1 /* 2 * Copyright (c) 1997, 2012, 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 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "oops/objArrayKlass.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/compile.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/machnode.hpp" 36 #include "opto/matcher.hpp" 37 #include "opto/memnode.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/phaseX.hpp" 40 #include "opto/regmask.hpp" 41 42 // Portions of code courtesy of Clifford Click 43 44 // Optimization - Graph Style 45 46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 47 48 //============================================================================= 49 uint MemNode::size_of() const { return sizeof(*this); } 50 51 const TypePtr *MemNode::adr_type() const { 52 Node* adr = in(Address); 53 const TypePtr* cross_check = NULL; 54 DEBUG_ONLY(cross_check = _adr_type); 55 return calculate_adr_type(adr->bottom_type(), cross_check); 56 } 57 58 #ifndef PRODUCT 59 void MemNode::dump_spec(outputStream *st) const { 60 if (in(Address) == NULL) return; // node is dead 61 #ifndef ASSERT 62 // fake the missing field 63 const TypePtr* _adr_type = NULL; 64 if (in(Address) != NULL) 65 _adr_type = in(Address)->bottom_type()->isa_ptr(); 66 #endif 67 dump_adr_type(this, _adr_type, st); 68 69 Compile* C = Compile::current(); 70 if( C->alias_type(_adr_type)->is_volatile() ) 71 st->print(" Volatile!"); 72 } 73 74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 75 st->print(" @"); 76 if (adr_type == NULL) { 77 st->print("NULL"); 78 } else { 79 adr_type->dump_on(st); 80 Compile* C = Compile::current(); 81 Compile::AliasType* atp = NULL; 82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 83 if (atp == NULL) 84 st->print(", idx=?\?;"); 85 else if (atp->index() == Compile::AliasIdxBot) 86 st->print(", idx=Bot;"); 87 else if (atp->index() == Compile::AliasIdxTop) 88 st->print(", idx=Top;"); 89 else if (atp->index() == Compile::AliasIdxRaw) 90 st->print(", idx=Raw;"); 91 else { 92 ciField* field = atp->field(); 93 if (field) { 94 st->print(", name="); 95 field->print_name_on(st); 96 } 97 st->print(", idx=%d;", atp->index()); 98 } 99 } 100 } 101 102 extern void print_alias_types(); 103 104 #endif 105 106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 107 const TypeOopPtr *tinst = t_adr->isa_oopptr(); 108 if (tinst == NULL || !tinst->is_known_instance_field()) 109 return mchain; // don't try to optimize non-instance types 110 uint instance_id = tinst->instance_id(); 111 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 112 Node *prev = NULL; 113 Node *result = mchain; 114 while (prev != result) { 115 prev = result; 116 if (result == start_mem) 117 break; // hit one of our sentinels 118 // skip over a call which does not affect this memory slice 119 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 120 Node *proj_in = result->in(0); 121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 122 break; // hit one of our sentinels 123 } else if (proj_in->is_Call()) { 124 CallNode *call = proj_in->as_Call(); 125 if (!call->may_modify(t_adr, phase)) { 126 result = call->in(TypeFunc::Memory); 127 } 128 } else if (proj_in->is_Initialize()) { 129 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 130 // Stop if this is the initialization for the object instance which 131 // which contains this memory slice, otherwise skip over it. 132 if (alloc != NULL && alloc->_idx != instance_id) { 133 result = proj_in->in(TypeFunc::Memory); 134 } 135 } else if (proj_in->is_MemBar()) { 136 result = proj_in->in(TypeFunc::Memory); 137 } else { 138 assert(false, "unexpected projection"); 139 } 140 } else if (result->is_ClearArray()) { 141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) { 142 // Can not bypass initialization of the instance 143 // we are looking for. 144 break; 145 } 146 // Otherwise skip it (the call updated 'result' value). 147 } else if (result->is_MergeMem()) { 148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 149 } 150 } 151 return result; 152 } 153 154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 155 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); 157 PhaseIterGVN *igvn = phase->is_IterGVN(); 158 Node *result = mchain; 159 result = optimize_simple_memory_chain(result, t_adr, phase); 160 if (is_instance && igvn != NULL && result->is_Phi()) { 161 PhiNode *mphi = result->as_Phi(); 162 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 163 const TypePtr *t = mphi->adr_type(); 164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 166 t->is_oopptr()->cast_to_exactness(true) 167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 168 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 169 // clone the Phi with our address type 170 result = mphi->split_out_instance(t_adr, igvn); 171 } else { 172 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 173 } 174 } 175 return result; 176 } 177 178 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 179 uint alias_idx = phase->C->get_alias_index(tp); 180 Node *mem = mmem; 181 #ifdef ASSERT 182 { 183 // Check that current type is consistent with the alias index used during graph construction 184 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 185 bool consistent = adr_check == NULL || adr_check->empty() || 186 phase->C->must_alias(adr_check, alias_idx ); 187 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 188 if( !consistent && adr_check != NULL && !adr_check->empty() && 189 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 190 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 191 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 192 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 193 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 194 // don't assert if it is dead code. 195 consistent = true; 196 } 197 if( !consistent ) { 198 st->print("alias_idx==%d, adr_check==", alias_idx); 199 if( adr_check == NULL ) { 200 st->print("NULL"); 201 } else { 202 adr_check->dump(); 203 } 204 st->cr(); 205 print_alias_types(); 206 assert(consistent, "adr_check must match alias idx"); 207 } 208 } 209 #endif 210 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 211 // means an array I have not precisely typed yet. Do not do any 212 // alias stuff with it any time soon. 213 const TypeOopPtr *toop = tp->isa_oopptr(); 214 if( tp->base() != Type::AnyPtr && 215 !(toop && 216 toop->klass() != NULL && 217 toop->klass()->is_java_lang_Object() && 218 toop->offset() == Type::OffsetBot) ) { 219 // compress paths and change unreachable cycles to TOP 220 // If not, we can update the input infinitely along a MergeMem cycle 221 // Equivalent code in PhiNode::Ideal 222 Node* m = phase->transform(mmem); 223 // If transformed to a MergeMem, get the desired slice 224 // Otherwise the returned node represents memory for every slice 225 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 226 // Update input if it is progress over what we have now 227 } 228 return mem; 229 } 230 231 //--------------------------Ideal_common--------------------------------------- 232 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 233 // Unhook non-raw memories from complete (macro-expanded) initializations. 234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 235 // If our control input is a dead region, kill all below the region 236 Node *ctl = in(MemNode::Control); 237 if (ctl && remove_dead_region(phase, can_reshape)) 238 return this; 239 ctl = in(MemNode::Control); 240 // Don't bother trying to transform a dead node 241 if (ctl && ctl->is_top()) return NodeSentinel; 242 243 PhaseIterGVN *igvn = phase->is_IterGVN(); 244 // Wait if control on the worklist. 245 if (ctl && can_reshape && igvn != NULL) { 246 Node* bol = NULL; 247 Node* cmp = NULL; 248 if (ctl->in(0)->is_If()) { 249 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 250 bol = ctl->in(0)->in(1); 251 if (bol->is_Bool()) 252 cmp = ctl->in(0)->in(1)->in(1); 253 } 254 if (igvn->_worklist.member(ctl) || 255 (bol != NULL && igvn->_worklist.member(bol)) || 256 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 257 // This control path may be dead. 258 // Delay this memory node transformation until the control is processed. 259 phase->is_IterGVN()->_worklist.push(this); 260 return NodeSentinel; // caller will return NULL 261 } 262 } 263 // Ignore if memory is dead, or self-loop 264 Node *mem = in(MemNode::Memory); 265 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 266 assert(mem != this, "dead loop in MemNode::Ideal"); 267 268 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 269 // This memory slice may be dead. 270 // Delay this mem node transformation until the memory is processed. 271 phase->is_IterGVN()->_worklist.push(this); 272 return NodeSentinel; // caller will return NULL 273 } 274 275 Node *address = in(MemNode::Address); 276 const Type *t_adr = phase->type(address); 277 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 278 279 if (can_reshape && igvn != NULL && 280 (igvn->_worklist.member(address) || 281 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { 282 // The address's base and type may change when the address is processed. 283 // Delay this mem node transformation until the address is processed. 284 phase->is_IterGVN()->_worklist.push(this); 285 return NodeSentinel; // caller will return NULL 286 } 287 288 // Do NOT remove or optimize the next lines: ensure a new alias index 289 // is allocated for an oop pointer type before Escape Analysis. 290 // Note: C++ will not remove it since the call has side effect. 291 if (t_adr->isa_oopptr()) { 292 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 293 } 294 295 #ifdef ASSERT 296 Node* base = NULL; 297 if (address->is_AddP()) 298 base = address->in(AddPNode::Base); 299 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 300 !t_adr->isa_rawptr()) { 301 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 302 Compile* C = phase->C; 303 tty->cr(); 304 tty->print_cr("===== NULL+offs not RAW address ====="); 305 if (C->is_dead_node(this->_idx)) tty->print_cr("'this' is dead"); 306 if ((ctl != NULL) && C->is_dead_node(ctl->_idx)) tty->print_cr("'ctl' is dead"); 307 if (C->is_dead_node(mem->_idx)) tty->print_cr("'mem' is dead"); 308 if (C->is_dead_node(address->_idx)) tty->print_cr("'address' is dead"); 309 if (C->is_dead_node(base->_idx)) tty->print_cr("'base' is dead"); 310 tty->cr(); 311 base->dump(1); 312 tty->cr(); 313 this->dump(2); 314 tty->print("this->adr_type(): "); adr_type()->dump(); tty->cr(); 315 tty->print("phase->type(address): "); t_adr->dump(); tty->cr(); 316 tty->print("phase->type(base): "); phase->type(address)->dump(); tty->cr(); 317 tty->cr(); 318 } 319 assert(base == NULL || t_adr->isa_rawptr() || 320 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); 321 #endif 322 323 // Avoid independent memory operations 324 Node* old_mem = mem; 325 326 // The code which unhooks non-raw memories from complete (macro-expanded) 327 // initializations was removed. After macro-expansion all stores catched 328 // by Initialize node became raw stores and there is no information 329 // which memory slices they modify. So it is unsafe to move any memory 330 // operation above these stores. Also in most cases hooked non-raw memories 331 // were already unhooked by using information from detect_ptr_independence() 332 // and find_previous_store(). 333 334 if (mem->is_MergeMem()) { 335 MergeMemNode* mmem = mem->as_MergeMem(); 336 const TypePtr *tp = t_adr->is_ptr(); 337 338 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 339 } 340 341 if (mem != old_mem) { 342 set_req(MemNode::Memory, mem); 343 if (can_reshape && old_mem->outcnt() == 0) { 344 igvn->_worklist.push(old_mem); 345 } 346 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 347 return this; 348 } 349 350 // let the subclass continue analyzing... 351 return NULL; 352 } 353 354 // Helper function for proving some simple control dominations. 355 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 356 // Already assumes that 'dom' is available at 'sub', and that 'sub' 357 // is not a constant (dominated by the method's StartNode). 358 // Used by MemNode::find_previous_store to prove that the 359 // control input of a memory operation predates (dominates) 360 // an allocation it wants to look past. 361 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 362 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 363 return false; // Conservative answer for dead code 364 365 // Check 'dom'. Skip Proj and CatchProj nodes. 366 dom = dom->find_exact_control(dom); 367 if (dom == NULL || dom->is_top()) 368 return false; // Conservative answer for dead code 369 370 if (dom == sub) { 371 // For the case when, for example, 'sub' is Initialize and the original 372 // 'dom' is Proj node of the 'sub'. 373 return false; 374 } 375 376 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 377 return true; 378 379 // 'dom' dominates 'sub' if its control edge and control edges 380 // of all its inputs dominate or equal to sub's control edge. 381 382 // Currently 'sub' is either Allocate, Initialize or Start nodes. 383 // Or Region for the check in LoadNode::Ideal(); 384 // 'sub' should have sub->in(0) != NULL. 385 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 386 sub->is_Region(), "expecting only these nodes"); 387 388 // Get control edge of 'sub'. 389 Node* orig_sub = sub; 390 sub = sub->find_exact_control(sub->in(0)); 391 if (sub == NULL || sub->is_top()) 392 return false; // Conservative answer for dead code 393 394 assert(sub->is_CFG(), "expecting control"); 395 396 if (sub == dom) 397 return true; 398 399 if (sub->is_Start() || sub->is_Root()) 400 return false; 401 402 { 403 // Check all control edges of 'dom'. 404 405 ResourceMark rm; 406 Arena* arena = Thread::current()->resource_area(); 407 Node_List nlist(arena); 408 Unique_Node_List dom_list(arena); 409 410 dom_list.push(dom); 411 bool only_dominating_controls = false; 412 413 for (uint next = 0; next < dom_list.size(); next++) { 414 Node* n = dom_list.at(next); 415 if (n == orig_sub) 416 return false; // One of dom's inputs dominated by sub. 417 if (!n->is_CFG() && n->pinned()) { 418 // Check only own control edge for pinned non-control nodes. 419 n = n->find_exact_control(n->in(0)); 420 if (n == NULL || n->is_top()) 421 return false; // Conservative answer for dead code 422 assert(n->is_CFG(), "expecting control"); 423 dom_list.push(n); 424 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 425 only_dominating_controls = true; 426 } else if (n->is_CFG()) { 427 if (n->dominates(sub, nlist)) 428 only_dominating_controls = true; 429 else 430 return false; 431 } else { 432 // First, own control edge. 433 Node* m = n->find_exact_control(n->in(0)); 434 if (m != NULL) { 435 if (m->is_top()) 436 return false; // Conservative answer for dead code 437 dom_list.push(m); 438 } 439 // Now, the rest of edges. 440 uint cnt = n->req(); 441 for (uint i = 1; i < cnt; i++) { 442 m = n->find_exact_control(n->in(i)); 443 if (m == NULL || m->is_top()) 444 continue; 445 dom_list.push(m); 446 } 447 } 448 } 449 return only_dominating_controls; 450 } 451 } 452 453 //---------------------detect_ptr_independence--------------------------------- 454 // Used by MemNode::find_previous_store to prove that two base 455 // pointers are never equal. 456 // The pointers are accompanied by their associated allocations, 457 // if any, which have been previously discovered by the caller. 458 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 459 Node* p2, AllocateNode* a2, 460 PhaseTransform* phase) { 461 // Attempt to prove that these two pointers cannot be aliased. 462 // They may both manifestly be allocations, and they should differ. 463 // Or, if they are not both allocations, they can be distinct constants. 464 // Otherwise, one is an allocation and the other a pre-existing value. 465 if (a1 == NULL && a2 == NULL) { // neither an allocation 466 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 467 } else if (a1 != NULL && a2 != NULL) { // both allocations 468 return (a1 != a2); 469 } else if (a1 != NULL) { // one allocation a1 470 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 471 return all_controls_dominate(p2, a1); 472 } else { //(a2 != NULL) // one allocation a2 473 return all_controls_dominate(p1, a2); 474 } 475 return false; 476 } 477 478 479 // The logic for reordering loads and stores uses four steps: 480 // (a) Walk carefully past stores and initializations which we 481 // can prove are independent of this load. 482 // (b) Observe that the next memory state makes an exact match 483 // with self (load or store), and locate the relevant store. 484 // (c) Ensure that, if we were to wire self directly to the store, 485 // the optimizer would fold it up somehow. 486 // (d) Do the rewiring, and return, depending on some other part of 487 // the optimizer to fold up the load. 488 // This routine handles steps (a) and (b). Steps (c) and (d) are 489 // specific to loads and stores, so they are handled by the callers. 490 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 491 // 492 Node* MemNode::find_previous_store(PhaseTransform* phase) { 493 Node* ctrl = in(MemNode::Control); 494 Node* adr = in(MemNode::Address); 495 intptr_t offset = 0; 496 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 497 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 498 499 if (offset == Type::OffsetBot) 500 return NULL; // cannot unalias unless there are precise offsets 501 502 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 503 504 intptr_t size_in_bytes = memory_size(); 505 506 Node* mem = in(MemNode::Memory); // start searching here... 507 508 int cnt = 50; // Cycle limiter 509 for (;;) { // While we can dance past unrelated stores... 510 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 511 512 if (mem->is_Store()) { 513 Node* st_adr = mem->in(MemNode::Address); 514 intptr_t st_offset = 0; 515 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 516 if (st_base == NULL) 517 break; // inscrutable pointer 518 if (st_offset != offset && st_offset != Type::OffsetBot) { 519 const int MAX_STORE = BytesPerLong; 520 if (st_offset >= offset + size_in_bytes || 521 st_offset <= offset - MAX_STORE || 522 st_offset <= offset - mem->as_Store()->memory_size()) { 523 // Success: The offsets are provably independent. 524 // (You may ask, why not just test st_offset != offset and be done? 525 // The answer is that stores of different sizes can co-exist 526 // in the same sequence of RawMem effects. We sometimes initialize 527 // a whole 'tile' of array elements with a single jint or jlong.) 528 mem = mem->in(MemNode::Memory); 529 continue; // (a) advance through independent store memory 530 } 531 } 532 if (st_base != base && 533 detect_ptr_independence(base, alloc, 534 st_base, 535 AllocateNode::Ideal_allocation(st_base, phase), 536 phase)) { 537 // Success: The bases are provably independent. 538 mem = mem->in(MemNode::Memory); 539 continue; // (a) advance through independent store memory 540 } 541 542 // (b) At this point, if the bases or offsets do not agree, we lose, 543 // since we have not managed to prove 'this' and 'mem' independent. 544 if (st_base == base && st_offset == offset) { 545 return mem; // let caller handle steps (c), (d) 546 } 547 548 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 549 InitializeNode* st_init = mem->in(0)->as_Initialize(); 550 AllocateNode* st_alloc = st_init->allocation(); 551 if (st_alloc == NULL) 552 break; // something degenerated 553 bool known_identical = false; 554 bool known_independent = false; 555 if (alloc == st_alloc) 556 known_identical = true; 557 else if (alloc != NULL) 558 known_independent = true; 559 else if (all_controls_dominate(this, st_alloc)) 560 known_independent = true; 561 562 if (known_independent) { 563 // The bases are provably independent: Either they are 564 // manifestly distinct allocations, or else the control 565 // of this load dominates the store's allocation. 566 int alias_idx = phase->C->get_alias_index(adr_type()); 567 if (alias_idx == Compile::AliasIdxRaw) { 568 mem = st_alloc->in(TypeFunc::Memory); 569 } else { 570 mem = st_init->memory(alias_idx); 571 } 572 continue; // (a) advance through independent store memory 573 } 574 575 // (b) at this point, if we are not looking at a store initializing 576 // the same allocation we are loading from, we lose. 577 if (known_identical) { 578 // From caller, can_see_stored_value will consult find_captured_store. 579 return mem; // let caller handle steps (c), (d) 580 } 581 582 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 583 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 584 if (mem->is_Proj() && mem->in(0)->is_Call()) { 585 CallNode *call = mem->in(0)->as_Call(); 586 if (!call->may_modify(addr_t, phase)) { 587 mem = call->in(TypeFunc::Memory); 588 continue; // (a) advance through independent call memory 589 } 590 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 591 mem = mem->in(0)->in(TypeFunc::Memory); 592 continue; // (a) advance through independent MemBar memory 593 } else if (mem->is_ClearArray()) { 594 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 595 // (the call updated 'mem' value) 596 continue; // (a) advance through independent allocation memory 597 } else { 598 // Can not bypass initialization of the instance 599 // we are looking for. 600 return mem; 601 } 602 } else if (mem->is_MergeMem()) { 603 int alias_idx = phase->C->get_alias_index(adr_type()); 604 mem = mem->as_MergeMem()->memory_at(alias_idx); 605 continue; // (a) advance through independent MergeMem memory 606 } 607 } 608 609 // Unless there is an explicit 'continue', we must bail out here, 610 // because 'mem' is an inscrutable memory state (e.g., a call). 611 break; 612 } 613 614 return NULL; // bail out 615 } 616 617 //----------------------calculate_adr_type------------------------------------- 618 // Helper function. Notices when the given type of address hits top or bottom. 