 // At this point, new_st might have swallowed a pre-existing store 3129 // at the same offset, or perhaps new_st might have disappeared, 3130 // if it redundantly stored the same value (or zero to fresh memory). 3131 3132 // In any case, wire it in: 3133 set_req(i, new_st); 3134 3135 // The caller may now kill the old guy. 3136 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3137 assert(check_st == new_st || check_st == NULL, "must be findable"); 3138 assert(!is_complete(), ""); 3139 return new_st; 3140 } 3141 3142 static bool store_constant(jlong* tiles, int num_tiles, 3143 intptr_t st_off, int st_size, 3144 jlong con) { 3145 if ((st_off & (st_size-1)) != 0) 3146 return false; // strange store offset (assume size==2**N) 3147 address addr = (address)tiles + st_off; 3148 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3149 switch (st_size) { 3150 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3151 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3152 case sizeof(jint): *(jint*) addr = (jint) con; break; 3153 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3154 default: return false; // strange store size (detect size!=2**N here) 3155 } 3156 return true; // return success to caller 3157 } 3158 3159 // Coalesce subword constants into int constants and possibly 3160 // into long constants. The goal, if the CPU permits, 3161 // is to initialize the object with a small number of 64-bit tiles. 3162 // Also, convert floating-point constants to bit patterns. 3163 // Non-constants are not relevant to this pass. 3164 // 3165 // In terms of the running example on InitializeNode::InitializeNode 3166 // and InitializeNode::capture_store, here is the transformation 3167 // of rawstore1 and rawstore2 into rawstore12: 3168 // alloc = (Allocate ...) 3169 // rawoop = alloc.RawAddress 3170 // tile12 = 0x00010002 3171 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3172 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3173 // 3174 void 3175 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3176 Node* size_in_bytes, 3177 PhaseGVN* phase) { 3178 Compile* C = phase->C; 3179 3180 assert(stores_are_sane(phase), ""); 3181 // Note: After this pass, they are not completely sane, 3182 // since there may be some overlaps. 3183 3184 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3185 3186 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3187 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3188 size_limit = MIN2(size_limit, ti_limit); 3189 size_limit = align_size_up(size_limit, BytesPerLong); 3190 int num_tiles = size_limit / BytesPerLong; 3191 3192 // allocate space for the tile map: 3193 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3194 jlong tiles_buf[small_len]; 3195 Node* nodes_buf[small_len]; 3196 jlong inits_buf[small_len]; 3197 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3198 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3199 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3200 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3201 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3202 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3203 // tiles: exact bitwise model of all primitive constants 3204 // nodes: last constant-storing node subsumed into the tiles model 3205 // inits: which bytes (in each tile) are touched by any initializations 3206 3207 //// Pass A: Fill in the tile model with any relevant stores. 3208 3209 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3210 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3211 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3212 Node* zmem = zero_memory(); // initially zero memory state 3213 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3214 Node* st = in(i); 3215 intptr_t st_off = get_store_offset(st, phase); 3216 3217 // Figure out the store's offset and constant value: 3218 if (st_off < header_size) continue; //skip (ignore header) 3219 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3220 int st_size = st->as_Store()->memory_size(); 3221 if (st_off + st_size > size_limit) break; 3222 3223 // Record which bytes are touched, whether by constant or not. 3224 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3225 continue; // skip (strange store size) 3226 3227 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3228 if (!val->singleton()) continue; //skip (non-con store) 3229 BasicType type = val->basic_type(); 3230 3231 jlong con = 0; 3232 switch (type) { 3233 case T_INT: con = val->is_int()->get_con(); break; 3234 case T_LONG: con = val->is_long()->get_con(); break; 3235 case T_FLOAT: con = jint_cast(val->getf()); break; 3236 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3237 default: continue; //skip (odd store type) 3238 } 3239 3240 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3241 st->Opcode() == Op_StoreL) { 3242 continue; // This StoreL is already optimal. 3243 } 3244 3245 // Store down the constant. 