1 /* 2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "oops/objArrayKlass.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/compile.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/machnode.hpp" 36 #include "opto/matcher.hpp" 37 #include "opto/memnode.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/phaseX.hpp" 40 #include "opto/regmask.hpp" 41 42 // Portions of code courtesy of Clifford Click 43 44 // Optimization - Graph Style 45 46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 47 48 //============================================================================= 49 uint MemNode::size_of() const { return sizeof(*this); } 50 51 const TypePtr *MemNode::adr_type() const { 52 Node* adr = in(Address); 53 const TypePtr* cross_check = NULL; 54 DEBUG_ONLY(cross_check = _adr_type); 55 return calculate_adr_type(adr->bottom_type(), cross_check); 56 } 57 58 #ifndef PRODUCT 59 void MemNode::dump_spec(outputStream *st) const { 60 if (in(Address) == NULL) return; // node is dead 61 #ifndef ASSERT 62 // fake the missing field 63 const TypePtr* _adr_type = NULL; 64 if (in(Address) != NULL) 65 _adr_type = in(Address)->bottom_type()->isa_ptr(); 66 #endif 67 dump_adr_type(this, _adr_type, st); 68 69 Compile* C = Compile::current(); 70 if( C->alias_type(_adr_type)->is_volatile() ) 71 st->print(" Volatile!"); 72 } 73 74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 75 st->print(" @"); 76 if (adr_type == NULL) { 77 st->print("NULL"); 78 } else { 79 adr_type->dump_on(st); 80 Compile* C = Compile::current(); 81 Compile::AliasType* atp = NULL; 82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 83 if (atp == NULL) 84 st->print(", idx=?\?;"); 85 else if (atp->index() == Compile::AliasIdxBot) 86 st->print(", idx=Bot;"); 87 else if (atp->index() == Compile::AliasIdxTop) 88 st->print(", idx=Top;"); 89 else if (atp->index() == Compile::AliasIdxRaw) 90 st->print(", idx=Raw;"); 91 else { 92 ciField* field = atp->field(); 93 if (field) { 94 st->print(", name="); 95 field->print_name_on(st); 96 } 97 st->print(", idx=%d;", atp->index()); 98 } 99 } 100 } 101 102 extern void print_alias_types(); 103 104 #endif 105 106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 107 assert((t_oop != NULL), "sanity"); 108 bool is_instance = t_oop->is_known_instance_field(); 109 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 110 (load != NULL) && load->is_Load() && 111 (phase->is_IterGVN() != NULL); 112 if (!(is_instance || is_boxed_value_load)) 113 return mchain; // don't try to optimize non-instance types 114 uint instance_id = t_oop->instance_id(); 115 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 116 Node *prev = NULL; 117 Node *result = mchain; 118 while (prev != result) { 119 prev = result; 120 if (result == start_mem) 121 break; // hit one of our sentinels 122 // skip over a call which does not affect this memory slice 123 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 124 Node *proj_in = result->in(0); 125 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 126 break; // hit one of our sentinels 127 } else if (proj_in->is_Call()) { 128 CallNode *call = proj_in->as_Call(); 129 if (!call->may_modify(t_oop, phase)) { // returns false for instances 130 result = call->in(TypeFunc::Memory); 131 } 132 } else if (proj_in->is_Initialize()) { 133 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 134 // Stop if this is the initialization for the object instance which 135 // which contains this memory slice, otherwise skip over it. 136 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 137 break; 138 } 139 if (is_instance) { 140 result = proj_in->in(TypeFunc::Memory); 141 } else if (is_boxed_value_load) { 142 Node* klass = alloc->in(AllocateNode::KlassNode); 143 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 144 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 145 result = proj_in->in(TypeFunc::Memory); // not related allocation 146 } 147 } 148 } else if (proj_in->is_MemBar()) { 149 result = proj_in->in(TypeFunc::Memory); 150 } else { 151 assert(false, "unexpected projection"); 152 } 153 } else if (result->is_ClearArray()) { 154 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 155 // Can not bypass initialization of the instance 156 // we are looking for. 157 break; 158 } 159 // Otherwise skip it (the call updated 'result' value). 160 } else if (result->is_MergeMem()) { 161 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 162 } 163 } 164 return result; 165 } 166 167 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 168 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 169 if (t_oop == NULL) 170 return mchain; // don't try to optimize non-oop types 171 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 172 bool is_instance = t_oop->is_known_instance_field(); 173 PhaseIterGVN *igvn = phase->is_IterGVN(); 174 if (is_instance && igvn != NULL && result->is_Phi()) { 175 PhiNode *mphi = result->as_Phi(); 176 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 177 const TypePtr *t = mphi->adr_type(); 178 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 179 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 180 t->is_oopptr()->cast_to_exactness(true) 181 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 182 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 183 // clone the Phi with our address type 184 result = mphi->split_out_instance(t_adr, igvn); 185 } else { 186 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 187 } 188 } 189 return result; 190 } 191 192 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 193 uint alias_idx = phase->C->get_alias_index(tp); 194 Node *mem = mmem; 195 #ifdef ASSERT 196 { 197 // Check that current type is consistent with the alias index used during graph construction 198 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 199 bool consistent = adr_check == NULL || adr_check->empty() || 200 phase->C->must_alias(adr_check, alias_idx ); 201 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 202 if( !consistent && adr_check != NULL && !adr_check->empty() && 203 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 204 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 205 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 206 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 207 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 208 // don't assert if it is dead code. 209 consistent = true; 210 } 211 if( !consistent ) { 212 st->print("alias_idx==%d, adr_check==", alias_idx); 213 if( adr_check == NULL ) { 214 st->print("NULL"); 215 } else { 216 adr_check->dump(); 217 } 218 st->cr(); 219 print_alias_types(); 220 assert(consistent, "adr_check must match alias idx"); 221 } 222 } 223 #endif 224 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 225 // means an array I have not precisely typed yet. Do not do any 226 // alias stuff with it any time soon. 227 const TypeOopPtr *toop = tp->isa_oopptr(); 228 if( tp->base() != Type::AnyPtr && 229 !(toop && 230 toop->klass() != NULL && 231 toop->klass()->is_java_lang_Object() && 232 toop->offset() == Type::OffsetBot) ) { 233 // compress paths and change unreachable cycles to TOP 234 // If not, we can update the input infinitely along a MergeMem cycle 235 // Equivalent code in PhiNode::Ideal 236 Node* m = phase->transform(mmem); 237 // If transformed to a MergeMem, get the desired slice 238 // Otherwise the returned node represents memory for every slice 239 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 240 // Update input if it is progress over what we have now 241 } 242 return mem; 243 } 244 245 //--------------------------Ideal_common--------------------------------------- 246 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 247 // Unhook non-raw memories from complete (macro-expanded) initializations. 248 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 249 // If our control input is a dead region, kill all below the region 250 Node *ctl = in(MemNode::Control); 251 if (ctl && remove_dead_region(phase, can_reshape)) 252 return this; 253 ctl = in(MemNode::Control); 254 // Don't bother trying to transform a dead node 255 if (ctl && ctl->is_top()) return NodeSentinel; 256 257 PhaseIterGVN *igvn = phase->is_IterGVN(); 258 // Wait if control on the worklist. 259 if (ctl && can_reshape && igvn != NULL) { 260 Node* bol = NULL; 261 Node* cmp = NULL; 262 if (ctl->in(0)->is_If()) { 263 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 264 bol = ctl->in(0)->in(1); 265 if (bol->is_Bool()) 266 cmp = ctl->in(0)->in(1)->in(1); 267 } 268 if (igvn->_worklist.member(ctl) || 269 (bol != NULL && igvn->_worklist.member(bol)) || 270 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 271 // This control path may be dead. 272 // Delay this memory node transformation until the control is processed. 273 phase->is_IterGVN()->_worklist.push(this); 274 return NodeSentinel; // caller will return NULL 275 } 276 } 277 // Ignore if memory is dead, or self-loop 278 Node *mem = in(MemNode::Memory); 279 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 280 assert(mem != this, "dead loop in MemNode::Ideal"); 281 282 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 283 // This memory slice may be dead. 284 // Delay this mem node transformation until the memory is processed. 285 phase->is_IterGVN()->_worklist.push(this); 286 return NodeSentinel; // caller will return NULL 287 } 288 289 Node *address = in(MemNode::Address); 290 const Type *t_adr = phase->type(address); 291 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 292 293 if (can_reshape && igvn != NULL && 294 (igvn->_worklist.member(address) || 295 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { 296 // The address's base and type may change when the address is processed. 297 // Delay this mem node transformation until the address is processed. 298 phase->is_IterGVN()->_worklist.push(this); 299 return NodeSentinel; // caller will return NULL 300 } 301 302 // Do NOT remove or optimize the next lines: ensure a new alias index 303 // is allocated for an oop pointer type before Escape Analysis. 304 // Note: C++ will not remove it since the call has side effect. 305 if (t_adr->isa_oopptr()) { 306 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 307 } 308 309 Node* base = NULL; 310 if (address->is_AddP()) { 311 base = address->in(AddPNode::Base); 312 } 313 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 314 !t_adr->isa_rawptr()) { 315 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 316 // Skip this node optimization if its address has TOP base. 317 return NodeSentinel; // caller will return NULL 318 } 319 320 // Avoid independent memory operations 321 Node* old_mem = mem; 322 323 // The code which unhooks non-raw memories from complete (macro-expanded) 324 // initializations was removed. After macro-expansion all stores catched 325 // by Initialize node became raw stores and there is no information 326 // which memory slices they modify. So it is unsafe to move any memory 327 // operation above these stores. Also in most cases hooked non-raw memories 328 // were already unhooked by using information from detect_ptr_independence() 329 // and find_previous_store(). 330 331 if (mem->is_MergeMem()) { 332 MergeMemNode* mmem = mem->as_MergeMem(); 333 const TypePtr *tp = t_adr->is_ptr(); 334 335 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 336 } 337 338 if (mem != old_mem) { 339 set_req(MemNode::Memory, mem); 340 if (can_reshape && old_mem->outcnt() == 0) { 341 igvn->_worklist.push(old_mem); 342 } 343 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 344 return this; 345 } 346 347 // let the subclass continue analyzing... 348 return NULL; 349 } 350 351 // Helper function for proving some simple control dominations. 352 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 353 // Already assumes that 'dom' is available at 'sub', and that 'sub' 354 // is not a constant (dominated by the method's StartNode). 355 // Used by MemNode::find_previous_store to prove that the 356 // control input of a memory operation predates (dominates) 357 // an allocation it wants to look past. 358 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 359 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 360 return false; // Conservative answer for dead code 361 362 // Check 'dom'. Skip Proj and CatchProj nodes. 363 dom = dom->find_exact_control(dom); 364 if (dom == NULL || dom->is_top()) 365 return false; // Conservative answer for dead code 366 367 if (dom == sub) { 368 // For the case when, for example, 'sub' is Initialize and the original 369 // 'dom' is Proj node of the 'sub'. 370 return false; 371 } 372 373 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 374 return true; 375 376 // 'dom' dominates 'sub' if its control edge and control edges 377 // of all its inputs dominate or equal to sub's control edge. 378 379 // Currently 'sub' is either Allocate, Initialize or Start nodes. 380 // Or Region for the check in LoadNode::Ideal(); 381 // 'sub' should have sub->in(0) != NULL. 382 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 383 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 384 385 // Get control edge of 'sub'. 386 Node* orig_sub = sub; 387 sub = sub->find_exact_control(sub->in(0)); 388 if (sub == NULL || sub->is_top()) 389 return false; // Conservative answer for dead code 390 391 assert(sub->is_CFG(), "expecting control"); 392 393 if (sub == dom) 394 return true; 395 396 if (sub->is_Start() || sub->is_Root()) 397 return false; 398 399 { 400 // Check all control edges of 'dom'. 401 402 ResourceMark rm; 403 Arena* arena = Thread::current()->resource_area(); 404 Node_List nlist(arena); 405 Unique_Node_List dom_list(arena); 406 407 dom_list.push(dom); 408 bool only_dominating_controls = false; 409 410 for (uint next = 0; next < dom_list.size(); next++) { 411 Node* n = dom_list.at(next); 412 if (n == orig_sub) 413 return false; // One of dom's inputs dominated by sub. 414 if (!n->is_CFG() && n->pinned()) { 415 // Check only own control edge for pinned non-control nodes. 416 n = n->find_exact_control(n->in(0)); 417 if (n == NULL || n->is_top()) 418 return false; // Conservative answer for dead code 419 assert(n->is_CFG(), "expecting control"); 420 dom_list.push(n); 421 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 422 only_dominating_controls = true; 423 } else if (n->is_CFG()) { 424 if (n->dominates(sub, nlist)) 425 only_dominating_controls = true; 426 else 427 return false; 428 } else { 429 // First, own control edge. 430 Node* m = n->find_exact_control(n->in(0)); 431 if (m != NULL) { 432 if (m->is_top()) 433 return false; // Conservative answer for dead code 434 dom_list.push(m); 435 } 436 // Now, the rest of edges. 437 uint cnt = n->req(); 438 for (uint i = 1; i < cnt; i++) { 439 m = n->find_exact_control(n->in(i)); 440 if (m == NULL || m->is_top()) 441 continue; 442 dom_list.push(m); 443 } 444 } 445 } 446 return only_dominating_controls; 447 } 448 } 449 450 //---------------------detect_ptr_independence--------------------------------- 451 // Used by MemNode::find_previous_store to prove that two base 452 // pointers are never equal. 453 // The pointers are accompanied by their associated allocations, 454 // if any, which have been previously discovered by the caller. 455 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 456 Node* p2, AllocateNode* a2, 457 PhaseTransform* phase) { 458 // Attempt to prove that these two pointers cannot be aliased. 459 // They may both manifestly be allocations, and they should differ. 460 // Or, if they are not both allocations, they can be distinct constants. 461 // Otherwise, one is an allocation and the other a pre-existing value. 462 if (a1 == NULL && a2 == NULL) { // neither an allocation 463 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 464 } else if (a1 != NULL && a2 != NULL) { // both allocations 465 return (a1 != a2); 466 } else if (a1 != NULL) { // one allocation a1 467 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 468 return all_controls_dominate(p2, a1); 469 } else { //(a2 != NULL) // one allocation a2 470 return all_controls_dominate(p1, a2); 471 } 472 return false; 473 } 474 475 476 // The logic for reordering loads and stores uses four steps: 477 // (a) Walk carefully past stores and initializations which we 478 // can prove are independent of this load. 479 // (b) Observe that the next memory state makes an exact match 480 // with self (load or store), and locate the relevant store. 481 // (c) Ensure that, if we were to wire self directly to the store, 482 // the optimizer would fold it up somehow. 483 // (d) Do the rewiring, and return, depending on some other part of 484 // the optimizer to fold up the load. 485 // This routine handles steps (a) and (b). Steps (c) and (d) are 486 // specific to loads and stores, so they are handled by the callers. 487 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 488 // 489 Node* MemNode::find_previous_store(PhaseTransform* phase) { 490 Node* ctrl = in(MemNode::Control); 491 Node* adr = in(MemNode::Address); 492 intptr_t offset = 0; 493 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 494 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 495 496 if (offset == Type::OffsetBot) 497 return NULL; // cannot unalias unless there are precise offsets 498 499 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 500 501 intptr_t size_in_bytes = memory_size(); 502 503 Node* mem = in(MemNode::Memory); // start searching here... 504 505 int cnt = 50; // Cycle limiter 506 for (;;) { // While we can dance past unrelated stores... 507 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 508 509 if (mem->is_Store()) { 510 Node* st_adr = mem->in(MemNode::Address); 511 intptr_t st_offset = 0; 512 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 513 if (st_base == NULL) 514 break; // inscrutable pointer 515 if (st_offset != offset && st_offset != Type::OffsetBot) { 516 const int MAX_STORE = BytesPerLong; 517 if (st_offset >= offset + size_in_bytes || 518 st_offset <= offset - MAX_STORE || 519 st_offset <= offset - mem->as_Store()->memory_size()) { 520 // Success: The offsets are provably independent. 521 // (You may ask, why not just test st_offset != offset and be done? 522 // The answer is that stores of different sizes can co-exist 523 // in the same sequence of RawMem effects. We sometimes initialize 524 // a whole 'tile' of array elements with a single jint or jlong.) 525 mem = mem->in(MemNode::Memory); 526 continue; // (a) advance through independent store memory 527 } 528 } 529 if (st_base != base && 530 detect_ptr_independence(base, alloc, 531 st_base, 532 AllocateNode::Ideal_allocation(st_base, phase), 533 phase)) { 534 // Success: The bases are provably independent. 535 mem = mem->in(MemNode::Memory); 536 continue; // (a) advance through independent store memory 537 } 538 539 // (b) At this point, if the bases or offsets do not agree, we lose, 540 // since we have not managed to prove 'this' and 'mem' independent. 541 if (st_base == base && st_offset == offset) { 542 return mem; // let caller handle steps (c), (d) 543 } 544 545 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 546 InitializeNode* st_init = mem->in(0)->as_Initialize(); 547 AllocateNode* st_alloc = st_init->allocation(); 548 if (st_alloc == NULL) 549 break; // something degenerated 550 bool known_identical = false; 551 bool known_independent = false; 552 if (alloc == st_alloc) 553 known_identical = true; 554 else if (alloc != NULL) 555 known_independent = true; 556 else if (all_controls_dominate(this, st_alloc)) 557 known_independent = true; 558 559 if (known_independent) { 560 // The bases are provably independent: Either they are 561 // manifestly distinct allocations, or else the control 562 // of this load dominates the store's allocation. 563 int alias_idx = phase->C->get_alias_index(adr_type()); 564 if (alias_idx == Compile::AliasIdxRaw) { 565 mem = st_alloc->in(TypeFunc::Memory); 566 } else { 567 mem = st_init->memory(alias_idx); 568 } 569 continue; // (a) advance through independent store memory 570 } 571 572 // (b) at this point, if we are not looking at a store initializing 573 // the same allocation we are loading from, we lose. 574 if (known_identical) { 575 // From caller, can_see_stored_value will consult find_captured_store. 