619 // Also, asserts a cross-check of the type against the expected address type. 620 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 621 if (t == Type::TOP) return NULL; // does not touch memory any more? 622 #ifdef PRODUCT 623 cross_check = NULL; 624 #else 625 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 626 #endif 627 const TypePtr* tp = t->isa_ptr(); 628 if (tp == NULL) { 629 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 630 return TypePtr::BOTTOM; // touches lots of memory 631 } else { 632 #ifdef ASSERT 633 // %%%% [phh] We don't check the alias index if cross_check is 634 // TypeRawPtr::BOTTOM. Needs to be investigated. 635 if (cross_check != NULL && 636 cross_check != TypePtr::BOTTOM && 637 cross_check != TypeRawPtr::BOTTOM) { 638 // Recheck the alias index, to see if it has changed (due to a bug). 639 Compile* C = Compile::current(); 640 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 641 "must stay in the original alias category"); 642 // The type of the address must be contained in the adr_type, 643 // disregarding "null"-ness. 644 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 645 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 646 assert(cross_check->meet(tp_notnull) == cross_check, 647 "real address must not escape from expected memory type"); 648 } 649 #endif 650 return tp; 651 } 652 } 653 654 //------------------------adr_phi_is_loop_invariant---------------------------- 655 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 656 // loop is loop invariant. Make a quick traversal of Phi and associated 657 // CastPP nodes, looking to see if they are a closed group within the loop. 658 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 659 // The idea is that the phi-nest must boil down to only CastPP nodes 660 // with the same data. This implies that any path into the loop already 661 // includes such a CastPP, and so the original cast, whatever its input, 662 // must be covered by an equivalent cast, with an earlier control input. 663 ResourceMark rm; 664 665 // The loop entry input of the phi should be the unique dominating 666 // node for every Phi/CastPP in the loop. 667 Unique_Node_List closure; 668 closure.push(adr_phi->in(LoopNode::EntryControl)); 669 670 // Add the phi node and the cast to the worklist. 671 Unique_Node_List worklist; 672 worklist.push(adr_phi); 673 if( cast != NULL ){ 674 if( !cast->is_ConstraintCast() ) return false; 675 worklist.push(cast); 676 } 677 678 // Begin recursive walk of phi nodes. 679 while( worklist.size() ){ 680 // Take a node off the worklist 681 Node *n = worklist.pop(); 682 if( !closure.member(n) ){ 683 // Add it to the closure. 684 closure.push(n); 685 // Make a sanity check to ensure we don't waste too much time here. 686 if( closure.size() > 20) return false; 687 // This node is OK if: 688 // - it is a cast of an identical value 689 // - or it is a phi node (then we add its inputs to the worklist) 690 // Otherwise, the node is not OK, and we presume the cast is not invariant 691 if( n->is_ConstraintCast() ){ 692 worklist.push(n->in(1)); 693 } else if( n->is_Phi() ) { 694 for( uint i = 1; i < n->req(); i++ ) { 695 worklist.push(n->in(i)); 696 } 697 } else { 698 return false; 699 } 700 } 701 } 702 703 // Quit when the worklist is empty, and we've found no offending nodes. 704 return true; 705 } 706 707 //------------------------------Ideal_DU_postCCP------------------------------- 708 // Find any cast-away of null-ness and keep its control. Null cast-aways are 709 // going away in this pass and we need to make this memory op depend on the 710 // gating null check. 711 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 712 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 713 } 714 715 // I tried to leave the CastPP's in. This makes the graph more accurate in 716 // some sense; we get to keep around the knowledge that an oop is not-null 717 // after some test. Alas, the CastPP's interfere with GVN (some values are 718 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 719 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 720 // some of the more trivial cases in the optimizer. Removing more useless 721 // Phi's started allowing Loads to illegally float above null checks. I gave 722 // up on this approach. CNC 10/20/2000 723 // This static method may be called not from MemNode (EncodePNode calls it). 724 // Only the control edge of the node 'n' might be updated. 725 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 726 Node *skipped_cast = NULL; 727 // Need a null check? Regular static accesses do not because they are 728 // from constant addresses. Array ops are gated by the range check (which 729 // always includes a NULL check). Just check field ops. 730 if( n->in(MemNode::Control) == NULL ) { 731 // Scan upwards for the highest location we can place this memory op. 732 while( true ) { 733 switch( adr->Opcode() ) { 734 735 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 736 adr = adr->in(AddPNode::Base); 737 continue; 738 739 case Op_DecodeN: // No change to NULL-ness, so peek thru 740 case Op_DecodeNKlass: 741 adr = adr->in(1); 742 continue; 743 744 case Op_EncodeP: 745 case Op_EncodePKlass: 746 // EncodeP node's control edge could be set by this method 747 // when EncodeP node depends on CastPP node. 748 // 749 // Use its control edge for memory op because EncodeP may go away 750 // later when it is folded with following or preceding DecodeN node. 751 if (adr->in(0) == NULL) { 752 // Keep looking for cast nodes. 753 adr = adr->in(1); 754 continue; 755 } 756 ccp->hash_delete(n); 757 n->set_req(MemNode::Control, adr->in(0)); 758 ccp->hash_insert(n); 759 return n; 760 761 case Op_CastPP: 762 // If the CastPP is useless, just peek on through it. 763 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 764 // Remember the cast that we've peeked though. If we peek 765 // through more than one, then we end up remembering the highest 766 // one, that is, if in a loop, the one closest to the top. 767 skipped_cast = adr; 768 adr = adr->in(1); 769 continue; 770 } 771 // CastPP is going away in this pass! We need this memory op to be 772 // control-dependent on the test that is guarding the CastPP. 773 ccp->hash_delete(n); 774 n->set_req(MemNode::Control, adr->in(0)); 775 ccp->hash_insert(n); 776 return n; 777 778 case Op_Phi: 779 // Attempt to float above a Phi to some dominating point. 780 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 781 // If we've already peeked through a Cast (which could have set the 782 // control), we can't float above a Phi, because the skipped Cast 783 // may not be loop invariant. 784 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 785 adr = adr->in(1); 786 continue; 787 } 788 } 789 790 // Intentional fallthrough! 791 792 // No obvious dominating point. The mem op is pinned below the Phi 793 // by the Phi itself. If the Phi goes away (no true value is merged) 794 // then the mem op can float, but not indefinitely. It must be pinned 795 // behind the controls leading to the Phi. 796 case Op_CheckCastPP: 797 // These usually stick around to change address type, however a 798 // useless one can be elided and we still need to pick up a control edge 799 if (adr->in(0) == NULL) { 800 // This CheckCastPP node has NO control and is likely useless. But we 801 // need check further up the ancestor chain for a control input to keep 802 // the node in place. 4959717. 803 skipped_cast = adr; 804 adr = adr->in(1); 805 continue; 806 } 807 ccp->hash_delete(n); 808 n->set_req(MemNode::Control, adr->in(0)); 809 ccp->hash_insert(n); 810 return n; 811 812 // List of "safe" opcodes; those that implicitly block the memory 813 // op below any null check. 814 case Op_CastX2P: // no null checks on native pointers 815 case Op_Parm: // 'this' pointer is not null 816 case Op_LoadP: // Loading from within a klass 817 case Op_LoadN: // Loading from within a klass 818 case Op_LoadKlass: // Loading from within a klass 819 case Op_LoadNKlass: // Loading from within a klass 820 case Op_ConP: // Loading from a klass 821 case Op_ConN: // Loading from a klass 822 case Op_ConNKlass: // Loading from a klass 823 case Op_CreateEx: // Sucking up the guts of an exception oop 824 case Op_Con: // Reading from TLS 825 case Op_CMoveP: // CMoveP is pinned 826 case Op_CMoveN: // CMoveN is pinned 827 break; // No progress 828 829 case Op_Proj: // Direct call to an allocation routine 830 case Op_SCMemProj: // Memory state from store conditional ops 831 #ifdef ASSERT 832 { 833 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 834 const Node* call = adr->in(0); 835 if (call->is_CallJava()) { 836 const CallJavaNode* call_java = call->as_CallJava(); 837 const TypeTuple *r = call_java->tf()->range(); 838 assert(r->cnt() > TypeFunc::Parms, "must return value"); 839 const Type* ret_type = r->field_at(TypeFunc::Parms); 840 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 841 // We further presume that this is one of 842 // new_instance_Java, new_array_Java, or 843 // the like, but do not assert for this. 844 } else if (call->is_Allocate()) { 845 // similar case to new_instance_Java, etc. 846 } else if (!call->is_CallLeaf()) { 847 // Projections from fetch_oop (OSR) are allowed as well. 848 ShouldNotReachHere(); 849 } 850 } 851 #endif 852 break; 853 default: 854 ShouldNotReachHere(); 855 } 856 break; 857 } 858 } 859 860 return NULL; // No progress 861 } 862 863 864 //============================================================================= 865 uint LoadNode::size_of() const { return sizeof(*this); } 866 uint LoadNode::cmp( const Node &n ) const 867 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 868 const Type *LoadNode::bottom_type() const { return _type; } 869 uint LoadNode::ideal_reg() const { 870 return _type->ideal_reg(); 871 } 872 873 #ifndef PRODUCT 874 void LoadNode::dump_spec(outputStream *st) const { 875 MemNode::dump_spec(st); 876 if( !Verbose && !WizardMode ) { 877 // standard dump does this in Verbose and WizardMode 878 st->print(" #"); _type->dump_on(st); 879 } 880 } 881 #endif 882 883 #ifdef ASSERT 884 //----------------------------is_immutable_value------------------------------- 885 // Helper function to allow a raw load without control edge for some cases 886 bool LoadNode::is_immutable_value(Node* adr) { 887 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 888 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 889 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 890 in_bytes(JavaThread::osthread_offset()))); 891 } 892 #endif 893 894 //----------------------------LoadNode::make----------------------------------- 895 // Polymorphic factory method: 896 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 897 Compile* C = gvn.C; 898 899 // sanity check the alias category against the created node type 900 assert(!(adr_type->isa_oopptr() && 901 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 902 "use LoadKlassNode instead"); 903 assert(!(adr_type->isa_aryptr() && 904 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 905 "use LoadRangeNode instead"); 906 // Check control edge of raw loads 907 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 908 // oop will be recorded in oop map if load crosses safepoint 909 rt->isa_oopptr() || is_immutable_value(adr), 910 "raw memory operations should have control edge"); 911 switch (bt) { 912 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 913 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 914 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 915 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 916 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 917 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 918 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt ); 919 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt ); 920 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 921 case T_OBJECT: 922 #ifdef _LP64 923 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 924 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 925 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr()); 926 } else 927 #endif 928 { 929 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 930 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 931 } 932 } 933 ShouldNotReachHere(); 934 return (LoadNode*)NULL; 935 } 936 937 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 938 bool require_atomic = true; 939 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 940 } 941 942 943 944 945 //------------------------------hash------------------------------------------- 946 uint LoadNode::hash() const { 947 // unroll addition of interesting fields 948 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 949 } 950 951 //---------------------------can_see_stored_value------------------------------ 952 // This routine exists to make sure this set of tests is done the same 953 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 954 // will change the graph shape in a way which makes memory alive twice at the 955 // same time (uses the Oracle model of aliasing), then some 956 // LoadXNode::Identity will fold things back to the equivalence-class model 957 // of aliasing. 958 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 959 Node* ld_adr = in(MemNode::Address); 960 961 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 962 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 963 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 964 atp->field() != NULL && !atp->field()->is_volatile()) { 965 uint alias_idx = atp->index(); 966 bool final = atp->field()->is_final(); 967 Node* result = NULL; 968 Node* current = st; 969 // Skip through chains of MemBarNodes checking the MergeMems for 970 // new states for the slice of this load. Stop once any other 971 // kind of node is encountered. Loads from final memory can skip 972 // through any kind of MemBar but normal loads shouldn't skip 973 // through MemBarAcquire since the could allow them to move out of 974 // a synchronized region. 975 while (current->is_Proj()) { 976 int opc = current->in(0)->Opcode(); 977 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) || 978 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder || 979 opc == Op_MemBarReleaseLock) { 980 Node* mem = current->in(0)->in(TypeFunc::Memory); 981 if (mem->is_MergeMem()) { 982 MergeMemNode* merge = mem->as_MergeMem(); 983 Node* new_st = merge->memory_at(alias_idx); 984 if (new_st == merge->base_memory()) { 985 // Keep searching 986 current = merge->base_memory(); 987 continue; 988 } 989 // Save the new memory state for the slice and fall through 990 // to exit. 991 result = new_st; 992 } 993 } 994 break; 995 } 996 if (result != NULL) { 997 st = result; 998 } 999 } 1000 1001 1002 // Loop around twice in the case Load -> Initialize -> Store. 1003 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1004 for (int trip = 0; trip <= 1; trip++) { 1005 1006 if (st->is_Store()) { 1007 Node* st_adr = st->in(MemNode::Address); 1008 if (!phase->eqv(st_adr, ld_adr)) { 1009 // Try harder before giving up... Match raw and non-raw pointers. 1010 intptr_t st_off = 0; 1011 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 1012 if (alloc == NULL) return NULL; 1013 intptr_t ld_off = 0; 1014 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 1015 if (alloc != allo2) return NULL; 1016 if (ld_off != st_off) return NULL; 1017 // At this point we have proven something like this setup: 1018 // A = Allocate(...) 1019 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 1020 // S = StoreQ(, AddP(, A.Parm , #Off), V) 1021 // (Actually, we haven't yet proven the Q's are the same.) 1022 // In other words, we are loading from a casted version of 1023 // the same pointer-and-offset that we stored to. 1024 // Thus, we are able to replace L by V. 1025 } 1026 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1027 if (store_Opcode() != st->Opcode()) 1028 return NULL; 1029 return st->in(MemNode::ValueIn); 1030 } 1031 1032 intptr_t offset = 0; // scratch 1033 1034 // A load from a freshly-created object always returns zero. 1035 // (This can happen after LoadNode::Ideal resets the load's memory input 1036 // to find_captured_store, which returned InitializeNode::zero_memory.) 1037 if (st->is_Proj() && st->in(0)->is_Allocate() && 1038 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 1039 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 1040 // return a zero value for the load's basic type 1041 // (This is one of the few places where a generic PhaseTransform 1042 // can create new nodes. Think of it as lazily manifesting 1043 // virtually pre-existing constants.) 1044 return phase->zerocon(memory_type()); 1045 } 1046 1047 // A load from an initialization barrier can match a captured store. 1048 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1049 InitializeNode* init = st->in(0)->as_Initialize(); 1050 AllocateNode* alloc = init->allocation(); 1051 if (alloc != NULL && 1052 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 1053 // examine a captured store value 1054 st = init->find_captured_store(offset, memory_size(), phase); 1055 if (st != NULL) 1056 continue; // take one more trip around 1057 } 1058 } 1059 1060 break; 1061 } 1062 1063 return NULL; 1064 } 1065 1066 //----------------------is_instance_field_load_with_local_phi------------------ 1067 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1068 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 1069 in(MemNode::Address)->is_AddP() ) { 1070 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 1071 // Only instances. 1072 if( t_oop != NULL && t_oop->is_known_instance_field() && 1073 t_oop->offset() != Type::OffsetBot && 1074 t_oop->offset() != Type::OffsetTop) { 1075 return true; 1076 } 1077 } 1078 return false; 1079 } 1080 1081 //------------------------------Identity--------------------------------------- 1082 // Loads are identity if previous store is to same address 1083 Node *LoadNode::Identity( PhaseTransform *phase ) { 1084 // If the previous store-maker is the right kind of Store, and the store is 1085 // to the same address, then we are equal to the value stored. 1086 Node* mem = in(MemNode::Memory); 1087 Node* value = can_see_stored_value(mem, phase); 1088 if( value ) { 1089 // byte, short & char stores truncate naturally. 1090 // A load has to load the truncated value which requires 1091 // some sort of masking operation and that requires an 1092 // Ideal call instead of an Identity call. 1093 if (memory_size() < BytesPerInt) { 1094 // If the input to the store does not fit with the load's result type, 1095 // it must be truncated via an Ideal call. 1096 if (!phase->type(value)->higher_equal(phase->type(this))) 1097 return this; 1098 } 1099 // (This works even when value is a Con, but LoadNode::Value 1100 // usually runs first, producing the singleton type of the Con.) 1101 return value; 1102 } 1103 1104 // Search for an existing data phi which was generated before for the same 1105 // instance's field to avoid infinite generation of phis in a loop. 1106 Node *region = mem->in(0); 1107 if (is_instance_field_load_with_local_phi(region)) { 1108 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 1109 int this_index = phase->C->get_alias_index(addr_t); 1110 int this_offset = addr_t->offset(); 1111 int this_id = addr_t->is_oopptr()->instance_id(); 1112 const Type* this_type = bottom_type(); 1113 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1114 Node* phi = region->fast_out(i); 1115 if (phi->is_Phi() && phi != mem && 1116 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 1117 return phi; 1118 } 1119 } 1120 } 1121 1122 return this; 1123 } 1124 1125 1126 // Returns true if the AliasType refers to the field that holds the 1127 // cached box array. Currently only handles the IntegerCache case. 1128 static bool is_autobox_cache(Compile::AliasType* atp) { 1129 if (atp != NULL && atp->field() != NULL) { 1130 ciField* field = atp->field(); 1131 ciSymbol* klass = field->holder()->name(); 1132 if (field->name() == ciSymbol::cache_field_name() && 1133 field->holder()->uses_default_loader() && 1134 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1135 return true; 1136 } 1137 } 1138 return false; 1139 } 1140 1141 // Fetch the base value in the autobox array 1142 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1143 if (atp != NULL && atp->field() != NULL) { 1144 ciField* field = atp->field(); 1145 ciSymbol* klass = field->holder()->name(); 1146 if (field->name() == ciSymbol::cache_field_name() && 1147 field->holder()->uses_default_loader() && 1148 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1149 assert(field->is_constant(), "what?"); 1150 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1151 // Fetch the box object at the base of the array and get its value 1152 ciInstance* box = array->obj_at(0)->as_instance(); 1153 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1154 if (ik->nof_nonstatic_fields() == 1) { 1155 // This should be true nonstatic_field_at requires calling 1156 // nof_nonstatic_fields so check it anyway 1157 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1158 cache_offset = c.as_int(); 1159 } 1160 return true; 1161 } 1162 } 1163 return false; 1164 } 1165 1166 // Returns true if the AliasType refers to the value field of an 1167 // autobox object. Currently only handles Integer. 1168 static bool is_autobox_object(Compile::AliasType* atp) { 1169 if (atp != NULL && atp->field() != NULL) { 1170 ciField* field = atp->field(); 1171 ciSymbol* klass = field->holder()->name(); 1172 if (field->name() == ciSymbol::value_name() && 1173 field->holder()->uses_default_loader() && 1174 klass == ciSymbol::java_lang_Integer()) { 1175 return true; 1176 } 1177 } 1178 return false; 1179 } 1180 1181 1182 // We're loading from an object which has autobox behaviour. 