3246 store_constant(tiles, num_tiles, st_off, st_size, con); 3247 3248 intptr_t j = st_off >> LogBytesPerLong; 3249 3250 if (type == T_INT && st_size == BytesPerInt 3251 && (st_off & BytesPerInt) == BytesPerInt) { 3252 jlong lcon = tiles[j]; 3253 if (!Matcher::isSimpleConstant64(lcon) && 3254 st->Opcode() == Op_StoreI) { 3255 // This StoreI is already optimal by itself. 3256 jint* intcon = (jint*) &tiles[j]; 3257 intcon[1] = 0; // undo the store_constant() 3258 3259 // If the previous store is also optimal by itself, back up and 3260 // undo the action of the previous loop iteration... if we can. 3261 // But if we can't, just let the previous half take care of itself. 3262 st = nodes[j]; 3263 st_off -= BytesPerInt; 3264 con = intcon[0]; 3265 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3266 assert(st_off >= header_size, "still ignoring header"); 3267 assert(get_store_offset(st, phase) == st_off, "must be"); 3268 assert(in(i-1) == zmem, "must be"); 3269 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3270 assert(con == tcon->is_int()->get_con(), "must be"); 3271 // Undo the effects of the previous loop trip, which swallowed st: 3272 intcon[0] = 0; // undo store_constant() 3273 set_req(i-1, st); // undo set_req(i, zmem) 3274 nodes[j] = NULL; // undo nodes[j] = st 3275 --old_subword; // undo ++old_subword 3276 } 3277 continue; // This StoreI is already optimal. 3278 } 3279 } 3280 3281 // This store is not needed. 3282 set_req(i, zmem); 3283 nodes[j] = st; // record for the moment 3284 if (st_size < BytesPerLong) // something has changed 3285 ++old_subword; // includes int/float, but who's counting... 3286 else ++old_long; 3287 } 3288 3289 if ((old_subword + old_long) == 0) 3290 return; // nothing more to do 3291 3292 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3293 // Be sure to insert them before overlapping non-constant stores. 3294 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3295 for (int j = 0; j < num_tiles; j++) { 3296 jlong con = tiles[j]; 3297 jlong init = inits[j]; 3298 if (con == 0) continue; 3299 jint con0, con1; // split the constant, address-wise 3300 jint init0, init1; // split the init map, address-wise 3301 { union { jlong con; jint intcon[2]; } u; 3302 u.con = con; 3303 con0 = u.intcon[0]; 3304 con1 = u.intcon[1]; 3305 u.con = init; 3306 init0 = u.intcon[0]; 3307 init1 = u.intcon[1]; 3308 } 3309 3310 Node* old = nodes[j]; 3311 assert(old != NULL, "need the prior store"); 3312 intptr_t offset = (j * BytesPerLong); 3313 3314 bool split = !Matcher::isSimpleConstant64(con); 3315 3316 if (offset < header_size) { 3317 assert(offset + BytesPerInt >= header_size, "second int counts"); 3318 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3319 split = true; // only the second word counts 3320 // Example: int a[] = { 42 ... } 3321 } else if (con0 == 0 && init0 == -1) { 3322 split = true; // first word is covered by full inits 3323 // Example: int a[] = { ... foo(), 42 ... } 3324 } else if (con1 == 0 && init1 == -1) { 3325 split = true; // second word is covered by full inits 3326 // Example: int a[] = { ... 42, foo() ... } 3327 } 3328 3329 // Here's a case where init0 is neither 0 nor -1: 3330 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3331 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3332 // In this case the tile is not split; it is (jlong)42. 3333 // The big tile is stored down, and then the foo() value is inserted. 3334 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3335 3336 Node* ctl = old->in(MemNode::Control); 3337 Node* adr = make_raw_address(offset, phase); 3338 const TypePtr* atp = TypeRawPtr::BOTTOM; 3339 3340 // One or two coalesced stores to plop down. 3341 Node* st[2]; 3342 intptr_t off[2]; 3343 int nst = 0; 3344 if (!split) { 3345 ++new_long; 3346 off[nst] = offset; 3347 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3348 phase->longcon(con), T_LONG); 3349 } else { 3350 // Omit either if it is a zero. 3351 if (con0 != 0) { 3352 ++new_int; 3353 off[nst] = offset; 3354 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3355 phase->intcon(con0), T_INT); 3356 } 3357 if (con1 != 0) { 3358 ++new_int; 3359 offset += BytesPerInt; 3360 adr = make_raw_address(offset, phase); 3361 off[nst] = offset; 3362 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3363 phase->intcon(con1), T_INT); 3364 } 3365 } 3366 3367 // Insert second store first, then the first before the second. 3368 // Insert each one just before any overlapping non-constant stores. 