576 return mem; // let caller handle steps (c), (d) 577 } 578 579 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 580 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 581 if (mem->is_Proj() && mem->in(0)->is_Call()) { 582 CallNode *call = mem->in(0)->as_Call(); 583 if (!call->may_modify(addr_t, phase)) { 584 mem = call->in(TypeFunc::Memory); 585 continue; // (a) advance through independent call memory 586 } 587 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 588 mem = mem->in(0)->in(TypeFunc::Memory); 589 continue; // (a) advance through independent MemBar memory 590 } else if (mem->is_ClearArray()) { 591 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 592 // (the call updated 'mem' value) 593 continue; // (a) advance through independent allocation memory 594 } else { 595 // Can not bypass initialization of the instance 596 // we are looking for. 597 return mem; 598 } 599 } else if (mem->is_MergeMem()) { 600 int alias_idx = phase->C->get_alias_index(adr_type()); 601 mem = mem->as_MergeMem()->memory_at(alias_idx); 602 continue; // (a) advance through independent MergeMem memory 603 } 604 } 605 606 // Unless there is an explicit 'continue', we must bail out here, 607 // because 'mem' is an inscrutable memory state (e.g., a call). 608 break; 609 } 610 611 return NULL; // bail out 612 } 613 614 //----------------------calculate_adr_type------------------------------------- 615 // Helper function. Notices when the given type of address hits top or bottom. 616 // Also, asserts a cross-check of the type against the expected address type. 617 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 618 if (t == Type::TOP) return NULL; // does not touch memory any more? 619 #ifdef PRODUCT 620 cross_check = NULL; 621 #else 622 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 623 #endif 624 const TypePtr* tp = t->isa_ptr(); 625 if (tp == NULL) { 626 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 627 return TypePtr::BOTTOM; // touches lots of memory 628 } else { 629 #ifdef ASSERT 630 // %%%% [phh] We don't check the alias index if cross_check is 631 // TypeRawPtr::BOTTOM. Needs to be investigated. 632 if (cross_check != NULL && 633 cross_check != TypePtr::BOTTOM && 634 cross_check != TypeRawPtr::BOTTOM) { 635 // Recheck the alias index, to see if it has changed (due to a bug). 636 Compile* C = Compile::current(); 637 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 638 "must stay in the original alias category"); 639 // The type of the address must be contained in the adr_type, 640 // disregarding "null"-ness. 641 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 642 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 643 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 644 "real address must not escape from expected memory type"); 645 } 646 #endif 647 return tp; 648 } 649 } 650 651 //------------------------adr_phi_is_loop_invariant---------------------------- 652 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 653 // loop is loop invariant. Make a quick traversal of Phi and associated 654 // CastPP nodes, looking to see if they are a closed group within the loop. 655 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 656 // The idea is that the phi-nest must boil down to only CastPP nodes 657 // with the same data. This implies that any path into the loop already 658 // includes such a CastPP, and so the original cast, whatever its input, 659 // must be covered by an equivalent cast, with an earlier control input. 660 ResourceMark rm; 661 662 // The loop entry input of the phi should be the unique dominating 663 // node for every Phi/CastPP in the loop. 664 Unique_Node_List closure; 665 closure.push(adr_phi->in(LoopNode::EntryControl)); 666 667 // Add the phi node and the cast to the worklist. 668 Unique_Node_List worklist; 669 worklist.push(adr_phi); 670 if( cast != NULL ){ 671 if( !cast->is_ConstraintCast() ) return false; 672 worklist.push(cast); 673 } 674 675 // Begin recursive walk of phi nodes. 676 while( worklist.size() ){ 677 // Take a node off the worklist 678 Node *n = worklist.pop(); 679 if( !closure.member(n) ){ 680 // Add it to the closure. 681 closure.push(n); 682 // Make a sanity check to ensure we don't waste too much time here. 683 if( closure.size() > 20) return false; 684 // This node is OK if: 685 // - it is a cast of an identical value 686 // - or it is a phi node (then we add its inputs to the worklist) 687 // Otherwise, the node is not OK, and we presume the cast is not invariant 688 if( n->is_ConstraintCast() ){ 689 worklist.push(n->in(1)); 690 } else if( n->is_Phi() ) { 691 for( uint i = 1; i < n->req(); i++ ) { 692 worklist.push(n->in(i)); 693 } 694 } else { 695 return false; 696 } 697 } 698 } 699 700 // Quit when the worklist is empty, and we've found no offending nodes. 701 return true; 702 } 703 704 //------------------------------Ideal_DU_postCCP------------------------------- 705 // Find any cast-away of null-ness and keep its control. Null cast-aways are 706 // going away in this pass and we need to make this memory op depend on the 707 // gating null check. 708 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 709 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 710 } 711 712 // I tried to leave the CastPP's in. This makes the graph more accurate in 713 // some sense; we get to keep around the knowledge that an oop is not-null 714 // after some test. Alas, the CastPP's interfere with GVN (some values are 715 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 716 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 717 // some of the more trivial cases in the optimizer. Removing more useless 718 // Phi's started allowing Loads to illegally float above null checks. I gave 719 // up on this approach. CNC 10/20/2000 720 // This static method may be called not from MemNode (EncodePNode calls it). 721 // Only the control edge of the node 'n' might be updated. 722 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 723 Node *skipped_cast = NULL; 724 // Need a null check? Regular static accesses do not because they are 725 // from constant addresses. Array ops are gated by the range check (which 726 // always includes a NULL check). Just check field ops. 727 if( n->in(MemNode::Control) == NULL ) { 728 // Scan upwards for the highest location we can place this memory op. 729 while( true ) { 730 switch( adr->Opcode() ) { 731 732 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 733 adr = adr->in(AddPNode::Base); 734 continue; 735 736 case Op_DecodeN: // No change to NULL-ness, so peek thru 737 case Op_DecodeNKlass: 738 adr = adr->in(1); 739 continue; 740 741 case Op_EncodeP: 742 case Op_EncodePKlass: 743 // EncodeP node's control edge could be set by this method 744 // when EncodeP node depends on CastPP node. 745 // 746 // Use its control edge for memory op because EncodeP may go away 747 // later when it is folded with following or preceding DecodeN node. 748 if (adr->in(0) == NULL) { 749 // Keep looking for cast nodes. 750 adr = adr->in(1); 751 continue; 752 } 753 ccp->hash_delete(n); 754 n->set_req(MemNode::Control, adr->in(0)); 755 ccp->hash_insert(n); 756 return n; 757 758 case Op_CastPP: 759 // If the CastPP is useless, just peek on through it. 760 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 761 // Remember the cast that we've peeked though. If we peek 762 // through more than one, then we end up remembering the highest 763 // one, that is, if in a loop, the one closest to the top. 764 skipped_cast = adr; 765 adr = adr->in(1); 766 continue; 767 } 768 // CastPP is going away in this pass! We need this memory op to be 769 // control-dependent on the test that is guarding the CastPP. 770 ccp->hash_delete(n); 771 n->set_req(MemNode::Control, adr->in(0)); 772 ccp->hash_insert(n); 773 return n; 774 775 case Op_Phi: 776 // Attempt to float above a Phi to some dominating point. 777 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 778 // If we've already peeked through a Cast (which could have set the 779 // control), we can't float above a Phi, because the skipped Cast 780 // may not be loop invariant. 781 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 782 adr = adr->in(1); 783 continue; 784 } 785 } 786 787 // Intentional fallthrough! 788 789 // No obvious dominating point. The mem op is pinned below the Phi 790 // by the Phi itself. If the Phi goes away (no true value is merged) 791 // then the mem op can float, but not indefinitely. It must be pinned 792 // behind the controls leading to the Phi. 793 case Op_CheckCastPP: 794 // These usually stick around to change address type, however a 795 // useless one can be elided and we still need to pick up a control edge 796 if (adr->in(0) == NULL) { 797 // This CheckCastPP node has NO control and is likely useless. But we 798 // need check further up the ancestor chain for a control input to keep 799 // the node in place. 4959717. 800 skipped_cast = adr; 801 adr = adr->in(1); 802 continue; 803 } 804 ccp->hash_delete(n); 805 n->set_req(MemNode::Control, adr->in(0)); 806 ccp->hash_insert(n); 807 return n; 808 809 // List of "safe" opcodes; those that implicitly block the memory 810 // op below any null check. 811 case Op_CastX2P: // no null checks on native pointers 812 case Op_Parm: // 'this' pointer is not null 813 case Op_LoadP: // Loading from within a klass 814 case Op_LoadN: // Loading from within a klass 815 case Op_LoadKlass: // Loading from within a klass 816 case Op_LoadNKlass: // Loading from within a klass 817 case Op_ConP: // Loading from a klass 818 case Op_ConN: // Loading from a klass 819 case Op_ConNKlass: // Loading from a klass 820 case Op_CreateEx: // Sucking up the guts of an exception oop 821 case Op_Con: // Reading from TLS 822 case Op_CMoveP: // CMoveP is pinned 823 case Op_CMoveN: // CMoveN is pinned 824 break; // No progress 825 826 case Op_Proj: // Direct call to an allocation routine 827 case Op_SCMemProj: // Memory state from store conditional ops 828 #ifdef ASSERT 829 { 830 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 831 const Node* call = adr->in(0); 832 if (call->is_CallJava()) { 833 const CallJavaNode* call_java = call->as_CallJava(); 834 const TypeTuple *r = call_java->tf()->range(); 835 assert(r->cnt() > TypeFunc::Parms, "must return value"); 836 const Type* ret_type = r->field_at(TypeFunc::Parms); 837 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 838 // We further presume that this is one of 839 // new_instance_Java, new_array_Java, or 840 // the like, but do not assert for this. 841 } else if (call->is_Allocate()) { 842 // similar case to new_instance_Java, etc. 843 } else if (!call->is_CallLeaf()) { 844 // Projections from fetch_oop (OSR) are allowed as well. 845 ShouldNotReachHere(); 846 } 847 } 848 #endif 849 break; 850 default: 851 ShouldNotReachHere(); 852 } 853 break; 854 } 855 } 856 857 return NULL; // No progress 858 } 859 860 861 //============================================================================= 862 // Should LoadNode::Ideal() attempt to remove control edges? 863 bool LoadNode::can_remove_control() const { 864 return true; 865 } 866 uint LoadNode::size_of() const { return sizeof(*this); } 867 uint LoadNode::cmp( const Node &n ) const 868 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 869 const Type *LoadNode::bottom_type() const { return _type; } 870 uint LoadNode::ideal_reg() const { 871 return _type->ideal_reg(); 872 } 873 874 #ifndef PRODUCT 875 void LoadNode::dump_spec(outputStream *st) const { 876 MemNode::dump_spec(st); 877 if( !Verbose && !WizardMode ) { 878 // standard dump does this in Verbose and WizardMode 879 st->print(" #"); _type->dump_on(st); 880 } 881 if (!_depends_only_on_test) { 882 st->print(" (does not depend only on test)"); 883 } 884 } 885 #endif 886 887 #ifdef ASSERT 888 //----------------------------is_immutable_value------------------------------- 889 // Helper function to allow a raw load without control edge for some cases 890 bool LoadNode::is_immutable_value(Node* adr) { 891 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 892 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 893 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 894 in_bytes(JavaThread::osthread_offset()))); 895 } 896 #endif 897 898 //----------------------------LoadNode::make----------------------------------- 899 // Polymorphic factory method: 900 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) { 901 Compile* C = gvn.C; 902 903 // sanity check the alias category against the created node type 904 assert(!(adr_type->isa_oopptr() && 905 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 906 "use LoadKlassNode instead"); 907 assert(!(adr_type->isa_aryptr() && 908 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 909 "use LoadRangeNode instead"); 910 // Check control edge of raw loads 911 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 912 // oop will be recorded in oop map if load crosses safepoint 913 rt->isa_oopptr() || is_immutable_value(adr), 914 "raw memory operations should have control edge"); 915 switch (bt) { 916 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 917 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 918 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 919 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 920 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 921 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); 922 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); 923 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); 924 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 925 case T_OBJECT: 926 #ifdef _LP64 927 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 928 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency)); 929 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr()); 930 } else 931 #endif 932 { 933 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 934 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo, control_dependency); 935 } 936 } 937 ShouldNotReachHere(); 938 return (LoadNode*)NULL; 939 } 940 941 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) { 942 bool require_atomic = true; 943 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 944 } 945 946 LoadDNode* LoadDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) { 947 bool require_atomic = true; 948 return new (C) LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 949 } 950 951 952 953 //------------------------------hash------------------------------------------- 954 uint LoadNode::hash() const { 955 // unroll addition of interesting fields 956 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 957 } 958 959 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 960 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 961 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 962 bool is_stable_ary = FoldStableValues && 963 (tp != NULL) && (tp->isa_aryptr() != NULL) && 964 tp->isa_aryptr()->is_stable(); 965 966 return (eliminate_boxing && non_volatile) || is_stable_ary; 967 } 968 969 return false; 970 } 971 972 //---------------------------can_see_stored_value------------------------------ 973 // This routine exists to make sure this set of tests is done the same 974 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 975 // will change the graph shape in a way which makes memory alive twice at the 976 // same time (uses the Oracle model of aliasing), then some 977 // LoadXNode::Identity will fold things back to the equivalence-class model 978 // of aliasing. 979 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 980 Node* ld_adr = in(MemNode::Address); 981 intptr_t ld_off = 0; 982 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 983 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 984 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 985 // This is more general than load from boxing objects. 986 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 987 uint alias_idx = atp->index(); 988 bool final = !atp->is_rewritable(); 989 Node* result = NULL; 990 Node* current = st; 991 // Skip through chains of MemBarNodes checking the MergeMems for 992 // new states for the slice of this load. Stop once any other 993 // kind of node is encountered. Loads from final memory can skip 994 // through any kind of MemBar but normal loads shouldn't skip 995 // through MemBarAcquire since the could allow them to move out of 996 // a synchronized region. 997 while (current->is_Proj()) { 998 int opc = current->in(0)->Opcode(); 999 if ((final && (opc == Op_MemBarAcquire || 1000 opc == Op_MemBarAcquireLock || 1001 opc == Op_LoadFence)) || 1002 opc == Op_MemBarRelease || 1003 opc == Op_StoreFence || 1004 opc == Op_MemBarReleaseLock || 1005 opc == Op_MemBarCPUOrder) { 1006 Node* mem = current->in(0)->in(TypeFunc::Memory); 1007 if (mem->is_MergeMem()) { 1008 MergeMemNode* merge = mem->as_MergeMem(); 1009 Node* new_st = merge->memory_at(alias_idx); 1010 if (new_st == merge->base_memory()) { 1011 // Keep searching 1012 current = new_st; 1013 continue; 1014 } 1015 // Save the new memory state for the slice and fall through 1016 // to exit. 1017 result = new_st; 1018 } 1019 } 1020 break; 1021 } 1022 if (result != NULL) { 1023 st = result; 1024 } 1025 } 1026 1027 // Loop around twice in the case Load -> Initialize -> Store. 1028 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1029 for (int trip = 0; trip <= 1; trip++) { 1030 1031 if (st->is_Store()) { 1032 Node* st_adr = st->in(MemNode::Address); 1033 if (!phase->eqv(st_adr, ld_adr)) { 1034 // Try harder before giving up... Match raw and non-raw pointers. 1035 intptr_t st_off = 0; 1036 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 1037 if (alloc == NULL) return NULL; 1038 if (alloc != ld_alloc) return NULL; 1039 if (ld_off != st_off) return NULL; 1040 // At this point we have proven something like this setup: 1041 // A = Allocate(...) 1042 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 1043 // S = StoreQ(, AddP(, A.Parm , #Off), V) 1044 // (Actually, we haven't yet proven the Q's are the same.) 1045 // In other words, we are loading from a casted version of 1046 // the same pointer-and-offset that we stored to. 1047 // Thus, we are able to replace L by V. 1048 } 1049 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1050 if (store_Opcode() != st->Opcode()) 1051 return NULL; 1052 return st->in(MemNode::ValueIn); 1053 } 1054 1055 // A load from a freshly-created object always returns zero. 1056 // (This can happen after LoadNode::Ideal resets the load's memory input 1057 // to find_captured_store, which returned InitializeNode::zero_memory.) 1058 if (st->is_Proj() && st->in(0)->is_Allocate() && 1059 (st->in(0) == ld_alloc) && 1060 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1061 // return a zero value for the load's basic type 1062 // (This is one of the few places where a generic PhaseTransform 1063 // can create new nodes. Think of it as lazily manifesting 1064 // virtually pre-existing constants.) 1065 return phase->zerocon(memory_type()); 1066 } 1067 1068 // A load from an initialization barrier can match a captured store. 1069 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1070 InitializeNode* init = st->in(0)->as_Initialize(); 1071 AllocateNode* alloc = init->allocation(); 1072 if ((alloc != NULL) && (alloc == ld_alloc)) { 1073 // examine a captured store value 1074 st = init->find_captured_store(ld_off, memory_size(), phase); 1075 if (st != NULL) 1076 continue; // take one more trip around 1077 } 1078 } 1079 1080 // Load boxed value from result of valueOf() call is input parameter. 1081 if (this->is_Load() && ld_adr->is_AddP() && 1082 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1083 intptr_t ignore = 0; 1084 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1085 if (base != NULL && base->is_Proj() && 1086 base->as_Proj()->_con == TypeFunc::Parms && 1087 base->in(0)->is_CallStaticJava() && 1088 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1089 return base->in(0)->in(TypeFunc::Parms); 1090 } 1091 } 1092 1093 break; 1094 } 1095 1096 return NULL; 1097 } 1098 1099 //----------------------is_instance_field_load_with_local_phi------------------ 1100 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1101 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1102 in(Address)->is_AddP() ) { 1103 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1104 // Only instances and boxed values. 1105 if( t_oop != NULL && 1106 (t_oop->is_ptr_to_boxed_value() || 1107 t_oop->is_known_instance_field()) && 1108 t_oop->offset() != Type::OffsetBot && 1109 t_oop->offset() != Type::OffsetTop) { 1110 return true; 1111 } 1112 } 1113 return false; 1114 } 1115 1116 //------------------------------Identity--------------------------------------- 1117 // Loads are identity if previous store is to same address 1118 Node *LoadNode::Identity( PhaseTransform *phase ) { 1119 // If the previous store-maker is the right kind of Store, and the store is 1120 // to the same address, then we are equal to the value stored. 1121 Node* mem = in(Memory); 1122 Node* value = can_see_stored_value(mem, phase); 1123 if( value ) { 1124 // byte, short & char stores truncate naturally. 