1183 // If this object is result of a valueOf call we'll have a phi 1184 // merging a newly allocated object and a load from the cache. 1185 // We want to replace this load with the original incoming 1186 // argument to the valueOf call. 1187 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1188 Node* base = in(Address)->in(AddPNode::Base); 1189 if (base->is_Phi() && base->req() == 3) { 1190 AllocateNode* allocation = NULL; 1191 int allocation_index = -1; 1192 int load_index = -1; 1193 for (uint i = 1; i < base->req(); i++) { 1194 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1195 if (allocation != NULL) { 1196 allocation_index = i; 1197 load_index = 3 - allocation_index; 1198 break; 1199 } 1200 } 1201 bool has_load = ( allocation != NULL && 1202 (base->in(load_index)->is_Load() || 1203 base->in(load_index)->is_DecodeN() && 1204 base->in(load_index)->in(1)->is_Load()) ); 1205 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1206 // Push the loads from the phi that comes from valueOf up 1207 // through it to allow elimination of the loads and the recovery 1208 // of the original value. 1209 Node* mem_phi = in(Memory); 1210 Node* offset = in(Address)->in(AddPNode::Offset); 1211 Node* region = base->in(0); 1212 1213 Node* in1 = clone(); 1214 Node* in1_addr = in1->in(Address)->clone(); 1215 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1216 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1217 in1_addr->set_req(AddPNode::Offset, offset); 1218 in1->set_req(0, region->in(allocation_index)); 1219 in1->set_req(Address, in1_addr); 1220 in1->set_req(Memory, mem_phi->in(allocation_index)); 1221 1222 Node* in2 = clone(); 1223 Node* in2_addr = in2->in(Address)->clone(); 1224 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1225 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1226 in2_addr->set_req(AddPNode::Offset, offset); 1227 in2->set_req(0, region->in(load_index)); 1228 in2->set_req(Address, in2_addr); 1229 in2->set_req(Memory, mem_phi->in(load_index)); 1230 1231 in1_addr = phase->transform(in1_addr); 1232 in1 = phase->transform(in1); 1233 in2_addr = phase->transform(in2_addr); 1234 in2 = phase->transform(in2); 1235 1236 PhiNode* result = PhiNode::make_blank(region, this); 1237 result->set_req(allocation_index, in1); 1238 result->set_req(load_index, in2); 1239 return result; 1240 } 1241 } else if (base->is_Load() || 1242 base->is_DecodeN() && base->in(1)->is_Load()) { 1243 if (base->is_DecodeN()) { 1244 // Get LoadN node which loads cached Integer object 1245 base = base->in(1); 1246 } 1247 // Eliminate the load of Integer.value for integers from the cache 1248 // array by deriving the value from the index into the array. 1249 // Capture the offset of the load and then reverse the computation. 1250 Node* load_base = base->in(Address)->in(AddPNode::Base); 1251 if (load_base->is_DecodeN()) { 1252 // Get LoadN node which loads IntegerCache.cache field 1253 load_base = load_base->in(1); 1254 } 1255 if (load_base != NULL) { 1256 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1257 intptr_t cache_offset; 1258 int shift = -1; 1259 Node* cache = NULL; 1260 if (is_autobox_cache(atp)) { 1261 shift = exact_log2(type2aelembytes(T_OBJECT)); 1262 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1263 } 1264 if (cache != NULL && base->in(Address)->is_AddP()) { 1265 Node* elements[4]; 1266 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1267 int cache_low; 1268 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1269 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1270 // Add up all the offsets making of the address of the load 1271 Node* result = elements[0]; 1272 for (int i = 1; i < count; i++) { 1273 result = phase->transform(new (phase->C) AddXNode(result, elements[i])); 1274 } 1275 // Remove the constant offset from the address and then 1276 // remove the scaling of the offset to recover the original index. 1277 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-offset))); 1278 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1279 // Peel the shift off directly but wrap it in a dummy node 1280 // since Ideal can't return existing nodes 1281 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0)); 1282 } else { 1283 result = new (phase->C) RShiftXNode(result, phase->intcon(shift)); 1284 } 1285 #ifdef _LP64 1286 result = new (phase->C) ConvL2INode(phase->transform(result)); 1287 #endif 1288 return result; 1289 } 1290 } 1291 } 1292 } 1293 return NULL; 1294 } 1295 1296 //------------------------------split_through_phi------------------------------ 1297 // Split instance field load through Phi. 1298 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1299 Node* mem = in(MemNode::Memory); 1300 Node* address = in(MemNode::Address); 1301 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1302 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1303 1304 assert(mem->is_Phi() && (t_oop != NULL) && 1305 t_oop->is_known_instance_field(), "invalide conditions"); 1306 1307 Node *region = mem->in(0); 1308 if (region == NULL) { 1309 return NULL; // Wait stable graph 1310 } 1311 uint cnt = mem->req(); 1312 for (uint i = 1; i < cnt; i++) { 1313 Node* rc = region->in(i); 1314 if (rc == NULL || phase->type(rc) == Type::TOP) 1315 return NULL; // Wait stable graph 1316 Node *in = mem->in(i); 1317 if (in == NULL) { 1318 return NULL; // Wait stable graph 1319 } 1320 } 1321 // Check for loop invariant. 1322 if (cnt == 3) { 1323 for (uint i = 1; i < cnt; i++) { 1324 Node *in = mem->in(i); 1325 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1326 if (m == mem) { 1327 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1328 return this; 1329 } 1330 } 1331 } 1332 // Split through Phi (see original code in loopopts.cpp). 1333 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1334 1335 // Do nothing here if Identity will find a value 1336 // (to avoid infinite chain of value phis generation). 1337 if (!phase->eqv(this, this->Identity(phase))) 1338 return NULL; 1339 1340 // Skip the split if the region dominates some control edge of the address. 1341 if (!MemNode::all_controls_dominate(address, region)) 1342 return NULL; 1343 1344 const Type* this_type = this->bottom_type(); 1345 int this_index = phase->C->get_alias_index(addr_t); 1346 int this_offset = addr_t->offset(); 1347 int this_iid = addr_t->is_oopptr()->instance_id(); 1348 PhaseIterGVN *igvn = phase->is_IterGVN(); 1349 Node *phi = new (igvn->C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1350 for (uint i = 1; i < region->req(); i++) { 1351 Node *x; 1352 Node* the_clone = NULL; 1353 if (region->in(i) == phase->C->top()) { 1354 x = phase->C->top(); // Dead path? Use a dead data op 1355 } else { 1356 x = this->clone(); // Else clone up the data op 1357 the_clone = x; // Remember for possible deletion. 1358 // Alter data node to use pre-phi inputs 1359 if (this->in(0) == region) { 1360 x->set_req(0, region->in(i)); 1361 } else { 1362 x->set_req(0, NULL); 1363 } 1364 for (uint j = 1; j < this->req(); j++) { 1365 Node *in = this->in(j); 1366 if (in->is_Phi() && in->in(0) == region) 1367 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 1368 } 1369 } 1370 // Check for a 'win' on some paths 1371 const Type *t = x->Value(igvn); 1372 1373 bool singleton = t->singleton(); 1374 1375 // See comments in PhaseIdealLoop::split_thru_phi(). 1376 if (singleton && t == Type::TOP) { 1377 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1378 } 1379 1380 if (singleton) { 1381 x = igvn->makecon(t); 1382 } else { 1383 // We now call Identity to try to simplify the cloned node. 1384 // Note that some Identity methods call phase->type(this). 1385 // Make sure that the type array is big enough for 1386 // our new node, even though we may throw the node away. 1387 // (This tweaking with igvn only works because x is a new node.) 1388 igvn->set_type(x, t); 1389 // If x is a TypeNode, capture any more-precise type permanently into Node 1390 // otherwise it will be not updated during igvn->transform since 1391 // igvn->type(x) is set to x->Value() already. 1392 x->raise_bottom_type(t); 1393 Node *y = x->Identity(igvn); 1394 if (y != x) { 1395 x = y; 1396 } else { 1397 y = igvn->hash_find(x); 1398 if (y) { 1399 x = y; 1400 } else { 1401 // Else x is a new node we are keeping 1402 // We do not need register_new_node_with_optimizer 1403 // because set_type has already been called. 1404 igvn->_worklist.push(x); 1405 } 1406 } 1407 } 1408 if (x != the_clone && the_clone != NULL) 1409 igvn->remove_dead_node(the_clone); 1410 phi->set_req(i, x); 1411 } 1412 // Record Phi 1413 igvn->register_new_node_with_optimizer(phi); 1414 return phi; 1415 } 1416 1417 //------------------------------Ideal------------------------------------------ 1418 // If the load is from Field memory and the pointer is non-null, we can 1419 // zero out the control input. 1420 // If the offset is constant and the base is an object allocation, 1421 // try to hook me up to the exact initializing store. 1422 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1423 Node* p = MemNode::Ideal_common(phase, can_reshape); 1424 if (p) return (p == NodeSentinel) ? NULL : p; 1425 1426 Node* ctrl = in(MemNode::Control); 1427 Node* address = in(MemNode::Address); 1428 1429 // Skip up past a SafePoint control. Cannot do this for Stores because 1430 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1431 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1432 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1433 ctrl = ctrl->in(0); 1434 set_req(MemNode::Control,ctrl); 1435 } 1436 1437 intptr_t ignore = 0; 1438 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1439 if (base != NULL 1440 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1441 // Check for useless control edge in some common special cases 1442 if (in(MemNode::Control) != NULL 1443 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1444 && all_controls_dominate(base, phase->C->start())) { 1445 // A method-invariant, non-null address (constant or 'this' argument). 1446 set_req(MemNode::Control, NULL); 1447 } 1448 1449 if (EliminateAutoBox && can_reshape) { 1450 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); 1451 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1452 if (is_autobox_object(atp)) { 1453 Node* result = eliminate_autobox(phase); 1454 if (result != NULL) return result; 1455 } 1456 } 1457 } 1458 1459 Node* mem = in(MemNode::Memory); 1460 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1461 1462 if (addr_t != NULL) { 1463 // try to optimize our memory input 1464 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1465 if (opt_mem != mem) { 1466 set_req(MemNode::Memory, opt_mem); 1467 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1468 return this; 1469 } 1470 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1471 if (can_reshape && opt_mem->is_Phi() && 1472 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1473 PhaseIterGVN *igvn = phase->is_IterGVN(); 1474 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1475 // Delay this transformation until memory Phi is processed. 1476 phase->is_IterGVN()->_worklist.push(this); 1477 return NULL; 1478 } 1479 // Split instance field load through Phi. 1480 Node* result = split_through_phi(phase); 1481 if (result != NULL) return result; 1482 } 1483 } 1484 1485 // Check for prior store with a different base or offset; make Load 1486 // independent. Skip through any number of them. Bail out if the stores 1487 // are in an endless dead cycle and report no progress. This is a key 1488 // transform for Reflection. However, if after skipping through the Stores 1489 // we can't then fold up against a prior store do NOT do the transform as 1490 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1491 // array memory alive twice: once for the hoisted Load and again after the 1492 // bypassed Store. This situation only works if EVERYBODY who does 1493 // anti-dependence work knows how to bypass. I.e. we need all 1494 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1495 // the alias index stuff. So instead, peek through Stores and IFF we can 1496 // fold up, do so. 1497 Node* prev_mem = find_previous_store(phase); 1498 // Steps (a), (b): Walk past independent stores to find an exact match. 1499 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1500 // (c) See if we can fold up on the spot, but don't fold up here. 1501 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1502 // just return a prior value, which is done by Identity calls. 1503 if (can_see_stored_value(prev_mem, phase)) { 1504 // Make ready for step (d): 1505 set_req(MemNode::Memory, prev_mem); 1506 return this; 1507 } 1508 } 1509 1510 return NULL; // No further progress 1511 } 1512 1513 // Helper to recognize certain Klass fields which are invariant across 1514 // some group of array types (e.g., int[] or all T[] where T < Object). 1515 const Type* 1516 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1517 ciKlass* klass) const { 1518 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1519 // The field is Klass::_modifier_flags. Return its (constant) value. 1520 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1521 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1522 return TypeInt::make(klass->modifier_flags()); 1523 } 1524 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1525 // The field is Klass::_access_flags. Return its (constant) value. 1526 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1527 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1528 return TypeInt::make(klass->access_flags()); 1529 } 1530 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1531 // The field is Klass::_layout_helper. Return its constant value if known. 1532 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1533 return TypeInt::make(klass->layout_helper()); 1534 } 1535 1536 // No match. 1537 return NULL; 1538 } 1539 1540 //------------------------------Value----------------------------------------- 1541 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1542 // Either input is TOP ==> the result is TOP 1543 Node* mem = in(MemNode::Memory); 1544 const Type *t1 = phase->type(mem); 1545 if (t1 == Type::TOP) return Type::TOP; 1546 Node* adr = in(MemNode::Address); 1547 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1548 if (tp == NULL || tp->empty()) return Type::TOP; 1549 int off = tp->offset(); 1550 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1551 Compile* C = phase->C; 1552 1553 // Try to guess loaded type from pointer type 1554 if (tp->base() == Type::AryPtr) { 1555 const Type *t = tp->is_aryptr()->elem(); 1556 // Don't do this for integer types. There is only potential profit if 1557 // the element type t is lower than _type; that is, for int types, if _type is 1558 // more restrictive than t. This only happens here if one is short and the other 1559 // char (both 16 bits), and in those cases we've made an intentional decision 1560 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1561 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1562 // 1563 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1564 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1565 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1566 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1567 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1568 // In fact, that could have been the original type of p1, and p1 could have 1569 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1570 // expression (LShiftL quux 3) independently optimized to the constant 8. 1571 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1572 && (_type->isa_vect() == NULL) 1573 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1574 // t might actually be lower than _type, if _type is a unique 1575 // concrete subclass of abstract class t. 1576 // Make sure the reference is not into the header, by comparing 1577 // the offset against the offset of the start of the array's data. 1578 // Different array types begin at slightly different offsets (12 vs. 16). 1579 // We choose T_BYTE as an example base type that is least restrictive 1580 // as to alignment, which will therefore produce the smallest 1581 // possible base offset. 1582 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1583 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1584 const Type* jt = t->join(_type); 1585 // In any case, do not allow the join, per se, to empty out the type. 1586 if (jt->empty() && !t->empty()) { 1587 // This can happen if a interface-typed array narrows to a class type. 1588 jt = _type; 1589 } 1590 1591 if (EliminateAutoBox && adr->is_AddP()) { 1592 // The pointers in the autobox arrays are always non-null 1593 Node* base = adr->in(AddPNode::Base); 1594 if (base != NULL && 1595 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { 1596 Compile::AliasType* atp = C->alias_type(base->adr_type()); 1597 if (is_autobox_cache(atp)) { 1598 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1599 } 1600 } 1601 } 1602 return jt; 1603 } 1604 } 1605 } else if (tp->base() == Type::InstPtr) { 1606 ciEnv* env = C->env(); 1607 const TypeInstPtr* tinst = tp->is_instptr(); 1608 ciKlass* klass = tinst->klass(); 1609 assert( off != Type::OffsetBot || 1610 // arrays can be cast to Objects 1611 tp->is_oopptr()->klass()->is_java_lang_Object() || 1612 // unsafe field access may not have a constant offset 1613 C->has_unsafe_access(), 1614 "Field accesses must be precise" ); 1615 // For oop loads, we expect the _type to be precise 1616 if (klass == env->String_klass() && 1617 adr->is_AddP() && off != Type::OffsetBot) { 1618 // For constant Strings treat the final fields as compile time constants. 1619 Node* base = adr->in(AddPNode::Base); 1620 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1621 if (t != NULL && t->singleton()) { 1622 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1623 if (field != NULL && field->is_final()) { 1624 ciObject* string = t->const_oop(); 1625 ciConstant constant = string->as_instance()->field_value(field); 1626 if (constant.basic_type() == T_INT) { 1627 return TypeInt::make(constant.as_int()); 1628 } else if (constant.basic_type() == T_ARRAY) { 1629 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1630 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1631 } else { 1632 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1633 } 1634 } 1635 } 1636 } 1637 } 1638 // Optimizations for constant objects 1639 ciObject* const_oop = tinst->const_oop(); 1640 if (const_oop != NULL) { 1641 // For constant CallSites treat the target field as a compile time constant. 1642 if (const_oop->is_call_site()) { 1643 ciCallSite* call_site = const_oop->as_call_site(); 1644 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1645 if (field != NULL && field->is_call_site_target()) { 1646 ciMethodHandle* target = call_site->get_target(); 1647 if (target != NULL) { // just in case 1648 ciConstant constant(T_OBJECT, target); 1649 const Type* t; 1650 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1651 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1652 } else { 1653 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1654 } 1655 // Add a dependence for invalidation of the optimization. 1656 if (!call_site->is_constant_call_site()) { 1657 C->dependencies()->assert_call_site_target_value(call_site, target); 1658 } 1659 return t; 1660 } 1661 } 1662 } 1663 } 1664 } else if (tp->base() == Type::KlassPtr) { 1665 assert( off != Type::OffsetBot || 1666 // arrays can be cast to Objects 1667 tp->is_klassptr()->klass()->is_java_lang_Object() || 1668 // also allow array-loading from the primary supertype 1669 // array during subtype checks 1670 Opcode() == Op_LoadKlass, 1671 "Field accesses must be precise" ); 1672 // For klass/static loads, we expect the _type to be precise 1673 } 1674 1675 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1676 if (tkls != NULL && !StressReflectiveCode) { 1677 ciKlass* klass = tkls->klass(); 1678 if (klass->is_loaded() && tkls->klass_is_exact()) { 1679 // We are loading a field from a Klass metaobject whose identity 1680 // is known at compile time (the type is "exact" or "precise"). 1681 // Check for fields we know are maintained as constants by the VM. 1682 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1683 // The field is Klass::_super_check_offset. Return its (constant) value. 1684 // (Folds up type checking code.) 1685 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1686 return TypeInt::make(klass->super_check_offset()); 1687 } 1688 // Compute index into primary_supers array 1689 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1690 // Check for overflowing; use unsigned compare to handle the negative case. 1691 if( depth < ciKlass::primary_super_limit() ) { 1692 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1693 // (Folds up type checking code.) 1694 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1695 ciKlass *ss = klass->super_of_depth(depth); 1696 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1697 } 1698 const Type* aift = load_array_final_field(tkls, klass); 1699 if (aift != NULL) return aift; 1700 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset()) 1701 && klass->is_array_klass()) { 1702 // The field is ArrayKlass::_component_mirror. Return its (constant) value. 1703 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1704 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1705 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1706 } 1707 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1708 // The field is Klass::_java_mirror. Return its (constant) value. 1709 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1710 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1711 return TypeInstPtr::make(klass->java_mirror()); 1712 } 1713 } 1714 1715 // We can still check if we are loading from the primary_supers array at a 1716 // shallow enough depth. Even though the klass is not exact, entries less 1717 // than or equal to its super depth are correct. 