1 /* 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "oops/objArrayKlass.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/compile.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/machnode.hpp" 36 #include "opto/matcher.hpp" 37 #include "opto/memnode.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/phaseX.hpp" 40 #include "opto/regmask.hpp" 41 42 // Portions of code courtesy of Clifford Click 43 44 // Optimization - Graph Style 45 46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 47 48 //============================================================================= 49 uint MemNode::size_of() const { return sizeof(*this); } 50 51 const TypePtr *MemNode::adr_type() const { 52 Node* adr = in(Address); 53 const TypePtr* cross_check = NULL; 54 DEBUG_ONLY(cross_check = _adr_type); 55 return calculate_adr_type(adr->bottom_type(), cross_check); 56 } 57 58 #ifndef PRODUCT 59 void MemNode::dump_spec(outputStream *st) const { 60 if (in(Address) == NULL) return; // node is dead 61 #ifndef ASSERT 62 // fake the missing field 63 const TypePtr* _adr_type = NULL; 64 if (in(Address) != NULL) 65 _adr_type = in(Address)->bottom_type()->isa_ptr(); 66 #endif 67 dump_adr_type(this, _adr_type, st); 68 69 Compile* C = Compile::current(); 70 if( C->alias_type(_adr_type)->is_volatile() ) 71 st->print(" Volatile!"); 72 } 73 74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 75 st->print(" @"); 76 if (adr_type == NULL) { 77 st->print("NULL"); 78 } else { 79 adr_type->dump_on(st); 80 Compile* C = Compile::current(); 81 Compile::AliasType* atp = NULL; 82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 83 if (atp == NULL) 84 st->print(", idx=?\?;"); 85 else if (atp->index() == Compile::AliasIdxBot) 86 st->print(", idx=Bot;"); 87 else if (atp->index() == Compile::AliasIdxTop) 88 st->print(", idx=Top;"); 89 else if (atp->index() == Compile::AliasIdxRaw) 90 st->print(", idx=Raw;"); 91 else { 92 ciField* field = atp->field(); 93 if (field) { 94 st->print(", name="); 95 field->print_name_on(st); 96 } 97 st->print(", idx=%d;", atp->index()); 98 } 99 } 100 } 101 102 extern void print_alias_types(); 103 104 #endif 105 106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 107 const TypeOopPtr *tinst = t_adr->isa_oopptr(); 108 if (tinst == NULL || !tinst->is_known_instance_field()) 109 return mchain; // don't try to optimize non-instance types 110 uint instance_id = tinst->instance_id(); 111 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 112 Node *prev = NULL; 113 Node *result = mchain; 114 while (prev != result) { 115 prev = result; 116 if (result == start_mem) 117 break; // hit one of our sentinels 118 // skip over a call which does not affect this memory slice 119 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 120 Node *proj_in = result->in(0); 121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 122 break; // hit one of our sentinels 123 } else if (proj_in->is_Call()) { 124 CallNode *call = proj_in->as_Call(); 125 if (!call->may_modify(t_adr, phase)) { 126 result = call->in(TypeFunc::Memory); 127 } 128 } else if (proj_in->is_Initialize()) { 129 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 130 // Stop if this is the initialization for the object instance which 131 // which contains this memory slice, otherwise skip over it. 132 if (alloc != NULL && alloc->_idx != instance_id) { 133 result = proj_in->in(TypeFunc::Memory); 134 } 135 } else if (proj_in->is_MemBar()) { 136 result = proj_in->in(TypeFunc::Memory); 137 } else { 138 assert(false, "unexpected projection"); 139 } 140 } else if (result->is_ClearArray()) { 141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) { 142 // Can not bypass initialization of the instance 143 // we are looking for. 144 break; 145 } 146 // Otherwise skip it (the call updated 'result' value). 147 } else if (result->is_MergeMem()) { 148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 149 } 150 } 151 return result; 152 } 153 154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 155 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); 157 PhaseIterGVN *igvn = phase->is_IterGVN(); 158 Node *result = mchain; 159 result = optimize_simple_memory_chain(result, t_adr, phase); 160 if (is_instance && igvn != NULL && result->is_Phi()) { 161 PhiNode *mphi = result->as_Phi(); 162 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 163 const TypePtr *t = mphi->adr_type(); 164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 166 t->is_oopptr()->cast_to_exactness(true) 167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 168 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 169 // clone the Phi with our address type 170 result = mphi->split_out_instance(t_adr, igvn); 171 } else { 172 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 173 } 174 } 175 return result; 176 } 177 178 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 179 uint alias_idx = phase->C->get_alias_index(tp); 180 Node *mem = mmem; 181 #ifdef ASSERT 182 { 183 // Check that current type is consistent with the alias index used during graph construction 184 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 185 bool consistent = adr_check == NULL || adr_check->empty() || 186 phase->C->must_alias(adr_check, alias_idx ); 187 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 188 if( !consistent && adr_check != NULL && !adr_check->empty() && 189 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 190 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 191 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 192 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 193 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 194 // don't assert if it is dead code. 195 consistent = true; 196 } 197 if( !consistent ) { 198 st->print("alias_idx==%d, adr_check==", alias_idx); 199 if( adr_check == NULL ) { 200 st->print("NULL"); 201 } else { 202 adr_check->dump(); 203 } 204 st->cr(); 205 print_alias_types(); 206 assert(consistent, "adr_check must match alias idx"); 207 } 208 } 209 #endif 210 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 211 // means an array I have not precisely typed yet. Do not do any 212 // alias stuff with it any time soon. 213 const TypeOopPtr *toop = tp->isa_oopptr(); 214 if( tp->base() != Type::AnyPtr && 215 !(toop && 216 toop->klass() != NULL && 217 toop->klass()->is_java_lang_Object() && 218 toop->offset() == Type::OffsetBot) ) { 219 // compress paths and change unreachable cycles to TOP 220 // If not, we can update the input infinitely along a MergeMem cycle 221 // Equivalent code in PhiNode::Ideal 222 Node* m = phase->transform(mmem); 223 // If transformed to a MergeMem, get the desired slice 224 // Otherwise the returned node represents memory for every slice 225 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 226 // Update input if it is progress over what we have now 227 } 228 return mem; 229 } 230 231 //--------------------------Ideal_common--------------------------------------- 232 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 233 // Unhook non-raw memories from complete (macro-expanded) initializations. 234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 235 // If our control input is a dead region, kill all below the region 236 Node *ctl = in(MemNode::Control); 237 if (ctl && remove_dead_region(phase, can_reshape)) 238 return this; 239 ctl = in(MemNode::Control); 240 // Don't bother trying to transform a dead node 241 if( ctl && ctl->is_top() ) return NodeSentinel; 242 243 PhaseIterGVN *igvn = phase->is_IterGVN(); 244 // Wait if control on the worklist. 245 if (ctl && can_reshape && igvn != NULL) { 246 Node* bol = NULL; 247 Node* cmp = NULL; 248 if (ctl->in(0)->is_If()) { 249 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 250 bol = ctl->in(0)->in(1); 251 if (bol->is_Bool()) 252 cmp = ctl->in(0)->in(1)->in(1); 253 } 254 if (igvn->_worklist.member(ctl) || 255 (bol != NULL && igvn->_worklist.member(bol)) || 256 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 257 // This control path may be dead. 258 // Delay this memory node transformation until the control is processed. 259 phase->is_IterGVN()->_worklist.push(this); 260 return NodeSentinel; // caller will return NULL 261 } 262 } 263 // Ignore if memory is dead, or self-loop 264 Node *mem = in(MemNode::Memory); 265 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 266 assert( mem != this, "dead loop in MemNode::Ideal" ); 267 268 Node *address = in(MemNode::Address); 269 const Type *t_adr = phase->type( address ); 270 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 271 272 if( can_reshape && igvn != NULL && 273 (igvn->_worklist.member(address) || 274 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) { 275 // The address's base and type may change when the address is processed. 276 // Delay this mem node transformation until the address is processed. 277 phase->is_IterGVN()->_worklist.push(this); 278 return NodeSentinel; // caller will return NULL 279 } 280 281 // Do NOT remove or optimize the next lines: ensure a new alias index 282 // is allocated for an oop pointer type before Escape Analysis. 283 // Note: C++ will not remove it since the call has side effect. 284 if ( t_adr->isa_oopptr() ) { 285 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 286 } 287 288 #ifdef ASSERT 289 Node* base = NULL; 290 if (address->is_AddP()) 291 base = address->in(AddPNode::Base); 292 assert(base == NULL || t_adr->isa_rawptr() || 293 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); 294 #endif 295 296 // Avoid independent memory operations 297 Node* old_mem = mem; 298 299 // The code which unhooks non-raw memories from complete (macro-expanded) 300 // initializations was removed. After macro-expansion all stores catched 301 // by Initialize node became raw stores and there is no information 302 // which memory slices they modify. So it is unsafe to move any memory 303 // operation above these stores. Also in most cases hooked non-raw memories 304 // were already unhooked by using information from detect_ptr_independence() 305 // and find_previous_store(). 306 307 if (mem->is_MergeMem()) { 308 MergeMemNode* mmem = mem->as_MergeMem(); 309 const TypePtr *tp = t_adr->is_ptr(); 310 311 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 312 } 313 314 if (mem != old_mem) { 315 set_req(MemNode::Memory, mem); 316 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 317 return this; 318 } 319 320 // let the subclass continue analyzing... 321 return NULL; 322 } 323 324 // Helper function for proving some simple control dominations. 325 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 326 // Already assumes that 'dom' is available at 'sub', and that 'sub' 327 // is not a constant (dominated by the method's StartNode). 328 // Used by MemNode::find_previous_store to prove that the 329 // control input of a memory operation predates (dominates) 330 // an allocation it wants to look past. 331 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 332 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 333 return false; // Conservative answer for dead code 334 335 // Check 'dom'. Skip Proj and CatchProj nodes. 336 dom = dom->find_exact_control(dom); 337 if (dom == NULL || dom->is_top()) 338 return false; // Conservative answer for dead code 339 340 if (dom == sub) { 341 // For the case when, for example, 'sub' is Initialize and the original 342 // 'dom' is Proj node of the 'sub'. 343 return false; 344 } 345 346 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 347 return true; 348 349 // 'dom' dominates 'sub' if its control edge and control edges 350 // of all its inputs dominate or equal to sub's control edge. 351 352 // Currently 'sub' is either Allocate, Initialize or Start nodes. 353 // Or Region for the check in LoadNode::Ideal(); 354 // 'sub' should have sub->in(0) != NULL. 355 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 356 sub->is_Region(), "expecting only these nodes"); 357 358 // Get control edge of 'sub'. 359 Node* orig_sub = sub; 360 sub = sub->find_exact_control(sub->in(0)); 361 if (sub == NULL || sub->is_top()) 362 return false; // Conservative answer for dead code 363 364 assert(sub->is_CFG(), "expecting control"); 365 366 if (sub == dom) 367 return true; 368 369 if (sub->is_Start() || sub->is_Root()) 370 return false; 371 372 { 373 // Check all control edges of 'dom'. 374 375 ResourceMark rm; 376 Arena* arena = Thread::current()->resource_area(); 377 Node_List nlist(arena); 378 Unique_Node_List dom_list(arena); 379 380 dom_list.push(dom); 381 bool only_dominating_controls = false; 382 383 for (uint next = 0; next < dom_list.size(); next++) { 384 Node* n = dom_list.at(next); 385 if (n == orig_sub) 386 return false; // One of dom's inputs dominated by sub. 387 if (!n->is_CFG() && n->pinned()) { 388 // Check only own control edge for pinned non-control nodes. 389 n = n->find_exact_control(n->in(0)); 390 if (n == NULL || n->is_top()) 391 return false; // Conservative answer for dead code 392 assert(n->is_CFG(), "expecting control"); 393 dom_list.push(n); 394 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 395 only_dominating_controls = true; 396 } else if (n->is_CFG()) { 397 if (n->dominates(sub, nlist)) 398 only_dominating_controls = true; 399 else 400 return false; 401 } else { 402 // First, own control edge. 403 Node* m = n->find_exact_control(n->in(0)); 404 if (m != NULL) { 405 if (m->is_top()) 406 return false; // Conservative answer for dead code 407 dom_list.push(m); 408 } 409 // Now, the rest of edges. 410 uint cnt = n->req(); 411 for (uint i = 1; i < cnt; i++) { 412 m = n->find_exact_control(n->in(i)); 413 if (m == NULL || m->is_top()) 414 continue; 415 dom_list.push(m); 416 } 417 } 418 } 419 return only_dominating_controls; 420 } 421 } 422 423 //---------------------detect_ptr_independence--------------------------------- 424 // Used by MemNode::find_previous_store to prove that two base 425 // pointers are never equal. 426 // The pointers are accompanied by their associated allocations, 427 // if any, which have been previously discovered by the caller. 428 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 429 Node* p2, AllocateNode* a2, 430 PhaseTransform* phase) { 431 // Attempt to prove that these two pointers cannot be aliased. 432 // They may both manifestly be allocations, and they should differ. 433 // Or, if they are not both allocations, they can be distinct constants. 434 // Otherwise, one is an allocation and the other a pre-existing value. 435 if (a1 == NULL && a2 == NULL) { // neither an allocation 436 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 437 } else if (a1 != NULL && a2 != NULL) { // both allocations 438 return (a1 != a2); 439 } else if (a1 != NULL) { // one allocation a1 440 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 441 return all_controls_dominate(p2, a1); 442 } else { //(a2 != NULL) // one allocation a2 443 return all_controls_dominate(p1, a2); 444 } 445 return false; 446 } 447 448 449 // The logic for reordering loads and stores uses four steps: 450 // (a) Walk carefully past stores and initializations which we 451 // can prove are independent of this load. 452 // (b) Observe that the next memory state makes an exact match 453 // with self (load or store), and locate the relevant store. 454 // (c) Ensure that, if we were to wire self directly to the store, 455 // the optimizer would fold it up somehow. 456 // (d) Do the rewiring, and return, depending on some other part of 457 // the optimizer to fold up the load. 458 // This routine handles steps (a) and (b). Steps (c) and (d) are 459 // specific to loads and stores, so they are handled by the callers. 460 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 461 // 462 Node* MemNode::find_previous_store(PhaseTransform* phase) { 463 Node* ctrl = in(MemNode::Control); 464 Node* adr = in(MemNode::Address); 465 intptr_t offset = 0; 466 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 467 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 468 469 if (offset == Type::OffsetBot) 470 return NULL; // cannot unalias unless there are precise offsets 471 472 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 473 474 intptr_t size_in_bytes = memory_size(); 475 476 Node* mem = in(MemNode::Memory); // start searching here... 477 478 int cnt = 50; // Cycle limiter 479 for (;;) { // While we can dance past unrelated stores... 480 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 481 482 if (mem->is_Store()) { 483 Node* st_adr = mem->in(MemNode::Address); 484 intptr_t st_offset = 0; 485 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 486 if (st_base == NULL) 487 break; // inscrutable pointer 488 if (st_offset != offset && st_offset != Type::OffsetBot) { 489 const int MAX_STORE = BytesPerLong; 490 if (st_offset >= offset + size_in_bytes || 491 st_offset <= offset - MAX_STORE || 492 st_offset <= offset - mem->as_Store()->memory_size()) { 493 // Success: The offsets are provably independent. 494 // (You may ask, why not just test st_offset != offset and be done? 495 // The answer is that stores of different sizes can co-exist 496 // in the same sequence of RawMem effects. We sometimes initialize 497 // a whole 'tile' of array elements with a single jint or jlong.) 498 mem = mem->in(MemNode::Memory); 499 continue; // (a) advance through independent store memory 500 } 501 } 502 if (st_base != base && 503 detect_ptr_independence(base, alloc, 504 st_base, 505 AllocateNode::Ideal_allocation(st_base, phase), 506 phase)) { 507 // Success: The bases are provably independent. 508 mem = mem->in(MemNode::Memory); 509 continue; // (a) advance through independent store memory 510 } 511 512 // (b) At this point, if the bases or offsets do not agree, we lose, 513 // since we have not managed to prove 'this' and 'mem' independent. 514 if (st_base == base && st_offset == offset) { 515 return mem; // let caller handle steps (c), (d) 516 } 517 518 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 519 InitializeNode* st_init = mem->in(0)->as_Initialize(); 520 AllocateNode* st_alloc = st_init->allocation(); 521 if (st_alloc == NULL) 522 break; // something degenerated 523 bool known_identical = false; 524 bool known_independent = false; 525 if (alloc == st_alloc) 526 known_identical = true; 527 else if (alloc != NULL) 528 known_independent = true; 529 else if (all_controls_dominate(this, st_alloc)) 530 known_independent = true; 531 532 if (known_independent) { 533 // The bases are provably independent: Either they are 534 // manifestly distinct allocations, or else the control 535 // of this load dominates the store's allocation. 536 int alias_idx = phase->C->get_alias_index(adr_type()); 537 if (alias_idx == Compile::AliasIdxRaw) { 538 mem = st_alloc->in(TypeFunc::Memory); 539 } else { 540 mem = st_init->memory(alias_idx); 541 } 542 continue; // (a) advance through independent store memory 543 } 544 545 // (b) at this point, if we are not looking at a store initializing 546 // the same allocation we are loading from, we lose. 547 if (known_identical) { 548 // From caller, can_see_stored_value will consult find_captured_store. 549 return mem; // let caller handle steps (c), (d) 550 } 551 552 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 553 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 554 if (mem->is_Proj() && mem->in(0)->is_Call()) { 555 CallNode *call = mem->in(0)->as_Call(); 556 if (!call->may_modify(addr_t, phase)) { 557 mem = call->in(TypeFunc::Memory); 558 continue; // (a) advance through independent call memory 559 } 560 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 561 mem = mem->in(0)->in(TypeFunc::Memory); 562 continue; // (a) advance through independent MemBar memory 563 } else if (mem->is_ClearArray()) { 564 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 565 // (the call updated 'mem' value) 566 continue; // (a) advance through independent allocation memory 567 } else { 568 // Can not bypass initialization of the instance 569 // we are looking for. 570 return mem; 571 } 572 } else if (mem->is_MergeMem()) { 573 int alias_idx = phase->C->get_alias_index(adr_type()); 574 mem = mem->as_MergeMem()->memory_at(alias_idx); 575 continue; // (a) advance through independent MergeMem memory 576 } 577 } 578 579 // Unless there is an explicit 'continue', we must bail out here, 580 // because 'mem' is an inscrutable memory state (e.g., a call). 581 break; 582 } 583 584 return NULL; // bail out 585 } 586 587 //----------------------calculate_adr_type------------------------------------- 588 // Helper function. Notices when the given type of address hits top or bottom. 