1 /* 2 * Copyright (c) 1997, 2017, 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 "gc_implementation/shenandoah/shenandoahBrooksPointer.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/compile.hpp" 34 #include "opto/connode.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/memnode.hpp" 39 #include "opto/mulnode.hpp" 40 #include "opto/phaseX.hpp" 41 #include "opto/regmask.hpp" 42 #include "opto/shenandoahSupport.hpp" 43 44 // Portions of code courtesy of Clifford Click 45 46 // Optimization - Graph Style 47 48 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 49 50 //============================================================================= 51 uint MemNode::size_of() const { return sizeof(*this); } 52 53 const TypePtr *MemNode::adr_type() const { 54 Node* adr = in(Address); 55 const TypePtr* cross_check = NULL; 56 DEBUG_ONLY(cross_check = _adr_type); 57 return calculate_adr_type(adr->bottom_type(), cross_check); 58 } 59 60 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 61 if (adr != NULL) { 62 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 63 return true; 64 } 65 } 66 return false; 67 } 68 69 #ifndef PRODUCT 70 void MemNode::dump_spec(outputStream *st) const { 71 if (in(Address) == NULL) return; // node is dead 72 #ifndef ASSERT 73 // fake the missing field 74 const TypePtr* _adr_type = NULL; 75 if (in(Address) != NULL) 76 _adr_type = in(Address)->bottom_type()->isa_ptr(); 77 #endif 78 dump_adr_type(this, _adr_type, st); 79 80 Compile* C = Compile::current(); 81 if (C->alias_type(_adr_type)->is_volatile()) { 82 st->print(" Volatile!"); 83 } 84 if (_unaligned_access) { 85 st->print(" unaligned"); 86 } 87 if (_mismatched_access) { 88 st->print(" mismatched"); 89 } 90 } 91 92 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 93 st->print(" @"); 94 if (adr_type == NULL) { 95 st->print("NULL"); 96 } else { 97 adr_type->dump_on(st); 98 Compile* C = Compile::current(); 99 Compile::AliasType* atp = NULL; 100 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 101 if (atp == NULL) 102 st->print(", idx=?\?;"); 103 else if (atp->index() == Compile::AliasIdxBot) 104 st->print(", idx=Bot;"); 105 else if (atp->index() == Compile::AliasIdxTop) 106 st->print(", idx=Top;"); 107 else if (atp->index() == Compile::AliasIdxRaw) 108 st->print(", idx=Raw;"); 109 else { 110 ciField* field = atp->field(); 111 if (field) { 112 st->print(", name="); 113 field->print_name_on(st); 114 } 115 st->print(", idx=%d;", atp->index()); 116 } 117 } 118 } 119 120 extern void print_alias_types(); 121 122 #endif 123 124 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 125 assert((t_oop != NULL), "sanity"); 126 bool is_instance = t_oop->is_known_instance_field(); 127 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 128 (load != NULL) && load->is_Load() && 129 (phase->is_IterGVN() != NULL); 130 if (!(is_instance || is_boxed_value_load)) 131 return mchain; // don't try to optimize non-instance types 132 uint instance_id = t_oop->instance_id(); 133 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 134 Node *prev = NULL; 135 Node *result = mchain; 136 while (prev != result) { 137 prev = result; 138 if (result == start_mem) 139 break; // hit one of our sentinels 140 // skip over a call which does not affect this memory slice 141 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 142 Node *proj_in = result->in(0); 143 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 144 break; // hit one of our sentinels 145 } else if (proj_in->is_Call()) { 146 CallNode *call = proj_in->as_Call(); 147 if (!call->may_modify(t_oop, phase)) { // returns false for instances 148 result = call->in(TypeFunc::Memory); 149 } 150 } else if (proj_in->is_Initialize()) { 151 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 152 // Stop if this is the initialization for the object instance which 153 // which contains this memory slice, otherwise skip over it. 154 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 155 break; 156 } 157 if (is_instance) { 158 result = proj_in->in(TypeFunc::Memory); 159 } else if (is_boxed_value_load) { 160 Node* klass = alloc->in(AllocateNode::KlassNode); 161 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 162 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 163 result = proj_in->in(TypeFunc::Memory); // not related allocation 164 } 165 } 166 } else if (proj_in->is_MemBar()) { 167 result = proj_in->in(TypeFunc::Memory); 168 } else { 169 assert(false, "unexpected projection"); 170 } 171 } else if (result->is_ClearArray()) { 172 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 173 // Can not bypass initialization of the instance 174 // we are looking for. 175 break; 176 } 177 // Otherwise skip it (the call updated 'result' value). 178 } else if (result->is_MergeMem()) { 179 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 180 } 181 } 182 return result; 183 } 184 185 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 186 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 187 if (t_oop == NULL) 188 return mchain; // don't try to optimize non-oop types 189 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 190 bool is_instance = t_oop->is_known_instance_field(); 191 PhaseIterGVN *igvn = phase->is_IterGVN(); 192 if (is_instance && igvn != NULL && result->is_Phi()) { 193 PhiNode *mphi = result->as_Phi(); 194 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 195 const TypePtr *t = mphi->adr_type(); 196 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 197 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 198 t->is_oopptr()->cast_to_exactness(true) 199 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 200 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 201 // clone the Phi with our address type 202 result = mphi->split_out_instance(t_adr, igvn); 203 } else { 204 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 205 } 206 } 207 return result; 208 } 209 210 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 211 uint alias_idx = phase->C->get_alias_index(tp); 212 Node *mem = mmem; 213 #ifdef ASSERT 214 { 215 // Check that current type is consistent with the alias index used during graph construction 216 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 217 bool consistent = adr_check == NULL || adr_check->empty() || 218 phase->C->must_alias(adr_check, alias_idx ); 219 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 220 if( !consistent && adr_check != NULL && !adr_check->empty() && 221 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 222 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 223 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 224 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 225 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 226 // don't assert if it is dead code. 227 consistent = true; 228 } 229 if( !consistent ) { 230 st->print("alias_idx==%d, adr_check==", alias_idx); 231 if( adr_check == NULL ) { 232 st->print("NULL"); 233 } else { 234 adr_check->dump(); 235 } 236 st->cr(); 237 print_alias_types(); 238 assert(consistent, "adr_check must match alias idx"); 239 } 240 } 241 #endif 242 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 243 // means an array I have not precisely typed yet. Do not do any 244 // alias stuff with it any time soon. 245 const TypeOopPtr *toop = tp->isa_oopptr(); 246 if( tp->base() != Type::AnyPtr && 247 !(toop && 248 toop->klass() != NULL && 249 toop->klass()->is_java_lang_Object() && 250 toop->offset() == Type::OffsetBot) ) { 251 // compress paths and change unreachable cycles to TOP 252 // If not, we can update the input infinitely along a MergeMem cycle 253 // Equivalent code in PhiNode::Ideal 254 Node* m = phase->transform(mmem); 255 // If transformed to a MergeMem, get the desired slice 256 // Otherwise the returned node represents memory for every slice 257 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 258 // Update input if it is progress over what we have now 259 } 260 return mem; 261 } 262 263 //--------------------------Ideal_common--------------------------------------- 264 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 265 // Unhook non-raw memories from complete (macro-expanded) initializations. 266 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 267 // If our control input is a dead region, kill all below the region 268 Node *ctl = in(MemNode::Control); 269 if (ctl && remove_dead_region(phase, can_reshape)) 270 return this; 271 ctl = in(MemNode::Control); 272 // Don't bother trying to transform a dead node 273 if (ctl && ctl->is_top()) return NodeSentinel; 274 275 PhaseIterGVN *igvn = phase->is_IterGVN(); 276 // Wait if control on the worklist. 277 if (ctl && can_reshape && igvn != NULL) { 278 Node* bol = NULL; 279 Node* cmp = NULL; 280 if (ctl->in(0)->is_If()) { 281 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 282 bol = ctl->in(0)->in(1); 283 if (bol->is_Bool()) 284 cmp = ctl->in(0)->in(1)->in(1); 285 } 286 if (igvn->_worklist.member(ctl) || 287 (bol != NULL && igvn->_worklist.member(bol)) || 288 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 289 // This control path may be dead. 290 // Delay this memory node transformation until the control is processed. 291 phase->is_IterGVN()->_worklist.push(this); 292 return NodeSentinel; // caller will return NULL 293 } 294 } 295 // Ignore if memory is dead, or self-loop 296 Node *mem = in(MemNode::Memory); 297 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 298 assert(mem != this, "dead loop in MemNode::Ideal"); 299 300 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 301 // This memory slice may be dead. 302 // Delay this mem node transformation until the memory is processed. 303 phase->is_IterGVN()->_worklist.push(this); 304 return NodeSentinel; // caller will return NULL 305 } 306 307 Node *address = in(MemNode::Address); 308 const Type *t_adr = phase->type(address); 309 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 310 311 if (can_reshape && igvn != NULL && 312 (igvn->_worklist.member(address) || 313 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { 314 // The address's base and type may change when the address is processed. 315 // Delay this mem node transformation until the address is processed. 316 phase->is_IterGVN()->_worklist.push(this); 317 return NodeSentinel; // caller will return NULL 318 } 319 320 // Do NOT remove or optimize the next lines: ensure a new alias index 321 // is allocated for an oop pointer type before Escape Analysis. 322 // Note: C++ will not remove it since the call has side effect. 323 if (t_adr->isa_oopptr()) { 324 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 325 } 326 327 Node* base = NULL; 328 if (address->is_AddP()) { 329 base = address->in(AddPNode::Base); 330 } 331 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 332 !t_adr->isa_rawptr()) { 333 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 334 // Skip this node optimization if its address has TOP base. 335 return NodeSentinel; // caller will return NULL 336 } 337 338 // Avoid independent memory operations 339 Node* old_mem = mem; 340 341 // The code which unhooks non-raw memories from complete (macro-expanded) 342 // initializations was removed. After macro-expansion all stores catched 343 // by Initialize node became raw stores and there is no information 344 // which memory slices they modify. So it is unsafe to move any memory 345 // operation above these stores. Also in most cases hooked non-raw memories 346 // were already unhooked by using information from detect_ptr_independence() 347 // and find_previous_store(). 348 349 if (mem->is_MergeMem()) { 350 MergeMemNode* mmem = mem->as_MergeMem(); 351 const TypePtr *tp = t_adr->is_ptr(); 352 353 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 354 } 355 356 if (mem != old_mem) { 357 set_req(MemNode::Memory, mem); 358 if (can_reshape && old_mem->outcnt() == 0) { 359 igvn->_worklist.push(old_mem); 360 } 361 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 362 return this; 363 } 364 365 // let the subclass continue analyzing... 366 return NULL; 367 } 368 369 // Helper function for proving some simple control dominations. 370 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 371 // Already assumes that 'dom' is available at 'sub', and that 'sub' 372 // is not a constant (dominated by the method's StartNode). 373 // Used by MemNode::find_previous_store to prove that the 374 // control input of a memory operation predates (dominates) 375 // an allocation it wants to look past. 376 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 377 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 378 return false; // Conservative answer for dead code 379 380 // Check 'dom'. Skip Proj and CatchProj nodes. 381 dom = dom->find_exact_control(dom); 382 if (dom == NULL || dom->is_top()) 383 return false; // Conservative answer for dead code 384 385 if (dom == sub) { 386 // For the case when, for example, 'sub' is Initialize and the original 387 // 'dom' is Proj node of the 'sub'. 388 return false; 389 } 390 391 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 392 return true; 393 394 // 'dom' dominates 'sub' if its control edge and control edges 395 // of all its inputs dominate or equal to sub's control edge. 396 397 // Currently 'sub' is either Allocate, Initialize or Start nodes. 398 // Or Region for the check in LoadNode::Ideal(); 399 // 'sub' should have sub->in(0) != NULL. 400 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 401 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 402 403 // Get control edge of 'sub'. 404 Node* orig_sub = sub; 405 sub = sub->find_exact_control(sub->in(0)); 406 if (sub == NULL || sub->is_top()) 407 return false; // Conservative answer for dead code 408 409 assert(sub->is_CFG(), "expecting control"); 410 411 if (sub == dom) 412 return true; 413 414 if (sub->is_Start() || sub->is_Root()) 415 return false; 416 417 { 418 // Check all control edges of 'dom'. 419 420 ResourceMark rm; 421 Arena* arena = Thread::current()->resource_area(); 422 Node_List nlist(arena); 423 Unique_Node_List dom_list(arena); 424 425 dom_list.push(dom); 426 bool only_dominating_controls = false; 427 428 for (uint next = 0; next < dom_list.size(); next++) { 429 Node* n = dom_list.at(next); 430 if (n == orig_sub) 431 return false; // One of dom's inputs dominated by sub. 432 if (!n->is_CFG() && n->pinned()) { 433 // Check only own control edge for pinned non-control nodes. 434 n = n->find_exact_control(n->in(0)); 435 if (n == NULL || n->is_top()) 436 return false; // Conservative answer for dead code 437 assert(n->is_CFG(), "expecting control"); 438 dom_list.push(n); 439 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 440 only_dominating_controls = true; 441 } else if (n->is_CFG()) { 442 if (n->dominates(sub, nlist)) 443 only_dominating_controls = true; 444 else 445 return false; 446 } else { 447 // First, own control edge. 448 Node* m = n->find_exact_control(n->in(0)); 449 if (m != NULL) { 450 if (m->is_top()) 451 return false; // Conservative answer for dead code 452 dom_list.push(m); 453 } 454 // Now, the rest of edges. 455 uint cnt = n->req(); 456 for (uint i = 1; i < cnt; i++) { 457 m = n->find_exact_control(n->in(i)); 458 if (m == NULL || m->is_top()) 459 continue; 460 dom_list.push(m); 461 } 462 } 463 } 464 return only_dominating_controls; 465 } 466 } 467 468 //---------------------detect_ptr_independence--------------------------------- 469 // Used by MemNode::find_previous_store to prove that two base 470 // pointers are never equal. 471 // The pointers are accompanied by their associated allocations, 472 // if any, which have been previously discovered by the caller. 473 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 474 Node* p2, AllocateNode* a2, 475 PhaseTransform* phase) { 476 // Attempt to prove that these two pointers cannot be aliased. 477 // They may both manifestly be allocations, and they should differ. 478 // Or, if they are not both allocations, they can be distinct constants. 479 // Otherwise, one is an allocation and the other a pre-existing value. 480 if (a1 == NULL && a2 == NULL) { // neither an allocation 481 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 482 } else if (a1 != NULL && a2 != NULL) { // both allocations 483 return (a1 != a2); 484 } else if (a1 != NULL) { // one allocation a1 485 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 486 return all_controls_dominate(p2, a1); 487 } else { //(a2 != NULL) // one allocation a2 488 return all_controls_dominate(p1, a2); 489 } 490 return false; 491 } 492 493 494 // The logic for reordering loads and stores uses four steps: 495 // (a) Walk carefully past stores and initializations which we 496 // can prove are independent of this load. 497 // (b) Observe that the next memory state makes an exact match 498 // with self (load or store), and locate the relevant store. 499 // (c) Ensure that, if we were to wire self directly to the store, 500 // the optimizer would fold it up somehow. 501 // (d) Do the rewiring, and return, depending on some other part of 502 // the optimizer to fold up the load. 503 // This routine handles steps (a) and (b). Steps (c) and (d) are 504 // specific to loads and stores, so they are handled by the callers. 505 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 506 // 507 Node* MemNode::find_previous_store(PhaseTransform* phase) { 508 Node* ctrl = in(MemNode::Control); 509 Node* adr = in(MemNode::Address); 510 intptr_t offset = 0; 511 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 512 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 513 514 if (offset == Type::OffsetBot) 515 return NULL; // cannot unalias unless there are precise offsets 516 517 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 518 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 519 520 intptr_t size_in_bytes = memory_size(); 521 522 Node* mem = in(MemNode::Memory); // start searching here... 523 524 int cnt = 50; // Cycle limiter 525 for (;;) { // While we can dance past unrelated stores... 526 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 527 528 if (mem->is_Store()) { 529 Node* st_adr = mem->in(MemNode::Address); 530 intptr_t st_offset = 0; 531 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 532 if (st_base == NULL) 533 break; // inscrutable pointer 534 535 // For raw accesses it's not enough to prove that constant offsets don't intersect. 536 // We need the bases to be the equal in order for the offset check to make sense. 537 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 538 break; 539 } 540 541 if (st_offset != offset && st_offset != Type::OffsetBot) { 542 const int MAX_STORE = BytesPerLong; 543 if (st_offset >= offset + size_in_bytes || 544 st_offset <= offset - MAX_STORE || 545 st_offset <= offset - mem->as_Store()->memory_size()) { 546 // Success: The offsets are provably independent. 547 // (You may ask, why not just test st_offset != offset and be done? 548 // The answer is that stores of different sizes can co-exist 549 // in the same sequence of RawMem effects. We sometimes initialize 550 // a whole 'tile' of array elements with a single jint or jlong.) 551 mem = mem->in(MemNode::Memory); 552 continue; // (a) advance through independent store memory 553 } 554 } 555 if (st_base != base && 556 detect_ptr_independence(base, alloc, 557 st_base, 558 AllocateNode::Ideal_allocation(st_base, phase), 559 phase)) { 560 // Success: The bases are provably independent. 561 mem = mem->in(MemNode::Memory); 562 continue; // (a) advance through independent store memory 563 } 564 565 // (b) At this point, if the bases or offsets do not agree, we lose, 566 // since we have not managed to prove 'this' and 'mem' independent. 567 if (st_base == base && st_offset == offset) { 568 return mem; // let caller handle steps (c), (d) 569 } 570 571 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 572 InitializeNode* st_init = mem->in(0)->as_Initialize(); 573 AllocateNode* st_alloc = st_init->allocation(); 574 if (st_alloc == NULL) 575 break; // something degenerated 576 bool known_identical = false; 577 bool known_independent = false; 578 if (alloc == st_alloc) 579 known_identical = true; 580 else if (alloc != NULL) 581 known_independent = true; 582 else if (all_controls_dominate(this, st_alloc)) 583 known_independent = true; 584 585 if (known_independent) { 586 // The bases are provably independent: Either they are 587 // manifestly distinct allocations, or else the control 588 // of this load dominates the store's allocation. 589 int alias_idx = phase->C->get_alias_index(adr_type()); 590 if (alias_idx == Compile::AliasIdxRaw) { 591 mem = st_alloc->in(TypeFunc::Memory); 592 } else { 593 mem = st_init->memory(alias_idx); 594 } 595 continue; // (a) advance through independent store memory 596 } 597 598 // (b) at this point, if we are not looking at a store initializing 599 // the same allocation we are loading from, we lose. 600 if (known_identical) { 601 // From caller, can_see_stored_value will consult find_captured_store. 602 return mem; // let caller handle steps (c), (d) 603 } 604 605 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 606 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 607 if (mem->is_Proj() && mem->in(0)->is_Call()) { 608 CallNode *call = mem->in(0)->as_Call(); 609 if (!call->may_modify(addr_t, phase)) { 610 mem = call->in(TypeFunc::Memory); 611 continue; // (a) advance through independent call memory 612 } 613 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 614 mem = mem->in(0)->in(TypeFunc::Memory); 615 continue; // (a) advance through independent MemBar memory 616 } else if (mem->is_ClearArray()) { 617 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 618 // (the call updated 'mem' value) 619 continue; // (a) advance through independent allocation memory 620 } else { 621 // Can not bypass initialization of the instance 622 // we are looking for. 623 return mem; 624 } 625 } else if (mem->is_MergeMem()) { 626 int alias_idx = phase->C->get_alias_index(adr_type()); 627 mem = mem->as_MergeMem()->memory_at(alias_idx); 628 continue; // (a) advance through independent MergeMem memory 629 } 630 } 631 632 // Unless there is an explicit 'continue', we must bail out here, 633 // because 'mem' is an inscrutable memory state (e.g., a call). 