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 && (phase->type(address) != 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 assert(base == NULL || t_adr->isa_rawptr() || 300 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); 301 #endif 302 303 // Avoid independent memory operations 304 Node* old_mem = mem; 305 306 // The code which unhooks non-raw memories from complete (macro-expanded) 307 // initializations was removed. After macro-expansion all stores catched 308 // by Initialize node became raw stores and there is no information 309 // which memory slices they modify. So it is unsafe to move any memory 310 // operation above these stores. Also in most cases hooked non-raw memories 311 // were already unhooked by using information from detect_ptr_independence() 312 // and find_previous_store(). 313 314 if (mem->is_MergeMem()) { 315 MergeMemNode* mmem = mem->as_MergeMem(); 316 const TypePtr *tp = t_adr->is_ptr(); 317 318 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 319 } 320 321 if (mem != old_mem) { 322 set_req(MemNode::Memory, mem); 323 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 324 return this; 325 } 326 327 // let the subclass continue analyzing... 328 return NULL; 329 } 330 331 // Helper function for proving some simple control dominations. 332 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 333 // Already assumes that 'dom' is available at 'sub', and that 'sub' 334 // is not a constant (dominated by the method's StartNode). 335 // Used by MemNode::find_previous_store to prove that the 336 // control input of a memory operation predates (dominates) 337 // an allocation it wants to look past. 338 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 339 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 340 return false; // Conservative answer for dead code 341 342 // Check 'dom'. Skip Proj and CatchProj nodes. 343 dom = dom->find_exact_control(dom); 344 if (dom == NULL || dom->is_top()) 345 return false; // Conservative answer for dead code 346 347 if (dom == sub) { 348 // For the case when, for example, 'sub' is Initialize and the original 349 // 'dom' is Proj node of the 'sub'. 350 return false; 351 } 352 353 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 354 return true; 355 356 // 'dom' dominates 'sub' if its control edge and control edges 357 // of all its inputs dominate or equal to sub's control edge. 358 359 // Currently 'sub' is either Allocate, Initialize or Start nodes. 360 // Or Region for the check in LoadNode::Ideal(); 361 // 'sub' should have sub->in(0) != NULL. 362 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 363 sub->is_Region(), "expecting only these nodes"); 364 365 // Get control edge of 'sub'. 366 Node* orig_sub = sub; 367 sub = sub->find_exact_control(sub->in(0)); 368 if (sub == NULL || sub->is_top()) 369 return false; // Conservative answer for dead code 370 371 assert(sub->is_CFG(), "expecting control"); 372 373 if (sub == dom) 374 return true; 375 376 if (sub->is_Start() || sub->is_Root()) 377 return false; 378 379 { 380 // Check all control edges of 'dom'. 381 382 ResourceMark rm; 383 Arena* arena = Thread::current()->resource_area(); 384 Node_List nlist(arena); 385 Unique_Node_List dom_list(arena); 386 387 dom_list.push(dom); 388 bool only_dominating_controls = false; 389 390 for (uint next = 0; next < dom_list.size(); next++) { 391 Node* n = dom_list.at(next); 392 if (n == orig_sub) 393 return false; // One of dom's inputs dominated by sub. 394 if (!n->is_CFG() && n->pinned()) { 395 // Check only own control edge for pinned non-control nodes. 396 n = n->find_exact_control(n->in(0)); 397 if (n == NULL || n->is_top()) 398 return false; // Conservative answer for dead code 399 assert(n->is_CFG(), "expecting control"); 400 dom_list.push(n); 401 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 402 only_dominating_controls = true; 403 } else if (n->is_CFG()) { 404 if (n->dominates(sub, nlist)) 405 only_dominating_controls = true; 406 else 407 return false; 408 } else { 409 // First, own control edge. 410 Node* m = n->find_exact_control(n->in(0)); 411 if (m != NULL) { 412 if (m->is_top()) 413 return false; // Conservative answer for dead code 414 dom_list.push(m); 415 } 416 // Now, the rest of edges. 417 uint cnt = n->req(); 418 for (uint i = 1; i < cnt; i++) { 419 m = n->find_exact_control(n->in(i)); 420 if (m == NULL || m->is_top()) 421 continue; 422 dom_list.push(m); 423 } 424 } 425 } 426 return only_dominating_controls; 427 } 428 } 429 430 //---------------------detect_ptr_independence--------------------------------- 431 // Used by MemNode::find_previous_store to prove that two base 432 // pointers are never equal. 433 // The pointers are accompanied by their associated allocations, 434 // if any, which have been previously discovered by the caller. 435 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 436 Node* p2, AllocateNode* a2, 437 PhaseTransform* phase) { 438 // Attempt to prove that these two pointers cannot be aliased. 439 // They may both manifestly be allocations, and they should differ. 440 // Or, if they are not both allocations, they can be distinct constants. 441 // Otherwise, one is an allocation and the other a pre-existing value. 442 if (a1 == NULL && a2 == NULL) { // neither an allocation 443 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 444 } else if (a1 != NULL && a2 != NULL) { // both allocations 445 return (a1 != a2); 446 } else if (a1 != NULL) { // one allocation a1 447 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 448 return all_controls_dominate(p2, a1); 449 } else { //(a2 != NULL) // one allocation a2 450 return all_controls_dominate(p1, a2); 451 } 452 return false; 453 } 454 455 456 // The logic for reordering loads and stores uses four steps: 457 // (a) Walk carefully past stores and initializations which we 458 // can prove are independent of this load. 459 // (b) Observe that the next memory state makes an exact match 460 // with self (load or store), and locate the relevant store. 461 // (c) Ensure that, if we were to wire self directly to the store, 462 // the optimizer would fold it up somehow. 463 // (d) Do the rewiring, and return, depending on some other part of 464 // the optimizer to fold up the load. 465 // This routine handles steps (a) and (b). Steps (c) and (d) are 466 // specific to loads and stores, so they are handled by the callers. 467 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 468 // 469 Node* MemNode::find_previous_store(PhaseTransform* phase) { 470 Node* ctrl = in(MemNode::Control); 471 Node* adr = in(MemNode::Address); 472 intptr_t offset = 0; 473 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 474 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 475 476 if (offset == Type::OffsetBot) 477 return NULL; // cannot unalias unless there are precise offsets 478 479 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 480 481 intptr_t size_in_bytes = memory_size(); 482 483 Node* mem = in(MemNode::Memory); // start searching here... 484 485 int cnt = 50; // Cycle limiter 486 for (;;) { // While we can dance past unrelated stores... 487 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 488 489 if (mem->is_Store()) { 490 Node* st_adr = mem->in(MemNode::Address); 491 intptr_t st_offset = 0; 492 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 493 if (st_base == NULL) 494 break; // inscrutable pointer 495 if (st_offset != offset && st_offset != Type::OffsetBot) { 496 const int MAX_STORE = BytesPerLong; 497 if (st_offset >= offset + size_in_bytes || 498 st_offset <= offset - MAX_STORE || 499 st_offset <= offset - mem->as_Store()->memory_size()) { 500 // Success: The offsets are provably independent. 501 // (You may ask, why not just test st_offset != offset and be done? 502 // The answer is that stores of different sizes can co-exist 503 // in the same sequence of RawMem effects. We sometimes initialize 504 // a whole 'tile' of array elements with a single jint or jlong.) 505 mem = mem->in(MemNode::Memory); 506 continue; // (a) advance through independent store memory 507 } 508 } 509 if (st_base != base && 510 detect_ptr_independence(base, alloc, 511 st_base, 512 AllocateNode::Ideal_allocation(st_base, phase), 513 phase)) { 514 // Success: The bases are provably independent. 515 mem = mem->in(MemNode::Memory); 516 continue; // (a) advance through independent store memory 517 } 518 519 // (b) At this point, if the bases or offsets do not agree, we lose, 520 // since we have not managed to prove 'this' and 'mem' independent. 521 if (st_base == base && st_offset == offset) { 522 return mem; // let caller handle steps (c), (d) 523 } 524 525 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 526 InitializeNode* st_init = mem->in(0)->as_Initialize(); 527 AllocateNode* st_alloc = st_init->allocation(); 528 if (st_alloc == NULL) 529 break; // something degenerated 530 bool known_identical = false; 531 bool known_independent = false; 532 if (alloc == st_alloc) 533 known_identical = true; 534 else if (alloc != NULL) 535 known_independent = true; 536 else if (all_controls_dominate(this, st_alloc)) 537 known_independent = true; 538 539 if (known_independent) { 540 // The bases are provably independent: Either they are 541 // manifestly distinct allocations, or else the control 542 // of this load dominates the store's allocation. 543 int alias_idx = phase->C->get_alias_index(adr_type()); 544 if (alias_idx == Compile::AliasIdxRaw) { 545 mem = st_alloc->in(TypeFunc::Memory); 546 } else { 547 mem = st_init->memory(alias_idx); 548 } 549 continue; // (a) advance through independent store memory 550 } 551 552 // (b) at this point, if we are not looking at a store initializing 553 // the same allocation we are loading from, we lose. 554 if (known_identical) { 555 // From caller, can_see_stored_value will consult find_captured_store. 556 return mem; // let caller handle steps (c), (d) 557 } 558 559 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 560 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 561 if (mem->is_Proj() && mem->in(0)->is_Call()) { 562 CallNode *call = mem->in(0)->as_Call(); 563 if (!call->may_modify(addr_t, phase)) { 564 mem = call->in(TypeFunc::Memory); 565 continue; // (a) advance through independent call memory 566 } 567 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 568 mem = mem->in(0)->in(TypeFunc::Memory); 569 continue; // (a) advance through independent MemBar memory 570 } else if (mem->is_ClearArray()) { 571 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 572 // (the call updated 'mem' value) 573 continue; // (a) advance through independent allocation memory 574 } else { 575 // Can not bypass initialization of the instance 576 // we are looking for. 577 return mem; 578 } 579 } else if (mem->is_MergeMem()) { 580 int alias_idx = phase->C->get_alias_index(adr_type()); 581 mem = mem->as_MergeMem()->memory_at(alias_idx); 582 continue; // (a) advance through independent MergeMem memory 583 } 584 } 585 586 // Unless there is an explicit 'continue', we must bail out here, 587 // because 'mem' is an inscrutable memory state (e.g., a call). 588 break; 589 } 590 591 return NULL; // bail out 592 } 593 594 //----------------------calculate_adr_type------------------------------------- 595 // Helper function. Notices when the given type of address hits top or bottom. 596 // Also, asserts a cross-check of the type against the expected address type. 597 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 598 if (t == Type::TOP) return NULL; // does not touch memory any more? 599 #ifdef PRODUCT 600 cross_check = NULL; 601 #else 602 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 603 #endif 604 const TypePtr* tp = t->isa_ptr(); 605 if (tp == NULL) { 606 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 607 return TypePtr::BOTTOM; // touches lots of memory 608 } else { 609 #ifdef ASSERT 610 // %%%% [phh] We don't check the alias index if cross_check is 611 // TypeRawPtr::BOTTOM. Needs to be investigated. 612 if (cross_check != NULL && 613 cross_check != TypePtr::BOTTOM && 614 cross_check != TypeRawPtr::BOTTOM) { 615 // Recheck the alias index, to see if it has changed (due to a bug). 616 Compile* C = Compile::current(); 617 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 618 "must stay in the original alias category"); 619 // The type of the address must be contained in the adr_type, 620 // disregarding "null"-ness. 621 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 622 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 623 assert(cross_check->meet(tp_notnull) == cross_check, 624 "real address must not escape from expected memory type"); 625 } 626 #endif 627 return tp; 628 } 629 } 630 631 //------------------------adr_phi_is_loop_invariant---------------------------- 632 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 633 // loop is loop invariant. Make a quick traversal of Phi and associated 634 // CastPP nodes, looking to see if they are a closed group within the loop. 635 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 636 // The idea is that the phi-nest must boil down to only CastPP nodes 637 // with the same data. This implies that any path into the loop already 638 // includes such a CastPP, and so the original cast, whatever its input, 639 // must be covered by an equivalent cast, with an earlier control input. 640 ResourceMark rm; 641 642 // The loop entry input of the phi should be the unique dominating 643 // node for every Phi/CastPP in the loop. 644 Unique_Node_List closure; 645 closure.push(adr_phi->in(LoopNode::EntryControl)); 646 647 // Add the phi node and the cast to the worklist. 648 Unique_Node_List worklist; 649 worklist.push(adr_phi); 650 if( cast != NULL ){ 651 if( !cast->is_ConstraintCast() ) return false; 652 worklist.push(cast); 653 } 654 655 // Begin recursive walk of phi nodes. 656 while( worklist.size() ){ 657 // Take a node off the worklist 658 Node *n = worklist.pop(); 659 if( !closure.member(n) ){ 660 // Add it to the closure. 661 closure.push(n); 662 // Make a sanity check to ensure we don't waste too much time here. 663 if( closure.size() > 20) return false; 664 // This node is OK if: 665 // - it is a cast of an identical value 666 // - or it is a phi node (then we add its inputs to the worklist) 667 // Otherwise, the node is not OK, and we presume the cast is not invariant 668 if( n->is_ConstraintCast() ){ 669 worklist.push(n->in(1)); 670 } else if( n->is_Phi() ) { 671 for( uint i = 1; i < n->req(); i++ ) { 672 worklist.push(n->in(i)); 673 } 674 } else { 675 return false; 676 } 677 } 678 } 679 680 // Quit when the worklist is empty, and we've found no offending nodes. 681 return true; 682 } 683 684 //------------------------------Ideal_DU_postCCP------------------------------- 685 // Find any cast-away of null-ness and keep its control. Null cast-aways are 686 // going away in this pass and we need to make this memory op depend on the 687 // gating null check. 688 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 689 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 690 } 691 692 // I tried to leave the CastPP's in. This makes the graph more accurate in 693 // some sense; we get to keep around the knowledge that an oop is not-null 694 // after some test. Alas, the CastPP's interfere with GVN (some values are 695 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 696 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 697 // some of the more trivial cases in the optimizer. Removing more useless 698 // Phi's started allowing Loads to illegally float above null checks. I gave 699 // up on this approach. CNC 10/20/2000 700 // This static method may be called not from MemNode (EncodePNode calls it). 701 // Only the control edge of the node 'n' might be updated. 702 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 703 Node *skipped_cast = NULL; 704 // Need a null check? Regular static accesses do not because they are 705 // from constant addresses. Array ops are gated by the range check (which 706 // always includes a NULL check). Just check field ops. 707 if( n->in(MemNode::Control) == NULL ) { 708 // Scan upwards for the highest location we can place this memory op. 709 while( true ) { 710 switch( adr->Opcode() ) { 711 712 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 713 adr = adr->in(AddPNode::Base); 714 continue; 715 716 case Op_DecodeN: // No change to NULL-ness, so peek thru 717 adr = adr->in(1); 718 continue; 719 720 case Op_CastPP: 721 // If the CastPP is useless, just peek on through it. 722 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 723 // Remember the cast that we've peeked though. If we peek 724 // through more than one, then we end up remembering the highest 725 // one, that is, if in a loop, the one closest to the top. 726 skipped_cast = adr; 727 adr = adr->in(1); 728 continue; 729 } 730 // CastPP is going away in this pass! We need this memory op to be 731 // control-dependent on the test that is guarding the CastPP. 732 ccp->hash_delete(n); 733 n->set_req(MemNode::Control, adr->in(0)); 734 ccp->hash_insert(n); 735 return n; 736 737 case Op_Phi: 738 // Attempt to float above a Phi to some dominating point. 739 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 740 // If we've already peeked through a Cast (which could have set the 741 // control), we can't float above a Phi, because the skipped Cast 742 // may not be loop invariant. 743 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 744 adr = adr->in(1); 745 continue; 746 } 747 } 748 749 // Intentional fallthrough! 750 751 // No obvious dominating point. The mem op is pinned below the Phi 752 // by the Phi itself. If the Phi goes away (no true value is merged) 753 // then the mem op can float, but not indefinitely. It must be pinned 754 // behind the controls leading to the Phi. 755 case Op_CheckCastPP: 756 // These usually stick around to change address type, however a 757 // useless one can be elided and we still need to pick up a control edge 758 if (adr->in(0) == NULL) { 759 // This CheckCastPP node has NO control and is likely useless. But we 760 // need check further up the ancestor chain for a control input to keep 761 // the node in place. 4959717. 762 skipped_cast = adr; 763 adr = adr->in(1); 764 continue; 765 } 766 ccp->hash_delete(n); 767 n->set_req(MemNode::Control, adr->in(0)); 768 ccp->hash_insert(n); 769 return n; 770 771 // List of "safe" opcodes; those that implicitly block the memory 772 // op below any null check. 