1125 // A load has to load the truncated value which requires 1126 // some sort of masking operation and that requires an 1127 // Ideal call instead of an Identity call. 1128 if (memory_size() < BytesPerInt) { 1129 // If the input to the store does not fit with the load's result type, 1130 // it must be truncated via an Ideal call. 1131 if (!phase->type(value)->higher_equal(phase->type(this))) 1132 return this; 1133 } 1134 // (This works even when value is a Con, but LoadNode::Value 1135 // usually runs first, producing the singleton type of the Con.) 1136 return value; 1137 } 1138 1139 // Search for an existing data phi which was generated before for the same 1140 // instance's field to avoid infinite generation of phis in a loop. 1141 Node *region = mem->in(0); 1142 if (is_instance_field_load_with_local_phi(region)) { 1143 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1144 int this_index = phase->C->get_alias_index(addr_t); 1145 int this_offset = addr_t->offset(); 1146 int this_iid = addr_t->instance_id(); 1147 if (!addr_t->is_known_instance() && 1148 addr_t->is_ptr_to_boxed_value()) { 1149 // Use _idx of address base (could be Phi node) for boxed values. 1150 intptr_t ignore = 0; 1151 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1152 this_iid = base->_idx; 1153 } 1154 const Type* this_type = bottom_type(); 1155 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1156 Node* phi = region->fast_out(i); 1157 if (phi->is_Phi() && phi != mem && 1158 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1159 return phi; 1160 } 1161 } 1162 } 1163 1164 return this; 1165 } 1166 1167 // We're loading from an object which has autobox behaviour. 1168 // If this object is result of a valueOf call we'll have a phi 1169 // merging a newly allocated object and a load from the cache. 1170 // We want to replace this load with the original incoming 1171 // argument to the valueOf call. 1172 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1173 assert(phase->C->eliminate_boxing(), "sanity"); 1174 intptr_t ignore = 0; 1175 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1176 if ((base == NULL) || base->is_Phi()) { 1177 // Push the loads from the phi that comes from valueOf up 1178 // through it to allow elimination of the loads and the recovery 1179 // of the original value. It is done in split_through_phi(). 1180 return NULL; 1181 } else if (base->is_Load() || 1182 base->is_DecodeN() && base->in(1)->is_Load()) { 1183 // Eliminate the load of boxed value for integer types from the cache 1184 // array by deriving the value from the index into the array. 1185 // Capture the offset of the load and then reverse the computation. 1186 1187 // Get LoadN node which loads a boxing object from 'cache' array. 1188 if (base->is_DecodeN()) { 1189 base = base->in(1); 1190 } 1191 if (!base->in(Address)->is_AddP()) { 1192 return NULL; // Complex address 1193 } 1194 AddPNode* address = base->in(Address)->as_AddP(); 1195 Node* cache_base = address->in(AddPNode::Base); 1196 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1197 // Get ConP node which is static 'cache' field. 1198 cache_base = cache_base->in(1); 1199 } 1200 if ((cache_base != NULL) && cache_base->is_Con()) { 1201 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1202 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1203 Node* elements[4]; 1204 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1205 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1206 if ((count > 0) && elements[0]->is_Con() && 1207 ((count == 1) || 1208 (count == 2) && elements[1]->Opcode() == Op_LShiftX && 1209 elements[1]->in(2) == phase->intcon(shift))) { 1210 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1211 // Fetch the box object cache[0] at the base of the array and get its value 1212 ciInstance* box = array->obj_at(0)->as_instance(); 1213 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1214 assert(ik->is_box_klass(), "sanity"); 1215 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1216 if (ik->nof_nonstatic_fields() == 1) { 1217 // This should be true nonstatic_field_at requires calling 1218 // nof_nonstatic_fields so check it anyway 1219 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1220 BasicType bt = c.basic_type(); 1221 // Only integer types have boxing cache. 1222 assert(bt == T_BOOLEAN || bt == T_CHAR || 1223 bt == T_BYTE || bt == T_SHORT || 1224 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt))); 1225 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1226 if (cache_low != (int)cache_low) { 1227 return NULL; // should not happen since cache is array indexed by value 1228 } 1229 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1230 if (offset != (int)offset) { 1231 return NULL; // should not happen since cache is array indexed by value 1232 } 1233 // Add up all the offsets making of the address of the load 1234 Node* result = elements[0]; 1235 for (int i = 1; i < count; i++) { 1236 result = phase->transform(new (phase->C) AddXNode(result, elements[i])); 1237 } 1238 // Remove the constant offset from the address and then 1239 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset))); 1240 // remove the scaling of the offset to recover the original index. 1241 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1242 // Peel the shift off directly but wrap it in a dummy node 1243 // since Ideal can't return existing nodes 1244 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0)); 1245 } else if (result->is_Add() && result->in(2)->is_Con() && 1246 result->in(1)->Opcode() == Op_LShiftX && 1247 result->in(1)->in(2) == phase->intcon(shift)) { 1248 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1249 // but for boxing cache access we know that X<<Z will not overflow 1250 // (there is range check) so we do this optimizatrion by hand here. 1251 Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift)); 1252 result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1253 } else { 1254 result = new (phase->C) RShiftXNode(result, phase->intcon(shift)); 1255 } 1256 #ifdef _LP64 1257 if (bt != T_LONG) { 1258 result = new (phase->C) ConvL2INode(phase->transform(result)); 1259 } 1260 #else 1261 if (bt == T_LONG) { 1262 result = new (phase->C) ConvI2LNode(phase->transform(result)); 1263 } 1264 #endif 1265 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1266 // Need to preserve unboxing load type if it is unsigned. 1267 switch(this->Opcode()) { 1268 case Op_LoadUB: 1269 result = new (phase->C) AndINode(phase->transform(result), phase->intcon(0xFF)); 1270 break; 1271 case Op_LoadUS: 1272 result = new (phase->C) AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1273 break; 1274 } 1275 return result; 1276 } 1277 } 1278 } 1279 } 1280 } 1281 return NULL; 1282 } 1283 1284 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1285 Node* region = phi->in(0); 1286 if (region == NULL) { 1287 return false; // Wait stable graph 1288 } 1289 uint cnt = phi->req(); 1290 for (uint i = 1; i < cnt; i++) { 1291 Node* rc = region->in(i); 1292 if (rc == NULL || phase->type(rc) == Type::TOP) 1293 return false; // Wait stable graph 1294 Node* in = phi->in(i); 1295 if (in == NULL || phase->type(in) == Type::TOP) 1296 return false; // Wait stable graph 1297 } 1298 return true; 1299 } 1300 //------------------------------split_through_phi------------------------------ 1301 // Split instance or boxed field load through Phi. 1302 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1303 Node* mem = in(Memory); 1304 Node* address = in(Address); 1305 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1306 1307 assert((t_oop != NULL) && 1308 (t_oop->is_known_instance_field() || 1309 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1310 1311 Compile* C = phase->C; 1312 intptr_t ignore = 0; 1313 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1314 bool base_is_phi = (base != NULL) && base->is_Phi(); 1315 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1316 (base != NULL) && (base == address->in(AddPNode::Base)) && 1317 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1318 1319 if (!((mem->is_Phi() || base_is_phi) && 1320 (load_boxed_values || t_oop->is_known_instance_field()))) { 1321 return NULL; // memory is not Phi 1322 } 1323 1324 if (mem->is_Phi()) { 1325 if (!stable_phi(mem->as_Phi(), phase)) { 1326 return NULL; // Wait stable graph 1327 } 1328 uint cnt = mem->req(); 1329 // Check for loop invariant memory. 1330 if (cnt == 3) { 1331 for (uint i = 1; i < cnt; i++) { 1332 Node* in = mem->in(i); 1333 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1334 if (m == mem) { 1335 set_req(Memory, mem->in(cnt - i)); 1336 return this; // made change 1337 } 1338 } 1339 } 1340 } 1341 if (base_is_phi) { 1342 if (!stable_phi(base->as_Phi(), phase)) { 1343 return NULL; // Wait stable graph 1344 } 1345 uint cnt = base->req(); 1346 // Check for loop invariant memory. 1347 if (cnt == 3) { 1348 for (uint i = 1; i < cnt; i++) { 1349 if (base->in(i) == base) { 1350 return NULL; // Wait stable graph 1351 } 1352 } 1353 } 1354 } 1355 1356 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1357 1358 // Split through Phi (see original code in loopopts.cpp). 1359 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1360 1361 // Do nothing here if Identity will find a value 1362 // (to avoid infinite chain of value phis generation). 1363 if (!phase->eqv(this, this->Identity(phase))) 1364 return NULL; 1365 1366 // Select Region to split through. 1367 Node* region; 1368 if (!base_is_phi) { 1369 assert(mem->is_Phi(), "sanity"); 1370 region = mem->in(0); 1371 // Skip if the region dominates some control edge of the address. 1372 if (!MemNode::all_controls_dominate(address, region)) 1373 return NULL; 1374 } else if (!mem->is_Phi()) { 1375 assert(base_is_phi, "sanity"); 1376 region = base->in(0); 1377 // Skip if the region dominates some control edge of the memory. 1378 if (!MemNode::all_controls_dominate(mem, region)) 1379 return NULL; 1380 } else if (base->in(0) != mem->in(0)) { 1381 assert(base_is_phi && mem->is_Phi(), "sanity"); 1382 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1383 region = base->in(0); 1384 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1385 region = mem->in(0); 1386 } else { 1387 return NULL; // complex graph 1388 } 1389 } else { 1390 assert(base->in(0) == mem->in(0), "sanity"); 1391 region = mem->in(0); 1392 } 1393 1394 const Type* this_type = this->bottom_type(); 1395 int this_index = C->get_alias_index(t_oop); 1396 int this_offset = t_oop->offset(); 1397 int this_iid = t_oop->instance_id(); 1398 if (!t_oop->is_known_instance() && load_boxed_values) { 1399 // Use _idx of address base for boxed values. 1400 this_iid = base->_idx; 1401 } 1402 PhaseIterGVN* igvn = phase->is_IterGVN(); 1403 Node* phi = new (C) PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1404 for (uint i = 1; i < region->req(); i++) { 1405 Node* x; 1406 Node* the_clone = NULL; 1407 if (region->in(i) == C->top()) { 1408 x = C->top(); // Dead path? Use a dead data op 1409 } else { 1410 x = this->clone(); // Else clone up the data op 1411 the_clone = x; // Remember for possible deletion. 1412 // Alter data node to use pre-phi inputs 1413 if (this->in(0) == region) { 1414 x->set_req(0, region->in(i)); 1415 } else { 1416 x->set_req(0, NULL); 1417 } 1418 if (mem->is_Phi() && (mem->in(0) == region)) { 1419 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1420 } 1421 if (address->is_Phi() && address->in(0) == region) { 1422 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1423 } 1424 if (base_is_phi && (base->in(0) == region)) { 1425 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1426 Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1427 x->set_req(Address, adr_x); 1428 } 1429 } 1430 // Check for a 'win' on some paths 1431 const Type *t = x->Value(igvn); 1432 1433 bool singleton = t->singleton(); 1434 1435 // See comments in PhaseIdealLoop::split_thru_phi(). 1436 if (singleton && t == Type::TOP) { 1437 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1438 } 1439 1440 if (singleton) { 1441 x = igvn->makecon(t); 1442 } else { 1443 // We now call Identity to try to simplify the cloned node. 1444 // Note that some Identity methods call phase->type(this). 1445 // Make sure that the type array is big enough for 1446 // our new node, even though we may throw the node away. 1447 // (This tweaking with igvn only works because x is a new node.) 1448 igvn->set_type(x, t); 1449 // If x is a TypeNode, capture any more-precise type permanently into Node 1450 // otherwise it will be not updated during igvn->transform since 1451 // igvn->type(x) is set to x->Value() already. 1452 x->raise_bottom_type(t); 1453 Node *y = x->Identity(igvn); 1454 if (y != x) { 1455 x = y; 1456 } else { 1457 y = igvn->hash_find_insert(x); 1458 if (y) { 1459 x = y; 1460 } else { 1461 // Else x is a new node we are keeping 1462 // We do not need register_new_node_with_optimizer 1463 // because set_type has already been called. 1464 igvn->_worklist.push(x); 1465 } 1466 } 1467 } 1468 if (x != the_clone && the_clone != NULL) { 1469 igvn->remove_dead_node(the_clone); 1470 } 1471 phi->set_req(i, x); 1472 } 1473 // Record Phi 1474 igvn->register_new_node_with_optimizer(phi); 1475 return phi; 1476 } 1477 1478 //------------------------------Ideal------------------------------------------ 1479 // If the load is from Field memory and the pointer is non-null, it might be possible to 1480 // zero out the control input. 1481 // If the offset is constant and the base is an object allocation, 1482 // try to hook me up to the exact initializing store. 1483 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1484 Node* p = MemNode::Ideal_common(phase, can_reshape); 1485 if (p) return (p == NodeSentinel) ? NULL : p; 1486 1487 Node* ctrl = in(MemNode::Control); 1488 Node* address = in(MemNode::Address); 1489 1490 // Skip up past a SafePoint control. Cannot do this for Stores because 1491 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1492 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1493 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1494 ctrl = ctrl->in(0); 1495 set_req(MemNode::Control,ctrl); 1496 } 1497 1498 intptr_t ignore = 0; 1499 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1500 if (base != NULL 1501 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1502 // Check for useless control edge in some common special cases 1503 if (in(MemNode::Control) != NULL 1504 && can_remove_control() 1505 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1506 && all_controls_dominate(base, phase->C->start())) { 1507 // A method-invariant, non-null address (constant or 'this' argument). 1508 set_req(MemNode::Control, NULL); 1509 } 1510 } 1511 1512 Node* mem = in(MemNode::Memory); 1513 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1514 1515 if (can_reshape && (addr_t != NULL)) { 1516 // try to optimize our memory input 1517 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1518 if (opt_mem != mem) { 1519 set_req(MemNode::Memory, opt_mem); 1520 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1521 return this; 1522 } 1523 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1524 if ((t_oop != NULL) && 1525 (t_oop->is_known_instance_field() || 1526 t_oop->is_ptr_to_boxed_value())) { 1527 PhaseIterGVN *igvn = phase->is_IterGVN(); 1528 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1529 // Delay this transformation until memory Phi is processed. 1530 phase->is_IterGVN()->_worklist.push(this); 1531 return NULL; 1532 } 1533 // Split instance field load through Phi. 1534 Node* result = split_through_phi(phase); 1535 if (result != NULL) return result; 1536 1537 if (t_oop->is_ptr_to_boxed_value()) { 1538 Node* result = eliminate_autobox(phase); 1539 if (result != NULL) return result; 1540 } 1541 } 1542 } 1543 1544 // Check for prior store with a different base or offset; make Load 1545 // independent. Skip through any number of them. Bail out if the stores 1546 // are in an endless dead cycle and report no progress. This is a key 1547 // transform for Reflection. However, if after skipping through the Stores 1548 // we can't then fold up against a prior store do NOT do the transform as 1549 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1550 // array memory alive twice: once for the hoisted Load and again after the 1551 // bypassed Store. This situation only works if EVERYBODY who does 1552 // anti-dependence work knows how to bypass. I.e. we need all 1553 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1554 // the alias index stuff. So instead, peek through Stores and IFF we can 1555 // fold up, do so. 1556 Node* prev_mem = find_previous_store(phase); 1557 // Steps (a), (b): Walk past independent stores to find an exact match. 1558 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1559 // (c) See if we can fold up on the spot, but don't fold up here. 1560 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1561 // just return a prior value, which is done by Identity calls. 1562 if (can_see_stored_value(prev_mem, phase)) { 1563 // Make ready for step (d): 1564 set_req(MemNode::Memory, prev_mem); 1565 return this; 1566 } 1567 } 1568 1569 return NULL; // No further progress 1570 } 1571 1572 // Helper to recognize certain Klass fields which are invariant across 1573 // some group of array types (e.g., int[] or all T[] where T < Object). 1574 const Type* 1575 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1576 ciKlass* klass) const { 1577 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1578 // The field is Klass::_modifier_flags. Return its (constant) value. 1579 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1580 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1581 return TypeInt::make(klass->modifier_flags()); 1582 } 1583 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1584 // The field is Klass::_access_flags. Return its (constant) value. 1585 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1586 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1587 return TypeInt::make(klass->access_flags()); 1588 } 1589 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1590 // The field is Klass::_layout_helper. Return its constant value if known. 1591 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1592 return TypeInt::make(klass->layout_helper()); 1593 } 1594 1595 // No match. 1596 return NULL; 1597 } 1598 1599 // Try to constant-fold a stable array element. 1600 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) { 1601 assert(ary->const_oop(), "array should be constant"); 1602 assert(ary->is_stable(), "array should be stable"); 1603 1604 // Decode the results of GraphKit::array_element_address. 1605 ciArray* aobj = ary->const_oop()->as_array(); 1606 ciConstant con = aobj->element_value_by_offset(off); 1607 1608 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) { 1609 const Type* con_type = Type::make_from_constant(con); 1610 if (con_type != NULL) { 1611 if (con_type->isa_aryptr()) { 1612 // Join with the array element type, in case it is also stable. 1613 int dim = ary->stable_dimension(); 1614 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1); 1615 } 1616 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) { 1617 con_type = con_type->make_narrowoop(); 1618 } 1619 #ifndef PRODUCT 1620 if (TraceIterativeGVN) { 1621 tty->print("FoldStableValues: array element [off=%d]: con_type=", off); 1622 con_type->dump(); tty->cr(); 1623 } 1624 #endif //PRODUCT 1625 return con_type; 1626 } 1627 } 1628 return NULL; 1629 } 1630 1631 //------------------------------Value----------------------------------------- 1632 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1633 // Either input is TOP ==> the result is TOP 1634 Node* mem = in(MemNode::Memory); 1635 const Type *t1 = phase->type(mem); 1636 if (t1 == Type::TOP) return Type::TOP; 1637 Node* adr = in(MemNode::Address); 1638 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1639 if (tp == NULL || tp->empty()) return Type::TOP; 1640 int off = tp->offset(); 1641 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1642 Compile* C = phase->C; 1643 1644 // Try to guess loaded type from pointer type 1645 if (tp->isa_aryptr()) { 1646 const TypeAryPtr* ary = tp->is_aryptr(); 1647 const Type* t = ary->elem(); 1648 1649 // Determine whether the reference is beyond the header or not, by comparing 1650 // the offset against the offset of the start of the array's data. 1651 // Different array types begin at slightly different offsets (12 vs. 16). 1652 // We choose T_BYTE as an example base type that is least restrictive 1653 // as to alignment, which will therefore produce the smallest 1654 // possible base offset. 1655 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1656 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1657 1658 // Try to constant-fold a stable array element. 