634 break; 635 } 636 637 return NULL; // bail out 638 } 639 640 //----------------------calculate_adr_type------------------------------------- 641 // Helper function. Notices when the given type of address hits top or bottom. 642 // Also, asserts a cross-check of the type against the expected address type. 643 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 644 if (t == Type::TOP) return NULL; // does not touch memory any more? 645 #ifdef PRODUCT 646 cross_check = NULL; 647 #else 648 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 649 #endif 650 const TypePtr* tp = t->isa_ptr(); 651 if (tp == NULL) { 652 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 653 return TypePtr::BOTTOM; // touches lots of memory 654 } else { 655 #ifdef ASSERT 656 // %%%% [phh] We don't check the alias index if cross_check is 657 // TypeRawPtr::BOTTOM. Needs to be investigated. 658 if (cross_check != NULL && 659 cross_check != TypePtr::BOTTOM && 660 cross_check != TypeRawPtr::BOTTOM) { 661 // Recheck the alias index, to see if it has changed (due to a bug). 662 Compile* C = Compile::current(); 663 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 664 "must stay in the original alias category"); 665 // The type of the address must be contained in the adr_type, 666 // disregarding "null"-ness. 667 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 668 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 669 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 670 "real address must not escape from expected memory type"); 671 } 672 #endif 673 return tp; 674 } 675 } 676 677 //============================================================================= 678 // Should LoadNode::Ideal() attempt to remove control edges? 679 bool LoadNode::can_remove_control() const { 680 return true; 681 } 682 uint LoadNode::size_of() const { return sizeof(*this); } 683 uint LoadNode::cmp( const Node &n ) const 684 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 685 const Type *LoadNode::bottom_type() const { return _type; } 686 uint LoadNode::ideal_reg() const { 687 return _type->ideal_reg(); 688 } 689 690 #ifndef PRODUCT 691 void LoadNode::dump_spec(outputStream *st) const { 692 MemNode::dump_spec(st); 693 if( !Verbose && !WizardMode ) { 694 // standard dump does this in Verbose and WizardMode 695 st->print(" #"); _type->dump_on(st); 696 } 697 if (!_depends_only_on_test) { 698 st->print(" (does not depend only on test)"); 699 } 700 } 701 #endif 702 703 #ifdef ASSERT 704 //----------------------------is_immutable_value------------------------------- 705 // Helper function to allow a raw load without control edge for some cases 706 bool LoadNode::is_immutable_value(Node* adr) { 707 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 708 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 709 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 710 in_bytes(JavaThread::osthread_offset()))); 711 } 712 #endif 713 714 //----------------------------LoadNode::make----------------------------------- 715 // Polymorphic factory method: 716 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, ControlDependency control_dependency) { 717 Compile* C = gvn.C; 718 719 // sanity check the alias category against the created node type 720 assert(!(adr_type->isa_oopptr() && 721 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 722 "use LoadKlassNode instead"); 723 assert(!(adr_type->isa_aryptr() && 724 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 725 "use LoadRangeNode instead"); 726 // Check control edge of raw loads 727 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 728 // oop will be recorded in oop map if load crosses safepoint 729 rt->isa_oopptr() || is_immutable_value(adr), 730 "raw memory operations should have control edge"); 731 switch (bt) { 732 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 733 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 734 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 735 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 736 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 737 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); 738 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); 739 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); 740 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 741 case T_OBJECT: 742 #ifdef _LP64 743 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 744 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency)); 745 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr()); 746 } else 747 #endif 748 { 749 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 750 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo, control_dependency); 751 } 752 } 753 ShouldNotReachHere(); 754 return (LoadNode*)NULL; 755 } 756 757 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) { 758 bool require_atomic = true; 759 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 760 } 761 762 LoadDNode* LoadDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) { 763 bool require_atomic = true; 764 return new (C) LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 765 } 766 767 768 769 //------------------------------hash------------------------------------------- 770 uint LoadNode::hash() const { 771 // unroll addition of interesting fields 772 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 773 } 774 775 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 776 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 777 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 778 bool is_stable_ary = FoldStableValues && 779 (tp != NULL) && (tp->isa_aryptr() != NULL) && 780 tp->isa_aryptr()->is_stable(); 781 782 return (eliminate_boxing && non_volatile) || is_stable_ary; 783 } 784 785 return false; 786 } 787 788 //---------------------------can_see_stored_value------------------------------ 789 // This routine exists to make sure this set of tests is done the same 790 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 791 // will change the graph shape in a way which makes memory alive twice at the 792 // same time (uses the Oracle model of aliasing), then some 793 // LoadXNode::Identity will fold things back to the equivalence-class model 794 // of aliasing. 795 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 796 Node* ld_adr = in(MemNode::Address); 797 intptr_t ld_off = 0; 798 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 799 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 800 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 801 // This is more general than load from boxing objects. 802 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 803 uint alias_idx = atp->index(); 804 bool final = !atp->is_rewritable(); 805 Node* result = NULL; 806 Node* current = st; 807 // Skip through chains of MemBarNodes checking the MergeMems for 808 // new states for the slice of this load. Stop once any other 809 // kind of node is encountered. Loads from final memory can skip 810 // through any kind of MemBar but normal loads shouldn't skip 811 // through MemBarAcquire since the could allow them to move out of 812 // a synchronized region. 813 while (current->is_Proj()) { 814 int opc = current->in(0)->Opcode(); 815 if ((final && (opc == Op_MemBarAcquire || 816 opc == Op_MemBarAcquireLock || 817 opc == Op_LoadFence)) || 818 opc == Op_MemBarRelease || 819 opc == Op_StoreFence || 820 opc == Op_MemBarReleaseLock || 821 opc == Op_MemBarCPUOrder) { 822 Node* mem = current->in(0)->in(TypeFunc::Memory); 823 if (mem->is_MergeMem()) { 824 MergeMemNode* merge = mem->as_MergeMem(); 825 Node* new_st = merge->memory_at(alias_idx); 826 if (new_st == merge->base_memory()) { 827 // Keep searching 828 current = new_st; 829 continue; 830 } 831 // Save the new memory state for the slice and fall through 832 // to exit. 833 result = new_st; 834 } 835 } 836 break; 837 } 838 if (result != NULL) { 839 st = result; 840 } 841 } 842 843 // Loop around twice in the case Load -> Initialize -> Store. 844 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 845 for (int trip = 0; trip <= 1; trip++) { 846 847 if (st->is_Store()) { 848 Node* st_adr = st->in(MemNode::Address); 849 if (!phase->eqv(st_adr, ld_adr)) { 850 // Try harder before giving up... Match raw and non-raw pointers. 851 intptr_t st_off = 0; 852 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 853 if (alloc == NULL) return NULL; 854 if (alloc != ld_alloc) return NULL; 855 if (ld_off != st_off) return NULL; 856 // At this point we have proven something like this setup: 857 // A = Allocate(...) 858 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 859 // S = StoreQ(, AddP(, A.Parm , #Off), V) 860 // (Actually, we haven't yet proven the Q's are the same.) 861 // In other words, we are loading from a casted version of 862 // the same pointer-and-offset that we stored to. 863 // Thus, we are able to replace L by V. 864 } 865 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 866 if (store_Opcode() != st->Opcode()) 867 return NULL; 868 return st->in(MemNode::ValueIn); 869 } 870 871 // A load from a freshly-created object always returns zero. 872 // (This can happen after LoadNode::Ideal resets the load's memory input 873 // to find_captured_store, which returned InitializeNode::zero_memory.) 874 if (st->is_Proj() && st->in(0)->is_Allocate() && 875 (st->in(0) == ld_alloc) && 876 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 877 // return a zero value for the load's basic type 878 // (This is one of the few places where a generic PhaseTransform 879 // can create new nodes. Think of it as lazily manifesting 880 // virtually pre-existing constants.) 881 return phase->zerocon(memory_type()); 882 } 883 884 // A load from an initialization barrier can match a captured store. 885 if (st->is_Proj() && st->in(0)->is_Initialize()) { 886 InitializeNode* init = st->in(0)->as_Initialize(); 887 AllocateNode* alloc = init->allocation(); 888 if ((alloc != NULL) && (alloc == ld_alloc)) { 889 // examine a captured store value 890 st = init->find_captured_store(ld_off, memory_size(), phase); 891 if (st != NULL) 892 continue; // take one more trip around 893 } 894 } 895 896 // Load boxed value from result of valueOf() call is input parameter. 897 if (this->is_Load() && ld_adr->is_AddP() && 898 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 899 intptr_t ignore = 0; 900 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 901 base = ShenandoahBarrierNode::skip_through_barrier(base); 902 if (base != NULL && base->is_Proj() && 903 base->as_Proj()->_con == TypeFunc::Parms && 904 base->in(0)->is_CallStaticJava() && 905 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 906 return base->in(0)->in(TypeFunc::Parms); 907 } 908 } 909 910 break; 911 } 912 913 return NULL; 914 } 915 916 //----------------------is_instance_field_load_with_local_phi------------------ 917 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 918 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 919 in(Address)->is_AddP() ) { 920 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 921 // Only instances and boxed values. 922 if( t_oop != NULL && 923 (t_oop->is_ptr_to_boxed_value() || 924 t_oop->is_known_instance_field()) && 925 t_oop->offset() != Type::OffsetBot && 926 t_oop->offset() != Type::OffsetTop) { 927 return true; 928 } 929 } 930 return false; 931 } 932 933 //------------------------------Identity--------------------------------------- 934 // Loads are identity if previous store is to same address 935 Node *LoadNode::Identity( PhaseTransform *phase ) { 936 // If the previous store-maker is the right kind of Store, and the store is 937 // to the same address, then we are equal to the value stored. 938 Node* mem = in(Memory); 939 Node* value = can_see_stored_value(mem, phase); 940 if( value ) { 941 // byte, short & char stores truncate naturally. 942 // A load has to load the truncated value which requires 943 // some sort of masking operation and that requires an 944 // Ideal call instead of an Identity call. 945 if (memory_size() < BytesPerInt) { 946 // If the input to the store does not fit with the load's result type, 947 // it must be truncated via an Ideal call. 948 if (!phase->type(value)->higher_equal(phase->type(this))) 949 return this; 950 } 951 PhaseIterGVN* igvn = phase->is_IterGVN(); 952 if (UseShenandoahGC && 953 igvn != NULL && 954 value->is_Phi() && 955 value->req() > 2 && 956 value->in(1) != NULL && 957 value->in(1)->is_ShenandoahBarrier()) { 958 if (igvn->_worklist.member(value) || 959 igvn->_worklist.member(value->in(0)) || 960 (value->in(0)->in(1) != NULL && 961 value->in(0)->in(1)->is_IfProj() && 962 (igvn->_worklist.member(value->in(0)->in(1)) || 963 (value->in(0)->in(1)->in(0) != NULL && 964 igvn->_worklist.member(value->in(0)->in(1)->in(0)))))) { 965 igvn->_worklist.push(this); 966 return this; 967 } 968 } 969 // (This works even when value is a Con, but LoadNode::Value 970 // usually runs first, producing the singleton type of the Con.) 971 if (UseShenandoahGC) { 972 Node* value_no_barrier = ShenandoahBarrierNode::skip_through_barrier(value->Opcode() == Op_EncodeP ? value->in(1) : value); 973 if (value->Opcode() == Op_EncodeP) { 974 if (value_no_barrier != value->in(1)) { 975 Node* encode = value->clone(); 976 encode->set_req(1, value_no_barrier); 977 encode = phase->transform(encode); 978 return encode; 979 } 980 } else { 981 return value_no_barrier; 982 } 983 } 984 985 return value; 986 } 987 988 // Search for an existing data phi which was generated before for the same 989 // instance's field to avoid infinite generation of phis in a loop. 990 Node *region = mem->in(0); 991 if (is_instance_field_load_with_local_phi(region)) { 992 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 993 int this_index = phase->C->get_alias_index(addr_t); 994 int this_offset = addr_t->offset(); 995 int this_iid = addr_t->instance_id(); 996 if (!addr_t->is_known_instance() && 997 addr_t->is_ptr_to_boxed_value()) { 998 // Use _idx of address base (could be Phi node) for boxed values. 999 intptr_t ignore = 0; 1000 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1001 if (base == NULL) { 1002 return this; 1003 } 1004 this_iid = base->_idx; 1005 } 1006 const Type* this_type = bottom_type(); 1007 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1008 Node* phi = region->fast_out(i); 1009 if (phi->is_Phi() && phi != mem && 1010 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1011 return phi; 1012 } 1013 } 1014 } 1015 1016 return this; 1017 } 1018 1019 // We're loading from an object which has autobox behaviour. 1020 // If this object is result of a valueOf call we'll have a phi 1021 // merging a newly allocated object and a load from the cache. 1022 // We want to replace this load with the original incoming 1023 // argument to the valueOf call. 1024 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1025 assert(phase->C->eliminate_boxing(), "sanity"); 1026 intptr_t ignore = 0; 1027 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1028 if ((base == NULL) || base->is_Phi()) { 1029 // Push the loads from the phi that comes from valueOf up 1030 // through it to allow elimination of the loads and the recovery 1031 // of the original value. It is done in split_through_phi(). 1032 return NULL; 1033 } else if (base->is_Load() || 1034 base->is_DecodeN() && base->in(1)->is_Load()) { 1035 // Eliminate the load of boxed value for integer types from the cache 1036 // array by deriving the value from the index into the array. 1037 // Capture the offset of the load and then reverse the computation. 1038 1039 // Get LoadN node which loads a boxing object from 'cache' array. 1040 if (base->is_DecodeN()) { 1041 base = base->in(1); 1042 } 1043 if (!base->in(Address)->is_AddP()) { 1044 return NULL; // Complex address 1045 } 1046 AddPNode* address = base->in(Address)->as_AddP(); 1047 Node* cache_base = ShenandoahBarrierNode::skip_through_barrier(address->in(AddPNode::Base)); 1048 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1049 // Get ConP node which is static 'cache' field. 1050 cache_base = cache_base->in(1); 1051 } 1052 if ((cache_base != NULL) && cache_base->is_Con()) { 1053 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1054 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1055 Node* elements[4]; 1056 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1057 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1058 if ((count > 0) && elements[0]->is_Con() && 1059 ((count == 1) || 1060 (count == 2) && elements[1]->Opcode() == Op_LShiftX && 1061 elements[1]->in(2) == phase->intcon(shift))) { 1062 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1063 // Fetch the box object cache[0] at the base of the array and get its value 1064 ciInstance* box = array->obj_at(0)->as_instance(); 1065 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1066 assert(ik->is_box_klass(), "sanity"); 1067 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1068 if (ik->nof_nonstatic_fields() == 1) { 1069 // This should be true nonstatic_field_at requires calling 1070 // nof_nonstatic_fields so check it anyway 1071 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1072 BasicType bt = c.basic_type(); 1073 // Only integer types have boxing cache. 1074 assert(bt == T_BOOLEAN || bt == T_CHAR || 1075 bt == T_BYTE || bt == T_SHORT || 1076 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt))); 1077 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1078 if (cache_low != (int)cache_low) { 1079 return NULL; // should not happen since cache is array indexed by value 1080 } 1081 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1082 if (offset != (int)offset) { 1083 return NULL; // should not happen since cache is array indexed by value 1084 } 1085 // Add up all the offsets making of the address of the load 1086 Node* result = elements[0]; 1087 for (int i = 1; i < count; i++) { 1088 result = phase->transform(new (phase->C) AddXNode(result, elements[i])); 1089 } 1090 // Remove the constant offset from the address and then 1091 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset))); 1092 // remove the scaling of the offset to recover the original index. 1093 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1094 // Peel the shift off directly but wrap it in a dummy node 1095 // since Ideal can't return existing nodes 1096 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0)); 1097 } else if (result->is_Add() && result->in(2)->is_Con() && 1098 result->in(1)->Opcode() == Op_LShiftX && 1099 result->in(1)->in(2) == phase->intcon(shift)) { 1100 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1101 // but for boxing cache access we know that X<<Z will not overflow 1102 // (there is range check) so we do this optimizatrion by hand here. 1103 Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift)); 1104 result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1105 } else { 1106 result = new (phase->C) RShiftXNode(result, phase->intcon(shift)); 1107 } 1108 #ifdef _LP64 1109 if (bt != T_LONG) { 1110 result = new (phase->C) ConvL2INode(phase->transform(result)); 1111 } 1112 #else 1113 if (bt == T_LONG) { 1114 result = new (phase->C) ConvI2LNode(phase->transform(result)); 1115 } 1116 #endif 1117 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1118 // Need to preserve unboxing load type if it is unsigned. 1119 switch(this->Opcode()) { 1120 case Op_LoadUB: 1121 result = new (phase->C) AndINode(phase->transform(result), phase->intcon(0xFF)); 1122 break; 1123 case Op_LoadUS: 1124 result = new (phase->C) AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1125 break; 1126 } 1127 return result; 1128 } 1129 } 1130 } 1131 } 1132 } 1133 return NULL; 1134 } 1135 1136 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1137 Node* region = phi->in(0); 1138 if (region == NULL) { 1139 return false; // Wait stable graph 1140 } 1141 uint cnt = phi->req(); 1142 for (uint i = 1; i < cnt; i++) { 1143 Node* rc = region->in(i); 1144 if (rc == NULL || phase->type(rc) == Type::TOP) 1145 return false; // Wait stable graph 1146 Node* in = phi->in(i); 1147 if (in == NULL || phase->type(in) == Type::TOP) 1148 return false; // Wait stable graph 1149 } 1150 return true; 1151 } 1152 //------------------------------split_through_phi------------------------------ 1153 // Split instance or boxed field load through Phi. 1154 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1155 Node* mem = in(Memory); 1156 Node* address = in(Address); 1157 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1158 1159 assert((t_oop != NULL) && 1160 (t_oop->is_known_instance_field() || 1161 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1162 1163 Compile* C = phase->C; 1164 intptr_t ignore = 0; 1165 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1166 bool base_is_phi = (base != NULL) && base->is_Phi(); 1167 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1168 (base != NULL) && (base == address->in(AddPNode::Base)) && 1169 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1170 1171 if (!((mem->is_Phi() || base_is_phi) && 1172 (load_boxed_values || t_oop->is_known_instance_field()))) { 1173 return NULL; // memory is not Phi 1174 } 1175 1176 if (mem->is_Phi()) { 1177 if (!stable_phi(mem->as_Phi(), phase)) { 1178 return NULL; // Wait stable graph 1179 } 1180 uint cnt = mem->req(); 1181 // Check for loop invariant memory. 1182 if (cnt == 3) { 1183 for (uint i = 1; i < cnt; i++) { 1184 Node* in = mem->in(i); 1185 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1186 if (m == mem) { 1187 set_req(Memory, mem->in(cnt - i)); 1188 return this; // made change 1189 } 1190 } 1191 } 1192 } 1193 if (base_is_phi) { 1194 if (!stable_phi(base->as_Phi(), phase)) { 1195 return NULL; // Wait stable graph 1196 } 1197 uint cnt = base->req(); 1198 // Check for loop invariant memory. 