589 // Also, asserts a cross-check of the type against the expected address type. 590 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 591 if (t == Type::TOP) return NULL; // does not touch memory any more? 592 #ifdef PRODUCT 593 cross_check = NULL; 594 #else 595 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 596 #endif 597 const TypePtr* tp = t->isa_ptr(); 598 if (tp == NULL) { 599 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 600 return TypePtr::BOTTOM; // touches lots of memory 601 } else { 602 #ifdef ASSERT 603 // %%%% [phh] We don't check the alias index if cross_check is 604 // TypeRawPtr::BOTTOM. Needs to be investigated. 605 if (cross_check != NULL && 606 cross_check != TypePtr::BOTTOM && 607 cross_check != TypeRawPtr::BOTTOM) { 608 // Recheck the alias index, to see if it has changed (due to a bug). 609 Compile* C = Compile::current(); 610 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 611 "must stay in the original alias category"); 612 // The type of the address must be contained in the adr_type, 613 // disregarding "null"-ness. 614 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 615 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 616 assert(cross_check->meet(tp_notnull) == cross_check, 617 "real address must not escape from expected memory type"); 618 } 619 #endif 620 return tp; 621 } 622 } 623 624 //------------------------adr_phi_is_loop_invariant---------------------------- 625 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 626 // loop is loop invariant. Make a quick traversal of Phi and associated 627 // CastPP nodes, looking to see if they are a closed group within the loop. 628 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 629 // The idea is that the phi-nest must boil down to only CastPP nodes 630 // with the same data. This implies that any path into the loop already 631 // includes such a CastPP, and so the original cast, whatever its input, 632 // must be covered by an equivalent cast, with an earlier control input. 633 ResourceMark rm; 634 635 // The loop entry input of the phi should be the unique dominating 636 // node for every Phi/CastPP in the loop. 637 Unique_Node_List closure; 638 closure.push(adr_phi->in(LoopNode::EntryControl)); 639 640 // Add the phi node and the cast to the worklist. 641 Unique_Node_List worklist; 642 worklist.push(adr_phi); 643 if( cast != NULL ){ 644 if( !cast->is_ConstraintCast() ) return false; 645 worklist.push(cast); 646 } 647 648 // Begin recursive walk of phi nodes. 649 while( worklist.size() ){ 650 // Take a node off the worklist 651 Node *n = worklist.pop(); 652 if( !closure.member(n) ){ 653 // Add it to the closure. 654 closure.push(n); 655 // Make a sanity check to ensure we don't waste too much time here. 656 if( closure.size() > 20) return false; 657 // This node is OK if: 658 // - it is a cast of an identical value 659 // - or it is a phi node (then we add its inputs to the worklist) 660 // Otherwise, the node is not OK, and we presume the cast is not invariant 661 if( n->is_ConstraintCast() ){ 662 worklist.push(n->in(1)); 663 } else if( n->is_Phi() ) { 664 for( uint i = 1; i < n->req(); i++ ) { 665 worklist.push(n->in(i)); 666 } 667 } else { 668 return false; 669 } 670 } 671 } 672 673 // Quit when the worklist is empty, and we've found no offending nodes. 674 return true; 675 } 676 677 //------------------------------Ideal_DU_postCCP------------------------------- 678 // Find any cast-away of null-ness and keep its control. Null cast-aways are 679 // going away in this pass and we need to make this memory op depend on the 680 // gating null check. 681 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 682 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 683 } 684 685 // I tried to leave the CastPP's in. This makes the graph more accurate in 686 // some sense; we get to keep around the knowledge that an oop is not-null 687 // after some test. Alas, the CastPP's interfere with GVN (some values are 688 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 689 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 690 // some of the more trivial cases in the optimizer. Removing more useless 691 // Phi's started allowing Loads to illegally float above null checks. I gave 692 // up on this approach. CNC 10/20/2000 693 // This static method may be called not from MemNode (EncodePNode calls it). 694 // Only the control edge of the node 'n' might be updated. 695 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 696 Node *skipped_cast = NULL; 697 // Need a null check? Regular static accesses do not because they are 698 // from constant addresses. Array ops are gated by the range check (which 699 // always includes a NULL check). Just check field ops. 700 if( n->in(MemNode::Control) == NULL ) { 701 // Scan upwards for the highest location we can place this memory op. 702 while( true ) { 703 switch( adr->Opcode() ) { 704 705 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 706 adr = adr->in(AddPNode::Base); 707 continue; 708 709 case Op_DecodeN: // No change to NULL-ness, so peek thru 710 adr = adr->in(1); 711 continue; 712 713 case Op_CastPP: 714 // If the CastPP is useless, just peek on through it. 715 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 716 // Remember the cast that we've peeked though. If we peek 717 // through more than one, then we end up remembering the highest 718 // one, that is, if in a loop, the one closest to the top. 719 skipped_cast = adr; 720 adr = adr->in(1); 721 continue; 722 } 723 // CastPP is going away in this pass! We need this memory op to be 724 // control-dependent on the test that is guarding the CastPP. 725 ccp->hash_delete(n); 726 n->set_req(MemNode::Control, adr->in(0)); 727 ccp->hash_insert(n); 728 return n; 729 730 case Op_Phi: 731 // Attempt to float above a Phi to some dominating point. 732 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 733 // If we've already peeked through a Cast (which could have set the 734 // control), we can't float above a Phi, because the skipped Cast 735 // may not be loop invariant. 736 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 737 adr = adr->in(1); 738 continue; 739 } 740 } 741 742 // Intentional fallthrough! 743 744 // No obvious dominating point. The mem op is pinned below the Phi 745 // by the Phi itself. If the Phi goes away (no true value is merged) 746 // then the mem op can float, but not indefinitely. It must be pinned 747 // behind the controls leading to the Phi. 748 case Op_CheckCastPP: 749 // These usually stick around to change address type, however a 750 // useless one can be elided and we still need to pick up a control edge 751 if (adr->in(0) == NULL) { 752 // This CheckCastPP node has NO control and is likely useless. But we 753 // need check further up the ancestor chain for a control input to keep 754 // the node in place. 4959717. 755 skipped_cast = adr; 756 adr = adr->in(1); 757 continue; 758 } 759 ccp->hash_delete(n); 760 n->set_req(MemNode::Control, adr->in(0)); 761 ccp->hash_insert(n); 762 return n; 763 764 // List of "safe" opcodes; those that implicitly block the memory 765 // op below any null check. 766 case Op_CastX2P: // no null checks on native pointers 767 case Op_Parm: // 'this' pointer is not null 768 case Op_LoadP: // Loading from within a klass 769 case Op_LoadN: // Loading from within a klass 770 case Op_LoadKlass: // Loading from within a klass 771 case Op_LoadNKlass: // Loading from within a klass 772 case Op_ConP: // Loading from a klass 773 case Op_ConN: // Loading from a klass 774 case Op_CreateEx: // Sucking up the guts of an exception oop 775 case Op_Con: // Reading from TLS 776 case Op_CMoveP: // CMoveP is pinned 777 case Op_CMoveN: // CMoveN is pinned 778 break; // No progress 779 780 case Op_Proj: // Direct call to an allocation routine 781 case Op_SCMemProj: // Memory state from store conditional ops 782 #ifdef ASSERT 783 { 784 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 785 const Node* call = adr->in(0); 786 if (call->is_CallJava()) { 787 const CallJavaNode* call_java = call->as_CallJava(); 788 const TypeTuple *r = call_java->tf()->range(); 789 assert(r->cnt() > TypeFunc::Parms, "must return value"); 790 const Type* ret_type = r->field_at(TypeFunc::Parms); 791 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 792 // We further presume that this is one of 793 // new_instance_Java, new_array_Java, or 794 // the like, but do not assert for this. 795 } else if (call->is_Allocate()) { 796 // similar case to new_instance_Java, etc. 797 } else if (!call->is_CallLeaf()) { 798 // Projections from fetch_oop (OSR) are allowed as well. 799 ShouldNotReachHere(); 800 } 801 } 802 #endif 803 break; 804 default: 805 ShouldNotReachHere(); 806 } 807 break; 808 } 809 } 810 811 return NULL; // No progress 812 } 813 814 815 //============================================================================= 816 uint LoadNode::size_of() const { return sizeof(*this); } 817 uint LoadNode::cmp( const Node &n ) const 818 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 819 const Type *LoadNode::bottom_type() const { return _type; } 820 uint LoadNode::ideal_reg() const { 821 return Matcher::base2reg[_type->base()]; 822 } 823 824 #ifndef PRODUCT 825 void LoadNode::dump_spec(outputStream *st) const { 826 MemNode::dump_spec(st); 827 if( !Verbose && !WizardMode ) { 828 // standard dump does this in Verbose and WizardMode 829 st->print(" #"); _type->dump_on(st); 830 } 831 } 832 #endif 833 834 #ifdef ASSERT 835 //----------------------------is_immutable_value------------------------------- 836 // Helper function to allow a raw load without control edge for some cases 837 bool LoadNode::is_immutable_value(Node* adr) { 838 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 839 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 840 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 841 in_bytes(JavaThread::osthread_offset()))); 842 } 843 #endif 844 845 //----------------------------LoadNode::make----------------------------------- 846 // Polymorphic factory method: 847 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 848 Compile* C = gvn.C; 849 850 // sanity check the alias category against the created node type 851 assert(!(adr_type->isa_oopptr() && 852 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 853 "use LoadKlassNode instead"); 854 assert(!(adr_type->isa_aryptr() && 855 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 856 "use LoadRangeNode instead"); 857 // Check control edge of raw loads 858 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 859 // oop will be recorded in oop map if load crosses safepoint 860 rt->isa_oopptr() || is_immutable_value(adr), 861 "raw memory operations should have control edge"); 862 switch (bt) { 863 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 864 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 865 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 866 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 867 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 868 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 869 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 870 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 871 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 872 case T_OBJECT: 873 #ifdef _LP64 874 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 875 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 876 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 877 } else 878 #endif 879 { 880 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 881 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 882 } 883 } 884 ShouldNotReachHere(); 885 return (LoadNode*)NULL; 886 } 887 888 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 889 bool require_atomic = true; 890 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 891 } 892 893 894 895 896 //------------------------------hash------------------------------------------- 897 uint LoadNode::hash() const { 898 // unroll addition of interesting fields 899 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 900 } 901 902 //---------------------------can_see_stored_value------------------------------ 903 // This routine exists to make sure this set of tests is done the same 904 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 905 // will change the graph shape in a way which makes memory alive twice at the 906 // same time (uses the Oracle model of aliasing), then some 907 // LoadXNode::Identity will fold things back to the equivalence-class model 908 // of aliasing. 909 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 910 Node* ld_adr = in(MemNode::Address); 911 912 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 913 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 914 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 915 atp->field() != NULL && !atp->field()->is_volatile()) { 916 uint alias_idx = atp->index(); 917 bool final = atp->field()->is_final(); 918 Node* result = NULL; 919 Node* current = st; 920 // Skip through chains of MemBarNodes checking the MergeMems for 921 // new states for the slice of this load. Stop once any other 922 // kind of node is encountered. Loads from final memory can skip 923 // through any kind of MemBar but normal loads shouldn't skip 924 // through MemBarAcquire since the could allow them to move out of 925 // a synchronized region. 926 while (current->is_Proj()) { 927 int opc = current->in(0)->Opcode(); 928 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) || 929 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder || 930 opc == Op_MemBarReleaseLock) { 931 Node* mem = current->in(0)->in(TypeFunc::Memory); 932 if (mem->is_MergeMem()) { 933 MergeMemNode* merge = mem->as_MergeMem(); 934 Node* new_st = merge->memory_at(alias_idx); 935 if (new_st == merge->base_memory()) { 936 // Keep searching 937 current = merge->base_memory(); 938 continue; 939 } 940 // Save the new memory state for the slice and fall through 941 // to exit. 942 result = new_st; 943 } 944 } 945 break; 946 } 947 if (result != NULL) { 948 st = result; 949 } 950 } 951 952 953 // Loop around twice in the case Load -> Initialize -> Store. 954 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 955 for (int trip = 0; trip <= 1; trip++) { 956 957 if (st->is_Store()) { 958 Node* st_adr = st->in(MemNode::Address); 959 if (!phase->eqv(st_adr, ld_adr)) { 960 // Try harder before giving up... Match raw and non-raw pointers. 961 intptr_t st_off = 0; 962 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 963 if (alloc == NULL) return NULL; 964 intptr_t ld_off = 0; 965 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 966 if (alloc != allo2) return NULL; 967 if (ld_off != st_off) return NULL; 968 // At this point we have proven something like this setup: 969 // A = Allocate(...) 970 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 971 // S = StoreQ(, AddP(, A.Parm , #Off), V) 972 // (Actually, we haven't yet proven the Q's are the same.) 973 // In other words, we are loading from a casted version of 974 // the same pointer-and-offset that we stored to. 975 // Thus, we are able to replace L by V. 976 } 977 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 978 if (store_Opcode() != st->Opcode()) 979 return NULL; 980 return st->in(MemNode::ValueIn); 981 } 982 983 intptr_t offset = 0; // scratch 984 985 // A load from a freshly-created object always returns zero. 986 // (This can happen after LoadNode::Ideal resets the load's memory input 987 // to find_captured_store, which returned InitializeNode::zero_memory.) 988 if (st->is_Proj() && st->in(0)->is_Allocate() && 989 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 990 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 991 // return a zero value for the load's basic type 992 // (This is one of the few places where a generic PhaseTransform 993 // can create new nodes. Think of it as lazily manifesting 994 // virtually pre-existing constants.) 995 return phase->zerocon(memory_type()); 996 } 997 998 // A load from an initialization barrier can match a captured store. 999 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1000 InitializeNode* init = st->in(0)->as_Initialize(); 1001 AllocateNode* alloc = init->allocation(); 1002 if (alloc != NULL && 1003 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 1004 // examine a captured store value 1005 st = init->find_captured_store(offset, memory_size(), phase); 1006 if (st != NULL) 1007 continue; // take one more trip around 1008 } 1009 } 1010 1011 break; 1012 } 1013 1014 return NULL; 1015 } 1016 1017 //----------------------is_instance_field_load_with_local_phi------------------ 1018 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1019 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 1020 in(MemNode::Address)->is_AddP() ) { 1021 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 1022 // Only instances. 1023 if( t_oop != NULL && t_oop->is_known_instance_field() && 1024 t_oop->offset() != Type::OffsetBot && 1025 t_oop->offset() != Type::OffsetTop) { 1026 return true; 1027 } 1028 } 1029 return false; 1030 } 1031 1032 //------------------------------Identity--------------------------------------- 1033 // Loads are identity if previous store is to same address 1034 Node *LoadNode::Identity( PhaseTransform *phase ) { 1035 // If the previous store-maker is the right kind of Store, and the store is 1036 // to the same address, then we are equal to the value stored. 1037 Node* mem = in(MemNode::Memory); 1038 Node* value = can_see_stored_value(mem, phase); 1039 if( value ) { 1040 // byte, short & char stores truncate naturally. 1041 // A load has to load the truncated value which requires 1042 // some sort of masking operation and that requires an 1043 // Ideal call instead of an Identity call. 1044 if (memory_size() < BytesPerInt) { 1045 // If the input to the store does not fit with the load's result type, 1046 // it must be truncated via an Ideal call. 1047 if (!phase->type(value)->higher_equal(phase->type(this))) 1048 return this; 1049 } 1050 // (This works even when value is a Con, but LoadNode::Value 1051 // usually runs first, producing the singleton type of the Con.) 1052 return value; 1053 } 1054 1055 // Search for an existing data phi which was generated before for the same 1056 // instance's field to avoid infinite generation of phis in a loop. 1057 Node *region = mem->in(0); 1058 if (is_instance_field_load_with_local_phi(region)) { 1059 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 1060 int this_index = phase->C->get_alias_index(addr_t); 1061 int this_offset = addr_t->offset(); 1062 int this_id = addr_t->is_oopptr()->instance_id(); 1063 const Type* this_type = bottom_type(); 1064 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1065 Node* phi = region->fast_out(i); 1066 if (phi->is_Phi() && phi != mem && 1067 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 1068 return phi; 1069 } 1070 } 1071 } 1072 1073 return this; 1074 } 1075 1076 1077 // Returns true if the AliasType refers to the field that holds the 1078 // cached box array. Currently only handles the IntegerCache case. 1079 static bool is_autobox_cache(Compile::AliasType* atp) { 1080 if (atp != NULL && atp->field() != NULL) { 1081 ciField* field = atp->field(); 1082 ciSymbol* klass = field->holder()->name(); 1083 if (field->name() == ciSymbol::cache_field_name() && 1084 field->holder()->uses_default_loader() && 1085 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1086 return true; 1087 } 1088 } 1089 return false; 1090 } 1091 1092 // Fetch the base value in the autobox array 1093 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1094 if (atp != NULL && atp->field() != NULL) { 1095 ciField* field = atp->field(); 1096 ciSymbol* klass = field->holder()->name(); 1097 if (field->name() == ciSymbol::cache_field_name() && 1098 field->holder()->uses_default_loader() && 1099 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1100 assert(field->is_constant(), "what?"); 1101 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1102 // Fetch the box object at the base of the array and get its value 1103 ciInstance* box = array->obj_at(0)->as_instance(); 1104 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1105 if (ik->nof_nonstatic_fields() == 1) { 1106 // This should be true nonstatic_field_at requires calling 1107 // nof_nonstatic_fields so check it anyway 1108 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1109 cache_offset = c.as_int(); 1110 } 1111 return true; 1112 } 1113 } 1114 return false; 1115 } 1116 1117 // Returns true if the AliasType refers to the value field of an 1118 // autobox object. Currently only handles Integer. 