1718 if (klass->is_loaded() ) { 1719 ciType *inner = klass; 1720 while( inner->is_obj_array_klass() ) 1721 inner = inner->as_obj_array_klass()->base_element_type(); 1722 if( inner->is_instance_klass() && 1723 !inner->as_instance_klass()->flags().is_interface() ) { 1724 // Compute index into primary_supers array 1725 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1726 // Check for overflowing; use unsigned compare to handle the negative case. 1727 if( depth < ciKlass::primary_super_limit() && 1728 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1729 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1730 // (Folds up type checking code.) 1731 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1732 ciKlass *ss = klass->super_of_depth(depth); 1733 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1734 } 1735 } 1736 } 1737 1738 // If the type is enough to determine that the thing is not an array, 1739 // we can give the layout_helper a positive interval type. 1740 // This will help short-circuit some reflective code. 1741 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1742 && !klass->is_array_klass() // not directly typed as an array 1743 && !klass->is_interface() // specifically not Serializable & Cloneable 1744 && !klass->is_java_lang_Object() // not the supertype of all T[] 1745 ) { 1746 // Note: When interfaces are reliable, we can narrow the interface 1747 // test to (klass != Serializable && klass != Cloneable). 1748 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1749 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1750 // The key property of this type is that it folds up tests 1751 // for array-ness, since it proves that the layout_helper is positive. 1752 // Thus, a generic value like the basic object layout helper works fine. 1753 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1754 } 1755 } 1756 1757 // If we are loading from a freshly-allocated object, produce a zero, 1758 // if the load is provably beyond the header of the object. 1759 // (Also allow a variable load from a fresh array to produce zero.) 1760 const TypeOopPtr *tinst = tp->isa_oopptr(); 1761 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1762 if (ReduceFieldZeroing || is_instance) { 1763 Node* value = can_see_stored_value(mem,phase); 1764 if (value != NULL && value->is_Con()) { 1765 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1766 return value->bottom_type(); 1767 } 1768 } 1769 1770 if (is_instance) { 1771 // If we have an instance type and our memory input is the 1772 // programs's initial memory state, there is no matching store, 1773 // so just return a zero of the appropriate type 1774 Node *mem = in(MemNode::Memory); 1775 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1776 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1777 return Type::get_zero_type(_type->basic_type()); 1778 } 1779 } 1780 return _type; 1781 } 1782 1783 //------------------------------match_edge------------------------------------- 1784 // Do we Match on this edge index or not? Match only the address. 1785 uint LoadNode::match_edge(uint idx) const { 1786 return idx == MemNode::Address; 1787 } 1788 1789 //--------------------------LoadBNode::Ideal-------------------------------------- 1790 // 1791 // If the previous store is to the same address as this load, 1792 // and the value stored was larger than a byte, replace this load 1793 // with the value stored truncated to a byte. If no truncation is 1794 // needed, the replacement is done in LoadNode::Identity(). 1795 // 1796 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1797 Node* mem = in(MemNode::Memory); 1798 Node* value = can_see_stored_value(mem,phase); 1799 if( value && !phase->type(value)->higher_equal( _type ) ) { 1800 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) ); 1801 return new (phase->C) RShiftINode(result, phase->intcon(24)); 1802 } 1803 // Identity call will handle the case where truncation is not needed. 1804 return LoadNode::Ideal(phase, can_reshape); 1805 } 1806 1807 const Type* LoadBNode::Value(PhaseTransform *phase) const { 1808 Node* mem = in(MemNode::Memory); 1809 Node* value = can_see_stored_value(mem,phase); 1810 if (value != NULL && value->is_Con() && 1811 !value->bottom_type()->higher_equal(_type)) { 1812 // If the input to the store does not fit with the load's result type, 1813 // it must be truncated. We can't delay until Ideal call since 1814 // a singleton Value is needed for split_thru_phi optimization. 1815 int con = value->get_int(); 1816 return TypeInt::make((con << 24) >> 24); 1817 } 1818 return LoadNode::Value(phase); 1819 } 1820 1821 //--------------------------LoadUBNode::Ideal------------------------------------- 1822 // 1823 // If the previous store is to the same address as this load, 1824 // and the value stored was larger than a byte, replace this load 1825 // with the value stored truncated to a byte. If no truncation is 1826 // needed, the replacement is done in LoadNode::Identity(). 1827 // 1828 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1829 Node* mem = in(MemNode::Memory); 1830 Node* value = can_see_stored_value(mem, phase); 1831 if (value && !phase->type(value)->higher_equal(_type)) 1832 return new (phase->C) AndINode(value, phase->intcon(0xFF)); 1833 // Identity call will handle the case where truncation is not needed. 1834 return LoadNode::Ideal(phase, can_reshape); 1835 } 1836 1837 const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1838 Node* mem = in(MemNode::Memory); 1839 Node* value = can_see_stored_value(mem,phase); 1840 if (value != NULL && value->is_Con() && 1841 !value->bottom_type()->higher_equal(_type)) { 1842 // If the input to the store does not fit with the load's result type, 1843 // it must be truncated. We can't delay until Ideal call since 1844 // a singleton Value is needed for split_thru_phi optimization. 1845 int con = value->get_int(); 1846 return TypeInt::make(con & 0xFF); 1847 } 1848 return LoadNode::Value(phase); 1849 } 1850 1851 //--------------------------LoadUSNode::Ideal------------------------------------- 1852 // 1853 // If the previous store is to the same address as this load, 1854 // and the value stored was larger than a char, replace this load 1855 // with the value stored truncated to a char. If no truncation is 1856 // needed, the replacement is done in LoadNode::Identity(). 1857 // 1858 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1859 Node* mem = in(MemNode::Memory); 1860 Node* value = can_see_stored_value(mem,phase); 1861 if( value && !phase->type(value)->higher_equal( _type ) ) 1862 return new (phase->C) AndINode(value,phase->intcon(0xFFFF)); 1863 // Identity call will handle the case where truncation is not needed. 1864 return LoadNode::Ideal(phase, can_reshape); 1865 } 1866 1867 const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1868 Node* mem = in(MemNode::Memory); 1869 Node* value = can_see_stored_value(mem,phase); 1870 if (value != NULL && value->is_Con() && 1871 !value->bottom_type()->higher_equal(_type)) { 1872 // If the input to the store does not fit with the load's result type, 1873 // it must be truncated. We can't delay until Ideal call since 1874 // a singleton Value is needed for split_thru_phi optimization. 1875 int con = value->get_int(); 1876 return TypeInt::make(con & 0xFFFF); 1877 } 1878 return LoadNode::Value(phase); 1879 } 1880 1881 //--------------------------LoadSNode::Ideal-------------------------------------- 1882 // 1883 // If the previous store is to the same address as this load, 1884 // and the value stored was larger than a short, replace this load 1885 // with the value stored truncated to a short. If no truncation is 1886 // needed, the replacement is done in LoadNode::Identity(). 1887 // 1888 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1889 Node* mem = in(MemNode::Memory); 1890 Node* value = can_see_stored_value(mem,phase); 1891 if( value && !phase->type(value)->higher_equal( _type ) ) { 1892 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) ); 1893 return new (phase->C) RShiftINode(result, phase->intcon(16)); 1894 } 1895 // Identity call will handle the case where truncation is not needed. 1896 return LoadNode::Ideal(phase, can_reshape); 1897 } 1898 1899 const Type* LoadSNode::Value(PhaseTransform *phase) const { 1900 Node* mem = in(MemNode::Memory); 1901 Node* value = can_see_stored_value(mem,phase); 1902 if (value != NULL && value->is_Con() && 1903 !value->bottom_type()->higher_equal(_type)) { 1904 // If the input to the store does not fit with the load's result type, 1905 // it must be truncated. We can't delay until Ideal call since 1906 // a singleton Value is needed for split_thru_phi optimization. 1907 int con = value->get_int(); 1908 return TypeInt::make((con << 16) >> 16); 1909 } 1910 return LoadNode::Value(phase); 1911 } 1912 1913 //============================================================================= 1914 //----------------------------LoadKlassNode::make------------------------------ 1915 // Polymorphic factory method: 1916 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1917 Compile* C = gvn.C; 1918 Node *ctl = NULL; 1919 // sanity check the alias category against the created node type 1920 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 1921 assert(adr_type != NULL, "expecting TypeKlassPtr"); 1922 #ifdef _LP64 1923 if (adr_type->is_ptr_to_narrowklass()) { 1924 assert(UseCompressedKlassPointers, "no compressed klasses"); 1925 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass())); 1926 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 1927 } 1928 #endif 1929 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1930 return new (C) LoadKlassNode(ctl, mem, adr, at, tk); 1931 } 1932 1933 //------------------------------Value------------------------------------------ 1934 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1935 return klass_value_common(phase); 1936 } 1937 1938 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1939 // Either input is TOP ==> the result is TOP 1940 const Type *t1 = phase->type( in(MemNode::Memory) ); 1941 if (t1 == Type::TOP) return Type::TOP; 1942 Node *adr = in(MemNode::Address); 1943 const Type *t2 = phase->type( adr ); 1944 if (t2 == Type::TOP) return Type::TOP; 1945 const TypePtr *tp = t2->is_ptr(); 1946 if (TypePtr::above_centerline(tp->ptr()) || 1947 tp->ptr() == TypePtr::Null) return Type::TOP; 1948 1949 // Return a more precise klass, if possible 1950 const TypeInstPtr *tinst = tp->isa_instptr(); 1951 if (tinst != NULL) { 1952 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1953 int offset = tinst->offset(); 1954 if (ik == phase->C->env()->Class_klass() 1955 && (offset == java_lang_Class::klass_offset_in_bytes() || 1956 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1957 // We are loading a special hidden field from a Class mirror object, 1958 // the field which points to the VM's Klass metaobject. 1959 ciType* t = tinst->java_mirror_type(); 1960 // java_mirror_type returns non-null for compile-time Class constants. 1961 if (t != NULL) { 1962 // constant oop => constant klass 1963 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1964 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1965 } 1966 if (!t->is_klass()) { 1967 // a primitive Class (e.g., int.class) has NULL for a klass field 1968 return TypePtr::NULL_PTR; 1969 } 1970 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1971 return TypeKlassPtr::make(t->as_klass()); 1972 } 1973 // non-constant mirror, so we can't tell what's going on 1974 } 1975 if( !ik->is_loaded() ) 1976 return _type; // Bail out if not loaded 1977 if (offset == oopDesc::klass_offset_in_bytes()) { 1978 if (tinst->klass_is_exact()) { 1979 return TypeKlassPtr::make(ik); 1980 } 1981 // See if we can become precise: no subklasses and no interface 1982 // (Note: We need to support verified interfaces.) 1983 if (!ik->is_interface() && !ik->has_subklass()) { 1984 //assert(!UseExactTypes, "this code should be useless with exact types"); 1985 // Add a dependence; if any subclass added we need to recompile 1986 if (!ik->is_final()) { 1987 // %%% should use stronger assert_unique_concrete_subtype instead 1988 phase->C->dependencies()->assert_leaf_type(ik); 1989 } 1990 // Return precise klass 1991 return TypeKlassPtr::make(ik); 1992 } 1993 1994 // Return root of possible klass 1995 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1996 } 1997 } 1998 1999 // Check for loading klass from an array 2000 const TypeAryPtr *tary = tp->isa_aryptr(); 2001 if( tary != NULL ) { 2002 ciKlass *tary_klass = tary->klass(); 2003 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2004 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2005 if (tary->klass_is_exact()) { 2006 return TypeKlassPtr::make(tary_klass); 2007 } 2008 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2009 // If the klass is an object array, we defer the question to the 2010 // array component klass. 2011 if( ak->is_obj_array_klass() ) { 2012 assert( ak->is_loaded(), "" ); 2013 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2014 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2015 ciInstanceKlass* ik = base_k->as_instance_klass(); 2016 // See if we can become precise: no subklasses and no interface 2017 if (!ik->is_interface() && !ik->has_subklass()) { 2018 //assert(!UseExactTypes, "this code should be useless with exact types"); 2019 // Add a dependence; if any subclass added we need to recompile 2020 if (!ik->is_final()) { 2021 phase->C->dependencies()->assert_leaf_type(ik); 2022 } 2023 // Return precise array klass 2024 return TypeKlassPtr::make(ak); 2025 } 2026 } 2027 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2028 } else { // Found a type-array? 2029 //assert(!UseExactTypes, "this code should be useless with exact types"); 2030 assert( ak->is_type_array_klass(), "" ); 2031 return TypeKlassPtr::make(ak); // These are always precise 2032 } 2033 } 2034 } 2035 2036 // Check for loading klass from an array klass 2037 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2038 if (tkls != NULL && !StressReflectiveCode) { 2039 ciKlass* klass = tkls->klass(); 2040 if( !klass->is_loaded() ) 2041 return _type; // Bail out if not loaded 2042 if( klass->is_obj_array_klass() && 2043 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2044 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2045 // // Always returning precise element type is incorrect, 2046 // // e.g., element type could be object and array may contain strings 2047 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2048 2049 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2050 // according to the element type's subclassing. 2051 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2052 } 2053 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2054 tkls->offset() == in_bytes(Klass::super_offset())) { 2055 ciKlass* sup = klass->as_instance_klass()->super(); 2056 // The field is Klass::_super. Return its (constant) value. 2057 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2058 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2059 } 2060 } 2061 2062 // Bailout case 2063 return LoadNode::Value(phase); 2064 } 2065 2066 //------------------------------Identity--------------------------------------- 2067 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2068 // Also feed through the klass in Allocate(...klass...)._klass. 2069 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2070 return klass_identity_common(phase); 2071 } 2072 2073 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2074 Node* x = LoadNode::Identity(phase); 2075 if (x != this) return x; 2076 2077 // Take apart the address into an oop and and offset. 2078 // Return 'this' if we cannot. 2079 Node* adr = in(MemNode::Address); 2080 intptr_t offset = 0; 2081 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2082 if (base == NULL) return this; 2083 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2084 if (toop == NULL) return this; 2085 2086 // We can fetch the klass directly through an AllocateNode. 2087 // This works even if the klass is not constant (clone or newArray). 2088 if (offset == oopDesc::klass_offset_in_bytes()) { 2089 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2090 if (allocated_klass != NULL) { 2091 return allocated_klass; 2092 } 2093 } 2094 2095 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2096 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass. 2097 // See inline_native_Class_query for occurrences of these patterns. 2098 // Java Example: x.getClass().isAssignableFrom(y) 2099 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 2100 // 2101 // This improves reflective code, often making the Class 2102 // mirror go completely dead. (Current exception: Class 2103 // mirrors may appear in debug info, but we could clean them out by 2104 // introducing a new debug info operator for Klass*.java_mirror). 2105 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2106 && (offset == java_lang_Class::klass_offset_in_bytes() || 2107 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2108 // We are loading a special hidden field from a Class mirror, 2109 // the field which points to its Klass or ArrayKlass metaobject. 2110 if (base->is_Load()) { 2111 Node* adr2 = base->in(MemNode::Address); 2112 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2113 if (tkls != NULL && !tkls->empty() 2114 && (tkls->klass()->is_instance_klass() || 2115 tkls->klass()->is_array_klass()) 2116 && adr2->is_AddP() 2117 ) { 2118 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2119 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2120 mirror_field = in_bytes(ArrayKlass::component_mirror_offset()); 2121 } 2122 if (tkls->offset() == mirror_field) { 2123 return adr2->in(AddPNode::Base); 2124 } 2125 } 2126 } 2127 } 2128 2129 return this; 2130 } 2131 2132 2133 //------------------------------Value------------------------------------------ 2134 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2135 const Type *t = klass_value_common(phase); 2136 if (t == Type::TOP) 2137 return t; 2138 2139 return t->make_narrowklass(); 2140 } 2141 2142 //------------------------------Identity--------------------------------------- 2143 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2144 // Also feed through the klass in Allocate(...klass...)._klass. 2145 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2146 Node *x = klass_identity_common(phase); 2147 2148 const Type *t = phase->type( x ); 2149 if( t == Type::TOP ) return x; 2150 if( t->isa_narrowklass()) return x; 2151 assert (!t->isa_narrowoop(), "no narrow oop here"); 2152 2153 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass())); 2154 } 2155 2156 //------------------------------Value----------------------------------------- 2157 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2158 // Either input is TOP ==> the result is TOP 2159 const Type *t1 = phase->type( in(MemNode::Memory) ); 2160 if( t1 == Type::TOP ) return Type::TOP; 2161 Node *adr = in(MemNode::Address); 2162 const Type *t2 = phase->type( adr ); 2163 if( t2 == Type::TOP ) return Type::TOP; 2164 const TypePtr *tp = t2->is_ptr(); 2165 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2166 const TypeAryPtr *tap = tp->isa_aryptr(); 2167 if( !tap ) return _type; 2168 return tap->size(); 2169 } 2170 2171 //-------------------------------Ideal--------------------------------------- 2172 // Feed through the length in AllocateArray(...length...)._length. 2173 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2174 Node* p = MemNode::Ideal_common(phase, can_reshape); 2175 if (p) return (p == NodeSentinel) ? NULL : p; 2176 2177 // Take apart the address into an oop and and offset. 2178 // Return 'this' if we cannot. 2179 Node* adr = in(MemNode::Address); 2180 intptr_t offset = 0; 2181 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2182 if (base == NULL) return NULL; 2183 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2184 if (tary == NULL) return NULL; 2185 2186 // We can fetch the length directly through an AllocateArrayNode. 2187 // This works even if the length is not constant (clone or newArray). 2188 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2189 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2190 if (alloc != NULL) { 2191 Node* allocated_length = alloc->Ideal_length(); 2192 Node* len = alloc->make_ideal_length(tary, phase); 2193 if (allocated_length != len) { 2194 // New CastII improves on this. 2195 return len; 2196 } 2197 } 2198 } 2199 2200 return NULL; 2201 } 2202 2203 //------------------------------Identity--------------------------------------- 2204 // Feed through the length in AllocateArray(...length...)._length. 2205 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2206 Node* x = LoadINode::Identity(phase); 2207 if (x != this) return x; 2208 2209 // Take apart the address into an oop and and offset. 2210 // Return 'this' if we cannot. 2211 Node* adr = in(MemNode::Address); 2212 intptr_t offset = 0; 2213 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2214 if (base == NULL) return this; 2215 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2216 if (tary == NULL) return this; 2217 2218 // We can fetch the length directly through an AllocateArrayNode. 2219 // This works even if the length is not constant (clone or newArray). 2220 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2221 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2222 if (alloc != NULL) { 2223 Node* allocated_length = alloc->Ideal_length(); 2224 // Do not allow make_ideal_length to allocate a CastII node. 2225 Node* len = alloc->make_ideal_length(tary, phase, false); 2226 if (allocated_length == len) { 2227 // Return allocated_length only if it would not be improved by a CastII. 