1199 if (cnt == 3) { 1200 for (uint i = 1; i < cnt; i++) { 1201 if (base->in(i) == base) { 1202 return NULL; // Wait stable graph 1203 } 1204 } 1205 } 1206 } 1207 1208 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1209 1210 // Split through Phi (see original code in loopopts.cpp). 1211 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1212 1213 // Do nothing here if Identity will find a value 1214 // (to avoid infinite chain of value phis generation). 1215 if (!phase->eqv(this, this->Identity(phase))) 1216 return NULL; 1217 1218 // Select Region to split through. 1219 Node* region; 1220 if (!base_is_phi) { 1221 assert(mem->is_Phi(), "sanity"); 1222 region = mem->in(0); 1223 // Skip if the region dominates some control edge of the address. 1224 if (!MemNode::all_controls_dominate(address, region)) 1225 return NULL; 1226 } else if (!mem->is_Phi()) { 1227 assert(base_is_phi, "sanity"); 1228 region = base->in(0); 1229 // Skip if the region dominates some control edge of the memory. 1230 if (!MemNode::all_controls_dominate(mem, region)) 1231 return NULL; 1232 } else if (base->in(0) != mem->in(0)) { 1233 assert(base_is_phi && mem->is_Phi(), "sanity"); 1234 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1235 region = base->in(0); 1236 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1237 region = mem->in(0); 1238 } else { 1239 return NULL; // complex graph 1240 } 1241 } else { 1242 assert(base->in(0) == mem->in(0), "sanity"); 1243 region = mem->in(0); 1244 } 1245 1246 const Type* this_type = this->bottom_type(); 1247 int this_index = C->get_alias_index(t_oop); 1248 int this_offset = t_oop->offset(); 1249 int this_iid = t_oop->instance_id(); 1250 if (!t_oop->is_known_instance() && load_boxed_values) { 1251 // Use _idx of address base for boxed values. 1252 this_iid = base->_idx; 1253 } 1254 PhaseIterGVN* igvn = phase->is_IterGVN(); 1255 Node* phi = new (C) PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1256 for (uint i = 1; i < region->req(); i++) { 1257 Node* x; 1258 Node* the_clone = NULL; 1259 if (region->in(i) == C->top()) { 1260 x = C->top(); // Dead path? Use a dead data op 1261 } else { 1262 x = this->clone(); // Else clone up the data op 1263 the_clone = x; // Remember for possible deletion. 1264 // Alter data node to use pre-phi inputs 1265 if (this->in(0) == region) { 1266 x->set_req(0, region->in(i)); 1267 } else { 1268 x->set_req(0, NULL); 1269 } 1270 if (mem->is_Phi() && (mem->in(0) == region)) { 1271 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1272 } 1273 if (address->is_Phi() && address->in(0) == region) { 1274 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1275 } 1276 if (base_is_phi && (base->in(0) == region)) { 1277 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1278 Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1279 x->set_req(Address, adr_x); 1280 } 1281 } 1282 // Check for a 'win' on some paths 1283 const Type *t = x->Value(igvn); 1284 1285 bool singleton = t->singleton(); 1286 1287 // See comments in PhaseIdealLoop::split_thru_phi(). 1288 if (singleton && t == Type::TOP) { 1289 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1290 } 1291 1292 if (singleton) { 1293 x = igvn->makecon(t); 1294 } else { 1295 // We now call Identity to try to simplify the cloned node. 1296 // Note that some Identity methods call phase->type(this). 1297 // Make sure that the type array is big enough for 1298 // our new node, even though we may throw the node away. 1299 // (This tweaking with igvn only works because x is a new node.) 1300 igvn->set_type(x, t); 1301 // If x is a TypeNode, capture any more-precise type permanently into Node 1302 // otherwise it will be not updated during igvn->transform since 1303 // igvn->type(x) is set to x->Value() already. 1304 x->raise_bottom_type(t); 1305 Node *y = x->Identity(igvn); 1306 if (y != x) { 1307 x = y; 1308 } else { 1309 y = igvn->hash_find_insert(x); 1310 if (y) { 1311 x = y; 1312 } else { 1313 // Else x is a new node we are keeping 1314 // We do not need register_new_node_with_optimizer 1315 // because set_type has already been called. 1316 igvn->_worklist.push(x); 1317 } 1318 } 1319 } 1320 if (x != the_clone && the_clone != NULL) { 1321 igvn->remove_dead_node(the_clone); 1322 } 1323 phi->set_req(i, x); 1324 } 1325 // Record Phi 1326 igvn->register_new_node_with_optimizer(phi); 1327 return phi; 1328 } 1329 1330 //------------------------------Ideal------------------------------------------ 1331 // If the load is from Field memory and the pointer is non-null, it might be possible to 1332 // zero out the control input. 1333 // If the offset is constant and the base is an object allocation, 1334 // try to hook me up to the exact initializing store. 1335 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1336 Node* p = MemNode::Ideal_common(phase, can_reshape); 1337 if (p) return (p == NodeSentinel) ? NULL : p; 1338 1339 Node* ctrl = in(MemNode::Control); 1340 Node* address = in(MemNode::Address); 1341 1342 // Skip up past a SafePoint control. Cannot do this for Stores because 1343 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1344 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1345 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1346 ctrl = ctrl->in(0); 1347 set_req(MemNode::Control,ctrl); 1348 } 1349 1350 intptr_t ignore = 0; 1351 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1352 if (base != NULL 1353 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1354 // Check for useless control edge in some common special cases 1355 if (in(MemNode::Control) != NULL 1356 && can_remove_control() 1357 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1358 && all_controls_dominate(base, phase->C->start())) { 1359 // A method-invariant, non-null address (constant or 'this' argument). 1360 set_req(MemNode::Control, NULL); 1361 } 1362 } 1363 1364 Node* mem = in(MemNode::Memory); 1365 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1366 1367 if (can_reshape && (addr_t != NULL)) { 1368 // try to optimize our memory input 1369 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1370 if (opt_mem != mem) { 1371 set_req(MemNode::Memory, opt_mem); 1372 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1373 return this; 1374 } 1375 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1376 if ((t_oop != NULL) && 1377 (t_oop->is_known_instance_field() || 1378 t_oop->is_ptr_to_boxed_value())) { 1379 PhaseIterGVN *igvn = phase->is_IterGVN(); 1380 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1381 // Delay this transformation until memory Phi is processed. 1382 phase->is_IterGVN()->_worklist.push(this); 1383 return NULL; 1384 } 1385 // Split instance field load through Phi. 1386 Node* result = split_through_phi(phase); 1387 if (result != NULL) return result; 1388 1389 if (t_oop->is_ptr_to_boxed_value()) { 1390 Node* result = eliminate_autobox(phase); 1391 if (result != NULL) return result; 1392 } 1393 } 1394 } 1395 1396 // Check for prior store with a different base or offset; make Load 1397 // independent. Skip through any number of them. Bail out if the stores 1398 // are in an endless dead cycle and report no progress. This is a key 1399 // transform for Reflection. However, if after skipping through the Stores 1400 // we can't then fold up against a prior store do NOT do the transform as 1401 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1402 // array memory alive twice: once for the hoisted Load and again after the 1403 // bypassed Store. This situation only works if EVERYBODY who does 1404 // anti-dependence work knows how to bypass. I.e. we need all 1405 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1406 // the alias index stuff. So instead, peek through Stores and IFF we can 1407 // fold up, do so. 1408 Node* prev_mem = find_previous_store(phase); 1409 // Steps (a), (b): Walk past independent stores to find an exact match. 1410 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1411 // (c) See if we can fold up on the spot, but don't fold up here. 1412 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1413 // just return a prior value, which is done by Identity calls. 1414 if (can_see_stored_value(prev_mem, phase)) { 1415 // Make ready for step (d): 1416 set_req(MemNode::Memory, prev_mem); 1417 return this; 1418 } 1419 } 1420 1421 return NULL; // No further progress 1422 } 1423 1424 // Helper to recognize certain Klass fields which are invariant across 1425 // some group of array types (e.g., int[] or all T[] where T < Object). 1426 const Type* 1427 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1428 ciKlass* klass) const { 1429 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1430 // The field is Klass::_modifier_flags. Return its (constant) value. 1431 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1432 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1433 return TypeInt::make(klass->modifier_flags()); 1434 } 1435 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1436 // The field is Klass::_access_flags. Return its (constant) value. 1437 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1438 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1439 return TypeInt::make(klass->access_flags()); 1440 } 1441 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1442 // The field is Klass::_layout_helper. Return its constant value if known. 1443 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1444 return TypeInt::make(klass->layout_helper()); 1445 } 1446 1447 // No match. 1448 return NULL; 1449 } 1450 1451 // Try to constant-fold a stable array element. 1452 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) { 1453 assert(ary->const_oop(), "array should be constant"); 1454 assert(ary->is_stable(), "array should be stable"); 1455 1456 // Decode the results of GraphKit::array_element_address. 1457 ciArray* aobj = ary->const_oop()->as_array(); 1458 ciConstant con = aobj->element_value_by_offset(off); 1459 1460 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) { 1461 const Type* con_type = Type::make_from_constant(con); 1462 if (con_type != NULL) { 1463 if (con_type->isa_aryptr()) { 1464 // Join with the array element type, in case it is also stable. 1465 int dim = ary->stable_dimension(); 1466 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1); 1467 } 1468 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) { 1469 con_type = con_type->make_narrowoop(); 1470 } 1471 #ifndef PRODUCT 1472 if (TraceIterativeGVN) { 1473 tty->print("FoldStableValues: array element [off=%d]: con_type=", off); 1474 con_type->dump(); tty->cr(); 1475 } 1476 #endif //PRODUCT 1477 return con_type; 1478 } 1479 } 1480 return NULL; 1481 } 1482 1483 //------------------------------Value----------------------------------------- 1484 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1485 // Either input is TOP ==> the result is TOP 1486 Node* mem = in(MemNode::Memory); 1487 const Type *t1 = phase->type(mem); 1488 if (t1 == Type::TOP) return Type::TOP; 1489 Node* adr = in(MemNode::Address); 1490 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1491 if (tp == NULL || tp->empty()) return Type::TOP; 1492 int off = tp->offset(); 1493 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1494 Compile* C = phase->C; 1495 1496 // Try to guess loaded type from pointer type 1497 if (tp->isa_aryptr()) { 1498 const TypeAryPtr* ary = tp->is_aryptr(); 1499 const Type* t = ary->elem(); 1500 1501 // Determine whether the reference is beyond the header or not, by comparing 1502 // the offset against the offset of the start of the array's data. 1503 // Different array types begin at slightly different offsets (12 vs. 16). 1504 // We choose T_BYTE as an example base type that is least restrictive 1505 // as to alignment, which will therefore produce the smallest 1506 // possible base offset. 1507 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1508 const bool off_beyond_header = (off != ShenandoahBrooksPointer::byte_offset() || !UseShenandoahGC) && ((uint)off >= (uint)min_base_off); 1509 1510 // Try to constant-fold a stable array element. 1511 if (FoldStableValues && ary->is_stable()) { 1512 // Make sure the reference is not into the header and the offset is constant 1513 ciObject* aobj = NULL; 1514 if (UseShenandoahGC && adr->is_AddP() && !adr->in(AddPNode::Base)->is_top()) { 1515 Node* base = ShenandoahBarrierNode::skip_through_barrier(adr->in(AddPNode::Base)); 1516 if (!base->is_top()) { 1517 ary = phase->type(base)->is_aryptr(); 1518 aobj = ary->const_oop(); 1519 } 1520 } else { 1521 aobj = ary->const_oop(); 1522 } 1523 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1524 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type()); 1525 if (con_type != NULL) { 1526 return con_type; 1527 } 1528 } 1529 } 1530 1531 // Don't do this for integer types. There is only potential profit if 1532 // the element type t is lower than _type; that is, for int types, if _type is 1533 // more restrictive than t. This only happens here if one is short and the other 1534 // char (both 16 bits), and in those cases we've made an intentional decision 1535 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1536 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1537 // 1538 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1539 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1540 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1541 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1542 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1543 // In fact, that could have been the original type of p1, and p1 could have 1544 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1545 // expression (LShiftL quux 3) independently optimized to the constant 8. 1546 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1547 && (_type->isa_vect() == NULL) 1548 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1549 // t might actually be lower than _type, if _type is a unique 1550 // concrete subclass of abstract class t. 1551 if (off_beyond_header) { // is the offset beyond the header? 1552 const Type* jt = t->join_speculative(_type); 1553 // In any case, do not allow the join, per se, to empty out the type. 1554 if (jt->empty() && !t->empty()) { 1555 // This can happen if a interface-typed array narrows to a class type. 1556 jt = _type; 1557 } 1558 #ifdef ASSERT 1559 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1560 // The pointers in the autobox arrays are always non-null 1561 Node* base = adr->in(AddPNode::Base); 1562 if ((base != NULL) && base->is_DecodeN()) { 1563 // Get LoadN node which loads IntegerCache.cache field 1564 base = base->in(1); 1565 } 1566 if ((base != NULL) && base->is_Con()) { 1567 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1568 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1569 // It could be narrow oop 1570 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1571 } 1572 } 1573 } 1574 #endif 1575 return jt; 1576 } 1577 } 1578 } else if (tp->base() == Type::InstPtr) { 1579 ciEnv* env = C->env(); 1580 const TypeInstPtr* tinst = tp->is_instptr(); 1581 ciKlass* klass = tinst->klass(); 1582 assert( off != Type::OffsetBot || 1583 // arrays can be cast to Objects 1584 tp->is_oopptr()->klass()->is_java_lang_Object() || 1585 // unsafe field access may not have a constant offset 1586 C->has_unsafe_access(), 1587 "Field accesses must be precise" ); 1588 // For oop loads, we expect the _type to be precise 1589 if (klass == env->String_klass() && 1590 adr->is_AddP() && off != Type::OffsetBot) { 1591 // For constant Strings treat the final fields as compile time constants. 1592 Node* base = adr->in(AddPNode::Base); 1593 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1594 if (t != NULL && t->singleton()) { 1595 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1596 if (field != NULL && field->is_final()) { 1597 ciObject* string = t->const_oop(); 1598 ciConstant constant = string->as_instance()->field_value(field); 1599 if (constant.basic_type() == T_INT) { 1600 return TypeInt::make(constant.as_int()); 1601 } else if (constant.basic_type() == T_ARRAY) { 1602 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1603 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1604 } else { 1605 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1606 } 1607 } 1608 } 1609 } 1610 } 1611 // Optimizations for constant objects 1612 ciObject* const_oop = NULL; 1613 if (UseShenandoahGC && adr->is_AddP() && !adr->in(AddPNode::Base)->is_top()) { 1614 Node* base = ShenandoahBarrierNode::skip_through_barrier(adr->in(AddPNode::Base)); 1615 if (phase->type(base) != Type::TOP) { 1616 const TypePtr* base_t = phase->type(base)->is_ptr(); 1617 if (base_t != TypePtr::NULL_PTR) { 1618 tinst = base_t->is_instptr(); 1619 const_oop = tinst->const_oop(); 1620 } 1621 } 1622 } else { 1623 const_oop = tinst->const_oop(); 1624 } 1625 if (const_oop != NULL) { 1626 // For constant Boxed value treat the target field as a compile time constant. 1627 if (tinst->is_ptr_to_boxed_value()) { 1628 return tinst->get_const_boxed_value(); 1629 } else 1630 // For constant CallSites treat the target field as a compile time constant. 1631 if (const_oop->is_call_site()) { 1632 ciCallSite* call_site = const_oop->as_call_site(); 1633 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1634 if (field != NULL && field->is_call_site_target()) { 1635 ciMethodHandle* target = call_site->get_target(); 1636 if (target != NULL) { // just in case 1637 ciConstant constant(T_OBJECT, target); 1638 const Type* t; 1639 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1640 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1641 } else { 1642 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1643 } 1644 // Add a dependence for invalidation of the optimization. 1645 if (!call_site->is_constant_call_site()) { 1646 C->dependencies()->assert_call_site_target_value(call_site, target); 1647 } 1648 return t; 1649 } 1650 } 1651 } 1652 } 1653 } else if (tp->base() == Type::KlassPtr) { 1654 assert( off != Type::OffsetBot || 1655 // arrays can be cast to Objects 1656 tp->is_klassptr()->klass()->is_java_lang_Object() || 1657 // also allow array-loading from the primary supertype 1658 // array during subtype checks 1659 Opcode() == Op_LoadKlass, 1660 "Field accesses must be precise" ); 1661 // For klass/static loads, we expect the _type to be precise 1662 } 1663 1664 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1665 if (tkls != NULL && !StressReflectiveCode) { 1666 ciKlass* klass = tkls->klass(); 1667 if (klass->is_loaded() && tkls->klass_is_exact()) { 1668 // We are loading a field from a Klass metaobject whose identity 1669 // is known at compile time (the type is "exact" or "precise"). 1670 // Check for fields we know are maintained as constants by the VM. 1671 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1672 // The field is Klass::_super_check_offset. Return its (constant) value. 1673 // (Folds up type checking code.) 1674 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1675 return TypeInt::make(klass->super_check_offset()); 1676 } 1677 // Compute index into primary_supers array 1678 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1679 // Check for overflowing; use unsigned compare to handle the negative case. 1680 if( depth < ciKlass::primary_super_limit() ) { 1681 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1682 // (Folds up type checking code.) 1683 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1684 ciKlass *ss = klass->super_of_depth(depth); 1685 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1686 } 1687 const Type* aift = load_array_final_field(tkls, klass); 1688 if (aift != NULL) return aift; 1689 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset()) 1690 && klass->is_array_klass()) { 1691 // The field is ArrayKlass::_component_mirror. Return its (constant) value. 1692 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1693 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1694 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1695 } 1696 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1697 // The field is Klass::_java_mirror. Return its (constant) value. 1698 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1699 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1700 return TypeInstPtr::make(klass->java_mirror()); 1701 } 1702 } 1703 1704 // We can still check if we are loading from the primary_supers array at a 1705 // shallow enough depth. Even though the klass is not exact, entries less 1706 // than or equal to its super depth are correct. 1707 if (klass->is_loaded() ) { 1708 ciType *inner = klass; 1709 while( inner->is_obj_array_klass() ) 1710 inner = inner->as_obj_array_klass()->base_element_type(); 1711 if( inner->is_instance_klass() && 1712 !inner->as_instance_klass()->flags().is_interface() ) { 1713 // Compute index into primary_supers array 1714 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1715 // Check for overflowing; use unsigned compare to handle the negative case. 1716 if( depth < ciKlass::primary_super_limit() && 1717 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1718 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1719 // (Folds up type checking code.) 1720 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1721 ciKlass *ss = klass->super_of_depth(depth); 1722 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1723 } 1724 } 1725 } 1726 1727 // If the type is enough to determine that the thing is not an array, 1728 // we can give the layout_helper a positive interval type. 1729 // This will help short-circuit some reflective code. 1730 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1731 && !klass->is_array_klass() // not directly typed as an array 1732 && !klass->is_interface() // specifically not Serializable & Cloneable 1733 && !klass->is_java_lang_Object() // not the supertype of all T[] 1734 ) { 1735 // Note: When interfaces are reliable, we can narrow the interface 1736 // test to (klass != Serializable && klass != Cloneable). 