773 case Op_CastX2P: // no null checks on native pointers 774 case Op_Parm: // 'this' pointer is not null 775 case Op_LoadP: // Loading from within a klass 776 case Op_LoadN: // Loading from within a klass 777 case Op_LoadKlass: // Loading from within a klass 778 case Op_LoadNKlass: // Loading from within a klass 779 case Op_ConP: // Loading from a klass 780 case Op_ConN: // Loading from a klass 781 case Op_CreateEx: // Sucking up the guts of an exception oop 782 case Op_Con: // Reading from TLS 783 case Op_CMoveP: // CMoveP is pinned 784 case Op_CMoveN: // CMoveN is pinned 785 break; // No progress 786 787 case Op_Proj: // Direct call to an allocation routine 788 case Op_SCMemProj: // Memory state from store conditional ops 789 #ifdef ASSERT 790 { 791 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 792 const Node* call = adr->in(0); 793 if (call->is_CallJava()) { 794 const CallJavaNode* call_java = call->as_CallJava(); 795 const TypeTuple *r = call_java->tf()->range(); 796 assert(r->cnt() > TypeFunc::Parms, "must return value"); 797 const Type* ret_type = r->field_at(TypeFunc::Parms); 798 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 799 // We further presume that this is one of 800 // new_instance_Java, new_array_Java, or 801 // the like, but do not assert for this. 802 } else if (call->is_Allocate()) { 803 // similar case to new_instance_Java, etc. 804 } else if (!call->is_CallLeaf()) { 805 // Projections from fetch_oop (OSR) are allowed as well. 806 ShouldNotReachHere(); 807 } 808 } 809 #endif 810 break; 811 default: 812 ShouldNotReachHere(); 813 } 814 break; 815 } 816 } 817 818 return NULL; // No progress 819 } 820 821 822 //============================================================================= 823 uint LoadNode::size_of() const { return sizeof(*this); } 824 uint LoadNode::cmp( const Node &n ) const 825 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 826 const Type *LoadNode::bottom_type() const { return _type; } 827 uint LoadNode::ideal_reg() const { 828 return Matcher::base2reg[_type->base()]; 829 } 830 831 #ifndef PRODUCT 832 void LoadNode::dump_spec(outputStream *st) const { 833 MemNode::dump_spec(st); 834 if( !Verbose && !WizardMode ) { 835 // standard dump does this in Verbose and WizardMode 836 st->print(" #"); _type->dump_on(st); 837 } 838 } 839 #endif 840 841 #ifdef ASSERT 842 //----------------------------is_immutable_value------------------------------- 843 // Helper function to allow a raw load without control edge for some cases 844 bool LoadNode::is_immutable_value(Node* adr) { 845 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 846 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 847 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 848 in_bytes(JavaThread::osthread_offset()))); 849 } 850 #endif 851 852 //----------------------------LoadNode::make----------------------------------- 853 // Polymorphic factory method: 854 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 855 Compile* C = gvn.C; 856 857 // sanity check the alias category against the created node type 858 assert(!(adr_type->isa_oopptr() && 859 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 860 "use LoadKlassNode instead"); 861 assert(!(adr_type->isa_aryptr() && 862 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 863 "use LoadRangeNode instead"); 864 // Check control edge of raw loads 865 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 866 // oop will be recorded in oop map if load crosses safepoint 867 rt->isa_oopptr() || is_immutable_value(adr), 868 "raw memory operations should have control edge"); 869 switch (bt) { 870 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 871 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 872 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 873 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 874 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 875 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 876 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 877 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 878 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 879 case T_OBJECT: 880 #ifdef _LP64 881 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 882 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 883 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 884 } else 885 #endif 886 { 887 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 888 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 889 } 890 } 891 ShouldNotReachHere(); 892 return (LoadNode*)NULL; 893 } 894 895 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 896 bool require_atomic = true; 897 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 898 } 899 900 901 902 903 //------------------------------hash------------------------------------------- 904 uint LoadNode::hash() const { 905 // unroll addition of interesting fields 906 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 907 } 908 909 //---------------------------can_see_stored_value------------------------------ 910 // This routine exists to make sure this set of tests is done the same 911 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 912 // will change the graph shape in a way which makes memory alive twice at the 913 // same time (uses the Oracle model of aliasing), then some 914 // LoadXNode::Identity will fold things back to the equivalence-class model 915 // of aliasing. 916 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 917 Node* ld_adr = in(MemNode::Address); 918 919 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 920 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 921 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 922 atp->field() != NULL && !atp->field()->is_volatile()) { 923 uint alias_idx = atp->index(); 924 bool final = atp->field()->is_final(); 925 Node* result = NULL; 926 Node* current = st; 927 // Skip through chains of MemBarNodes checking the MergeMems for 928 // new states for the slice of this load. Stop once any other 929 // kind of node is encountered. Loads from final memory can skip 930 // through any kind of MemBar but normal loads shouldn't skip 931 // through MemBarAcquire since the could allow them to move out of 932 // a synchronized region. 933 while (current->is_Proj()) { 934 int opc = current->in(0)->Opcode(); 935 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) || 936 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder || 937 opc == Op_MemBarReleaseLock) { 938 Node* mem = current->in(0)->in(TypeFunc::Memory); 939 if (mem->is_MergeMem()) { 940 MergeMemNode* merge = mem->as_MergeMem(); 941 Node* new_st = merge->memory_at(alias_idx); 942 if (new_st == merge->base_memory()) { 943 // Keep searching 944 current = merge->base_memory(); 945 continue; 946 } 947 // Save the new memory state for the slice and fall through 948 // to exit. 949 result = new_st; 950 } 951 } 952 break; 953 } 954 if (result != NULL) { 955 st = result; 956 } 957 } 958 959 960 // Loop around twice in the case Load -> Initialize -> Store. 961 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 962 for (int trip = 0; trip <= 1; trip++) { 963 964 if (st->is_Store()) { 965 Node* st_adr = st->in(MemNode::Address); 966 if (!phase->eqv(st_adr, ld_adr)) { 967 // Try harder before giving up... Match raw and non-raw pointers. 968 intptr_t st_off = 0; 969 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 970 if (alloc == NULL) return NULL; 971 intptr_t ld_off = 0; 972 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 973 if (alloc != allo2) return NULL; 974 if (ld_off != st_off) return NULL; 975 // At this point we have proven something like this setup: 976 // A = Allocate(...) 977 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 978 // S = StoreQ(, AddP(, A.Parm , #Off), V) 979 // (Actually, we haven't yet proven the Q's are the same.) 980 // In other words, we are loading from a casted version of 981 // the same pointer-and-offset that we stored to. 982 // Thus, we are able to replace L by V. 983 } 984 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 985 if (store_Opcode() != st->Opcode()) 986 return NULL; 987 return st->in(MemNode::ValueIn); 988 } 989 990 intptr_t offset = 0; // scratch 991 992 // A load from a freshly-created object always returns zero. 993 // (This can happen after LoadNode::Ideal resets the load's memory input 994 // to find_captured_store, which returned InitializeNode::zero_memory.) 995 if (st->is_Proj() && st->in(0)->is_Allocate() && 996 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 997 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 998 // return a zero value for the load's basic type 999 // (This is one of the few places where a generic PhaseTransform 1000 // can create new nodes. Think of it as lazily manifesting 1001 // virtually pre-existing constants.) 1002 return phase->zerocon(memory_type()); 1003 } 1004 1005 // A load from an initialization barrier can match a captured store. 1006 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1007 InitializeNode* init = st->in(0)->as_Initialize(); 1008 AllocateNode* alloc = init->allocation(); 1009 if (alloc != NULL && 1010 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 1011 // examine a captured store value 1012 st = init->find_captured_store(offset, memory_size(), phase); 1013 if (st != NULL) 1014 continue; // take one more trip around 1015 } 1016 } 1017 1018 break; 1019 } 1020 1021 return NULL; 1022 } 1023 1024 //----------------------is_instance_field_load_with_local_phi------------------ 1025 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1026 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 1027 in(MemNode::Address)->is_AddP() ) { 1028 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 1029 // Only instances. 1030 if( t_oop != NULL && t_oop->is_known_instance_field() && 1031 t_oop->offset() != Type::OffsetBot && 1032 t_oop->offset() != Type::OffsetTop) { 1033 return true; 1034 } 1035 } 1036 return false; 1037 } 1038 1039 //------------------------------Identity--------------------------------------- 1040 // Loads are identity if previous store is to same address 1041 Node *LoadNode::Identity( PhaseTransform *phase ) { 1042 // If the previous store-maker is the right kind of Store, and the store is 1043 // to the same address, then we are equal to the value stored. 1044 Node* mem = in(MemNode::Memory); 1045 Node* value = can_see_stored_value(mem, phase); 1046 if( value ) { 1047 // byte, short & char stores truncate naturally. 1048 // A load has to load the truncated value which requires 1049 // some sort of masking operation and that requires an 1050 // Ideal call instead of an Identity call. 1051 if (memory_size() < BytesPerInt) { 1052 // If the input to the store does not fit with the load's result type, 1053 // it must be truncated via an Ideal call. 1054 if (!phase->type(value)->higher_equal(phase->type(this))) 1055 return this; 1056 } 1057 // (This works even when value is a Con, but LoadNode::Value 1058 // usually runs first, producing the singleton type of the Con.) 1059 return value; 1060 } 1061 1062 // Search for an existing data phi which was generated before for the same 1063 // instance's field to avoid infinite generation of phis in a loop. 1064 Node *region = mem->in(0); 1065 if (is_instance_field_load_with_local_phi(region)) { 1066 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 1067 int this_index = phase->C->get_alias_index(addr_t); 1068 int this_offset = addr_t->offset(); 1069 int this_id = addr_t->is_oopptr()->instance_id(); 1070 const Type* this_type = bottom_type(); 1071 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1072 Node* phi = region->fast_out(i); 1073 if (phi->is_Phi() && phi != mem && 1074 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 1075 return phi; 1076 } 1077 } 1078 } 1079 1080 return this; 1081 } 1082 1083 1084 // Returns true if the AliasType refers to the field that holds the 1085 // cached box array. Currently only handles the IntegerCache case. 1086 static bool is_autobox_cache(Compile::AliasType* atp) { 1087 if (atp != NULL && atp->field() != NULL) { 1088 ciField* field = atp->field(); 1089 ciSymbol* klass = field->holder()->name(); 1090 if (field->name() == ciSymbol::cache_field_name() && 1091 field->holder()->uses_default_loader() && 1092 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1093 return true; 1094 } 1095 } 1096 return false; 1097 } 1098 1099 // Fetch the base value in the autobox array 1100 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1101 if (atp != NULL && atp->field() != NULL) { 1102 ciField* field = atp->field(); 1103 ciSymbol* klass = field->holder()->name(); 1104 if (field->name() == ciSymbol::cache_field_name() && 1105 field->holder()->uses_default_loader() && 1106 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1107 assert(field->is_constant(), "what?"); 1108 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1109 // Fetch the box object at the base of the array and get its value 1110 ciInstance* box = array->obj_at(0)->as_instance(); 1111 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1112 if (ik->nof_nonstatic_fields() == 1) { 1113 // This should be true nonstatic_field_at requires calling 1114 // nof_nonstatic_fields so check it anyway 1115 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1116 cache_offset = c.as_int(); 1117 } 1118 return true; 1119 } 1120 } 1121 return false; 1122 } 1123 1124 // Returns true if the AliasType refers to the value field of an 1125 // autobox object. Currently only handles Integer. 1126 static bool is_autobox_object(Compile::AliasType* atp) { 1127 if (atp != NULL && atp->field() != NULL) { 1128 ciField* field = atp->field(); 1129 ciSymbol* klass = field->holder()->name(); 1130 if (field->name() == ciSymbol::value_name() && 1131 field->holder()->uses_default_loader() && 1132 klass == ciSymbol::java_lang_Integer()) { 1133 return true; 1134 } 1135 } 1136 return false; 1137 } 1138 1139 1140 // We're loading from an object which has autobox behaviour. 1141 // If this object is result of a valueOf call we'll have a phi 1142 // merging a newly allocated object and a load from the cache. 1143 // We want to replace this load with the original incoming 1144 // argument to the valueOf call. 1145 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1146 Node* base = in(Address)->in(AddPNode::Base); 1147 if (base->is_Phi() && base->req() == 3) { 1148 AllocateNode* allocation = NULL; 1149 int allocation_index = -1; 1150 int load_index = -1; 1151 for (uint i = 1; i < base->req(); i++) { 1152 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1153 if (allocation != NULL) { 1154 allocation_index = i; 1155 load_index = 3 - allocation_index; 1156 break; 1157 } 1158 } 1159 bool has_load = ( allocation != NULL && 1160 (base->in(load_index)->is_Load() || 1161 base->in(load_index)->is_DecodeN() && 1162 base->in(load_index)->in(1)->is_Load()) ); 1163 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1164 // Push the loads from the phi that comes from valueOf up 1165 // through it to allow elimination of the loads and the recovery 1166 // of the original value. 1167 Node* mem_phi = in(Memory); 1168 Node* offset = in(Address)->in(AddPNode::Offset); 1169 Node* region = base->in(0); 1170 1171 Node* in1 = clone(); 1172 Node* in1_addr = in1->in(Address)->clone(); 1173 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1174 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1175 in1_addr->set_req(AddPNode::Offset, offset); 1176 in1->set_req(0, region->in(allocation_index)); 1177 in1->set_req(Address, in1_addr); 1178 in1->set_req(Memory, mem_phi->in(allocation_index)); 1179 1180 Node* in2 = clone(); 1181 Node* in2_addr = in2->in(Address)->clone(); 1182 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1183 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1184 in2_addr->set_req(AddPNode::Offset, offset); 1185 in2->set_req(0, region->in(load_index)); 1186 in2->set_req(Address, in2_addr); 1187 in2->set_req(Memory, mem_phi->in(load_index)); 1188 1189 in1_addr = phase->transform(in1_addr); 1190 in1 = phase->transform(in1); 1191 in2_addr = phase->transform(in2_addr); 1192 in2 = phase->transform(in2); 1193 1194 PhiNode* result = PhiNode::make_blank(region, this); 1195 result->set_req(allocation_index, in1); 1196 result->set_req(load_index, in2); 1197 return result; 1198 } 1199 } else if (base->is_Load() || 1200 base->is_DecodeN() && base->in(1)->is_Load()) { 1201 if (base->is_DecodeN()) { 1202 // Get LoadN node which loads cached Integer object 1203 base = base->in(1); 1204 } 1205 // Eliminate the load of Integer.value for integers from the cache 1206 // array by deriving the value from the index into the array. 1207 // Capture the offset of the load and then reverse the computation. 1208 Node* load_base = base->in(Address)->in(AddPNode::Base); 1209 if (load_base->is_DecodeN()) { 1210 // Get LoadN node which loads IntegerCache.cache field 1211 load_base = load_base->in(1); 1212 } 1213 if (load_base != NULL) { 1214 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1215 intptr_t cache_offset; 1216 int shift = -1; 1217 Node* cache = NULL; 1218 if (is_autobox_cache(atp)) { 1219 shift = exact_log2(type2aelembytes(T_OBJECT)); 1220 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1221 } 1222 if (cache != NULL && base->in(Address)->is_AddP()) { 1223 Node* elements[4]; 1224 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1225 int cache_low; 1226 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1227 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1228 // Add up all the offsets making of the address of the load 1229 Node* result = elements[0]; 1230 for (int i = 1; i < count; i++) { 1231 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1232 } 1233 // Remove the constant offset from the address and then 1234 // remove the scaling of the offset to recover the original index. 1235 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1236 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1237 // Peel the shift off directly but wrap it in a dummy node 1238 // since Ideal can't return existing nodes 1239 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1240 } else { 1241 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1242 } 1243 #ifdef _LP64 1244 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1245 #endif 1246 return result; 1247 } 1248 } 1249 } 1250 } 1251 return NULL; 1252 } 1253 1254 //------------------------------split_through_phi------------------------------ 1255 // Split instance field load through Phi. 1256 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1257 Node* mem = in(MemNode::Memory); 1258 Node* address = in(MemNode::Address); 1259 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1260 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1261 1262 assert(mem->is_Phi() && (t_oop != NULL) && 1263 t_oop->is_known_instance_field(), "invalide conditions"); 1264 1265 Node *region = mem->in(0); 1266 if (region == NULL) { 1267 return NULL; // Wait stable graph 1268 } 1269 uint cnt = mem->req(); 1270 for (uint i = 1; i < cnt; i++) { 1271 Node* rc = region->in(i); 1272 if (rc == NULL || phase->type(rc) == Type::TOP) 1273 return NULL; // Wait stable graph 1274 Node *in = mem->in(i); 1275 if (in == NULL) { 1276 return NULL; // Wait stable graph 1277 } 1278 } 1279 // Check for loop invariant. 1280 if (cnt == 3) { 1281 for (uint i = 1; i < cnt; i++) { 1282 Node *in = mem->in(i); 1283 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1284 if (m == mem) { 1285 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1286 return this; 1287 } 1288 } 1289 } 1290 // Split through Phi (see original code in loopopts.cpp). 1291 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1292 1293 // Do nothing here if Identity will find a value 1294 // (to avoid infinite chain of value phis generation). 1295 if (!phase->eqv(this, this->Identity(phase))) 1296 return NULL; 1297 1298 // Skip the split if the region dominates some control edge of the address. 1299 if (!MemNode::all_controls_dominate(address, region)) 1300 return NULL; 1301 1302 const Type* this_type = this->bottom_type(); 1303 int this_index = phase->C->get_alias_index(addr_t); 1304 int this_offset = addr_t->offset(); 1305 int this_iid = addr_t->is_oopptr()->instance_id(); 1306 PhaseIterGVN *igvn = phase->is_IterGVN(); 1307 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1308 for (uint i = 1; i < region->req(); i++) { 1309 Node *x; 1310 Node* the_clone = NULL; 1311 if (region->in(i) == phase->C->top()) { 1312 x = phase->C->top(); // Dead path? Use a dead data op 1313 } else { 1314 x = this->clone(); // Else clone up the data op 1315 the_clone = x; // Remember for possible deletion. 1316 // Alter data node to use pre-phi inputs 1317 if (this->in(0) == region) { 1318 x->set_req(0, region->in(i)); 1319 } else { 1320 x->set_req(0, NULL); 1321 } 1322 for (uint j = 1; j < this->req(); j++) { 1323 Node *in = this->in(j); 1324 if (in->is_Phi() && in->in(0) == region) 1325 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 1326 } 1327 } 1328 // Check for a 'win' on some paths 1329 const Type *t = x->Value(igvn); 1330 1331 bool singleton = t->singleton(); 1332 1333 // See comments in PhaseIdealLoop::split_thru_phi(). 