1119 static bool is_autobox_object(Compile::AliasType* atp) { 1120 if (atp != NULL && atp->field() != NULL) { 1121 ciField* field = atp->field(); 1122 ciSymbol* klass = field->holder()->name(); 1123 if (field->name() == ciSymbol::value_name() && 1124 field->holder()->uses_default_loader() && 1125 klass == ciSymbol::java_lang_Integer()) { 1126 return true; 1127 } 1128 } 1129 return false; 1130 } 1131 1132 1133 // We're loading from an object which has autobox behaviour. 1134 // If this object is result of a valueOf call we'll have a phi 1135 // merging a newly allocated object and a load from the cache. 1136 // We want to replace this load with the original incoming 1137 // argument to the valueOf call. 1138 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1139 Node* base = in(Address)->in(AddPNode::Base); 1140 if (base->is_Phi() && base->req() == 3) { 1141 AllocateNode* allocation = NULL; 1142 int allocation_index = -1; 1143 int load_index = -1; 1144 for (uint i = 1; i < base->req(); i++) { 1145 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1146 if (allocation != NULL) { 1147 allocation_index = i; 1148 load_index = 3 - allocation_index; 1149 break; 1150 } 1151 } 1152 bool has_load = ( allocation != NULL && 1153 (base->in(load_index)->is_Load() || 1154 base->in(load_index)->is_DecodeN() && 1155 base->in(load_index)->in(1)->is_Load()) ); 1156 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1157 // Push the loads from the phi that comes from valueOf up 1158 // through it to allow elimination of the loads and the recovery 1159 // of the original value. 1160 Node* mem_phi = in(Memory); 1161 Node* offset = in(Address)->in(AddPNode::Offset); 1162 Node* region = base->in(0); 1163 1164 Node* in1 = clone(); 1165 Node* in1_addr = in1->in(Address)->clone(); 1166 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1167 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1168 in1_addr->set_req(AddPNode::Offset, offset); 1169 in1->set_req(0, region->in(allocation_index)); 1170 in1->set_req(Address, in1_addr); 1171 in1->set_req(Memory, mem_phi->in(allocation_index)); 1172 1173 Node* in2 = clone(); 1174 Node* in2_addr = in2->in(Address)->clone(); 1175 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1176 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1177 in2_addr->set_req(AddPNode::Offset, offset); 1178 in2->set_req(0, region->in(load_index)); 1179 in2->set_req(Address, in2_addr); 1180 in2->set_req(Memory, mem_phi->in(load_index)); 1181 1182 in1_addr = phase->transform(in1_addr); 1183 in1 = phase->transform(in1); 1184 in2_addr = phase->transform(in2_addr); 1185 in2 = phase->transform(in2); 1186 1187 PhiNode* result = PhiNode::make_blank(region, this); 1188 result->set_req(allocation_index, in1); 1189 result->set_req(load_index, in2); 1190 return result; 1191 } 1192 } else if (base->is_Load() || 1193 base->is_DecodeN() && base->in(1)->is_Load()) { 1194 if (base->is_DecodeN()) { 1195 // Get LoadN node which loads cached Integer object 1196 base = base->in(1); 1197 } 1198 // Eliminate the load of Integer.value for integers from the cache 1199 // array by deriving the value from the index into the array. 1200 // Capture the offset of the load and then reverse the computation. 1201 Node* load_base = base->in(Address)->in(AddPNode::Base); 1202 if (load_base->is_DecodeN()) { 1203 // Get LoadN node which loads IntegerCache.cache field 1204 load_base = load_base->in(1); 1205 } 1206 if (load_base != NULL) { 1207 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1208 intptr_t cache_offset; 1209 int shift = -1; 1210 Node* cache = NULL; 1211 if (is_autobox_cache(atp)) { 1212 shift = exact_log2(type2aelembytes(T_OBJECT)); 1213 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1214 } 1215 if (cache != NULL && base->in(Address)->is_AddP()) { 1216 Node* elements[4]; 1217 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1218 int cache_low; 1219 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1220 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1221 // Add up all the offsets making of the address of the load 1222 Node* result = elements[0]; 1223 for (int i = 1; i < count; i++) { 1224 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1225 } 1226 // Remove the constant offset from the address and then 1227 // remove the scaling of the offset to recover the original index. 1228 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1229 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1230 // Peel the shift off directly but wrap it in a dummy node 1231 // since Ideal can't return existing nodes 1232 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1233 } else { 1234 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1235 } 1236 #ifdef _LP64 1237 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1238 #endif 1239 return result; 1240 } 1241 } 1242 } 1243 } 1244 return NULL; 1245 } 1246 1247 //------------------------------split_through_phi------------------------------ 1248 // Split instance field load through Phi. 1249 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1250 Node* mem = in(MemNode::Memory); 1251 Node* address = in(MemNode::Address); 1252 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1253 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1254 1255 assert(mem->is_Phi() && (t_oop != NULL) && 1256 t_oop->is_known_instance_field(), "invalide conditions"); 1257 1258 Node *region = mem->in(0); 1259 if (region == NULL) { 1260 return NULL; // Wait stable graph 1261 } 1262 uint cnt = mem->req(); 1263 for (uint i = 1; i < cnt; i++) { 1264 Node* rc = region->in(i); 1265 if (rc == NULL || phase->type(rc) == Type::TOP) 1266 return NULL; // Wait stable graph 1267 Node *in = mem->in(i); 1268 if (in == NULL) { 1269 return NULL; // Wait stable graph 1270 } 1271 } 1272 // Check for loop invariant. 1273 if (cnt == 3) { 1274 for (uint i = 1; i < cnt; i++) { 1275 Node *in = mem->in(i); 1276 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1277 if (m == mem) { 1278 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1279 return this; 1280 } 1281 } 1282 } 1283 // Split through Phi (see original code in loopopts.cpp). 1284 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1285 1286 // Do nothing here if Identity will find a value 1287 // (to avoid infinite chain of value phis generation). 1288 if (!phase->eqv(this, this->Identity(phase))) 1289 return NULL; 1290 1291 // Skip the split if the region dominates some control edge of the address. 1292 if (!MemNode::all_controls_dominate(address, region)) 1293 return NULL; 1294 1295 const Type* this_type = this->bottom_type(); 1296 int this_index = phase->C->get_alias_index(addr_t); 1297 int this_offset = addr_t->offset(); 1298 int this_iid = addr_t->is_oopptr()->instance_id(); 1299 PhaseIterGVN *igvn = phase->is_IterGVN(); 1300 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1301 for (uint i = 1; i < region->req(); i++) { 1302 Node *x; 1303 Node* the_clone = NULL; 1304 if (region->in(i) == phase->C->top()) { 1305 x = phase->C->top(); // Dead path? Use a dead data op 1306 } else { 1307 x = this->clone(); // Else clone up the data op 1308 the_clone = x; // Remember for possible deletion. 1309 // Alter data node to use pre-phi inputs 1310 if (this->in(0) == region) { 1311 x->set_req(0, region->in(i)); 1312 } else { 1313 x->set_req(0, NULL); 1314 } 1315 for (uint j = 1; j < this->req(); j++) { 1316 Node *in = this->in(j); 1317 if (in->is_Phi() && in->in(0) == region) 1318 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 1319 } 1320 } 1321 // Check for a 'win' on some paths 1322 const Type *t = x->Value(igvn); 1323 1324 bool singleton = t->singleton(); 1325 1326 // See comments in PhaseIdealLoop::split_thru_phi(). 1327 if (singleton && t == Type::TOP) { 1328 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1329 } 1330 1331 if (singleton) { 1332 x = igvn->makecon(t); 1333 } else { 1334 // We now call Identity to try to simplify the cloned node. 1335 // Note that some Identity methods call phase->type(this). 1336 // Make sure that the type array is big enough for 1337 // our new node, even though we may throw the node away. 1338 // (This tweaking with igvn only works because x is a new node.) 1339 igvn->set_type(x, t); 1340 // If x is a TypeNode, capture any more-precise type permanently into Node 1341 // otherwise it will be not updated during igvn->transform since 1342 // igvn->type(x) is set to x->Value() already. 1343 x->raise_bottom_type(t); 1344 Node *y = x->Identity(igvn); 1345 if (y != x) { 1346 x = y; 1347 } else { 1348 y = igvn->hash_find(x); 1349 if (y) { 1350 x = y; 1351 } else { 1352 // Else x is a new node we are keeping 1353 // We do not need register_new_node_with_optimizer 1354 // because set_type has already been called. 1355 igvn->_worklist.push(x); 1356 } 1357 } 1358 } 1359 if (x != the_clone && the_clone != NULL) 1360 igvn->remove_dead_node(the_clone); 1361 phi->set_req(i, x); 1362 } 1363 // Record Phi 1364 igvn->register_new_node_with_optimizer(phi); 1365 return phi; 1366 } 1367 1368 //------------------------------Ideal------------------------------------------ 1369 // If the load is from Field memory and the pointer is non-null, we can 1370 // zero out the control input. 1371 // If the offset is constant and the base is an object allocation, 1372 // try to hook me up to the exact initializing store. 1373 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1374 Node* p = MemNode::Ideal_common(phase, can_reshape); 1375 if (p) return (p == NodeSentinel) ? NULL : p; 1376 1377 Node* ctrl = in(MemNode::Control); 1378 Node* address = in(MemNode::Address); 1379 1380 // Skip up past a SafePoint control. Cannot do this for Stores because 1381 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1382 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1383 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1384 ctrl = ctrl->in(0); 1385 set_req(MemNode::Control,ctrl); 1386 } 1387 1388 intptr_t ignore = 0; 1389 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1390 if (base != NULL 1391 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1392 // Check for useless control edge in some common special cases 1393 if (in(MemNode::Control) != NULL 1394 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1395 && all_controls_dominate(base, phase->C->start())) { 1396 // A method-invariant, non-null address (constant or 'this' argument). 1397 set_req(MemNode::Control, NULL); 1398 } 1399 1400 if (EliminateAutoBox && can_reshape) { 1401 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); 1402 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1403 if (is_autobox_object(atp)) { 1404 Node* result = eliminate_autobox(phase); 1405 if (result != NULL) return result; 1406 } 1407 } 1408 } 1409 1410 Node* mem = in(MemNode::Memory); 1411 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1412 1413 if (addr_t != NULL) { 1414 // try to optimize our memory input 1415 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1416 if (opt_mem != mem) { 1417 set_req(MemNode::Memory, opt_mem); 1418 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1419 return this; 1420 } 1421 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1422 if (can_reshape && opt_mem->is_Phi() && 1423 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1424 PhaseIterGVN *igvn = phase->is_IterGVN(); 1425 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1426 // Delay this transformation until memory Phi is processed. 1427 phase->is_IterGVN()->_worklist.push(this); 1428 return NULL; 1429 } 1430 // Split instance field load through Phi. 1431 Node* result = split_through_phi(phase); 1432 if (result != NULL) return result; 1433 } 1434 } 1435 1436 // Check for prior store with a different base or offset; make Load 1437 // independent. Skip through any number of them. Bail out if the stores 1438 // are in an endless dead cycle and report no progress. This is a key 1439 // transform for Reflection. However, if after skipping through the Stores 1440 // we can't then fold up against a prior store do NOT do the transform as 1441 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1442 // array memory alive twice: once for the hoisted Load and again after the 1443 // bypassed Store. This situation only works if EVERYBODY who does 1444 // anti-dependence work knows how to bypass. I.e. we need all 1445 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1446 // the alias index stuff. So instead, peek through Stores and IFF we can 1447 // fold up, do so. 1448 Node* prev_mem = find_previous_store(phase); 1449 // Steps (a), (b): Walk past independent stores to find an exact match. 1450 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1451 // (c) See if we can fold up on the spot, but don't fold up here. 1452 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1453 // just return a prior value, which is done by Identity calls. 1454 if (can_see_stored_value(prev_mem, phase)) { 1455 // Make ready for step (d): 1456 set_req(MemNode::Memory, prev_mem); 1457 return this; 1458 } 1459 } 1460 1461 return NULL; // No further progress 1462 } 1463 1464 // Helper to recognize certain Klass fields which are invariant across 1465 // some group of array types (e.g., int[] or all T[] where T < Object). 1466 const Type* 1467 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1468 ciKlass* klass) const { 1469 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1470 // The field is Klass::_modifier_flags. Return its (constant) value. 1471 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1472 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1473 return TypeInt::make(klass->modifier_flags()); 1474 } 1475 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1476 // The field is Klass::_access_flags. Return its (constant) value. 1477 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1478 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1479 return TypeInt::make(klass->access_flags()); 1480 } 1481 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1482 // The field is Klass::_layout_helper. Return its constant value if known. 1483 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1484 return TypeInt::make(klass->layout_helper()); 1485 } 1486 1487 // No match. 1488 return NULL; 1489 } 1490 1491 //------------------------------Value----------------------------------------- 1492 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1493 // Either input is TOP ==> the result is TOP 1494 Node* mem = in(MemNode::Memory); 1495 const Type *t1 = phase->type(mem); 1496 if (t1 == Type::TOP) return Type::TOP; 1497 Node* adr = in(MemNode::Address); 1498 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1499 if (tp == NULL || tp->empty()) return Type::TOP; 1500 int off = tp->offset(); 1501 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1502 Compile* C = phase->C; 1503 1504 // Try to guess loaded type from pointer type 1505 if (tp->base() == Type::AryPtr) { 1506 const Type *t = tp->is_aryptr()->elem(); 1507 // Don't do this for integer types. There is only potential profit if 1508 // the element type t is lower than _type; that is, for int types, if _type is 1509 // more restrictive than t. This only happens here if one is short and the other 1510 // char (both 16 bits), and in those cases we've made an intentional decision 1511 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1512 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1513 // 1514 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1515 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1516 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1517 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1518 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1519 // In fact, that could have been the original type of p1, and p1 could have 1520 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1521 // expression (LShiftL quux 3) independently optimized to the constant 8. 1522 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1523 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1524 // t might actually be lower than _type, if _type is a unique 1525 // concrete subclass of abstract class t. 1526 // Make sure the reference is not into the header, by comparing 1527 // the offset against the offset of the start of the array's data. 1528 // Different array types begin at slightly different offsets (12 vs. 16). 1529 // We choose T_BYTE as an example base type that is least restrictive 1530 // as to alignment, which will therefore produce the smallest 1531 // possible base offset. 1532 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1533 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1534 const Type* jt = t->join(_type); 1535 // In any case, do not allow the join, per se, to empty out the type. 1536 if (jt->empty() && !t->empty()) { 1537 // This can happen if a interface-typed array narrows to a class type. 1538 jt = _type; 1539 } 1540 1541 if (EliminateAutoBox && adr->is_AddP()) { 1542 // The pointers in the autobox arrays are always non-null 1543 Node* base = adr->in(AddPNode::Base); 1544 if (base != NULL && 1545 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { 1546 Compile::AliasType* atp = C->alias_type(base->adr_type()); 1547 if (is_autobox_cache(atp)) { 1548 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1549 } 1550 } 1551 } 1552 return jt; 1553 } 1554 } 1555 } else if (tp->base() == Type::InstPtr) { 1556 ciEnv* env = C->env(); 1557 const TypeInstPtr* tinst = tp->is_instptr(); 1558 ciKlass* klass = tinst->klass(); 1559 assert( off != Type::OffsetBot || 1560 // arrays can be cast to Objects 1561 tp->is_oopptr()->klass()->is_java_lang_Object() || 1562 // unsafe field access may not have a constant offset 1563 C->has_unsafe_access(), 1564 "Field accesses must be precise" ); 1565 // For oop loads, we expect the _type to be precise 1566 if (klass == env->String_klass() && 1567 adr->is_AddP() && off != Type::OffsetBot) { 1568 // For constant Strings treat the final fields as compile time constants. 1569 Node* base = adr->in(AddPNode::Base); 1570 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1571 if (t != NULL && t->singleton()) { 1572 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1573 if (field != NULL && field->is_final()) { 1574 ciObject* string = t->const_oop(); 1575 ciConstant constant = string->as_instance()->field_value(field); 1576 if (constant.basic_type() == T_INT) { 1577 return TypeInt::make(constant.as_int()); 1578 } else if (constant.basic_type() == T_ARRAY) { 1579 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1580 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1581 } else { 1582 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1583 } 1584 } 1585 } 1586 } 1587 } 1588 // Optimizations for constant objects 1589 ciObject* const_oop = tinst->const_oop(); 1590 if (const_oop != NULL) { 1591 // For constant CallSites treat the target field as a compile time constant. 1592 if (const_oop->is_call_site()) { 1593 ciCallSite* call_site = const_oop->as_call_site(); 1594 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1595 if (field != NULL && field->is_call_site_target()) { 1596 ciMethodHandle* target = call_site->get_target(); 1597 if (target != NULL) { // just in case 1598 ciConstant constant(T_OBJECT, target); 1599 const Type* t; 1600 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1601 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1602 } else { 1603 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1604 } 1605 // Add a dependence for invalidation of the optimization. 1606 if (!call_site->is_constant_call_site()) { 1607 C->dependencies()->assert_call_site_target_value(call_site, target); 1608 } 1609 return t; 1610 } 1611 } 1612 } 1613 } 1614 } else if (tp->base() == Type::KlassPtr) { 1615 assert( off != Type::OffsetBot || 1616 // arrays can be cast to Objects 1617 tp->is_klassptr()->klass()->is_java_lang_Object() || 1618 // also allow array-loading from the primary supertype 1619 // array during subtype checks 1620 Opcode() == Op_LoadKlass, 1621 "Field accesses must be precise" ); 1622 // For klass/static loads, we expect the _type to be precise 1623 } 1624 1625 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1626 if (tkls != NULL && !StressReflectiveCode) { 1627 ciKlass* klass = tkls->klass(); 1628 if (klass->is_loaded() && tkls->klass_is_exact()) { 1629 // We are loading a field from a Klass metaobject whose identity 1630 // is known at compile time (the type is "exact" or "precise"). 