1659 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) { 1660 // Make sure the reference is not into the header and the offset is constant 1661 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1662 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type()); 1663 if (con_type != NULL) { 1664 return con_type; 1665 } 1666 } 1667 } 1668 1669 // Don't do this for integer types. There is only potential profit if 1670 // the element type t is lower than _type; that is, for int types, if _type is 1671 // more restrictive than t. This only happens here if one is short and the other 1672 // char (both 16 bits), and in those cases we've made an intentional decision 1673 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1674 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1675 // 1676 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1677 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1678 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1679 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1680 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1681 // In fact, that could have been the original type of p1, and p1 could have 1682 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1683 // expression (LShiftL quux 3) independently optimized to the constant 8. 1684 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1685 && (_type->isa_vect() == NULL) 1686 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1687 // t might actually be lower than _type, if _type is a unique 1688 // concrete subclass of abstract class t. 1689 if (off_beyond_header) { // is the offset beyond the header? 1690 const Type* jt = t->join_speculative(_type); 1691 // In any case, do not allow the join, per se, to empty out the type. 1692 if (jt->empty() && !t->empty()) { 1693 // This can happen if a interface-typed array narrows to a class type. 1694 jt = _type; 1695 } 1696 #ifdef ASSERT 1697 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1698 // The pointers in the autobox arrays are always non-null 1699 Node* base = adr->in(AddPNode::Base); 1700 if ((base != NULL) && base->is_DecodeN()) { 1701 // Get LoadN node which loads IntegerCache.cache field 1702 base = base->in(1); 1703 } 1704 if ((base != NULL) && base->is_Con()) { 1705 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1706 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1707 // It could be narrow oop 1708 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1709 } 1710 } 1711 } 1712 #endif 1713 return jt; 1714 } 1715 } 1716 } else if (tp->base() == Type::InstPtr) { 1717 ciEnv* env = C->env(); 1718 const TypeInstPtr* tinst = tp->is_instptr(); 1719 ciKlass* klass = tinst->klass(); 1720 assert( off != Type::OffsetBot || 1721 // arrays can be cast to Objects 1722 tp->is_oopptr()->klass()->is_java_lang_Object() || 1723 // unsafe field access may not have a constant offset 1724 C->has_unsafe_access(), 1725 "Field accesses must be precise" ); 1726 // For oop loads, we expect the _type to be precise 1727 if (klass == env->String_klass() && 1728 adr->is_AddP() && off != Type::OffsetBot) { 1729 // For constant Strings treat the final fields as compile time constants. 1730 Node* base = adr->in(AddPNode::Base); 1731 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1732 if (t != NULL && t->singleton()) { 1733 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1734 if (field != NULL && field->is_final()) { 1735 ciObject* string = t->const_oop(); 1736 ciConstant constant = string->as_instance()->field_value(field); 1737 if (constant.basic_type() == T_INT) { 1738 return TypeInt::make(constant.as_int()); 1739 } else if (constant.basic_type() == T_ARRAY) { 1740 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1741 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1742 } else { 1743 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1744 } 1745 } 1746 } 1747 } 1748 } 1749 // Optimizations for constant objects 1750 ciObject* const_oop = tinst->const_oop(); 1751 if (const_oop != NULL) { 1752 // For constant Boxed value treat the target field as a compile time constant. 1753 if (tinst->is_ptr_to_boxed_value()) { 1754 return tinst->get_const_boxed_value(); 1755 } else 1756 // For constant CallSites treat the target field as a compile time constant. 1757 if (const_oop->is_call_site()) { 1758 ciCallSite* call_site = const_oop->as_call_site(); 1759 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1760 if (field != NULL && field->is_call_site_target()) { 1761 ciMethodHandle* target = call_site->get_target(); 1762 if (target != NULL) { // just in case 1763 ciConstant constant(T_OBJECT, target); 1764 const Type* t; 1765 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1766 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1767 } else { 1768 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1769 } 1770 // Add a dependence for invalidation of the optimization. 1771 if (!call_site->is_constant_call_site()) { 1772 C->dependencies()->assert_call_site_target_value(call_site, target); 1773 } 1774 return t; 1775 } 1776 } 1777 } 1778 } 1779 } else if (tp->base() == Type::KlassPtr) { 1780 assert( off != Type::OffsetBot || 1781 // arrays can be cast to Objects 1782 tp->is_klassptr()->klass()->is_java_lang_Object() || 1783 // also allow array-loading from the primary supertype 1784 // array during subtype checks 1785 Opcode() == Op_LoadKlass, 1786 "Field accesses must be precise" ); 1787 // For klass/static loads, we expect the _type to be precise 1788 } 1789 1790 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1791 if (tkls != NULL && !StressReflectiveCode) { 1792 ciKlass* klass = tkls->klass(); 1793 if (klass->is_loaded() && tkls->klass_is_exact()) { 1794 // We are loading a field from a Klass metaobject whose identity 1795 // is known at compile time (the type is "exact" or "precise"). 1796 // Check for fields we know are maintained as constants by the VM. 1797 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1798 // The field is Klass::_super_check_offset. Return its (constant) value. 1799 // (Folds up type checking code.) 1800 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1801 return TypeInt::make(klass->super_check_offset()); 1802 } 1803 // Compute index into primary_supers array 1804 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1805 // Check for overflowing; use unsigned compare to handle the negative case. 1806 if( depth < ciKlass::primary_super_limit() ) { 1807 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1808 // (Folds up type checking code.) 1809 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1810 ciKlass *ss = klass->super_of_depth(depth); 1811 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1812 } 1813 const Type* aift = load_array_final_field(tkls, klass); 1814 if (aift != NULL) return aift; 1815 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset()) 1816 && klass->is_array_klass()) { 1817 // The field is ArrayKlass::_component_mirror. Return its (constant) value. 1818 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1819 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1820 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1821 } 1822 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1823 // The field is Klass::_java_mirror. Return its (constant) value. 1824 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1825 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1826 return TypeInstPtr::make(klass->java_mirror()); 1827 } 1828 } 1829 1830 // We can still check if we are loading from the primary_supers array at a 1831 // shallow enough depth. Even though the klass is not exact, entries less 1832 // than or equal to its super depth are correct. 1833 if (klass->is_loaded() ) { 1834 ciType *inner = klass; 1835 while( inner->is_obj_array_klass() ) 1836 inner = inner->as_obj_array_klass()->base_element_type(); 1837 if( inner->is_instance_klass() && 1838 !inner->as_instance_klass()->flags().is_interface() ) { 1839 // Compute index into primary_supers array 1840 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1841 // Check for overflowing; use unsigned compare to handle the negative case. 1842 if( depth < ciKlass::primary_super_limit() && 1843 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1844 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1845 // (Folds up type checking code.) 1846 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1847 ciKlass *ss = klass->super_of_depth(depth); 1848 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1849 } 1850 } 1851 } 1852 1853 // If the type is enough to determine that the thing is not an array, 1854 // we can give the layout_helper a positive interval type. 1855 // This will help short-circuit some reflective code. 1856 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1857 && !klass->is_array_klass() // not directly typed as an array 1858 && !klass->is_interface() // specifically not Serializable & Cloneable 1859 && !klass->is_java_lang_Object() // not the supertype of all T[] 1860 ) { 1861 // Note: When interfaces are reliable, we can narrow the interface 1862 // test to (klass != Serializable && klass != Cloneable). 1863 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1864 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1865 // The key property of this type is that it folds up tests 1866 // for array-ness, since it proves that the layout_helper is positive. 1867 // Thus, a generic value like the basic object layout helper works fine. 1868 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1869 } 1870 } 1871 1872 // If we are loading from a freshly-allocated object, produce a zero, 1873 // if the load is provably beyond the header of the object. 1874 // (Also allow a variable load from a fresh array to produce zero.) 1875 const TypeOopPtr *tinst = tp->isa_oopptr(); 1876 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1877 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1878 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1879 Node* value = can_see_stored_value(mem,phase); 1880 if (value != NULL && value->is_Con()) { 1881 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1882 return value->bottom_type(); 1883 } 1884 } 1885 1886 if (is_instance) { 1887 // If we have an instance type and our memory input is the 1888 // programs's initial memory state, there is no matching store, 1889 // so just return a zero of the appropriate type 1890 Node *mem = in(MemNode::Memory); 1891 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1892 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1893 return Type::get_zero_type(_type->basic_type()); 1894 } 1895 } 1896 return _type; 1897 } 1898 1899 //------------------------------match_edge------------------------------------- 1900 // Do we Match on this edge index or not? Match only the address. 1901 uint LoadNode::match_edge(uint idx) const { 1902 return idx == MemNode::Address; 1903 } 1904 1905 //--------------------------LoadBNode::Ideal-------------------------------------- 1906 // 1907 // If the previous store is to the same address as this load, 1908 // and the value stored was larger than a byte, replace this load 1909 // with the value stored truncated to a byte. If no truncation is 1910 // needed, the replacement is done in LoadNode::Identity(). 1911 // 1912 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1913 Node* mem = in(MemNode::Memory); 1914 Node* value = can_see_stored_value(mem,phase); 1915 if( value && !phase->type(value)->higher_equal( _type ) ) { 1916 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) ); 1917 return new (phase->C) RShiftINode(result, phase->intcon(24)); 1918 } 1919 // Identity call will handle the case where truncation is not needed. 1920 return LoadNode::Ideal(phase, can_reshape); 1921 } 1922 1923 const Type* LoadBNode::Value(PhaseTransform *phase) const { 1924 Node* mem = in(MemNode::Memory); 1925 Node* value = can_see_stored_value(mem,phase); 1926 if (value != NULL && value->is_Con() && 1927 !value->bottom_type()->higher_equal(_type)) { 1928 // If the input to the store does not fit with the load's result type, 1929 // it must be truncated. We can't delay until Ideal call since 1930 // a singleton Value is needed for split_thru_phi optimization. 1931 int con = value->get_int(); 1932 return TypeInt::make((con << 24) >> 24); 1933 } 1934 return LoadNode::Value(phase); 1935 } 1936 1937 //--------------------------LoadUBNode::Ideal------------------------------------- 1938 // 1939 // If the previous store is to the same address as this load, 1940 // and the value stored was larger than a byte, replace this load 1941 // with the value stored truncated to a byte. If no truncation is 1942 // needed, the replacement is done in LoadNode::Identity(). 1943 // 1944 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1945 Node* mem = in(MemNode::Memory); 1946 Node* value = can_see_stored_value(mem, phase); 1947 if (value && !phase->type(value)->higher_equal(_type)) 1948 return new (phase->C) AndINode(value, phase->intcon(0xFF)); 1949 // Identity call will handle the case where truncation is not needed. 1950 return LoadNode::Ideal(phase, can_reshape); 1951 } 1952 1953 const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1954 Node* mem = in(MemNode::Memory); 1955 Node* value = can_see_stored_value(mem,phase); 1956 if (value != NULL && value->is_Con() && 1957 !value->bottom_type()->higher_equal(_type)) { 1958 // If the input to the store does not fit with the load's result type, 1959 // it must be truncated. We can't delay until Ideal call since 1960 // a singleton Value is needed for split_thru_phi optimization. 1961 int con = value->get_int(); 1962 return TypeInt::make(con & 0xFF); 1963 } 1964 return LoadNode::Value(phase); 1965 } 1966 1967 //--------------------------LoadUSNode::Ideal------------------------------------- 1968 // 1969 // If the previous store is to the same address as this load, 1970 // and the value stored was larger than a char, replace this load 1971 // with the value stored truncated to a char. If no truncation is 1972 // needed, the replacement is done in LoadNode::Identity(). 1973 // 1974 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1975 Node* mem = in(MemNode::Memory); 1976 Node* value = can_see_stored_value(mem,phase); 1977 if( value && !phase->type(value)->higher_equal( _type ) ) 1978 return new (phase->C) AndINode(value,phase->intcon(0xFFFF)); 1979 // Identity call will handle the case where truncation is not needed. 1980 return LoadNode::Ideal(phase, can_reshape); 1981 } 1982 1983 const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1984 Node* mem = in(MemNode::Memory); 1985 Node* value = can_see_stored_value(mem,phase); 1986 if (value != NULL && value->is_Con() && 1987 !value->bottom_type()->higher_equal(_type)) { 1988 // If the input to the store does not fit with the load's result type, 1989 // it must be truncated. We can't delay until Ideal call since 1990 // a singleton Value is needed for split_thru_phi optimization. 1991 int con = value->get_int(); 1992 return TypeInt::make(con & 0xFFFF); 1993 } 1994 return LoadNode::Value(phase); 1995 } 1996 1997 //--------------------------LoadSNode::Ideal-------------------------------------- 1998 // 1999 // If the previous store is to the same address as this load, 2000 // and the value stored was larger than a short, replace this load 2001 // with the value stored truncated to a short. If no truncation is 2002 // needed, the replacement is done in LoadNode::Identity(). 2003 // 2004 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2005 Node* mem = in(MemNode::Memory); 2006 Node* value = can_see_stored_value(mem,phase); 2007 if( value && !phase->type(value)->higher_equal( _type ) ) { 2008 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) ); 2009 return new (phase->C) RShiftINode(result, phase->intcon(16)); 2010 } 2011 // Identity call will handle the case where truncation is not needed. 2012 return LoadNode::Ideal(phase, can_reshape); 2013 } 2014 2015 const Type* LoadSNode::Value(PhaseTransform *phase) const { 2016 Node* mem = in(MemNode::Memory); 2017 Node* value = can_see_stored_value(mem,phase); 2018 if (value != NULL && value->is_Con() && 2019 !value->bottom_type()->higher_equal(_type)) { 2020 // If the input to the store does not fit with the load's result type, 2021 // it must be truncated. We can't delay until Ideal call since 2022 // a singleton Value is needed for split_thru_phi optimization. 2023 int con = value->get_int(); 2024 return TypeInt::make((con << 16) >> 16); 2025 } 2026 return LoadNode::Value(phase); 2027 } 2028 2029 //============================================================================= 2030 //----------------------------LoadKlassNode::make------------------------------ 2031 // Polymorphic factory method: 2032 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk) { 2033 Compile* C = gvn.C; 2034 // sanity check the alias category against the created node type 2035 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2036 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2037 #ifdef _LP64 2038 if (adr_type->is_ptr_to_narrowklass()) { 2039 assert(UseCompressedClassPointers, "no compressed klasses"); 2040 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2041 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2042 } 2043 #endif 2044 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2045 return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2046 } 2047 2048 //------------------------------Value------------------------------------------ 2049 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 2050 return klass_value_common(phase); 2051 } 2052 2053 // In most cases, LoadKlassNode does not have the control input set. If the control 2054 // input is set, it must not be removed (by LoadNode::Ideal()). 2055 bool LoadKlassNode::can_remove_control() const { 2056 return false; 2057 } 2058 2059 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 2060 // Either input is TOP ==> the result is TOP 2061 const Type *t1 = phase->type( in(MemNode::Memory) ); 2062 if (t1 == Type::TOP) return Type::TOP; 2063 Node *adr = in(MemNode::Address); 2064 const Type *t2 = phase->type( adr ); 2065 if (t2 == Type::TOP) return Type::TOP; 2066 const TypePtr *tp = t2->is_ptr(); 2067 if (TypePtr::above_centerline(tp->ptr()) || 2068 tp->ptr() == TypePtr::Null) return Type::TOP; 2069 2070 // Return a more precise klass, if possible 2071 const TypeInstPtr *tinst = tp->isa_instptr(); 2072 if (tinst != NULL) { 2073 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2074 int offset = tinst->offset(); 2075 if (ik == phase->C->env()->Class_klass() 2076 && (offset == java_lang_Class::klass_offset_in_bytes() || 2077 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2078 // We are loading a special hidden field from a Class mirror object, 2079 // the field which points to the VM's Klass metaobject. 2080 ciType* t = tinst->java_mirror_type(); 2081 // java_mirror_type returns non-null for compile-time Class constants. 2082 if (t != NULL) { 2083 // constant oop => constant klass 2084 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2085 if (t->is_void()) { 2086 // We cannot create a void array. Since void is a primitive type return null 2087 // klass. Users of this result need to do a null check on the returned klass. 2088 return TypePtr::NULL_PTR; 2089 } 2090 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2091 } 2092 if (!t->is_klass()) { 2093 // a primitive Class (e.g., int.class) has NULL for a klass field 2094 return TypePtr::NULL_PTR; 2095 } 2096 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2097 return TypeKlassPtr::make(t->as_klass()); 2098 } 2099 // non-constant mirror, so we can't tell what's going on 2100 } 2101 if( !ik->is_loaded() ) 2102 return _type; // Bail out if not loaded 2103 if (offset == oopDesc::klass_offset_in_bytes()) { 2104 if (tinst->klass_is_exact()) { 2105 return TypeKlassPtr::make(ik); 2106 } 2107 // See if we can become precise: no subklasses and no interface 2108 // (Note: We need to support verified interfaces.) 2109 if (!ik->is_interface() && !ik->has_subklass()) { 2110 //assert(!UseExactTypes, "this code should be useless with exact types"); 2111 // Add a dependence; if any subclass added we need to recompile 2112 if (!ik->is_final()) { 2113 // %%% should use stronger assert_unique_concrete_subtype instead 2114 phase->C->dependencies()->assert_leaf_type(ik); 2115 } 2116 // Return precise klass 2117 return TypeKlassPtr::make(ik); 2118 } 2119 2120 // Return root of possible klass 2121 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2122 } 2123 } 2124 2125 // Check for loading klass from an array 2126 const TypeAryPtr *tary = tp->isa_aryptr(); 2127 if( tary != NULL ) { 2128 ciKlass *tary_klass = tary->klass(); 2129 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2130 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2131 if (tary->klass_is_exact()) { 2132 return TypeKlassPtr::make(tary_klass); 2133 } 2134 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2135 // If the klass is an object array, we defer the question to the 2136 // array component klass. 