3369 while (nst > 0) { 3370 Node* st1 = st[--nst]; 3371 C->copy_node_notes_to(st1, old); 3372 st1 = phase->transform(st1); 3373 offset = off[nst]; 3374 assert(offset >= header_size, "do not smash header"); 3375 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3376 guarantee(ins_idx != 0, "must re-insert constant store"); 3377 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3378 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3379 set_req(--ins_idx, st1); 3380 else 3381 ins_req(ins_idx, st1); 3382 } 3383 } 3384 3385 if (PrintCompilation && WizardMode) 3386 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3387 old_subword, old_long, new_int, new_long); 3388 if (C->log() != NULL) 3389 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3390 old_subword, old_long, new_int, new_long); 3391 3392 // Clean up any remaining occurrences of zmem: 3393 remove_extra_zeroes(); 3394 } 3395 3396 // Explore forward from in(start) to find the first fully initialized 3397 // word, and return its offset. Skip groups of subword stores which 3398 // together initialize full words. If in(start) is itself part of a 3399 // fully initialized word, return the offset of in(start). If there 3400 // are no following full-word stores, or if something is fishy, return 3401 // a negative value. 3402 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3403 int int_map = 0; 3404 intptr_t int_map_off = 0; 3405 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3406 3407 for (uint i = start, limit = req(); i < limit; i++) { 3408 Node* st = in(i); 3409 3410 intptr_t st_off = get_store_offset(st, phase); 3411 if (st_off < 0) break; // return conservative answer 3412 3413 int st_size = st->as_Store()->memory_size(); 3414 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3415 return st_off; // we found a complete word init 3416 } 3417 3418 // update the map: 3419 3420 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3421 if (this_int_off != int_map_off) { 3422 // reset the map: 3423 int_map = 0; 3424 int_map_off = this_int_off; 3425 } 3426 3427 int subword_off = st_off - this_int_off; 3428 int_map |= right_n_bits(st_size) << subword_off; 3429 if ((int_map & FULL_MAP) == FULL_MAP) { 3430 return this_int_off; // we found a complete word init 3431 } 3432 3433 // Did this store hit or cross the word boundary? 3434 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3435 if (next_int_off == this_int_off + BytesPerInt) { 3436 // We passed the current int, without fully initializing it. 3437 int_map_off = next_int_off; 3438 int_map >>= BytesPerInt; 3439 } else if (next_int_off > this_int_off + BytesPerInt) { 3440 // We passed the current and next int. 3441 return this_int_off + BytesPerInt; 3442 } 3443 } 3444 3445 return -1; 3446 } 3447 3448 3449 // Called when the associated AllocateNode is expanded into CFG. 3450 // At this point, we may perform additional optimizations. 3451 // Linearize the stores by ascending offset, to make memory 3452 // activity as coherent as possible. 3453 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3454 intptr_t header_size, 3455 Node* size_in_bytes, 3456 PhaseGVN* phase) { 3457 assert(!is_complete(), "not already complete"); 3458 assert(stores_are_sane(phase), ""); 3459 assert(allocation() != NULL, "must be present"); 3460 3461 remove_extra_zeroes(); 3462 3463 if (ReduceFieldZeroing || ReduceBulkZeroing) 3464 // reduce instruction count for common initialization patterns 3465 coalesce_subword_stores(header_size, size_in_bytes, phase); 3466 3467 Node* zmem = zero_memory(); // initially zero memory state 3468 Node* inits = zmem; // accumulating a linearized chain of inits 3469 #ifdef ASSERT 3470 intptr_t first_offset = allocation()->minimum_header_size(); 3471 intptr_t last_init_off = first_offset; // previous init offset 3472 intptr_t last_init_end = first_offset; // previous init offset+size 3473 intptr_t last_tile_end = first_offset; // previous tile offset+size 3474 #endif 3475 intptr_t zeroes_done = header_size; 3476 3477 bool do_zeroing = true; // we might give up if inits are very sparse 3478 int big_init_gaps = 0; // how many large gaps have we seen? 3479 3480 if (ZeroTLAB) do_zeroing = false; 3481 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3482 3483 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3484 Node* st = in(i); 3485 intptr_t st_off = get_store_offset(st, phase); 3486 if (st_off < 0) 3487 break; // unknown junk in the inits 3488 if (st->in(MemNode::Memory) != zmem) 3489 break; // complicated store chains somehow in list 3490 3491 int st_size = st->as_Store()->memory_size(); 3492 intptr_t next_init_off = st_off + st_size; 3493 3494 if (do_zeroing && zeroes_done < next_init_off) { 3495 // See if this store needs a zero before it or under it. 3496 intptr_t zeroes_needed = st_off; 3497 3498 if (st_size < BytesPerInt) { 3499 // Look for subword stores which only partially initialize words. 3500 // If we find some, we must lay down some word-level zeroes first, 3501 // underneath the subword stores. 3502 // 3503 // Examples: 3504 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3505 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3506 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3507 // 3508 // Note: coalesce_subword_stores may have already done this, 3509 // if it was prompted by constant non-zero subword initializers. 3510 // But this case can still arise with non-constant stores. 