1334 if (singleton && t == Type::TOP) { 1335 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1336 } 1337 1338 if (singleton) { 1339 x = igvn->makecon(t); 1340 } else { 1341 // We now call Identity to try to simplify the cloned node. 1342 // Note that some Identity methods call phase->type(this). 1343 // Make sure that the type array is big enough for 1344 // our new node, even though we may throw the node away. 1345 // (This tweaking with igvn only works because x is a new node.) 1346 igvn->set_type(x, t); 1347 // If x is a TypeNode, capture any more-precise type permanently into Node 1348 // otherwise it will be not updated during igvn->transform since 1349 // igvn->type(x) is set to x->Value() already. 1350 x->raise_bottom_type(t); 1351 Node *y = x->Identity(igvn); 1352 if (y != x) { 1353 x = y; 1354 } else { 1355 y = igvn->hash_find(x); 1356 if (y) { 1357 x = y; 1358 } else { 1359 // Else x is a new node we are keeping 1360 // We do not need register_new_node_with_optimizer 1361 // because set_type has already been called. 1362 igvn->_worklist.push(x); 1363 } 1364 } 1365 } 1366 if (x != the_clone && the_clone != NULL) 1367 igvn->remove_dead_node(the_clone); 1368 phi->set_req(i, x); 1369 } 1370 // Record Phi 1371 igvn->register_new_node_with_optimizer(phi); 1372 return phi; 1373 } 1374 1375 //------------------------------Ideal------------------------------------------ 1376 // If the load is from Field memory and the pointer is non-null, we can 1377 // zero out the control input. 1378 // If the offset is constant and the base is an object allocation, 1379 // try to hook me up to the exact initializing store. 1380 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1381 Node* p = MemNode::Ideal_common(phase, can_reshape); 1382 if (p) return (p == NodeSentinel) ? NULL : p; 1383 1384 Node* ctrl = in(MemNode::Control); 1385 Node* address = in(MemNode::Address); 1386 1387 // Skip up past a SafePoint control. Cannot do this for Stores because 1388 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1389 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1390 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1391 ctrl = ctrl->in(0); 1392 set_req(MemNode::Control,ctrl); 1393 } 1394 1395 intptr_t ignore = 0; 1396 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1397 if (base != NULL 1398 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1399 // Check for useless control edge in some common special cases 1400 if (in(MemNode::Control) != NULL 1401 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1402 && all_controls_dominate(base, phase->C->start())) { 1403 // A method-invariant, non-null address (constant or 'this' argument). 1404 set_req(MemNode::Control, NULL); 1405 } 1406 1407 if (EliminateAutoBox && can_reshape) { 1408 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); 1409 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1410 if (is_autobox_object(atp)) { 1411 Node* result = eliminate_autobox(phase); 1412 if (result != NULL) return result; 1413 } 1414 } 1415 } 1416 1417 Node* mem = in(MemNode::Memory); 1418 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1419 1420 if (addr_t != NULL) { 1421 // try to optimize our memory input 1422 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1423 if (opt_mem != mem) { 1424 set_req(MemNode::Memory, opt_mem); 1425 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1426 return this; 1427 } 1428 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1429 if (can_reshape && opt_mem->is_Phi() && 1430 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1431 PhaseIterGVN *igvn = phase->is_IterGVN(); 1432 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1433 // Delay this transformation until memory Phi is processed. 1434 phase->is_IterGVN()->_worklist.push(this); 1435 return NULL; 1436 } 1437 // Split instance field load through Phi. 1438 Node* result = split_through_phi(phase); 1439 if (result != NULL) return result; 1440 } 1441 } 1442 1443 // Check for prior store with a different base or offset; make Load 1444 // independent. Skip through any number of them. Bail out if the stores 1445 // are in an endless dead cycle and report no progress. This is a key 1446 // transform for Reflection. However, if after skipping through the Stores 1447 // we can't then fold up against a prior store do NOT do the transform as 1448 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1449 // array memory alive twice: once for the hoisted Load and again after the 1450 // bypassed Store. This situation only works if EVERYBODY who does 1451 // anti-dependence work knows how to bypass. I.e. we need all 1452 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1453 // the alias index stuff. So instead, peek through Stores and IFF we can 1454 // fold up, do so. 1455 Node* prev_mem = find_previous_store(phase); 1456 // Steps (a), (b): Walk past independent stores to find an exact match. 1457 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1458 // (c) See if we can fold up on the spot, but don't fold up here. 1459 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1460 // just return a prior value, which is done by Identity calls. 1461 if (can_see_stored_value(prev_mem, phase)) { 1462 // Make ready for step (d): 1463 set_req(MemNode::Memory, prev_mem); 1464 return this; 1465 } 1466 } 1467 1468 return NULL; // No further progress 1469 } 1470 1471 // Helper to recognize certain Klass fields which are invariant across 1472 // some group of array types (e.g., int[] or all T[] where T < Object). 1473 const Type* 1474 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1475 ciKlass* klass) const { 1476 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1477 // The field is Klass::_modifier_flags. Return its (constant) value. 1478 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1479 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1480 return TypeInt::make(klass->modifier_flags()); 1481 } 1482 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1483 // The field is Klass::_access_flags. Return its (constant) value. 1484 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1485 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1486 return TypeInt::make(klass->access_flags()); 1487 } 1488 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1489 // The field is Klass::_layout_helper. Return its constant value if known. 1490 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1491 return TypeInt::make(klass->layout_helper()); 1492 } 1493 1494 // No match. 1495 return NULL; 1496 } 1497 1498 //------------------------------Value----------------------------------------- 1499 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1500 // Either input is TOP ==> the result is TOP 1501 Node* mem = in(MemNode::Memory); 1502 const Type *t1 = phase->type(mem); 1503 if (t1 == Type::TOP) return Type::TOP; 1504 Node* adr = in(MemNode::Address); 1505 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1506 if (tp == NULL || tp->empty()) return Type::TOP; 1507 int off = tp->offset(); 1508 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1509 Compile* C = phase->C; 1510 1511 // Try to guess loaded type from pointer type 1512 if (tp->base() == Type::AryPtr) { 1513 const Type *t = tp->is_aryptr()->elem(); 1514 // Don't do this for integer types. There is only potential profit if 1515 // the element type t is lower than _type; that is, for int types, if _type is 1516 // more restrictive than t. This only happens here if one is short and the other 1517 // char (both 16 bits), and in those cases we've made an intentional decision 1518 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1519 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1520 // 1521 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1522 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1523 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1524 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1525 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1526 // In fact, that could have been the original type of p1, and p1 could have 1527 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1528 // expression (LShiftL quux 3) independently optimized to the constant 8. 1529 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1530 && (_type->isa_vect() == NULL) 1531 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1532 // t might actually be lower than _type, if _type is a unique 1533 // concrete subclass of abstract class t. 1534 // Make sure the reference is not into the header, by comparing 1535 // the offset against the offset of the start of the array's data. 1536 // Different array types begin at slightly different offsets (12 vs. 16). 1537 // We choose T_BYTE as an example base type that is least restrictive 1538 // as to alignment, which will therefore produce the smallest 1539 // possible base offset. 1540 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1541 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1542 const Type* jt = t->join(_type); 1543 // In any case, do not allow the join, per se, to empty out the type. 1544 if (jt->empty() && !t->empty()) { 1545 // This can happen if a interface-typed array narrows to a class type. 1546 jt = _type; 1547 } 1548 1549 if (EliminateAutoBox && adr->is_AddP()) { 1550 // The pointers in the autobox arrays are always non-null 1551 Node* base = adr->in(AddPNode::Base); 1552 if (base != NULL && 1553 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { 1554 Compile::AliasType* atp = C->alias_type(base->adr_type()); 1555 if (is_autobox_cache(atp)) { 1556 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1557 } 1558 } 1559 } 1560 return jt; 1561 } 1562 } 1563 } else if (tp->base() == Type::InstPtr) { 1564 ciEnv* env = C->env(); 1565 const TypeInstPtr* tinst = tp->is_instptr(); 1566 ciKlass* klass = tinst->klass(); 1567 assert( off != Type::OffsetBot || 1568 // arrays can be cast to Objects 1569 tp->is_oopptr()->klass()->is_java_lang_Object() || 1570 // unsafe field access may not have a constant offset 1571 C->has_unsafe_access(), 1572 "Field accesses must be precise" ); 1573 // For oop loads, we expect the _type to be precise 1574 if (klass == env->String_klass() && 1575 adr->is_AddP() && off != Type::OffsetBot) { 1576 // For constant Strings treat the final fields as compile time constants. 1577 Node* base = adr->in(AddPNode::Base); 1578 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1579 if (t != NULL && t->singleton()) { 1580 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1581 if (field != NULL && field->is_final()) { 1582 ciObject* string = t->const_oop(); 1583 ciConstant constant = string->as_instance()->field_value(field); 1584 if (constant.basic_type() == T_INT) { 1585 return TypeInt::make(constant.as_int()); 1586 } else if (constant.basic_type() == T_ARRAY) { 1587 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1588 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1589 } else { 1590 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1591 } 1592 } 1593 } 1594 } 1595 } 1596 // Optimizations for constant objects 1597 ciObject* const_oop = tinst->const_oop(); 1598 if (const_oop != NULL) { 1599 // For constant CallSites treat the target field as a compile time constant. 1600 if (const_oop->is_call_site()) { 1601 ciCallSite* call_site = const_oop->as_call_site(); 1602 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1603 if (field != NULL && field->is_call_site_target()) { 1604 ciMethodHandle* target = call_site->get_target(); 1605 if (target != NULL) { // just in case 1606 ciConstant constant(T_OBJECT, target); 1607 const Type* t; 1608 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1609 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1610 } else { 1611 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1612 } 1613 // Add a dependence for invalidation of the optimization. 1614 if (!call_site->is_constant_call_site()) { 1615 C->dependencies()->assert_call_site_target_value(call_site, target); 1616 } 1617 return t; 1618 } 1619 } 1620 } 1621 } 1622 } else if (tp->base() == Type::KlassPtr) { 1623 assert( off != Type::OffsetBot || 1624 // arrays can be cast to Objects 1625 tp->is_klassptr()->klass()->is_java_lang_Object() || 1626 // also allow array-loading from the primary supertype 1627 // array during subtype checks 1628 Opcode() == Op_LoadKlass, 1629 "Field accesses must be precise" ); 1630 // For klass/static loads, we expect the _type to be precise 1631 } 1632 1633 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1634 if (tkls != NULL && !StressReflectiveCode) { 1635 ciKlass* klass = tkls->klass(); 1636 if (klass->is_loaded() && tkls->klass_is_exact()) { 1637 // We are loading a field from a Klass metaobject whose identity 1638 // is known at compile time (the type is "exact" or "precise"). 1639 // Check for fields we know are maintained as constants by the VM. 1640 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1641 // The field is Klass::_super_check_offset. Return its (constant) value. 1642 // (Folds up type checking code.) 1643 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1644 return TypeInt::make(klass->super_check_offset()); 1645 } 1646 // Compute index into primary_supers array 1647 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(klassOop); 1648 // Check for overflowing; use unsigned compare to handle the negative case. 1649 if( depth < ciKlass::primary_super_limit() ) { 1650 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1651 // (Folds up type checking code.) 1652 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1653 ciKlass *ss = klass->super_of_depth(depth); 1654 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1655 } 1656 const Type* aift = load_array_final_field(tkls, klass); 1657 if (aift != NULL) return aift; 1658 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) 1659 && klass->is_array_klass()) { 1660 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1661 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1662 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1663 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1664 } 1665 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1666 // The field is Klass::_java_mirror. Return its (constant) value. 1667 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1668 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1669 return TypeInstPtr::make(klass->java_mirror()); 1670 } 1671 } 1672 1673 // We can still check if we are loading from the primary_supers array at a 1674 // shallow enough depth. Even though the klass is not exact, entries less 1675 // than or equal to its super depth are correct. 1676 if (klass->is_loaded() ) { 1677 ciType *inner = klass->klass(); 1678 while( inner->is_obj_array_klass() ) 1679 inner = inner->as_obj_array_klass()->base_element_type(); 1680 if( inner->is_instance_klass() && 1681 !inner->as_instance_klass()->flags().is_interface() ) { 1682 // Compute index into primary_supers array 1683 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(klassOop); 1684 // Check for overflowing; use unsigned compare to handle the negative case. 1685 if( depth < ciKlass::primary_super_limit() && 1686 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1687 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1688 // (Folds up type checking code.) 1689 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1690 ciKlass *ss = klass->super_of_depth(depth); 1691 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1692 } 1693 } 1694 } 1695 1696 // If the type is enough to determine that the thing is not an array, 1697 // we can give the layout_helper a positive interval type. 1698 // This will help short-circuit some reflective code. 1699 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1700 && !klass->is_array_klass() // not directly typed as an array 1701 && !klass->is_interface() // specifically not Serializable & Cloneable 1702 && !klass->is_java_lang_Object() // not the supertype of all T[] 1703 ) { 1704 // Note: When interfaces are reliable, we can narrow the interface 1705 // test to (klass != Serializable && klass != Cloneable). 1706 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1707 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1708 // The key property of this type is that it folds up tests 1709 // for array-ness, since it proves that the layout_helper is positive. 1710 // Thus, a generic value like the basic object layout helper works fine. 1711 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1712 } 1713 } 1714 1715 // If we are loading from a freshly-allocated object, produce a zero, 1716 // if the load is provably beyond the header of the object. 1717 // (Also allow a variable load from a fresh array to produce zero.) 1718 const TypeOopPtr *tinst = tp->isa_oopptr(); 1719 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1720 if (ReduceFieldZeroing || is_instance) { 1721 Node* value = can_see_stored_value(mem,phase); 1722 if (value != NULL && value->is_Con()) { 1723 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1724 return value->bottom_type(); 1725 } 1726 } 1727 1728 if (is_instance) { 1729 // If we have an instance type and our memory input is the 1730 // programs's initial memory state, there is no matching store, 1731 // so just return a zero of the appropriate type 1732 Node *mem = in(MemNode::Memory); 1733 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1734 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1735 return Type::get_zero_type(_type->basic_type()); 1736 } 1737 } 1738 return _type; 1739 } 1740 1741 //------------------------------match_edge------------------------------------- 1742 // Do we Match on this edge index or not? Match only the address. 1743 uint LoadNode::match_edge(uint idx) const { 1744 return idx == MemNode::Address; 1745 } 1746 1747 //--------------------------LoadBNode::Ideal-------------------------------------- 1748 // 1749 // If the previous store is to the same address as this load, 1750 // and the value stored was larger than a byte, replace this load 1751 // with the value stored truncated to a byte. If no truncation is 1752 // needed, the replacement is done in LoadNode::Identity(). 1753 // 1754 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1755 Node* mem = in(MemNode::Memory); 1756 Node* value = can_see_stored_value(mem,phase); 1757 if( value && !phase->type(value)->higher_equal( _type ) ) { 1758 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1759 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1760 } 1761 // Identity call will handle the case where truncation is not needed. 1762 return LoadNode::Ideal(phase, can_reshape); 1763 } 1764 1765 const Type* LoadBNode::Value(PhaseTransform *phase) const { 1766 Node* mem = in(MemNode::Memory); 1767 Node* value = can_see_stored_value(mem,phase); 1768 if (value != NULL && value->is_Con() && 1769 !value->bottom_type()->higher_equal(_type)) { 1770 // If the input to the store does not fit with the load's result type, 1771 // it must be truncated. We can't delay until Ideal call since 1772 // a singleton Value is needed for split_thru_phi optimization. 1773 int con = value->get_int(); 1774 return TypeInt::make((con << 24) >> 24); 1775 } 1776 return LoadNode::Value(phase); 1777 } 1778 1779 //--------------------------LoadUBNode::Ideal------------------------------------- 1780 // 1781 // If the previous store is to the same address as this load, 1782 // and the value stored was larger than a byte, replace this load 1783 // with the value stored truncated to a byte. If no truncation is 1784 // needed, the replacement is done in LoadNode::Identity(). 