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