2137 if( ak->is_obj_array_klass() ) { 2138 assert( ak->is_loaded(), "" ); 2139 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2140 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2141 ciInstanceKlass* ik = base_k->as_instance_klass(); 2142 // See if we can become precise: no subklasses and no interface 2143 if (!ik->is_interface() && !ik->has_subklass()) { 2144 //assert(!UseExactTypes, "this code should be useless with exact types"); 2145 // Add a dependence; if any subclass added we need to recompile 2146 if (!ik->is_final()) { 2147 phase->C->dependencies()->assert_leaf_type(ik); 2148 } 2149 // Return precise array klass 2150 return TypeKlassPtr::make(ak); 2151 } 2152 } 2153 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2154 } else { // Found a type-array? 2155 //assert(!UseExactTypes, "this code should be useless with exact types"); 2156 assert( ak->is_type_array_klass(), "" ); 2157 return TypeKlassPtr::make(ak); // These are always precise 2158 } 2159 } 2160 } 2161 2162 // Check for loading klass from an array klass 2163 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2164 if (tkls != NULL && !StressReflectiveCode) { 2165 ciKlass* klass = tkls->klass(); 2166 if( !klass->is_loaded() ) 2167 return _type; // Bail out if not loaded 2168 if( klass->is_obj_array_klass() && 2169 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2170 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2171 // // Always returning precise element type is incorrect, 2172 // // e.g., element type could be object and array may contain strings 2173 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2174 2175 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2176 // according to the element type's subclassing. 2177 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2178 } 2179 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2180 tkls->offset() == in_bytes(Klass::super_offset())) { 2181 ciKlass* sup = klass->as_instance_klass()->super(); 2182 // The field is Klass::_super. Return its (constant) value. 2183 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2184 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2185 } 2186 } 2187 2188 // Bailout case 2189 return LoadNode::Value(phase); 2190 } 2191 2192 //------------------------------Identity--------------------------------------- 2193 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2194 // Also feed through the klass in Allocate(...klass...)._klass. 2195 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2196 return klass_identity_common(phase); 2197 } 2198 2199 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2200 Node* x = LoadNode::Identity(phase); 2201 if (x != this) return x; 2202 2203 // Take apart the address into an oop and and offset. 2204 // Return 'this' if we cannot. 2205 Node* adr = in(MemNode::Address); 2206 intptr_t offset = 0; 2207 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2208 if (base == NULL) return this; 2209 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2210 if (toop == NULL) return this; 2211 2212 // We can fetch the klass directly through an AllocateNode. 2213 // This works even if the klass is not constant (clone or newArray). 2214 if (offset == oopDesc::klass_offset_in_bytes()) { 2215 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2216 if (allocated_klass != NULL) { 2217 return allocated_klass; 2218 } 2219 } 2220 2221 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2222 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass. 2223 // See inline_native_Class_query for occurrences of these patterns. 2224 // Java Example: x.getClass().isAssignableFrom(y) 2225 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 2226 // 2227 // This improves reflective code, often making the Class 2228 // mirror go completely dead. (Current exception: Class 2229 // mirrors may appear in debug info, but we could clean them out by 2230 // introducing a new debug info operator for Klass*.java_mirror). 2231 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2232 && (offset == java_lang_Class::klass_offset_in_bytes() || 2233 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2234 // We are loading a special hidden field from a Class mirror, 2235 // the field which points to its Klass or ArrayKlass metaobject. 2236 if (base->is_Load()) { 2237 Node* adr2 = base->in(MemNode::Address); 2238 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2239 if (tkls != NULL && !tkls->empty() 2240 && (tkls->klass()->is_instance_klass() || 2241 tkls->klass()->is_array_klass()) 2242 && adr2->is_AddP() 2243 ) { 2244 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2245 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2246 mirror_field = in_bytes(ArrayKlass::component_mirror_offset()); 2247 } 2248 if (tkls->offset() == mirror_field) { 2249 return adr2->in(AddPNode::Base); 2250 } 2251 } 2252 } 2253 } 2254 2255 return this; 2256 } 2257 2258 2259 //------------------------------Value------------------------------------------ 2260 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2261 const Type *t = klass_value_common(phase); 2262 if (t == Type::TOP) 2263 return t; 2264 2265 return t->make_narrowklass(); 2266 } 2267 2268 //------------------------------Identity--------------------------------------- 2269 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2270 // Also feed through the klass in Allocate(...klass...)._klass. 2271 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2272 Node *x = klass_identity_common(phase); 2273 2274 const Type *t = phase->type( x ); 2275 if( t == Type::TOP ) return x; 2276 if( t->isa_narrowklass()) return x; 2277 assert (!t->isa_narrowoop(), "no narrow oop here"); 2278 2279 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass())); 2280 } 2281 2282 //------------------------------Value----------------------------------------- 2283 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2284 // Either input is TOP ==> the result is TOP 2285 const Type *t1 = phase->type( in(MemNode::Memory) ); 2286 if( t1 == Type::TOP ) return Type::TOP; 2287 Node *adr = in(MemNode::Address); 2288 const Type *t2 = phase->type( adr ); 2289 if( t2 == Type::TOP ) return Type::TOP; 2290 const TypePtr *tp = t2->is_ptr(); 2291 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2292 const TypeAryPtr *tap = tp->isa_aryptr(); 2293 if( !tap ) return _type; 2294 return tap->size(); 2295 } 2296 2297 //-------------------------------Ideal--------------------------------------- 2298 // Feed through the length in AllocateArray(...length...)._length. 2299 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2300 Node* p = MemNode::Ideal_common(phase, can_reshape); 2301 if (p) return (p == NodeSentinel) ? NULL : p; 2302 2303 // Take apart the address into an oop and and offset. 2304 // Return 'this' if we cannot. 2305 Node* adr = in(MemNode::Address); 2306 intptr_t offset = 0; 2307 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2308 if (base == NULL) return NULL; 2309 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2310 if (tary == NULL) return NULL; 2311 2312 // We can fetch the length directly through an AllocateArrayNode. 2313 // This works even if the length is not constant (clone or newArray). 2314 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2315 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2316 if (alloc != NULL) { 2317 Node* allocated_length = alloc->Ideal_length(); 2318 Node* len = alloc->make_ideal_length(tary, phase); 2319 if (allocated_length != len) { 2320 // New CastII improves on this. 2321 return len; 2322 } 2323 } 2324 } 2325 2326 return NULL; 2327 } 2328 2329 //------------------------------Identity--------------------------------------- 2330 // Feed through the length in AllocateArray(...length...)._length. 2331 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2332 Node* x = LoadINode::Identity(phase); 2333 if (x != this) return x; 2334 2335 // Take apart the address into an oop and and offset. 2336 // Return 'this' if we cannot. 2337 Node* adr = in(MemNode::Address); 2338 intptr_t offset = 0; 2339 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2340 if (base == NULL) return this; 2341 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2342 if (tary == NULL) return this; 2343 2344 // We can fetch the length directly through an AllocateArrayNode. 2345 // This works even if the length is not constant (clone or newArray). 2346 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2347 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2348 if (alloc != NULL) { 2349 Node* allocated_length = alloc->Ideal_length(); 2350 // Do not allow make_ideal_length to allocate a CastII node. 2351 Node* len = alloc->make_ideal_length(tary, phase, false); 2352 if (allocated_length == len) { 2353 // Return allocated_length only if it would not be improved by a CastII. 2354 return allocated_length; 2355 } 2356 } 2357 } 2358 2359 return this; 2360 2361 } 2362 2363 //============================================================================= 2364 //---------------------------StoreNode::make----------------------------------- 2365 // Polymorphic factory method: 2366 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2367 assert((mo == unordered || mo == release), "unexpected"); 2368 Compile* C = gvn.C; 2369 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2370 ctl != NULL, "raw memory operations should have control edge"); 2371 2372 switch (bt) { 2373 case T_BOOLEAN: val = gvn.transform(new (C) AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2374 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo); 2375 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo); 2376 case T_CHAR: 2377 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo); 2378 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo); 2379 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo); 2380 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo); 2381 case T_METADATA: 2382 case T_ADDRESS: 2383 case T_OBJECT: 2384 #ifdef _LP64 2385 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2386 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2387 return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo); 2388 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2389 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2390 adr->bottom_type()->isa_rawptr())) { 2391 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2392 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2393 } 2394 #endif 2395 { 2396 return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo); 2397 } 2398 } 2399 ShouldNotReachHere(); 2400 return (StoreNode*)NULL; 2401 } 2402 2403 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2404 bool require_atomic = true; 2405 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2406 } 2407 2408 StoreDNode* StoreDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2409 bool require_atomic = true; 2410 return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2411 } 2412 2413 2414 //--------------------------bottom_type---------------------------------------- 2415 const Type *StoreNode::bottom_type() const { 2416 return Type::MEMORY; 2417 } 2418 2419 //------------------------------hash------------------------------------------- 2420 uint StoreNode::hash() const { 2421 // unroll addition of interesting fields 2422 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2423 2424 // Since they are not commoned, do not hash them: 2425 return NO_HASH; 2426 } 2427 2428 //------------------------------Ideal------------------------------------------ 2429 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2430 // When a store immediately follows a relevant allocation/initialization, 2431 // try to capture it into the initialization, or hoist it above. 2432 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2433 Node* p = MemNode::Ideal_common(phase, can_reshape); 2434 if (p) return (p == NodeSentinel) ? NULL : p; 2435 2436 Node* mem = in(MemNode::Memory); 2437 Node* address = in(MemNode::Address); 2438 2439 // Back-to-back stores to same address? Fold em up. Generally 2440 // unsafe if I have intervening uses... Also disallowed for StoreCM 2441 // since they must follow each StoreP operation. Redundant StoreCMs 2442 // are eliminated just before matching in final_graph_reshape. 2443 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && 2444 mem->Opcode() != Op_StoreCM) { 2445 // Looking at a dead closed cycle of memory? 2446 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2447 2448 assert(Opcode() == mem->Opcode() || 2449 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2450 "no mismatched stores, except on raw memory"); 2451 2452 if (mem->outcnt() == 1 && // check for intervening uses 2453 mem->as_Store()->memory_size() <= this->memory_size()) { 2454 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2455 // For example, 'mem' might be the final state at a conditional return. 2456 // Or, 'mem' might be used by some node which is live at the same time 2457 // 'this' is live, which might be unschedulable. So, require exactly 2458 // ONE user, the 'this' store, until such time as we clone 'mem' for 2459 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2460 if (can_reshape) { // (%%% is this an anachronism?) 2461 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2462 phase->is_IterGVN()); 2463 } else { 2464 // It's OK to do this in the parser, since DU info is always accurate, 2465 // and the parser always refers to nodes via SafePointNode maps. 2466 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2467 } 2468 return this; 2469 } 2470 } 2471 2472 // Capture an unaliased, unconditional, simple store into an initializer. 2473 // Or, if it is independent of the allocation, hoist it above the allocation. 2474 if (ReduceFieldZeroing && /*can_reshape &&*/ 2475 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2476 InitializeNode* init = mem->in(0)->as_Initialize(); 2477 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2478 if (offset > 0) { 2479 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2480 // If the InitializeNode captured me, it made a raw copy of me, 2481 // and I need to disappear. 2482 if (moved != NULL) { 2483 // %%% hack to ensure that Ideal returns a new node: 2484 mem = MergeMemNode::make(phase->C, mem); 2485 return mem; // fold me away 2486 } 2487 } 2488 } 2489 2490 return NULL; // No further progress 2491 } 2492 2493 //------------------------------Value----------------------------------------- 2494 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2495 // Either input is TOP ==> the result is TOP 2496 const Type *t1 = phase->type( in(MemNode::Memory) ); 2497 if( t1 == Type::TOP ) return Type::TOP; 2498 const Type *t2 = phase->type( in(MemNode::Address) ); 2499 if( t2 == Type::TOP ) return Type::TOP; 2500 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2501 if( t3 == Type::TOP ) return Type::TOP; 2502 return Type::MEMORY; 2503 } 2504 2505 //------------------------------Identity--------------------------------------- 2506 // Remove redundant stores: 2507 // Store(m, p, Load(m, p)) changes to m. 2508 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2509 Node *StoreNode::Identity( PhaseTransform *phase ) { 2510 Node* mem = in(MemNode::Memory); 2511 Node* adr = in(MemNode::Address); 2512 Node* val = in(MemNode::ValueIn); 2513 2514 // Load then Store? Then the Store is useless 2515 if (val->is_Load() && 2516 val->in(MemNode::Address)->eqv_uncast(adr) && 2517 val->in(MemNode::Memory )->eqv_uncast(mem) && 2518 val->as_Load()->store_Opcode() == Opcode()) { 2519 return mem; 2520 } 2521 2522 // Two stores in a row of the same value? 2523 if (mem->is_Store() && 2524 mem->in(MemNode::Address)->eqv_uncast(adr) && 2525 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2526 mem->Opcode() == Opcode()) { 2527 return mem; 2528 } 2529 2530 // Store of zero anywhere into a freshly-allocated object? 2531 // Then the store is useless. 2532 // (It must already have been captured by the InitializeNode.) 2533 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2534 // a newly allocated object is already all-zeroes everywhere 2535 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2536 return mem; 2537 } 2538 2539 // the store may also apply to zero-bits in an earlier object 2540 Node* prev_mem = find_previous_store(phase); 2541 // Steps (a), (b): Walk past independent stores to find an exact match. 2542 if (prev_mem != NULL) { 2543 Node* prev_val = can_see_stored_value(prev_mem, phase); 2544 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2545 // prev_val and val might differ by a cast; it would be good 2546 // to keep the more informative of the two. 2547 return mem; 2548 } 2549 } 2550 } 2551 2552 return this; 2553 } 2554 2555 //------------------------------match_edge------------------------------------- 2556 // Do we Match on this edge index or not? Match only memory & value 2557 uint StoreNode::match_edge(uint idx) const { 2558 return idx == MemNode::Address || idx == MemNode::ValueIn; 2559 } 2560 2561 //------------------------------cmp-------------------------------------------- 2562 // Do not common stores up together. They generally have to be split 2563 // back up anyways, so do not bother. 2564 uint StoreNode::cmp( const Node &n ) const { 2565 return (&n == this); // Always fail except on self 2566 } 2567 2568 //------------------------------Ideal_masked_input----------------------------- 2569 // Check for a useless mask before a partial-word store 2570 // (StoreB ... (AndI valIn conIa) ) 2571 // If (conIa & mask == mask) this simplifies to 2572 // (StoreB ... (valIn) ) 2573 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2574 Node *val = in(MemNode::ValueIn); 2575 if( val->Opcode() == Op_AndI ) { 2576 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2577 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2578 set_req(MemNode::ValueIn, val->in(1)); 2579 return this; 2580 } 2581 } 2582 return NULL; 2583 } 2584 2585 2586 //------------------------------Ideal_sign_extended_input---------------------- 2587 // Check for useless sign-extension before a partial-word store 2588 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2589 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2590 // (StoreB ... (valIn) ) 2591 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2592 Node *val = in(MemNode::ValueIn); 2593 if( val->Opcode() == Op_RShiftI ) { 2594 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2595 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2596 Node *shl = val->in(1); 2597 if( shl->Opcode() == Op_LShiftI ) { 2598 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2599 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2600 set_req(MemNode::ValueIn, shl->in(1)); 2601 return this; 2602 } 2603 } 2604 } 2605 } 2606 return NULL; 2607 } 2608 2609 //------------------------------value_never_loaded----------------------------------- 2610 // Determine whether there are any possible loads of the value stored. 