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