2228 return allocated_length; 2229 } 2230 } 2231 } 2232 2233 return this; 2234 2235 } 2236 2237 //============================================================================= 2238 //---------------------------StoreNode::make----------------------------------- 2239 // Polymorphic factory method: 2240 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 2241 Compile* C = gvn.C; 2242 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2243 ctl != NULL, "raw memory operations should have control edge"); 2244 2245 switch (bt) { 2246 case T_BOOLEAN: 2247 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val); 2248 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val); 2249 case T_CHAR: 2250 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val); 2251 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val); 2252 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val); 2253 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val); 2254 case T_METADATA: 2255 case T_ADDRESS: 2256 case T_OBJECT: 2257 #ifdef _LP64 2258 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2259 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2260 return new (C) StoreNNode(ctl, mem, adr, adr_type, val); 2261 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2262 (UseCompressedKlassPointers && val->bottom_type()->isa_klassptr() && 2263 adr->bottom_type()->isa_rawptr())) { 2264 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2265 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val); 2266 } 2267 #endif 2268 { 2269 return new (C) StorePNode(ctl, mem, adr, adr_type, val); 2270 } 2271 } 2272 ShouldNotReachHere(); 2273 return (StoreNode*)NULL; 2274 } 2275 2276 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2277 bool require_atomic = true; 2278 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2279 } 2280 2281 2282 //--------------------------bottom_type---------------------------------------- 2283 const Type *StoreNode::bottom_type() const { 2284 return Type::MEMORY; 2285 } 2286 2287 //------------------------------hash------------------------------------------- 2288 uint StoreNode::hash() const { 2289 // unroll addition of interesting fields 2290 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2291 2292 // Since they are not commoned, do not hash them: 2293 return NO_HASH; 2294 } 2295 2296 //------------------------------Ideal------------------------------------------ 2297 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2298 // When a store immediately follows a relevant allocation/initialization, 2299 // try to capture it into the initialization, or hoist it above. 2300 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2301 Node* p = MemNode::Ideal_common(phase, can_reshape); 2302 if (p) return (p == NodeSentinel) ? NULL : p; 2303 2304 Node* mem = in(MemNode::Memory); 2305 Node* address = in(MemNode::Address); 2306 2307 // Back-to-back stores to same address? Fold em up. Generally 2308 // unsafe if I have intervening uses... Also disallowed for StoreCM 2309 // since they must follow each StoreP operation. Redundant StoreCMs 2310 // are eliminated just before matching in final_graph_reshape. 2311 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && 2312 mem->Opcode() != Op_StoreCM) { 2313 // Looking at a dead closed cycle of memory? 2314 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2315 2316 assert(Opcode() == mem->Opcode() || 2317 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2318 "no mismatched stores, except on raw memory"); 2319 2320 if (mem->outcnt() == 1 && // check for intervening uses 2321 mem->as_Store()->memory_size() <= this->memory_size()) { 2322 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2323 // For example, 'mem' might be the final state at a conditional return. 2324 // Or, 'mem' might be used by some node which is live at the same time 2325 // 'this' is live, which might be unschedulable. So, require exactly 2326 // ONE user, the 'this' store, until such time as we clone 'mem' for 2327 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2328 if (can_reshape) { // (%%% is this an anachronism?) 2329 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2330 phase->is_IterGVN()); 2331 } else { 2332 // It's OK to do this in the parser, since DU info is always accurate, 2333 // and the parser always refers to nodes via SafePointNode maps. 2334 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2335 } 2336 return this; 2337 } 2338 } 2339 2340 // Capture an unaliased, unconditional, simple store into an initializer. 2341 // Or, if it is independent of the allocation, hoist it above the allocation. 2342 if (ReduceFieldZeroing && /*can_reshape &&*/ 2343 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2344 InitializeNode* init = mem->in(0)->as_Initialize(); 2345 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2346 if (offset > 0) { 2347 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2348 // If the InitializeNode captured me, it made a raw copy of me, 2349 // and I need to disappear. 2350 if (moved != NULL) { 2351 // %%% hack to ensure that Ideal returns a new node: 2352 mem = MergeMemNode::make(phase->C, mem); 2353 return mem; // fold me away 2354 } 2355 } 2356 } 2357 2358 return NULL; // No further progress 2359 } 2360 2361 //------------------------------Value----------------------------------------- 2362 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2363 // Either input is TOP ==> the result is TOP 2364 const Type *t1 = phase->type( in(MemNode::Memory) ); 2365 if( t1 == Type::TOP ) return Type::TOP; 2366 const Type *t2 = phase->type( in(MemNode::Address) ); 2367 if( t2 == Type::TOP ) return Type::TOP; 2368 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2369 if( t3 == Type::TOP ) return Type::TOP; 2370 return Type::MEMORY; 2371 } 2372 2373 //------------------------------Identity--------------------------------------- 2374 // Remove redundant stores: 2375 // Store(m, p, Load(m, p)) changes to m. 2376 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2377 Node *StoreNode::Identity( PhaseTransform *phase ) { 2378 Node* mem = in(MemNode::Memory); 2379 Node* adr = in(MemNode::Address); 2380 Node* val = in(MemNode::ValueIn); 2381 2382 // Load then Store? Then the Store is useless 2383 if (val->is_Load() && 2384 val->in(MemNode::Address)->eqv_uncast(adr) && 2385 val->in(MemNode::Memory )->eqv_uncast(mem) && 2386 val->as_Load()->store_Opcode() == Opcode()) { 2387 return mem; 2388 } 2389 2390 // Two stores in a row of the same value? 2391 if (mem->is_Store() && 2392 mem->in(MemNode::Address)->eqv_uncast(adr) && 2393 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2394 mem->Opcode() == Opcode()) { 2395 return mem; 2396 } 2397 2398 // Store of zero anywhere into a freshly-allocated object? 2399 // Then the store is useless. 2400 // (It must already have been captured by the InitializeNode.) 2401 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2402 // a newly allocated object is already all-zeroes everywhere 2403 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2404 return mem; 2405 } 2406 2407 // the store may also apply to zero-bits in an earlier object 2408 Node* prev_mem = find_previous_store(phase); 2409 // Steps (a), (b): Walk past independent stores to find an exact match. 2410 if (prev_mem != NULL) { 2411 Node* prev_val = can_see_stored_value(prev_mem, phase); 2412 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2413 // prev_val and val might differ by a cast; it would be good 2414 // to keep the more informative of the two. 2415 return mem; 2416 } 2417 } 2418 } 2419 2420 return this; 2421 } 2422 2423 //------------------------------match_edge------------------------------------- 2424 // Do we Match on this edge index or not? Match only memory & value 2425 uint StoreNode::match_edge(uint idx) const { 2426 return idx == MemNode::Address || idx == MemNode::ValueIn; 2427 } 2428 2429 //------------------------------cmp-------------------------------------------- 2430 // Do not common stores up together. They generally have to be split 2431 // back up anyways, so do not bother. 2432 uint StoreNode::cmp( const Node &n ) const { 2433 return (&n == this); // Always fail except on self 2434 } 2435 2436 //------------------------------Ideal_masked_input----------------------------- 2437 // Check for a useless mask before a partial-word store 2438 // (StoreB ... (AndI valIn conIa) ) 2439 // If (conIa & mask == mask) this simplifies to 2440 // (StoreB ... (valIn) ) 2441 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2442 Node *val = in(MemNode::ValueIn); 2443 if( val->Opcode() == Op_AndI ) { 2444 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2445 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2446 set_req(MemNode::ValueIn, val->in(1)); 2447 return this; 2448 } 2449 } 2450 return NULL; 2451 } 2452 2453 2454 //------------------------------Ideal_sign_extended_input---------------------- 2455 // Check for useless sign-extension before a partial-word store 2456 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2457 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2458 // (StoreB ... (valIn) ) 2459 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2460 Node *val = in(MemNode::ValueIn); 2461 if( val->Opcode() == Op_RShiftI ) { 2462 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2463 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2464 Node *shl = val->in(1); 2465 if( shl->Opcode() == Op_LShiftI ) { 2466 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2467 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2468 set_req(MemNode::ValueIn, shl->in(1)); 2469 return this; 2470 } 2471 } 2472 } 2473 } 2474 return NULL; 2475 } 2476 2477 //------------------------------value_never_loaded----------------------------------- 2478 // Determine whether there are any possible loads of the value stored. 2479 // For simplicity, we actually check if there are any loads from the 2480 // address stored to, not just for loads of the value stored by this node. 2481 // 2482 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2483 Node *adr = in(Address); 2484 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2485 if (adr_oop == NULL) 2486 return false; 2487 if (!adr_oop->is_known_instance_field()) 2488 return false; // if not a distinct instance, there may be aliases of the address 2489 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2490 Node *use = adr->fast_out(i); 2491 int opc = use->Opcode(); 2492 if (use->is_Load() || use->is_LoadStore()) { 2493 return false; 2494 } 2495 } 2496 return true; 2497 } 2498 2499 //============================================================================= 2500 //------------------------------Ideal------------------------------------------ 2501 // If the store is from an AND mask that leaves the low bits untouched, then 2502 // we can skip the AND operation. If the store is from a sign-extension 2503 // (a left shift, then right shift) we can skip both. 2504 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2505 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2506 if( progress != NULL ) return progress; 2507 2508 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2509 if( progress != NULL ) return progress; 2510 2511 // Finally check the default case 2512 return StoreNode::Ideal(phase, can_reshape); 2513 } 2514 2515 //============================================================================= 2516 //------------------------------Ideal------------------------------------------ 2517 // If the store is from an AND mask that leaves the low bits untouched, then 2518 // we can skip the AND operation 2519 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2520 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2521 if( progress != NULL ) return progress; 2522 2523 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2524 if( progress != NULL ) return progress; 2525 2526 // Finally check the default case 2527 return StoreNode::Ideal(phase, can_reshape); 2528 } 2529 2530 //============================================================================= 2531 //------------------------------Identity--------------------------------------- 2532 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2533 // No need to card mark when storing a null ptr 2534 Node* my_store = in(MemNode::OopStore); 2535 if (my_store->is_Store()) { 2536 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2537 if( t1 == TypePtr::NULL_PTR ) { 2538 return in(MemNode::Memory); 2539 } 2540 } 2541 return this; 2542 } 2543 2544 //============================================================================= 2545 //------------------------------Ideal--------------------------------------- 2546 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2547 Node* progress = StoreNode::Ideal(phase, can_reshape); 2548 if (progress != NULL) return progress; 2549 2550 Node* my_store = in(MemNode::OopStore); 2551 if (my_store->is_MergeMem()) { 2552 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2553 set_req(MemNode::OopStore, mem); 2554 return this; 2555 } 2556 2557 return NULL; 2558 } 2559 2560 //------------------------------Value----------------------------------------- 2561 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2562 // Either input is TOP ==> the result is TOP 2563 const Type *t = phase->type( in(MemNode::Memory) ); 2564 if( t == Type::TOP ) return Type::TOP; 2565 t = phase->type( in(MemNode::Address) ); 2566 if( t == Type::TOP ) return Type::TOP; 2567 t = phase->type( in(MemNode::ValueIn) ); 2568 if( t == Type::TOP ) return Type::TOP; 2569 // If extra input is TOP ==> the result is TOP 2570 t = phase->type( in(MemNode::OopStore) ); 2571 if( t == Type::TOP ) return Type::TOP; 2572 2573 return StoreNode::Value( phase ); 2574 } 2575 2576 2577 //============================================================================= 2578 //----------------------------------SCMemProjNode------------------------------ 2579 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2580 { 2581 return bottom_type(); 2582 } 2583 2584 //============================================================================= 2585 //----------------------------------LoadStoreNode------------------------------ 2586 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2587 : Node(required), 2588 _type(rt), 2589 _adr_type(at) 2590 { 2591 init_req(MemNode::Control, c ); 2592 init_req(MemNode::Memory , mem); 2593 init_req(MemNode::Address, adr); 2594 init_req(MemNode::ValueIn, val); 2595 init_class_id(Class_LoadStore); 2596 } 2597 2598 uint LoadStoreNode::ideal_reg() const { 2599 return _type->ideal_reg(); 2600 } 2601 2602 bool LoadStoreNode::result_not_used() const { 2603 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2604 Node *x = fast_out(i); 2605 if (x->Opcode() == Op_SCMemProj) continue; 2606 return false; 2607 } 2608 return true; 2609 } 2610 2611 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2612 2613 //============================================================================= 2614 //----------------------------------LoadStoreConditionalNode-------------------- 2615 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2616 init_req(ExpectedIn, ex ); 2617 } 2618 2619 //============================================================================= 2620 //-------------------------------adr_type-------------------------------------- 2621 // Do we Match on this edge index or not? Do not match memory 2622 const TypePtr* ClearArrayNode::adr_type() const { 2623 Node *adr = in(3); 2624 return MemNode::calculate_adr_type(adr->bottom_type()); 2625 } 2626 2627 //------------------------------match_edge------------------------------------- 2628 // Do we Match on this edge index or not? Do not match memory 2629 uint ClearArrayNode::match_edge(uint idx) const { 2630 return idx > 1; 2631 } 2632 2633 //------------------------------Identity--------------------------------------- 2634 // Clearing a zero length array does nothing 2635 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2636 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2637 } 2638 2639 //------------------------------Idealize--------------------------------------- 2640 // Clearing a short array is faster with stores 2641 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2642 const int unit = BytesPerLong; 2643 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2644 if (!t) return NULL; 2645 if (!t->is_con()) return NULL; 2646 intptr_t raw_count = t->get_con(); 2647 intptr_t size = raw_count; 2648 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2649 // Clearing nothing uses the Identity call. 2650 // Negative clears are possible on dead ClearArrays 2651 // (see jck test stmt114.stmt11402.val). 2652 if (size <= 0 || size % unit != 0) return NULL; 2653 intptr_t count = size / unit; 2654 // Length too long; use fast hardware clear 2655 if (size > Matcher::init_array_short_size) return NULL; 2656 Node *mem = in(1); 2657 if( phase->type(mem)==Type::TOP ) return NULL; 2658 Node *adr = in(3); 2659 const Type* at = phase->type(adr); 2660 if( at==Type::TOP ) return NULL; 2661 const TypePtr* atp = at->isa_ptr(); 2662 // adjust atp to be the correct array element address type 2663 if (atp == NULL) atp = TypePtr::BOTTOM; 2664 else atp = atp->add_offset(Type::OffsetBot); 2665 // Get base for derived pointer purposes 2666 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2667 Node *base = adr->in(1); 2668 2669 Node *zero = phase->makecon(TypeLong::ZERO); 2670 Node *off = phase->MakeConX(BytesPerLong); 2671 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero); 2672 count--; 2673 while( count-- ) { 2674 mem = phase->transform(mem); 2675 adr = phase->transform(new (phase->C) AddPNode(base,adr,off)); 2676 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero); 2677 } 2678 return mem; 2679 } 2680 2681 //----------------------------step_through---------------------------------- 2682 // Return allocation input memory edge if it is different instance 2683 // or itself if it is the one we are looking for. 2684 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2685 Node* n = *np; 2686 assert(n->is_ClearArray(), "sanity"); 2687 intptr_t offset; 2688 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2689 // This method is called only before Allocate nodes are expanded during 2690 // macro nodes expansion. Before that ClearArray nodes are only generated 2691 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2692 assert(alloc != NULL, "should have allocation"); 2693 if (alloc->_idx == instance_id) { 2694 // Can not bypass initialization of the instance we are looking for. 2695 return false; 2696 } 2697 // Otherwise skip it. 2698 InitializeNode* init = alloc->initialization(); 2699 if (init != NULL) 2700 *np = init->in(TypeFunc::Memory); 2701 else 2702 *np = alloc->in(TypeFunc::Memory); 2703 return true; 2704 } 2705 2706 //----------------------------clear_memory------------------------------------- 2707 // Generate code to initialize object storage to zero. 2708 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2709 intptr_t start_offset, 2710 Node* end_offset, 2711 PhaseGVN* phase) { 2712 Compile* C = phase->C; 2713 intptr_t offset = start_offset; 2714 2715 int unit = BytesPerLong; 2716 if ((offset % unit) != 0) { 2717 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset)); 2718 adr = phase->transform(adr); 2719 const TypePtr* atp = TypeRawPtr::BOTTOM; 2720 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2721 mem = phase->transform(mem); 2722 offset += BytesPerInt; 2723 } 2724 assert((offset % unit) == 0, ""); 2725 2726 // Initialize the remaining stuff, if any, with a ClearArray. 