3511 3512 intptr_t next_full_store = find_next_fullword_store(i, phase); 3513 3514 // In the examples above: 3515 // in(i) p q r s x y z 3516 // st_off 12 13 14 15 12 13 14 3517 // st_size 1 1 1 1 1 1 1 3518 // next_full_s. 12 16 16 16 16 16 16 3519 // z's_done 12 16 16 16 12 16 12 3520 // z's_needed 12 16 16 16 16 16 16 3521 // zsize 0 0 0 0 4 0 4 3522 if (next_full_store < 0) { 3523 // Conservative tack: Zero to end of current word. 3524 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3525 } else { 3526 // Zero to beginning of next fully initialized word. 3527 // Or, don't zero at all, if we are already in that word. 3528 assert(next_full_store >= zeroes_needed, "must go forward"); 3529 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3530 zeroes_needed = next_full_store; 3531 } 3532 } 3533 3534 if (zeroes_needed > zeroes_done) { 3535 intptr_t zsize = zeroes_needed - zeroes_done; 3536 // Do some incremental zeroing on rawmem, in parallel with inits. 3537 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3538 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3539 zeroes_done, zeroes_needed, 3540 phase); 3541 zeroes_done = zeroes_needed; 3542 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3543 do_zeroing = false; // leave the hole, next time 3544 } 3545 } 3546 3547 // Collect the store and move on: 3548 st->set_req(MemNode::Memory, inits); 3549 inits = st; // put it on the linearized chain 3550 set_req(i, zmem); // unhook from previous position 3551 3552 if (zeroes_done == st_off) 3553 zeroes_done = next_init_off; 3554 3555 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3556 3557 #ifdef ASSERT 3558 // Various order invariants. Weaker than stores_are_sane because 3559 // a large constant tile can be filled in by smaller non-constant stores. 3560 assert(st_off >= last_init_off, "inits do not reverse"); 3561 last_init_off = st_off; 3562 const Type* val = NULL; 3563 if (st_size >= BytesPerInt && 3564 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3565 (int)val->basic_type() < (int)T_OBJECT) { 3566 assert(st_off >= last_tile_end, "tiles do not overlap"); 3567 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3568 last_tile_end = MAX2(last_tile_end, next_init_off); 3569 } else { 3570 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3571 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3572 assert(st_off >= last_init_end, "inits do not overlap"); 3573 last_init_end = next_init_off; // it's a non-tile 3574 } 3575 #endif //ASSERT 3576 } 3577 3578 remove_extra_zeroes(); // clear out all the zmems left over 3579 add_req(inits); 3580 3581 if (!ZeroTLAB) { 3582 // If anything remains to be zeroed, zero it all now. 3583 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3584 // if it is the last unused 4 bytes of an instance, forget about it 3585 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3586 if (zeroes_done + BytesPerLong >= size_limit) { 3587 assert(allocation() != NULL, ""); 3588 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3589 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3590 if (zeroes_done == k->layout_helper()) 3591 zeroes_done = size_limit; 3592 } 3593 if (zeroes_done < size_limit) { 3594 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3595 zeroes_done, size_in_bytes, phase); 3596 } 3597 } 3598 3599 set_complete(phase); 3600 return rawmem; 3601 } 3602 3603 3604 #ifdef ASSERT 3605 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3606 if (is_complete()) 3607 return true; // stores could be anything at this point 3608 assert(allocation() != NULL, "must be present"); 3609 intptr_t last_off = allocation()->minimum_header_size(); 3610 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3611 Node* st = in(i); 3612 intptr_t st_off = get_store_offset(st, phase); 3613 if (st_off < 0) continue; // ignore dead garbage 3614 if (last_off > st_off) { 3615 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3616 this->dump(2); 3617 assert(false, "ascending store offsets"); 3618 return false; 3619 } 3620 last_off = st_off + st->as_Store()->memory_size(); 3621 } 3622 return true; 3623 } 3624 #endif //ASSERT 3625 3626 3627 3628 3629 //============================MergeMemNode===================================== 3630 // 3631 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3632 // contributing store or call operations. Each contributor provides the memory 3633 // state for a particular "alias type" (see Compile::alias_type). For example, 3634 // if a MergeMem has an input X for alias category #6, then any memory reference 3635 // to alias category #6 may use X as its memory state input, as an exact equivalent 3636 // to using the MergeMem as a whole. 3637 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3638 // 3639 // (Here, the <N> notation gives the index of the relevant adr_type.) 3640 // 3641 // In one special case (and more cases in the future), alias categories overlap. 