2611 // For simplicity, we actually check if there are any loads from the 2612 // address stored to, not just for loads of the value stored by this node. 2613 // 2614 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2615 Node *adr = in(Address); 2616 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2617 if (adr_oop == NULL) 2618 return false; 2619 if (!adr_oop->is_known_instance_field()) 2620 return false; // if not a distinct instance, there may be aliases of the address 2621 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2622 Node *use = adr->fast_out(i); 2623 int opc = use->Opcode(); 2624 if (use->is_Load() || use->is_LoadStore()) { 2625 return false; 2626 } 2627 } 2628 return true; 2629 } 2630 2631 //============================================================================= 2632 //------------------------------Ideal------------------------------------------ 2633 // If the store is from an AND mask that leaves the low bits untouched, then 2634 // we can skip the AND operation. If the store is from a sign-extension 2635 // (a left shift, then right shift) we can skip both. 2636 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2637 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2638 if( progress != NULL ) return progress; 2639 2640 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2641 if( progress != NULL ) return progress; 2642 2643 // Finally check the default case 2644 return StoreNode::Ideal(phase, can_reshape); 2645 } 2646 2647 //============================================================================= 2648 //------------------------------Ideal------------------------------------------ 2649 // If the store is from an AND mask that leaves the low bits untouched, then 2650 // we can skip the AND operation 2651 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2652 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2653 if( progress != NULL ) return progress; 2654 2655 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2656 if( progress != NULL ) return progress; 2657 2658 // Finally check the default case 2659 return StoreNode::Ideal(phase, can_reshape); 2660 } 2661 2662 //============================================================================= 2663 //------------------------------Identity--------------------------------------- 2664 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2665 // No need to card mark when storing a null ptr 2666 Node* my_store = in(MemNode::OopStore); 2667 if (my_store->is_Store()) { 2668 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2669 if( t1 == TypePtr::NULL_PTR ) { 2670 return in(MemNode::Memory); 2671 } 2672 } 2673 return this; 2674 } 2675 2676 //============================================================================= 2677 //------------------------------Ideal--------------------------------------- 2678 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2679 Node* progress = StoreNode::Ideal(phase, can_reshape); 2680 if (progress != NULL) return progress; 2681 2682 Node* my_store = in(MemNode::OopStore); 2683 if (my_store->is_MergeMem()) { 2684 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2685 set_req(MemNode::OopStore, mem); 2686 return this; 2687 } 2688 2689 return NULL; 2690 } 2691 2692 //------------------------------Value----------------------------------------- 2693 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2694 // Either input is TOP ==> the result is TOP 2695 const Type *t = phase->type( in(MemNode::Memory) ); 2696 if( t == Type::TOP ) return Type::TOP; 2697 t = phase->type( in(MemNode::Address) ); 2698 if( t == Type::TOP ) return Type::TOP; 2699 t = phase->type( in(MemNode::ValueIn) ); 2700 if( t == Type::TOP ) return Type::TOP; 2701 // If extra input is TOP ==> the result is TOP 2702 t = phase->type( in(MemNode::OopStore) ); 2703 if( t == Type::TOP ) return Type::TOP; 2704 2705 return StoreNode::Value( phase ); 2706 } 2707 2708 2709 //============================================================================= 2710 //----------------------------------SCMemProjNode------------------------------ 2711 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2712 { 2713 return bottom_type(); 2714 } 2715 2716 //============================================================================= 2717 //----------------------------------LoadStoreNode------------------------------ 2718 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2719 : Node(required), 2720 _type(rt), 2721 _adr_type(at) 2722 { 2723 init_req(MemNode::Control, c ); 2724 init_req(MemNode::Memory , mem); 2725 init_req(MemNode::Address, adr); 2726 init_req(MemNode::ValueIn, val); 2727 init_class_id(Class_LoadStore); 2728 } 2729 2730 uint LoadStoreNode::ideal_reg() const { 2731 return _type->ideal_reg(); 2732 } 2733 2734 bool LoadStoreNode::result_not_used() const { 2735 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2736 Node *x = fast_out(i); 2737 if (x->Opcode() == Op_SCMemProj) continue; 2738 return false; 2739 } 2740 return true; 2741 } 2742 2743 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2744 2745 //============================================================================= 2746 //----------------------------------LoadStoreConditionalNode-------------------- 2747 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2748 init_req(ExpectedIn, ex ); 2749 } 2750 2751 //============================================================================= 2752 //-------------------------------adr_type-------------------------------------- 2753 // Do we Match on this edge index or not? Do not match memory 2754 const TypePtr* ClearArrayNode::adr_type() const { 2755 Node *adr = in(3); 2756 return MemNode::calculate_adr_type(adr->bottom_type()); 2757 } 2758 2759 //------------------------------match_edge------------------------------------- 2760 // Do we Match on this edge index or not? Do not match memory 2761 uint ClearArrayNode::match_edge(uint idx) const { 2762 return idx > 1; 2763 } 2764 2765 //------------------------------Identity--------------------------------------- 2766 // Clearing a zero length array does nothing 2767 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2768 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2769 } 2770 2771 //------------------------------Idealize--------------------------------------- 2772 // Clearing a short array is faster with stores 2773 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2774 const int unit = BytesPerLong; 2775 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2776 if (!t) return NULL; 2777 if (!t->is_con()) return NULL; 2778 intptr_t raw_count = t->get_con(); 2779 intptr_t size = raw_count; 2780 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2781 // Clearing nothing uses the Identity call. 2782 // Negative clears are possible on dead ClearArrays 2783 // (see jck test stmt114.stmt11402.val). 2784 if (size <= 0 || size % unit != 0) return NULL; 2785 intptr_t count = size / unit; 2786 // Length too long; use fast hardware clear 2787 if (size > Matcher::init_array_short_size) return NULL; 2788 Node *mem = in(1); 2789 if( phase->type(mem)==Type::TOP ) return NULL; 2790 Node *adr = in(3); 2791 const Type* at = phase->type(adr); 2792 if( at==Type::TOP ) return NULL; 2793 const TypePtr* atp = at->isa_ptr(); 2794 // adjust atp to be the correct array element address type 2795 if (atp == NULL) atp = TypePtr::BOTTOM; 2796 else atp = atp->add_offset(Type::OffsetBot); 2797 // Get base for derived pointer purposes 2798 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2799 Node *base = adr->in(1); 2800 2801 Node *zero = phase->makecon(TypeLong::ZERO); 2802 Node *off = phase->MakeConX(BytesPerLong); 2803 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2804 count--; 2805 while( count-- ) { 2806 mem = phase->transform(mem); 2807 adr = phase->transform(new (phase->C) AddPNode(base,adr,off)); 2808 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2809 } 2810 return mem; 2811 } 2812 2813 //----------------------------step_through---------------------------------- 2814 // Return allocation input memory edge if it is different instance 2815 // or itself if it is the one we are looking for. 2816 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2817 Node* n = *np; 2818 assert(n->is_ClearArray(), "sanity"); 2819 intptr_t offset; 2820 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2821 // This method is called only before Allocate nodes are expanded during 2822 // macro nodes expansion. Before that ClearArray nodes are only generated 2823 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2824 assert(alloc != NULL, "should have allocation"); 2825 if (alloc->_idx == instance_id) { 2826 // Can not bypass initialization of the instance we are looking for. 2827 return false; 2828 } 2829 // Otherwise skip it. 2830 InitializeNode* init = alloc->initialization(); 2831 if (init != NULL) 2832 *np = init->in(TypeFunc::Memory); 2833 else 2834 *np = alloc->in(TypeFunc::Memory); 2835 return true; 2836 } 2837 2838 //----------------------------clear_memory------------------------------------- 2839 // Generate code to initialize object storage to zero. 2840 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2841 intptr_t start_offset, 2842 Node* end_offset, 2843 PhaseGVN* phase) { 2844 Compile* C = phase->C; 2845 intptr_t offset = start_offset; 2846 2847 int unit = BytesPerLong; 2848 if ((offset % unit) != 0) { 2849 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset)); 2850 adr = phase->transform(adr); 2851 const TypePtr* atp = TypeRawPtr::BOTTOM; 2852 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2853 mem = phase->transform(mem); 2854 offset += BytesPerInt; 2855 } 2856 assert((offset % unit) == 0, ""); 2857 2858 // Initialize the remaining stuff, if any, with a ClearArray. 2859 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2860 } 2861 2862 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2863 Node* start_offset, 2864 Node* end_offset, 2865 PhaseGVN* phase) { 2866 if (start_offset == end_offset) { 2867 // nothing to do 2868 return mem; 2869 } 2870 2871 Compile* C = phase->C; 2872 int unit = BytesPerLong; 2873 Node* zbase = start_offset; 2874 Node* zend = end_offset; 2875 2876 // Scale to the unit required by the CPU: 2877 if (!Matcher::init_array_count_is_in_bytes) { 2878 Node* shift = phase->intcon(exact_log2(unit)); 2879 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) ); 2880 zend = phase->transform( new(C) URShiftXNode(zend, shift) ); 2881 } 2882 2883 // Bulk clear double-words 2884 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) ); 2885 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) ); 2886 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr); 2887 return phase->transform(mem); 2888 } 2889 2890 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2891 intptr_t start_offset, 2892 intptr_t end_offset, 2893 PhaseGVN* phase) { 2894 if (start_offset == end_offset) { 2895 // nothing to do 2896 return mem; 2897 } 2898 2899 Compile* C = phase->C; 2900 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2901 intptr_t done_offset = end_offset; 2902 if ((done_offset % BytesPerLong) != 0) { 2903 done_offset -= BytesPerInt; 2904 } 2905 if (done_offset > start_offset) { 2906 mem = clear_memory(ctl, mem, dest, 2907 start_offset, phase->MakeConX(done_offset), phase); 2908 } 2909 if (done_offset < end_offset) { // emit the final 32-bit store 2910 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2911 adr = phase->transform(adr); 2912 const TypePtr* atp = TypeRawPtr::BOTTOM; 2913 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2914 mem = phase->transform(mem); 2915 done_offset += BytesPerInt; 2916 } 2917 assert(done_offset == end_offset, ""); 2918 return mem; 2919 } 2920 2921 //============================================================================= 2922 // Do not match memory edge. 2923 uint StrIntrinsicNode::match_edge(uint idx) const { 2924 return idx == 2 || idx == 3; 2925 } 2926 2927 //------------------------------Ideal------------------------------------------ 2928 // Return a node which is more "ideal" than the current node. Strip out 2929 // control copies 2930 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2931 if (remove_dead_region(phase, can_reshape)) return this; 2932 // Don't bother trying to transform a dead node 2933 if (in(0) && in(0)->is_top()) return NULL; 2934 2935 if (can_reshape) { 2936 Node* mem = phase->transform(in(MemNode::Memory)); 2937 // If transformed to a MergeMem, get the desired slice 2938 uint alias_idx = phase->C->get_alias_index(adr_type()); 2939 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; 2940 if (mem != in(MemNode::Memory)) { 2941 set_req(MemNode::Memory, mem); 2942 return this; 2943 } 2944 } 2945 return NULL; 2946 } 2947 2948 //------------------------------Value------------------------------------------ 2949 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { 2950 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; 2951 return bottom_type(); 2952 } 2953 2954 //============================================================================= 2955 //------------------------------match_edge------------------------------------- 2956 // Do not match memory edge 2957 uint EncodeISOArrayNode::match_edge(uint idx) const { 2958 return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len) 2959 } 2960 2961 //------------------------------Ideal------------------------------------------ 2962 // Return a node which is more "ideal" than the current node. Strip out 2963 // control copies 2964 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2965 return remove_dead_region(phase, can_reshape) ? this : NULL; 2966 } 2967 2968 //------------------------------Value------------------------------------------ 2969 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const { 2970 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; 2971 return bottom_type(); 2972 } 2973 2974 //============================================================================= 2975 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2976 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2977 _adr_type(C->get_adr_type(alias_idx)) 2978 { 2979 init_class_id(Class_MemBar); 2980 Node* top = C->top(); 2981 init_req(TypeFunc::I_O,top); 2982 init_req(TypeFunc::FramePtr,top); 2983 init_req(TypeFunc::ReturnAdr,top); 2984 if (precedent != NULL) 2985 init_req(TypeFunc::Parms, precedent); 2986 } 2987 2988 //------------------------------cmp-------------------------------------------- 2989 uint MemBarNode::hash() const { return NO_HASH; } 2990 uint MemBarNode::cmp( const Node &n ) const { 2991 return (&n == this); // Always fail except on self 2992 } 2993 2994 //------------------------------make------------------------------------------- 2995 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2996 switch (opcode) { 2997 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn); 2998 case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn); 2999 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn); 3000 case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn); 3001 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn); 3002 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn); 3003 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn); 3004 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn); 3005 case Op_Initialize: return new(C) InitializeNode(C, atp, pn); 3006 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn); 3007 default: ShouldNotReachHere(); return NULL; 3008 } 3009 } 3010 3011 //------------------------------Ideal------------------------------------------ 3012 // Return a node which is more "ideal" than the current node. Strip out 3013 // control copies 3014 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3015 if (remove_dead_region(phase, can_reshape)) return this; 3016 // Don't bother trying to transform a dead node 3017 if (in(0) && in(0)->is_top()) { 3018 return NULL; 3019 } 3020 3021 // Eliminate volatile MemBars for scalar replaced objects. 3022 if (can_reshape && req() == (Precedent+1)) { 3023 bool eliminate = false; 3024 int opc = Opcode(); 3025 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3026 // Volatile field loads and stores. 3027 Node* my_mem = in(MemBarNode::Precedent); 3028 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3029 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3030 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3031 // replace this Precedent (decodeN) with the Load instead. 3032 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3033 Node* load_node = my_mem->in(1); 3034 set_req(MemBarNode::Precedent, load_node); 3035 phase->is_IterGVN()->_worklist.push(my_mem); 3036 my_mem = load_node; 3037 } else { 3038 assert(my_mem->unique_out() == this, "sanity"); 3039 del_req(Precedent); 3040 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3041 my_mem = NULL; 3042 } 3043 } 3044 if (my_mem != NULL && my_mem->is_Mem()) { 3045 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3046 // Check for scalar replaced object reference. 3047 if( t_oop != NULL && t_oop->is_known_instance_field() && 3048 t_oop->offset() != Type::OffsetBot && 3049 t_oop->offset() != Type::OffsetTop) { 3050 eliminate = true; 3051 } 3052 } 3053 } else if (opc == Op_MemBarRelease) { 3054 // Final field stores. 3055 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3056 if ((alloc != NULL) && alloc->is_Allocate() && 3057 alloc->as_Allocate()->_is_non_escaping) { 3058 // The allocated object does not escape. 3059 eliminate = true; 3060 } 3061 } 3062 if (eliminate) { 3063 // Replace MemBar projections by its inputs. 3064 PhaseIterGVN* igvn = phase->is_IterGVN(); 3065 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3066 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3067 // Must return either the original node (now dead) or a new node 3068 // (Do not return a top here, since that would break the uniqueness of top.) 3069 return new (phase->C) ConINode(TypeInt::ZERO); 3070 } 3071 } 3072 return NULL; 3073 } 3074 3075 //------------------------------Value------------------------------------------ 3076 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 3077 if( !in(0) ) return Type::TOP; 3078 if( phase->type(in(0)) == Type::TOP ) 3079 return Type::TOP; 3080 return TypeTuple::MEMBAR; 3081 } 3082 3083 //------------------------------match------------------------------------------ 3084 // Construct projections for memory. 3085 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3086 switch (proj->_con) { 3087 case TypeFunc::Control: 3088 case TypeFunc::Memory: 3089 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3090 } 3091 ShouldNotReachHere(); 3092 return NULL; 3093 } 3094 3095 //===========================InitializeNode==================================== 3096 // SUMMARY: 3097 // This node acts as a memory barrier on raw memory, after some raw stores. 3098 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3099 // The Initialize can 'capture' suitably constrained stores as raw inits. 3100 // It can coalesce related raw stores into larger units (called 'tiles'). 3101 // It can avoid zeroing new storage for memory units which have raw inits. 3102 // At macro-expansion, it is marked 'complete', and does not optimize further. 3103 // 3104 // EXAMPLE: 3105 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3106 // ctl = incoming control; mem* = incoming memory 3107 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3108 // First allocate uninitialized memory and fill in the header: 3109 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3110 // ctl := alloc.Control; mem* := alloc.Memory* 3111 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3112 // Then initialize to zero the non-header parts of the raw memory block: 3113 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3114 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3115 // After the initialize node executes, the object is ready for service: 3116 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3117 // Suppose its body is immediately initialized as {1,2}: 3118 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3119 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3120 // mem.SLICE(#short[*]) := store2 3121 // 3122 // DETAILS: 3123 // An InitializeNode collects and isolates object initialization after 3124 // an AllocateNode and before the next possible safepoint. As a 3125 // memory barrier (MemBarNode), it keeps critical stores from drifting 3126 // down past any safepoint or any publication of the allocation. 3127 // Before this barrier, a newly-allocated object may have uninitialized bits. 3128 // After this barrier, it may be treated as a real oop, and GC is allowed. 3129 // 3130 // The semantics of the InitializeNode include an implicit zeroing of 3131 // the new object from object header to the end of the object. 3132 // (The object header and end are determined by the AllocateNode.) 3133 // 3134 // Certain stores may be added as direct inputs to the InitializeNode. 3135 // These stores must update raw memory, and they must be to addresses 3136 // derived from the raw address produced by AllocateNode, and with 3137 // a constant offset. They must be ordered by increasing offset. 3138 // The first one is at in(RawStores), the last at in(req()-1). 3139 // Unlike most memory operations, they are not linked in a chain, 3140 // but are displayed in parallel as users of the rawmem output of 3141 // the allocation. 3142 // 3143 // (See comments in InitializeNode::capture_store, which continue 3144 // the example given above.) 3145 // 3146 // When the associated Allocate is macro-expanded, the InitializeNode 3147 // may be rewritten to optimize collected stores. A ClearArrayNode 3148 // may also be created at that point to represent any required zeroing. 3149 // The InitializeNode is then marked 'complete', prohibiting further 3150 // capturing of nearby memory operations. 