1785 // 1786 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1787 Node* mem = in(MemNode::Memory); 1788 Node* value = can_see_stored_value(mem, phase); 1789 if (value && !phase->type(value)->higher_equal(_type)) 1790 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); 1791 // Identity call will handle the case where truncation is not needed. 1792 return LoadNode::Ideal(phase, can_reshape); 1793 } 1794 1795 const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1796 Node* mem = in(MemNode::Memory); 1797 Node* value = can_see_stored_value(mem,phase); 1798 if (value != NULL && value->is_Con() && 1799 !value->bottom_type()->higher_equal(_type)) { 1800 // If the input to the store does not fit with the load's result type, 1801 // it must be truncated. We can't delay until Ideal call since 1802 // a singleton Value is needed for split_thru_phi optimization. 1803 int con = value->get_int(); 1804 return TypeInt::make(con & 0xFF); 1805 } 1806 return LoadNode::Value(phase); 1807 } 1808 1809 //--------------------------LoadUSNode::Ideal------------------------------------- 1810 // 1811 // If the previous store is to the same address as this load, 1812 // and the value stored was larger than a char, replace this load 1813 // with the value stored truncated to a char. If no truncation is 1814 // needed, the replacement is done in LoadNode::Identity(). 1815 // 1816 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1817 Node* mem = in(MemNode::Memory); 1818 Node* value = can_see_stored_value(mem,phase); 1819 if( value && !phase->type(value)->higher_equal( _type ) ) 1820 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1821 // Identity call will handle the case where truncation is not needed. 1822 return LoadNode::Ideal(phase, can_reshape); 1823 } 1824 1825 const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1826 Node* mem = in(MemNode::Memory); 1827 Node* value = can_see_stored_value(mem,phase); 1828 if (value != NULL && value->is_Con() && 1829 !value->bottom_type()->higher_equal(_type)) { 1830 // If the input to the store does not fit with the load's result type, 1831 // it must be truncated. We can't delay until Ideal call since 1832 // a singleton Value is needed for split_thru_phi optimization. 1833 int con = value->get_int(); 1834 return TypeInt::make(con & 0xFFFF); 1835 } 1836 return LoadNode::Value(phase); 1837 } 1838 1839 //--------------------------LoadSNode::Ideal-------------------------------------- 1840 // 1841 // If the previous store is to the same address as this load, 1842 // and the value stored was larger than a short, replace this load 1843 // with the value stored truncated to a short. If no truncation is 1844 // needed, the replacement is done in LoadNode::Identity(). 1845 // 1846 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1847 Node* mem = in(MemNode::Memory); 1848 Node* value = can_see_stored_value(mem,phase); 1849 if( value && !phase->type(value)->higher_equal( _type ) ) { 1850 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1851 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1852 } 1853 // Identity call will handle the case where truncation is not needed. 1854 return LoadNode::Ideal(phase, can_reshape); 1855 } 1856 1857 const Type* LoadSNode::Value(PhaseTransform *phase) const { 1858 Node* mem = in(MemNode::Memory); 1859 Node* value = can_see_stored_value(mem,phase); 1860 if (value != NULL && value->is_Con() && 1861 !value->bottom_type()->higher_equal(_type)) { 1862 // If the input to the store does not fit with the load's result type, 1863 // it must be truncated. We can't delay until Ideal call since 1864 // a singleton Value is needed for split_thru_phi optimization. 1865 int con = value->get_int(); 1866 return TypeInt::make((con << 16) >> 16); 1867 } 1868 return LoadNode::Value(phase); 1869 } 1870 1871 //============================================================================= 1872 //----------------------------LoadKlassNode::make------------------------------ 1873 // Polymorphic factory method: 1874 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1875 Compile* C = gvn.C; 1876 Node *ctl = NULL; 1877 // sanity check the alias category against the created node type 1878 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1879 assert(adr_type != NULL, "expecting TypeOopPtr"); 1880 #ifdef _LP64 1881 if (adr_type->is_ptr_to_narrowoop()) { 1882 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1883 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1884 } 1885 #endif 1886 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1887 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1888 } 1889 1890 //------------------------------Value------------------------------------------ 1891 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1892 return klass_value_common(phase); 1893 } 1894 1895 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1896 // Either input is TOP ==> the result is TOP 1897 const Type *t1 = phase->type( in(MemNode::Memory) ); 1898 if (t1 == Type::TOP) return Type::TOP; 1899 Node *adr = in(MemNode::Address); 1900 const Type *t2 = phase->type( adr ); 1901 if (t2 == Type::TOP) return Type::TOP; 1902 const TypePtr *tp = t2->is_ptr(); 1903 if (TypePtr::above_centerline(tp->ptr()) || 1904 tp->ptr() == TypePtr::Null) return Type::TOP; 1905 1906 // Return a more precise klass, if possible 1907 const TypeInstPtr *tinst = tp->isa_instptr(); 1908 if (tinst != NULL) { 1909 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1910 int offset = tinst->offset(); 1911 if (ik == phase->C->env()->Class_klass() 1912 && (offset == java_lang_Class::klass_offset_in_bytes() || 1913 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1914 // We are loading a special hidden field from a Class mirror object, 1915 // the field which points to the VM's Klass metaobject. 1916 ciType* t = tinst->java_mirror_type(); 1917 // java_mirror_type returns non-null for compile-time Class constants. 1918 if (t != NULL) { 1919 // constant oop => constant klass 1920 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1921 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1922 } 1923 if (!t->is_klass()) { 1924 // a primitive Class (e.g., int.class) has NULL for a klass field 1925 return TypePtr::NULL_PTR; 1926 } 1927 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1928 return TypeKlassPtr::make(t->as_klass()); 1929 } 1930 // non-constant mirror, so we can't tell what's going on 1931 } 1932 if( !ik->is_loaded() ) 1933 return _type; // Bail out if not loaded 1934 if (offset == oopDesc::klass_offset_in_bytes()) { 1935 if (tinst->klass_is_exact()) { 1936 return TypeKlassPtr::make(ik); 1937 } 1938 // See if we can become precise: no subklasses and no interface 1939 // (Note: We need to support verified interfaces.) 1940 if (!ik->is_interface() && !ik->has_subklass()) { 1941 //assert(!UseExactTypes, "this code should be useless with exact types"); 1942 // Add a dependence; if any subclass added we need to recompile 1943 if (!ik->is_final()) { 1944 // %%% should use stronger assert_unique_concrete_subtype instead 1945 phase->C->dependencies()->assert_leaf_type(ik); 1946 } 1947 // Return precise klass 1948 return TypeKlassPtr::make(ik); 1949 } 1950 1951 // Return root of possible klass 1952 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1953 } 1954 } 1955 1956 // Check for loading klass from an array 1957 const TypeAryPtr *tary = tp->isa_aryptr(); 1958 if( tary != NULL ) { 1959 ciKlass *tary_klass = tary->klass(); 1960 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1961 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1962 if (tary->klass_is_exact()) { 1963 return TypeKlassPtr::make(tary_klass); 1964 } 1965 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1966 // If the klass is an object array, we defer the question to the 1967 // array component klass. 1968 if( ak->is_obj_array_klass() ) { 1969 assert( ak->is_loaded(), "" ); 1970 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1971 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1972 ciInstanceKlass* ik = base_k->as_instance_klass(); 1973 // See if we can become precise: no subklasses and no interface 1974 if (!ik->is_interface() && !ik->has_subklass()) { 1975 //assert(!UseExactTypes, "this code should be useless with exact types"); 1976 // Add a dependence; if any subclass added we need to recompile 1977 if (!ik->is_final()) { 1978 phase->C->dependencies()->assert_leaf_type(ik); 1979 } 1980 // Return precise array klass 1981 return TypeKlassPtr::make(ak); 1982 } 1983 } 1984 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1985 } else { // Found a type-array? 1986 //assert(!UseExactTypes, "this code should be useless with exact types"); 1987 assert( ak->is_type_array_klass(), "" ); 1988 return TypeKlassPtr::make(ak); // These are always precise 1989 } 1990 } 1991 } 1992 1993 // Check for loading klass from an array klass 1994 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1995 if (tkls != NULL && !StressReflectiveCode) { 1996 ciKlass* klass = tkls->klass(); 1997 if( !klass->is_loaded() ) 1998 return _type; // Bail out if not loaded 1999 if( klass->is_obj_array_klass() && 2000 tkls->offset() == in_bytes(objArrayKlass::element_klass_offset())) { 2001 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2002 // // Always returning precise element type is incorrect, 2003 // // e.g., element type could be object and array may contain strings 2004 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2005 2006 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2007 // according to the element type's subclassing. 2008 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2009 } 2010 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2011 tkls->offset() == in_bytes(Klass::super_offset())) { 2012 ciKlass* sup = klass->as_instance_klass()->super(); 2013 // The field is Klass::_super. Return its (constant) value. 2014 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2015 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2016 } 2017 } 2018 2019 // Bailout case 2020 return LoadNode::Value(phase); 2021 } 2022 2023 //------------------------------Identity--------------------------------------- 2024 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2025 // Also feed through the klass in Allocate(...klass...)._klass. 2026 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2027 return klass_identity_common(phase); 2028 } 2029 2030 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2031 Node* x = LoadNode::Identity(phase); 2032 if (x != this) return x; 2033 2034 // Take apart the address into an oop and and offset. 2035 // Return 'this' if we cannot. 2036 Node* adr = in(MemNode::Address); 2037 intptr_t offset = 0; 2038 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2039 if (base == NULL) return this; 2040 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2041 if (toop == NULL) return this; 2042 2043 // We can fetch the klass directly through an AllocateNode. 2044 // This works even if the klass is not constant (clone or newArray). 2045 if (offset == oopDesc::klass_offset_in_bytes()) { 2046 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2047 if (allocated_klass != NULL) { 2048 return allocated_klass; 2049 } 2050 } 2051 2052 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 2053 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 2054 // See inline_native_Class_query for occurrences of these patterns. 2055 // Java Example: x.getClass().isAssignableFrom(y) 2056 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 2057 // 2058 // This improves reflective code, often making the Class 2059 // mirror go completely dead. (Current exception: Class 2060 // mirrors may appear in debug info, but we could clean them out by 2061 // introducing a new debug info operator for klassOop.java_mirror). 2062 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2063 && (offset == java_lang_Class::klass_offset_in_bytes() || 2064 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2065 // We are loading a special hidden field from a Class mirror, 2066 // the field which points to its Klass or arrayKlass metaobject. 2067 if (base->is_Load()) { 2068 Node* adr2 = base->in(MemNode::Address); 2069 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2070 if (tkls != NULL && !tkls->empty() 2071 && (tkls->klass()->is_instance_klass() || 2072 tkls->klass()->is_array_klass()) 2073 && adr2->is_AddP() 2074 ) { 2075 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2076 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2077 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 2078 } 2079 if (tkls->offset() == mirror_field) { 2080 return adr2->in(AddPNode::Base); 2081 } 2082 } 2083 } 2084 } 2085 2086 return this; 2087 } 2088 2089 2090 //------------------------------Value------------------------------------------ 2091 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2092 const Type *t = klass_value_common(phase); 2093 if (t == Type::TOP) 2094 return t; 2095 2096 return t->make_narrowoop(); 2097 } 2098 2099 //------------------------------Identity--------------------------------------- 2100 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2101 // Also feed through the klass in Allocate(...klass...)._klass. 2102 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2103 Node *x = klass_identity_common(phase); 2104 2105 const Type *t = phase->type( x ); 2106 if( t == Type::TOP ) return x; 2107 if( t->isa_narrowoop()) return x; 2108 2109 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 2110 } 2111 2112 //------------------------------Value----------------------------------------- 2113 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2114 // Either input is TOP ==> the result is TOP 2115 const Type *t1 = phase->type( in(MemNode::Memory) ); 2116 if( t1 == Type::TOP ) return Type::TOP; 2117 Node *adr = in(MemNode::Address); 2118 const Type *t2 = phase->type( adr ); 2119 if( t2 == Type::TOP ) return Type::TOP; 2120 const TypePtr *tp = t2->is_ptr(); 2121 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2122 const TypeAryPtr *tap = tp->isa_aryptr(); 2123 if( !tap ) return _type; 2124 return tap->size(); 2125 } 2126 2127 //-------------------------------Ideal--------------------------------------- 2128 // Feed through the length in AllocateArray(...length...)._length. 2129 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2130 Node* p = MemNode::Ideal_common(phase, can_reshape); 2131 if (p) return (p == NodeSentinel) ? NULL : p; 2132 2133 // Take apart the address into an oop and and offset. 2134 // Return 'this' if we cannot. 2135 Node* adr = in(MemNode::Address); 2136 intptr_t offset = 0; 2137 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2138 if (base == NULL) return NULL; 2139 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2140 if (tary == NULL) return NULL; 2141 2142 // We can fetch the length directly through an AllocateArrayNode. 2143 // This works even if the length is not constant (clone or newArray). 2144 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2145 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2146 if (alloc != NULL) { 2147 Node* allocated_length = alloc->Ideal_length(); 2148 Node* len = alloc->make_ideal_length(tary, phase); 2149 if (allocated_length != len) { 2150 // New CastII improves on this. 2151 return len; 2152 } 2153 } 2154 } 2155 2156 return NULL; 2157 } 2158 2159 //------------------------------Identity--------------------------------------- 2160 // Feed through the length in AllocateArray(...length...)._length. 2161 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2162 Node* x = LoadINode::Identity(phase); 2163 if (x != this) return x; 2164 2165 // Take apart the address into an oop and and offset. 2166 // Return 'this' if we cannot. 2167 Node* adr = in(MemNode::Address); 2168 intptr_t offset = 0; 2169 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2170 if (base == NULL) return this; 2171 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2172 if (tary == NULL) return this; 2173 2174 // We can fetch the length directly through an AllocateArrayNode. 2175 // This works even if the length is not constant (clone or newArray). 2176 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2177 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2178 if (alloc != NULL) { 2179 Node* allocated_length = alloc->Ideal_length(); 2180 // Do not allow make_ideal_length to allocate a CastII node. 2181 Node* len = alloc->make_ideal_length(tary, phase, false); 2182 if (allocated_length == len) { 2183 // Return allocated_length only if it would not be improved by a CastII. 2184 return allocated_length; 2185 } 2186 } 2187 } 2188 2189 return this; 2190 2191 } 2192 2193 //============================================================================= 2194 //---------------------------StoreNode::make----------------------------------- 2195 // Polymorphic factory method: 2196 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 2197 Compile* C = gvn.C; 2198 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2199 ctl != NULL, "raw memory operations should have control edge"); 2200 2201 switch (bt) { 2202 case T_BOOLEAN: 2203 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 2204 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 2205 case T_CHAR: 2206 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 2207 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 2208 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 2209 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 2210 case T_ADDRESS: 2211 case T_OBJECT: 2212 #ifdef _LP64 2213 if (adr->bottom_type()->is_ptr_to_narrowoop() || 2214 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 2215 adr->bottom_type()->isa_rawptr())) { 2216 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2217 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 2218 } else 2219 #endif 2220 { 2221 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 2222 } 2223 } 2224 ShouldNotReachHere(); 2225 return (StoreNode*)NULL; 2226 } 2227 2228 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2229 bool require_atomic = true; 2230 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2231 } 2232 2233 2234 //--------------------------bottom_type---------------------------------------- 2235 const Type *StoreNode::bottom_type() const { 2236 return Type::MEMORY; 2237 } 2238 2239 //------------------------------hash------------------------------------------- 2240 uint StoreNode::hash() const { 2241 // unroll addition of interesting fields 2242 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2243 2244 // Since they are not commoned, do not hash them: 2245 return NO_HASH; 2246 } 2247 2248 //------------------------------Ideal------------------------------------------ 2249 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2250 // When a store immediately follows a relevant allocation/initialization, 2251 // try to capture it into the initialization, or hoist it above. 2252 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2253 Node* p = MemNode::Ideal_common(phase, can_reshape); 2254 if (p) return (p == NodeSentinel) ? NULL : p; 2255 2256 Node* mem = in(MemNode::Memory); 2257 Node* address = in(MemNode::Address); 2258 2259 // Back-to-back stores to same address? Fold em up. Generally 2260 // unsafe if I have intervening uses... Also disallowed for StoreCM 2261 // since they must follow each StoreP operation. Redundant StoreCMs 2262 // are eliminated just before matching in final_graph_reshape. 2263 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && 2264 mem->Opcode() != Op_StoreCM) { 2265 // Looking at a dead closed cycle of memory? 