1737 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1738 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1739 // The key property of this type is that it folds up tests 1740 // for array-ness, since it proves that the layout_helper is positive. 1741 // Thus, a generic value like the basic object layout helper works fine. 1742 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1743 } 1744 } 1745 1746 // If we are loading from a freshly-allocated object, produce a zero, 1747 // if the load is provably beyond the header of the object. 1748 // (Also allow a variable load from a fresh array to produce zero.) 1749 const TypeOopPtr *tinst = tp->isa_oopptr(); 1750 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1751 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1752 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1753 Node* value = can_see_stored_value(mem,phase); 1754 if (value != NULL && value->is_Con()) { 1755 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1756 return value->bottom_type(); 1757 } 1758 } 1759 1760 if (is_instance) { 1761 // If we have an instance type and our memory input is the 1762 // programs's initial memory state, there is no matching store, 1763 // so just return a zero of the appropriate type 1764 Node *mem = in(MemNode::Memory); 1765 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1766 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1767 return Type::get_zero_type(_type->basic_type()); 1768 } 1769 } 1770 return _type; 1771 } 1772 1773 //------------------------------match_edge------------------------------------- 1774 // Do we Match on this edge index or not? Match only the address. 1775 uint LoadNode::match_edge(uint idx) const { 1776 return idx == MemNode::Address; 1777 } 1778 1779 //--------------------------LoadBNode::Ideal-------------------------------------- 1780 // 1781 // If the previous store is to the same address as this load, 1782 // and the value stored was larger than a byte, replace this load 1783 // with the value stored truncated to a byte. If no truncation is 1784 // needed, the replacement is done in LoadNode::Identity(). 1785 // 1786 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1787 Node* mem = in(MemNode::Memory); 1788 Node* value = can_see_stored_value(mem,phase); 1789 if( value && !phase->type(value)->higher_equal( _type ) ) { 1790 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) ); 1791 return new (phase->C) RShiftINode(result, phase->intcon(24)); 1792 } 1793 // Identity call will handle the case where truncation is not needed. 1794 return LoadNode::Ideal(phase, can_reshape); 1795 } 1796 1797 const Type* LoadBNode::Value(PhaseTransform *phase) const { 1798 Node* mem = in(MemNode::Memory); 1799 Node* value = can_see_stored_value(mem,phase); 1800 if (value != NULL && value->is_Con() && 1801 !value->bottom_type()->higher_equal(_type)) { 1802 // If the input to the store does not fit with the load's result type, 1803 // it must be truncated. We can't delay until Ideal call since 1804 // a singleton Value is needed for split_thru_phi optimization. 1805 int con = value->get_int(); 1806 return TypeInt::make((con << 24) >> 24); 1807 } 1808 return LoadNode::Value(phase); 1809 } 1810 1811 //--------------------------LoadUBNode::Ideal------------------------------------- 1812 // 1813 // If the previous store is to the same address as this load, 1814 // and the value stored was larger than a byte, replace this load 1815 // with the value stored truncated to a byte. If no truncation is 1816 // needed, the replacement is done in LoadNode::Identity(). 1817 // 1818 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1819 Node* mem = in(MemNode::Memory); 1820 Node* value = can_see_stored_value(mem, phase); 1821 if (value && !phase->type(value)->higher_equal(_type)) 1822 return new (phase->C) AndINode(value, phase->intcon(0xFF)); 1823 // Identity call will handle the case where truncation is not needed. 1824 return LoadNode::Ideal(phase, can_reshape); 1825 } 1826 1827 const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1828 Node* mem = in(MemNode::Memory); 1829 Node* value = can_see_stored_value(mem,phase); 1830 if (value != NULL && value->is_Con() && 1831 !value->bottom_type()->higher_equal(_type)) { 1832 // If the input to the store does not fit with the load's result type, 1833 // it must be truncated. We can't delay until Ideal call since 1834 // a singleton Value is needed for split_thru_phi optimization. 1835 int con = value->get_int(); 1836 return TypeInt::make(con & 0xFF); 1837 } 1838 return LoadNode::Value(phase); 1839 } 1840 1841 //--------------------------LoadUSNode::Ideal------------------------------------- 1842 // 1843 // If the previous store is to the same address as this load, 1844 // and the value stored was larger than a char, replace this load 1845 // with the value stored truncated to a char. If no truncation is 1846 // needed, the replacement is done in LoadNode::Identity(). 1847 // 1848 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1849 Node* mem = in(MemNode::Memory); 1850 Node* value = can_see_stored_value(mem,phase); 1851 if( value && !phase->type(value)->higher_equal( _type ) ) 1852 return new (phase->C) AndINode(value,phase->intcon(0xFFFF)); 1853 // Identity call will handle the case where truncation is not needed. 1854 return LoadNode::Ideal(phase, can_reshape); 1855 } 1856 1857 const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1858 Node* mem = in(MemNode::Memory); 1859 Node* value = can_see_stored_value(mem,phase); 1860 if (value != NULL && value->is_Con() && 1861 !value->bottom_type()->higher_equal(_type)) { 1862 // If the input to the store does not fit with the load's result type, 1863 // it must be truncated. We can't delay until Ideal call since 1864 // a singleton Value is needed for split_thru_phi optimization. 1865 int con = value->get_int(); 1866 return TypeInt::make(con & 0xFFFF); 1867 } 1868 return LoadNode::Value(phase); 1869 } 1870 1871 //--------------------------LoadSNode::Ideal-------------------------------------- 1872 // 1873 // If the previous store is to the same address as this load, 1874 // and the value stored was larger than a short, replace this load 1875 // with the value stored truncated to a short. If no truncation is 1876 // needed, the replacement is done in LoadNode::Identity(). 1877 // 1878 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1879 Node* mem = in(MemNode::Memory); 1880 Node* value = can_see_stored_value(mem,phase); 1881 if( value && !phase->type(value)->higher_equal( _type ) ) { 1882 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) ); 1883 return new (phase->C) RShiftINode(result, phase->intcon(16)); 1884 } 1885 // Identity call will handle the case where truncation is not needed. 1886 return LoadNode::Ideal(phase, can_reshape); 1887 } 1888 1889 const Type* LoadSNode::Value(PhaseTransform *phase) const { 1890 Node* mem = in(MemNode::Memory); 1891 Node* value = can_see_stored_value(mem,phase); 1892 if (value != NULL && value->is_Con() && 1893 !value->bottom_type()->higher_equal(_type)) { 1894 // If the input to the store does not fit with the load's result type, 1895 // it must be truncated. We can't delay until Ideal call since 1896 // a singleton Value is needed for split_thru_phi optimization. 1897 int con = value->get_int(); 1898 return TypeInt::make((con << 16) >> 16); 1899 } 1900 return LoadNode::Value(phase); 1901 } 1902 1903 //============================================================================= 1904 //----------------------------LoadKlassNode::make------------------------------ 1905 // Polymorphic factory method: 1906 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk) { 1907 Compile* C = gvn.C; 1908 // sanity check the alias category against the created node type 1909 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 1910 assert(adr_type != NULL, "expecting TypeKlassPtr"); 1911 #ifdef _LP64 1912 if (adr_type->is_ptr_to_narrowklass()) { 1913 assert(UseCompressedClassPointers, "no compressed klasses"); 1914 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 1915 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 1916 } 1917 #endif 1918 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1919 return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 1920 } 1921 1922 //------------------------------Value------------------------------------------ 1923 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1924 return klass_value_common(phase); 1925 } 1926 1927 // In most cases, LoadKlassNode does not have the control input set. If the control 1928 // input is set, it must not be removed (by LoadNode::Ideal()). 1929 bool LoadKlassNode::can_remove_control() const { 1930 return false; 1931 } 1932 1933 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1934 // Either input is TOP ==> the result is TOP 1935 const Type *t1 = phase->type( in(MemNode::Memory) ); 1936 if (t1 == Type::TOP) return Type::TOP; 1937 Node *adr = in(MemNode::Address); 1938 const Type *t2 = phase->type( adr ); 1939 if (t2 == Type::TOP) return Type::TOP; 1940 const TypePtr *tp = t2->is_ptr(); 1941 if (TypePtr::above_centerline(tp->ptr()) || 1942 tp->ptr() == TypePtr::Null) return Type::TOP; 1943 1944 // Return a more precise klass, if possible 1945 const TypeInstPtr *tinst = tp->isa_instptr(); 1946 if (tinst != NULL) { 1947 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1948 int offset = tinst->offset(); 1949 if (ik == phase->C->env()->Class_klass() 1950 && (offset == java_lang_Class::klass_offset_in_bytes() || 1951 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1952 // We are loading a special hidden field from a Class mirror object, 1953 // the field which points to the VM's Klass metaobject. 1954 ciType* t = tinst->java_mirror_type(); 1955 // java_mirror_type returns non-null for compile-time Class constants. 1956 if (t != NULL) { 1957 // constant oop => constant klass 1958 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1959 if (t->is_void()) { 1960 // We cannot create a void array. Since void is a primitive type return null 1961 // klass. Users of this result need to do a null check on the returned klass. 1962 return TypePtr::NULL_PTR; 1963 } 1964 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1965 } 1966 if (!t->is_klass()) { 1967 // a primitive Class (e.g., int.class) has NULL for a klass field 1968 return TypePtr::NULL_PTR; 1969 } 1970 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1971 return TypeKlassPtr::make(t->as_klass()); 1972 } 1973 // non-constant mirror, so we can't tell what's going on 1974 } 1975 if( !ik->is_loaded() ) 1976 return _type; // Bail out if not loaded 1977 if (offset == oopDesc::klass_offset_in_bytes()) { 1978 if (tinst->klass_is_exact()) { 1979 return TypeKlassPtr::make(ik); 1980 } 1981 // See if we can become precise: no subklasses and no interface 1982 // (Note: We need to support verified interfaces.) 1983 if (!ik->is_interface() && !ik->has_subklass()) { 1984 //assert(!UseExactTypes, "this code should be useless with exact types"); 1985 // Add a dependence; if any subclass added we need to recompile 1986 if (!ik->is_final()) { 1987 // %%% should use stronger assert_unique_concrete_subtype instead 1988 phase->C->dependencies()->assert_leaf_type(ik); 1989 } 1990 // Return precise klass 1991 return TypeKlassPtr::make(ik); 1992 } 1993 1994 // Return root of possible klass 1995 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1996 } 1997 } 1998 1999 // Check for loading klass from an array 2000 const TypeAryPtr *tary = tp->isa_aryptr(); 2001 if( tary != NULL ) { 2002 ciKlass *tary_klass = tary->klass(); 2003 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2004 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2005 if (tary->klass_is_exact()) { 2006 return TypeKlassPtr::make(tary_klass); 2007 } 2008 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2009 // If the klass is an object array, we defer the question to the 2010 // array component klass. 2011 if( ak->is_obj_array_klass() ) { 2012 assert( ak->is_loaded(), "" ); 2013 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2014 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2015 ciInstanceKlass* ik = base_k->as_instance_klass(); 2016 // See if we can become precise: no subklasses and no interface 2017 if (!ik->is_interface() && !ik->has_subklass()) { 2018 //assert(!UseExactTypes, "this code should be useless with exact types"); 2019 // Add a dependence; if any subclass added we need to recompile 2020 if (!ik->is_final()) { 2021 phase->C->dependencies()->assert_leaf_type(ik); 2022 } 2023 // Return precise array klass 2024 return TypeKlassPtr::make(ak); 2025 } 2026 } 2027 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2028 } else { // Found a type-array? 2029 //assert(!UseExactTypes, "this code should be useless with exact types"); 2030 assert( ak->is_type_array_klass(), "" ); 2031 return TypeKlassPtr::make(ak); // These are always precise 2032 } 2033 } 2034 } 2035 2036 // Check for loading klass from an array klass 2037 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2038 if (tkls != NULL && !StressReflectiveCode) { 2039 ciKlass* klass = tkls->klass(); 2040 if( !klass->is_loaded() ) 2041 return _type; // Bail out if not loaded 2042 if( klass->is_obj_array_klass() && 2043 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2044 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2045 // // Always returning precise element type is incorrect, 2046 // // e.g., element type could be object and array may contain strings 2047 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2048 2049 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2050 // according to the element type's subclassing. 2051 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2052 } 2053 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2054 tkls->offset() == in_bytes(Klass::super_offset())) { 2055 ciKlass* sup = klass->as_instance_klass()->super(); 2056 // The field is Klass::_super. Return its (constant) value. 2057 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2058 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2059 } 2060 } 2061 2062 // Bailout case 2063 return LoadNode::Value(phase); 2064 } 2065 2066 //------------------------------Identity--------------------------------------- 2067 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2068 // Also feed through the klass in Allocate(...klass...)._klass. 2069 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2070 return klass_identity_common(phase); 2071 } 2072 2073 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2074 Node* x = LoadNode::Identity(phase); 2075 if (x != this) return x; 2076 2077 // Take apart the address into an oop and and offset. 2078 // Return 'this' if we cannot. 2079 Node* adr = in(MemNode::Address); 2080 intptr_t offset = 0; 2081 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2082 if (base == NULL) return this; 2083 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2084 if (toop == NULL) return this; 2085 2086 // We can fetch the klass directly through an AllocateNode. 2087 // This works even if the klass is not constant (clone or newArray). 2088 if (offset == oopDesc::klass_offset_in_bytes()) { 2089 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2090 if (allocated_klass != NULL) { 2091 return allocated_klass; 2092 } 2093 } 2094 2095 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2096 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass. 2097 // See inline_native_Class_query for occurrences of these patterns. 2098 // Java Example: x.getClass().isAssignableFrom(y) 2099 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 2100 // 2101 // This improves reflective code, often making the Class 2102 // mirror go completely dead. (Current exception: Class 2103 // mirrors may appear in debug info, but we could clean them out by 2104 // introducing a new debug info operator for Klass*.java_mirror). 2105 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2106 && (offset == java_lang_Class::klass_offset_in_bytes() || 2107 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2108 // We are loading a special hidden field from a Class mirror, 2109 // the field which points to its Klass or ArrayKlass metaobject. 2110 if (base->is_Load()) { 2111 Node* adr2 = base->in(MemNode::Address); 2112 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2113 if (tkls != NULL && !tkls->empty() 2114 && (tkls->klass()->is_instance_klass() || 2115 tkls->klass()->is_array_klass()) 2116 && adr2->is_AddP() 2117 ) { 2118 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2119 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2120 mirror_field = in_bytes(ArrayKlass::component_mirror_offset()); 2121 } 2122 if (tkls->offset() == mirror_field) { 2123 return adr2->in(AddPNode::Base); 2124 } 2125 } 2126 } 2127 } 2128 2129 return this; 2130 } 2131 2132 2133 //------------------------------Value------------------------------------------ 2134 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2135 const Type *t = klass_value_common(phase); 2136 if (t == Type::TOP) 2137 return t; 2138 2139 return t->make_narrowklass(); 2140 } 2141 2142 //------------------------------Identity--------------------------------------- 2143 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2144 // Also feed through the klass in Allocate(...klass...)._klass. 2145 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2146 Node *x = klass_identity_common(phase); 2147 2148 const Type *t = phase->type( x ); 2149 if( t == Type::TOP ) return x; 2150 if( t->isa_narrowklass()) return x; 2151 assert (!t->isa_narrowoop(), "no narrow oop here"); 2152 2153 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass())); 2154 } 2155 2156 //------------------------------Value----------------------------------------- 2157 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2158 // Either input is TOP ==> the result is TOP 2159 const Type *t1 = phase->type( in(MemNode::Memory) ); 2160 if( t1 == Type::TOP ) return Type::TOP; 2161 Node *adr = in(MemNode::Address); 2162 const Type *t2 = phase->type( adr ); 2163 if( t2 == Type::TOP ) return Type::TOP; 2164 const TypePtr *tp = t2->is_ptr(); 2165 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2166 const TypeAryPtr *tap = tp->isa_aryptr(); 2167 if( !tap ) return _type; 2168 return tap->size(); 2169 } 2170 2171 //-------------------------------Ideal--------------------------------------- 2172 // Feed through the length in AllocateArray(...length...)._length. 2173 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2174 Node* p = MemNode::Ideal_common(phase, can_reshape); 2175 if (p) return (p == NodeSentinel) ? NULL : p; 2176 2177 // Take apart the address into an oop and and offset. 2178 // Return 'this' if we cannot. 2179 Node* adr = in(MemNode::Address); 2180 intptr_t offset = 0; 2181 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2182 if (base == NULL) return NULL; 2183 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2184 if (tary == NULL) return NULL; 2185 2186 // We can fetch the length directly through an AllocateArrayNode. 2187 // This works even if the length is not constant (clone or newArray). 2188 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2189 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2190 if (alloc != NULL) { 2191 Node* allocated_length = alloc->Ideal_length(); 2192 Node* len = alloc->make_ideal_length(tary, phase); 2193 if (allocated_length != len) { 2194 // New CastII improves on this. 2195 return len; 2196 } 2197 } 2198 } 2199 2200 return NULL; 2201 } 2202 2203 //------------------------------Identity--------------------------------------- 2204 // Feed through the length in AllocateArray(...length...)._length. 2205 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2206 Node* x = LoadINode::Identity(phase); 2207 if (x != this) return x; 2208 2209 // Take apart the address into an oop and and offset. 2210 // Return 'this' if we cannot. 2211 Node* adr = in(MemNode::Address); 2212 intptr_t offset = 0; 2213 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2214 if (base == NULL) return this; 2215 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2216 if (tary == NULL) return this; 2217 2218 // We can fetch the length directly through an AllocateArrayNode. 2219 // This works even if the length is not constant (clone or newArray). 2220 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2221 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2222 if (alloc != NULL) { 2223 Node* allocated_length = alloc->Ideal_length(); 2224 // Do not allow make_ideal_length to allocate a CastII node. 2225 Node* len = alloc->make_ideal_length(tary, phase, false); 2226 if (allocated_length == len) { 2227 // Return allocated_length only if it would not be improved by a CastII. 