1631 // Check for fields we know are maintained as constants by the VM. 1632 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1633 // The field is Klass::_super_check_offset. Return its (constant) value. 1634 // (Folds up type checking code.) 1635 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1636 return TypeInt::make(klass->super_check_offset()); 1637 } 1638 // Compute index into primary_supers array 1639 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1640 // Check for overflowing; use unsigned compare to handle the negative case. 1641 if( depth < ciKlass::primary_super_limit() ) { 1642 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1643 // (Folds up type checking code.) 1644 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1645 ciKlass *ss = klass->super_of_depth(depth); 1646 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1647 } 1648 const Type* aift = load_array_final_field(tkls, klass); 1649 if (aift != NULL) return aift; 1650 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1651 && klass->is_array_klass()) { 1652 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1653 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1654 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1655 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1656 } 1657 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1658 // The field is Klass::_java_mirror. Return its (constant) value. 1659 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1660 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1661 return TypeInstPtr::make(klass->java_mirror()); 1662 } 1663 } 1664 1665 // We can still check if we are loading from the primary_supers array at a 1666 // shallow enough depth. Even though the klass is not exact, entries less 1667 // than or equal to its super depth are correct. 1668 if (klass->is_loaded() ) { 1669 ciType *inner = klass->klass(); 1670 while( inner->is_obj_array_klass() ) 1671 inner = inner->as_obj_array_klass()->base_element_type(); 1672 if( inner->is_instance_klass() && 1673 !inner->as_instance_klass()->flags().is_interface() ) { 1674 // Compute index into primary_supers array 1675 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1676 // Check for overflowing; use unsigned compare to handle the negative case. 1677 if( depth < ciKlass::primary_super_limit() && 1678 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1679 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1680 // (Folds up type checking code.) 1681 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1682 ciKlass *ss = klass->super_of_depth(depth); 1683 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1684 } 1685 } 1686 } 1687 1688 // If the type is enough to determine that the thing is not an array, 1689 // we can give the layout_helper a positive interval type. 1690 // This will help short-circuit some reflective code. 1691 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1692 && !klass->is_array_klass() // not directly typed as an array 1693 && !klass->is_interface() // specifically not Serializable & Cloneable 1694 && !klass->is_java_lang_Object() // not the supertype of all T[] 1695 ) { 1696 // Note: When interfaces are reliable, we can narrow the interface 1697 // test to (klass != Serializable && klass != Cloneable). 1698 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1699 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1700 // The key property of this type is that it folds up tests 1701 // for array-ness, since it proves that the layout_helper is positive. 1702 // Thus, a generic value like the basic object layout helper works fine. 1703 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1704 } 1705 } 1706 1707 // If we are loading from a freshly-allocated object, produce a zero, 1708 // if the load is provably beyond the header of the object. 1709 // (Also allow a variable load from a fresh array to produce zero.) 1710 const TypeOopPtr *tinst = tp->isa_oopptr(); 1711 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1712 if (ReduceFieldZeroing || is_instance) { 1713 Node* value = can_see_stored_value(mem,phase); 1714 if (value != NULL && value->is_Con()) 1715 return value->bottom_type(); 1716 } 1717 1718 if (is_instance) { 1719 // If we have an instance type and our memory input is the 1720 // programs's initial memory state, there is no matching store, 1721 // so just return a zero of the appropriate type 1722 Node *mem = in(MemNode::Memory); 1723 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1724 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1725 return Type::get_zero_type(_type->basic_type()); 1726 } 1727 } 1728 return _type; 1729 } 1730 1731 //------------------------------match_edge------------------------------------- 1732 // Do we Match on this edge index or not? Match only the address. 1733 uint LoadNode::match_edge(uint idx) const { 1734 return idx == MemNode::Address; 1735 } 1736 1737 //--------------------------LoadBNode::Ideal-------------------------------------- 1738 // 1739 // If the previous store is to the same address as this load, 1740 // and the value stored was larger than a byte, replace this load 1741 // with the value stored truncated to a byte. If no truncation is 1742 // needed, the replacement is done in LoadNode::Identity(). 1743 // 1744 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1745 Node* mem = in(MemNode::Memory); 1746 Node* value = can_see_stored_value(mem,phase); 1747 if( value && !phase->type(value)->higher_equal( _type ) ) { 1748 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1749 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1750 } 1751 // Identity call will handle the case where truncation is not needed. 1752 return LoadNode::Ideal(phase, can_reshape); 1753 } 1754 1755 //--------------------------LoadUBNode::Ideal------------------------------------- 1756 // 1757 // If the previous store is to the same address as this load, 1758 // and the value stored was larger than a byte, replace this load 1759 // with the value stored truncated to a byte. If no truncation is 1760 // needed, the replacement is done in LoadNode::Identity(). 1761 // 1762 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1763 Node* mem = in(MemNode::Memory); 1764 Node* value = can_see_stored_value(mem, phase); 1765 if (value && !phase->type(value)->higher_equal(_type)) 1766 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); 1767 // Identity call will handle the case where truncation is not needed. 1768 return LoadNode::Ideal(phase, can_reshape); 1769 } 1770 1771 //--------------------------LoadUSNode::Ideal------------------------------------- 1772 // 1773 // If the previous store is to the same address as this load, 1774 // and the value stored was larger than a char, replace this load 1775 // with the value stored truncated to a char. If no truncation is 1776 // needed, the replacement is done in LoadNode::Identity(). 1777 // 1778 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1779 Node* mem = in(MemNode::Memory); 1780 Node* value = can_see_stored_value(mem,phase); 1781 if( value && !phase->type(value)->higher_equal( _type ) ) 1782 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1783 // Identity call will handle the case where truncation is not needed. 1784 return LoadNode::Ideal(phase, can_reshape); 1785 } 1786 1787 //--------------------------LoadSNode::Ideal-------------------------------------- 1788 // 1789 // If the previous store is to the same address as this load, 1790 // and the value stored was larger than a short, replace this load 1791 // with the value stored truncated to a short. If no truncation is 1792 // needed, the replacement is done in LoadNode::Identity(). 1793 // 1794 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1795 Node* mem = in(MemNode::Memory); 1796 Node* value = can_see_stored_value(mem,phase); 1797 if( value && !phase->type(value)->higher_equal( _type ) ) { 1798 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1799 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1800 } 1801 // Identity call will handle the case where truncation is not needed. 1802 return LoadNode::Ideal(phase, can_reshape); 1803 } 1804 1805 //============================================================================= 1806 //----------------------------LoadKlassNode::make------------------------------ 1807 // Polymorphic factory method: 1808 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1809 Compile* C = gvn.C; 1810 Node *ctl = NULL; 1811 // sanity check the alias category against the created node type 1812 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1813 assert(adr_type != NULL, "expecting TypeOopPtr"); 1814 #ifdef _LP64 1815 if (adr_type->is_ptr_to_narrowoop()) { 1816 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1817 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1818 } 1819 #endif 1820 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1821 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1822 } 1823 1824 //------------------------------Value------------------------------------------ 1825 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1826 return klass_value_common(phase); 1827 } 1828 1829 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1830 // Either input is TOP ==> the result is TOP 1831 const Type *t1 = phase->type( in(MemNode::Memory) ); 1832 if (t1 == Type::TOP) return Type::TOP; 1833 Node *adr = in(MemNode::Address); 1834 const Type *t2 = phase->type( adr ); 1835 if (t2 == Type::TOP) return Type::TOP; 1836 const TypePtr *tp = t2->is_ptr(); 1837 if (TypePtr::above_centerline(tp->ptr()) || 1838 tp->ptr() == TypePtr::Null) return Type::TOP; 1839 1840 // Return a more precise klass, if possible 1841 const TypeInstPtr *tinst = tp->isa_instptr(); 1842 if (tinst != NULL) { 1843 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1844 int offset = tinst->offset(); 1845 if (ik == phase->C->env()->Class_klass() 1846 && (offset == java_lang_Class::klass_offset_in_bytes() || 1847 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1848 // We are loading a special hidden field from a Class mirror object, 1849 // the field which points to the VM's Klass metaobject. 1850 ciType* t = tinst->java_mirror_type(); 1851 // java_mirror_type returns non-null for compile-time Class constants. 1852 if (t != NULL) { 1853 // constant oop => constant klass 1854 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1855 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1856 } 1857 if (!t->is_klass()) { 1858 // a primitive Class (e.g., int.class) has NULL for a klass field 1859 return TypePtr::NULL_PTR; 1860 } 1861 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1862 return TypeKlassPtr::make(t->as_klass()); 1863 } 1864 // non-constant mirror, so we can't tell what's going on 1865 } 1866 if( !ik->is_loaded() ) 1867 return _type; // Bail out if not loaded 1868 if (offset == oopDesc::klass_offset_in_bytes()) { 1869 if (tinst->klass_is_exact()) { 1870 return TypeKlassPtr::make(ik); 1871 } 1872 // See if we can become precise: no subklasses and no interface 1873 // (Note: We need to support verified interfaces.) 1874 if (!ik->is_interface() && !ik->has_subklass()) { 1875 //assert(!UseExactTypes, "this code should be useless with exact types"); 1876 // Add a dependence; if any subclass added we need to recompile 1877 if (!ik->is_final()) { 1878 // %%% should use stronger assert_unique_concrete_subtype instead 1879 phase->C->dependencies()->assert_leaf_type(ik); 1880 } 1881 // Return precise klass 1882 return TypeKlassPtr::make(ik); 1883 } 1884 1885 // Return root of possible klass 1886 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1887 } 1888 } 1889 1890 // Check for loading klass from an array 1891 const TypeAryPtr *tary = tp->isa_aryptr(); 1892 if( tary != NULL ) { 1893 ciKlass *tary_klass = tary->klass(); 1894 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1895 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1896 if (tary->klass_is_exact()) { 1897 return TypeKlassPtr::make(tary_klass); 1898 } 1899 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1900 // If the klass is an object array, we defer the question to the 1901 // array component klass. 1902 if( ak->is_obj_array_klass() ) { 1903 assert( ak->is_loaded(), "" ); 1904 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1905 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1906 ciInstanceKlass* ik = base_k->as_instance_klass(); 1907 // See if we can become precise: no subklasses and no interface 1908 if (!ik->is_interface() && !ik->has_subklass()) { 1909 //assert(!UseExactTypes, "this code should be useless with exact types"); 1910 // Add a dependence; if any subclass added we need to recompile 1911 if (!ik->is_final()) { 1912 phase->C->dependencies()->assert_leaf_type(ik); 1913 } 1914 // Return precise array klass 1915 return TypeKlassPtr::make(ak); 1916 } 1917 } 1918 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1919 } else { // Found a type-array? 1920 //assert(!UseExactTypes, "this code should be useless with exact types"); 1921 assert( ak->is_type_array_klass(), "" ); 1922 return TypeKlassPtr::make(ak); // These are always precise 1923 } 1924 } 1925 } 1926 1927 // Check for loading klass from an array klass 1928 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1929 if (tkls != NULL && !StressReflectiveCode) { 1930 ciKlass* klass = tkls->klass(); 1931 if( !klass->is_loaded() ) 1932 return _type; // Bail out if not loaded 1933 if( klass->is_obj_array_klass() && 1934 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1935 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1936 // // Always returning precise element type is incorrect, 1937 // // e.g., element type could be object and array may contain strings 1938 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1939 1940 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1941 // according to the element type's subclassing. 1942 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1943 } 1944 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1945 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1946 ciKlass* sup = klass->as_instance_klass()->super(); 1947 // The field is Klass::_super. Return its (constant) value. 1948 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1949 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1950 } 1951 } 1952 1953 // Bailout case 1954 return LoadNode::Value(phase); 1955 } 1956 1957 //------------------------------Identity--------------------------------------- 1958 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1959 // Also feed through the klass in Allocate(...klass...)._klass. 1960 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1961 return klass_identity_common(phase); 1962 } 1963 1964 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 1965 Node* x = LoadNode::Identity(phase); 1966 if (x != this) return x; 1967 1968 // Take apart the address into an oop and and offset. 1969 // Return 'this' if we cannot. 1970 Node* adr = in(MemNode::Address); 1971 intptr_t offset = 0; 1972 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1973 if (base == NULL) return this; 1974 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1975 if (toop == NULL) return this; 1976 1977 // We can fetch the klass directly through an AllocateNode. 1978 // This works even if the klass is not constant (clone or newArray). 1979 if (offset == oopDesc::klass_offset_in_bytes()) { 1980 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1981 if (allocated_klass != NULL) { 1982 return allocated_klass; 1983 } 1984 } 1985 1986 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1987 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1988 // See inline_native_Class_query for occurrences of these patterns. 1989 // Java Example: x.getClass().isAssignableFrom(y) 1990 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1991 // 1992 // This improves reflective code, often making the Class 1993 // mirror go completely dead. (Current exception: Class 1994 // mirrors may appear in debug info, but we could clean them out by 1995 // introducing a new debug info operator for klassOop.java_mirror). 1996 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1997 && (offset == java_lang_Class::klass_offset_in_bytes() || 1998 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1999 // We are loading a special hidden field from a Class mirror, 2000 // the field which points to its Klass or arrayKlass metaobject. 2001 if (base->is_Load()) { 2002 Node* adr2 = base->in(MemNode::Address); 2003 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2004 if (tkls != NULL && !tkls->empty() 2005 && (tkls->klass()->is_instance_klass() || 2006 tkls->klass()->is_array_klass()) 2007 && adr2->is_AddP() 2008 ) { 2009 int mirror_field = Klass::java_mirror_offset_in_bytes(); 2010 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2011 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 2012 } 2013 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 2014 return adr2->in(AddPNode::Base); 2015 } 2016 } 2017 } 2018 } 2019 2020 return this; 2021 } 2022 2023 2024 //------------------------------Value------------------------------------------ 2025 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2026 const Type *t = klass_value_common(phase); 2027 if (t == Type::TOP) 2028 return t; 2029 2030 return t->make_narrowoop(); 2031 } 2032 2033 //------------------------------Identity--------------------------------------- 2034 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2035 // Also feed through the klass in Allocate(...klass...)._klass. 2036 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2037 Node *x = klass_identity_common(phase); 2038 2039 const Type *t = phase->type( x ); 2040 if( t == Type::TOP ) return x; 2041 if( t->isa_narrowoop()) return x; 2042 2043 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 2044 } 2045 2046 //------------------------------Value----------------------------------------- 2047 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2048 // Either input is TOP ==> the result is TOP 2049 const Type *t1 = phase->type( in(MemNode::Memory) ); 2050 if( t1 == Type::TOP ) return Type::TOP; 2051 Node *adr = in(MemNode::Address); 2052 const Type *t2 = phase->type( adr ); 2053 if( t2 == Type::TOP ) return Type::TOP; 2054 const TypePtr *tp = t2->is_ptr(); 2055 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2056 const TypeAryPtr *tap = tp->isa_aryptr(); 2057 if( !tap ) return _type; 2058 return tap->size(); 2059 } 2060 2061 //-------------------------------Ideal--------------------------------------- 2062 // Feed through the length in AllocateArray(...length...)._length. 2063 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2064 Node* p = MemNode::Ideal_common(phase, can_reshape); 2065 if (p) return (p == NodeSentinel) ? NULL : p; 2066 2067 // Take apart the address into an oop and and offset. 2068 // Return 'this' if we cannot. 2069 Node* adr = in(MemNode::Address); 2070 intptr_t offset = 0; 2071 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2072 if (base == NULL) return NULL; 2073 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2074 if (tary == NULL) return NULL; 2075 2076 // We can fetch the length directly through an AllocateArrayNode. 2077 // This works even if the length is not constant (clone or newArray). 2078 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2079 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2080 if (alloc != NULL) { 2081 Node* allocated_length = alloc->Ideal_length(); 2082 Node* len = alloc->make_ideal_length(tary, phase); 2083 if (allocated_length != len) { 2084 // New CastII improves on this. 2085 return len; 2086 } 2087 } 2088 } 2089 2090 return NULL; 2091 } 2092 2093 //------------------------------Identity--------------------------------------- 2094 // Feed through the length in AllocateArray(...length...)._length. 2095 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2096 Node* x = LoadINode::Identity(phase); 2097 if (x != this) return x; 2098 2099 // Take apart the address into an oop and and offset. 2100 // Return 'this' if we cannot. 2101 Node* adr = in(MemNode::Address); 2102 intptr_t offset = 0; 2103 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2104 if (base == NULL) return this; 2105 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2106 if (tary == NULL) return this; 2107 2108 // We can fetch the length directly through an AllocateArrayNode. 2109 // This works even if the length is not constant (clone or newArray). 