2266 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2267 2268 assert(Opcode() == mem->Opcode() || 2269 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2270 "no mismatched stores, except on raw memory"); 2271 2272 if (mem->outcnt() == 1 && // check for intervening uses 2273 mem->as_Store()->memory_size() <= this->memory_size()) { 2274 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2275 // For example, 'mem' might be the final state at a conditional return. 2276 // Or, 'mem' might be used by some node which is live at the same time 2277 // 'this' is live, which might be unschedulable. So, require exactly 2278 // ONE user, the 'this' store, until such time as we clone 'mem' for 2279 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2280 if (can_reshape) { // (%%% is this an anachronism?) 2281 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2282 phase->is_IterGVN()); 2283 } else { 2284 // It's OK to do this in the parser, since DU info is always accurate, 2285 // and the parser always refers to nodes via SafePointNode maps. 2286 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2287 } 2288 return this; 2289 } 2290 } 2291 2292 // Capture an unaliased, unconditional, simple store into an initializer. 2293 // Or, if it is independent of the allocation, hoist it above the allocation. 2294 if (ReduceFieldZeroing && /*can_reshape &&*/ 2295 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2296 InitializeNode* init = mem->in(0)->as_Initialize(); 2297 intptr_t offset = init->can_capture_store(this, phase); 2298 if (offset > 0) { 2299 Node* moved = init->capture_store(this, offset, phase); 2300 // If the InitializeNode captured me, it made a raw copy of me, 2301 // and I need to disappear. 2302 if (moved != NULL) { 2303 // %%% hack to ensure that Ideal returns a new node: 2304 mem = MergeMemNode::make(phase->C, mem); 2305 return mem; // fold me away 2306 } 2307 } 2308 } 2309 2310 return NULL; // No further progress 2311 } 2312 2313 //------------------------------Value----------------------------------------- 2314 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2315 // Either input is TOP ==> the result is TOP 2316 const Type *t1 = phase->type( in(MemNode::Memory) ); 2317 if( t1 == Type::TOP ) return Type::TOP; 2318 const Type *t2 = phase->type( in(MemNode::Address) ); 2319 if( t2 == Type::TOP ) return Type::TOP; 2320 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2321 if( t3 == Type::TOP ) return Type::TOP; 2322 return Type::MEMORY; 2323 } 2324 2325 //------------------------------Identity--------------------------------------- 2326 // Remove redundant stores: 2327 // Store(m, p, Load(m, p)) changes to m. 2328 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2329 Node *StoreNode::Identity( PhaseTransform *phase ) { 2330 Node* mem = in(MemNode::Memory); 2331 Node* adr = in(MemNode::Address); 2332 Node* val = in(MemNode::ValueIn); 2333 2334 // Load then Store? Then the Store is useless 2335 if (val->is_Load() && 2336 val->in(MemNode::Address)->eqv_uncast(adr) && 2337 val->in(MemNode::Memory )->eqv_uncast(mem) && 2338 val->as_Load()->store_Opcode() == Opcode()) { 2339 return mem; 2340 } 2341 2342 // Two stores in a row of the same value? 2343 if (mem->is_Store() && 2344 mem->in(MemNode::Address)->eqv_uncast(adr) && 2345 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2346 mem->Opcode() == Opcode()) { 2347 return mem; 2348 } 2349 2350 // Store of zero anywhere into a freshly-allocated object? 2351 // Then the store is useless. 2352 // (It must already have been captured by the InitializeNode.) 2353 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2354 // a newly allocated object is already all-zeroes everywhere 2355 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2356 return mem; 2357 } 2358 2359 // the store may also apply to zero-bits in an earlier object 2360 Node* prev_mem = find_previous_store(phase); 2361 // Steps (a), (b): Walk past independent stores to find an exact match. 2362 if (prev_mem != NULL) { 2363 Node* prev_val = can_see_stored_value(prev_mem, phase); 2364 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2365 // prev_val and val might differ by a cast; it would be good 2366 // to keep the more informative of the two. 2367 return mem; 2368 } 2369 } 2370 } 2371 2372 return this; 2373 } 2374 2375 //------------------------------match_edge------------------------------------- 2376 // Do we Match on this edge index or not? Match only memory & value 2377 uint StoreNode::match_edge(uint idx) const { 2378 return idx == MemNode::Address || idx == MemNode::ValueIn; 2379 } 2380 2381 //------------------------------cmp-------------------------------------------- 2382 // Do not common stores up together. They generally have to be split 2383 // back up anyways, so do not bother. 2384 uint StoreNode::cmp( const Node &n ) const { 2385 return (&n == this); // Always fail except on self 2386 } 2387 2388 //------------------------------Ideal_masked_input----------------------------- 2389 // Check for a useless mask before a partial-word store 2390 // (StoreB ... (AndI valIn conIa) ) 2391 // If (conIa & mask == mask) this simplifies to 2392 // (StoreB ... (valIn) ) 2393 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2394 Node *val = in(MemNode::ValueIn); 2395 if( val->Opcode() == Op_AndI ) { 2396 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2397 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2398 set_req(MemNode::ValueIn, val->in(1)); 2399 return this; 2400 } 2401 } 2402 return NULL; 2403 } 2404 2405 2406 //------------------------------Ideal_sign_extended_input---------------------- 2407 // Check for useless sign-extension before a partial-word store 2408 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2409 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2410 // (StoreB ... (valIn) ) 2411 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2412 Node *val = in(MemNode::ValueIn); 2413 if( val->Opcode() == Op_RShiftI ) { 2414 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2415 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2416 Node *shl = val->in(1); 2417 if( shl->Opcode() == Op_LShiftI ) { 2418 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2419 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2420 set_req(MemNode::ValueIn, shl->in(1)); 2421 return this; 2422 } 2423 } 2424 } 2425 } 2426 return NULL; 2427 } 2428 2429 //------------------------------value_never_loaded----------------------------------- 2430 // Determine whether there are any possible loads of the value stored. 2431 // For simplicity, we actually check if there are any loads from the 2432 // address stored to, not just for loads of the value stored by this node. 2433 // 2434 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2435 Node *adr = in(Address); 2436 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2437 if (adr_oop == NULL) 2438 return false; 2439 if (!adr_oop->is_known_instance_field()) 2440 return false; // if not a distinct instance, there may be aliases of the address 2441 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2442 Node *use = adr->fast_out(i); 2443 int opc = use->Opcode(); 2444 if (use->is_Load() || use->is_LoadStore()) { 2445 return false; 2446 } 2447 } 2448 return true; 2449 } 2450 2451 //============================================================================= 2452 //------------------------------Ideal------------------------------------------ 2453 // If the store is from an AND mask that leaves the low bits untouched, then 2454 // we can skip the AND operation. If the store is from a sign-extension 2455 // (a left shift, then right shift) we can skip both. 2456 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2457 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2458 if( progress != NULL ) return progress; 2459 2460 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2461 if( progress != NULL ) return progress; 2462 2463 // Finally check the default case 2464 return StoreNode::Ideal(phase, can_reshape); 2465 } 2466 2467 //============================================================================= 2468 //------------------------------Ideal------------------------------------------ 2469 // If the store is from an AND mask that leaves the low bits untouched, then 2470 // we can skip the AND operation 2471 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2472 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2473 if( progress != NULL ) return progress; 2474 2475 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2476 if( progress != NULL ) return progress; 2477 2478 // Finally check the default case 2479 return StoreNode::Ideal(phase, can_reshape); 2480 } 2481 2482 //============================================================================= 2483 //------------------------------Identity--------------------------------------- 2484 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2485 // No need to card mark when storing a null ptr 2486 Node* my_store = in(MemNode::OopStore); 2487 if (my_store->is_Store()) { 2488 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2489 if( t1 == TypePtr::NULL_PTR ) { 2490 return in(MemNode::Memory); 2491 } 2492 } 2493 return this; 2494 } 2495 2496 //============================================================================= 2497 //------------------------------Ideal--------------------------------------- 2498 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2499 Node* progress = StoreNode::Ideal(phase, can_reshape); 2500 if (progress != NULL) return progress; 2501 2502 Node* my_store = in(MemNode::OopStore); 2503 if (my_store->is_MergeMem()) { 2504 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2505 set_req(MemNode::OopStore, mem); 2506 return this; 2507 } 2508 2509 return NULL; 2510 } 2511 2512 //------------------------------Value----------------------------------------- 2513 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2514 // Either input is TOP ==> the result is TOP 2515 const Type *t = phase->type( in(MemNode::Memory) ); 2516 if( t == Type::TOP ) return Type::TOP; 2517 t = phase->type( in(MemNode::Address) ); 2518 if( t == Type::TOP ) return Type::TOP; 2519 t = phase->type( in(MemNode::ValueIn) ); 2520 if( t == Type::TOP ) return Type::TOP; 2521 // If extra input is TOP ==> the result is TOP 2522 t = phase->type( in(MemNode::OopStore) ); 2523 if( t == Type::TOP ) return Type::TOP; 2524 2525 return StoreNode::Value( phase ); 2526 } 2527 2528 2529 //============================================================================= 2530 //----------------------------------SCMemProjNode------------------------------ 2531 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2532 { 2533 return bottom_type(); 2534 } 2535 2536 //============================================================================= 2537 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2538 init_req(MemNode::Control, c ); 2539 init_req(MemNode::Memory , mem); 2540 init_req(MemNode::Address, adr); 2541 init_req(MemNode::ValueIn, val); 2542 init_req( ExpectedIn, ex ); 2543 init_class_id(Class_LoadStore); 2544 2545 } 2546 2547 //============================================================================= 2548 //-------------------------------adr_type-------------------------------------- 2549 // Do we Match on this edge index or not? Do not match memory 2550 const TypePtr* ClearArrayNode::adr_type() const { 2551 Node *adr = in(3); 2552 return MemNode::calculate_adr_type(adr->bottom_type()); 2553 } 2554 2555 //------------------------------match_edge------------------------------------- 2556 // Do we Match on this edge index or not? Do not match memory 2557 uint ClearArrayNode::match_edge(uint idx) const { 2558 return idx > 1; 2559 } 2560 2561 //------------------------------Identity--------------------------------------- 2562 // Clearing a zero length array does nothing 2563 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2564 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2565 } 2566 2567 //------------------------------Idealize--------------------------------------- 2568 // Clearing a short array is faster with stores 2569 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2570 const int unit = BytesPerLong; 2571 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2572 if (!t) return NULL; 2573 if (!t->is_con()) return NULL; 2574 intptr_t raw_count = t->get_con(); 2575 intptr_t size = raw_count; 2576 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2577 // Clearing nothing uses the Identity call. 2578 // Negative clears are possible on dead ClearArrays 2579 // (see jck test stmt114.stmt11402.val). 2580 if (size <= 0 || size % unit != 0) return NULL; 2581 intptr_t count = size / unit; 2582 // Length too long; use fast hardware clear 2583 if (size > Matcher::init_array_short_size) return NULL; 2584 Node *mem = in(1); 2585 if( phase->type(mem)==Type::TOP ) return NULL; 2586 Node *adr = in(3); 2587 const Type* at = phase->type(adr); 2588 if( at==Type::TOP ) return NULL; 2589 const TypePtr* atp = at->isa_ptr(); 2590 // adjust atp to be the correct array element address type 2591 if (atp == NULL) atp = TypePtr::BOTTOM; 2592 else atp = atp->add_offset(Type::OffsetBot); 2593 // Get base for derived pointer purposes 2594 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2595 Node *base = adr->in(1); 2596 2597 Node *zero = phase->makecon(TypeLong::ZERO); 2598 Node *off = phase->MakeConX(BytesPerLong); 2599 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2600 count--; 2601 while( count-- ) { 2602 mem = phase->transform(mem); 2603 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2604 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2605 } 2606 return mem; 2607 } 2608 2609 //----------------------------step_through---------------------------------- 2610 // Return allocation input memory edge if it is different instance 2611 // or itself if it is the one we are looking for. 2612 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2613 Node* n = *np; 2614 assert(n->is_ClearArray(), "sanity"); 2615 intptr_t offset; 2616 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2617 // This method is called only before Allocate nodes are expanded during 2618 // macro nodes expansion. Before that ClearArray nodes are only generated 2619 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2620 assert(alloc != NULL, "should have allocation"); 2621 if (alloc->_idx == instance_id) { 2622 // Can not bypass initialization of the instance we are looking for. 2623 return false; 2624 } 2625 // Otherwise skip it. 2626 InitializeNode* init = alloc->initialization(); 2627 if (init != NULL) 2628 *np = init->in(TypeFunc::Memory); 2629 else 2630 *np = alloc->in(TypeFunc::Memory); 2631 return true; 2632 } 2633 2634 //----------------------------clear_memory------------------------------------- 2635 // Generate code to initialize object storage to zero. 2636 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2637 intptr_t start_offset, 2638 Node* end_offset, 2639 PhaseGVN* phase) { 2640 Compile* C = phase->C; 2641 intptr_t offset = start_offset; 2642 2643 int unit = BytesPerLong; 2644 if ((offset % unit) != 0) { 2645 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2646 adr = phase->transform(adr); 2647 const TypePtr* atp = TypeRawPtr::BOTTOM; 2648 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2649 mem = phase->transform(mem); 2650 offset += BytesPerInt; 2651 } 2652 assert((offset % unit) == 0, ""); 2653 2654 // Initialize the remaining stuff, if any, with a ClearArray. 2655 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2656 } 2657 2658 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2659 Node* start_offset, 2660 Node* end_offset, 2661 PhaseGVN* phase) { 2662 if (start_offset == end_offset) { 2663 // nothing to do 2664 return mem; 2665 } 2666 2667 Compile* C = phase->C; 2668 int unit = BytesPerLong; 2669 Node* zbase = start_offset; 2670 Node* zend = end_offset; 2671 2672 // Scale to the unit required by the CPU: 2673 if (!Matcher::init_array_count_is_in_bytes) { 2674 Node* shift = phase->intcon(exact_log2(unit)); 2675 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2676 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2677 } 2678 2679 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2680 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2681 2682 // Bulk clear double-words 2683 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2684 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2685 return phase->transform(mem); 2686 } 2687 2688 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2689 intptr_t start_offset, 2690 intptr_t end_offset, 2691 PhaseGVN* phase) { 2692 if (start_offset == end_offset) { 2693 // nothing to do 2694 return mem; 2695 } 2696 2697 Compile* C = phase->C; 2698 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2699 intptr_t done_offset = end_offset; 2700 if ((done_offset % BytesPerLong) != 0) { 2701 done_offset -= BytesPerInt; 2702 } 2703 if (done_offset > start_offset) { 2704 mem = clear_memory(ctl, mem, dest, 2705 start_offset, phase->MakeConX(done_offset), phase); 2706 } 2707 if (done_offset < end_offset) { // emit the final 32-bit store 2708 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2709 adr = phase->transform(adr); 2710 const TypePtr* atp = TypeRawPtr::BOTTOM; 2711 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2712 mem = phase->transform(mem); 2713 done_offset += BytesPerInt; 2714 } 2715 assert(done_offset == end_offset, ""); 2716 return mem; 2717 } 2718 2719 //============================================================================= 2720 // Do not match memory edge. 2721 uint StrIntrinsicNode::match_edge(uint idx) const { 2722 return idx == 2 || idx == 3; 2723 } 2724 2725 //------------------------------Ideal------------------------------------------ 2726 // Return a node which is more "ideal" than the current node. Strip out 2727 // control copies 2728 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2729 if (remove_dead_region(phase, can_reshape)) return this; 2730 // Don't bother trying to transform a dead node 2731 if (in(0) && in(0)->is_top()) return NULL; 2732 2733 if (can_reshape) { 2734 Node* mem = phase->transform(in(MemNode::Memory)); 2735 // If transformed to a MergeMem, get the desired slice 2736 uint alias_idx = phase->C->get_alias_index(adr_type()); 2737 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; 2738 if (mem != in(MemNode::Memory)) { 2739 set_req(MemNode::Memory, mem); 2740 return this; 2741 } 2742 } 2743 return NULL; 2744 } 2745 2746 //------------------------------Value------------------------------------------ 2747 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { 2748 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; 2749 return bottom_type(); 2750 } 2751 2752 //============================================================================= 2753 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2754 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2755 _adr_type(C->get_adr_type(alias_idx)) 2756 { 2757 init_class_id(Class_MemBar); 2758 Node* top = C->top(); 2759 init_req(TypeFunc::I_O,top); 2760 init_req(TypeFunc::FramePtr,top); 2761 init_req(TypeFunc::ReturnAdr,top); 2762 if (precedent != NULL) 2763 init_req(TypeFunc::Parms, precedent); 2764 } 2765 2766 //------------------------------cmp-------------------------------------------- 2767 uint MemBarNode::hash() const { return NO_HASH; } 2768 uint MemBarNode::cmp( const Node &n ) const { 2769 return (&n == this); // Always fail except on self 2770 } 2771 2772 //------------------------------make------------------------------------------- 2773 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2774 int len = Precedent + (pn == NULL? 0: 1); 2775 switch (opcode) { 2776 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2777 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2778 case Op_MemBarAcquireLock: return new(C, len) MemBarAcquireLockNode(C, atp, pn); 2779 case Op_MemBarReleaseLock: return new(C, len) MemBarReleaseLockNode(C, atp, pn); 2780 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2781 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2782 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2783 case Op_MemBarStoreStore: return new(C, len) MemBarStoreStoreNode(C, atp, pn); 2784 default: ShouldNotReachHere(); return NULL; 2785 } 2786 } 2787 2788 //------------------------------Ideal------------------------------------------ 2789 // Return a node which is more "ideal" than the current node. Strip out 2790 // control copies 2791 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2792 if (remove_dead_region(phase, can_reshape)) return this; 2793 // Don't bother trying to transform a dead node 2794 if (in(0) && in(0)->is_top()) return NULL; 2795 2796 // Eliminate volatile MemBars for scalar replaced objects. 