3642 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3643 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3644 // it is exactly equivalent to that state W: 3645 // MergeMem(<Bot>: W) <==> W 3646 // 3647 // Usually, the merge has more than one input. In that case, where inputs 3648 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3649 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3650 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3651 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3652 // 3653 // A merge can take a "wide" memory state as one of its narrow inputs. 3654 // This simply means that the merge observes out only the relevant parts of 3655 // the wide input. That is, wide memory states arriving at narrow merge inputs 3656 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3657 // 3658 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3659 // and that memory slices "leak through": 3660 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3661 // 3662 // But, in such a cascade, repeated memory slices can "block the leak": 3663 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3664 // 3665 // In the last example, Y is not part of the combined memory state of the 3666 // outermost MergeMem. The system must, of course, prevent unschedulable 3667 // memory states from arising, so you can be sure that the state Y is somehow 3668 // a precursor to state Y'. 3669 // 3670 // 3671 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3672 // of each MergeMemNode array are exactly the numerical alias indexes, including 3673 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3674 // Compile::alias_type (and kin) produce and manage these indexes. 3675 // 3676 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3677 // (Note that this provides quick access to the top node inside MergeMem methods, 3678 // without the need to reach out via TLS to Compile::current.) 3679 // 3680 // As a consequence of what was just described, a MergeMem that represents a full 3681 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3682 // containing all alias categories. 3683 // 3684 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3685 // 3686 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3687 // a memory state for the alias type <N>, or else the top node, meaning that 3688 // there is no particular input for that alias type. Note that the length of 3689 // a MergeMem is variable, and may be extended at any time to accommodate new 3690 // memory states at larger alias indexes. When merges grow, they are of course 3691 // filled with "top" in the unused in() positions. 3692 // 3693 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3694 // (Top was chosen because it works smoothly with passes like GCM.) 3695 // 3696 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3697 // the type of random VM bits like TLS references.) Since it is always the 3698 // first non-Bot memory slice, some low-level loops use it to initialize an 3699 // index variable: for (i = AliasIdxRaw; i < req(); i++). 3700 // 3701 // 3702 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3703 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3704 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 3705 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3706 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3707 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3708 // 3709 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3710 // really that different from the other memory inputs. An abbreviation called 3711 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3712 // 3713 // 3714 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3715 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3716 // that "emerges though" the base memory will be marked as excluding the alias types 3717 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 3718 // 3719 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3720 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3721 // 3722 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 3723 // (It is currently unimplemented.) As you can see, the resulting merge is 3724 // actually a disjoint union of memory states, rather than an overlay. 3725 // 3726 3727 //------------------------------MergeMemNode----------------------------------- 3728 Node* MergeMemNode::make_empty_memory() { 3729 Node* empty_memory = (Node*) Compile::current()->top(); 3730 assert(empty_memory->is_top(), "correct sentinel identity"); 3731 return empty_memory; 3732 } 3733 3734 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3735 init_class_id(Class_MergeMem); 3736 // all inputs are nullified in Node::Node(int) 3737 // set_input(0, NULL); // no control input 3738 3739 // Initialize the edges uniformly to top, for starters. 3740 Node* empty_mem = make_empty_memory(); 3741 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3742 init_req(i,empty_mem); 3743 } 3744 assert(empty_memory() == empty_mem, ""); 3745 3746 if( new_base != NULL && new_base->is_MergeMem() ) { 3747 MergeMemNode* mdef = new_base->as_MergeMem(); 3748 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3749 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3750 mms.set_memory(mms.memory2()); 3751 } 3752 assert(base_memory() == mdef->base_memory(), ""); 3753 } else { 3754 set_base_memory(new_base); 3755 } 3756 } 3757 3758 // Make a new, untransformed MergeMem with the same base as 'mem'. 