2727 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2728 } 2729 2730 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2731 Node* start_offset, 2732 Node* end_offset, 2733 PhaseGVN* phase) { 2734 if (start_offset == end_offset) { 2735 // nothing to do 2736 return mem; 2737 } 2738 2739 Compile* C = phase->C; 2740 int unit = BytesPerLong; 2741 Node* zbase = start_offset; 2742 Node* zend = end_offset; 2743 2744 // Scale to the unit required by the CPU: 2745 if (!Matcher::init_array_count_is_in_bytes) { 2746 Node* shift = phase->intcon(exact_log2(unit)); 2747 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) ); 2748 zend = phase->transform( new(C) URShiftXNode(zend, shift) ); 2749 } 2750 2751 // Bulk clear double-words 2752 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) ); 2753 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) ); 2754 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr); 2755 return phase->transform(mem); 2756 } 2757 2758 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2759 intptr_t start_offset, 2760 intptr_t end_offset, 2761 PhaseGVN* phase) { 2762 if (start_offset == end_offset) { 2763 // nothing to do 2764 return mem; 2765 } 2766 2767 Compile* C = phase->C; 2768 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2769 intptr_t done_offset = end_offset; 2770 if ((done_offset % BytesPerLong) != 0) { 2771 done_offset -= BytesPerInt; 2772 } 2773 if (done_offset > start_offset) { 2774 mem = clear_memory(ctl, mem, dest, 2775 start_offset, phase->MakeConX(done_offset), phase); 2776 } 2777 if (done_offset < end_offset) { // emit the final 32-bit store 2778 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2779 adr = phase->transform(adr); 2780 const TypePtr* atp = TypeRawPtr::BOTTOM; 2781 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2782 mem = phase->transform(mem); 2783 done_offset += BytesPerInt; 2784 } 2785 assert(done_offset == end_offset, ""); 2786 return mem; 2787 } 2788 2789 //============================================================================= 2790 // Do not match memory edge. 2791 uint StrIntrinsicNode::match_edge(uint idx) const { 2792 return idx == 2 || idx == 3; 2793 } 2794 2795 //------------------------------Ideal------------------------------------------ 2796 // Return a node which is more "ideal" than the current node. Strip out 2797 // control copies 2798 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2799 if (remove_dead_region(phase, can_reshape)) return this; 2800 // Don't bother trying to transform a dead node 2801 if (in(0) && in(0)->is_top()) return NULL; 2802 2803 if (can_reshape) { 2804 Node* mem = phase->transform(in(MemNode::Memory)); 2805 // If transformed to a MergeMem, get the desired slice 2806 uint alias_idx = phase->C->get_alias_index(adr_type()); 2807 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; 2808 if (mem != in(MemNode::Memory)) { 2809 set_req(MemNode::Memory, mem); 2810 return this; 2811 } 2812 } 2813 return NULL; 2814 } 2815 2816 //------------------------------Value------------------------------------------ 2817 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { 2818 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; 2819 return bottom_type(); 2820 } 2821 2822 //============================================================================= 2823 //------------------------------match_edge------------------------------------- 2824 // Do not match memory edge 2825 uint EncodeISOArrayNode::match_edge(uint idx) const { 2826 return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len) 2827 } 2828 2829 //------------------------------Ideal------------------------------------------ 2830 // Return a node which is more "ideal" than the current node. Strip out 2831 // control copies 2832 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2833 return remove_dead_region(phase, can_reshape) ? this : NULL; 2834 } 2835 2836 //------------------------------Value------------------------------------------ 2837 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const { 2838 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; 2839 return bottom_type(); 2840 } 2841 2842 //============================================================================= 2843 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2844 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2845 _adr_type(C->get_adr_type(alias_idx)) 2846 { 2847 init_class_id(Class_MemBar); 2848 Node* top = C->top(); 2849 init_req(TypeFunc::I_O,top); 2850 init_req(TypeFunc::FramePtr,top); 2851 init_req(TypeFunc::ReturnAdr,top); 2852 if (precedent != NULL) 2853 init_req(TypeFunc::Parms, precedent); 2854 } 2855 2856 //------------------------------cmp-------------------------------------------- 2857 uint MemBarNode::hash() const { return NO_HASH; } 2858 uint MemBarNode::cmp( const Node &n ) const { 2859 return (&n == this); // Always fail except on self 2860 } 2861 2862 //------------------------------make------------------------------------------- 2863 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2864 switch (opcode) { 2865 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn); 2866 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn); 2867 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn); 2868 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn); 2869 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn); 2870 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn); 2871 case Op_Initialize: return new(C) InitializeNode(C, atp, pn); 2872 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn); 2873 default: ShouldNotReachHere(); return NULL; 2874 } 2875 } 2876 2877 //------------------------------Ideal------------------------------------------ 2878 // Return a node which is more "ideal" than the current node. Strip out 2879 // control copies 2880 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2881 if (remove_dead_region(phase, can_reshape)) return this; 2882 // Don't bother trying to transform a dead node 2883 if (in(0) && in(0)->is_top()) return NULL; 2884 2885 // Eliminate volatile MemBars for scalar replaced objects. 2886 if (can_reshape && req() == (Precedent+1) && 2887 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { 2888 // Volatile field loads and stores. 2889 Node* my_mem = in(MemBarNode::Precedent); 2890 if (my_mem != NULL && my_mem->is_Mem()) { 2891 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2892 // Check for scalar replaced object reference. 2893 if( t_oop != NULL && t_oop->is_known_instance_field() && 2894 t_oop->offset() != Type::OffsetBot && 2895 t_oop->offset() != Type::OffsetTop) { 2896 // Replace MemBar projections by its inputs. 2897 PhaseIterGVN* igvn = phase->is_IterGVN(); 2898 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2899 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2900 // Must return either the original node (now dead) or a new node 2901 // (Do not return a top here, since that would break the uniqueness of top.) 2902 return new (phase->C) ConINode(TypeInt::ZERO); 2903 } 2904 } 2905 } 2906 return NULL; 2907 } 2908 2909 //------------------------------Value------------------------------------------ 2910 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2911 if( !in(0) ) return Type::TOP; 2912 if( phase->type(in(0)) == Type::TOP ) 2913 return Type::TOP; 2914 return TypeTuple::MEMBAR; 2915 } 2916 2917 //------------------------------match------------------------------------------ 2918 // Construct projections for memory. 2919 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2920 switch (proj->_con) { 2921 case TypeFunc::Control: 2922 case TypeFunc::Memory: 2923 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2924 } 2925 ShouldNotReachHere(); 2926 return NULL; 2927 } 2928 2929 //===========================InitializeNode==================================== 2930 // SUMMARY: 2931 // This node acts as a memory barrier on raw memory, after some raw stores. 2932 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2933 // The Initialize can 'capture' suitably constrained stores as raw inits. 2934 // It can coalesce related raw stores into larger units (called 'tiles'). 2935 // It can avoid zeroing new storage for memory units which have raw inits. 2936 // At macro-expansion, it is marked 'complete', and does not optimize further. 2937 // 2938 // EXAMPLE: 2939 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2940 // ctl = incoming control; mem* = incoming memory 2941 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2942 // First allocate uninitialized memory and fill in the header: 2943 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2944 // ctl := alloc.Control; mem* := alloc.Memory* 2945 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2946 // Then initialize to zero the non-header parts of the raw memory block: 2947 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2948 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2949 // After the initialize node executes, the object is ready for service: 2950 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2951 // Suppose its body is immediately initialized as {1,2}: 2952 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2953 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2954 // mem.SLICE(#short[*]) := store2 2955 // 2956 // DETAILS: 2957 // An InitializeNode collects and isolates object initialization after 2958 // an AllocateNode and before the next possible safepoint. As a 2959 // memory barrier (MemBarNode), it keeps critical stores from drifting 2960 // down past any safepoint or any publication of the allocation. 2961 // Before this barrier, a newly-allocated object may have uninitialized bits. 2962 // After this barrier, it may be treated as a real oop, and GC is allowed. 2963 // 2964 // The semantics of the InitializeNode include an implicit zeroing of 2965 // the new object from object header to the end of the object. 2966 // (The object header and end are determined by the AllocateNode.) 2967 // 2968 // Certain stores may be added as direct inputs to the InitializeNode. 2969 // These stores must update raw memory, and they must be to addresses 2970 // derived from the raw address produced by AllocateNode, and with 2971 // a constant offset. They must be ordered by increasing offset. 2972 // The first one is at in(RawStores), the last at in(req()-1). 2973 // Unlike most memory operations, they are not linked in a chain, 2974 // but are displayed in parallel as users of the rawmem output of 2975 // the allocation. 2976 // 2977 // (See comments in InitializeNode::capture_store, which continue 2978 // the example given above.) 2979 // 2980 // When the associated Allocate is macro-expanded, the InitializeNode 2981 // may be rewritten to optimize collected stores. A ClearArrayNode 2982 // may also be created at that point to represent any required zeroing. 2983 // The InitializeNode is then marked 'complete', prohibiting further 2984 // capturing of nearby memory operations. 2985 // 2986 // During macro-expansion, all captured initializations which store 2987 // constant values of 32 bits or smaller are coalesced (if advantageous) 2988 // into larger 'tiles' 32 or 64 bits. This allows an object to be 2989 // initialized in fewer memory operations. Memory words which are 2990 // covered by neither tiles nor non-constant stores are pre-zeroed 2991 // by explicit stores of zero. (The code shape happens to do all 2992 // zeroing first, then all other stores, with both sequences occurring 2993 // in order of ascending offsets.) 2994 // 2995 // Alternatively, code may be inserted between an AllocateNode and its 2996 // InitializeNode, to perform arbitrary initialization of the new object. 2997 // E.g., the object copying intrinsics insert complex data transfers here. 2998 // The initialization must then be marked as 'complete' disable the 2999 // built-in zeroing semantics and the collection of initializing stores. 3000 // 3001 // While an InitializeNode is incomplete, reads from the memory state 3002 // produced by it are optimizable if they match the control edge and 3003 // new oop address associated with the allocation/initialization. 3004 // They return a stored value (if the offset matches) or else zero. 3005 // A write to the memory state, if it matches control and address, 3006 // and if it is to a constant offset, may be 'captured' by the 3007 // InitializeNode. It is cloned as a raw memory operation and rewired 3008 // inside the initialization, to the raw oop produced by the allocation. 3009 // Operations on addresses which are provably distinct (e.g., to 3010 // other AllocateNodes) are allowed to bypass the initialization. 3011 // 3012 // The effect of all this is to consolidate object initialization 3013 // (both arrays and non-arrays, both piecewise and bulk) into a 3014 // single location, where it can be optimized as a unit. 3015 // 3016 // Only stores with an offset less than TrackedInitializationLimit words 3017 // will be considered for capture by an InitializeNode. This puts a 3018 // reasonable limit on the complexity of optimized initializations. 3019 3020 //---------------------------InitializeNode------------------------------------ 3021 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3022 : _is_complete(Incomplete), _does_not_escape(false), 3023 MemBarNode(C, adr_type, rawoop) 3024 { 3025 init_class_id(Class_Initialize); 3026 3027 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3028 assert(in(RawAddress) == rawoop, "proper init"); 3029 // Note: allocation() can be NULL, for secondary initialization barriers 3030 } 3031 3032 // Since this node is not matched, it will be processed by the 3033 // register allocator. Declare that there are no constraints 3034 // on the allocation of the RawAddress edge. 3035 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3036 // This edge should be set to top, by the set_complete. But be conservative. 3037 if (idx == InitializeNode::RawAddress) 3038 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3039 return RegMask::Empty; 3040 } 3041 3042 Node* InitializeNode::memory(uint alias_idx) { 3043 Node* mem = in(Memory); 3044 if (mem->is_MergeMem()) { 3045 return mem->as_MergeMem()->memory_at(alias_idx); 3046 } else { 3047 // incoming raw memory is not split 3048 return mem; 3049 } 3050 } 3051 3052 bool InitializeNode::is_non_zero() { 3053 if (is_complete()) return false; 3054 remove_extra_zeroes(); 3055 return (req() > RawStores); 3056 } 3057 3058 void InitializeNode::set_complete(PhaseGVN* phase) { 3059 assert(!is_complete(), "caller responsibility"); 3060 _is_complete = Complete; 3061 3062 // After this node is complete, it contains a bunch of 3063 // raw-memory initializations. There is no need for 3064 // it to have anything to do with non-raw memory effects. 3065 // Therefore, tell all non-raw users to re-optimize themselves, 3066 // after skipping the memory effects of this initialization. 3067 PhaseIterGVN* igvn = phase->is_IterGVN(); 3068 if (igvn) igvn->add_users_to_worklist(this); 3069 } 3070 3071 // convenience function 3072 // return false if the init contains any stores already 3073 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3074 InitializeNode* init = initialization(); 3075 if (init == NULL || init->is_complete()) return false; 3076 init->remove_extra_zeroes(); 3077 // for now, if this allocation has already collected any inits, bail: 3078 if (init->is_non_zero()) return false; 3079 init->set_complete(phase); 3080 return true; 3081 } 3082 3083 void InitializeNode::remove_extra_zeroes() { 3084 if (req() == RawStores) return; 3085 Node* zmem = zero_memory(); 3086 uint fill = RawStores; 3087 for (uint i = fill; i < req(); i++) { 3088 Node* n = in(i); 3089 if (n->is_top() || n == zmem) continue; // skip 3090 if (fill < i) set_req(fill, n); // compact 3091 ++fill; 3092 } 3093 // delete any empty spaces created: 3094 while (fill < req()) { 3095 del_req(fill); 3096 } 3097 } 3098 3099 // Helper for remembering which stores go with which offsets. 3100 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3101 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3102 intptr_t offset = -1; 3103 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3104 phase, offset); 3105 if (base == NULL) return -1; // something is dead, 3106 if (offset < 0) return -1; // dead, dead 3107 return offset; 3108 } 3109 3110 // Helper for proving that an initialization expression is 3111 // "simple enough" to be folded into an object initialization. 3112 // Attempts to prove that a store's initial value 'n' can be captured 3113 // within the initialization without creating a vicious cycle, such as: 3114 // { Foo p = new Foo(); p.next = p; } 3115 // True for constants and parameters and small combinations thereof. 3116 bool InitializeNode::detect_init_independence(Node* n, 3117 bool st_is_pinned, 3118 int& count) { 3119 if (n == NULL) return true; // (can this really happen?) 3120 if (n->is_Proj()) n = n->in(0); 3121 if (n == this) return false; // found a cycle 3122 if (n->is_Con()) return true; 3123 if (n->is_Start()) return true; // params, etc., are OK 3124 if (n->is_Root()) return true; // even better 3125 3126 Node* ctl = n->in(0); 3127 if (ctl != NULL && !ctl->is_top()) { 3128 if (ctl->is_Proj()) ctl = ctl->in(0); 3129 if (ctl == this) return false; 3130 3131 // If we already know that the enclosing memory op is pinned right after 3132 // the init, then any control flow that the store has picked up 3133 // must have preceded the init, or else be equal to the init. 3134 // Even after loop optimizations (which might change control edges) 3135 // a store is never pinned *before* the availability of its inputs. 3136 if (!MemNode::all_controls_dominate(n, this)) 3137 return false; // failed to prove a good control 3138 3139 } 3140 3141 // Check data edges for possible dependencies on 'this'. 3142 if ((count += 1) > 20) return false; // complexity limit 3143 for (uint i = 1; i < n->req(); i++) { 3144 Node* m = n->in(i); 3145 if (m == NULL || m == n || m->is_top()) continue; 3146 uint first_i = n->find_edge(m); 3147 if (i != first_i) continue; // process duplicate edge just once 3148 if (!detect_init_independence(m, st_is_pinned, count)) { 3149 return false; 3150 } 3151 } 3152 3153 return true; 3154 } 3155 3156 // Here are all the checks a Store must pass before it can be moved into 3157 // an initialization. Returns zero if a check fails. 3158 // On success, returns the (constant) offset to which the store applies, 3159 // within the initialized memory. 3160 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3161 const int FAIL = 0; 3162 if (st->req() != MemNode::ValueIn + 1) 3163 return FAIL; // an inscrutable StoreNode (card mark?) 3164 Node* ctl = st->in(MemNode::Control); 3165 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3166 return FAIL; // must be unconditional after the initialization 3167 Node* mem = st->in(MemNode::Memory); 3168 if (!(mem->is_Proj() && mem->in(0) == this)) 3169 return FAIL; // must not be preceded by other stores 3170 Node* adr = st->in(MemNode::Address); 3171 intptr_t offset; 3172 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3173 if (alloc == NULL) 3174 return FAIL; // inscrutable address 3175 if (alloc != allocation()) 3176 return FAIL; // wrong allocation! (store needs to float up) 3177 Node* val = st->in(MemNode::ValueIn); 3178 int complexity_count = 0; 3179 if (!detect_init_independence(val, true, complexity_count)) 3180 return FAIL; // stored value must be 'simple enough' 3181 3182 // The Store can be captured only if nothing after the allocation 3183 // and before the Store is using the memory location that the store 3184 // overwrites. 3185 bool failed = false; 3186 // If is_complete_with_arraycopy() is true the shape of the graph is 3187 // well defined and is safe so no need for extra checks. 3188 if (!is_complete_with_arraycopy()) { 3189 // We are going to look at each use of the memory state following 3190 // the allocation to make sure nothing reads the memory that the 3191 // Store writes. 3192 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3193 int alias_idx = phase->C->get_alias_index(t_adr); 3194 ResourceMark rm; 3195 Unique_Node_List mems; 3196 mems.push(mem); 3197 Node* unique_merge = NULL; 3198 for (uint next = 0; next < mems.size(); ++next) { 3199 Node *m = mems.at(next); 3200 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3201 Node *n = m->fast_out(j); 3202 if (n->outcnt() == 0) { 3203 continue; 3204 } 3205 if (n == st) { 3206 continue; 3207 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3208 // If the control of this use is different from the control 3209 // of the Store which is right after the InitializeNode then 3210 // this node cannot be between the InitializeNode and the 3211 // Store. 3212 continue; 3213 } else if (n->is_MergeMem()) { 3214 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3215 // We can hit a MergeMemNode (that will likely go away 3216 // later) that is a direct use of the memory state 3217 // following the InitializeNode on the same slice as the 3218 // store node that we'd like to capture. We need to check 3219 // the uses of the MergeMemNode. 3220 mems.push(n); 3221 } 3222 } else if (n->is_Mem()) { 3223 Node* other_adr = n->in(MemNode::Address); 3224 if (other_adr == adr) { 3225 failed = true; 3226 break; 3227 } else { 3228 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3229 if (other_t_adr != NULL) { 3230 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3231 if (other_alias_idx == alias_idx) { 3232 // A load from the same memory slice as the store right 3233 // after the InitializeNode. We check the control of the 3234 // object/array that is loaded from. If it's the same as 3235 // the store control then we cannot capture the store. 3236 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3237 Node* base = other_adr; 3238 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); 3239 base = base->in(AddPNode::Base); 3240 if (base != NULL) { 3241 base = base->uncast(); 3242 if (base->is_Proj() && base->in(0) == alloc) { 3243 failed = true; 3244 break; 3245 } 3246 } 3247 } 3248 } 3249 } 3250 } else { 3251 failed = true; 3252 break; 3253 } 3254 } 3255 } 3256 } 3257 if (failed) { 3258 if (!can_reshape) { 3259 // We decided we couldn't capture the store during parsing. We 3260 // should try again during the next IGVN once the graph is 3261 // cleaner. 3262 phase->C->record_for_igvn(st); 3263 } 3264 return FAIL; 3265 } 3266 3267 return offset; // success 3268 } 3269 3270 // Find the captured store in(i) which corresponds to the range 3271 // [start..start+size) in the initialized object. 3272 // If there is one, return its index i. If there isn't, return the 3273 // negative of the index where it should be inserted. 3274 // Return 0 if the queried range overlaps an initialization boundary 3275 // or if dead code is encountered. 3276 // If size_in_bytes is zero, do not bother with overlap checks. 3277 int InitializeNode::captured_store_insertion_point(intptr_t start, 3278 int size_in_bytes, 3279 PhaseTransform* phase) { 3280 const int FAIL = 0, MAX_STORE = BytesPerLong; 3281 3282 if (is_complete()) 3283 return FAIL; // arraycopy got here first; punt 3284 3285 assert(allocation() != NULL, "must be present"); 3286 3287 // no negatives, no header fields: 3288 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3289 3290 // after a certain size, we bail out on tracking all the stores: 3291 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3292 if (start >= ti_limit) return FAIL; 3293 3294 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3295 if (i >= limit) return -(int)i; // not found; here is where to put it 3296 3297 Node* st = in(i); 3298 intptr_t st_off = get_store_offset(st, phase); 3299 if (st_off < 0) { 3300 if (st != zero_memory()) { 3301 return FAIL; // bail out if there is dead garbage 3302 } 3303 } else if (st_off > start) { 3304 // ...we are done, since stores are ordered 3305 if (st_off < start + size_in_bytes) { 3306 return FAIL; // the next store overlaps 3307 } 3308 return -(int)i; // not found; here is where to put it 3309 } else if (st_off < start) { 3310 if (size_in_bytes != 0 && 3311 start < st_off + MAX_STORE && 3312 start < st_off + st->as_Store()->memory_size()) { 3313 return FAIL; // the previous store overlaps 3314 } 3315 } else { 3316 if (size_in_bytes != 0 && 3317 st->as_Store()->memory_size() != size_in_bytes) { 3318 return FAIL; // mismatched store size 3319 } 3320 return i; 3321 } 3322 3323 ++i; 3324 } 3325 } 3326 3327 // Look for a captured store which initializes at the offset 'start' 3328 // with the given size. If there is no such store, and no other 3329 // initialization interferes, then return zero_memory (the memory 3330 // projection of the AllocateNode). 