2228 return allocated_length; 2229 } 2230 } 2231 } 2232 2233 return this; 2234 2235 } 2236 2237 //============================================================================= 2238 //---------------------------StoreNode::make----------------------------------- 2239 // Polymorphic factory method: 2240 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2241 assert((mo == unordered || mo == release), "unexpected"); 2242 Compile* C = gvn.C; 2243 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2244 ctl != NULL, "raw memory operations should have control edge"); 2245 2246 switch (bt) { 2247 case T_BOOLEAN: val = gvn.transform(new (C) AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2248 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo); 2249 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo); 2250 case T_CHAR: 2251 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo); 2252 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo); 2253 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo); 2254 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo); 2255 case T_METADATA: 2256 case T_ADDRESS: 2257 case T_OBJECT: 2258 #ifdef _LP64 2259 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2260 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2261 return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo); 2262 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2263 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2264 adr->bottom_type()->isa_rawptr())) { 2265 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2266 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2267 } 2268 #endif 2269 { 2270 return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo); 2271 } 2272 } 2273 ShouldNotReachHere(); 2274 return (StoreNode*)NULL; 2275 } 2276 2277 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2278 bool require_atomic = true; 2279 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2280 } 2281 2282 StoreDNode* StoreDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2283 bool require_atomic = true; 2284 return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2285 } 2286 2287 2288 //--------------------------bottom_type---------------------------------------- 2289 const Type *StoreNode::bottom_type() const { 2290 return Type::MEMORY; 2291 } 2292 2293 //------------------------------hash------------------------------------------- 2294 uint StoreNode::hash() const { 2295 // unroll addition of interesting fields 2296 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2297 2298 // Since they are not commoned, do not hash them: 2299 return NO_HASH; 2300 } 2301 2302 //------------------------------Ideal------------------------------------------ 2303 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2304 // When a store immediately follows a relevant allocation/initialization, 2305 // try to capture it into the initialization, or hoist it above. 2306 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2307 Node* p = MemNode::Ideal_common(phase, can_reshape); 2308 if (p) return (p == NodeSentinel) ? NULL : p; 2309 2310 Node* mem = in(MemNode::Memory); 2311 Node* address = in(MemNode::Address); 2312 2313 // Back-to-back stores to same address? Fold em up. Generally 2314 // unsafe if I have intervening uses... Also disallowed for StoreCM 2315 // since they must follow each StoreP operation. Redundant StoreCMs 2316 // are eliminated just before matching in final_graph_reshape. 2317 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && 2318 mem->Opcode() != Op_StoreCM) { 2319 // Looking at a dead closed cycle of memory? 2320 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2321 2322 assert(Opcode() == mem->Opcode() || 2323 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2324 "no mismatched stores, except on raw memory"); 2325 2326 if (mem->outcnt() == 1 && // check for intervening uses 2327 mem->as_Store()->memory_size() <= this->memory_size()) { 2328 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2329 // For example, 'mem' might be the final state at a conditional return. 2330 // Or, 'mem' might be used by some node which is live at the same time 2331 // 'this' is live, which might be unschedulable. So, require exactly 2332 // ONE user, the 'this' store, until such time as we clone 'mem' for 2333 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2334 if (can_reshape) { // (%%% is this an anachronism?) 2335 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2336 phase->is_IterGVN()); 2337 } else { 2338 // It's OK to do this in the parser, since DU info is always accurate, 2339 // and the parser always refers to nodes via SafePointNode maps. 2340 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2341 } 2342 return this; 2343 } 2344 } 2345 2346 // Capture an unaliased, unconditional, simple store into an initializer. 2347 // Or, if it is independent of the allocation, hoist it above the allocation. 2348 if (ReduceFieldZeroing && /*can_reshape &&*/ 2349 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2350 InitializeNode* init = mem->in(0)->as_Initialize(); 2351 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2352 if (offset > 0) { 2353 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2354 // If the InitializeNode captured me, it made a raw copy of me, 2355 // and I need to disappear. 2356 if (moved != NULL) { 2357 // %%% hack to ensure that Ideal returns a new node: 2358 mem = MergeMemNode::make(phase->C, mem); 2359 return mem; // fold me away 2360 } 2361 } 2362 } 2363 2364 return NULL; // No further progress 2365 } 2366 2367 //------------------------------Value----------------------------------------- 2368 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2369 // Either input is TOP ==> the result is TOP 2370 const Type *t1 = phase->type( in(MemNode::Memory) ); 2371 if( t1 == Type::TOP ) return Type::TOP; 2372 const Type *t2 = phase->type( in(MemNode::Address) ); 2373 if( t2 == Type::TOP ) return Type::TOP; 2374 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2375 if( t3 == Type::TOP ) return Type::TOP; 2376 return Type::MEMORY; 2377 } 2378 2379 //------------------------------Identity--------------------------------------- 2380 // Remove redundant stores: 2381 // Store(m, p, Load(m, p)) changes to m. 2382 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2383 Node *StoreNode::Identity( PhaseTransform *phase ) { 2384 Node* mem = in(MemNode::Memory); 2385 Node* adr = in(MemNode::Address); 2386 Node* val = in(MemNode::ValueIn); 2387 2388 // Load then Store? Then the Store is useless 2389 if (val->is_Load() && 2390 val->in(MemNode::Address)->eqv_uncast(adr) && 2391 val->in(MemNode::Memory )->eqv_uncast(mem) && 2392 val->as_Load()->store_Opcode() == Opcode()) { 2393 return mem; 2394 } 2395 2396 // Two stores in a row of the same value? 2397 if (mem->is_Store() && 2398 mem->in(MemNode::Address)->eqv_uncast(adr) && 2399 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2400 mem->Opcode() == Opcode()) { 2401 return mem; 2402 } 2403 2404 // Store of zero anywhere into a freshly-allocated object? 2405 // Then the store is useless. 2406 // (It must already have been captured by the InitializeNode.) 2407 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2408 // a newly allocated object is already all-zeroes everywhere 2409 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2410 return mem; 2411 } 2412 2413 // the store may also apply to zero-bits in an earlier object 2414 Node* prev_mem = find_previous_store(phase); 2415 // Steps (a), (b): Walk past independent stores to find an exact match. 2416 if (prev_mem != NULL) { 2417 Node* prev_val = can_see_stored_value(prev_mem, phase); 2418 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2419 // prev_val and val might differ by a cast; it would be good 2420 // to keep the more informative of the two. 2421 return mem; 2422 } 2423 } 2424 } 2425 2426 return this; 2427 } 2428 2429 //------------------------------match_edge------------------------------------- 2430 // Do we Match on this edge index or not? Match only memory & value 2431 uint StoreNode::match_edge(uint idx) const { 2432 return idx == MemNode::Address || idx == MemNode::ValueIn; 2433 } 2434 2435 //------------------------------cmp-------------------------------------------- 2436 // Do not common stores up together. They generally have to be split 2437 // back up anyways, so do not bother. 2438 uint StoreNode::cmp( const Node &n ) const { 2439 return (&n == this); // Always fail except on self 2440 } 2441 2442 //------------------------------Ideal_masked_input----------------------------- 2443 // Check for a useless mask before a partial-word store 2444 // (StoreB ... (AndI valIn conIa) ) 2445 // If (conIa & mask == mask) this simplifies to 2446 // (StoreB ... (valIn) ) 2447 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2448 Node *val = in(MemNode::ValueIn); 2449 if( val->Opcode() == Op_AndI ) { 2450 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2451 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2452 set_req(MemNode::ValueIn, val->in(1)); 2453 return this; 2454 } 2455 } 2456 return NULL; 2457 } 2458 2459 2460 //------------------------------Ideal_sign_extended_input---------------------- 2461 // Check for useless sign-extension before a partial-word store 2462 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2463 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2464 // (StoreB ... (valIn) ) 2465 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2466 Node *val = in(MemNode::ValueIn); 2467 if( val->Opcode() == Op_RShiftI ) { 2468 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2469 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2470 Node *shl = val->in(1); 2471 if( shl->Opcode() == Op_LShiftI ) { 2472 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2473 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2474 set_req(MemNode::ValueIn, shl->in(1)); 2475 return this; 2476 } 2477 } 2478 } 2479 } 2480 return NULL; 2481 } 2482 2483 //------------------------------value_never_loaded----------------------------------- 2484 // Determine whether there are any possible loads of the value stored. 2485 // For simplicity, we actually check if there are any loads from the 2486 // address stored to, not just for loads of the value stored by this node. 2487 // 2488 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2489 Node *adr = in(Address); 2490 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2491 if (adr_oop == NULL) 2492 return false; 2493 if (!adr_oop->is_known_instance_field()) 2494 return false; // if not a distinct instance, there may be aliases of the address 2495 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2496 Node *use = adr->fast_out(i); 2497 int opc = use->Opcode(); 2498 if (use->is_Load() || use->is_LoadStore()) { 2499 return false; 2500 } 2501 } 2502 return true; 2503 } 2504 2505 //============================================================================= 2506 //------------------------------Ideal------------------------------------------ 2507 // If the store is from an AND mask that leaves the low bits untouched, then 2508 // we can skip the AND operation. If the store is from a sign-extension 2509 // (a left shift, then right shift) we can skip both. 2510 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2511 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2512 if( progress != NULL ) return progress; 2513 2514 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2515 if( progress != NULL ) return progress; 2516 2517 // Finally check the default case 2518 return StoreNode::Ideal(phase, can_reshape); 2519 } 2520 2521 //============================================================================= 2522 //------------------------------Ideal------------------------------------------ 2523 // If the store is from an AND mask that leaves the low bits untouched, then 2524 // we can skip the AND operation 2525 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2526 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2527 if( progress != NULL ) return progress; 2528 2529 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2530 if( progress != NULL ) return progress; 2531 2532 // Finally check the default case 2533 return StoreNode::Ideal(phase, can_reshape); 2534 } 2535 2536 //============================================================================= 2537 //------------------------------Identity--------------------------------------- 2538 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2539 // No need to card mark when storing a null ptr 2540 Node* my_store = in(MemNode::OopStore); 2541 if (my_store->is_Store()) { 2542 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2543 if( t1 == TypePtr::NULL_PTR ) { 2544 return in(MemNode::Memory); 2545 } 2546 } 2547 return this; 2548 } 2549 2550 //============================================================================= 2551 //------------------------------Ideal--------------------------------------- 2552 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2553 Node* progress = StoreNode::Ideal(phase, can_reshape); 2554 if (progress != NULL) return progress; 2555 2556 Node* my_store = in(MemNode::OopStore); 2557 if (my_store->is_MergeMem()) { 2558 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2559 set_req(MemNode::OopStore, mem); 2560 return this; 2561 } 2562 2563 return NULL; 2564 } 2565 2566 //------------------------------Value----------------------------------------- 2567 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2568 // Either input is TOP ==> the result is TOP 2569 const Type *t = phase->type( in(MemNode::Memory) ); 2570 if( t == Type::TOP ) return Type::TOP; 2571 t = phase->type( in(MemNode::Address) ); 2572 if( t == Type::TOP ) return Type::TOP; 2573 t = phase->type( in(MemNode::ValueIn) ); 2574 if( t == Type::TOP ) return Type::TOP; 2575 // If extra input is TOP ==> the result is TOP 2576 t = phase->type( in(MemNode::OopStore) ); 2577 if( t == Type::TOP ) return Type::TOP; 2578 2579 return StoreNode::Value( phase ); 2580 } 2581 2582 2583 //============================================================================= 2584 //----------------------------------SCMemProjNode------------------------------ 2585 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2586 { 2587 return bottom_type(); 2588 } 2589 2590 //============================================================================= 2591 //----------------------------------LoadStoreNode------------------------------ 2592 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2593 : Node(required), 2594 _type(rt), 2595 _adr_type(at) 2596 { 2597 init_req(MemNode::Control, c ); 2598 init_req(MemNode::Memory , mem); 2599 init_req(MemNode::Address, adr); 2600 init_req(MemNode::ValueIn, val); 2601 init_class_id(Class_LoadStore); 2602 } 2603 2604 uint LoadStoreNode::ideal_reg() const { 2605 return _type->ideal_reg(); 2606 } 2607 2608 bool LoadStoreNode::result_not_used() const { 2609 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2610 Node *x = fast_out(i); 2611 if (x->Opcode() == Op_SCMemProj) continue; 2612 return false; 2613 } 2614 return true; 2615 } 2616 2617 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2618 2619 //============================================================================= 2620 //----------------------------------LoadStoreConditionalNode-------------------- 2621 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2622 init_req(ExpectedIn, ex ); 2623 } 2624 2625 //============================================================================= 2626 //-------------------------------adr_type-------------------------------------- 2627 // Do we Match on this edge index or not? Do not match memory 2628 const TypePtr* ClearArrayNode::adr_type() const { 2629 Node *adr = in(3); 2630 return MemNode::calculate_adr_type(adr->bottom_type()); 2631 } 2632 2633 //------------------------------match_edge------------------------------------- 2634 // Do we Match on this edge index or not? Do not match memory 2635 uint ClearArrayNode::match_edge(uint idx) const { 2636 return idx > 1; 2637 } 2638 2639 //------------------------------Identity--------------------------------------- 2640 // Clearing a zero length array does nothing 2641 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2642 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2643 } 2644 2645 //------------------------------Idealize--------------------------------------- 2646 // Clearing a short array is faster with stores 2647 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2648 const int unit = BytesPerLong; 2649 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2650 if (!t) return NULL; 2651 if (!t->is_con()) return NULL; 2652 intptr_t raw_count = t->get_con(); 2653 intptr_t size = raw_count; 2654 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2655 // Clearing nothing uses the Identity call. 2656 // Negative clears are possible on dead ClearArrays 2657 // (see jck test stmt114.stmt11402.val). 2658 if (size <= 0 || size % unit != 0) return NULL; 2659 intptr_t count = size / unit; 2660 // Length too long; use fast hardware clear 2661 if (size > Matcher::init_array_short_size) return NULL; 2662 Node *mem = in(1); 2663 if( phase->type(mem)==Type::TOP ) return NULL; 2664 Node *adr = in(3); 2665 const Type* at = phase->type(adr); 2666 if( at==Type::TOP ) return NULL; 2667 const TypePtr* atp = at->isa_ptr(); 2668 // adjust atp to be the correct array element address type 2669 if (atp == NULL) atp = TypePtr::BOTTOM; 2670 else atp = atp->add_offset(Type::OffsetBot); 2671 // Get base for derived pointer purposes 2672 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2673 Node *base = adr->in(1); 2674 2675 Node *zero = phase->makecon(TypeLong::ZERO); 2676 Node *off = phase->MakeConX(BytesPerLong); 2677 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2678 count--; 2679 while( count-- ) { 2680 mem = phase->transform(mem); 2681 adr = phase->transform(new (phase->C) AddPNode(base,adr,off)); 2682 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2683 } 2684 return mem; 2685 } 2686 2687 //----------------------------step_through---------------------------------- 2688 // Return allocation input memory edge if it is different instance 2689 // or itself if it is the one we are looking for. 2690 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2691 Node* n = *np; 2692 assert(n->is_ClearArray(), "sanity"); 2693 intptr_t offset; 2694 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2695 // This method is called only before Allocate nodes are expanded during 2696 // macro nodes expansion. Before that ClearArray nodes are only generated 2697 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2698 assert(alloc != NULL, "should have allocation"); 2699 if (alloc->_idx == instance_id) { 2700 // Can not bypass initialization of the instance we are looking for. 2701 return false; 2702 } 2703 // Otherwise skip it. 2704 InitializeNode* init = alloc->initialization(); 2705 if (init != NULL) 2706 *np = init->in(TypeFunc::Memory); 2707 else 2708 *np = alloc->in(TypeFunc::Memory); 2709 return true; 2710 } 2711 2712 //----------------------------clear_memory------------------------------------- 2713 // Generate code to initialize object storage to zero. 2714 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2715 intptr_t start_offset, 2716 Node* end_offset, 2717 PhaseGVN* phase) { 2718 Compile* C = phase->C; 2719 intptr_t offset = start_offset; 2720 2721 int unit = BytesPerLong; 2722 if ((offset % unit) != 0) { 2723 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset)); 2724 adr = phase->transform(adr); 2725 const TypePtr* atp = TypeRawPtr::BOTTOM; 2726 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2727 mem = phase->transform(mem); 2728 offset += BytesPerInt; 2729 } 2730 assert((offset % unit) == 0, ""); 2731 2732 // Initialize the remaining stuff, if any, with a ClearArray. 