3759 // If mem is itself a MergeMem, populate the result with the same edges. 3760 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3761 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3762 } 3763 3764 //------------------------------cmp-------------------------------------------- 3765 uint MergeMemNode::hash() const { return NO_HASH; } 3766 uint MergeMemNode::cmp( const Node &n ) const { 3767 return (&n == this); // Always fail except on self 3768 } 3769 3770 //------------------------------Identity--------------------------------------- 3771 Node* MergeMemNode::Identity(PhaseTransform *phase) { 3772 // Identity if this merge point does not record any interesting memory 3773 // disambiguations. 3774 Node* base_mem = base_memory(); 3775 Node* empty_mem = empty_memory(); 3776 if (base_mem != empty_mem) { // Memory path is not dead? 3777 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3778 Node* mem = in(i); 3779 if (mem != empty_mem && mem != base_mem) { 3780 return this; // Many memory splits; no change 3781 } 3782 } 3783 } 3784 return base_mem; // No memory splits; ID on the one true input 3785 } 3786 3787 //------------------------------Ideal------------------------------------------ 3788 // This method is invoked recursively on chains of MergeMem nodes 3789 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3790 // Remove chain'd MergeMems 3791 // 3792 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3793 // relative to the "in(Bot)". Since we are patching both at the same time, 3794 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3795 // but rewrite each "in(i)" relative to the new "in(Bot)". 3796 Node *progress = NULL; 3797 3798 3799 Node* old_base = base_memory(); 3800 Node* empty_mem = empty_memory(); 3801 if (old_base == empty_mem) 3802 return NULL; // Dead memory path. 3803 3804 MergeMemNode* old_mbase; 3805 if (old_base != NULL && old_base->is_MergeMem()) 3806 old_mbase = old_base->as_MergeMem(); 3807 else 3808 old_mbase = NULL; 3809 Node* new_base = old_base; 3810 3811 // simplify stacked MergeMems in base memory 3812 if (old_mbase) new_base = old_mbase->base_memory(); 3813 3814 // the base memory might contribute new slices beyond my req() 3815 if (old_mbase) grow_to_match(old_mbase); 3816 3817 // Look carefully at the base node if it is a phi. 3818 PhiNode* phi_base; 3819 if (new_base != NULL && new_base->is_Phi()) 3820 phi_base = new_base->as_Phi(); 3821 else 3822 phi_base = NULL; 3823 3824 Node* phi_reg = NULL; 3825 uint phi_len = (uint)-1; 3826 if (phi_base != NULL && !phi_base->is_copy()) { 3827 // do not examine phi if degraded to a copy 3828 phi_reg = phi_base->region(); 3829 phi_len = phi_base->req(); 3830 // see if the phi is unfinished 3831 for (uint i = 1; i < phi_len; i++) { 3832 if (phi_base->in(i) == NULL) { 3833 // incomplete phi; do not look at it yet! 3834 phi_reg = NULL; 3835 phi_len = (uint)-1; 3836 break; 3837 } 3838 } 3839 } 3840 3841 // Note: We do not call verify_sparse on entry, because inputs 3842 // can normalize to the base_memory via subsume_node or similar 3843 // mechanisms. This method repairs that damage. 3844 3845 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3846 3847 // Look at each slice. 3848 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3849 Node* old_in = in(i); 3850 // calculate the old memory value 3851 Node* old_mem = old_in; 3852 if (old_mem == empty_mem) old_mem = old_base; 3853 assert(old_mem == memory_at(i), ""); 3854 3855 // maybe update (reslice) the old memory value 3856 3857 // simplify stacked MergeMems 3858 Node* new_mem = old_mem; 3859 MergeMemNode* old_mmem; 3860 if (old_mem != NULL && old_mem->is_MergeMem()) 3861 old_mmem = old_mem->as_MergeMem(); 3862 else 3863 old_mmem = NULL; 3864 if (old_mmem == this) { 3865 // This can happen if loops break up and safepoints disappear. 3866 // A merge of BotPtr (default) with a RawPtr memory derived from a 3867 // safepoint can be rewritten to a merge of the same BotPtr with 3868 // the BotPtr phi coming into the loop. If that phi disappears 3869 // also, we can end up with a self-loop of the mergemem. 3870 // In general, if loops degenerate and memory effects disappear, 3871 // a mergemem can be left looking at itself. This simply means 3872 // that the mergemem's default should be used, since there is 3873 // no longer any apparent effect on this slice. 