3331 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3332 PhaseTransform* phase) { 3333 assert(stores_are_sane(phase), ""); 3334 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3335 if (i == 0) { 3336 return NULL; // something is dead 3337 } else if (i < 0) { 3338 return zero_memory(); // just primordial zero bits here 3339 } else { 3340 Node* st = in(i); // here is the store at this position 3341 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3342 return st; 3343 } 3344 } 3345 3346 // Create, as a raw pointer, an address within my new object at 'offset'. 3347 Node* InitializeNode::make_raw_address(intptr_t offset, 3348 PhaseTransform* phase) { 3349 Node* addr = in(RawAddress); 3350 if (offset != 0) { 3351 Compile* C = phase->C; 3352 addr = phase->transform( new (C) AddPNode(C->top(), addr, 3353 phase->MakeConX(offset)) ); 3354 } 3355 return addr; 3356 } 3357 3358 // Clone the given store, converting it into a raw store 3359 // initializing a field or element of my new object. 3360 // Caller is responsible for retiring the original store, 3361 // with subsume_node or the like. 3362 // 3363 // From the example above InitializeNode::InitializeNode, 3364 // here are the old stores to be captured: 3365 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3366 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3367 // 3368 // Here is the changed code; note the extra edges on init: 3369 // alloc = (Allocate ...) 3370 // rawoop = alloc.RawAddress 3371 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3372 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3373 // init = (Initialize alloc.Control alloc.Memory rawoop 3374 // rawstore1 rawstore2) 3375 // 3376 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3377 PhaseTransform* phase, bool can_reshape) { 3378 assert(stores_are_sane(phase), ""); 3379 3380 if (start < 0) return NULL; 3381 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3382 3383 Compile* C = phase->C; 3384 int size_in_bytes = st->memory_size(); 3385 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3386 if (i == 0) return NULL; // bail out 3387 Node* prev_mem = NULL; // raw memory for the captured store 3388 if (i > 0) { 3389 prev_mem = in(i); // there is a pre-existing store under this one 3390 set_req(i, C->top()); // temporarily disconnect it 3391 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3392 } else { 3393 i = -i; // no pre-existing store 3394 prev_mem = zero_memory(); // a slice of the newly allocated object 3395 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3396 set_req(--i, C->top()); // reuse this edge; it has been folded away 3397 else 3398 ins_req(i, C->top()); // build a new edge 3399 } 3400 Node* new_st = st->clone(); 3401 new_st->set_req(MemNode::Control, in(Control)); 3402 new_st->set_req(MemNode::Memory, prev_mem); 3403 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3404 new_st = phase->transform(new_st); 3405 3406 // At this point, new_st might have swallowed a pre-existing store 3407 // at the same offset, or perhaps new_st might have disappeared, 3408 // if it redundantly stored the same value (or zero to fresh memory). 3409 3410 // In any case, wire it in: 3411 set_req(i, new_st); 3412 3413 // The caller may now kill the old guy. 3414 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3415 assert(check_st == new_st || check_st == NULL, "must be findable"); 3416 assert(!is_complete(), ""); 3417 return new_st; 3418 } 3419 3420 static bool store_constant(jlong* tiles, int num_tiles, 3421 intptr_t st_off, int st_size, 3422 jlong con) { 3423 if ((st_off & (st_size-1)) != 0) 3424 return false; // strange store offset (assume size==2**N) 3425 address addr = (address)tiles + st_off; 3426 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3427 switch (st_size) { 3428 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3429 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3430 case sizeof(jint): *(jint*) addr = (jint) con; break; 3431 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3432 default: return false; // strange store size (detect size!=2**N here) 3433 } 3434 return true; // return success to caller 3435 } 3436 3437 // Coalesce subword constants into int constants and possibly 3438 // into long constants. The goal, if the CPU permits, 3439 // is to initialize the object with a small number of 64-bit tiles. 3440 // Also, convert floating-point constants to bit patterns. 3441 // Non-constants are not relevant to this pass. 3442 // 3443 // In terms of the running example on InitializeNode::InitializeNode 3444 // and InitializeNode::capture_store, here is the transformation 3445 // of rawstore1 and rawstore2 into rawstore12: 3446 // alloc = (Allocate ...) 3447 // rawoop = alloc.RawAddress 3448 // tile12 = 0x00010002 3449 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3450 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3451 // 3452 void 3453 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3454 Node* size_in_bytes, 3455 PhaseGVN* phase) { 3456 Compile* C = phase->C; 3457 3458 assert(stores_are_sane(phase), ""); 3459 // Note: After this pass, they are not completely sane, 3460 // since there may be some overlaps. 3461 3462 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3463 3464 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3465 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3466 size_limit = MIN2(size_limit, ti_limit); 3467 size_limit = align_size_up(size_limit, BytesPerLong); 3468 int num_tiles = size_limit / BytesPerLong; 3469 3470 // allocate space for the tile map: 3471 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3472 jlong tiles_buf[small_len]; 3473 Node* nodes_buf[small_len]; 3474 jlong inits_buf[small_len]; 3475 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3476 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3477 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3478 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3479 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3480 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3481 // tiles: exact bitwise model of all primitive constants 3482 // nodes: last constant-storing node subsumed into the tiles model 3483 // inits: which bytes (in each tile) are touched by any initializations 3484 3485 //// Pass A: Fill in the tile model with any relevant stores. 3486 3487 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3488 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3489 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3490 Node* zmem = zero_memory(); // initially zero memory state 3491 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3492 Node* st = in(i); 3493 intptr_t st_off = get_store_offset(st, phase); 3494 3495 // Figure out the store's offset and constant value: 3496 if (st_off < header_size) continue; //skip (ignore header) 3497 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3498 int st_size = st->as_Store()->memory_size(); 3499 if (st_off + st_size > size_limit) break; 3500 3501 // Record which bytes are touched, whether by constant or not. 3502 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3503 continue; // skip (strange store size) 3504 3505 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3506 if (!val->singleton()) continue; //skip (non-con store) 3507 BasicType type = val->basic_type(); 3508 3509 jlong con = 0; 3510 switch (type) { 3511 case T_INT: con = val->is_int()->get_con(); break; 3512 case T_LONG: con = val->is_long()->get_con(); break; 3513 case T_FLOAT: con = jint_cast(val->getf()); break; 3514 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3515 default: continue; //skip (odd store type) 3516 } 3517 3518 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3519 st->Opcode() == Op_StoreL) { 3520 continue; // This StoreL is already optimal. 3521 } 3522 3523 // Store down the constant. 3524 store_constant(tiles, num_tiles, st_off, st_size, con); 3525 3526 intptr_t j = st_off >> LogBytesPerLong; 3527 3528 if (type == T_INT && st_size == BytesPerInt 3529 && (st_off & BytesPerInt) == BytesPerInt) { 3530 jlong lcon = tiles[j]; 3531 if (!Matcher::isSimpleConstant64(lcon) && 3532 st->Opcode() == Op_StoreI) { 3533 // This StoreI is already optimal by itself. 3534 jint* intcon = (jint*) &tiles[j]; 3535 intcon[1] = 0; // undo the store_constant() 3536 3537 // If the previous store is also optimal by itself, back up and 3538 // undo the action of the previous loop iteration... if we can. 3539 // But if we can't, just let the previous half take care of itself. 3540 st = nodes[j]; 3541 st_off -= BytesPerInt; 3542 con = intcon[0]; 3543 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3544 assert(st_off >= header_size, "still ignoring header"); 3545 assert(get_store_offset(st, phase) == st_off, "must be"); 3546 assert(in(i-1) == zmem, "must be"); 3547 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3548 assert(con == tcon->is_int()->get_con(), "must be"); 3549 // Undo the effects of the previous loop trip, which swallowed st: 3550 intcon[0] = 0; // undo store_constant() 3551 set_req(i-1, st); // undo set_req(i, zmem) 3552 nodes[j] = NULL; // undo nodes[j] = st 3553 --old_subword; // undo ++old_subword 3554 } 3555 continue; // This StoreI is already optimal. 3556 } 3557 } 3558 3559 // This store is not needed. 3560 set_req(i, zmem); 3561 nodes[j] = st; // record for the moment 3562 if (st_size < BytesPerLong) // something has changed 3563 ++old_subword; // includes int/float, but who's counting... 3564 else ++old_long; 3565 } 3566 3567 if ((old_subword + old_long) == 0) 3568 return; // nothing more to do 3569 3570 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3571 // Be sure to insert them before overlapping non-constant stores. 3572 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3573 for (int j = 0; j < num_tiles; j++) { 3574 jlong con = tiles[j]; 3575 jlong init = inits[j]; 3576 if (con == 0) continue; 3577 jint con0, con1; // split the constant, address-wise 3578 jint init0, init1; // split the init map, address-wise 3579 { union { jlong con; jint intcon[2]; } u; 3580 u.con = con; 3581 con0 = u.intcon[0]; 3582 con1 = u.intcon[1]; 3583 u.con = init; 3584 init0 = u.intcon[0]; 3585 init1 = u.intcon[1]; 3586 } 3587 3588 Node* old = nodes[j]; 3589 assert(old != NULL, "need the prior store"); 3590 intptr_t offset = (j * BytesPerLong); 3591 3592 bool split = !Matcher::isSimpleConstant64(con); 3593 3594 if (offset < header_size) { 3595 assert(offset + BytesPerInt >= header_size, "second int counts"); 3596 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3597 split = true; // only the second word counts 3598 // Example: int a[] = { 42 ... } 3599 } else if (con0 == 0 && init0 == -1) { 3600 split = true; // first word is covered by full inits 3601 // Example: int a[] = { ... foo(), 42 ... } 3602 } else if (con1 == 0 && init1 == -1) { 3603 split = true; // second word is covered by full inits 3604 // Example: int a[] = { ... 42, foo() ... } 3605 } 3606 3607 // Here's a case where init0 is neither 0 nor -1: 3608 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3609 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3610 // In this case the tile is not split; it is (jlong)42. 3611 // The big tile is stored down, and then the foo() value is inserted. 3612 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3613 3614 Node* ctl = old->in(MemNode::Control); 3615 Node* adr = make_raw_address(offset, phase); 3616 const TypePtr* atp = TypeRawPtr::BOTTOM; 3617 3618 // One or two coalesced stores to plop down. 3619 Node* st[2]; 3620 intptr_t off[2]; 3621 int nst = 0; 3622 if (!split) { 3623 ++new_long; 3624 off[nst] = offset; 3625 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3626 phase->longcon(con), T_LONG); 3627 } else { 3628 // Omit either if it is a zero. 3629 if (con0 != 0) { 3630 ++new_int; 3631 off[nst] = offset; 3632 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3633 phase->intcon(con0), T_INT); 3634 } 3635 if (con1 != 0) { 3636 ++new_int; 3637 offset += BytesPerInt; 3638 adr = make_raw_address(offset, phase); 3639 off[nst] = offset; 3640 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3641 phase->intcon(con1), T_INT); 3642 } 3643 } 3644 3645 // Insert second store first, then the first before the second. 3646 // Insert each one just before any overlapping non-constant stores. 3647 while (nst > 0) { 3648 Node* st1 = st[--nst]; 3649 C->copy_node_notes_to(st1, old); 3650 st1 = phase->transform(st1); 3651 offset = off[nst]; 3652 assert(offset >= header_size, "do not smash header"); 3653 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3654 guarantee(ins_idx != 0, "must re-insert constant store"); 3655 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3656 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3657 set_req(--ins_idx, st1); 3658 else 3659 ins_req(ins_idx, st1); 3660 } 3661 } 3662 3663 if (PrintCompilation && WizardMode) 3664 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3665 old_subword, old_long, new_int, new_long); 3666 if (C->log() != NULL) 3667 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3668 old_subword, old_long, new_int, new_long); 3669 3670 // Clean up any remaining occurrences of zmem: 3671 remove_extra_zeroes(); 3672 } 3673 3674 // Explore forward from in(start) to find the first fully initialized 3675 // word, and return its offset. Skip groups of subword stores which 3676 // together initialize full words. If in(start) is itself part of a 3677 // fully initialized word, return the offset of in(start). If there 3678 // are no following full-word stores, or if something is fishy, return 3679 // a negative value. 3680 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3681 int int_map = 0; 3682 intptr_t int_map_off = 0; 3683 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3684 3685 for (uint i = start, limit = req(); i < limit; i++) { 3686 Node* st = in(i); 3687 3688 intptr_t st_off = get_store_offset(st, phase); 3689 if (st_off < 0) break; // return conservative answer 3690 3691 int st_size = st->as_Store()->memory_size(); 3692 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3693 return st_off; // we found a complete word init 3694 } 3695 3696 // update the map: 3697 3698 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3699 if (this_int_off != int_map_off) { 3700 // reset the map: 3701 int_map = 0; 3702 int_map_off = this_int_off; 3703 } 3704 3705 int subword_off = st_off - this_int_off; 3706 int_map |= right_n_bits(st_size) << subword_off; 3707 if ((int_map & FULL_MAP) == FULL_MAP) { 3708 return this_int_off; // we found a complete word init 3709 } 3710 3711 // Did this store hit or cross the word boundary? 3712 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3713 if (next_int_off == this_int_off + BytesPerInt) { 3714 // We passed the current int, without fully initializing it. 3715 int_map_off = next_int_off; 3716 int_map >>= BytesPerInt; 3717 } else if (next_int_off > this_int_off + BytesPerInt) { 3718 // We passed the current and next int. 3719 return this_int_off + BytesPerInt; 3720 } 3721 } 3722 3723 return -1; 3724 } 3725 3726 3727 // Called when the associated AllocateNode is expanded into CFG. 3728 // At this point, we may perform additional optimizations. 3729 // Linearize the stores by ascending offset, to make memory 3730 // activity as coherent as possible. 3731 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3732 intptr_t header_size, 3733 Node* size_in_bytes, 3734 PhaseGVN* phase) { 3735 assert(!is_complete(), "not already complete"); 3736 assert(stores_are_sane(phase), ""); 3737 assert(allocation() != NULL, "must be present"); 3738 3739 remove_extra_zeroes(); 3740 3741 if (ReduceFieldZeroing || ReduceBulkZeroing) 3742 // reduce instruction count for common initialization patterns 3743 coalesce_subword_stores(header_size, size_in_bytes, phase); 3744 3745 Node* zmem = zero_memory(); // initially zero memory state 3746 Node* inits = zmem; // accumulating a linearized chain of inits 3747 #ifdef ASSERT 3748 intptr_t first_offset = allocation()->minimum_header_size(); 3749 intptr_t last_init_off = first_offset; // previous init offset 3750 intptr_t last_init_end = first_offset; // previous init offset+size 3751 intptr_t last_tile_end = first_offset; // previous tile offset+size 3752 #endif 3753 intptr_t zeroes_done = header_size; 3754 3755 bool do_zeroing = true; // we might give up if inits are very sparse 3756 int big_init_gaps = 0; // how many large gaps have we seen? 3757 3758 if (ZeroTLAB) do_zeroing = false; 3759 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3760 3761 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3762 Node* st = in(i); 3763 intptr_t st_off = get_store_offset(st, phase); 3764 if (st_off < 0) 3765 break; // unknown junk in the inits 3766 if (st->in(MemNode::Memory) != zmem) 3767 break; // complicated store chains somehow in list 3768 3769 int st_size = st->as_Store()->memory_size(); 3770 intptr_t next_init_off = st_off + st_size; 3771 3772 if (do_zeroing && zeroes_done < next_init_off) { 3773 // See if this store needs a zero before it or under it. 3774 intptr_t zeroes_needed = st_off; 3775 3776 if (st_size < BytesPerInt) { 3777 // Look for subword stores which only partially initialize words. 3778 // If we find some, we must lay down some word-level zeroes first, 3779 // underneath the subword stores. 3780 // 3781 // Examples: 3782 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3783 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3784 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3785 // 3786 // Note: coalesce_subword_stores may have already done this, 3787 // if it was prompted by constant non-zero subword initializers. 3788 // But this case can still arise with non-constant stores. 3789 3790 intptr_t next_full_store = find_next_fullword_store(i, phase); 3791 3792 // In the examples above: 3793 // in(i) p q r s x y z 3794 // st_off 12 13 14 15 12 13 14 3795 // st_size 1 1 1 1 1 1 1 3796 // next_full_s. 12 16 16 16 16 16 16 3797 // z's_done 12 16 16 16 12 16 12 3798 // z's_needed 12 16 16 16 16 16 16 3799 // zsize 0 0 0 0 4 0 4 3800 if (next_full_store < 0) { 3801 // Conservative tack: Zero to end of current word. 3802 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3803 } else { 3804 // Zero to beginning of next fully initialized word. 3805 // Or, don't zero at all, if we are already in that word. 3806 assert(next_full_store >= zeroes_needed, "must go forward"); 3807 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3808 zeroes_needed = next_full_store; 3809 } 3810 } 3811 3812 if (zeroes_needed > zeroes_done) { 3813 intptr_t zsize = zeroes_needed - zeroes_done; 3814 // Do some incremental zeroing on rawmem, in parallel with inits. 3815 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3816 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3817 zeroes_done, zeroes_needed, 3818 phase); 3819 zeroes_done = zeroes_needed; 3820 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3821 do_zeroing = false; // leave the hole, next time 3822 } 3823 } 3824 3825 // Collect the store and move on: 3826 st->set_req(MemNode::Memory, inits); 3827 inits = st; // put it on the linearized chain 3828 set_req(i, zmem); // unhook from previous position 3829 3830 if (zeroes_done == st_off) 3831 zeroes_done = next_init_off; 3832 3833 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3834 3835 #ifdef ASSERT 3836 // Various order invariants. Weaker than stores_are_sane because 3837 // a large constant tile can be filled in by smaller non-constant stores. 3838 assert(st_off >= last_init_off, "inits do not reverse"); 3839 last_init_off = st_off; 3840 const Type* val = NULL; 3841 if (st_size >= BytesPerInt && 3842 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3843 (int)val->basic_type() < (int)T_OBJECT) { 3844 assert(st_off >= last_tile_end, "tiles do not overlap"); 3845 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3846 last_tile_end = MAX2(last_tile_end, next_init_off); 3847 } else { 3848 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3849 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3850 assert(st_off >= last_init_end, "inits do not overlap"); 3851 last_init_end = next_init_off; // it's a non-tile 3852 } 3853 #endif //ASSERT 3854 } 3855 3856 remove_extra_zeroes(); // clear out all the zmems left over 3857 add_req(inits); 3858 3859 if (!ZeroTLAB) { 3860 // If anything remains to be zeroed, zero it all now. 3861 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3862 // if it is the last unused 4 bytes of an instance, forget about it 3863 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3864 if (zeroes_done + BytesPerLong >= size_limit) { 3865 assert(allocation() != NULL, ""); 3866 if (allocation()->Opcode() == Op_Allocate) { 3867 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3868 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3869 if (zeroes_done == k->layout_helper()) 3870 zeroes_done = size_limit; 3871 } 3872 } 3873 if (zeroes_done < size_limit) { 3874 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3875 zeroes_done, size_in_bytes, phase); 3876 } 3877 } 3878 3879 set_complete(phase); 3880 return rawmem; 3881 } 3882 3883 3884 #ifdef ASSERT 3885 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3886 if (is_complete()) 3887 return true; // stores could be anything at this point 3888 assert(allocation() != NULL, "must be present"); 3889 intptr_t last_off = allocation()->minimum_header_size(); 3890 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3891 Node* st = in(i); 3892 intptr_t st_off = get_store_offset(st, phase); 3893 if (st_off < 0) continue; // ignore dead garbage 3894 if (last_off > st_off) { 3895 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3896 this->dump(2); 3897 assert(false, "ascending store offsets"); 3898 return false; 3899 } 3900 last_off = st_off + st->as_Store()->memory_size(); 3901 } 3902 return true; 3903 } 3904 #endif //ASSERT 3905 3906 3907 3908 3909 //============================MergeMemNode===================================== 3910 // 3911 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3912 // contributing store or call operations. Each contributor provides the memory 3913 // state for a particular "alias type" (see Compile::alias_type). For example, 3914 // if a MergeMem has an input X for alias category #6, then any memory reference 3915 // to alias category #6 may use X as its memory state input, as an exact equivalent 3916 // to using the MergeMem as a whole. 