2733 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2734 } 2735 2736 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2737 Node* start_offset, 2738 Node* end_offset, 2739 PhaseGVN* phase) { 2740 if (start_offset == end_offset) { 2741 // nothing to do 2742 return mem; 2743 } 2744 2745 Compile* C = phase->C; 2746 int unit = BytesPerLong; 2747 Node* zbase = start_offset; 2748 Node* zend = end_offset; 2749 2750 // Scale to the unit required by the CPU: 2751 if (!Matcher::init_array_count_is_in_bytes) { 2752 Node* shift = phase->intcon(exact_log2(unit)); 2753 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) ); 2754 zend = phase->transform( new(C) URShiftXNode(zend, shift) ); 2755 } 2756 2757 // Bulk clear double-words 2758 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) ); 2759 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) ); 2760 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr); 2761 return phase->transform(mem); 2762 } 2763 2764 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2765 intptr_t start_offset, 2766 intptr_t end_offset, 2767 PhaseGVN* phase) { 2768 if (start_offset == end_offset) { 2769 // nothing to do 2770 return mem; 2771 } 2772 2773 Compile* C = phase->C; 2774 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2775 intptr_t done_offset = end_offset; 2776 if ((done_offset % BytesPerLong) != 0) { 2777 done_offset -= BytesPerInt; 2778 } 2779 if (done_offset > start_offset) { 2780 mem = clear_memory(ctl, mem, dest, 2781 start_offset, phase->MakeConX(done_offset), phase); 2782 } 2783 if (done_offset < end_offset) { // emit the final 32-bit store 2784 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2785 adr = phase->transform(adr); 2786 const TypePtr* atp = TypeRawPtr::BOTTOM; 2787 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2788 mem = phase->transform(mem); 2789 done_offset += BytesPerInt; 2790 } 2791 assert(done_offset == end_offset, ""); 2792 return mem; 2793 } 2794 2795 //============================================================================= 2796 // Do not match memory edge. 2797 uint StrIntrinsicNode::match_edge(uint idx) const { 2798 return idx == 2 || idx == 3; 2799 } 2800 2801 //------------------------------Ideal------------------------------------------ 2802 // Return a node which is more "ideal" than the current node. Strip out 2803 // control copies 2804 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2805 if (remove_dead_region(phase, can_reshape)) return this; 2806 // Don't bother trying to transform a dead node 2807 if (in(0) && in(0)->is_top()) return NULL; 2808 2809 if (can_reshape) { 2810 Node* mem = phase->transform(in(MemNode::Memory)); 2811 // If transformed to a MergeMem, get the desired slice 2812 uint alias_idx = phase->C->get_alias_index(adr_type()); 2813 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; 2814 if (mem != in(MemNode::Memory)) { 2815 set_req(MemNode::Memory, mem); 2816 return this; 2817 } 2818 } 2819 return NULL; 2820 } 2821 2822 //------------------------------Value------------------------------------------ 2823 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { 2824 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; 2825 return bottom_type(); 2826 } 2827 2828 //============================================================================= 2829 //------------------------------match_edge------------------------------------- 2830 // Do not match memory edge 2831 uint EncodeISOArrayNode::match_edge(uint idx) const { 2832 return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len) 2833 } 2834 2835 //------------------------------Ideal------------------------------------------ 2836 // Return a node which is more "ideal" than the current node. Strip out 2837 // control copies 2838 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2839 return remove_dead_region(phase, can_reshape) ? this : NULL; 2840 } 2841 2842 //------------------------------Value------------------------------------------ 2843 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const { 2844 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; 2845 return bottom_type(); 2846 } 2847 2848 //============================================================================= 2849 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2850 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2851 _adr_type(C->get_adr_type(alias_idx)) 2852 { 2853 init_class_id(Class_MemBar); 2854 Node* top = C->top(); 2855 init_req(TypeFunc::I_O,top); 2856 init_req(TypeFunc::FramePtr,top); 2857 init_req(TypeFunc::ReturnAdr,top); 2858 if (precedent != NULL) 2859 init_req(TypeFunc::Parms, precedent); 2860 } 2861 2862 //------------------------------cmp-------------------------------------------- 2863 uint MemBarNode::hash() const { return NO_HASH; } 2864 uint MemBarNode::cmp( const Node &n ) const { 2865 return (&n == this); // Always fail except on self 2866 } 2867 2868 //------------------------------make------------------------------------------- 2869 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2870 switch (opcode) { 2871 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn); 2872 case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn); 2873 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn); 2874 case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn); 2875 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn); 2876 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn); 2877 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn); 2878 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn); 2879 case Op_Initialize: return new(C) InitializeNode(C, atp, pn); 2880 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn); 2881 default: ShouldNotReachHere(); return NULL; 2882 } 2883 } 2884 2885 //------------------------------Ideal------------------------------------------ 2886 // Return a node which is more "ideal" than the current node. Strip out 2887 // control copies 2888 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2889 if (remove_dead_region(phase, can_reshape)) return this; 2890 // Don't bother trying to transform a dead node 2891 if (in(0) && in(0)->is_top()) { 2892 return NULL; 2893 } 2894 2895 // Eliminate volatile MemBars for scalar replaced objects. 2896 if (can_reshape && req() == (Precedent+1)) { 2897 bool eliminate = false; 2898 int opc = Opcode(); 2899 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2900 // Volatile field loads and stores. 2901 Node* my_mem = in(MemBarNode::Precedent); 2902 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2903 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2904 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2905 // replace this Precedent (decodeN) with the Load instead. 2906 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2907 Node* load_node = my_mem->in(1); 2908 set_req(MemBarNode::Precedent, load_node); 2909 phase->is_IterGVN()->_worklist.push(my_mem); 2910 my_mem = load_node; 2911 } else { 2912 assert(my_mem->unique_out() == this, "sanity"); 2913 del_req(Precedent); 2914 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2915 my_mem = NULL; 2916 } 2917 } 2918 if (my_mem != NULL && my_mem->is_Mem()) { 2919 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2920 // Check for scalar replaced object reference. 2921 if( t_oop != NULL && t_oop->is_known_instance_field() && 2922 t_oop->offset() != Type::OffsetBot && 2923 t_oop->offset() != Type::OffsetTop) { 2924 eliminate = true; 2925 } 2926 } 2927 } else if (opc == Op_MemBarRelease) { 2928 // Final field stores. 2929 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 2930 if ((alloc != NULL) && alloc->is_Allocate() && 2931 AARCH64_ONLY ( alloc->as_Allocate()->does_not_escape_thread() ) 2932 NOT_AARCH64 ( alloc->as_Allocate()->_is_non_escaping ) 2933 ) { 2934 // The allocated object does not escape. 2935 eliminate = true; 2936 } 2937 } 2938 if (eliminate) { 2939 // Replace MemBar projections by its inputs. 2940 PhaseIterGVN* igvn = phase->is_IterGVN(); 2941 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2942 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2943 // Must return either the original node (now dead) or a new node 2944 // (Do not return a top here, since that would break the uniqueness of top.) 2945 return new (phase->C) ConINode(TypeInt::ZERO); 2946 } 2947 } 2948 return NULL; 2949 } 2950 2951 //------------------------------Value------------------------------------------ 2952 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2953 if( !in(0) ) return Type::TOP; 2954 if( phase->type(in(0)) == Type::TOP ) 2955 return Type::TOP; 2956 return TypeTuple::MEMBAR; 2957 } 2958 2959 //------------------------------match------------------------------------------ 2960 // Construct projections for memory. 2961 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2962 switch (proj->_con) { 2963 case TypeFunc::Control: 2964 case TypeFunc::Memory: 2965 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2966 } 2967 ShouldNotReachHere(); 2968 return NULL; 2969 } 2970 2971 //===========================InitializeNode==================================== 2972 // SUMMARY: 2973 // This node acts as a memory barrier on raw memory, after some raw stores. 2974 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2975 // The Initialize can 'capture' suitably constrained stores as raw inits. 2976 // It can coalesce related raw stores into larger units (called 'tiles'). 2977 // It can avoid zeroing new storage for memory units which have raw inits. 2978 // At macro-expansion, it is marked 'complete', and does not optimize further. 2979 // 2980 // EXAMPLE: 2981 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2982 // ctl = incoming control; mem* = incoming memory 2983 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2984 // First allocate uninitialized memory and fill in the header: 2985 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2986 // ctl := alloc.Control; mem* := alloc.Memory* 2987 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2988 // Then initialize to zero the non-header parts of the raw memory block: 2989 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2990 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2991 // After the initialize node executes, the object is ready for service: 2992 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2993 // Suppose its body is immediately initialized as {1,2}: 2994 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2995 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 2996 // mem.SLICE(#short[*]) := store2 2997 // 2998 // DETAILS: 2999 // An InitializeNode collects and isolates object initialization after 3000 // an AllocateNode and before the next possible safepoint. As a 3001 // memory barrier (MemBarNode), it keeps critical stores from drifting 3002 // down past any safepoint or any publication of the allocation. 3003 // Before this barrier, a newly-allocated object may have uninitialized bits. 3004 // After this barrier, it may be treated as a real oop, and GC is allowed. 3005 // 3006 // The semantics of the InitializeNode include an implicit zeroing of 3007 // the new object from object header to the end of the object. 3008 // (The object header and end are determined by the AllocateNode.) 3009 // 3010 // Certain stores may be added as direct inputs to the InitializeNode. 3011 // These stores must update raw memory, and they must be to addresses 3012 // derived from the raw address produced by AllocateNode, and with 3013 // a constant offset. They must be ordered by increasing offset. 3014 // The first one is at in(RawStores), the last at in(req()-1). 3015 // Unlike most memory operations, they are not linked in a chain, 3016 // but are displayed in parallel as users of the rawmem output of 3017 // the allocation. 3018 // 3019 // (See comments in InitializeNode::capture_store, which continue 3020 // the example given above.) 3021 // 3022 // When the associated Allocate is macro-expanded, the InitializeNode 3023 // may be rewritten to optimize collected stores. A ClearArrayNode 3024 // may also be created at that point to represent any required zeroing. 3025 // The InitializeNode is then marked 'complete', prohibiting further 3026 // capturing of nearby memory operations. 3027 // 3028 // During macro-expansion, all captured initializations which store 3029 // constant values of 32 bits or smaller are coalesced (if advantageous) 3030 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3031 // initialized in fewer memory operations. Memory words which are 3032 // covered by neither tiles nor non-constant stores are pre-zeroed 3033 // by explicit stores of zero. (The code shape happens to do all 3034 // zeroing first, then all other stores, with both sequences occurring 3035 // in order of ascending offsets.) 3036 // 3037 // Alternatively, code may be inserted between an AllocateNode and its 3038 // InitializeNode, to perform arbitrary initialization of the new object. 3039 // E.g., the object copying intrinsics insert complex data transfers here. 3040 // The initialization must then be marked as 'complete' disable the 3041 // built-in zeroing semantics and the collection of initializing stores. 3042 // 3043 // While an InitializeNode is incomplete, reads from the memory state 3044 // produced by it are optimizable if they match the control edge and 3045 // new oop address associated with the allocation/initialization. 3046 // They return a stored value (if the offset matches) or else zero. 3047 // A write to the memory state, if it matches control and address, 3048 // and if it is to a constant offset, may be 'captured' by the 3049 // InitializeNode. It is cloned as a raw memory operation and rewired 3050 // inside the initialization, to the raw oop produced by the allocation. 3051 // Operations on addresses which are provably distinct (e.g., to 3052 // other AllocateNodes) are allowed to bypass the initialization. 3053 // 3054 // The effect of all this is to consolidate object initialization 3055 // (both arrays and non-arrays, both piecewise and bulk) into a 3056 // single location, where it can be optimized as a unit. 3057 // 3058 // Only stores with an offset less than TrackedInitializationLimit words 3059 // will be considered for capture by an InitializeNode. This puts a 3060 // reasonable limit on the complexity of optimized initializations. 3061 3062 //---------------------------InitializeNode------------------------------------ 3063 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3064 : _is_complete(Incomplete), _does_not_escape(false), 3065 MemBarNode(C, adr_type, rawoop) 3066 { 3067 init_class_id(Class_Initialize); 3068 3069 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3070 assert(in(RawAddress) == rawoop, "proper init"); 3071 // Note: allocation() can be NULL, for secondary initialization barriers 3072 } 3073 3074 // Since this node is not matched, it will be processed by the 3075 // register allocator. Declare that there are no constraints 3076 // on the allocation of the RawAddress edge. 3077 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3078 // This edge should be set to top, by the set_complete. But be conservative. 3079 if (idx == InitializeNode::RawAddress) 3080 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3081 return RegMask::Empty; 3082 } 3083 3084 Node* InitializeNode::memory(uint alias_idx) { 3085 Node* mem = in(Memory); 3086 if (mem->is_MergeMem()) { 3087 return mem->as_MergeMem()->memory_at(alias_idx); 3088 } else { 3089 // incoming raw memory is not split 3090 return mem; 3091 } 3092 } 3093 3094 bool InitializeNode::is_non_zero() { 3095 if (is_complete()) return false; 3096 remove_extra_zeroes(); 3097 return (req() > RawStores); 3098 } 3099 3100 void InitializeNode::set_complete(PhaseGVN* phase) { 3101 assert(!is_complete(), "caller responsibility"); 3102 _is_complete = Complete; 3103 3104 // After this node is complete, it contains a bunch of 3105 // raw-memory initializations. There is no need for 3106 // it to have anything to do with non-raw memory effects. 3107 // Therefore, tell all non-raw users to re-optimize themselves, 3108 // after skipping the memory effects of this initialization. 3109 PhaseIterGVN* igvn = phase->is_IterGVN(); 3110 if (igvn) igvn->add_users_to_worklist(this); 3111 } 3112 3113 // convenience function 3114 // return false if the init contains any stores already 3115 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3116 InitializeNode* init = initialization(); 3117 if (init == NULL || init->is_complete()) return false; 3118 init->remove_extra_zeroes(); 3119 // for now, if this allocation has already collected any inits, bail: 3120 if (init->is_non_zero()) return false; 3121 init->set_complete(phase); 3122 return true; 3123 } 3124 3125 void InitializeNode::remove_extra_zeroes() { 3126 if (req() == RawStores) return; 3127 Node* zmem = zero_memory(); 3128 uint fill = RawStores; 3129 for (uint i = fill; i < req(); i++) { 3130 Node* n = in(i); 3131 if (n->is_top() || n == zmem) continue; // skip 3132 if (fill < i) set_req(fill, n); // compact 3133 ++fill; 3134 } 3135 // delete any empty spaces created: 3136 while (fill < req()) { 3137 del_req(fill); 3138 } 3139 } 3140 3141 // Helper for remembering which stores go with which offsets. 3142 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3143 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3144 intptr_t offset = -1; 3145 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3146 phase, offset); 3147 if (base == NULL) return -1; // something is dead, 3148 if (offset < 0) return -1; // dead, dead 3149 return offset; 3150 } 3151 3152 // Helper for proving that an initialization expression is 3153 // "simple enough" to be folded into an object initialization. 3154 // Attempts to prove that a store's initial value 'n' can be captured 3155 // within the initialization without creating a vicious cycle, such as: 3156 // { Foo p = new Foo(); p.next = p; } 3157 // True for constants and parameters and small combinations thereof. 3158 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3159 if (n == NULL) return true; // (can this really happen?) 3160 if (n->is_Proj()) n = n->in(0); 3161 if (n == this) return false; // found a cycle 3162 if (n->is_Con()) return true; 3163 if (n->is_Start()) return true; // params, etc., are OK 3164 if (n->is_Root()) return true; // even better 3165 3166 Node* ctl = n->in(0); 3167 if (ctl != NULL && !ctl->is_top()) { 3168 if (ctl->is_Proj()) ctl = ctl->in(0); 3169 if (ctl == this) return false; 3170 3171 // If we already know that the enclosing memory op is pinned right after 3172 // the init, then any control flow that the store has picked up 3173 // must have preceded the init, or else be equal to the init. 3174 // Even after loop optimizations (which might change control edges) 3175 // a store is never pinned *before* the availability of its inputs. 3176 if (!MemNode::all_controls_dominate(n, this)) 3177 return false; // failed to prove a good control 3178 } 3179 3180 // Check data edges for possible dependencies on 'this'. 3181 if ((count += 1) > 20) return false; // complexity limit 3182 for (uint i = 1; i < n->req(); i++) { 3183 Node* m = n->in(i); 3184 if (m == NULL || m == n || m->is_top()) continue; 3185 uint first_i = n->find_edge(m); 3186 if (i != first_i) continue; // process duplicate edge just once 3187 if (!detect_init_independence(m, count)) { 3188 return false; 3189 } 3190 } 3191 3192 return true; 3193 } 3194 3195 // Here are all the checks a Store must pass before it can be moved into 3196 // an initialization. Returns zero if a check fails. 3197 // On success, returns the (constant) offset to which the store applies, 3198 // within the initialized memory. 3199 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3200 const int FAIL = 0; 3201 if (st->is_unaligned_access()) { 3202 return FAIL; 3203 } 3204 if (st->req() != MemNode::ValueIn + 1) 3205 return FAIL; // an inscrutable StoreNode (card mark?) 3206 Node* ctl = st->in(MemNode::Control); 3207 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3208 return FAIL; // must be unconditional after the initialization 3209 Node* mem = st->in(MemNode::Memory); 3210 if (!(mem->is_Proj() && mem->in(0) == this)) 3211 return FAIL; // must not be preceded by other stores 3212 Node* adr = st->in(MemNode::Address); 3213 intptr_t offset; 3214 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3215 if (alloc == NULL) 3216 return FAIL; // inscrutable address 3217 if (alloc != allocation()) 3218 return FAIL; // wrong allocation! (store needs to float up) 3219 Node* val = st->in(MemNode::ValueIn); 3220 int complexity_count = 0; 3221 if (!detect_init_independence(val, complexity_count)) 3222 return FAIL; // stored value must be 'simple enough' 3223 3224 // The Store can be captured only if nothing after the allocation 3225 // and before the Store is using the memory location that the store 3226 // overwrites. 3227 bool failed = false; 3228 // If is_complete_with_arraycopy() is true the shape of the graph is 3229 // well defined and is safe so no need for extra checks. 3230 if (!is_complete_with_arraycopy()) { 3231 // We are going to look at each use of the memory state following 3232 // the allocation to make sure nothing reads the memory that the 3233 // Store writes. 3234 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3235 int alias_idx = phase->C->get_alias_index(t_adr); 3236 ResourceMark rm; 3237 Unique_Node_List mems; 3238 mems.push(mem); 3239 Node* unique_merge = NULL; 3240 for (uint next = 0; next < mems.size(); ++next) { 3241 Node *m = mems.at(next); 3242 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3243 Node *n = m->fast_out(j); 3244 if (n->outcnt() == 0) { 3245 continue; 3246 } 3247 if (n == st) { 3248 continue; 3249 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3250 // If the control of this use is different from the control 3251 // of the Store which is right after the InitializeNode then 3252 // this node cannot be between the InitializeNode and the 3253 // Store. 3254 continue; 3255 } else if (n->is_MergeMem()) { 3256 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3257 // We can hit a MergeMemNode (that will likely go away 3258 // later) that is a direct use of the memory state 3259 // following the InitializeNode on the same slice as the 3260 // store node that we'd like to capture. We need to check 3261 // the uses of the MergeMemNode. 3262 mems.push(n); 3263 } 3264 } else if (n->is_Mem()) { 3265 Node* other_adr = n->in(MemNode::Address); 3266 if (other_adr == adr) { 3267 failed = true; 3268 break; 3269 } else { 3270 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3271 if (other_t_adr != NULL) { 3272 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3273 if (other_alias_idx == alias_idx) { 3274 // A load from the same memory slice as the store right 3275 // after the InitializeNode. We check the control of the 3276 // object/array that is loaded from. If it's the same as 3277 // the store control then we cannot capture the store. 3278 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3279 Node* base = other_adr; 3280 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); 3281 base = base->in(AddPNode::Base); 3282 if (base != NULL) { 3283 base = base->uncast(); 3284 if (base->is_Proj() && base->in(0) == alloc) { 3285 failed = true; 3286 break; 3287 } 3288 } 3289 } 3290 } 3291 } 3292 } else { 3293 failed = true; 3294 break; 3295 } 3296 } 3297 } 3298 } 3299 if (failed) { 3300 if (!can_reshape) { 3301 // We decided we couldn't capture the store during parsing. We 3302 // should try again during the next IGVN once the graph is 3303 // cleaner. 3304 phase->C->record_for_igvn(st); 3305 } 3306 return FAIL; 3307 } 3308 3309 return offset; // success 3310 } 3311 3312 // Find the captured store in(i) which corresponds to the range 3313 // [start..start+size) in the initialized object. 3314 // If there is one, return its index i. If there isn't, return the 3315 // negative of the index where it should be inserted. 3316 // Return 0 if the queried range overlaps an initialization boundary 3317 // or if dead code is encountered. 3318 // If size_in_bytes is zero, do not bother with overlap checks. 