2110 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2111 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2112 if (alloc != NULL) { 2113 Node* allocated_length = alloc->Ideal_length(); 2114 // Do not allow make_ideal_length to allocate a CastII node. 2115 Node* len = alloc->make_ideal_length(tary, phase, false); 2116 if (allocated_length == len) { 2117 // Return allocated_length only if it would not be improved by a CastII. 2118 return allocated_length; 2119 } 2120 } 2121 } 2122 2123 return this; 2124 2125 } 2126 2127 //============================================================================= 2128 //---------------------------StoreNode::make----------------------------------- 2129 // Polymorphic factory method: 2130 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 2131 Compile* C = gvn.C; 2132 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2133 ctl != NULL, "raw memory operations should have control edge"); 2134 2135 switch (bt) { 2136 case T_BOOLEAN: 2137 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 2138 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 2139 case T_CHAR: 2140 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 2141 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 2142 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 2143 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 2144 case T_ADDRESS: 2145 case T_OBJECT: 2146 #ifdef _LP64 2147 if (adr->bottom_type()->is_ptr_to_narrowoop() || 2148 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 2149 adr->bottom_type()->isa_rawptr())) { 2150 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2151 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 2152 } else 2153 #endif 2154 { 2155 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 2156 } 2157 } 2158 ShouldNotReachHere(); 2159 return (StoreNode*)NULL; 2160 } 2161 2162 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2163 bool require_atomic = true; 2164 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2165 } 2166 2167 2168 //--------------------------bottom_type---------------------------------------- 2169 const Type *StoreNode::bottom_type() const { 2170 return Type::MEMORY; 2171 } 2172 2173 //------------------------------hash------------------------------------------- 2174 uint StoreNode::hash() const { 2175 // unroll addition of interesting fields 2176 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2177 2178 // Since they are not commoned, do not hash them: 2179 return NO_HASH; 2180 } 2181 2182 //------------------------------Ideal------------------------------------------ 2183 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2184 // When a store immediately follows a relevant allocation/initialization, 2185 // try to capture it into the initialization, or hoist it above. 2186 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2187 Node* p = MemNode::Ideal_common(phase, can_reshape); 2188 if (p) return (p == NodeSentinel) ? NULL : p; 2189 2190 Node* mem = in(MemNode::Memory); 2191 Node* address = in(MemNode::Address); 2192 2193 // Back-to-back stores to same address? Fold em up. Generally 2194 // unsafe if I have intervening uses... Also disallowed for StoreCM 2195 // since they must follow each StoreP operation. Redundant StoreCMs 2196 // are eliminated just before matching in final_graph_reshape. 2197 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address) && 2198 mem->Opcode() != Op_StoreCM) { 2199 // Looking at a dead closed cycle of memory? 2200 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2201 2202 assert(Opcode() == mem->Opcode() || 2203 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2204 "no mismatched stores, except on raw memory"); 2205 2206 if (mem->outcnt() == 1 && // check for intervening uses 2207 mem->as_Store()->memory_size() <= this->memory_size()) { 2208 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2209 // For example, 'mem' might be the final state at a conditional return. 2210 // Or, 'mem' might be used by some node which is live at the same time 2211 // 'this' is live, which might be unschedulable. So, require exactly 2212 // ONE user, the 'this' store, until such time as we clone 'mem' for 2213 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2214 if (can_reshape) { // (%%% is this an anachronism?) 2215 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2216 phase->is_IterGVN()); 2217 } else { 2218 // It's OK to do this in the parser, since DU info is always accurate, 2219 // and the parser always refers to nodes via SafePointNode maps. 2220 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2221 } 2222 return this; 2223 } 2224 } 2225 2226 // Capture an unaliased, unconditional, simple store into an initializer. 2227 // Or, if it is independent of the allocation, hoist it above the allocation. 2228 if (ReduceFieldZeroing && /*can_reshape &&*/ 2229 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2230 InitializeNode* init = mem->in(0)->as_Initialize(); 2231 intptr_t offset = init->can_capture_store(this, phase); 2232 if (offset > 0) { 2233 Node* moved = init->capture_store(this, offset, phase); 2234 // If the InitializeNode captured me, it made a raw copy of me, 2235 // and I need to disappear. 2236 if (moved != NULL) { 2237 // %%% hack to ensure that Ideal returns a new node: 2238 mem = MergeMemNode::make(phase->C, mem); 2239 return mem; // fold me away 2240 } 2241 } 2242 } 2243 2244 return NULL; // No further progress 2245 } 2246 2247 //------------------------------Value----------------------------------------- 2248 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2249 // Either input is TOP ==> the result is TOP 2250 const Type *t1 = phase->type( in(MemNode::Memory) ); 2251 if( t1 == Type::TOP ) return Type::TOP; 2252 const Type *t2 = phase->type( in(MemNode::Address) ); 2253 if( t2 == Type::TOP ) return Type::TOP; 2254 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2255 if( t3 == Type::TOP ) return Type::TOP; 2256 return Type::MEMORY; 2257 } 2258 2259 //------------------------------Identity--------------------------------------- 2260 // Remove redundant stores: 2261 // Store(m, p, Load(m, p)) changes to m. 2262 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2263 Node *StoreNode::Identity( PhaseTransform *phase ) { 2264 Node* mem = in(MemNode::Memory); 2265 Node* adr = in(MemNode::Address); 2266 Node* val = in(MemNode::ValueIn); 2267 2268 // Load then Store? Then the Store is useless 2269 if (val->is_Load() && 2270 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 2271 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 2272 val->as_Load()->store_Opcode() == Opcode()) { 2273 return mem; 2274 } 2275 2276 // Two stores in a row of the same value? 2277 if (mem->is_Store() && 2278 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 2279 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 2280 mem->Opcode() == Opcode()) { 2281 return mem; 2282 } 2283 2284 // Store of zero anywhere into a freshly-allocated object? 2285 // Then the store is useless. 2286 // (It must already have been captured by the InitializeNode.) 2287 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2288 // a newly allocated object is already all-zeroes everywhere 2289 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2290 return mem; 2291 } 2292 2293 // the store may also apply to zero-bits in an earlier object 2294 Node* prev_mem = find_previous_store(phase); 2295 // Steps (a), (b): Walk past independent stores to find an exact match. 2296 if (prev_mem != NULL) { 2297 Node* prev_val = can_see_stored_value(prev_mem, phase); 2298 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2299 // prev_val and val might differ by a cast; it would be good 2300 // to keep the more informative of the two. 2301 return mem; 2302 } 2303 } 2304 } 2305 2306 return this; 2307 } 2308 2309 //------------------------------match_edge------------------------------------- 2310 // Do we Match on this edge index or not? Match only memory & value 2311 uint StoreNode::match_edge(uint idx) const { 2312 return idx == MemNode::Address || idx == MemNode::ValueIn; 2313 } 2314 2315 //------------------------------cmp-------------------------------------------- 2316 // Do not common stores up together. They generally have to be split 2317 // back up anyways, so do not bother. 2318 uint StoreNode::cmp( const Node &n ) const { 2319 return (&n == this); // Always fail except on self 2320 } 2321 2322 //------------------------------Ideal_masked_input----------------------------- 2323 // Check for a useless mask before a partial-word store 2324 // (StoreB ... (AndI valIn conIa) ) 2325 // If (conIa & mask == mask) this simplifies to 2326 // (StoreB ... (valIn) ) 2327 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2328 Node *val = in(MemNode::ValueIn); 2329 if( val->Opcode() == Op_AndI ) { 2330 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2331 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2332 set_req(MemNode::ValueIn, val->in(1)); 2333 return this; 2334 } 2335 } 2336 return NULL; 2337 } 2338 2339 2340 //------------------------------Ideal_sign_extended_input---------------------- 2341 // Check for useless sign-extension before a partial-word store 2342 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2343 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2344 // (StoreB ... (valIn) ) 2345 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2346 Node *val = in(MemNode::ValueIn); 2347 if( val->Opcode() == Op_RShiftI ) { 2348 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2349 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2350 Node *shl = val->in(1); 2351 if( shl->Opcode() == Op_LShiftI ) { 2352 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2353 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2354 set_req(MemNode::ValueIn, shl->in(1)); 2355 return this; 2356 } 2357 } 2358 } 2359 } 2360 return NULL; 2361 } 2362 2363 //------------------------------value_never_loaded----------------------------------- 2364 // Determine whether there are any possible loads of the value stored. 2365 // For simplicity, we actually check if there are any loads from the 2366 // address stored to, not just for loads of the value stored by this node. 2367 // 2368 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2369 Node *adr = in(Address); 2370 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2371 if (adr_oop == NULL) 2372 return false; 2373 if (!adr_oop->is_known_instance_field()) 2374 return false; // if not a distinct instance, there may be aliases of the address 2375 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2376 Node *use = adr->fast_out(i); 2377 int opc = use->Opcode(); 2378 if (use->is_Load() || use->is_LoadStore()) { 2379 return false; 2380 } 2381 } 2382 return true; 2383 } 2384 2385 //============================================================================= 2386 //------------------------------Ideal------------------------------------------ 2387 // If the store is from an AND mask that leaves the low bits untouched, then 2388 // we can skip the AND operation. If the store is from a sign-extension 2389 // (a left shift, then right shift) we can skip both. 2390 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2391 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2392 if( progress != NULL ) return progress; 2393 2394 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2395 if( progress != NULL ) return progress; 2396 2397 // Finally check the default case 2398 return StoreNode::Ideal(phase, can_reshape); 2399 } 2400 2401 //============================================================================= 2402 //------------------------------Ideal------------------------------------------ 2403 // If the store is from an AND mask that leaves the low bits untouched, then 2404 // we can skip the AND operation 2405 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2406 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2407 if( progress != NULL ) return progress; 2408 2409 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2410 if( progress != NULL ) return progress; 2411 2412 // Finally check the default case 2413 return StoreNode::Ideal(phase, can_reshape); 2414 } 2415 2416 //============================================================================= 2417 //------------------------------Identity--------------------------------------- 2418 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2419 // No need to card mark when storing a null ptr 2420 Node* my_store = in(MemNode::OopStore); 2421 if (my_store->is_Store()) { 2422 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2423 if( t1 == TypePtr::NULL_PTR ) { 2424 return in(MemNode::Memory); 2425 } 2426 } 2427 return this; 2428 } 2429 2430 //============================================================================= 2431 //------------------------------Ideal--------------------------------------- 2432 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2433 Node* progress = StoreNode::Ideal(phase, can_reshape); 2434 if (progress != NULL) return progress; 2435 2436 Node* my_store = in(MemNode::OopStore); 2437 if (my_store->is_MergeMem()) { 2438 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2439 set_req(MemNode::OopStore, mem); 2440 return this; 2441 } 2442 2443 return NULL; 2444 } 2445 2446 //------------------------------Value----------------------------------------- 2447 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2448 // Either input is TOP ==> the result is TOP 2449 const Type *t = phase->type( in(MemNode::Memory) ); 2450 if( t == Type::TOP ) return Type::TOP; 2451 t = phase->type( in(MemNode::Address) ); 2452 if( t == Type::TOP ) return Type::TOP; 2453 t = phase->type( in(MemNode::ValueIn) ); 2454 if( t == Type::TOP ) return Type::TOP; 2455 // If extra input is TOP ==> the result is TOP 2456 t = phase->type( in(MemNode::OopStore) ); 2457 if( t == Type::TOP ) return Type::TOP; 2458 2459 return StoreNode::Value( phase ); 2460 } 2461 2462 2463 //============================================================================= 2464 //----------------------------------SCMemProjNode------------------------------ 2465 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2466 { 2467 return bottom_type(); 2468 } 2469 2470 //============================================================================= 2471 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2472 init_req(MemNode::Control, c ); 2473 init_req(MemNode::Memory , mem); 2474 init_req(MemNode::Address, adr); 2475 init_req(MemNode::ValueIn, val); 2476 init_req( ExpectedIn, ex ); 2477 init_class_id(Class_LoadStore); 2478 2479 } 2480 2481 //============================================================================= 2482 //-------------------------------adr_type-------------------------------------- 2483 // Do we Match on this edge index or not? Do not match memory 2484 const TypePtr* ClearArrayNode::adr_type() const { 2485 Node *adr = in(3); 2486 return MemNode::calculate_adr_type(adr->bottom_type()); 2487 } 2488 2489 //------------------------------match_edge------------------------------------- 2490 // Do we Match on this edge index or not? Do not match memory 2491 uint ClearArrayNode::match_edge(uint idx) const { 2492 return idx > 1; 2493 } 2494 2495 //------------------------------Identity--------------------------------------- 2496 // Clearing a zero length array does nothing 2497 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2498 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2499 } 2500 2501 //------------------------------Idealize--------------------------------------- 2502 // Clearing a short array is faster with stores 2503 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2504 const int unit = BytesPerLong; 2505 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2506 if (!t) return NULL; 2507 if (!t->is_con()) return NULL; 2508 intptr_t raw_count = t->get_con(); 2509 intptr_t size = raw_count; 2510 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2511 // Clearing nothing uses the Identity call. 2512 // Negative clears are possible on dead ClearArrays 2513 // (see jck test stmt114.stmt11402.val). 2514 if (size <= 0 || size % unit != 0) return NULL; 2515 intptr_t count = size / unit; 2516 // Length too long; use fast hardware clear 2517 if (size > Matcher::init_array_short_size) return NULL; 2518 Node *mem = in(1); 2519 if( phase->type(mem)==Type::TOP ) return NULL; 2520 Node *adr = in(3); 2521 const Type* at = phase->type(adr); 2522 if( at==Type::TOP ) return NULL; 2523 const TypePtr* atp = at->isa_ptr(); 2524 // adjust atp to be the correct array element address type 2525 if (atp == NULL) atp = TypePtr::BOTTOM; 2526 else atp = atp->add_offset(Type::OffsetBot); 2527 // Get base for derived pointer purposes 2528 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2529 Node *base = adr->in(1); 2530 2531 Node *zero = phase->makecon(TypeLong::ZERO); 2532 Node *off = phase->MakeConX(BytesPerLong); 2533 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2534 count--; 2535 while( count-- ) { 2536 mem = phase->transform(mem); 2537 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2538 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2539 } 2540 return mem; 2541 } 2542 2543 //----------------------------step_through---------------------------------- 2544 // Return allocation input memory edge if it is different instance 2545 // or itself if it is the one we are looking for. 2546 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2547 Node* n = *np; 2548 assert(n->is_ClearArray(), "sanity"); 2549 intptr_t offset; 2550 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2551 // This method is called only before Allocate nodes are expanded during 2552 // macro nodes expansion. Before that ClearArray nodes are only generated 2553 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2554 assert(alloc != NULL, "should have allocation"); 2555 if (alloc->_idx == instance_id) { 2556 // Can not bypass initialization of the instance we are looking for. 2557 return false; 2558 } 2559 // Otherwise skip it. 2560 InitializeNode* init = alloc->initialization(); 2561 if (init != NULL) 2562 *np = init->in(TypeFunc::Memory); 2563 else 2564 *np = alloc->in(TypeFunc::Memory); 2565 return true; 2566 } 2567 2568 //----------------------------clear_memory------------------------------------- 2569 // Generate code to initialize object storage to zero. 2570 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2571 intptr_t start_offset, 2572 Node* end_offset, 2573 PhaseGVN* phase) { 2574 Compile* C = phase->C; 2575 intptr_t offset = start_offset; 2576 2577 int unit = BytesPerLong; 2578 if ((offset % unit) != 0) { 2579 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2580 adr = phase->transform(adr); 2581 const TypePtr* atp = TypeRawPtr::BOTTOM; 2582 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2583 mem = phase->transform(mem); 2584 offset += BytesPerInt; 2585 } 2586 assert((offset % unit) == 0, ""); 2587 2588 // Initialize the remaining stuff, if any, with a ClearArray. 