3874 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3875 // from start. Update the input to TOP. 3876 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3877 } 3878 else if (old_mmem != NULL) { 3879 new_mem = old_mmem->memory_at(i); 3880 } 3881 // else preceding memory was not a MergeMem 3882 3883 // replace equivalent phis (unfortunately, they do not GVN together) 3884 if (new_mem != NULL && new_mem != new_base && 3885 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3886 if (new_mem->is_Phi()) { 3887 PhiNode* phi_mem = new_mem->as_Phi(); 3888 for (uint i = 1; i < phi_len; i++) { 3889 if (phi_base->in(i) != phi_mem->in(i)) { 3890 phi_mem = NULL; 3891 break; 3892 } 3893 } 3894 if (phi_mem != NULL) { 3895 // equivalent phi nodes; revert to the def 3896 new_mem = new_base; 3897 } 3898 } 3899 } 3900 3901 // maybe store down a new value 3902 Node* new_in = new_mem; 3903 if (new_in == new_base) new_in = empty_mem; 3904 3905 if (new_in != old_in) { 3906 // Warning: Do not combine this "if" with the previous "if" 3907 // A memory slice might have be be rewritten even if it is semantically 3908 // unchanged, if the base_memory value has changed. 3909 set_req(i, new_in); 3910 progress = this; // Report progress 3911 } 3912 } 3913 3914 if (new_base != old_base) { 3915 set_req(Compile::AliasIdxBot, new_base); 3916 // Don't use set_base_memory(new_base), because we need to update du. 3917 assert(base_memory() == new_base, ""); 3918 progress = this; 3919 } 3920 3921 if( base_memory() == this ) { 3922 // a self cycle indicates this memory path is dead 3923 set_req(Compile::AliasIdxBot, empty_mem); 3924 } 3925 3926 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3927 // Recursion must occur after the self cycle check above 3928 if( base_memory()->is_MergeMem() ) { 3929 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3930 Node *m = phase->transform(new_mbase); // Rollup any cycles 3931 if( m != NULL && (m->is_top() || 3932 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3933 // propagate rollup of dead cycle to self 3934 set_req(Compile::AliasIdxBot, empty_mem); 3935 } 3936 } 3937 3938 if( base_memory() == empty_mem ) { 3939 progress = this; 3940 // Cut inputs during Parse phase only. 3941 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3942 if( !can_reshape ) { 3943 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3944 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3945 } 3946 } 3947 } 3948 3949 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3950 // Check if PhiNode::Ideal's "Split phis through memory merges" 3951 // transform should be attempted. Look for this->phi->this cycle. 3952 uint merge_width = req(); 3953 if (merge_width > Compile::AliasIdxRaw) { 3954 PhiNode* phi = base_memory()->as_Phi(); 3955 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3956 if (phi->in(i) == this) { 3957 phase->is_IterGVN()->_worklist.push(phi); 3958 break; 3959 } 3960 } 3961 } 3962 } 3963 3964 assert(progress || verify_sparse(), "please, no dups of base"); 3965 return progress; 3966 } 3967 3968 //-------------------------set_base_memory------------------------------------- 3969 void MergeMemNode::set_base_memory(Node *new_base) { 3970 Node* empty_mem = empty_memory(); 3971 set_req(Compile::AliasIdxBot, new_base); 3972 assert(memory_at(req()) == new_base, "must set default memory"); 3973 // Clear out other occurrences of new_base: 3974 if (new_base != empty_mem) { 3975 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3976 if (in(i) == new_base) set_req(i, empty_mem); 3977 } 3978 } 3979 } 3980 3981 //------------------------------out_RegMask------------------------------------ 3982 const RegMask &MergeMemNode::out_RegMask() const { 3983 return RegMask::Empty; 3984 } 3985 3986 //------------------------------dump_spec-------------------------------------- 3987 #ifndef PRODUCT 3988 void MergeMemNode::dump_spec(outputStream *st) const { 3989 st->print(" {"); 3990 Node* base_mem = base_memory(); 3991 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3992 Node* mem = memory_at(i); 3993 if (mem == base_mem) { st->print(" -"); continue; } 3994 st->print( " N%d:", mem->_idx ); 3995 Compile::current()->get_adr_type(i)->dump_on(st); 3996 } 3997 st->print(" }"); 3998 } 3999 #endif // !PRODUCT 4000 4001 4002 #ifdef ASSERT 4003 static bool might_be_same(Node* a, Node* b) { 4004 if (a == b) return true; 4005 if (!(a->is_Phi() || b->is_Phi())) return false; 4006 // phis shift around during optimization 4007 return true; // pretty stupid... 4008 } 4009 4010 // verify a narrow slice (either incoming or outgoing) 4011 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4012 if (!VerifyAliases) return; // don't bother to verify unless requested 4013 if (is_error_reported()) return; // muzzle asserts when debugging an error 4014 if (Node::in_dump()) return; // muzzle asserts when printing 4015 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4016 assert(n != NULL, ""); 4017 // Elide intervening MergeMem's 4018 while (n->is_MergeMem()) { 4019 n = n->as_MergeMem()->memory_at(alias_idx); 4020 } 4021 Compile* C = Compile::current(); 4022 const TypePtr* n_adr_type = n->adr_type(); 4023 if (n == m->empty_memory()) { 4024 // Implicit copy of base_memory() 4025 } else if (n_adr_type != TypePtr::BOTTOM) { 4026 