3319 int InitializeNode::captured_store_insertion_point(intptr_t start, 3320 int size_in_bytes, 3321 PhaseTransform* phase) { 3322 const int FAIL = 0, MAX_STORE = BytesPerLong; 3323 3324 if (is_complete()) 3325 return FAIL; // arraycopy got here first; punt 3326 3327 assert(allocation() != NULL, "must be present"); 3328 3329 // no negatives, no header fields: 3330 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3331 3332 // after a certain size, we bail out on tracking all the stores: 3333 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3334 if (start >= ti_limit) return FAIL; 3335 3336 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3337 if (i >= limit) return -(int)i; // not found; here is where to put it 3338 3339 Node* st = in(i); 3340 intptr_t st_off = get_store_offset(st, phase); 3341 if (st_off < 0) { 3342 if (st != zero_memory()) { 3343 return FAIL; // bail out if there is dead garbage 3344 } 3345 } else if (st_off > start) { 3346 // ...we are done, since stores are ordered 3347 if (st_off < start + size_in_bytes) { 3348 return FAIL; // the next store overlaps 3349 } 3350 return -(int)i; // not found; here is where to put it 3351 } else if (st_off < start) { 3352 if (size_in_bytes != 0 && 3353 start < st_off + MAX_STORE && 3354 start < st_off + st->as_Store()->memory_size()) { 3355 return FAIL; // the previous store overlaps 3356 } 3357 } else { 3358 if (size_in_bytes != 0 && 3359 st->as_Store()->memory_size() != size_in_bytes) { 3360 return FAIL; // mismatched store size 3361 } 3362 return i; 3363 } 3364 3365 ++i; 3366 } 3367 } 3368 3369 // Look for a captured store which initializes at the offset 'start' 3370 // with the given size. If there is no such store, and no other 3371 // initialization interferes, then return zero_memory (the memory 3372 // projection of the AllocateNode). 3373 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3374 PhaseTransform* phase) { 3375 assert(stores_are_sane(phase), ""); 3376 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3377 if (i == 0) { 3378 return NULL; // something is dead 3379 } else if (i < 0) { 3380 return zero_memory(); // just primordial zero bits here 3381 } else { 3382 Node* st = in(i); // here is the store at this position 3383 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3384 return st; 3385 } 3386 } 3387 3388 // Create, as a raw pointer, an address within my new object at 'offset'. 3389 Node* InitializeNode::make_raw_address(intptr_t offset, 3390 PhaseTransform* phase) { 3391 Node* addr = in(RawAddress); 3392 if (offset != 0) { 3393 Compile* C = phase->C; 3394 addr = phase->transform( new (C) AddPNode(C->top(), addr, 3395 phase->MakeConX(offset)) ); 3396 } 3397 return addr; 3398 } 3399 3400 // Clone the given store, converting it into a raw store 3401 // initializing a field or element of my new object. 3402 // Caller is responsible for retiring the original store, 3403 // with subsume_node or the like. 3404 // 3405 // From the example above InitializeNode::InitializeNode, 3406 // here are the old stores to be captured: 3407 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3408 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3409 // 3410 // Here is the changed code; note the extra edges on init: 3411 // alloc = (Allocate ...) 3412 // rawoop = alloc.RawAddress 3413 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3414 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3415 // init = (Initialize alloc.Control alloc.Memory rawoop 3416 // rawstore1 rawstore2) 3417 // 3418 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3419 PhaseTransform* phase, bool can_reshape) { 3420 assert(stores_are_sane(phase), ""); 3421 3422 if (start < 0) return NULL; 3423 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3424 3425 Compile* C = phase->C; 3426 int size_in_bytes = st->memory_size(); 3427 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3428 if (i == 0) return NULL; // bail out 3429 Node* prev_mem = NULL; // raw memory for the captured store 3430 if (i > 0) { 3431 prev_mem = in(i); // there is a pre-existing store under this one 3432 set_req(i, C->top()); // temporarily disconnect it 3433 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3434 } else { 3435 i = -i; // no pre-existing store 3436 prev_mem = zero_memory(); // a slice of the newly allocated object 3437 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3438 set_req(--i, C->top()); // reuse this edge; it has been folded away 3439 else 3440 ins_req(i, C->top()); // build a new edge 3441 } 3442 Node* new_st = st->clone(); 3443 new_st->set_req(MemNode::Control, in(Control)); 3444 new_st->set_req(MemNode::Memory, prev_mem); 3445 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3446 new_st = phase->transform(new_st); 3447 3448 // At this point, new_st might have swallowed a pre-existing store 3449 // at the same offset, or perhaps new_st might have disappeared, 3450 // if it redundantly stored the same value (or zero to fresh memory). 3451 3452 // In any case, wire it in: 3453 set_req(i, new_st); 3454 3455 // The caller may now kill the old guy. 3456 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3457 assert(check_st == new_st || check_st == NULL, "must be findable"); 3458 assert(!is_complete(), ""); 3459 return new_st; 3460 } 3461 3462 static bool store_constant(jlong* tiles, int num_tiles, 3463 intptr_t st_off, int st_size, 3464 jlong con) { 3465 if ((st_off & (st_size-1)) != 0) 3466 return false; // strange store offset (assume size==2**N) 3467 address addr = (address)tiles + st_off; 3468 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3469 switch (st_size) { 3470 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3471 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3472 case sizeof(jint): *(jint*) addr = (jint) con; break; 3473 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3474 default: return false; // strange store size (detect size!=2**N here) 3475 } 3476 return true; // return success to caller 3477 } 3478 3479 // Coalesce subword constants into int constants and possibly 3480 // into long constants. The goal, if the CPU permits, 3481 // is to initialize the object with a small number of 64-bit tiles. 3482 // Also, convert floating-point constants to bit patterns. 3483 // Non-constants are not relevant to this pass. 3484 // 3485 // In terms of the running example on InitializeNode::InitializeNode 3486 // and InitializeNode::capture_store, here is the transformation 3487 // of rawstore1 and rawstore2 into rawstore12: 3488 // alloc = (Allocate ...) 3489 // rawoop = alloc.RawAddress 3490 // tile12 = 0x00010002 3491 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3492 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3493 // 3494 void 3495 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3496 Node* size_in_bytes, 3497 PhaseGVN* phase) { 3498 Compile* C = phase->C; 3499 3500 assert(stores_are_sane(phase), ""); 3501 // Note: After this pass, they are not completely sane, 3502 // since there may be some overlaps. 3503 3504 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3505 3506 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3507 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3508 size_limit = MIN2(size_limit, ti_limit); 3509 size_limit = align_size_up(size_limit, BytesPerLong); 3510 int num_tiles = size_limit / BytesPerLong; 3511 3512 // allocate space for the tile map: 3513 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3514 jlong tiles_buf[small_len]; 3515 Node* nodes_buf[small_len]; 3516 jlong inits_buf[small_len]; 3517 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3518 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3519 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3520 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3521 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3522 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3523 // tiles: exact bitwise model of all primitive constants 3524 // nodes: last constant-storing node subsumed into the tiles model 3525 // inits: which bytes (in each tile) are touched by any initializations 3526 3527 //// Pass A: Fill in the tile model with any relevant stores. 3528 3529 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3530 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3531 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3532 Node* zmem = zero_memory(); // initially zero memory state 3533 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3534 Node* st = in(i); 3535 intptr_t st_off = get_store_offset(st, phase); 3536 3537 // Figure out the store's offset and constant value: 3538 if (st_off < header_size) continue; //skip (ignore header) 3539 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3540 int st_size = st->as_Store()->memory_size(); 3541 if (st_off + st_size > size_limit) break; 3542 3543 // Record which bytes are touched, whether by constant or not. 3544 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3545 continue; // skip (strange store size) 3546 3547 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3548 if (!val->singleton()) continue; //skip (non-con store) 3549 BasicType type = val->basic_type(); 3550 3551 jlong con = 0; 3552 switch (type) { 3553 case T_INT: con = val->is_int()->get_con(); break; 3554 case T_LONG: con = val->is_long()->get_con(); break; 3555 case T_FLOAT: con = jint_cast(val->getf()); break; 3556 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3557 default: continue; //skip (odd store type) 3558 } 3559 3560 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3561 st->Opcode() == Op_StoreL) { 3562 continue; // This StoreL is already optimal. 3563 } 3564 3565 // Store down the constant. 3566 store_constant(tiles, num_tiles, st_off, st_size, con); 3567 3568 intptr_t j = st_off >> LogBytesPerLong; 3569 3570 if (type == T_INT && st_size == BytesPerInt 3571 && (st_off & BytesPerInt) == BytesPerInt) { 3572 jlong lcon = tiles[j]; 3573 if (!Matcher::isSimpleConstant64(lcon) && 3574 st->Opcode() == Op_StoreI) { 3575 // This StoreI is already optimal by itself. 3576 jint* intcon = (jint*) &tiles[j]; 3577 intcon[1] = 0; // undo the store_constant() 3578 3579 // If the previous store is also optimal by itself, back up and 3580 // undo the action of the previous loop iteration... if we can. 3581 // But if we can't, just let the previous half take care of itself. 3582 st = nodes[j]; 3583 st_off -= BytesPerInt; 3584 con = intcon[0]; 3585 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3586 assert(st_off >= header_size, "still ignoring header"); 3587 assert(get_store_offset(st, phase) == st_off, "must be"); 3588 assert(in(i-1) == zmem, "must be"); 3589 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3590 assert(con == tcon->is_int()->get_con(), "must be"); 3591 // Undo the effects of the previous loop trip, which swallowed st: 3592 intcon[0] = 0; // undo store_constant() 3593 set_req(i-1, st); // undo set_req(i, zmem) 3594 nodes[j] = NULL; // undo nodes[j] = st 3595 --old_subword; // undo ++old_subword 3596 } 3597 continue; // This StoreI is already optimal. 3598 } 3599 } 3600 3601 // This store is not needed. 3602 set_req(i, zmem); 3603 nodes[j] = st; // record for the moment 3604 if (st_size < BytesPerLong) // something has changed 3605 ++old_subword; // includes int/float, but who's counting... 3606 else ++old_long; 3607 } 3608 3609 if ((old_subword + old_long) == 0) 3610 return; // nothing more to do 3611 3612 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3613 // Be sure to insert them before overlapping non-constant stores. 3614 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3615 for (int j = 0; j < num_tiles; j++) { 3616 jlong con = tiles[j]; 3617 jlong init = inits[j]; 3618 if (con == 0) continue; 3619 jint con0, con1; // split the constant, address-wise 3620 jint init0, init1; // split the init map, address-wise 3621 { union { jlong con; jint intcon[2]; } u; 3622 u.con = con; 3623 con0 = u.intcon[0]; 3624 con1 = u.intcon[1]; 3625 u.con = init; 3626 init0 = u.intcon[0]; 3627 init1 = u.intcon[1]; 3628 } 3629 3630 Node* old = nodes[j]; 3631 assert(old != NULL, "need the prior store"); 3632 intptr_t offset = (j * BytesPerLong); 3633 3634 bool split = !Matcher::isSimpleConstant64(con); 3635 3636 if (offset < header_size) { 3637 assert(offset + BytesPerInt >= header_size, "second int counts"); 3638 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3639 split = true; // only the second word counts 3640 // Example: int a[] = { 42 ... } 3641 } else if (con0 == 0 && init0 == -1) { 3642 split = true; // first word is covered by full inits 3643 // Example: int a[] = { ... foo(), 42 ... } 3644 } else if (con1 == 0 && init1 == -1) { 3645 split = true; // second word is covered by full inits 3646 // Example: int a[] = { ... 42, foo() ... } 3647 } 3648 3649 // Here's a case where init0 is neither 0 nor -1: 3650 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3651 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3652 // In this case the tile is not split; it is (jlong)42. 3653 // The big tile is stored down, and then the foo() value is inserted. 3654 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3655 3656 Node* ctl = old->in(MemNode::Control); 3657 Node* adr = make_raw_address(offset, phase); 3658 const TypePtr* atp = TypeRawPtr::BOTTOM; 3659 3660 // One or two coalesced stores to plop down. 3661 Node* st[2]; 3662 intptr_t off[2]; 3663 int nst = 0; 3664 if (!split) { 3665 ++new_long; 3666 off[nst] = offset; 3667 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3668 phase->longcon(con), T_LONG, MemNode::unordered); 3669 } else { 3670 // Omit either if it is a zero. 3671 if (con0 != 0) { 3672 ++new_int; 3673 off[nst] = offset; 3674 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3675 phase->intcon(con0), T_INT, MemNode::unordered); 3676 } 3677 if (con1 != 0) { 3678 ++new_int; 3679 offset += BytesPerInt; 3680 adr = make_raw_address(offset, phase); 3681 off[nst] = offset; 3682 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3683 phase->intcon(con1), T_INT, MemNode::unordered); 3684 } 3685 } 3686 3687 // Insert second store first, then the first before the second. 3688 // Insert each one just before any overlapping non-constant stores. 3689 while (nst > 0) { 3690 Node* st1 = st[--nst]; 3691 C->copy_node_notes_to(st1, old); 3692 st1 = phase->transform(st1); 3693 offset = off[nst]; 3694 assert(offset >= header_size, "do not smash header"); 3695 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3696 guarantee(ins_idx != 0, "must re-insert constant store"); 3697 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3698 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3699 set_req(--ins_idx, st1); 3700 else 3701 ins_req(ins_idx, st1); 3702 } 3703 } 3704 3705 if (PrintCompilation && WizardMode) 3706 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3707 old_subword, old_long, new_int, new_long); 3708 if (C->log() != NULL) 3709 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3710 old_subword, old_long, new_int, new_long); 3711 3712 // Clean up any remaining occurrences of zmem: 3713 remove_extra_zeroes(); 3714 } 3715 3716 // Explore forward from in(start) to find the first fully initialized 3717 // word, and return its offset. Skip groups of subword stores which 3718 // together initialize full words. If in(start) is itself part of a 3719 // fully initialized word, return the offset of in(start). If there 3720 // are no following full-word stores, or if something is fishy, return 3721 // a negative value. 3722 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3723 int int_map = 0; 3724 intptr_t int_map_off = 0; 3725 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3726 3727 for (uint i = start, limit = req(); i < limit; i++) { 3728 Node* st = in(i); 3729 3730 intptr_t st_off = get_store_offset(st, phase); 3731 if (st_off < 0) break; // return conservative answer 3732 3733 int st_size = st->as_Store()->memory_size(); 3734 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3735 return st_off; // we found a complete word init 3736 } 3737 3738 // update the map: 3739 3740 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3741 if (this_int_off != int_map_off) { 3742 // reset the map: 3743 int_map = 0; 3744 int_map_off = this_int_off; 3745 } 3746 3747 int subword_off = st_off - this_int_off; 3748 int_map |= right_n_bits(st_size) << subword_off; 3749 if ((int_map & FULL_MAP) == FULL_MAP) { 3750 return this_int_off; // we found a complete word init 3751 } 3752 3753 // Did this store hit or cross the word boundary? 3754 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3755 if (next_int_off == this_int_off + BytesPerInt) { 3756 // We passed the current int, without fully initializing it. 3757 int_map_off = next_int_off; 3758 int_map >>= BytesPerInt; 3759 } else if (next_int_off > this_int_off + BytesPerInt) { 3760 // We passed the current and next int. 3761 return this_int_off + BytesPerInt; 3762 } 3763 } 3764 3765 return -1; 3766 } 3767 3768 3769 // Called when the associated AllocateNode is expanded into CFG. 3770 // At this point, we may perform additional optimizations. 3771 // Linearize the stores by ascending offset, to make memory 3772 // activity as coherent as possible. 3773 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3774 intptr_t header_size, 3775 Node* size_in_bytes, 3776 PhaseGVN* phase) { 3777 assert(!is_complete(), "not already complete"); 3778 assert(stores_are_sane(phase), ""); 3779 assert(allocation() != NULL, "must be present"); 3780 3781 remove_extra_zeroes(); 3782 3783 if (ReduceFieldZeroing || ReduceBulkZeroing) 3784 // reduce instruction count for common initialization patterns 3785 coalesce_subword_stores(header_size, size_in_bytes, phase); 3786 3787 Node* zmem = zero_memory(); // initially zero memory state 3788 Node* inits = zmem; // accumulating a linearized chain of inits 3789 #ifdef ASSERT 3790 intptr_t first_offset = allocation()->minimum_header_size(); 3791 intptr_t last_init_off = first_offset; // previous init offset 3792 intptr_t last_init_end = first_offset; // previous init offset+size 3793 intptr_t last_tile_end = first_offset; // previous tile offset+size 3794 #endif 3795 intptr_t zeroes_done = header_size; 3796 3797 bool do_zeroing = true; // we might give up if inits are very sparse 3798 int big_init_gaps = 0; // how many large gaps have we seen? 3799 3800 if (ZeroTLAB) do_zeroing = false; 3801 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3802 3803 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3804 Node* st = in(i); 3805 intptr_t st_off = get_store_offset(st, phase); 3806 if (st_off < 0) 3807 break; // unknown junk in the inits 3808 if (st->in(MemNode::Memory) != zmem) 3809 break; // complicated store chains somehow in list 3810 3811 int st_size = st->as_Store()->memory_size(); 3812 intptr_t next_init_off = st_off + st_size; 3813 3814 if (do_zeroing && zeroes_done < next_init_off) { 3815 // See if this store needs a zero before it or under it. 3816 intptr_t zeroes_needed = st_off; 3817 3818 if (st_size < BytesPerInt) { 3819 // Look for subword stores which only partially initialize words. 3820 // If we find some, we must lay down some word-level zeroes first, 3821 // underneath the subword stores. 3822 // 3823 // Examples: 3824 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3825 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3826 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3827 // 3828 // Note: coalesce_subword_stores may have already done this, 3829 // if it was prompted by constant non-zero subword initializers. 3830 // But this case can still arise with non-constant stores. 3831 3832 intptr_t next_full_store = find_next_fullword_store(i, phase); 3833 3834 // In the examples above: 3835 // in(i) p q r s x y z 3836 // st_off 12 13 14 15 12 13 14 3837 // st_size 1 1 1 1 1 1 1 3838 // next_full_s. 12 16 16 16 16 16 16 3839 // z's_done 12 16 16 16 12 16 12 3840 // z's_needed 12 16 16 16 16 16 16 3841 // zsize 0 0 0 0 4 0 4 3842 if (next_full_store < 0) { 3843 // Conservative tack: Zero to end of current word. 3844 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3845 } else { 3846 // Zero to beginning of next fully initialized word. 3847 // Or, don't zero at all, if we are already in that word. 3848 assert(next_full_store >= zeroes_needed, "must go forward"); 3849 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3850 zeroes_needed = next_full_store; 3851 } 3852 } 3853 3854 if (zeroes_needed > zeroes_done) { 3855 intptr_t zsize = zeroes_needed - zeroes_done; 3856 // Do some incremental zeroing on rawmem, in parallel with inits. 3857 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3858 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3859 zeroes_done, zeroes_needed, 3860 phase); 3861 zeroes_done = zeroes_needed; 3862 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3863 do_zeroing = false; // leave the hole, next time 3864 } 3865 } 3866 3867 // Collect the store and move on: 3868 st->set_req(MemNode::Memory, inits); 3869 inits = st; // put it on the linearized chain 3870 set_req(i, zmem); // unhook from previous position 3871 3872 if (zeroes_done == st_off) 3873 zeroes_done = next_init_off; 3874 3875 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3876 3877 #ifdef ASSERT 3878 // Various order invariants. Weaker than stores_are_sane because 3879 // a large constant tile can be filled in by smaller non-constant stores. 