2589 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2590 } 2591 2592 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2593 Node* start_offset, 2594 Node* end_offset, 2595 PhaseGVN* phase) { 2596 if (start_offset == end_offset) { 2597 // nothing to do 2598 return mem; 2599 } 2600 2601 Compile* C = phase->C; 2602 int unit = BytesPerLong; 2603 Node* zbase = start_offset; 2604 Node* zend = end_offset; 2605 2606 // Scale to the unit required by the CPU: 2607 if (!Matcher::init_array_count_is_in_bytes) { 2608 Node* shift = phase->intcon(exact_log2(unit)); 2609 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2610 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2611 } 2612 2613 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2614 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2615 2616 // Bulk clear double-words 2617 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2618 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2619 return phase->transform(mem); 2620 } 2621 2622 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2623 intptr_t start_offset, 2624 intptr_t end_offset, 2625 PhaseGVN* phase) { 2626 if (start_offset == end_offset) { 2627 // nothing to do 2628 return mem; 2629 } 2630 2631 Compile* C = phase->C; 2632 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2633 intptr_t done_offset = end_offset; 2634 if ((done_offset % BytesPerLong) != 0) { 2635 done_offset -= BytesPerInt; 2636 } 2637 if (done_offset > start_offset) { 2638 mem = clear_memory(ctl, mem, dest, 2639 start_offset, phase->MakeConX(done_offset), phase); 2640 } 2641 if (done_offset < end_offset) { // emit the final 32-bit store 2642 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2643 adr = phase->transform(adr); 2644 const TypePtr* atp = TypeRawPtr::BOTTOM; 2645 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2646 mem = phase->transform(mem); 2647 done_offset += BytesPerInt; 2648 } 2649 assert(done_offset == end_offset, ""); 2650 return mem; 2651 } 2652 2653 //============================================================================= 2654 // Do not match memory edge. 2655 uint StrIntrinsicNode::match_edge(uint idx) const { 2656 return idx == 2 || idx == 3; 2657 } 2658 2659 //------------------------------Ideal------------------------------------------ 2660 // Return a node which is more "ideal" than the current node. Strip out 2661 // control copies 2662 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2663 if (remove_dead_region(phase, can_reshape)) return this; 2664 2665 if (can_reshape) { 2666 Node* mem = phase->transform(in(MemNode::Memory)); 2667 // If transformed to a MergeMem, get the desired slice 2668 uint alias_idx = phase->C->get_alias_index(adr_type()); 2669 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; 2670 if (mem != in(MemNode::Memory)) { 2671 set_req(MemNode::Memory, mem); 2672 return this; 2673 } 2674 } 2675 return NULL; 2676 } 2677 2678 //============================================================================= 2679 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2680 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2681 _adr_type(C->get_adr_type(alias_idx)) 2682 { 2683 init_class_id(Class_MemBar); 2684 Node* top = C->top(); 2685 init_req(TypeFunc::I_O,top); 2686 init_req(TypeFunc::FramePtr,top); 2687 init_req(TypeFunc::ReturnAdr,top); 2688 if (precedent != NULL) 2689 init_req(TypeFunc::Parms, precedent); 2690 } 2691 2692 //------------------------------cmp-------------------------------------------- 2693 uint MemBarNode::hash() const { return NO_HASH; } 2694 uint MemBarNode::cmp( const Node &n ) const { 2695 return (&n == this); // Always fail except on self 2696 } 2697 2698 //------------------------------make------------------------------------------- 2699 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2700 int len = Precedent + (pn == NULL? 0: 1); 2701 switch (opcode) { 2702 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2703 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2704 case Op_MemBarAcquireLock: return new(C, len) MemBarAcquireLockNode(C, atp, pn); 2705 case Op_MemBarReleaseLock: return new(C, len) MemBarReleaseLockNode(C, atp, pn); 2706 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2707 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2708 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2709 default: ShouldNotReachHere(); return NULL; 2710 } 2711 } 2712 2713 //------------------------------Ideal------------------------------------------ 2714 // Return a node which is more "ideal" than the current node. Strip out 2715 // control copies 2716 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2717 if (remove_dead_region(phase, can_reshape)) return this; 2718 2719 // Eliminate volatile MemBars for scalar replaced objects. 2720 if (can_reshape && req() == (Precedent+1) && 2721 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { 2722 // Volatile field loads and stores. 2723 Node* my_mem = in(MemBarNode::Precedent); 2724 if (my_mem != NULL && my_mem->is_Mem()) { 2725 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2726 // Check for scalar replaced object reference. 2727 if( t_oop != NULL && t_oop->is_known_instance_field() && 2728 t_oop->offset() != Type::OffsetBot && 2729 t_oop->offset() != Type::OffsetTop) { 2730 // Replace MemBar projections by its inputs. 2731 PhaseIterGVN* igvn = phase->is_IterGVN(); 2732 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2733 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2734 // Must return either the original node (now dead) or a new node 2735 // (Do not return a top here, since that would break the uniqueness of top.) 2736 return new (phase->C, 1) ConINode(TypeInt::ZERO); 2737 } 2738 } 2739 } 2740 return NULL; 2741 } 2742 2743 //------------------------------Value------------------------------------------ 2744 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2745 if( !in(0) ) return Type::TOP; 2746 if( phase->type(in(0)) == Type::TOP ) 2747 return Type::TOP; 2748 return TypeTuple::MEMBAR; 2749 } 2750 2751 //------------------------------match------------------------------------------ 2752 // Construct projections for memory. 2753 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2754 switch (proj->_con) { 2755 case TypeFunc::Control: 2756 case TypeFunc::Memory: 2757 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2758 } 2759 ShouldNotReachHere(); 2760 return NULL; 2761 } 2762 2763 //===========================InitializeNode==================================== 2764 // SUMMARY: 2765 // This node acts as a memory barrier on raw memory, after some raw stores. 2766 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2767 // The Initialize can 'capture' suitably constrained stores as raw inits. 2768 // It can coalesce related raw stores into larger units (called 'tiles'). 2769 // It can avoid zeroing new storage for memory units which have raw inits. 2770 // At macro-expansion, it is marked 'complete', and does not optimize further. 2771 // 2772 // EXAMPLE: 2773 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2774 // ctl = incoming control; mem* = incoming memory 2775 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2776 // First allocate uninitialized memory and fill in the header: 2777 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2778 // ctl := alloc.Control; mem* := alloc.Memory* 2779 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2780 // Then initialize to zero the non-header parts of the raw memory block: 2781 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2782 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2783 // After the initialize node executes, the object is ready for service: 2784 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2785 // Suppose its body is immediately initialized as {1,2}: 2786 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2787 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2788 // mem.SLICE(#short[*]) := store2 2789 // 2790 // DETAILS: 2791 // An InitializeNode collects and isolates object initialization after 2792 // an AllocateNode and before the next possible safepoint. As a 2793 // memory barrier (MemBarNode), it keeps critical stores from drifting 2794 // down past any safepoint or any publication of the allocation. 2795 // Before this barrier, a newly-allocated object may have uninitialized bits. 2796 // After this barrier, it may be treated as a real oop, and GC is allowed. 2797 // 2798 // The semantics of the InitializeNode include an implicit zeroing of 2799 // the new object from object header to the end of the object. 2800 // (The object header and end are determined by the AllocateNode.) 2801 // 2802 // Certain stores may be added as direct inputs to the InitializeNode. 2803 // These stores must update raw memory, and they must be to addresses 2804 // derived from the raw address produced by AllocateNode, and with 2805 // a constant offset. They must be ordered by increasing offset. 2806 // The first one is at in(RawStores), the last at in(req()-1). 2807 // Unlike most memory operations, they are not linked in a chain, 2808 // but are displayed in parallel as users of the rawmem output of 2809 // the allocation. 2810 // 2811 // (See comments in InitializeNode::capture_store, which continue 2812 // the example given above.) 2813 // 2814 // When the associated Allocate is macro-expanded, the InitializeNode 2815 // may be rewritten to optimize collected stores. A ClearArrayNode 2816 // may also be created at that point to represent any required zeroing. 2817 // The InitializeNode is then marked 'complete', prohibiting further 2818 // capturing of nearby memory operations. 2819 // 2820 // During macro-expansion, all captured initializations which store 2821 // constant values of 32 bits or smaller are coalesced (if advantageous) 2822 // into larger 'tiles' 32 or 64 bits. This allows an object to be 2823 // initialized in fewer memory operations. Memory words which are 2824 // covered by neither tiles nor non-constant stores are pre-zeroed 2825 // by explicit stores of zero. (The code shape happens to do all 2826 // zeroing first, then all other stores, with both sequences occurring 2827 // in order of ascending offsets.) 2828 // 2829 // Alternatively, code may be inserted between an AllocateNode and its 2830 // InitializeNode, to perform arbitrary initialization of the new object. 2831 // E.g., the object copying intrinsics insert complex data transfers here. 2832 // The initialization must then be marked as 'complete' disable the 2833 // built-in zeroing semantics and the collection of initializing stores. 2834 // 2835 // While an InitializeNode is incomplete, reads from the memory state 2836 // produced by it are optimizable if they match the control edge and 2837 // new oop address associated with the allocation/initialization. 2838 // They return a stored value (if the offset matches) or else zero. 2839 // A write to the memory state, if it matches control and address, 2840 // and if it is to a constant offset, may be 'captured' by the 2841 // InitializeNode. It is cloned as a raw memory operation and rewired 2842 // inside the initialization, to the raw oop produced by the allocation. 2843 // Operations on addresses which are provably distinct (e.g., to 2844 // other AllocateNodes) are allowed to bypass the initialization. 2845 // 2846 // The effect of all this is to consolidate object initialization 2847 // (both arrays and non-arrays, both piecewise and bulk) into a 2848 // single location, where it can be optimized as a unit. 2849 // 2850 // Only stores with an offset less than TrackedInitializationLimit words 2851 // will be considered for capture by an InitializeNode. This puts a 2852 // reasonable limit on the complexity of optimized initializations. 2853 2854 //---------------------------InitializeNode------------------------------------ 2855 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2856 : _is_complete(Incomplete), 2857 MemBarNode(C, adr_type, rawoop) 2858 { 2859 init_class_id(Class_Initialize); 2860 2861 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2862 assert(in(RawAddress) == rawoop, "proper init"); 2863 // Note: allocation() can be NULL, for secondary initialization barriers 2864 } 2865 2866 // Since this node is not matched, it will be processed by the 2867 // register allocator. Declare that there are no constraints 2868 // on the allocation of the RawAddress edge. 2869 const RegMask &InitializeNode::in_RegMask(uint idx) const { 2870 // This edge should be set to top, by the set_complete. But be conservative. 2871 if (idx == InitializeNode::RawAddress) 2872 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2873 return RegMask::Empty; 2874 } 2875 2876 Node* InitializeNode::memory(uint alias_idx) { 2877 Node* mem = in(Memory); 2878 if (mem->is_MergeMem()) { 2879 return mem->as_MergeMem()->memory_at(alias_idx); 2880 } else { 2881 // incoming raw memory is not split 2882 return mem; 2883 } 2884 } 2885 2886 bool InitializeNode::is_non_zero() { 2887 if (is_complete()) return false; 2888 remove_extra_zeroes(); 2889 return (req() > RawStores); 2890 } 2891 2892 void InitializeNode::set_complete(PhaseGVN* phase) { 2893 assert(!is_complete(), "caller responsibility"); 2894 _is_complete = Complete; 2895 2896 // After this node is complete, it contains a bunch of 2897 // raw-memory initializations. There is no need for 2898 // it to have anything to do with non-raw memory effects. 2899 // Therefore, tell all non-raw users to re-optimize themselves, 2900 // after skipping the memory effects of this initialization. 2901 PhaseIterGVN* igvn = phase->is_IterGVN(); 2902 if (igvn) igvn->add_users_to_worklist(this); 2903 } 2904 2905 // convenience function 2906 // return false if the init contains any stores already 2907 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2908 InitializeNode* init = initialization(); 2909 if (init == NULL || init->is_complete()) return false; 2910 init->remove_extra_zeroes(); 2911 // for now, if this allocation has already collected any inits, bail: 2912 if (init->is_non_zero()) return false; 2913 init->set_complete(phase); 2914 return true; 2915 } 2916 2917 void InitializeNode::remove_extra_zeroes() { 2918 if (req() == RawStores) return; 2919 Node* zmem = zero_memory(); 2920 uint fill = RawStores; 2921 for (uint i = fill; i < req(); i++) { 2922 Node* n = in(i); 2923 if (n->is_top() || n == zmem) continue; // skip 2924 if (fill < i) set_req(fill, n); // compact 2925 ++fill; 2926 } 2927 // delete any empty spaces created: 2928 while (fill < req()) { 2929 del_req(fill); 2930 } 2931 } 2932 2933 // Helper for remembering which stores go with which offsets. 2934 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2935 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2936 intptr_t offset = -1; 2937 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2938 phase, offset); 2939 if (base == NULL) return -1; // something is dead, 2940 if (offset < 0) return -1; // dead, dead 2941 return offset; 2942 } 2943 2944 // Helper for proving that an initialization expression is 2945 // "simple enough" to be folded into an object initialization. 2946 // Attempts to prove that a store's initial value 'n' can be captured 2947 // within the initialization without creating a vicious cycle, such as: 2948 // { Foo p = new Foo(); p.next = p; } 2949 // True for constants and parameters and small combinations thereof. 2950 bool InitializeNode::detect_init_independence(Node* n, 2951 bool st_is_pinned, 2952 int& count) { 2953 if (n == NULL) return true; // (can this really happen?) 2954 if (n->is_Proj()) n = n->in(0); 2955 if (n == this) return false; // found a cycle 2956 if (n->is_Con()) return true; 2957 if (n->is_Start()) return true; // params, etc., are OK 2958 if (n->is_Root()) return true; // even better 2959 2960 Node* ctl = n->in(0); 2961 if (ctl != NULL && !ctl->is_top()) { 2962 if (ctl->is_Proj()) ctl = ctl->in(0); 2963 if (ctl == this) return false; 2964 2965 // If we already know that the enclosing memory op is pinned right after 2966 // the init, then any control flow that the store has picked up 2967 // must have preceded the init, or else be equal to the init. 2968 // Even after loop optimizations (which might change control edges) 2969 // a store is never pinned *before* the availability of its inputs. 2970 if (!MemNode::all_controls_dominate(n, this)) 2971 return false; // failed to prove a good control 2972 2973 } 2974 2975 // Check data edges for possible dependencies on 'this'. 2976 if ((count += 1) > 20) return false; // complexity limit 2977 for (uint i = 1; i < n->req(); i++) { 2978 Node* m = n->in(i); 2979 if (m == NULL || m == n || m->is_top()) continue; 2980 uint first_i = n->find_edge(m); 2981 if (i != first_i) continue; // process duplicate edge just once 2982 if (!detect_init_independence(m, st_is_pinned, count)) { 2983 return false; 2984 } 2985 } 2986 2987 return true; 2988 } 2989 2990 // Here are all the checks a Store must pass before it can be moved into 2991 // an initialization. Returns zero if a check fails. 2992 // On success, returns the (constant) offset to which the store applies, 2993 // within the initialized memory. 2994 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2995 const int FAIL = 0; 2996 if (st->req() != MemNode::ValueIn + 1) 2997 return FAIL; // an inscrutable StoreNode (card mark?) 2998 Node* ctl = st->in(MemNode::Control); 2999 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3000 return FAIL; // must be unconditional after the initialization 3001 Node* mem = st->in(MemNode::Memory); 3002 if (!(mem->is_Proj() && mem->in(0) == this)) 3003 return FAIL; // must not be preceded by other stores 3004 Node* adr = st->in(MemNode::Address); 3005 intptr_t offset; 3006 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3007 if (alloc == NULL) 3008 return FAIL; // inscrutable address 3009 if (alloc != allocation()) 3010 return FAIL; // wrong allocation! (store needs to float up) 3011 Node* val = st->in(MemNode::ValueIn); 3012 int complexity_count = 0; 3013 if (!detect_init_independence(val, true, complexity_count)) 3014 return FAIL; // stored value must be 'simple enough' 3015 3016 return offset; // success 3017 } 3018 3019 // Find the captured store in(i) which corresponds to the range 3020 // [start..start+size) in the initialized object. 3021 // If there is one, return its index i. If there isn't, return the 3022 // negative of the index where it should be inserted. 3023 // Return 0 if the queried range overlaps an initialization boundary 3024 // or if dead code is encountered. 3025 // If size_in_bytes is zero, do not bother with overlap checks. 3026 int InitializeNode::captured_store_insertion_point(intptr_t start, 3027 int size_in_bytes, 3028 PhaseTransform* phase) { 3029 const int FAIL = 0, MAX_STORE = BytesPerLong; 3030 3031 if (is_complete()) 3032 return FAIL; // arraycopy got here first; punt 3033 3034 assert(allocation() != NULL, "must be present"); 3035 3036 // no negatives, no header fields: 3037 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3038 3039 // after a certain size, we bail out on tracking all the stores: 3040 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3041 if (start >= ti_limit) return FAIL; 3042 3043 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3044 if (i >= limit) return -(int)i; // not found; here is where to put it 3045 3046 Node* st = in(i); 3047 intptr_t st_off = get_store_offset(st, phase); 3048 if (st_off < 0) { 3049 if (st != zero_memory()) { 3050 return FAIL; // bail out if there is dead garbage 3051 } 3052 } else if (st_off > start) { 3053 // ...we are done, since stores are ordered 3054 if (st_off < start + size_in_bytes) { 3055 return FAIL; // the next store overlaps 3056 } 3057 return -(int)i; // not found; here is where to put it 3058 } else if (st_off < start) { 3059 if (size_in_bytes != 0 && 3060 start < st_off + MAX_STORE && 3061 start < st_off + st->as_Store()->memory_size()) { 3062 return FAIL; // the previous store overlaps 3063 } 3064 } else { 3065 if (size_in_bytes != 0 && 3066 st->as_Store()->memory_size() != size_in_bytes) { 3067 return FAIL; // mismatched store size 3068 } 3069 return i; 3070 } 3071 3072 ++i; 3073 } 3074 } 3075 3076 // Look for a captured store which initializes at the offset 'start' 3077 // with the given size. If there is no such store, and no other 3078 // initialization interferes, then return zero_memory (the memory 3079 // projection of the AllocateNode). 3080 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3081 PhaseTransform* phase) { 3082 assert(stores_are_sane(phase), ""); 3083 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3084 if (i == 0) { 3085 return NULL; // something is dead 3086 } else if (i < 0) { 3087 return zero_memory(); // just primordial zero bits here 3088 } else { 3089 Node* st = in(i); // here is the store at this position 3090 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3091 return st; 3092 } 3093 } 3094 3095 // Create, as a raw pointer, an address within my new object at 'offset'. 3096 Node* InitializeNode::make_raw_address(intptr_t offset, 3097 PhaseTransform* phase) { 3098 Node* addr = in(RawAddress); 3099 if (offset != 0) { 3100 Compile* C = phase->C; 3101 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 3102 phase->MakeConX(offset)) ); 3103 } 3104 return addr; 3105 } 3106 3107 // Clone the given store, converting it into a raw store 3108 // initializing a field or element of my new object. 3109 // Caller is responsible for retiring the original store, 3110 // with subsume_node or the like. 3111 // 3112 // From the example above InitializeNode::InitializeNode, 3113 // here are the old stores to be captured: 3114 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3115 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3116 // 3117 // Here is the changed code; note the extra edges on init: 3118 // alloc = (Allocate ...) 