3151 // 3152 // During macro-expansion, all captured initializations which store 3153 // constant values of 32 bits or smaller are coalesced (if advantageous) 3154 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3155 // initialized in fewer memory operations. Memory words which are 3156 // covered by neither tiles nor non-constant stores are pre-zeroed 3157 // by explicit stores of zero. (The code shape happens to do all 3158 // zeroing first, then all other stores, with both sequences occurring 3159 // in order of ascending offsets.) 3160 // 3161 // Alternatively, code may be inserted between an AllocateNode and its 3162 // InitializeNode, to perform arbitrary initialization of the new object. 3163 // E.g., the object copying intrinsics insert complex data transfers here. 3164 // The initialization must then be marked as 'complete' disable the 3165 // built-in zeroing semantics and the collection of initializing stores. 3166 // 3167 // While an InitializeNode is incomplete, reads from the memory state 3168 // produced by it are optimizable if they match the control edge and 3169 // new oop address associated with the allocation/initialization. 3170 // They return a stored value (if the offset matches) or else zero. 3171 // A write to the memory state, if it matches control and address, 3172 // and if it is to a constant offset, may be 'captured' by the 3173 // InitializeNode. It is cloned as a raw memory operation and rewired 3174 // inside the initialization, to the raw oop produced by the allocation. 3175 // Operations on addresses which are provably distinct (e.g., to 3176 // other AllocateNodes) are allowed to bypass the initialization. 3177 // 3178 // The effect of all this is to consolidate object initialization 3179 // (both arrays and non-arrays, both piecewise and bulk) into a 3180 // single location, where it can be optimized as a unit. 3181 // 3182 // Only stores with an offset less than TrackedInitializationLimit words 3183 // will be considered for capture by an InitializeNode. This puts a 3184 // reasonable limit on the complexity of optimized initializations. 3185 3186 //---------------------------InitializeNode------------------------------------ 3187 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3188 : _is_complete(Incomplete), _does_not_escape(false), 3189 MemBarNode(C, adr_type, rawoop) 3190 { 3191 init_class_id(Class_Initialize); 3192 3193 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3194 assert(in(RawAddress) == rawoop, "proper init"); 3195 // Note: allocation() can be NULL, for secondary initialization barriers 3196 } 3197 3198 // Since this node is not matched, it will be processed by the 3199 // register allocator. Declare that there are no constraints 3200 // on the allocation of the RawAddress edge. 3201 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3202 // This edge should be set to top, by the set_complete. But be conservative. 3203 if (idx == InitializeNode::RawAddress) 3204 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3205 return RegMask::Empty; 3206 } 3207 3208 Node* InitializeNode::memory(uint alias_idx) { 3209 Node* mem = in(Memory); 3210 if (mem->is_MergeMem()) { 3211 return mem->as_MergeMem()->memory_at(alias_idx); 3212 } else { 3213 // incoming raw memory is not split 3214 return mem; 3215 } 3216 } 3217 3218 bool InitializeNode::is_non_zero() { 3219 if (is_complete()) return false; 3220 remove_extra_zeroes(); 3221 return (req() > RawStores); 3222 } 3223 3224 void InitializeNode::set_complete(PhaseGVN* phase) { 3225 assert(!is_complete(), "caller responsibility"); 3226 _is_complete = Complete; 3227 3228 // After this node is complete, it contains a bunch of 3229 // raw-memory initializations. There is no need for 3230 // it to have anything to do with non-raw memory effects. 3231 // Therefore, tell all non-raw users to re-optimize themselves, 3232 // after skipping the memory effects of this initialization. 3233 PhaseIterGVN* igvn = phase->is_IterGVN(); 3234 if (igvn) igvn->add_users_to_worklist(this); 3235 } 3236 3237 // convenience function 3238 // return false if the init contains any stores already 3239 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3240 InitializeNode* init = initialization(); 3241 if (init == NULL || init->is_complete()) return false; 3242 init->remove_extra_zeroes(); 3243 // for now, if this allocation has already collected any inits, bail: 3244 if (init->is_non_zero()) return false; 3245 init->set_complete(phase); 3246 return true; 3247 } 3248 3249 void InitializeNode::remove_extra_zeroes() { 3250 if (req() == RawStores) return; 3251 Node* zmem = zero_memory(); 3252 uint fill = RawStores; 3253 for (uint i = fill; i < req(); i++) { 3254 Node* n = in(i); 3255 if (n->is_top() || n == zmem) continue; // skip 3256 if (fill < i) set_req(fill, n); // compact 3257 ++fill; 3258 } 3259 // delete any empty spaces created: 3260 while (fill < req()) { 3261 del_req(fill); 3262 } 3263 } 3264 3265 // Helper for remembering which stores go with which offsets. 3266 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3267 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3268 intptr_t offset = -1; 3269 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3270 phase, offset); 3271 if (base == NULL) return -1; // something is dead, 3272 if (offset < 0) return -1; // dead, dead 3273 return offset; 3274 } 3275 3276 // Helper for proving that an initialization expression is 3277 // "simple enough" to be folded into an object initialization. 3278 // Attempts to prove that a store's initial value 'n' can be captured 3279 // within the initialization without creating a vicious cycle, such as: 3280 // { Foo p = new Foo(); p.next = p; } 3281 // True for constants and parameters and small combinations thereof. 3282 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3283 if (n == NULL) return true; // (can this really happen?) 3284 if (n->is_Proj()) n = n->in(0); 3285 if (n == this) return false; // found a cycle 3286 if (n->is_Con()) return true; 3287 if (n->is_Start()) return true; // params, etc., are OK 3288 if (n->is_Root()) return true; // even better 3289 3290 Node* ctl = n->in(0); 3291 if (ctl != NULL && !ctl->is_top()) { 3292 if (ctl->is_Proj()) ctl = ctl->in(0); 3293 if (ctl == this) return false; 3294 3295 // If we already know that the enclosing memory op is pinned right after 3296 // the init, then any control flow that the store has picked up 3297 // must have preceded the init, or else be equal to the init. 3298 // Even after loop optimizations (which might change control edges) 3299 // a store is never pinned *before* the availability of its inputs. 3300 if (!MemNode::all_controls_dominate(n, this)) 3301 return false; // failed to prove a good control 3302 } 3303 3304 // Check data edges for possible dependencies on 'this'. 3305 if ((count += 1) > 20) return false; // complexity limit 3306 for (uint i = 1; i < n->req(); i++) { 3307 Node* m = n->in(i); 3308 if (m == NULL || m == n || m->is_top()) continue; 3309 uint first_i = n->find_edge(m); 3310 if (i != first_i) continue; // process duplicate edge just once 3311 if (!detect_init_independence(m, count)) { 3312 return false; 3313 } 3314 } 3315 3316 return true; 3317 } 3318 3319 // Here are all the checks a Store must pass before it can be moved into 3320 // an initialization. Returns zero if a check fails. 3321 // On success, returns the (constant) offset to which the store applies, 3322 // within the initialized memory. 3323 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3324 const int FAIL = 0; 3325 if (st->req() != MemNode::ValueIn + 1) 3326 return FAIL; // an inscrutable StoreNode (card mark?) 3327 Node* ctl = st->in(MemNode::Control); 3328 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3329 return FAIL; // must be unconditional after the initialization 3330 Node* mem = st->in(MemNode::Memory); 3331 if (!(mem->is_Proj() && mem->in(0) == this)) 3332 return FAIL; // must not be preceded by other stores 3333 Node* adr = st->in(MemNode::Address); 3334 intptr_t offset; 3335 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3336 if (alloc == NULL) 3337 return FAIL; // inscrutable address 3338 if (alloc != allocation()) 3339 return FAIL; // wrong allocation! (store needs to float up) 3340 Node* val = st->in(MemNode::ValueIn); 3341 int complexity_count = 0; 3342 if (!detect_init_independence(val, complexity_count)) 3343 return FAIL; // stored value must be 'simple enough' 3344 3345 // The Store can be captured only if nothing after the allocation 3346 // and before the Store is using the memory location that the store 3347 // overwrites. 3348 bool failed = false; 3349 // If is_complete_with_arraycopy() is true the shape of the graph is 3350 // well defined and is safe so no need for extra checks. 3351 if (!is_complete_with_arraycopy()) { 3352 // We are going to look at each use of the memory state following 3353 // the allocation to make sure nothing reads the memory that the 3354 // Store writes. 3355 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3356 int alias_idx = phase->C->get_alias_index(t_adr); 3357 ResourceMark rm; 3358 Unique_Node_List mems; 3359 mems.push(mem); 3360 Node* unique_merge = NULL; 3361 for (uint next = 0; next < mems.size(); ++next) { 3362 Node *m = mems.at(next); 3363 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3364 Node *n = m->fast_out(j); 3365 if (n->outcnt() == 0) { 3366 continue; 3367 } 3368 if (n == st) { 3369 continue; 3370 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3371 // If the control of this use is different from the control 3372 // of the Store which is right after the InitializeNode then 3373 // this node cannot be between the InitializeNode and the 3374 // Store. 3375 continue; 3376 } else if (n->is_MergeMem()) { 3377 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3378 // We can hit a MergeMemNode (that will likely go away 3379 // later) that is a direct use of the memory state 3380 // following the InitializeNode on the same slice as the 3381 // store node that we'd like to capture. We need to check 3382 // the uses of the MergeMemNode. 3383 mems.push(n); 3384 } 3385 } else if (n->is_Mem()) { 3386 Node* other_adr = n->in(MemNode::Address); 3387 if (other_adr == adr) { 3388 failed = true; 3389 break; 3390 } else { 3391 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3392 if (other_t_adr != NULL) { 3393 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3394 if (other_alias_idx == alias_idx) { 3395 // A load from the same memory slice as the store right 3396 // after the InitializeNode. We check the control of the 3397 // object/array that is loaded from. If it's the same as 3398 // the store control then we cannot capture the store. 3399 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3400 Node* base = other_adr; 3401 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); 3402 base = base->in(AddPNode::Base); 3403 if (base != NULL) { 3404 base = base->uncast(); 3405 if (base->is_Proj() && base->in(0) == alloc) { 3406 failed = true; 3407 break; 3408 } 3409 } 3410 } 3411 } 3412 } 3413 } else { 3414 failed = true; 3415 break; 3416 } 3417 } 3418 } 3419 } 3420 if (failed) { 3421 if (!can_reshape) { 3422 // We decided we couldn't capture the store during parsing. We 3423 // should try again during the next IGVN once the graph is 3424 // cleaner. 3425 phase->C->record_for_igvn(st); 3426 } 3427 return FAIL; 3428 } 3429 3430 return offset; // success 3431 } 3432 3433 // Find the captured store in(i) which corresponds to the range 3434 // [start..start+size) in the initialized object. 3435 // If there is one, return its index i. If there isn't, return the 3436 // negative of the index where it should be inserted. 3437 // Return 0 if the queried range overlaps an initialization boundary 3438 // or if dead code is encountered. 3439 // If size_in_bytes is zero, do not bother with overlap checks. 3440 int InitializeNode::captured_store_insertion_point(intptr_t start, 3441 int size_in_bytes, 3442 PhaseTransform* phase) { 3443 const int FAIL = 0, MAX_STORE = BytesPerLong; 3444 3445 if (is_complete()) 3446 return FAIL; // arraycopy got here first; punt 3447 3448 assert(allocation() != NULL, "must be present"); 3449 3450 // no negatives, no header fields: 3451 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3452 3453 // after a certain size, we bail out on tracking all the stores: 3454 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3455 if (start >= ti_limit) return FAIL; 3456 3457 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3458 if (i >= limit) return -(int)i; // not found; here is where to put it 3459 3460 Node* st = in(i); 3461 intptr_t st_off = get_store_offset(st, phase); 3462 if (st_off < 0) { 3463 if (st != zero_memory()) { 3464 return FAIL; // bail out if there is dead garbage 3465 } 3466 } else if (st_off > start) { 3467 // ...we are done, since stores are ordered 3468 if (st_off < start + size_in_bytes) { 3469 return FAIL; // the next store overlaps 3470 } 3471 return -(int)i; // not found; here is where to put it 3472 } else if (st_off < start) { 3473 if (size_in_bytes != 0 && 3474 start < st_off + MAX_STORE && 3475 start < st_off + st->as_Store()->memory_size()) { 3476 return FAIL; // the previous store overlaps 3477 } 3478 } else { 3479 if (size_in_bytes != 0 && 3480 st->as_Store()->memory_size() != size_in_bytes) { 3481 return FAIL; // mismatched store size 3482 } 3483 return i; 3484 } 3485 3486 ++i; 3487 } 3488 } 3489 3490 // Look for a captured store which initializes at the offset 'start' 3491 // with the given size. If there is no such store, and no other 3492 // initialization interferes, then return zero_memory (the memory 3493 // projection of the AllocateNode). 3494 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3495 PhaseTransform* phase) { 3496 assert(stores_are_sane(phase), ""); 3497 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3498 if (i == 0) { 3499 return NULL; // something is dead 3500 } else if (i < 0) { 3501 return zero_memory(); // just primordial zero bits here 3502 } else { 3503 Node* st = in(i); // here is the store at this position 3504 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3505 return st; 3506 } 3507 } 3508 3509 // Create, as a raw pointer, an address within my new object at 'offset'. 3510 Node* InitializeNode::make_raw_address(intptr_t offset, 3511 PhaseTransform* phase) { 3512 Node* addr = in(RawAddress); 3513 if (offset != 0) { 3514 Compile* C = phase->C; 3515 addr = phase->transform( new (C) AddPNode(C->top(), addr, 3516 phase->MakeConX(offset)) ); 3517 } 3518 return addr; 3519 } 3520 3521 // Clone the given store, converting it into a raw store 3522 // initializing a field or element of my new object. 3523 // Caller is responsible for retiring the original store, 3524 // with subsume_node or the like. 3525 // 3526 // From the example above InitializeNode::InitializeNode, 3527 // here are the old stores to be captured: 3528 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3529 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3530 // 3531 // Here is the changed code; note the extra edges on init: 3532 // alloc = (Allocate ...) 3533 // rawoop = alloc.RawAddress 3534 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3535 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3536 // init = (Initialize alloc.Control alloc.Memory rawoop 3537 // rawstore1 rawstore2) 3538 // 3539 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3540 PhaseTransform* phase, bool can_reshape) { 3541 assert(stores_are_sane(phase), ""); 3542 3543 if (start < 0) return NULL; 3544 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3545 3546 Compile* C = phase->C; 3547 int size_in_bytes = st->memory_size(); 3548 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3549 if (i == 0) return NULL; // bail out 3550 Node* prev_mem = NULL; // raw memory for the captured store 3551 if (i > 0) { 3552 prev_mem = in(i); // there is a pre-existing store under this one 3553 set_req(i, C->top()); // temporarily disconnect it 3554 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3555 } else { 3556 i = -i; // no pre-existing store 3557 prev_mem = zero_memory(); // a slice of the newly allocated object 3558 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3559 set_req(--i, C->top()); // reuse this edge; it has been folded away 3560 else 3561 ins_req(i, C->top()); // build a new edge 3562 } 3563 Node* new_st = st->clone(); 3564 new_st->set_req(MemNode::Control, in(Control)); 3565 new_st->set_req(MemNode::Memory, prev_mem); 3566 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3567 new_st = phase->transform(new_st); 3568 3569 // At this point, new_st might have swallowed a pre-existing store 3570 // at the same offset, or perhaps new_st might have disappeared, 3571 // if it redundantly stored the same value (or zero to fresh memory). 3572 3573 // In any case, wire it in: 3574 set_req(i, new_st); 3575 3576 // The caller may now kill the old guy. 3577 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3578 assert(check_st == new_st || check_st == NULL, "must be findable"); 3579 assert(!is_complete(), ""); 3580 return new_st; 3581 } 3582 3583 static bool store_constant(jlong* tiles, int num_tiles, 3584 intptr_t st_off, int st_size, 3585 jlong con) { 3586 if ((st_off & (st_size-1)) != 0) 3587 return false; // strange store offset (assume size==2**N) 3588 address addr = (address)tiles + st_off; 3589 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3590 switch (st_size) { 3591 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3592 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3593 case sizeof(jint): *(jint*) addr = (jint) con; break; 3594 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3595 default: return false; // strange store size (detect size!=2**N here) 3596 } 3597 return true; // return success to caller 3598 } 3599 3600 // Coalesce subword constants into int constants and possibly 3601 // into long constants. The goal, if the CPU permits, 3602 // is to initialize the object with a small number of 64-bit tiles. 3603 // Also, convert floating-point constants to bit patterns. 3604 // Non-constants are not relevant to this pass. 3605 // 3606 // In terms of the running example on InitializeNode::InitializeNode 3607 // and InitializeNode::capture_store, here is the transformation 3608 // of rawstore1 and rawstore2 into rawstore12: 3609 // alloc = (Allocate ...) 3610 // rawoop = alloc.RawAddress 3611 // tile12 = 0x00010002 3612 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3613 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3614 // 3615 void 3616 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3617 Node* size_in_bytes, 3618 PhaseGVN* phase) { 3619 Compile* C = phase->C; 3620 3621 assert(stores_are_sane(phase), ""); 3622 // Note: After this pass, they are not completely sane, 3623 // since there may be some overlaps. 3624 3625 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3626 3627 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3628 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3629 size_limit = MIN2(size_limit, ti_limit); 3630 size_limit = align_size_up(size_limit, BytesPerLong); 3631 int num_tiles = size_limit / BytesPerLong; 3632 3633 // allocate space for the tile map: 3634 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3635 jlong tiles_buf[small_len]; 3636 Node* nodes_buf[small_len]; 3637 jlong inits_buf[small_len]; 3638 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3639 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3640 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3641 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3642 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3643 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3644 // tiles: exact bitwise model of all primitive constants 3645 // nodes: last constant-storing node subsumed into the tiles model 3646 // inits: which bytes (in each tile) are touched by any initializations 3647 3648 //// Pass A: Fill in the tile model with any relevant stores. 