2797 if (can_reshape && req() == (Precedent+1) && 2798 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { 2799 // Volatile field loads and stores. 2800 Node* my_mem = in(MemBarNode::Precedent); 2801 if (my_mem != NULL && my_mem->is_Mem()) { 2802 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2803 // Check for scalar replaced object reference. 2804 if( t_oop != NULL && t_oop->is_known_instance_field() && 2805 t_oop->offset() != Type::OffsetBot && 2806 t_oop->offset() != Type::OffsetTop) { 2807 // Replace MemBar projections by its inputs. 2808 PhaseIterGVN* igvn = phase->is_IterGVN(); 2809 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2810 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2811 // Must return either the original node (now dead) or a new node 2812 // (Do not return a top here, since that would break the uniqueness of top.) 2813 return new (phase->C, 1) ConINode(TypeInt::ZERO); 2814 } 2815 } 2816 } 2817 return NULL; 2818 } 2819 2820 //------------------------------Value------------------------------------------ 2821 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2822 if( !in(0) ) return Type::TOP; 2823 if( phase->type(in(0)) == Type::TOP ) 2824 return Type::TOP; 2825 return TypeTuple::MEMBAR; 2826 } 2827 2828 //------------------------------match------------------------------------------ 2829 // Construct projections for memory. 2830 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2831 switch (proj->_con) { 2832 case TypeFunc::Control: 2833 case TypeFunc::Memory: 2834 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2835 } 2836 ShouldNotReachHere(); 2837 return NULL; 2838 } 2839 2840 //===========================InitializeNode==================================== 2841 // SUMMARY: 2842 // This node acts as a memory barrier on raw memory, after some raw stores. 2843 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2844 // The Initialize can 'capture' suitably constrained stores as raw inits. 2845 // It can coalesce related raw stores into larger units (called 'tiles'). 2846 // It can avoid zeroing new storage for memory units which have raw inits. 2847 // At macro-expansion, it is marked 'complete', and does not optimize further. 2848 // 2849 // EXAMPLE: 2850 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2851 // ctl = incoming control; mem* = incoming memory 2852 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2853 // First allocate uninitialized memory and fill in the header: 2854 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2855 // ctl := alloc.Control; mem* := alloc.Memory* 2856 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2857 // Then initialize to zero the non-header parts of the raw memory block: 2858 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2859 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2860 // After the initialize node executes, the object is ready for service: 2861 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2862 // Suppose its body is immediately initialized as {1,2}: 2863 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2864 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2865 // mem.SLICE(#short[*]) := store2 2866 // 2867 // DETAILS: 2868 // An InitializeNode collects and isolates object initialization after 2869 // an AllocateNode and before the next possible safepoint. As a 2870 // memory barrier (MemBarNode), it keeps critical stores from drifting 2871 // down past any safepoint or any publication of the allocation. 2872 // Before this barrier, a newly-allocated object may have uninitialized bits. 2873 // After this barrier, it may be treated as a real oop, and GC is allowed. 2874 // 2875 // The semantics of the InitializeNode include an implicit zeroing of 2876 // the new object from object header to the end of the object. 2877 // (The object header and end are determined by the AllocateNode.) 2878 // 2879 // Certain stores may be added as direct inputs to the InitializeNode. 2880 // These stores must update raw memory, and they must be to addresses 2881 // derived from the raw address produced by AllocateNode, and with 2882 // a constant offset. They must be ordered by increasing offset. 2883 // The first one is at in(RawStores), the last at in(req()-1). 2884 // Unlike most memory operations, they are not linked in a chain, 2885 // but are displayed in parallel as users of the rawmem output of 2886 // the allocation. 2887 // 2888 // (See comments in InitializeNode::capture_store, which continue 2889 // the example given above.) 2890 // 2891 // When the associated Allocate is macro-expanded, the InitializeNode 2892 // may be rewritten to optimize collected stores. A ClearArrayNode 2893 // may also be created at that point to represent any required zeroing. 2894 // The InitializeNode is then marked 'complete', prohibiting further 2895 // capturing of nearby memory operations. 2896 // 2897 // During macro-expansion, all captured initializations which store 2898 // constant values of 32 bits or smaller are coalesced (if advantageous) 2899 // into larger 'tiles' 32 or 64 bits. This allows an object to be 2900 // initialized in fewer memory operations. Memory words which are 2901 // covered by neither tiles nor non-constant stores are pre-zeroed 2902 // by explicit stores of zero. (The code shape happens to do all 2903 // zeroing first, then all other stores, with both sequences occurring 2904 // in order of ascending offsets.) 2905 // 2906 // Alternatively, code may be inserted between an AllocateNode and its 2907 // InitializeNode, to perform arbitrary initialization of the new object. 2908 // E.g., the object copying intrinsics insert complex data transfers here. 2909 // The initialization must then be marked as 'complete' disable the 2910 // built-in zeroing semantics and the collection of initializing stores. 2911 // 2912 // While an InitializeNode is incomplete, reads from the memory state 2913 // produced by it are optimizable if they match the control edge and 2914 // new oop address associated with the allocation/initialization. 2915 // They return a stored value (if the offset matches) or else zero. 2916 // A write to the memory state, if it matches control and address, 2917 // and if it is to a constant offset, may be 'captured' by the 2918 // InitializeNode. It is cloned as a raw memory operation and rewired 2919 // inside the initialization, to the raw oop produced by the allocation. 2920 // Operations on addresses which are provably distinct (e.g., to 2921 // other AllocateNodes) are allowed to bypass the initialization. 2922 // 2923 // The effect of all this is to consolidate object initialization 2924 // (both arrays and non-arrays, both piecewise and bulk) into a 2925 // single location, where it can be optimized as a unit. 2926 // 2927 // Only stores with an offset less than TrackedInitializationLimit words 2928 // will be considered for capture by an InitializeNode. This puts a 2929 // reasonable limit on the complexity of optimized initializations. 2930 2931 //---------------------------InitializeNode------------------------------------ 2932 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2933 : _is_complete(Incomplete), _does_not_escape(false), 2934 MemBarNode(C, adr_type, rawoop) 2935 { 2936 init_class_id(Class_Initialize); 2937 2938 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2939 assert(in(RawAddress) == rawoop, "proper init"); 2940 // Note: allocation() can be NULL, for secondary initialization barriers 2941 } 2942 2943 // Since this node is not matched, it will be processed by the 2944 // register allocator. Declare that there are no constraints 2945 // on the allocation of the RawAddress edge. 2946 const RegMask &InitializeNode::in_RegMask(uint idx) const { 2947 // This edge should be set to top, by the set_complete. But be conservative. 2948 if (idx == InitializeNode::RawAddress) 2949 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2950 return RegMask::Empty; 2951 } 2952 2953 Node* InitializeNode::memory(uint alias_idx) { 2954 Node* mem = in(Memory); 2955 if (mem->is_MergeMem()) { 2956 return mem->as_MergeMem()->memory_at(alias_idx); 2957 } else { 2958 // incoming raw memory is not split 2959 return mem; 2960 } 2961 } 2962 2963 bool InitializeNode::is_non_zero() { 2964 if (is_complete()) return false; 2965 remove_extra_zeroes(); 2966 return (req() > RawStores); 2967 } 2968 2969 void InitializeNode::set_complete(PhaseGVN* phase) { 2970 assert(!is_complete(), "caller responsibility"); 2971 _is_complete = Complete; 2972 2973 // After this node is complete, it contains a bunch of 2974 // raw-memory initializations. There is no need for 2975 // it to have anything to do with non-raw memory effects. 2976 // Therefore, tell all non-raw users to re-optimize themselves, 2977 // after skipping the memory effects of this initialization. 2978 PhaseIterGVN* igvn = phase->is_IterGVN(); 2979 if (igvn) igvn->add_users_to_worklist(this); 2980 } 2981 2982 // convenience function 2983 // return false if the init contains any stores already 2984 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2985 InitializeNode* init = initialization(); 2986 if (init == NULL || init->is_complete()) return false; 2987 init->remove_extra_zeroes(); 2988 // for now, if this allocation has already collected any inits, bail: 2989 if (init->is_non_zero()) return false; 2990 init->set_complete(phase); 2991 return true; 2992 } 2993 2994 void InitializeNode::remove_extra_zeroes() { 2995 if (req() == RawStores) return; 2996 Node* zmem = zero_memory(); 2997 uint fill = RawStores; 2998 for (uint i = fill; i < req(); i++) { 2999 Node* n = in(i); 3000 if (n->is_top() || n == zmem) continue; // skip 3001 if (fill < i) set_req(fill, n); // compact 3002 ++fill; 3003 } 3004 // delete any empty spaces created: 3005 while (fill < req()) { 3006 del_req(fill); 3007 } 3008 } 3009 3010 // Helper for remembering which stores go with which offsets. 3011 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3012 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3013 intptr_t offset = -1; 3014 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3015 phase, offset); 3016 if (base == NULL) return -1; // something is dead, 3017 if (offset < 0) return -1; // dead, dead 3018 return offset; 3019 } 3020 3021 // Helper for proving that an initialization expression is 3022 // "simple enough" to be folded into an object initialization. 3023 // Attempts to prove that a store's initial value 'n' can be captured 3024 // within the initialization without creating a vicious cycle, such as: 3025 // { Foo p = new Foo(); p.next = p; } 3026 // True for constants and parameters and small combinations thereof. 3027 bool InitializeNode::detect_init_independence(Node* n, 3028 bool st_is_pinned, 3029 int& count) { 3030 if (n == NULL) return true; // (can this really happen?) 3031 if (n->is_Proj()) n = n->in(0); 3032 if (n == this) return false; // found a cycle 3033 if (n->is_Con()) return true; 3034 if (n->is_Start()) return true; // params, etc., are OK 3035 if (n->is_Root()) return true; // even better 3036 3037 Node* ctl = n->in(0); 3038 if (ctl != NULL && !ctl->is_top()) { 3039 if (ctl->is_Proj()) ctl = ctl->in(0); 3040 if (ctl == this) return false; 3041 3042 // If we already know that the enclosing memory op is pinned right after 3043 // the init, then any control flow that the store has picked up 3044 // must have preceded the init, or else be equal to the init. 3045 // Even after loop optimizations (which might change control edges) 3046 // a store is never pinned *before* the availability of its inputs. 3047 if (!MemNode::all_controls_dominate(n, this)) 3048 return false; // failed to prove a good control 3049 3050 } 3051 3052 // Check data edges for possible dependencies on 'this'. 3053 if ((count += 1) > 20) return false; // complexity limit 3054 for (uint i = 1; i < n->req(); i++) { 3055 Node* m = n->in(i); 3056 if (m == NULL || m == n || m->is_top()) continue; 3057 uint first_i = n->find_edge(m); 3058 if (i != first_i) continue; // process duplicate edge just once 3059 if (!detect_init_independence(m, st_is_pinned, count)) { 3060 return false; 3061 } 3062 } 3063 3064 return true; 3065 } 3066 3067 // Here are all the checks a Store must pass before it can be moved into 3068 // an initialization. Returns zero if a check fails. 3069 // On success, returns the (constant) offset to which the store applies, 3070 // within the initialized memory. 3071 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 3072 const int FAIL = 0; 3073 if (st->req() != MemNode::ValueIn + 1) 3074 return FAIL; // an inscrutable StoreNode (card mark?) 3075 Node* ctl = st->in(MemNode::Control); 3076 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3077 return FAIL; // must be unconditional after the initialization 3078 Node* mem = st->in(MemNode::Memory); 3079 if (!(mem->is_Proj() && mem->in(0) == this)) 3080 return FAIL; // must not be preceded by other stores 3081 Node* adr = st->in(MemNode::Address); 3082 intptr_t offset; 3083 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3084 if (alloc == NULL) 3085 return FAIL; // inscrutable address 3086 if (alloc != allocation()) 3087 return FAIL; // wrong allocation! (store needs to float up) 3088 Node* val = st->in(MemNode::ValueIn); 3089 int complexity_count = 0; 3090 if (!detect_init_independence(val, true, complexity_count)) 3091 return FAIL; // stored value must be 'simple enough' 3092 3093 return offset; // success 3094 } 3095 3096 // Find the captured store in(i) which corresponds to the range 3097 // [start..start+size) in the initialized object. 3098 // If there is one, return its index i. If there isn't, return the 3099 // negative of the index where it should be inserted. 3100 // Return 0 if the queried range overlaps an initialization boundary 3101 // or if dead code is encountered. 3102 // If size_in_bytes is zero, do not bother with overlap checks. 3103 int InitializeNode::captured_store_insertion_point(intptr_t start, 3104 int size_in_bytes, 3105 PhaseTransform* phase) { 3106 const int FAIL = 0, MAX_STORE = BytesPerLong; 3107 3108 if (is_complete()) 3109 return FAIL; // arraycopy got here first; punt 3110 3111 assert(allocation() != NULL, "must be present"); 3112 3113 // no negatives, no header fields: 3114 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3115 3116 // after a certain size, we bail out on tracking all the stores: 3117 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3118 if (start >= ti_limit) return FAIL; 3119 3120 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3121 if (i >= limit) return -(int)i; // not found; here is where to put it 3122 3123 Node* st = in(i); 3124 intptr_t st_off = get_store_offset(st, phase); 3125 if (st_off < 0) { 3126 if (st != zero_memory()) { 3127 return FAIL; // bail out if there is dead garbage 3128 } 3129 } else if (st_off > start) { 3130 // ...we are done, since stores are ordered 3131 if (st_off < start + size_in_bytes) { 3132 return FAIL; // the next store overlaps 3133 } 3134 return -(int)i; // not found; here is where to put it 3135 } else if (st_off < start) { 3136 if (size_in_bytes != 0 && 3137 start < st_off + MAX_STORE && 3138 start < st_off + st->as_Store()->memory_size()) { 3139 return FAIL; // the previous store overlaps 3140 } 3141 } else { 3142 if (size_in_bytes != 0 && 3143 st->as_Store()->memory_size() != size_in_bytes) { 3144 return FAIL; // mismatched store size 3145 } 3146 return i; 3147 } 3148 3149 ++i; 3150 } 3151 } 3152 3153 // Look for a captured store which initializes at the offset 'start' 3154 // with the given size. If there is no such store, and no other 3155 // initialization interferes, then return zero_memory (the memory 3156 // projection of the AllocateNode). 3157 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3158 PhaseTransform* phase) { 3159 assert(stores_are_sane(phase), ""); 3160 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3161 if (i == 0) { 3162 return NULL; // something is dead 3163 } else if (i < 0) { 3164 return zero_memory(); // just primordial zero bits here 3165 } else { 3166 Node* st = in(i); // here is the store at this position 3167 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3168 return st; 3169 } 3170 } 3171 3172 // Create, as a raw pointer, an address within my new object at 'offset'. 3173 Node* InitializeNode::make_raw_address(intptr_t offset, 3174 PhaseTransform* phase) { 3175 Node* addr = in(RawAddress); 3176 if (offset != 0) { 3177 Compile* C = phase->C; 3178 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 3179 phase->MakeConX(offset)) ); 3180 } 3181 return addr; 3182 } 3183 3184 // Clone the given store, converting it into a raw store 3185 // initializing a field or element of my new object. 3186 // Caller is responsible for retiring the original store, 3187 // with subsume_node or the like. 3188 // 3189 // From the example above InitializeNode::InitializeNode, 3190 // here are the old stores to be captured: 3191 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3192 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3193 // 3194 // Here is the changed code; note the extra edges on init: 3195 // alloc = (Allocate ...) 3196 // rawoop = alloc.RawAddress 3197 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3198 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3199 // init = (Initialize alloc.Control alloc.Memory rawoop 3200 // rawstore1 rawstore2) 3201 // 3202 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3203 PhaseTransform* phase) { 3204 assert(stores_are_sane(phase), ""); 3205 3206 if (start < 0) return NULL; 3207 assert(can_capture_store(st, phase) == start, "sanity"); 3208 3209 Compile* C = phase->C; 3210 int size_in_bytes = st->memory_size(); 3211 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3212 if (i == 0) return NULL; // bail out 3213 Node* prev_mem = NULL; // raw memory for the captured store 3214 if (i > 0) { 3215 prev_mem = in(i); // there is a pre-existing store under this one 3216 set_req(i, C->top()); // temporarily disconnect it 3217 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3218 } else { 3219 i = -i; // no pre-existing store 3220 prev_mem = zero_memory(); // a slice of the newly allocated object 3221 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3222 set_req(--i, C->top()); // reuse this edge; it has been folded away 3223 else 3224 ins_req(i, C->top()); // build a new edge 3225 } 3226 Node* new_st = st->clone(); 3227 new_st->set_req(MemNode::Control, in(Control)); 3228 new_st->set_req(MemNode::Memory, prev_mem); 3229 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3230 new_st = phase->transform(new_st); 3231 3232 // At this point, new_st might have swallowed a pre-existing store 3233 // at the same offset, or perhaps new_st might have disappeared, 3234 // if it redundantly stored the same value (or zero to fresh memory). 