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4027 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4028 } else { 4029 // A few places like make_runtime_call "know" that VM calls are narrow, 4030 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4031 bool expected_wide_mem = false; 4032 if (n == m->base_memory()) { 4033 expected_wide_mem = true; 4034 } else if (alias_idx == Compile::AliasIdxRaw || 4035 n == m->memory_at(Compile::AliasIdxRaw)) { 4036 expected_wide_mem = true; 4037 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4038 // memory can "leak through" calls on channels that 4039 // are write-once. Allow this also. 4040 expected_wide_mem = true; 4041 } 4042 assert(expected_wide_mem, "expected narrow slice replacement"); 4043 } 4044 } 4045 #else // !ASSERT 4046 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 4047 #endif 4048 4049 4050 //-----------------------------memory_at--------------------------------------- 4051 Node* MergeMemNode::memory_at(uint alias_idx) const { 4052 assert(alias_idx >= Compile::AliasIdxRaw || 4053 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4054 "must avoid base_memory and AliasIdxTop"); 4055 4056 // Otherwise, it is a narrow slice. 4057 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4058 Compile *C = Compile::current(); 4059 if (is_empty_memory(n)) { 4060 // the array is sparse; empty slots are the "top" node 4061 n = base_memory(); 4062 assert(Node::in_dump() 4063 || n == NULL || n->bottom_type() == Type::TOP 4064 || n->adr_type() == TypePtr::BOTTOM 4065 || n->adr_type() == TypeRawPtr::BOTTOM 4066 || Compile::current()->AliasLevel() == 0, 4067 "must be a wide memory"); 4068 // AliasLevel == 0 if we are organizing the memory states manually. 4069 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4070 } else { 4071 // make sure the stored slice is sane 4072 #ifdef ASSERT 4073 if (is_error_reported() || Node::in_dump()) { 4074 } else if (might_be_same(n, base_memory())) { 4075 // Give it a pass: It is a mostly harmless repetition of the base. 4076 // This can arise normally from node subsumption during optimization. 4077 } else { 4078 verify_memory_slice(this, alias_idx, n); 4079 } 4080 #endif 4081 } 4082 return n; 4083 } 4084 4085 //---------------------------set_memory_at------------------------------------- 4086 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4087 verify_memory_slice(this, alias_idx, n); 4088 Node* empty_mem = empty_memory(); 4089 if (n == base_memory()) n = empty_mem; // collapse default 4090 uint need_req = alias_idx+1; 4091 if (req() < need_req) { 4092 if (n == empty_mem) return; // already the default, so do not grow me 4093 // grow the sparse array 4094 do { 4095 add_req(empty_mem); 4096 } while (req() < need_req); 4097 } 4098 set_req( alias_idx, n ); 4099 } 4100 4101 4102 4103 //--------------------------iteration_setup------------------------------------ 4104 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4105 if (other != NULL) { 4106 grow_to_match(other); 4107 // invariant: the finite support of mm2 is within mm->req() 4108 #ifdef ASSERT 4109 for (uint i = req(); i < other->req(); i++) { 4110 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4111 } 4112 #endif 4113 } 4114 // Replace spurious copies of base_memory by top. 4115 Node* base_mem = base_memory(); 4116 if (base_mem != NULL && !base_mem->is_top()) { 4117 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4118 if (in(i) == base_mem) 4119 set_req(i, empty_memory()); 4120 } 4121 } 4122 } 4123 4124 //---------------------------grow_to_match------------------------------------- 4125 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4126 Node* empty_mem = empty_memory(); 4127 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4128 // look for the finite support of the other memory 4129 for (uint i = other->req(); --i >= req(); ) { 4130 if (other->in(i) != empty_mem) { 4131 uint new_len = i+1; 4132 while (req() < new_len) add_req(empty_mem); 4133 break; 4134 } 4135 } 4136 } 4137 4138 //---------------------------verify_sparse------------------------------------- 4139 #ifndef PRODUCT 4140 bool MergeMemNode::verify_sparse() const { 4141 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4142 Node* base_mem = base_memory(); 4143 // The following can happen in degenerate cases, since empty==top. 4144 if (is_empty_memory(base_mem)) return true; 4145 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4146 assert(in(i) != NULL, "sane slice"); 4147 if (in(i) == base_mem) return false; // should have been the sentinel value! 4148 } 4149 return true; 4150 } 4151 4152 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4153 Node* n; 4154 n = mm->in(idx); 4155 if (mem == n) return true; // might be empty_memory() 4156 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4157 if (mem == n) return true; 4158 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4159 if (mem == n) return true; 4160 if (n == NULL) break; 4161 } 4162 return false; 4163 } 4164 #endif // !PRODUCT