3119 // rawoop = alloc.RawAddress 3120 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3121 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3122 // init = (Initialize alloc.Control alloc.Memory rawoop 3123 // rawstore1 rawstore2) 3124 // 3125 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3126 PhaseTransform* phase) { 3127 assert(stores_are_sane(phase), ""); 3128 3129 if (start < 0) return NULL; 3130 assert(can_capture_store(st, phase) == start, "sanity"); 3131 3132 Compile* C = phase->C; 3133 int size_in_bytes = st->memory_size(); 3134 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3135 if (i == 0) return NULL; // bail out 3136 Node* prev_mem = NULL; // raw memory for the captured store 3137 if (i > 0) { 3138 prev_mem = in(i); // there is a pre-existing store under this one 3139 set_req(i, C->top()); // temporarily disconnect it 3140 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3141 } else { 3142 i = -i; // no pre-existing store 3143 prev_mem = zero_memory(); // a slice of the newly allocated object 3144 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3145 set_req(--i, C->top()); // reuse this edge; it has been folded away 3146 else 3147 ins_req(i, C->top()); // build a new edge 3148 } 3149 Node* new_st = st->clone(); 3150 new_st->set_req(MemNode::Control, in(Control)); 3151 new_st->set_req(MemNode::Memory, prev_mem); 3152 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3153 new_st = phase->transform(new_st); 3154 3155 // At this point, new_st might have swallowed a pre-existing store 3156 // at the same offset, or perhaps new_st might have disappeared, 3157 // if it redundantly stored the same value (or zero to fresh memory). 3158 3159 // In any case, wire it in: 3160 set_req(i, new_st); 3161 3162 // The caller may now kill the old guy. 3163 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3164 assert(check_st == new_st || check_st == NULL, "must be findable"); 3165 assert(!is_complete(), ""); 3166 return new_st; 3167 } 3168 3169 static bool store_constant(jlong* tiles, int num_tiles, 3170 intptr_t st_off, int st_size, 3171 jlong con) { 3172 if ((st_off & (st_size-1)) != 0) 3173 return false; // strange store offset (assume size==2**N) 3174 address addr = (address)tiles + st_off; 3175 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3176 switch (st_size) { 3177 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3178 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3179 case sizeof(jint): *(jint*) addr = (jint) con; break; 3180 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3181 default: return false; // strange store size (detect size!=2**N here) 3182 } 3183 return true; // return success to caller 3184 } 3185 3186 // Coalesce subword constants into int constants and possibly 3187 // into long constants. The goal, if the CPU permits, 3188 // is to initialize the object with a small number of 64-bit tiles. 3189 // Also, convert floating-point constants to bit patterns. 3190 // Non-constants are not relevant to this pass. 3191 // 3192 // In terms of the running example on InitializeNode::InitializeNode 3193 // and InitializeNode::capture_store, here is the transformation 3194 // of rawstore1 and rawstore2 into rawstore12: 3195 // alloc = (Allocate ...) 3196 // rawoop = alloc.RawAddress 3197 // tile12 = 0x00010002 3198 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3199 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3200 // 3201 void 3202 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3203 Node* size_in_bytes, 3204 PhaseGVN* phase) { 3205 Compile* C = phase->C; 3206 3207 assert(stores_are_sane(phase), ""); 3208 // Note: After this pass, they are not completely sane, 3209 // since there may be some overlaps. 3210 3211 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3212 3213 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3214 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3215 size_limit = MIN2(size_limit, ti_limit); 3216 size_limit = align_size_up(size_limit, BytesPerLong); 3217 int num_tiles = size_limit / BytesPerLong; 3218 3219 // allocate space for the tile map: 3220 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3221 jlong tiles_buf[small_len]; 3222 Node* nodes_buf[small_len]; 3223 jlong inits_buf[small_len]; 3224 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3225 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3226 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3227 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3228 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3229 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3230 // tiles: exact bitwise model of all primitive constants 3231 // nodes: last constant-storing node subsumed into the tiles model 3232 // inits: which bytes (in each tile) are touched by any initializations 3233 3234 //// Pass A: Fill in the tile model with any relevant stores. 3235 3236 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3237 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3238 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3239 Node* zmem = zero_memory(); // initially zero memory state 3240 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3241 Node* st = in(i); 3242 intptr_t st_off = get_store_offset(st, phase); 3243 3244 // Figure out the store's offset and constant value: 3245 if (st_off < header_size) continue; //skip (ignore header) 3246 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3247 int st_size = st->as_Store()->memory_size(); 3248 if (st_off + st_size > size_limit) break; 3249 3250 // Record which bytes are touched, whether by constant or not. 3251 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3252 continue; // skip (strange store size) 3253 3254 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3255 if (!val->singleton()) continue; //skip (non-con store) 3256 BasicType type = val->basic_type(); 3257 3258 jlong con = 0; 3259 switch (type) { 3260 case T_INT: con = val->is_int()->get_con(); break; 3261 case T_LONG: con = val->is_long()->get_con(); break; 3262 case T_FLOAT: con = jint_cast(val->getf()); break; 3263 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3264 default: continue; //skip (odd store type) 3265 } 3266 3267 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3268 st->Opcode() == Op_StoreL) { 3269 continue; // This StoreL is already optimal. 3270 } 3271 3272 // Store down the constant. 3273 store_constant(tiles, num_tiles, st_off, st_size, con); 3274 3275 intptr_t j = st_off >> LogBytesPerLong; 3276 3277 if (type == T_INT && st_size == BytesPerInt 3278 && (st_off & BytesPerInt) == BytesPerInt) { 3279 jlong lcon = tiles[j]; 3280 if (!Matcher::isSimpleConstant64(lcon) && 3281 st->Opcode() == Op_StoreI) { 3282 // This StoreI is already optimal by itself. 3283 jint* intcon = (jint*) &tiles[j]; 3284 intcon[1] = 0; // undo the store_constant() 3285 3286 // If the previous store is also optimal by itself, back up and 3287 // undo the action of the previous loop iteration... if we can. 3288 // But if we can't, just let the previous half take care of itself. 3289 st = nodes[j]; 3290 st_off -= BytesPerInt; 3291 con = intcon[0]; 3292 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3293 assert(st_off >= header_size, "still ignoring header"); 3294 assert(get_store_offset(st, phase) == st_off, "must be"); 3295 assert(in(i-1) == zmem, "must be"); 3296 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3297 assert(con == tcon->is_int()->get_con(), "must be"); 3298 // Undo the effects of the previous loop trip, which swallowed st: 3299 intcon[0] = 0; // undo store_constant() 3300 set_req(i-1, st); // undo set_req(i, zmem) 3301 nodes[j] = NULL; // undo nodes[j] = st 3302 --old_subword; // undo ++old_subword 3303 } 3304 continue; // This StoreI is already optimal. 3305 } 3306 } 3307 3308 // This store is not needed. 3309 set_req(i, zmem); 3310 nodes[j] = st; // record for the moment 3311 if (st_size < BytesPerLong) // something has changed 3312 ++old_subword; // includes int/float, but who's counting... 3313 else ++old_long; 3314 } 3315 3316 if ((old_subword + old_long) == 0) 3317 return; // nothing more to do 3318 3319 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3320 // Be sure to insert them before overlapping non-constant stores. 3321 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3322 for (int j = 0; j < num_tiles; j++) { 3323 jlong con = tiles[j]; 3324 jlong init = inits[j]; 3325 if (con == 0) continue; 3326 jint con0, con1; // split the constant, address-wise 3327 jint init0, init1; // split the init map, address-wise 3328 { union { jlong con; jint intcon[2]; } u; 3329 u.con = con; 3330 con0 = u.intcon[0]; 3331 con1 = u.intcon[1]; 3332 u.con = init; 3333 init0 = u.intcon[0]; 3334 init1 = u.intcon[1]; 3335 } 3336 3337 Node* old = nodes[j]; 3338 assert(old != NULL, "need the prior store"); 3339 intptr_t offset = (j * BytesPerLong); 3340 3341 bool split = !Matcher::isSimpleConstant64(con); 3342 3343 if (offset < header_size) { 3344 assert(offset + BytesPerInt >= header_size, "second int counts"); 3345 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3346 split = true; // only the second word counts 3347 // Example: int a[] = { 42 ... } 3348 } else if (con0 == 0 && init0 == -1) { 3349 split = true; // first word is covered by full inits 3350 // Example: int a[] = { ... foo(), 42 ... } 3351 } else if (con1 == 0 && init1 == -1) { 3352 split = true; // second word is covered by full inits 3353 // Example: int a[] = { ... 42, foo() ... } 3354 } 3355 3356 // Here's a case where init0 is neither 0 nor -1: 3357 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3358 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3359 // In this case the tile is not split; it is (jlong)42. 3360 // The big tile is stored down, and then the foo() value is inserted. 3361 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3362 3363 Node* ctl = old->in(MemNode::Control); 3364 Node* adr = make_raw_address(offset, phase); 3365 const TypePtr* atp = TypeRawPtr::BOTTOM; 3366 3367 // One or two coalesced stores to plop down. 3368 Node* st[2]; 3369 intptr_t off[2]; 3370 int nst = 0; 3371 if (!split) { 3372 ++new_long; 3373 off[nst] = offset; 3374 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3375 phase->longcon(con), T_LONG); 3376 } else { 3377 // Omit either if it is a zero. 3378 if (con0 != 0) { 3379 ++new_int; 3380 off[nst] = offset; 3381 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3382 phase->intcon(con0), T_INT); 3383 } 3384 if (con1 != 0) { 3385 ++new_int; 3386 offset += BytesPerInt; 3387 adr = make_raw_address(offset, phase); 3388 off[nst] = offset; 3389 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3390 phase->intcon(con1), T_INT); 3391 } 3392 } 3393 3394 // Insert second store first, then the first before the second. 3395 // Insert each one just before any overlapping non-constant stores. 3396 while (nst > 0) { 3397 Node* st1 = st[--nst]; 3398 C->copy_node_notes_to(st1, old); 3399 st1 = phase->transform(st1); 3400 offset = off[nst]; 3401 assert(offset >= header_size, "do not smash header"); 3402 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3403 guarantee(ins_idx != 0, "must re-insert constant store"); 3404 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3405 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3406 set_req(--ins_idx, st1); 3407 else 3408 ins_req(ins_idx, st1); 3409 } 3410 } 3411 3412 if (PrintCompilation && WizardMode) 3413 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3414 old_subword, old_long, new_int, new_long); 3415 if (C->log() != NULL) 3416 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3417 old_subword, old_long, new_int, new_long); 3418 3419 // Clean up any remaining occurrences of zmem: 3420 remove_extra_zeroes(); 3421 } 3422 3423 // Explore forward from in(start) to find the first fully initialized 3424 // word, and return its offset. Skip groups of subword stores which 3425 // together initialize full words. If in(start) is itself part of a 3426 // fully initialized word, return the offset of in(start). If there 3427 // are no following full-word stores, or if something is fishy, return 3428 // a negative value. 3429 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3430 int int_map = 0; 3431 intptr_t int_map_off = 0; 3432 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3433 3434 for (uint i = start, limit = req(); i < limit; i++) { 3435 Node* st = in(i); 3436 3437 intptr_t st_off = get_store_offset(st, phase); 3438 if (st_off < 0) break; // return conservative answer 3439 3440 int st_size = st->as_Store()->memory_size(); 3441 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3442 return st_off; // we found a complete word init 3443 } 3444 3445 // update the map: 3446 3447 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3448 if (this_int_off != int_map_off) { 3449 // reset the map: 3450 int_map = 0; 3451 int_map_off = this_int_off; 3452 } 3453 3454 int subword_off = st_off - this_int_off; 3455 int_map |= right_n_bits(st_size) << subword_off; 3456 if ((int_map & FULL_MAP) == FULL_MAP) { 3457 return this_int_off; // we found a complete word init 3458 } 3459 3460 // Did this store hit or cross the word boundary? 3461 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3462 if (next_int_off == this_int_off + BytesPerInt) { 3463 // We passed the current int, without fully initializing it. 3464 int_map_off = next_int_off; 3465 int_map >>= BytesPerInt; 3466 } else if (next_int_off > this_int_off + BytesPerInt) { 3467 // We passed the current and next int. 3468 return this_int_off + BytesPerInt; 3469 } 3470 } 3471 3472 return -1; 3473 } 3474 3475 3476 // Called when the associated AllocateNode is expanded into CFG. 3477 // At this point, we may perform additional optimizations. 3478 // Linearize the stores by ascending offset, to make memory 3479 // activity as coherent as possible. 3480 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3481 intptr_t header_size, 3482 Node* size_in_bytes, 3483 PhaseGVN* phase) { 3484 assert(!is_complete(), "not already complete"); 3485 assert(stores_are_sane(phase), ""); 3486 assert(allocation() != NULL, "must be present"); 3487 3488 remove_extra_zeroes(); 3489 3490 if (ReduceFieldZeroing || ReduceBulkZeroing) 3491 // reduce instruction count for common initialization patterns 3492 coalesce_subword_stores(header_size, size_in_bytes, phase); 3493 3494 Node* zmem = zero_memory(); // initially zero memory state 3495 Node* inits = zmem; // accumulating a linearized chain of inits 3496 #ifdef ASSERT 3497 intptr_t first_offset = allocation()->minimum_header_size(); 3498 intptr_t last_init_off = first_offset; // previous init offset 3499 intptr_t last_init_end = first_offset; // previous init offset+size 3500 intptr_t last_tile_end = first_offset; // previous tile offset+size 3501 #endif 3502 intptr_t zeroes_done = header_size; 3503 3504 bool do_zeroing = true; // we might give up if inits are very sparse 3505 int big_init_gaps = 0; // how many large gaps have we seen? 3506 3507 if (ZeroTLAB) do_zeroing = false; 3508 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3509 3510 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3511 Node* st = in(i); 3512 intptr_t st_off = get_store_offset(st, phase); 3513 if (st_off < 0) 3514 break; // unknown junk in the inits 3515 if (st->in(MemNode::Memory) != zmem) 3516 break; // complicated store chains somehow in list 3517 3518 int st_size = st->as_Store()->memory_size(); 3519 intptr_t next_init_off = st_off + st_size; 3520 3521 if (do_zeroing && zeroes_done < next_init_off) { 3522 // See if this store needs a zero before it or under it. 3523 intptr_t zeroes_needed = st_off; 3524 3525 if (st_size < BytesPerInt) { 3526 // Look for subword stores which only partially initialize words. 3527 // If we find some, we must lay down some word-level zeroes first, 3528 // underneath the subword stores. 3529 // 3530 // Examples: 3531 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3532 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3533 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3534 // 3535 // Note: coalesce_subword_stores may have already done this, 3536 // if it was prompted by constant non-zero subword initializers. 3537 // But this case can still arise with non-constant stores. 3538 3539 intptr_t next_full_store = find_next_fullword_store(i, phase); 3540 3541 // In the examples above: 3542 // in(i) p q r s x y z 3543 // st_off 12 13 14 15 12 13 14 3544 // st_size 1 1 1 1 1 1 1 3545 // next_full_s. 12 16 16 16 16 16 16 3546 // z's_done 12 16 16 16 12 16 12 3547 // z's_needed 12 16 16 16 16 16 16 3548 // zsize 0 0 0 0 4 0 4 3549 if (next_full_store < 0) { 3550 // Conservative tack: Zero to end of current word. 3551 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3552 } else { 3553 // Zero to beginning of next fully initialized word. 3554 // Or, don't zero at all, if we are already in that word. 3555 assert(next_full_store >= zeroes_needed, "must go forward"); 3556 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3557 zeroes_needed = next_full_store; 3558 } 3559 } 3560 3561 if (zeroes_needed > zeroes_done) { 3562 intptr_t zsize = zeroes_needed - zeroes_done; 3563 // Do some incremental zeroing on rawmem, in parallel with inits. 3564 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3565 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3566 zeroes_done, zeroes_needed, 3567 phase); 3568 zeroes_done = zeroes_needed; 3569 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3570 do_zeroing = false; // leave the hole, next time 3571 } 3572 } 3573 3574 // Collect the store and move on: 3575 st->set_req(MemNode::Memory, inits); 3576 inits = st; // put it on the linearized chain 3577 set_req(i, zmem); // unhook from previous position 3578 3579 if (zeroes_done == st_off) 3580 zeroes_done = next_init_off; 3581 3582 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3583 3584 #ifdef ASSERT 3585 // Various order invariants. Weaker than stores_are_sane because 3586 // a large constant tile can be filled in by smaller non-constant stores. 3587 assert(st_off >= last_init_off, "inits do not reverse"); 3588 last_init_off = st_off; 3589 const Type* val = NULL; 3590 if (st_size >= BytesPerInt && 3591 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3592 (int)val->basic_type() < (int)T_OBJECT) { 3593 assert(st_off >= last_tile_end, "tiles do not overlap"); 3594 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3595 last_tile_end = MAX2(last_tile_end, next_init_off); 3596 } else { 3597 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3598 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3599 assert(st_off >= last_init_end, "inits do not overlap"); 3600 last_init_end = next_init_off; // it's a non-tile 3601 } 3602 #endif //ASSERT 3603 } 3604 3605 remove_extra_zeroes(); // clear out all the zmems left over 3606 add_req(inits); 3607 3608 if (!ZeroTLAB) { 3609 // If anything remains to be zeroed, zero it all now. 3610 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3611 // if it is the last unused 4 bytes of an instance, forget about it 3612 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3613 if (zeroes_done + BytesPerLong >= size_limit) { 3614 assert(allocation() != NULL, ""); 3615 if (allocation()->Opcode() == Op_Allocate) { 3616 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3617 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3618 if (zeroes_done == k->layout_helper()) 3619 zeroes_done = size_limit; 3620 } 3621 } 3622 if (zeroes_done < size_limit) { 3623 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3624 zeroes_done, size_in_bytes, phase); 3625 } 3626 } 3627 3628 set_complete(phase); 3629 return rawmem; 3630 } 3631 3632 3633 #ifdef ASSERT 3634 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3635 if (is_complete()) 3636 return true; // stores could be anything at this point 3637 assert(allocation() != NULL, "must be present"); 3638 intptr_t last_off = allocation()->minimum_header_size(); 3639 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3640 Node* st = in(i); 3641 intptr_t st_off = get_store_offset(st, phase); 3642 if (st_off < 0) continue; // ignore dead garbage 3643 if (last_off > st_off) { 3644 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3645 this->dump(2); 3646 assert(false, "ascending store offsets"); 3647 return false; 3648 } 3649 last_off = st_off + st->as_Store()->memory_size(); 3650 } 3651 return true; 3652 } 3653 #endif //ASSERT 3654 3655 3656 3657 3658 //============================MergeMemNode===================================== 3659 // 3660 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3661 // contributing store or call operations. Each contributor provides the memory 3662 // state for a particular "alias type" (see Compile::alias_type). For example, 3663 // if a MergeMem has an input X for alias category #6, then any memory reference 3664 // to alias category #6 may use X as its memory state input, as an exact equivalent 3665 // to using the MergeMem as a whole. 