3880 assert(st_off >= last_init_off, "inits do not reverse"); 3881 last_init_off = st_off; 3882 const Type* val = NULL; 3883 if (st_size >= BytesPerInt && 3884 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3885 (int)val->basic_type() < (int)T_OBJECT) { 3886 assert(st_off >= last_tile_end, "tiles do not overlap"); 3887 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3888 last_tile_end = MAX2(last_tile_end, next_init_off); 3889 } else { 3890 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3891 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3892 assert(st_off >= last_init_end, "inits do not overlap"); 3893 last_init_end = next_init_off; // it's a non-tile 3894 } 3895 #endif //ASSERT 3896 } 3897 3898 remove_extra_zeroes(); // clear out all the zmems left over 3899 add_req(inits); 3900 3901 if (!ZeroTLAB) { 3902 // If anything remains to be zeroed, zero it all now. 3903 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3904 // if it is the last unused 4 bytes of an instance, forget about it 3905 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3906 if (zeroes_done + BytesPerLong >= size_limit) { 3907 AllocateNode* alloc = allocation(); 3908 assert(alloc != NULL, "must be present"); 3909 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 3910 Node* klass_node = alloc->in(AllocateNode::KlassNode); 3911 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3912 if (zeroes_done == k->layout_helper()) 3913 zeroes_done = size_limit; 3914 } 3915 } 3916 if (zeroes_done < size_limit) { 3917 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3918 zeroes_done, size_in_bytes, phase); 3919 } 3920 } 3921 3922 set_complete(phase); 3923 return rawmem; 3924 } 3925 3926 3927 #ifdef ASSERT 3928 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3929 if (is_complete()) 3930 return true; // stores could be anything at this point 3931 assert(allocation() != NULL, "must be present"); 3932 intptr_t last_off = allocation()->minimum_header_size(); 3933 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3934 Node* st = in(i); 3935 intptr_t st_off = get_store_offset(st, phase); 3936 if (st_off < 0) continue; // ignore dead garbage 3937 if (last_off > st_off) { 3938 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 3939 this->dump(2); 3940 assert(false, "ascending store offsets"); 3941 return false; 3942 } 3943 last_off = st_off + st->as_Store()->memory_size(); 3944 } 3945 return true; 3946 } 3947 #endif //ASSERT 3948 3949 3950 3951 3952 //============================MergeMemNode===================================== 3953 // 3954 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3955 // contributing store or call operations. Each contributor provides the memory 3956 // state for a particular "alias type" (see Compile::alias_type). For example, 3957 // if a MergeMem has an input X for alias category #6, then any memory reference 3958 // to alias category #6 may use X as its memory state input, as an exact equivalent 3959 // to using the MergeMem as a whole. 3960 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3961 // 3962 // (Here, the <N> notation gives the index of the relevant adr_type.) 3963 // 3964 // In one special case (and more cases in the future), alias categories overlap. 3965 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3966 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3967 // it is exactly equivalent to that state W: 3968 // MergeMem(<Bot>: W) <==> W 3969 // 3970 // Usually, the merge has more than one input. In that case, where inputs 3971 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3972 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3973 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3974 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3975 // 3976 // A merge can take a "wide" memory state as one of its narrow inputs. 3977 // This simply means that the merge observes out only the relevant parts of 3978 // the wide input. That is, wide memory states arriving at narrow merge inputs 3979 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3980 // 3981 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3982 // and that memory slices "leak through": 3983 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3984 // 3985 // But, in such a cascade, repeated memory slices can "block the leak": 3986 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3987 // 3988 // In the last example, Y is not part of the combined memory state of the 3989 // outermost MergeMem. The system must, of course, prevent unschedulable 3990 // memory states from arising, so you can be sure that the state Y is somehow 3991 // a precursor to state Y'. 3992 // 3993 // 3994 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3995 // of each MergeMemNode array are exactly the numerical alias indexes, including 3996 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3997 // Compile::alias_type (and kin) produce and manage these indexes. 3998 // 3999 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4000 // (Note that this provides quick access to the top node inside MergeMem methods, 4001 // without the need to reach out via TLS to Compile::current.) 4002 // 4003 // As a consequence of what was just described, a MergeMem that represents a full 4004 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4005 // containing all alias categories. 4006 // 4007 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4008 // 4009 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4010 // a memory state for the alias type <N>, or else the top node, meaning that 4011 // there is no particular input for that alias type. Note that the length of 4012 // a MergeMem is variable, and may be extended at any time to accommodate new 4013 // memory states at larger alias indexes. When merges grow, they are of course 4014 // filled with "top" in the unused in() positions. 4015 // 4016 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4017 // (Top was chosen because it works smoothly with passes like GCM.) 4018 // 4019 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4020 // the type of random VM bits like TLS references.) Since it is always the 4021 // first non-Bot memory slice, some low-level loops use it to initialize an 4022 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4023 // 4024 // 4025 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4026 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4027 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4028 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4029 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4030 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4031 // 4032 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4033 // really that different from the other memory inputs. An abbreviation called 4034 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4035 // 4036 // 4037 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4038 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4039 // that "emerges though" the base memory will be marked as excluding the alias types 4040 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4041 // 4042 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4043 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4044 // 4045 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4046 // (It is currently unimplemented.) As you can see, the resulting merge is 4047 // actually a disjoint union of memory states, rather than an overlay. 4048 // 4049 4050 //------------------------------MergeMemNode----------------------------------- 4051 Node* MergeMemNode::make_empty_memory() { 4052 Node* empty_memory = (Node*) Compile::current()->top(); 4053 assert(empty_memory->is_top(), "correct sentinel identity"); 4054 return empty_memory; 4055 } 4056 4057 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4058 init_class_id(Class_MergeMem); 4059 // all inputs are nullified in Node::Node(int) 4060 // set_input(0, NULL); // no control input 4061 4062 // Initialize the edges uniformly to top, for starters. 4063 Node* empty_mem = make_empty_memory(); 4064 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4065 init_req(i,empty_mem); 4066 } 4067 assert(empty_memory() == empty_mem, ""); 4068 4069 if( new_base != NULL && new_base->is_MergeMem() ) { 4070 MergeMemNode* mdef = new_base->as_MergeMem(); 4071 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4072 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4073 mms.set_memory(mms.memory2()); 4074 } 4075 assert(base_memory() == mdef->base_memory(), ""); 4076 } else { 4077 set_base_memory(new_base); 4078 } 4079 } 4080 4081 // Make a new, untransformed MergeMem with the same base as 'mem'. 4082 // If mem is itself a MergeMem, populate the result with the same edges. 4083 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 4084 return new(C) MergeMemNode(mem); 4085 } 4086 4087 //------------------------------cmp-------------------------------------------- 4088 uint MergeMemNode::hash() const { return NO_HASH; } 4089 uint MergeMemNode::cmp( const Node &n ) const { 4090 return (&n == this); // Always fail except on self 4091 } 4092 4093 //------------------------------Identity--------------------------------------- 4094 Node* MergeMemNode::Identity(PhaseTransform *phase) { 4095 // Identity if this merge point does not record any interesting memory 4096 // disambiguations. 4097 Node* base_mem = base_memory(); 4098 Node* empty_mem = empty_memory(); 4099 if (base_mem != empty_mem) { // Memory path is not dead? 4100 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4101 Node* mem = in(i); 4102 if (mem != empty_mem && mem != base_mem) { 4103 return this; // Many memory splits; no change 4104 } 4105 } 4106 } 4107 return base_mem; // No memory splits; ID on the one true input 4108 } 4109 4110 //------------------------------Ideal------------------------------------------ 4111 // This method is invoked recursively on chains of MergeMem nodes 4112 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4113 // Remove chain'd MergeMems 4114 // 4115 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4116 // relative to the "in(Bot)". Since we are patching both at the same time, 4117 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4118 // but rewrite each "in(i)" relative to the new "in(Bot)". 4119 Node *progress = NULL; 4120 4121 4122 Node* old_base = base_memory(); 4123 Node* empty_mem = empty_memory(); 4124 if (old_base == empty_mem) 4125 return NULL; // Dead memory path. 4126 4127 MergeMemNode* old_mbase; 4128 if (old_base != NULL && old_base->is_MergeMem()) 4129 old_mbase = old_base->as_MergeMem(); 4130 else 4131 old_mbase = NULL; 4132 Node* new_base = old_base; 4133 4134 // simplify stacked MergeMems in base memory 4135 if (old_mbase) new_base = old_mbase->base_memory(); 4136 4137 // the base memory might contribute new slices beyond my req() 4138 if (old_mbase) grow_to_match(old_mbase); 4139 4140 // Look carefully at the base node if it is a phi. 4141 PhiNode* phi_base; 4142 if (new_base != NULL && new_base->is_Phi()) 4143 phi_base = new_base->as_Phi(); 4144 else 4145 phi_base = NULL; 4146 4147 Node* phi_reg = NULL; 4148 uint phi_len = (uint)-1; 4149 if (phi_base != NULL && !phi_base->is_copy()) { 4150 // do not examine phi if degraded to a copy 4151 phi_reg = phi_base->region(); 4152 phi_len = phi_base->req(); 4153 // see if the phi is unfinished 4154 for (uint i = 1; i < phi_len; i++) { 4155 if (phi_base->in(i) == NULL) { 4156 // incomplete phi; do not look at it yet! 4157 phi_reg = NULL; 4158 phi_len = (uint)-1; 4159 break; 4160 } 4161 } 4162 } 4163 4164 // Note: We do not call verify_sparse on entry, because inputs 4165 // can normalize to the base_memory via subsume_node or similar 4166 // mechanisms. This method repairs that damage. 4167 4168 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4169 4170 // Look at each slice. 4171 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4172 Node* old_in = in(i); 4173 // calculate the old memory value 4174 Node* old_mem = old_in; 4175 if (old_mem == empty_mem) old_mem = old_base; 4176 assert(old_mem == memory_at(i), ""); 4177 4178 // maybe update (reslice) the old memory value 4179 4180 // simplify stacked MergeMems 4181 Node* new_mem = old_mem; 4182 MergeMemNode* old_mmem; 4183 if (old_mem != NULL && old_mem->is_MergeMem()) 4184 old_mmem = old_mem->as_MergeMem(); 4185 else 4186 old_mmem = NULL; 4187 if (old_mmem == this) { 4188 // This can happen if loops break up and safepoints disappear. 4189 // A merge of BotPtr (default) with a RawPtr memory derived from a 4190 // safepoint can be rewritten to a merge of the same BotPtr with 4191 // the BotPtr phi coming into the loop. If that phi disappears 4192 // also, we can end up with a self-loop of the mergemem. 4193 // In general, if loops degenerate and memory effects disappear, 4194 // a mergemem can be left looking at itself. This simply means 4195 // that the mergemem's default should be used, since there is 4196 // no longer any apparent effect on this slice. 4197 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4198 // from start. Update the input to TOP. 4199 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4200 } 4201 else if (old_mmem != NULL) { 4202 new_mem = old_mmem->memory_at(i); 4203 } 4204 // else preceding memory was not a MergeMem 4205 4206 // replace equivalent phis (unfortunately, they do not GVN together) 4207 if (new_mem != NULL && new_mem != new_base && 4208 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4209 if (new_mem->is_Phi()) { 4210 PhiNode* phi_mem = new_mem->as_Phi(); 4211 for (uint i = 1; i < phi_len; i++) { 4212 if (phi_base->in(i) != phi_mem->in(i)) { 4213 phi_mem = NULL; 4214 break; 4215 } 4216 } 4217 if (phi_mem != NULL) { 4218 // equivalent phi nodes; revert to the def 4219 new_mem = new_base; 4220 } 4221 } 4222 } 4223 4224 // maybe store down a new value 4225 Node* new_in = new_mem; 4226 if (new_in == new_base) new_in = empty_mem; 4227 4228 if (new_in != old_in) { 4229 // Warning: Do not combine this "if" with the previous "if" 4230 // A memory slice might have be be rewritten even if it is semantically 4231 // unchanged, if the base_memory value has changed. 4232 set_req(i, new_in); 4233 progress = this; // Report progress 4234 } 4235 } 4236 4237 if (new_base != old_base) { 4238 set_req(Compile::AliasIdxBot, new_base); 4239 // Don't use set_base_memory(new_base), because we need to update du. 4240 assert(base_memory() == new_base, ""); 4241 progress = this; 4242 } 4243 4244 if( base_memory() == this ) { 4245 // a self cycle indicates this memory path is dead 4246 set_req(Compile::AliasIdxBot, empty_mem); 4247 } 4248 4249 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4250 // Recursion must occur after the self cycle check above 4251 if( base_memory()->is_MergeMem() ) { 4252 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4253 Node *m = phase->transform(new_mbase); // Rollup any cycles 4254 if( m != NULL && (m->is_top() || 4255 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4256 // propagate rollup of dead cycle to self 4257 set_req(Compile::AliasIdxBot, empty_mem); 4258 } 4259 } 4260 4261 if( base_memory() == empty_mem ) { 4262 progress = this; 4263 // Cut inputs during Parse phase only. 4264 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4265 if( !can_reshape ) { 4266 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4267 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4268 } 4269 } 4270 } 4271 4272 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4273 // Check if PhiNode::Ideal's "Split phis through memory merges" 4274 // transform should be attempted. Look for this->phi->this cycle. 4275 uint merge_width = req(); 4276 if (merge_width > Compile::AliasIdxRaw) { 4277 PhiNode* phi = base_memory()->as_Phi(); 4278 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4279 if (phi->in(i) == this) { 4280 phase->is_IterGVN()->_worklist.push(phi); 4281 break; 4282 } 4283 } 4284 } 4285 } 4286 4287 assert(progress || verify_sparse(), "please, no dups of base"); 4288 return progress; 4289 } 4290 4291 //-------------------------set_base_memory------------------------------------- 4292 void MergeMemNode::set_base_memory(Node *new_base) { 4293 Node* empty_mem = empty_memory(); 4294 set_req(Compile::AliasIdxBot, new_base); 4295 assert(memory_at(req()) == new_base, "must set default memory"); 4296 // Clear out other occurrences of new_base: 4297 if (new_base != empty_mem) { 4298 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4299 if (in(i) == new_base) set_req(i, empty_mem); 4300 } 4301 } 4302 } 4303 4304 //------------------------------out_RegMask------------------------------------ 4305 const RegMask &MergeMemNode::out_RegMask() const { 4306 return RegMask::Empty; 4307 } 4308 4309 //------------------------------dump_spec-------------------------------------- 4310 #ifndef PRODUCT 4311 void MergeMemNode::dump_spec(outputStream *st) const { 4312 st->print(" {"); 4313 Node* base_mem = base_memory(); 4314 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4315 Node* mem = memory_at(i); 4316 if (mem == base_mem) { st->print(" -"); continue; } 4317 st->print( " N%d:", mem->_idx ); 4318 Compile::current()->get_adr_type(i)->dump_on(st); 4319 } 4320 st->print(" }"); 4321 } 4322 #endif // !PRODUCT 4323 4324 4325 #ifdef ASSERT 4326 static bool might_be_same(Node* a, Node* b) { 4327 if (a == b) return true; 4328 if (!(a->is_Phi() || b->is_Phi())) return false; 4329 // phis shift around during optimization 4330 return true; // pretty stupid... 4331 } 4332 4333 // verify a narrow slice (either incoming or outgoing) 4334 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4335 if (!VerifyAliases) return; // don't bother to verify unless requested 4336 if (is_error_reported()) return; // muzzle asserts when debugging an error 4337 if (Node::in_dump()) return; // muzzle asserts when printing 4338 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4339 assert(n != NULL, ""); 4340 // Elide intervening MergeMem's 4341 while (n->is_MergeMem()) { 4342 n = n->as_MergeMem()->memory_at(alias_idx); 4343 } 4344 Compile* C = Compile::current(); 4345 const TypePtr* n_adr_type = n->adr_type(); 4346 if (n == m->empty_memory()) { 4347 // Implicit copy of base_memory() 4348 } else if (n_adr_type != TypePtr::BOTTOM) { 4349 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4350 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4351 } else { 4352 // A few places like make_runtime_call "know" that VM calls are narrow, 4353 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4354 bool expected_wide_mem = false; 4355 if (n == m->base_memory()) { 4356 expected_wide_mem = true; 4357 } else if (alias_idx == Compile::AliasIdxRaw || 4358 n == m->memory_at(Compile::AliasIdxRaw)) { 4359 expected_wide_mem = true; 4360 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4361 // memory can "leak through" calls on channels that 4362 // are write-once. Allow this also. 4363 expected_wide_mem = true; 4364 } 4365 assert(expected_wide_mem, "expected narrow slice replacement"); 4366 } 4367 } 4368 #else // !ASSERT 4369 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4370 #endif 4371 4372 4373 //-----------------------------memory_at--------------------------------------- 4374 Node* MergeMemNode::memory_at(uint alias_idx) const { 4375 assert(alias_idx >= Compile::AliasIdxRaw || 4376 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4377 "must avoid base_memory and AliasIdxTop"); 4378 4379 // Otherwise, it is a narrow slice. 4380 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4381 Compile *C = Compile::current(); 4382 if (is_empty_memory(n)) { 4383 // the array is sparse; empty slots are the "top" node 4384 n = base_memory(); 4385 assert(Node::in_dump() 4386 || n == NULL || n->bottom_type() == Type::TOP 4387 || n->adr_type() == NULL // address is TOP 4388 || n->adr_type() == TypePtr::BOTTOM 4389 || n->adr_type() == TypeRawPtr::BOTTOM 4390 || Compile::current()->AliasLevel() == 0, 4391 "must be a wide memory"); 4392 // AliasLevel == 0 if we are organizing the memory states manually. 4393 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4394 } else { 4395 // make sure the stored slice is sane 4396 #ifdef ASSERT 4397 if (is_error_reported() || Node::in_dump()) { 4398 } else if (might_be_same(n, base_memory())) { 4399 // Give it a pass: It is a mostly harmless repetition of the base. 4400 // This can arise normally from node subsumption during optimization. 4401 } else { 4402 verify_memory_slice(this, alias_idx, n); 4403 } 4404 #endif 4405 } 4406 return n; 4407 } 4408 4409 //---------------------------set_memory_at------------------------------------- 4410 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4411 verify_memory_slice(this, alias_idx, n); 4412 Node* empty_mem = empty_memory(); 4413 if (n == base_memory()) n = empty_mem; // collapse default 4414 uint need_req = alias_idx+1; 4415 if (req() < need_req) { 4416 if (n == empty_mem) return; // already the default, so do not grow me 4417 // grow the sparse array 4418 do { 4419 add_req(empty_mem); 4420 } while (req() < need_req); 4421 } 4422 set_req( alias_idx, n ); 4423 } 4424 4425 4426 4427 //--------------------------iteration_setup------------------------------------ 4428 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4429 if (other != NULL) { 4430 grow_to_match(other); 4431 // invariant: the finite support of mm2 is within mm->req() 4432 #ifdef ASSERT 4433 for (uint i = req(); i < other->req(); i++) { 4434 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4435 } 4436 #endif 4437 } 4438 // Replace spurious copies of base_memory by top. 4439 Node* base_mem = base_memory(); 4440 if (base_mem != NULL && !base_mem->is_top()) { 4441 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4442 if (in(i) == base_mem) 4443 set_req(i, empty_memory()); 4444 } 4445 } 4446 } 4447 4448 //---------------------------grow_to_match------------------------------------- 4449 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4450 Node* empty_mem = empty_memory(); 4451 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4452 // look for the finite support of the other memory 4453 for (uint i = other->req(); --i >= req(); ) { 4454 if (other->in(i) != empty_mem) { 4455 uint new_len = i+1; 4456 while (req() < new_len) add_req(empty_mem); 4457 break; 4458 } 4459 } 4460 } 4461 4462 //---------------------------verify_sparse------------------------------------- 4463 #ifndef PRODUCT 4464 bool MergeMemNode::verify_sparse() const { 4465 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4466 Node* base_mem = base_memory(); 4467 // The following can happen in degenerate cases, since empty==top. 4468 if (is_empty_memory(base_mem)) return true; 4469 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4470 assert(in(i) != NULL, "sane slice"); 4471 if (in(i) == base_mem) return false; // should have been the sentinel value! 4472 } 4473 return true; 4474 } 4475 4476 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4477 Node* n; 4478 n = mm->in(idx); 4479 if (mem == n) return true; // might be empty_memory() 4480 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4481 if (mem == n) return true; 4482 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4483 if (mem == n) return true; 4484 if (n == NULL) break; 4485 } 4486 return false; 4487 } 4488 #endif // !PRODUCT