3235 3236 // In any case, wire it in: 3237 set_req(i, new_st); 3238 3239 // The caller may now kill the old guy. 3240 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3241 assert(check_st == new_st || check_st == NULL, "must be findable"); 3242 assert(!is_complete(), ""); 3243 return new_st; 3244 } 3245 3246 static bool store_constant(jlong* tiles, int num_tiles, 3247 intptr_t st_off, int st_size, 3248 jlong con) { 3249 if ((st_off & (st_size-1)) != 0) 3250 return false; // strange store offset (assume size==2**N) 3251 address addr = (address)tiles + st_off; 3252 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3253 switch (st_size) { 3254 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3255 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3256 case sizeof(jint): *(jint*) addr = (jint) con; break; 3257 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3258 default: return false; // strange store size (detect size!=2**N here) 3259 } 3260 return true; // return success to caller 3261 } 3262 3263 // Coalesce subword constants into int constants and possibly 3264 // into long constants. The goal, if the CPU permits, 3265 // is to initialize the object with a small number of 64-bit tiles. 3266 // Also, convert floating-point constants to bit patterns. 3267 // Non-constants are not relevant to this pass. 3268 // 3269 // In terms of the running example on InitializeNode::InitializeNode 3270 // and InitializeNode::capture_store, here is the transformation 3271 // of rawstore1 and rawstore2 into rawstore12: 3272 // alloc = (Allocate ...) 3273 // rawoop = alloc.RawAddress 3274 // tile12 = 0x00010002 3275 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3276 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3277 // 3278 void 3279 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3280 Node* size_in_bytes, 3281 PhaseGVN* phase) { 3282 Compile* C = phase->C; 3283 3284 assert(stores_are_sane(phase), ""); 3285 // Note: After this pass, they are not completely sane, 3286 // since there may be some overlaps. 3287 3288 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3289 3290 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3291 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3292 size_limit = MIN2(size_limit, ti_limit); 3293 size_limit = align_size_up(size_limit, BytesPerLong); 3294 int num_tiles = size_limit / BytesPerLong; 3295 3296 // allocate space for the tile map: 3297 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3298 jlong tiles_buf[small_len]; 3299 Node* nodes_buf[small_len]; 3300 jlong inits_buf[small_len]; 3301 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3302 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3303 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3304 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3305 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3306 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3307 // tiles: exact bitwise model of all primitive constants 3308 // nodes: last constant-storing node subsumed into the tiles model 3309 // inits: which bytes (in each tile) are touched by any initializations 3310 3311 //// Pass A: Fill in the tile model with any relevant stores. 3312 3313 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3314 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3315 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3316 Node* zmem = zero_memory(); // initially zero memory state 3317 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3318 Node* st = in(i); 3319 intptr_t st_off = get_store_offset(st, phase); 3320 3321 // Figure out the store's offset and constant value: 3322 if (st_off < header_size) continue; //skip (ignore header) 3323 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3324 int st_size = st->as_Store()->memory_size(); 3325 if (st_off + st_size > size_limit) break; 3326 3327 // Record which bytes are touched, whether by constant or not. 3328 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3329 continue; // skip (strange store size) 3330 3331 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3332 if (!val->singleton()) continue; //skip (non-con store) 3333 BasicType type = val->basic_type(); 3334 3335 jlong con = 0; 3336 switch (type) { 3337 case T_INT: con = val->is_int()->get_con(); break; 3338 case T_LONG: con = val->is_long()->get_con(); break; 3339 case T_FLOAT: con = jint_cast(val->getf()); break; 3340 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3341 default: continue; //skip (odd store type) 3342 } 3343 3344 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3345 st->Opcode() == Op_StoreL) { 3346 continue; // This StoreL is already optimal. 3347 } 3348 3349 // Store down the constant. 3350 store_constant(tiles, num_tiles, st_off, st_size, con); 3351 3352 intptr_t j = st_off >> LogBytesPerLong; 3353 3354 if (type == T_INT && st_size == BytesPerInt 3355 && (st_off & BytesPerInt) == BytesPerInt) { 3356 jlong lcon = tiles[j]; 3357 if (!Matcher::isSimpleConstant64(lcon) && 3358 st->Opcode() == Op_StoreI) { 3359 // This StoreI is already optimal by itself. 3360 jint* intcon = (jint*) &tiles[j]; 3361 intcon[1] = 0; // undo the store_constant() 3362 3363 // If the previous store is also optimal by itself, back up and 3364 // undo the action of the previous loop iteration... if we can. 3365 // But if we can't, just let the previous half take care of itself. 3366 st = nodes[j]; 3367 st_off -= BytesPerInt; 3368 con = intcon[0]; 3369 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3370 assert(st_off >= header_size, "still ignoring header"); 3371 assert(get_store_offset(st, phase) == st_off, "must be"); 3372 assert(in(i-1) == zmem, "must be"); 3373 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3374 assert(con == tcon->is_int()->get_con(), "must be"); 3375 // Undo the effects of the previous loop trip, which swallowed st: 3376 intcon[0] = 0; // undo store_constant() 3377 set_req(i-1, st); // undo set_req(i, zmem) 3378 nodes[j] = NULL; // undo nodes[j] = st 3379 --old_subword; // undo ++old_subword 3380 } 3381 continue; // This StoreI is already optimal. 3382 } 3383 } 3384 3385 // This store is not needed. 3386 set_req(i, zmem); 3387 nodes[j] = st; // record for the moment 3388 if (st_size < BytesPerLong) // something has changed 3389 ++old_subword; // includes int/float, but who's counting... 3390 else ++old_long; 3391 } 3392 3393 if ((old_subword + old_long) == 0) 3394 return; // nothing more to do 3395 3396 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3397 // Be sure to insert them before overlapping non-constant stores. 3398 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3399 for (int j = 0; j < num_tiles; j++) { 3400 jlong con = tiles[j]; 3401 jlong init = inits[j]; 3402 if (con == 0) continue; 3403 jint con0, con1; // split the constant, address-wise 3404 jint init0, init1; // split the init map, address-wise 3405 { union { jlong con; jint intcon[2]; } u; 3406 u.con = con; 3407 con0 = u.intcon[0]; 3408 con1 = u.intcon[1]; 3409 u.con = init; 3410 init0 = u.intcon[0]; 3411 init1 = u.intcon[1]; 3412 } 3413 3414 Node* old = nodes[j]; 3415 assert(old != NULL, "need the prior store"); 3416 intptr_t offset = (j * BytesPerLong); 3417 3418 bool split = !Matcher::isSimpleConstant64(con); 3419 3420 if (offset < header_size) { 3421 assert(offset + BytesPerInt >= header_size, "second int counts"); 3422 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3423 split = true; // only the second word counts 3424 // Example: int a[] = { 42 ... } 3425 } else if (con0 == 0 && init0 == -1) { 3426 split = true; // first word is covered by full inits 3427 // Example: int a[] = { ... foo(), 42 ... } 3428 } else if (con1 == 0 && init1 == -1) { 3429 split = true; // second word is covered by full inits 3430 // Example: int a[] = { ... 42, foo() ... } 3431 } 3432 3433 // Here's a case where init0 is neither 0 nor -1: 3434 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3435 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3436 // In this case the tile is not split; it is (jlong)42. 3437 // The big tile is stored down, and then the foo() value is inserted. 3438 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3439 3440 Node* ctl = old->in(MemNode::Control); 3441 Node* adr = make_raw_address(offset, phase); 3442 const TypePtr* atp = TypeRawPtr::BOTTOM; 3443 3444 // One or two coalesced stores to plop down. 3445 Node* st[2]; 3446 intptr_t off[2]; 3447 int nst = 0; 3448 if (!split) { 3449 ++new_long; 3450 off[nst] = offset; 3451 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3452 phase->longcon(con), T_LONG); 3453 } else { 3454 // Omit either if it is a zero. 3455 if (con0 != 0) { 3456 ++new_int; 3457 off[nst] = offset; 3458 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3459 phase->intcon(con0), T_INT); 3460 } 3461 if (con1 != 0) { 3462 ++new_int; 3463 offset += BytesPerInt; 3464 adr = make_raw_address(offset, phase); 3465 off[nst] = offset; 3466 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3467 phase->intcon(con1), T_INT); 3468 } 3469 } 3470 3471 // Insert second store first, then the first before the second. 3472 // Insert each one just before any overlapping non-constant stores. 3473 while (nst > 0) { 3474 Node* st1 = st[--nst]; 3475 C->copy_node_notes_to(st1, old); 3476 st1 = phase->transform(st1); 3477 offset = off[nst]; 3478 assert(offset >= header_size, "do not smash header"); 3479 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3480 guarantee(ins_idx != 0, "must re-insert constant store"); 3481 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3482 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3483 set_req(--ins_idx, st1); 3484 else 3485 ins_req(ins_idx, st1); 3486 } 3487 } 3488 3489 if (PrintCompilation && WizardMode) 3490 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3491 old_subword, old_long, new_int, new_long); 3492 if (C->log() != NULL) 3493 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3494 old_subword, old_long, new_int, new_long); 3495 3496 // Clean up any remaining occurrences of zmem: 3497 remove_extra_zeroes(); 3498 } 3499 3500 // Explore forward from in(start) to find the first fully initialized 3501 // word, and return its offset. Skip groups of subword stores which 3502 // together initialize full words. If in(start) is itself part of a 3503 // fully initialized word, return the offset of in(start). If there 3504 // are no following full-word stores, or if something is fishy, return 3505 // a negative value. 3506 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3507 int int_map = 0; 3508 intptr_t int_map_off = 0; 3509 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3510 3511 for (uint i = start, limit = req(); i < limit; i++) { 3512 Node* st = in(i); 3513 3514 intptr_t st_off = get_store_offset(st, phase); 3515 if (st_off < 0) break; // return conservative answer 3516 3517 int st_size = st->as_Store()->memory_size(); 3518 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3519 return st_off; // we found a complete word init 3520 } 3521 3522 // update the map: 3523 3524 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3525 if (this_int_off != int_map_off) { 3526 // reset the map: 3527 int_map = 0; 3528 int_map_off = this_int_off; 3529 } 3530 3531 int subword_off = st_off - this_int_off; 3532 int_map |= right_n_bits(st_size) << subword_off; 3533 if ((int_map & FULL_MAP) == FULL_MAP) { 3534 return this_int_off; // we found a complete word init 3535 } 3536 3537 // Did this store hit or cross the word boundary? 3538 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3539 if (next_int_off == this_int_off + BytesPerInt) { 3540 // We passed the current int, without fully initializing it. 3541 int_map_off = next_int_off; 3542 int_map >>= BytesPerInt; 3543 } else if (next_int_off > this_int_off + BytesPerInt) { 3544 // We passed the current and next int. 3545 return this_int_off + BytesPerInt; 3546 } 3547 } 3548 3549 return -1; 3550 } 3551 3552 3553 // Called when the associated AllocateNode is expanded into CFG. 3554 // At this point, we may perform additional optimizations. 3555 // Linearize the stores by ascending offset, to make memory 3556 // activity as coherent as possible. 3557 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3558 intptr_t header_size, 3559 Node* size_in_bytes, 3560 PhaseGVN* phase) { 3561 assert(!is_complete(), "not already complete"); 3562 assert(stores_are_sane(phase), ""); 3563 assert(allocation() != NULL, "must be present"); 3564 3565 remove_extra_zeroes(); 3566 3567 if (ReduceFieldZeroing || ReduceBulkZeroing) 3568 // reduce instruction count for common initialization patterns 3569 coalesce_subword_stores(header_size, size_in_bytes, phase); 3570 3571 Node* zmem = zero_memory(); // initially zero memory state 3572 Node* inits = zmem; // accumulating a linearized chain of inits 3573 #ifdef ASSERT 3574 intptr_t first_offset = allocation()->minimum_header_size(); 3575 intptr_t last_init_off = first_offset; // previous init offset 3576 intptr_t last_init_end = first_offset; // previous init offset+size 3577 intptr_t last_tile_end = first_offset; // previous tile offset+size 3578 #endif 3579 intptr_t zeroes_done = header_size; 3580 3581 bool do_zeroing = true; // we might give up if inits are very sparse 3582 int big_init_gaps = 0; // how many large gaps have we seen? 3583 3584 if (ZeroTLAB) do_zeroing = false; 3585 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3586 3587 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3588 Node* st = in(i); 3589 intptr_t st_off = get_store_offset(st, phase); 3590 if (st_off < 0) 3591 break; // unknown junk in the inits 3592 if (st->in(MemNode::Memory) != zmem) 3593 break; // complicated store chains somehow in list 3594 3595 int st_size = st->as_Store()->memory_size(); 3596 intptr_t next_init_off = st_off + st_size; 3597 3598 if (do_zeroing && zeroes_done < next_init_off) { 3599 // See if this store needs a zero before it or under it. 3600 intptr_t zeroes_needed = st_off; 3601 3602 if (st_size < BytesPerInt) { 3603 // Look for subword stores which only partially initialize words. 3604 // If we find some, we must lay down some word-level zeroes first, 3605 // underneath the subword stores. 3606 // 3607 // Examples: 3608 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3609 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3610 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3611 // 3612 // Note: coalesce_subword_stores may have already done this, 3613 // if it was prompted by constant non-zero subword initializers. 3614 // But this case can still arise with non-constant stores. 3615 3616 intptr_t next_full_store = find_next_fullword_store(i, phase); 3617 3618 // In the examples above: 3619 // in(i) p q r s x y z 3620 // st_off 12 13 14 15 12 13 14 3621 // st_size 1 1 1 1 1 1 1 3622 // next_full_s. 12 16 16 16 16 16 16 3623 // z's_done 12 16 16 16 12 16 12 3624 // z's_needed 12 16 16 16 16 16 16 3625 // zsize 0 0 0 0 4 0 4 3626 if (next_full_store < 0) { 3627 // Conservative tack: Zero to end of current word. 3628 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3629 } else { 3630 // Zero to beginning of next fully initialized word. 3631 // Or, don't zero at all, if we are already in that word. 3632 assert(next_full_store >= zeroes_needed, "must go forward"); 3633 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3634 zeroes_needed = next_full_store; 3635 } 3636 } 3637 3638 if (zeroes_needed > zeroes_done) { 3639 intptr_t zsize = zeroes_needed - zeroes_done; 3640 // Do some incremental zeroing on rawmem, in parallel with inits. 3641 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3642 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3643 zeroes_done, zeroes_needed, 3644 phase); 3645 zeroes_done = zeroes_needed; 3646 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3647 do_zeroing = false; // leave the hole, next time 3648 } 3649 } 3650 3651 // Collect the store and move on: 3652 st->set_req(MemNode::Memory, inits); 3653 inits = st; // put it on the linearized chain 3654 set_req(i, zmem); // unhook from previous position 3655 3656 if (zeroes_done == st_off) 3657 zeroes_done = next_init_off; 3658 3659 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3660 3661 #ifdef ASSERT 3662 // Various order invariants. Weaker than stores_are_sane because 3663 // a large constant tile can be filled in by smaller non-constant stores. 3664 assert(st_off >= last_init_off, "inits do not reverse"); 3665 last_init_off = st_off; 3666 const Type* val = NULL; 3667 if (st_size >= BytesPerInt && 3668 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3669 (int)val->basic_type() < (int)T_OBJECT) { 3670 assert(st_off >= last_tile_end, "tiles do not overlap"); 3671 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3672 last_tile_end = MAX2(last_tile_end, next_init_off); 3673 } else { 3674 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3675 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3676 assert(st_off >= last_init_end, "inits do not overlap"); 3677 last_init_end = next_init_off; // it's a non-tile 3678 } 3679 #endif //ASSERT 3680 } 3681 3682 remove_extra_zeroes(); // clear out all the zmems left over 3683 add_req(inits); 3684 3685 if (!ZeroTLAB) { 3686 // If anything remains to be zeroed, zero it all now. 3687 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3688 // if it is the last unused 4 bytes of an instance, forget about it 3689 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3690 if (zeroes_done + BytesPerLong >= size_limit) { 3691 assert(allocation() != NULL, ""); 3692 if (allocation()->Opcode() == Op_Allocate) { 3693 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3694 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3695 if (zeroes_done == k->layout_helper()) 3696 zeroes_done = size_limit; 3697 } 3698 } 3699 if (zeroes_done < size_limit) { 3700 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3701 zeroes_done, size_in_bytes, phase); 3702 } 3703 } 3704 3705 set_complete(phase); 3706 return rawmem; 3707 } 3708 3709 3710 #ifdef ASSERT 3711 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3712 if (is_complete()) 3713 return true; // stores could be anything at this point 3714 assert(allocation() != NULL, "must be present"); 3715 intptr_t last_off = allocation()->minimum_header_size(); 3716 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3717 Node* st = in(i); 3718 intptr_t st_off = get_store_offset(st, phase); 3719 if (st_off < 0) continue; // ignore dead garbage 3720 if (last_off > st_off) { 3721 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3722 this->dump(2); 3723 assert(false, "ascending store offsets"); 3724 return false; 3725 } 3726 last_off = st_off + st->as_Store()->memory_size(); 3727 } 3728 return true; 3729 } 3730 #endif //ASSERT 3731 3732 3733 3734 3735 //============================MergeMemNode===================================== 3736 // 3737 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3738 // contributing store or call operations. Each contributor provides the memory 3739 // state for a particular "alias type" (see Compile::alias_type). For example, 3740 // if a MergeMem has an input X for alias category #6, then any memory reference 3741 // to alias category #6 may use X as its memory state input, as an exact equivalent 3742 // to using the MergeMem as a whole. 3743 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3744 // 3745 // (Here, the <N> notation gives the index of the relevant adr_type.) 3746 // 3747 // In one special case (and more cases in the future), alias categories overlap. 3748 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3749 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3750 // it is exactly equivalent to that state W: 3751 // MergeMem(<Bot>: W) <==> W 3752 // 3753 // Usually, the merge has more than one input. In that case, where inputs 3754 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3755 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3756 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3757 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3758 // 3759 // A merge can take a "wide" memory state as one of its narrow inputs. 