3917 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3918 // 3919 // (Here, the <N> notation gives the index of the relevant adr_type.) 3920 // 3921 // In one special case (and more cases in the future), alias categories overlap. 3922 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3923 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3924 // it is exactly equivalent to that state W: 3925 // MergeMem(<Bot>: W) <==> W 3926 // 3927 // Usually, the merge has more than one input. In that case, where inputs 3928 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3929 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3930 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3931 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3932 // 3933 // A merge can take a "wide" memory state as one of its narrow inputs. 3934 // This simply means that the merge observes out only the relevant parts of 3935 // the wide input. That is, wide memory states arriving at narrow merge inputs 3936 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3937 // 3938 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3939 // and that memory slices "leak through": 3940 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3941 // 3942 // But, in such a cascade, repeated memory slices can "block the leak": 3943 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3944 // 3945 // In the last example, Y is not part of the combined memory state of the 3946 // outermost MergeMem. The system must, of course, prevent unschedulable 3947 // memory states from arising, so you can be sure that the state Y is somehow 3948 // a precursor to state Y'. 3949 // 3950 // 3951 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3952 // of each MergeMemNode array are exactly the numerical alias indexes, including 3953 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3954 // Compile::alias_type (and kin) produce and manage these indexes. 3955 // 3956 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3957 // (Note that this provides quick access to the top node inside MergeMem methods, 3958 // without the need to reach out via TLS to Compile::current.) 3959 // 3960 // As a consequence of what was just described, a MergeMem that represents a full 3961 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3962 // containing all alias categories. 3963 // 3964 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3965 // 3966 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3967 // a memory state for the alias type <N>, or else the top node, meaning that 3968 // there is no particular input for that alias type. Note that the length of 3969 // a MergeMem is variable, and may be extended at any time to accommodate new 3970 // memory states at larger alias indexes. When merges grow, they are of course 3971 // filled with "top" in the unused in() positions. 3972 // 3973 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3974 // (Top was chosen because it works smoothly with passes like GCM.) 3975 // 3976 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3977 // the type of random VM bits like TLS references.) Since it is always the 3978 // first non-Bot memory slice, some low-level loops use it to initialize an 3979 // index variable: for (i = AliasIdxRaw; i < req(); i++). 3980 // 3981 // 3982 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3983 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3984 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 3985 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3986 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3987 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3988 // 3989 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3990 // really that different from the other memory inputs. An abbreviation called 3991 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3992 // 3993 // 3994 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3995 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3996 // that "emerges though" the base memory will be marked as excluding the alias types 3997 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 3998 // 3999 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4000 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4001 // 4002 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4003 // (It is currently unimplemented.) As you can see, the resulting merge is 4004 // actually a disjoint union of memory states, rather than an overlay. 4005 // 4006 4007 //------------------------------MergeMemNode----------------------------------- 4008 Node* MergeMemNode::make_empty_memory() { 4009 Node* empty_memory = (Node*) Compile::current()->top(); 4010 assert(empty_memory->is_top(), "correct sentinel identity"); 4011 return empty_memory; 4012 } 4013 4014 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4015 init_class_id(Class_MergeMem); 4016 // all inputs are nullified in Node::Node(int) 4017 // set_input(0, NULL); // no control input 4018 4019 // Initialize the edges uniformly to top, for starters. 4020 Node* empty_mem = make_empty_memory(); 4021 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4022 init_req(i,empty_mem); 4023 } 4024 assert(empty_memory() == empty_mem, ""); 4025 4026 if( new_base != NULL && new_base->is_MergeMem() ) { 4027 MergeMemNode* mdef = new_base->as_MergeMem(); 4028 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4029 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4030 mms.set_memory(mms.memory2()); 4031 } 4032 assert(base_memory() == mdef->base_memory(), ""); 4033 } else { 4034 set_base_memory(new_base); 4035 } 4036 } 4037 4038 // Make a new, untransformed MergeMem with the same base as 'mem'. 4039 // If mem is itself a MergeMem, populate the result with the same edges. 4040 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 4041 return new(C) MergeMemNode(mem); 4042 } 4043 4044 //------------------------------cmp-------------------------------------------- 4045 uint MergeMemNode::hash() const { return NO_HASH; } 4046 uint MergeMemNode::cmp( const Node &n ) const { 4047 return (&n == this); // Always fail except on self 4048 } 4049 4050 //------------------------------Identity--------------------------------------- 4051 Node* MergeMemNode::Identity(PhaseTransform *phase) { 4052 // Identity if this merge point does not record any interesting memory 4053 // disambiguations. 4054 Node* base_mem = base_memory(); 4055 Node* empty_mem = empty_memory(); 4056 if (base_mem != empty_mem) { // Memory path is not dead? 4057 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4058 Node* mem = in(i); 4059 if (mem != empty_mem && mem != base_mem) { 4060 return this; // Many memory splits; no change 4061 } 4062 } 4063 } 4064 return base_mem; // No memory splits; ID on the one true input 4065 } 4066 4067 //------------------------------Ideal------------------------------------------ 4068 // This method is invoked recursively on chains of MergeMem nodes 4069 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4070 // Remove chain'd MergeMems 4071 // 4072 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4073 // relative to the "in(Bot)". Since we are patching both at the same time, 4074 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4075 // but rewrite each "in(i)" relative to the new "in(Bot)". 4076 Node *progress = NULL; 4077 4078 4079 Node* old_base = base_memory(); 4080 Node* empty_mem = empty_memory(); 4081 if (old_base == empty_mem) 4082 return NULL; // Dead memory path. 4083 4084 MergeMemNode* old_mbase; 4085 if (old_base != NULL && old_base->is_MergeMem()) 4086 old_mbase = old_base->as_MergeMem(); 4087 else 4088 old_mbase = NULL; 4089 Node* new_base = old_base; 4090 4091 // simplify stacked MergeMems in base memory 4092 if (old_mbase) new_base = old_mbase->base_memory(); 4093 4094 // the base memory might contribute new slices beyond my req() 4095 if (old_mbase) grow_to_match(old_mbase); 4096 4097 // Look carefully at the base node if it is a phi. 4098 PhiNode* phi_base; 4099 if (new_base != NULL && new_base->is_Phi()) 4100 phi_base = new_base->as_Phi(); 4101 else 4102 phi_base = NULL; 4103 4104 Node* phi_reg = NULL; 4105 uint phi_len = (uint)-1; 4106 if (phi_base != NULL && !phi_base->is_copy()) { 4107 // do not examine phi if degraded to a copy 4108 phi_reg = phi_base->region(); 4109 phi_len = phi_base->req(); 4110 // see if the phi is unfinished 4111 for (uint i = 1; i < phi_len; i++) { 4112 if (phi_base->in(i) == NULL) { 4113 // incomplete phi; do not look at it yet! 4114 phi_reg = NULL; 4115 phi_len = (uint)-1; 4116 break; 4117 } 4118 } 4119 } 4120 4121 // Note: We do not call verify_sparse on entry, because inputs 4122 // can normalize to the base_memory via subsume_node or similar 4123 // mechanisms. This method repairs that damage. 4124 4125 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4126 4127 // Look at each slice. 4128 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4129 Node* old_in = in(i); 4130 // calculate the old memory value 4131 Node* old_mem = old_in; 4132 if (old_mem == empty_mem) old_mem = old_base; 4133 assert(old_mem == memory_at(i), ""); 4134 4135 // maybe update (reslice) the old memory value 4136 4137 // simplify stacked MergeMems 4138 Node* new_mem = old_mem; 4139 MergeMemNode* old_mmem; 4140 if (old_mem != NULL && old_mem->is_MergeMem()) 4141 old_mmem = old_mem->as_MergeMem(); 4142 else 4143 old_mmem = NULL; 4144 if (old_mmem == this) { 4145 // This can happen if loops break up and safepoints disappear. 4146 // A merge of BotPtr (default) with a RawPtr memory derived from a 4147 // safepoint can be rewritten to a merge of the same BotPtr with 4148 // the BotPtr phi coming into the loop. If that phi disappears 4149 // also, we can end up with a self-loop of the mergemem. 4150 // In general, if loops degenerate and memory effects disappear, 4151 // a mergemem can be left looking at itself. This simply means 4152 // that the mergemem's default should be used, since there is 4153 // no longer any apparent effect on this slice. 4154 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4155 // from start. Update the input to TOP. 4156 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4157 } 4158 else if (old_mmem != NULL) { 4159 new_mem = old_mmem->memory_at(i); 4160 } 4161 // else preceding memory was not a MergeMem 4162 4163 // replace equivalent phis (unfortunately, they do not GVN together) 4164 if (new_mem != NULL && new_mem != new_base && 4165 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4166 if (new_mem->is_Phi()) { 4167 PhiNode* phi_mem = new_mem->as_Phi(); 4168 for (uint i = 1; i < phi_len; i++) { 4169 if (phi_base->in(i) != phi_mem->in(i)) { 4170 phi_mem = NULL; 4171 break; 4172 } 4173 } 4174 if (phi_mem != NULL) { 4175 // equivalent phi nodes; revert to the def 4176 new_mem = new_base; 4177 } 4178 } 4179 } 4180 4181 // maybe store down a new value 4182 Node* new_in = new_mem; 4183 if (new_in == new_base) new_in = empty_mem; 4184 4185 if (new_in != old_in) { 4186 // Warning: Do not combine this "if" with the previous "if" 4187 // A memory slice might have be be rewritten even if it is semantically 4188 // unchanged, if the base_memory value has changed. 4189 set_req(i, new_in); 4190 progress = this; // Report progress 4191 } 4192 } 4193 4194 if (new_base != old_base) { 4195 set_req(Compile::AliasIdxBot, new_base); 4196 // Don't use set_base_memory(new_base), because we need to update du. 4197 assert(base_memory() == new_base, ""); 4198 progress = this; 4199 } 4200 4201 if( base_memory() == this ) { 4202 // a self cycle indicates this memory path is dead 4203 set_req(Compile::AliasIdxBot, empty_mem); 4204 } 4205 4206 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4207 // Recursion must occur after the self cycle check above 4208 if( base_memory()->is_MergeMem() ) { 4209 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4210 Node *m = phase->transform(new_mbase); // Rollup any cycles 4211 if( m != NULL && (m->is_top() || 4212 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4213 // propagate rollup of dead cycle to self 4214 set_req(Compile::AliasIdxBot, empty_mem); 4215 } 4216 } 4217 4218 if( base_memory() == empty_mem ) { 4219 progress = this; 4220 // Cut inputs during Parse phase only. 4221 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4222 if( !can_reshape ) { 4223 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4224 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4225 } 4226 } 4227 } 4228 4229 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4230 // Check if PhiNode::Ideal's "Split phis through memory merges" 4231 // transform should be attempted. Look for this->phi->this cycle. 4232 uint merge_width = req(); 4233 if (merge_width > Compile::AliasIdxRaw) { 4234 PhiNode* phi = base_memory()->as_Phi(); 4235 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4236 if (phi->in(i) == this) { 4237 phase->is_IterGVN()->_worklist.push(phi); 4238 break; 4239 } 4240 } 4241 } 4242 } 4243 4244 assert(progress || verify_sparse(), "please, no dups of base"); 4245 return progress; 4246 } 4247 4248 //-------------------------set_base_memory------------------------------------- 4249 void MergeMemNode::set_base_memory(Node *new_base) { 4250 Node* empty_mem = empty_memory(); 4251 set_req(Compile::AliasIdxBot, new_base); 4252 assert(memory_at(req()) == new_base, "must set default memory"); 4253 // Clear out other occurrences of new_base: 4254 if (new_base != empty_mem) { 4255 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4256 if (in(i) == new_base) set_req(i, empty_mem); 4257 } 4258 } 4259 } 4260 4261 //------------------------------out_RegMask------------------------------------ 4262 const RegMask &MergeMemNode::out_RegMask() const { 4263 return RegMask::Empty; 4264 } 4265 4266 //------------------------------dump_spec-------------------------------------- 4267 #ifndef PRODUCT 4268 void MergeMemNode::dump_spec(outputStream *st) const { 4269 st->print(" {"); 4270 Node* base_mem = base_memory(); 4271 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4272 Node* mem = memory_at(i); 4273 if (mem == base_mem) { st->print(" -"); continue; } 4274 st->print( " N%d:", mem->_idx ); 4275 Compile::current()->get_adr_type(i)->dump_on(st); 4276 } 4277 st->print(" }"); 4278 } 4279 #endif // !PRODUCT 4280 4281 4282 #ifdef ASSERT 4283 static bool might_be_same(Node* a, Node* b) { 4284 if (a == b) return true; 4285 if (!(a->is_Phi() || b->is_Phi())) return false; 4286 // phis shift around during optimization 4287 return true; // pretty stupid... 4288 } 4289 4290 // verify a narrow slice (either incoming or outgoing) 4291 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4292 if (!VerifyAliases) return; // don't bother to verify unless requested 4293 if (is_error_reported()) return; // muzzle asserts when debugging an error 4294 if (Node::in_dump()) return; // muzzle asserts when printing 4295 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4296 assert(n != NULL, ""); 4297 // Elide intervening MergeMem's 4298 while (n->is_MergeMem()) { 4299 n = n->as_MergeMem()->memory_at(alias_idx); 4300 } 4301 Compile* C = Compile::current(); 4302 const TypePtr* n_adr_type = n->adr_type(); 4303 if (n == m->empty_memory()) { 4304 // Implicit copy of base_memory() 4305 } else if (n_adr_type != TypePtr::BOTTOM) { 4306 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4307 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4308 } else { 4309 // A few places like make_runtime_call "know" that VM calls are narrow, 4310 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4311 bool expected_wide_mem = false; 4312 if (n == m->base_memory()) { 4313 expected_wide_mem = true; 4314 } else if (alias_idx == Compile::AliasIdxRaw || 4315 n == m->memory_at(Compile::AliasIdxRaw)) { 4316 expected_wide_mem = true; 4317 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4318 // memory can "leak through" calls on channels that 4319 // are write-once. Allow this also. 4320 expected_wide_mem = true; 4321 } 4322 assert(expected_wide_mem, "expected narrow slice replacement"); 4323 } 4324 } 4325 #else // !ASSERT 4326 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 4327 #endif 4328 4329 4330 //-----------------------------memory_at--------------------------------------- 4331 Node* MergeMemNode::memory_at(uint alias_idx) const { 4332 assert(alias_idx >= Compile::AliasIdxRaw || 4333 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4334 "must avoid base_memory and AliasIdxTop"); 4335 4336 // Otherwise, it is a narrow slice. 4337 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4338 Compile *C = Compile::current(); 4339 if (is_empty_memory(n)) { 4340 // the array is sparse; empty slots are the "top" node 4341 n = base_memory(); 4342 assert(Node::in_dump() 4343 || n == NULL || n->bottom_type() == Type::TOP 4344 || n->adr_type() == NULL // address is TOP 4345 || n->adr_type() == TypePtr::BOTTOM 4346 || n->adr_type() == TypeRawPtr::BOTTOM 4347 || Compile::current()->AliasLevel() == 0, 4348 "must be a wide memory"); 4349 // AliasLevel == 0 if we are organizing the memory states manually. 4350 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4351 } else { 4352 // make sure the stored slice is sane 4353 #ifdef ASSERT 4354 if (is_error_reported() || Node::in_dump()) { 4355 } else if (might_be_same(n, base_memory())) { 4356 // Give it a pass: It is a mostly harmless repetition of the base. 4357 // This can arise normally from node subsumption during optimization. 4358 } else { 4359 verify_memory_slice(this, alias_idx, n); 4360 } 4361 #endif 4362 } 4363 return n; 4364 } 4365 4366 //---------------------------set_memory_at------------------------------------- 4367 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4368 verify_memory_slice(this, alias_idx, n); 4369 Node* empty_mem = empty_memory(); 4370 if (n == base_memory()) n = empty_mem; // collapse default 4371 uint need_req = alias_idx+1; 4372 if (req() < need_req) { 4373 if (n == empty_mem) return; // already the default, so do not grow me 4374 // grow the sparse array 4375 do { 4376 add_req(empty_mem); 4377 } while (req() < need_req); 4378 } 4379 set_req( alias_idx, n ); 4380 } 4381 4382 4383 4384 //--------------------------iteration_setup------------------------------------ 4385 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4386 if (other != NULL) { 4387 grow_to_match(other); 4388 // invariant: the finite support of mm2 is within mm->req() 4389 #ifdef ASSERT 4390 for (uint i = req(); i < other->req(); i++) { 4391 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4392 } 4393 #endif 4394 } 4395 // Replace spurious copies of base_memory by top. 4396 Node* base_mem = base_memory(); 4397 if (base_mem != NULL && !base_mem->is_top()) { 4398 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4399 if (in(i) == base_mem) 4400 set_req(i, empty_memory()); 4401 } 4402 } 4403 } 4404 4405 //---------------------------grow_to_match------------------------------------- 4406 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4407 Node* empty_mem = empty_memory(); 4408 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4409 // look for the finite support of the other memory 4410 for (uint i = other->req(); --i >= req(); ) { 4411 if (other->in(i) != empty_mem) { 4412 uint new_len = i+1; 4413 while (req() < new_len) add_req(empty_mem); 4414 break; 4415 } 4416 } 4417 } 4418 4419 //---------------------------verify_sparse------------------------------------- 4420 #ifndef PRODUCT 4421 bool MergeMemNode::verify_sparse() const { 4422 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4423 Node* base_mem = base_memory(); 4424 // The following can happen in degenerate cases, since empty==top. 4425 if (is_empty_memory(base_mem)) return true; 4426 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4427 assert(in(i) != NULL, "sane slice"); 4428 if (in(i) == base_mem) return false; // should have been the sentinel value! 4429 } 4430 return true; 4431 } 4432 4433 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4434 Node* n; 4435 n = mm->in(idx); 4436 if (mem == n) return true; // might be empty_memory() 4437 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4438 if (mem == n) return true; 4439 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4440 if (mem == n) return true; 4441 if (n == NULL) break; 4442 } 4443 return false; 4444 } 4445 #endif // !PRODUCT