3666 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3667 // 3668 // (Here, the <N> notation gives the index of the relevant adr_type.) 3669 // 3670 // In one special case (and more cases in the future), alias categories overlap. 3671 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3672 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3673 // it is exactly equivalent to that state W: 3674 // MergeMem(<Bot>: W) <==> W 3675 // 3676 // Usually, the merge has more than one input. In that case, where inputs 3677 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3678 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3679 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3680 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3681 // 3682 // A merge can take a "wide" memory state as one of its narrow inputs. 3683 // This simply means that the merge observes out only the relevant parts of 3684 // the wide input. That is, wide memory states arriving at narrow merge inputs 3685 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3686 // 3687 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3688 // and that memory slices "leak through": 3689 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3690 // 3691 // But, in such a cascade, repeated memory slices can "block the leak": 3692 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3693 // 3694 // In the last example, Y is not part of the combined memory state of the 3695 // outermost MergeMem. The system must, of course, prevent unschedulable 3696 // memory states from arising, so you can be sure that the state Y is somehow 3697 // a precursor to state Y'. 3698 // 3699 // 3700 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3701 // of each MergeMemNode array are exactly the numerical alias indexes, including 3702 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3703 // Compile::alias_type (and kin) produce and manage these indexes. 3704 // 3705 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3706 // (Note that this provides quick access to the top node inside MergeMem methods, 3707 // without the need to reach out via TLS to Compile::current.) 3708 // 3709 // As a consequence of what was just described, a MergeMem that represents a full 3710 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3711 // containing all alias categories. 3712 // 3713 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3714 // 3715 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3716 // a memory state for the alias type <N>, or else the top node, meaning that 3717 // there is no particular input for that alias type. Note that the length of 3718 // a MergeMem is variable, and may be extended at any time to accommodate new 3719 // memory states at larger alias indexes. When merges grow, they are of course 3720 // filled with "top" in the unused in() positions. 3721 // 3722 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3723 // (Top was chosen because it works smoothly with passes like GCM.) 3724 // 3725 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3726 // the type of random VM bits like TLS references.) Since it is always the 3727 // first non-Bot memory slice, some low-level loops use it to initialize an 3728 // index variable: for (i = AliasIdxRaw; i < req(); i++). 3729 // 3730 // 3731 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3732 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3733 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 3734 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3735 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3736 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3737 // 3738 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3739 // really that different from the other memory inputs. An abbreviation called 3740 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3741 // 3742 // 3743 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3744 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3745 // that "emerges though" the base memory will be marked as excluding the alias types 3746 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 3747 // 3748 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3749 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3750 // 3751 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 3752 // (It is currently unimplemented.) As you can see, the resulting merge is 3753 // actually a disjoint union of memory states, rather than an overlay. 3754 // 3755 3756 //------------------------------MergeMemNode----------------------------------- 3757 Node* MergeMemNode::make_empty_memory() { 3758 Node* empty_memory = (Node*) Compile::current()->top(); 3759 assert(empty_memory->is_top(), "correct sentinel identity"); 3760 return empty_memory; 3761 } 3762 3763 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3764 init_class_id(Class_MergeMem); 3765 // all inputs are nullified in Node::Node(int) 3766 // set_input(0, NULL); // no control input 3767 3768 // Initialize the edges uniformly to top, for starters. 3769 Node* empty_mem = make_empty_memory(); 3770 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3771 init_req(i,empty_mem); 3772 } 3773 assert(empty_memory() == empty_mem, ""); 3774 3775 if( new_base != NULL && new_base->is_MergeMem() ) { 3776 MergeMemNode* mdef = new_base->as_MergeMem(); 3777 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3778 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3779 mms.set_memory(mms.memory2()); 3780 } 3781 assert(base_memory() == mdef->base_memory(), ""); 3782 } else { 3783 set_base_memory(new_base); 3784 } 3785 } 3786 3787 // Make a new, untransformed MergeMem with the same base as 'mem'. 3788 // If mem is itself a MergeMem, populate the result with the same edges. 3789 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3790 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3791 } 3792 3793 //------------------------------cmp-------------------------------------------- 3794 uint MergeMemNode::hash() const { return NO_HASH; } 3795 uint MergeMemNode::cmp( const Node &n ) const { 3796 return (&n == this); // Always fail except on self 3797 } 3798 3799 //------------------------------Identity--------------------------------------- 3800 Node* MergeMemNode::Identity(PhaseTransform *phase) { 3801 // Identity if this merge point does not record any interesting memory 3802 // disambiguations. 3803 Node* base_mem = base_memory(); 3804 Node* empty_mem = empty_memory(); 3805 if (base_mem != empty_mem) { // Memory path is not dead? 3806 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3807 Node* mem = in(i); 3808 if (mem != empty_mem && mem != base_mem) { 3809 return this; // Many memory splits; no change 3810 } 3811 } 3812 } 3813 return base_mem; // No memory splits; ID on the one true input 3814 } 3815 3816 //------------------------------Ideal------------------------------------------ 3817 // This method is invoked recursively on chains of MergeMem nodes 3818 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3819 // Remove chain'd MergeMems 3820 // 3821 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3822 // relative to the "in(Bot)". Since we are patching both at the same time, 3823 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3824 // but rewrite each "in(i)" relative to the new "in(Bot)". 3825 Node *progress = NULL; 3826 3827 3828 Node* old_base = base_memory(); 3829 Node* empty_mem = empty_memory(); 3830 if (old_base == empty_mem) 3831 return NULL; // Dead memory path. 3832 3833 MergeMemNode* old_mbase; 3834 if (old_base != NULL && old_base->is_MergeMem()) 3835 old_mbase = old_base->as_MergeMem(); 3836 else 3837 old_mbase = NULL; 3838 Node* new_base = old_base; 3839 3840 // simplify stacked MergeMems in base memory 3841 if (old_mbase) new_base = old_mbase->base_memory(); 3842 3843 // the base memory might contribute new slices beyond my req() 3844 if (old_mbase) grow_to_match(old_mbase); 3845 3846 // Look carefully at the base node if it is a phi. 3847 PhiNode* phi_base; 3848 if (new_base != NULL && new_base->is_Phi()) 3849 phi_base = new_base->as_Phi(); 3850 else 3851 phi_base = NULL; 3852 3853 Node* phi_reg = NULL; 3854 uint phi_len = (uint)-1; 3855 if (phi_base != NULL && !phi_base->is_copy()) { 3856 // do not examine phi if degraded to a copy 3857 phi_reg = phi_base->region(); 3858 phi_len = phi_base->req(); 3859 // see if the phi is unfinished 3860 for (uint i = 1; i < phi_len; i++) { 3861 if (phi_base->in(i) == NULL) { 3862 // incomplete phi; do not look at it yet! 3863 phi_reg = NULL; 3864 phi_len = (uint)-1; 3865 break; 3866 } 3867 } 3868 } 3869 3870 // Note: We do not call verify_sparse on entry, because inputs 3871 // can normalize to the base_memory via subsume_node or similar 3872 // mechanisms. This method repairs that damage. 3873 3874 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3875 3876 // Look at each slice. 3877 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3878 Node* old_in = in(i); 3879 // calculate the old memory value 3880 Node* old_mem = old_in; 3881 if (old_mem == empty_mem) old_mem = old_base; 3882 assert(old_mem == memory_at(i), ""); 3883 3884 // maybe update (reslice) the old memory value 3885 3886 // simplify stacked MergeMems 3887 Node* new_mem = old_mem; 3888 MergeMemNode* old_mmem; 3889 if (old_mem != NULL && old_mem->is_MergeMem()) 3890 old_mmem = old_mem->as_MergeMem(); 3891 else 3892 old_mmem = NULL; 3893 if (old_mmem == this) { 3894 // This can happen if loops break up and safepoints disappear. 3895 // A merge of BotPtr (default) with a RawPtr memory derived from a 3896 // safepoint can be rewritten to a merge of the same BotPtr with 3897 // the BotPtr phi coming into the loop. If that phi disappears 3898 // also, we can end up with a self-loop of the mergemem. 3899 // In general, if loops degenerate and memory effects disappear, 3900 // a mergemem can be left looking at itself. This simply means 3901 // that the mergemem's default should be used, since there is 3902 // no longer any apparent effect on this slice. 3903 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3904 // from start. Update the input to TOP. 3905 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3906 } 3907 else if (old_mmem != NULL) { 3908 new_mem = old_mmem->memory_at(i); 3909 } 3910 // else preceding memory was not a MergeMem 3911 3912 // replace equivalent phis (unfortunately, they do not GVN together) 3913 if (new_mem != NULL && new_mem != new_base && 3914 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3915 if (new_mem->is_Phi()) { 3916 PhiNode* phi_mem = new_mem->as_Phi(); 3917 for (uint i = 1; i < phi_len; i++) { 3918 if (phi_base->in(i) != phi_mem->in(i)) { 3919 phi_mem = NULL; 3920 break; 3921 } 3922 } 3923 if (phi_mem != NULL) { 3924 // equivalent phi nodes; revert to the def 3925 new_mem = new_base; 3926 } 3927 } 3928 } 3929 3930 // maybe store down a new value 3931 Node* new_in = new_mem; 3932 if (new_in == new_base) new_in = empty_mem; 3933 3934 if (new_in != old_in) { 3935 // Warning: Do not combine this "if" with the previous "if" 3936 // A memory slice might have be be rewritten even if it is semantically 3937 // unchanged, if the base_memory value has changed. 3938 set_req(i, new_in); 3939 progress = this; // Report progress 3940 } 3941 } 3942 3943 if (new_base != old_base) { 3944 set_req(Compile::AliasIdxBot, new_base); 3945 // Don't use set_base_memory(new_base), because we need to update du. 3946 assert(base_memory() == new_base, ""); 3947 progress = this; 3948 } 3949 3950 if( base_memory() == this ) { 3951 // a self cycle indicates this memory path is dead 3952 set_req(Compile::AliasIdxBot, empty_mem); 3953 } 3954 3955 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3956 // Recursion must occur after the self cycle check above 3957 if( base_memory()->is_MergeMem() ) { 3958 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3959 Node *m = phase->transform(new_mbase); // Rollup any cycles 3960 if( m != NULL && (m->is_top() || 3961 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3962 // propagate rollup of dead cycle to self 3963 set_req(Compile::AliasIdxBot, empty_mem); 3964 } 3965 } 3966 3967 if( base_memory() == empty_mem ) { 3968 progress = this; 3969 // Cut inputs during Parse phase only. 3970 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3971 if( !can_reshape ) { 3972 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3973 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3974 } 3975 } 3976 } 3977 3978 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3979 // Check if PhiNode::Ideal's "Split phis through memory merges" 3980 // transform should be attempted. Look for this->phi->this cycle. 3981 uint merge_width = req(); 3982 if (merge_width > Compile::AliasIdxRaw) { 3983 PhiNode* phi = base_memory()->as_Phi(); 3984 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3985 if (phi->in(i) == this) { 3986 phase->is_IterGVN()->_worklist.push(phi); 3987 break; 3988 } 3989 } 3990 } 3991 } 3992 3993 assert(progress || verify_sparse(), "please, no dups of base"); 3994 return progress; 3995 } 3996 3997 //-------------------------set_base_memory------------------------------------- 3998 void MergeMemNode::set_base_memory(Node *new_base) { 3999 Node* empty_mem = empty_memory(); 4000 set_req(Compile::AliasIdxBot, new_base); 4001 assert(memory_at(req()) == new_base, "must set default memory"); 4002 // Clear out other occurrences of new_base: 4003 if (new_base != empty_mem) { 4004 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4005 if (in(i) == new_base) set_req(i, empty_mem); 4006 } 4007 } 4008 } 4009 4010 //------------------------------out_RegMask------------------------------------ 4011 const RegMask &MergeMemNode::out_RegMask() const { 4012 return RegMask::Empty; 4013 } 4014 4015 //------------------------------dump_spec-------------------------------------- 4016 #ifndef PRODUCT 4017 void MergeMemNode::dump_spec(outputStream *st) const { 4018 st->print(" {"); 4019 Node* base_mem = base_memory(); 4020 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4021 Node* mem = memory_at(i); 4022 if (mem == base_mem) { st->print(" -"); continue; } 4023 st->print( " N%d:", mem->_idx ); 4024 Compile::current()->get_adr_type(i)->dump_on(st); 4025 } 4026 st->print(" }"); 4027 } 4028 #endif // !PRODUCT 4029 4030 4031 #ifdef ASSERT 4032 static bool might_be_same(Node* a, Node* b) { 4033 if (a == b) return true; 4034 if (!(a->is_Phi() || b->is_Phi())) return false; 4035 // phis shift around during optimization 4036 return true; // pretty stupid... 4037 } 4038 4039 // verify a narrow slice (either incoming or outgoing) 4040 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4041 if (!VerifyAliases) return; // don't bother to verify unless requested 4042 if (is_error_reported()) return; // muzzle asserts when debugging an error 4043 if (Node::in_dump()) return; // muzzle asserts when printing 4044 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4045 assert(n != NULL, ""); 4046 // Elide intervening MergeMem's 4047 while (n->is_MergeMem()) { 4048 n = n->as_MergeMem()->memory_at(alias_idx); 4049 } 4050 Compile* C = Compile::current(); 4051 const TypePtr* n_adr_type = n->adr_type(); 4052 if (n == m->empty_memory()) { 4053 // Implicit copy of base_memory() 4054 } else if (n_adr_type != TypePtr::BOTTOM) { 4055 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4056 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4057 } else { 4058 // A few places like make_runtime_call "know" that VM calls are narrow, 4059 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4060 bool expected_wide_mem = false; 4061 if (n == m->base_memory()) { 4062 expected_wide_mem = true; 4063 } else if (alias_idx == Compile::AliasIdxRaw || 4064 n == m->memory_at(Compile::AliasIdxRaw)) { 4065 expected_wide_mem = true; 4066 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4067 // memory can "leak through" calls on channels that 4068 // are write-once. Allow this also. 4069 expected_wide_mem = true; 4070 } 4071 assert(expected_wide_mem, "expected narrow slice replacement"); 4072 } 4073 } 4074 #else // !ASSERT 4075 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 4076 #endif 4077 4078 4079 //-----------------------------memory_at--------------------------------------- 4080 Node* MergeMemNode::memory_at(uint alias_idx) const { 4081 assert(alias_idx >= Compile::AliasIdxRaw || 4082 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4083 "must avoid base_memory and AliasIdxTop"); 4084 4085 // Otherwise, it is a narrow slice. 4086 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4087 Compile *C = Compile::current(); 4088 if (is_empty_memory(n)) { 4089 // the array is sparse; empty slots are the "top" node 4090 n = base_memory(); 4091 assert(Node::in_dump() 4092 || n == NULL || n->bottom_type() == Type::TOP 4093 || n->adr_type() == NULL // address is TOP 4094 || n->adr_type() == TypePtr::BOTTOM 4095 || n->adr_type() == TypeRawPtr::BOTTOM 4096 || Compile::current()->AliasLevel() == 0, 4097 "must be a wide memory"); 4098 // AliasLevel == 0 if we are organizing the memory states manually. 4099 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4100 } else { 4101 // make sure the stored slice is sane 4102 #ifdef ASSERT 4103 if (is_error_reported() || Node::in_dump()) { 4104 } else if (might_be_same(n, base_memory())) { 4105 // Give it a pass: It is a mostly harmless repetition of the base. 4106 // This can arise normally from node subsumption during optimization. 4107 } else { 4108 verify_memory_slice(this, alias_idx, n); 4109 } 4110 #endif 4111 } 4112 return n; 4113 } 4114 4115 //---------------------------set_memory_at------------------------------------- 4116 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4117 verify_memory_slice(this, alias_idx, n); 4118 Node* empty_mem = empty_memory(); 4119 if (n == base_memory()) n = empty_mem; // collapse default 4120 uint need_req = alias_idx+1; 4121 if (req() < need_req) { 4122 if (n == empty_mem) return; // already the default, so do not grow me 4123 // grow the sparse array 4124 do { 4125 add_req(empty_mem); 4126 } while (req() < need_req); 4127 } 4128 set_req( alias_idx, n ); 4129 } 4130 4131 4132 4133 //--------------------------iteration_setup------------------------------------ 4134 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4135 if (other != NULL) { 4136 grow_to_match(other); 4137 // invariant: the finite support of mm2 is within mm->req() 4138 #ifdef ASSERT 4139 for (uint i = req(); i < other->req(); i++) { 4140 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4141 } 4142 #endif 4143 } 4144 // Replace spurious copies of base_memory by top. 4145 Node* base_mem = base_memory(); 4146 if (base_mem != NULL && !base_mem->is_top()) { 4147 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4148 if (in(i) == base_mem) 4149 set_req(i, empty_memory()); 4150 } 4151 } 4152 } 4153 4154 //---------------------------grow_to_match------------------------------------- 4155 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4156 Node* empty_mem = empty_memory(); 4157 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4158 // look for the finite support of the other memory 4159 for (uint i = other->req(); --i >= req(); ) { 4160 if (other->in(i) != empty_mem) { 4161 uint new_len = i+1; 4162 while (req() < new_len) add_req(empty_mem); 4163 break; 4164 } 4165 } 4166 } 4167 4168 //---------------------------verify_sparse------------------------------------- 4169 #ifndef PRODUCT 4170 bool MergeMemNode::verify_sparse() const { 4171 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4172 Node* base_mem = base_memory(); 4173 // The following can happen in degenerate cases, since empty==top. 4174 if (is_empty_memory(base_mem)) return true; 4175 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4176 assert(in(i) != NULL, "sane slice"); 4177 if (in(i) == base_mem) return false; // should have been the sentinel value! 4178 } 4179 return true; 4180 } 4181 4182 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4183 Node* n; 4184 n = mm->in(idx); 4185 if (mem == n) return true; // might be empty_memory() 4186 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4187 if (mem == n) return true; 4188 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4189 if (mem == n) return true; 4190 if (n == NULL) break; 4191 } 4192 return false; 4193 } 4194 #endif // !PRODUCT