3760 // This simply means that the merge observes out only the relevant parts of 3761 // the wide input. That is, wide memory states arriving at narrow merge inputs 3762 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3763 // 3764 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3765 // and that memory slices "leak through": 3766 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3767 // 3768 // But, in such a cascade, repeated memory slices can "block the leak": 3769 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3770 // 3771 // In the last example, Y is not part of the combined memory state of the 3772 // outermost MergeMem. The system must, of course, prevent unschedulable 3773 // memory states from arising, so you can be sure that the state Y is somehow 3774 // a precursor to state Y'. 3775 // 3776 // 3777 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3778 // of each MergeMemNode array are exactly the numerical alias indexes, including 3779 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3780 // Compile::alias_type (and kin) produce and manage these indexes. 3781 // 3782 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3783 // (Note that this provides quick access to the top node inside MergeMem methods, 3784 // without the need to reach out via TLS to Compile::current.) 3785 // 3786 // As a consequence of what was just described, a MergeMem that represents a full 3787 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3788 // containing all alias categories. 3789 // 3790 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3791 // 3792 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3793 // a memory state for the alias type <N>, or else the top node, meaning that 3794 // there is no particular input for that alias type. Note that the length of 3795 // a MergeMem is variable, and may be extended at any time to accommodate new 3796 // memory states at larger alias indexes. When merges grow, they are of course 3797 // filled with "top" in the unused in() positions. 3798 // 3799 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3800 // (Top was chosen because it works smoothly with passes like GCM.) 3801 // 3802 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3803 // the type of random VM bits like TLS references.) Since it is always the 3804 // first non-Bot memory slice, some low-level loops use it to initialize an 3805 // index variable: for (i = AliasIdxRaw; i < req(); i++). 3806 // 3807 // 3808 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3809 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3810 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 3811 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3812 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3813 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3814 // 3815 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3816 // really that different from the other memory inputs. An abbreviation called 3817 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3818 // 3819 // 3820 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3821 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3822 // that "emerges though" the base memory will be marked as excluding the alias types 3823 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 3824 // 3825 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3826 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3827 // 3828 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 3829 // (It is currently unimplemented.) As you can see, the resulting merge is 3830 // actually a disjoint union of memory states, rather than an overlay. 3831 // 3832 3833 //------------------------------MergeMemNode----------------------------------- 3834 Node* MergeMemNode::make_empty_memory() { 3835 Node* empty_memory = (Node*) Compile::current()->top(); 3836 assert(empty_memory->is_top(), "correct sentinel identity"); 3837 return empty_memory; 3838 } 3839 3840 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3841 init_class_id(Class_MergeMem); 3842 // all inputs are nullified in Node::Node(int) 3843 // set_input(0, NULL); // no control input 3844 3845 // Initialize the edges uniformly to top, for starters. 3846 Node* empty_mem = make_empty_memory(); 3847 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3848 init_req(i,empty_mem); 3849 } 3850 assert(empty_memory() == empty_mem, ""); 3851 3852 if( new_base != NULL && new_base->is_MergeMem() ) { 3853 MergeMemNode* mdef = new_base->as_MergeMem(); 3854 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3855 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3856 mms.set_memory(mms.memory2()); 3857 } 3858 assert(base_memory() == mdef->base_memory(), ""); 3859 } else { 3860 set_base_memory(new_base); 3861 } 3862 } 3863 3864 // Make a new, untransformed MergeMem with the same base as 'mem'. 3865 // If mem is itself a MergeMem, populate the result with the same edges. 3866 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3867 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3868 } 3869 3870 //------------------------------cmp-------------------------------------------- 3871 uint MergeMemNode::hash() const { return NO_HASH; } 3872 uint MergeMemNode::cmp( const Node &n ) const { 3873 return (&n == this); // Always fail except on self 3874 } 3875 3876 //------------------------------Identity--------------------------------------- 3877 Node* MergeMemNode::Identity(PhaseTransform *phase) { 3878 // Identity if this merge point does not record any interesting memory 3879 // disambiguations. 3880 Node* base_mem = base_memory(); 3881 Node* empty_mem = empty_memory(); 3882 if (base_mem != empty_mem) { // Memory path is not dead? 3883 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3884 Node* mem = in(i); 3885 if (mem != empty_mem && mem != base_mem) { 3886 return this; // Many memory splits; no change 3887 } 3888 } 3889 } 3890 return base_mem; // No memory splits; ID on the one true input 3891 } 3892 3893 //------------------------------Ideal------------------------------------------ 3894 // This method is invoked recursively on chains of MergeMem nodes 3895 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3896 // Remove chain'd MergeMems 3897 // 3898 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3899 // relative to the "in(Bot)". Since we are patching both at the same time, 3900 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3901 // but rewrite each "in(i)" relative to the new "in(Bot)". 3902 Node *progress = NULL; 3903 3904 3905 Node* old_base = base_memory(); 3906 Node* empty_mem = empty_memory(); 3907 if (old_base == empty_mem) 3908 return NULL; // Dead memory path. 3909 3910 MergeMemNode* old_mbase; 3911 if (old_base != NULL && old_base->is_MergeMem()) 3912 old_mbase = old_base->as_MergeMem(); 3913 else 3914 old_mbase = NULL; 3915 Node* new_base = old_base; 3916 3917 // simplify stacked MergeMems in base memory 3918 if (old_mbase) new_base = old_mbase->base_memory(); 3919 3920 // the base memory might contribute new slices beyond my req() 3921 if (old_mbase) grow_to_match(old_mbase); 3922 3923 // Look carefully at the base node if it is a phi. 3924 PhiNode* phi_base; 3925 if (new_base != NULL && new_base->is_Phi()) 3926 phi_base = new_base->as_Phi(); 3927 else 3928 phi_base = NULL; 3929 3930 Node* phi_reg = NULL; 3931 uint phi_len = (uint)-1; 3932 if (phi_base != NULL && !phi_base->is_copy()) { 3933 // do not examine phi if degraded to a copy 3934 phi_reg = phi_base->region(); 3935 phi_len = phi_base->req(); 3936 // see if the phi is unfinished 3937 for (uint i = 1; i < phi_len; i++) { 3938 if (phi_base->in(i) == NULL) { 3939 // incomplete phi; do not look at it yet! 3940 phi_reg = NULL; 3941 phi_len = (uint)-1; 3942 break; 3943 } 3944 } 3945 } 3946 3947 // Note: We do not call verify_sparse on entry, because inputs 3948 // can normalize to the base_memory via subsume_node or similar 3949 // mechanisms. This method repairs that damage. 3950 3951 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3952 3953 // Look at each slice. 3954 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3955 Node* old_in = in(i); 3956 // calculate the old memory value 3957 Node* old_mem = old_in; 3958 if (old_mem == empty_mem) old_mem = old_base; 3959 assert(old_mem == memory_at(i), ""); 3960 3961 // maybe update (reslice) the old memory value 3962 3963 // simplify stacked MergeMems 3964 Node* new_mem = old_mem; 3965 MergeMemNode* old_mmem; 3966 if (old_mem != NULL && old_mem->is_MergeMem()) 3967 old_mmem = old_mem->as_MergeMem(); 3968 else 3969 old_mmem = NULL; 3970 if (old_mmem == this) { 3971 // This can happen if loops break up and safepoints disappear. 3972 // A merge of BotPtr (default) with a RawPtr memory derived from a 3973 // safepoint can be rewritten to a merge of the same BotPtr with 3974 // the BotPtr phi coming into the loop. If that phi disappears 3975 // also, we can end up with a self-loop of the mergemem. 3976 // In general, if loops degenerate and memory effects disappear, 3977 // a mergemem can be left looking at itself. This simply means 3978 // that the mergemem's default should be used, since there is 3979 // no longer any apparent effect on this slice. 3980 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3981 // from start. Update the input to TOP. 3982 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3983 } 3984 else if (old_mmem != NULL) { 3985 new_mem = old_mmem->memory_at(i); 3986 } 3987 // else preceding memory was not a MergeMem 3988 3989 // replace equivalent phis (unfortunately, they do not GVN together) 3990 if (new_mem != NULL && new_mem != new_base && 3991 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3992 if (new_mem->is_Phi()) { 3993 PhiNode* phi_mem = new_mem->as_Phi(); 3994 for (uint i = 1; i < phi_len; i++) { 3995 if (phi_base->in(i) != phi_mem->in(i)) { 3996 phi_mem = NULL; 3997 break; 3998 } 3999 } 4000 if (phi_mem != NULL) { 4001 // equivalent phi nodes; revert to the def 4002 new_mem = new_base; 4003 } 4004 } 4005 } 4006 4007 // maybe store down a new value 4008 Node* new_in = new_mem; 4009 if (new_in == new_base) new_in = empty_mem; 4010 4011 if (new_in != old_in) { 4012 // Warning: Do not combine this "if" with the previous "if" 4013 // A memory slice might have be be rewritten even if it is semantically 4014 // unchanged, if the base_memory value has changed. 4015 set_req(i, new_in); 4016 progress = this; // Report progress 4017 } 4018 } 4019 4020 if (new_base != old_base) { 4021 set_req(Compile::AliasIdxBot, new_base); 4022 // Don't use set_base_memory(new_base), because we need to update du. 4023 assert(base_memory() == new_base, ""); 4024 progress = this; 4025 } 4026 4027 if( base_memory() == this ) { 4028 // a self cycle indicates this memory path is dead 4029 set_req(Compile::AliasIdxBot, empty_mem); 4030 } 4031 4032 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4033 // Recursion must occur after the self cycle check above 4034 if( base_memory()->is_MergeMem() ) { 4035 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4036 Node *m = phase->transform(new_mbase); // Rollup any cycles 4037 if( m != NULL && (m->is_top() || 4038 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4039 // propagate rollup of dead cycle to self 4040 set_req(Compile::AliasIdxBot, empty_mem); 4041 } 4042 } 4043 4044 if( base_memory() == empty_mem ) { 4045 progress = this; 4046 // Cut inputs during Parse phase only. 4047 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4048 if( !can_reshape ) { 4049 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4050 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4051 } 4052 } 4053 } 4054 4055 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4056 // Check if PhiNode::Ideal's "Split phis through memory merges" 4057 // transform should be attempted. Look for this->phi->this cycle. 4058 uint merge_width = req(); 4059 if (merge_width > Compile::AliasIdxRaw) { 4060 PhiNode* phi = base_memory()->as_Phi(); 4061 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4062 if (phi->in(i) == this) { 4063 phase->is_IterGVN()->_worklist.push(phi); 4064 break; 4065 } 4066 } 4067 } 4068 } 4069 4070 assert(progress || verify_sparse(), "please, no dups of base"); 4071 return progress; 4072 } 4073 4074 //-------------------------set_base_memory------------------------------------- 4075 void MergeMemNode::set_base_memory(Node *new_base) { 4076 Node* empty_mem = empty_memory(); 4077 set_req(Compile::AliasIdxBot, new_base); 4078 assert(memory_at(req()) == new_base, "must set default memory"); 4079 // Clear out other occurrences of new_base: 4080 if (new_base != empty_mem) { 4081 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4082 if (in(i) == new_base) set_req(i, empty_mem); 4083 } 4084 } 4085 } 4086 4087 //------------------------------out_RegMask------------------------------------ 4088 const RegMask &MergeMemNode::out_RegMask() const { 4089 return RegMask::Empty; 4090 } 4091 4092 //------------------------------dump_spec-------------------------------------- 4093 #ifndef PRODUCT 4094 void MergeMemNode::dump_spec(outputStream *st) const { 4095 st->print(" {"); 4096 Node* base_mem = base_memory(); 4097 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4098 Node* mem = memory_at(i); 4099 if (mem == base_mem) { st->print(" -"); continue; } 4100 st->print( " N%d:", mem->_idx ); 4101 Compile::current()->get_adr_type(i)->dump_on(st); 4102 } 4103 st->print(" }"); 4104 } 4105 #endif // !PRODUCT 4106 4107 4108 #ifdef ASSERT 4109 static bool might_be_same(Node* a, Node* b) { 4110 if (a == b) return true; 4111 if (!(a->is_Phi() || b->is_Phi())) return false; 4112 // phis shift around during optimization 4113 return true; // pretty stupid... 4114 } 4115 4116 // verify a narrow slice (either incoming or outgoing) 4117 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4118 if (!VerifyAliases) return; // don't bother to verify unless requested 4119 if (is_error_reported()) return; // muzzle asserts when debugging an error 4120 if (Node::in_dump()) return; // muzzle asserts when printing 4121 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4122 assert(n != NULL, ""); 4123 // Elide intervening MergeMem's 4124 while (n->is_MergeMem()) { 4125 n = n->as_MergeMem()->memory_at(alias_idx); 4126 } 4127 Compile* C = Compile::current(); 4128 const TypePtr* n_adr_type = n->adr_type(); 4129 if (n == m->empty_memory()) { 4130 // Implicit copy of base_memory() 4131 } else if (n_adr_type != TypePtr::BOTTOM) { 4132 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4133 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4134 } else { 4135 // A few places like make_runtime_call "know" that VM calls are narrow, 4136 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4137 bool expected_wide_mem = false; 4138 if (n == m->base_memory()) { 4139 expected_wide_mem = true; 4140 } else if (alias_idx == Compile::AliasIdxRaw || 4141 n == m->memory_at(Compile::AliasIdxRaw)) { 4142 expected_wide_mem = true; 4143 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4144 // memory can "leak through" calls on channels that 4145 // are write-once. Allow this also. 4146 expected_wide_mem = true; 4147 } 4148 assert(expected_wide_mem, "expected narrow slice replacement"); 4149 } 4150 } 4151 #else // !ASSERT 4152 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 4153 #endif 4154 4155 4156 //-----------------------------memory_at--------------------------------------- 4157 Node* MergeMemNode::memory_at(uint alias_idx) const { 4158 assert(alias_idx >= Compile::AliasIdxRaw || 4159 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4160 "must avoid base_memory and AliasIdxTop"); 4161 4162 // Otherwise, it is a narrow slice. 4163 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4164 Compile *C = Compile::current(); 4165 if (is_empty_memory(n)) { 4166 // the array is sparse; empty slots are the "top" node 4167 n = base_memory(); 4168 assert(Node::in_dump() 4169 || n == NULL || n->bottom_type() == Type::TOP 4170 || n->adr_type() == NULL // address is TOP 4171 || n->adr_type() == TypePtr::BOTTOM 4172 || n->adr_type() == TypeRawPtr::BOTTOM 4173 || Compile::current()->AliasLevel() == 0, 4174 "must be a wide memory"); 4175 // AliasLevel == 0 if we are organizing the memory states manually. 4176 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4177 } else { 4178 // make sure the stored slice is sane 4179 #ifdef ASSERT 4180 if (is_error_reported() || Node::in_dump()) { 4181 } else if (might_be_same(n, base_memory())) { 4182 // Give it a pass: It is a mostly harmless repetition of the base. 4183 // This can arise normally from node subsumption during optimization. 4184 } else { 4185 verify_memory_slice(this, alias_idx, n); 4186 } 4187 #endif 4188 } 4189 return n; 4190 } 4191 4192 //---------------------------set_memory_at------------------------------------- 4193 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4194 verify_memory_slice(this, alias_idx, n); 4195 Node* empty_mem = empty_memory(); 4196 if (n == base_memory()) n = empty_mem; // collapse default 4197 uint need_req = alias_idx+1; 4198 if (req() < need_req) { 4199 if (n == empty_mem) return; // already the default, so do not grow me 4200 // grow the sparse array 4201 do { 4202 add_req(empty_mem); 4203 } while (req() < need_req); 4204 } 4205 set_req( alias_idx, n ); 4206 } 4207 4208 4209 4210 //--------------------------iteration_setup------------------------------------ 4211 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4212 if (other != NULL) { 4213 grow_to_match(other); 4214 // invariant: the finite support of mm2 is within mm->req() 4215 #ifdef ASSERT 4216 for (uint i = req(); i < other->req(); i++) { 4217 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4218 } 4219 #endif 4220 } 4221 // Replace spurious copies of base_memory by top. 4222 Node* base_mem = base_memory(); 4223 if (base_mem != NULL && !base_mem->is_top()) { 4224 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4225 if (in(i) == base_mem) 4226 set_req(i, empty_memory()); 4227 } 4228 } 4229 } 4230 4231 //---------------------------grow_to_match------------------------------------- 4232 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4233 Node* empty_mem = empty_memory(); 4234 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4235 // look for the finite support of the other memory 4236 for (uint i = other->req(); --i >= req(); ) { 4237 if (other->in(i) != empty_mem) { 4238 uint new_len = i+1; 4239 while (req() < new_len) add_req(empty_mem); 4240 break; 4241 } 4242 } 4243 } 4244 4245 //---------------------------verify_sparse------------------------------------- 4246 #ifndef PRODUCT 4247 bool MergeMemNode::verify_sparse() const { 4248 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4249 Node* base_mem = base_memory(); 4250 // The following can happen in degenerate cases, since empty==top. 4251 if (is_empty_memory(base_mem)) return true; 4252 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4253 assert(in(i) != NULL, "sane slice"); 4254 if (in(i) == base_mem) return false; // should have been the sentinel value! 4255 } 4256 return true; 4257 } 4258 4259 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4260 Node* n; 4261 n = mm->in(idx); 4262 if (mem == n) return true; // might be empty_memory() 4263 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4264 if (mem == n) return true; 4265 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4266 if (mem == n) return true; 4267 if (n == NULL) break; 4268 } 4269 return false; 4270 } 4271 #endif // !PRODUCT