3649 3650 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3651 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3652 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3653 Node* zmem = zero_memory(); // initially zero memory state 3654 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3655 Node* st = in(i); 3656 intptr_t st_off = get_store_offset(st, phase); 3657 3658 // Figure out the store's offset and constant value: 3659 if (st_off < header_size) continue; //skip (ignore header) 3660 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3661 int st_size = st->as_Store()->memory_size(); 3662 if (st_off + st_size > size_limit) break; 3663 3664 // Record which bytes are touched, whether by constant or not. 3665 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3666 continue; // skip (strange store size) 3667 3668 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3669 if (!val->singleton()) continue; //skip (non-con store) 3670 BasicType type = val->basic_type(); 3671 3672 jlong con = 0; 3673 switch (type) { 3674 case T_INT: con = val->is_int()->get_con(); break; 3675 case T_LONG: con = val->is_long()->get_con(); break; 3676 case T_FLOAT: con = jint_cast(val->getf()); break; 3677 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3678 default: continue; //skip (odd store type) 3679 } 3680 3681 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3682 st->Opcode() == Op_StoreL) { 3683 continue; // This StoreL is already optimal. 3684 } 3685 3686 // Store down the constant. 3687 store_constant(tiles, num_tiles, st_off, st_size, con); 3688 3689 intptr_t j = st_off >> LogBytesPerLong; 3690 3691 if (type == T_INT && st_size == BytesPerInt 3692 && (st_off & BytesPerInt) == BytesPerInt) { 3693 jlong lcon = tiles[j]; 3694 if (!Matcher::isSimpleConstant64(lcon) && 3695 st->Opcode() == Op_StoreI) { 3696 // This StoreI is already optimal by itself. 3697 jint* intcon = (jint*) &tiles[j]; 3698 intcon[1] = 0; // undo the store_constant() 3699 3700 // If the previous store is also optimal by itself, back up and 3701 // undo the action of the previous loop iteration... if we can. 3702 // But if we can't, just let the previous half take care of itself. 3703 st = nodes[j]; 3704 st_off -= BytesPerInt; 3705 con = intcon[0]; 3706 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3707 assert(st_off >= header_size, "still ignoring header"); 3708 assert(get_store_offset(st, phase) == st_off, "must be"); 3709 assert(in(i-1) == zmem, "must be"); 3710 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3711 assert(con == tcon->is_int()->get_con(), "must be"); 3712 // Undo the effects of the previous loop trip, which swallowed st: 3713 intcon[0] = 0; // undo store_constant() 3714 set_req(i-1, st); // undo set_req(i, zmem) 3715 nodes[j] = NULL; // undo nodes[j] = st 3716 --old_subword; // undo ++old_subword 3717 } 3718 continue; // This StoreI is already optimal. 3719 } 3720 } 3721 3722 // This store is not needed. 3723 set_req(i, zmem); 3724 nodes[j] = st; // record for the moment 3725 if (st_size < BytesPerLong) // something has changed 3726 ++old_subword; // includes int/float, but who's counting... 3727 else ++old_long; 3728 } 3729 3730 if ((old_subword + old_long) == 0) 3731 return; // nothing more to do 3732 3733 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3734 // Be sure to insert them before overlapping non-constant stores. 3735 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3736 for (int j = 0; j < num_tiles; j++) { 3737 jlong con = tiles[j]; 3738 jlong init = inits[j]; 3739 if (con == 0) continue; 3740 jint con0, con1; // split the constant, address-wise 3741 jint init0, init1; // split the init map, address-wise 3742 { union { jlong con; jint intcon[2]; } u; 3743 u.con = con; 3744 con0 = u.intcon[0]; 3745 con1 = u.intcon[1]; 3746 u.con = init; 3747 init0 = u.intcon[0]; 3748 init1 = u.intcon[1]; 3749 } 3750 3751 Node* old = nodes[j]; 3752 assert(old != NULL, "need the prior store"); 3753 intptr_t offset = (j * BytesPerLong); 3754 3755 bool split = !Matcher::isSimpleConstant64(con); 3756 3757 if (offset < header_size) { 3758 assert(offset + BytesPerInt >= header_size, "second int counts"); 3759 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3760 split = true; // only the second word counts 3761 // Example: int a[] = { 42 ... } 3762 } else if (con0 == 0 && init0 == -1) { 3763 split = true; // first word is covered by full inits 3764 // Example: int a[] = { ... foo(), 42 ... } 3765 } else if (con1 == 0 && init1 == -1) { 3766 split = true; // second word is covered by full inits 3767 // Example: int a[] = { ... 42, foo() ... } 3768 } 3769 3770 // Here's a case where init0 is neither 0 nor -1: 3771 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3772 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3773 // In this case the tile is not split; it is (jlong)42. 3774 // The big tile is stored down, and then the foo() value is inserted. 3775 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3776 3777 Node* ctl = old->in(MemNode::Control); 3778 Node* adr = make_raw_address(offset, phase); 3779 const TypePtr* atp = TypeRawPtr::BOTTOM; 3780 3781 // One or two coalesced stores to plop down. 3782 Node* st[2]; 3783 intptr_t off[2]; 3784 int nst = 0; 3785 if (!split) { 3786 ++new_long; 3787 off[nst] = offset; 3788 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3789 phase->longcon(con), T_LONG, MemNode::unordered); 3790 } else { 3791 // Omit either if it is a zero. 3792 if (con0 != 0) { 3793 ++new_int; 3794 off[nst] = offset; 3795 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3796 phase->intcon(con0), T_INT, MemNode::unordered); 3797 } 3798 if (con1 != 0) { 3799 ++new_int; 3800 offset += BytesPerInt; 3801 adr = make_raw_address(offset, phase); 3802 off[nst] = offset; 3803 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3804 phase->intcon(con1), T_INT, MemNode::unordered); 3805 } 3806 } 3807 3808 // Insert second store first, then the first before the second. 3809 // Insert each one just before any overlapping non-constant stores. 3810 while (nst > 0) { 3811 Node* st1 = st[--nst]; 3812 C->copy_node_notes_to(st1, old); 3813 st1 = phase->transform(st1); 3814 offset = off[nst]; 3815 assert(offset >= header_size, "do not smash header"); 3816 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3817 guarantee(ins_idx != 0, "must re-insert constant store"); 3818 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3819 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3820 set_req(--ins_idx, st1); 3821 else 3822 ins_req(ins_idx, st1); 3823 } 3824 } 3825 3826 if (PrintCompilation && WizardMode) 3827 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3828 old_subword, old_long, new_int, new_long); 3829 if (C->log() != NULL) 3830 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3831 old_subword, old_long, new_int, new_long); 3832 3833 // Clean up any remaining occurrences of zmem: 3834 remove_extra_zeroes(); 3835 } 3836 3837 // Explore forward from in(start) to find the first fully initialized 3838 // word, and return its offset. Skip groups of subword stores which 3839 // together initialize full words. If in(start) is itself part of a 3840 // fully initialized word, return the offset of in(start). If there 3841 // are no following full-word stores, or if something is fishy, return 3842 // a negative value. 3843 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3844 int int_map = 0; 3845 intptr_t int_map_off = 0; 3846 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3847 3848 for (uint i = start, limit = req(); i < limit; i++) { 3849 Node* st = in(i); 3850 3851 intptr_t st_off = get_store_offset(st, phase); 3852 if (st_off < 0) break; // return conservative answer 3853 3854 int st_size = st->as_Store()->memory_size(); 3855 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3856 return st_off; // we found a complete word init 3857 } 3858 3859 // update the map: 3860 3861 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3862 if (this_int_off != int_map_off) { 3863 // reset the map: 3864 int_map = 0; 3865 int_map_off = this_int_off; 3866 } 3867 3868 int subword_off = st_off - this_int_off; 3869 int_map |= right_n_bits(st_size) << subword_off; 3870 if ((int_map & FULL_MAP) == FULL_MAP) { 3871 return this_int_off; // we found a complete word init 3872 } 3873 3874 // Did this store hit or cross the word boundary? 3875 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3876 if (next_int_off == this_int_off + BytesPerInt) { 3877 // We passed the current int, without fully initializing it. 3878 int_map_off = next_int_off; 3879 int_map >>= BytesPerInt; 3880 } else if (next_int_off > this_int_off + BytesPerInt) { 3881 // We passed the current and next int. 3882 return this_int_off + BytesPerInt; 3883 } 3884 } 3885 3886 return -1; 3887 } 3888 3889 3890 // Called when the associated AllocateNode is expanded into CFG. 3891 // At this point, we may perform additional optimizations. 3892 // Linearize the stores by ascending offset, to make memory 3893 // activity as coherent as possible. 3894 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3895 intptr_t header_size, 3896 Node* size_in_bytes, 3897 PhaseGVN* phase) { 3898 assert(!is_complete(), "not already complete"); 3899 assert(stores_are_sane(phase), ""); 3900 assert(allocation() != NULL, "must be present"); 3901 3902 remove_extra_zeroes(); 3903 3904 if (ReduceFieldZeroing || ReduceBulkZeroing) 3905 // reduce instruction count for common initialization patterns 3906 coalesce_subword_stores(header_size, size_in_bytes, phase); 3907 3908 Node* zmem = zero_memory(); // initially zero memory state 3909 Node* inits = zmem; // accumulating a linearized chain of inits 3910 #ifdef ASSERT 3911 intptr_t first_offset = allocation()->minimum_header_size(); 3912 intptr_t last_init_off = first_offset; // previous init offset 3913 intptr_t last_init_end = first_offset; // previous init offset+size 3914 intptr_t last_tile_end = first_offset; // previous tile offset+size 3915 #endif 3916 intptr_t zeroes_done = header_size; 3917 3918 bool do_zeroing = true; // we might give up if inits are very sparse 3919 int big_init_gaps = 0; // how many large gaps have we seen? 3920 3921 if (ZeroTLAB) do_zeroing = false; 3922 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3923 3924 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3925 Node* st = in(i); 3926 intptr_t st_off = get_store_offset(st, phase); 3927 if (st_off < 0) 3928 break; // unknown junk in the inits 3929 if (st->in(MemNode::Memory) != zmem) 3930 break; // complicated store chains somehow in list 3931 3932 int st_size = st->as_Store()->memory_size(); 3933 intptr_t next_init_off = st_off + st_size; 3934 3935 if (do_zeroing && zeroes_done < next_init_off) { 3936 // See if this store needs a zero before it or under it. 3937 intptr_t zeroes_needed = st_off; 3938 3939 if (st_size < BytesPerInt) { 3940 // Look for subword stores which only partially initialize words. 3941 // If we find some, we must lay down some word-level zeroes first, 3942 // underneath the subword stores. 3943 // 3944 // Examples: 3945 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3946 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3947 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3948 // 3949 // Note: coalesce_subword_stores may have already done this, 3950 // if it was prompted by constant non-zero subword initializers. 3951 // But this case can still arise with non-constant stores. 3952 3953 intptr_t next_full_store = find_next_fullword_store(i, phase); 3954 3955 // In the examples above: 3956 // in(i) p q r s x y z 3957 // st_off 12 13 14 15 12 13 14 3958 // st_size 1 1 1 1 1 1 1 3959 // next_full_s. 12 16 16 16 16 16 16 3960 // z's_done 12 16 16 16 12 16 12 3961 // z's_needed 12 16 16 16 16 16 16 3962 // zsize 0 0 0 0 4 0 4 3963 if (next_full_store < 0) { 3964 // Conservative tack: Zero to end of current word. 3965 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3966 } else { 3967 // Zero to beginning of next fully initialized word. 3968 // Or, don't zero at all, if we are already in that word. 3969 assert(next_full_store >= zeroes_needed, "must go forward"); 3970 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3971 zeroes_needed = next_full_store; 3972 } 3973 } 3974 3975 if (zeroes_needed > zeroes_done) { 3976 intptr_t zsize = zeroes_needed - zeroes_done; 3977 // Do some incremental zeroing on rawmem, in parallel with inits. 3978 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3979 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3980 zeroes_done, zeroes_needed, 3981 phase); 3982 zeroes_done = zeroes_needed; 3983 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3984 do_zeroing = false; // leave the hole, next time 3985 } 3986 } 3987 3988 // Collect the store and move on: 3989 st->set_req(MemNode::Memory, inits); 3990 inits = st; // put it on the linearized chain 3991 set_req(i, zmem); // unhook from previous position 3992 3993 if (zeroes_done == st_off) 3994 zeroes_done = next_init_off; 3995 3996 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3997 3998 #ifdef ASSERT 3999 // Various order invariants. Weaker than stores_are_sane because 4000 // a large constant tile can be filled in by smaller non-constant stores. 4001 assert(st_off >= last_init_off, "inits do not reverse"); 4002 last_init_off = st_off; 4003 const Type* val = NULL; 4004 if (st_size >= BytesPerInt && 4005 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4006 (int)val->basic_type() < (int)T_OBJECT) { 4007 assert(st_off >= last_tile_end, "tiles do not overlap"); 4008 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4009 last_tile_end = MAX2(last_tile_end, next_init_off); 4010 } else { 4011 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 4012 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4013 assert(st_off >= last_init_end, "inits do not overlap"); 4014 last_init_end = next_init_off; // it's a non-tile 4015 } 4016 #endif //ASSERT 4017 } 4018 4019 remove_extra_zeroes(); // clear out all the zmems left over 4020 add_req(inits); 4021 4022 if (!ZeroTLAB) { 4023 // If anything remains to be zeroed, zero it all now. 4024 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 4025 // if it is the last unused 4 bytes of an instance, forget about it 4026 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4027 if (zeroes_done + BytesPerLong >= size_limit) { 4028 assert(allocation() != NULL, ""); 4029 if (allocation()->Opcode() == Op_Allocate) { 4030 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 4031 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4032 if (zeroes_done == k->layout_helper()) 4033 zeroes_done = size_limit; 4034 } 4035 } 4036 if (zeroes_done < size_limit) { 4037 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4038 zeroes_done, size_in_bytes, phase); 4039 } 4040 } 4041 4042 set_complete(phase); 4043 return rawmem; 4044 } 4045 4046 4047 #ifdef ASSERT 4048 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4049 if (is_complete()) 4050 return true; // stores could be anything at this point 4051 assert(allocation() != NULL, "must be present"); 4052 intptr_t last_off = allocation()->minimum_header_size(); 4053 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4054 Node* st = in(i); 4055 intptr_t st_off = get_store_offset(st, phase); 4056 if (st_off < 0) continue; // ignore dead garbage 4057 if (last_off > st_off) { 4058 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4059 this->dump(2); 4060 assert(false, "ascending store offsets"); 4061 return false; 4062 } 4063 last_off = st_off + st->as_Store()->memory_size(); 4064 } 4065 return true; 4066 } 4067 #endif //ASSERT 4068 4069 4070 4071 4072 //============================MergeMemNode===================================== 4073 // 4074 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4075 // contributing store or call operations. Each contributor provides the memory 4076 // state for a particular "alias type" (see Compile::alias_type). For example, 4077 // if a MergeMem has an input X for alias category #6, then any memory reference 4078 // to alias category #6 may use X as its memory state input, as an exact equivalent 4079 // to using the MergeMem as a whole. 4080 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4081 // 4082 // (Here, the <N> notation gives the index of the relevant adr_type.) 4083 // 4084 // In one special case (and more cases in the future), alias categories overlap. 4085 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4086 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4087 // it is exactly equivalent to that state W: 4088 // MergeMem(<Bot>: W) <==> W 4089 // 4090 // Usually, the merge has more than one input. In that case, where inputs 4091 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4092 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4093 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4094 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4095 // 4096 // A merge can take a "wide" memory state as one of its narrow inputs. 4097 // This simply means that the merge observes out only the relevant parts of 4098 // the wide input. That is, wide memory states arriving at narrow merge inputs 4099 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4100 // 4101 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4102 // and that memory slices "leak through": 4103 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4104 // 4105 // But, in such a cascade, repeated memory slices can "block the leak": 4106 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4107 // 4108 // In the last example, Y is not part of the combined memory state of the 4109 // outermost MergeMem. The system must, of course, prevent unschedulable 4110 // memory states from arising, so you can be sure that the state Y is somehow 4111 // a precursor to state Y'. 4112 // 4113 // 4114 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4115 // of each MergeMemNode array are exactly the numerical alias indexes, including 4116 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4117 // Compile::alias_type (and kin) produce and manage these indexes. 4118 // 4119 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4120 // (Note that this provides quick access to the top node inside MergeMem methods, 4121 // without the need to reach out via TLS to Compile::current.) 4122 // 4123 // As a consequence of what was just described, a MergeMem that represents a full 4124 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4125 // containing all alias categories. 4126 // 4127 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4128 // 4129 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4130 // a memory state for the alias type <N>, or else the top node, meaning that 4131 // there is no particular input for that alias type. Note that the length of 4132 // a MergeMem is variable, and may be extended at any time to accommodate new 4133 // memory states at larger alias indexes. When merges grow, they are of course 4134 // filled with "top" in the unused in() positions. 4135 // 4136 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4137 // (Top was chosen because it works smoothly with passes like GCM.) 4138 // 4139 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4140 // the type of random VM bits like TLS references.) Since it is always the 4141 // first non-Bot memory slice, some low-level loops use it to initialize an 4142 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4143 // 4144 // 4145 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4146 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4147 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4148 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4149 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4150 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4151 // 4152 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4153 // really that different from the other memory inputs. An abbreviation called 4154 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4155 // 4156 // 4157 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4158 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4159 // that "emerges though" the base memory will be marked as excluding the alias types 4160 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4161 // 4162 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4163 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4164 // 4165 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4166 // (It is currently unimplemented.) As you can see, the resulting merge is 4167 // actually a disjoint union of memory states, rather than an overlay. 4168 // 4169 4170 //------------------------------MergeMemNode----------------------------------- 4171 Node* MergeMemNode::make_empty_memory() { 4172 Node* empty_memory = (Node*) Compile::current()->top(); 4173 assert(empty_memory->is_top(), "correct sentinel identity"); 4174 return empty_memory; 4175 } 4176 4177 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4178 init_class_id(Class_MergeMem); 4179 // all inputs are nullified in Node::Node(int) 4180 // set_input(0, NULL); // no control input 4181 4182 // Initialize the edges uniformly to top, for starters. 4183 Node* empty_mem = make_empty_memory(); 4184 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4185 init_req(i,empty_mem); 4186 } 4187 assert(empty_memory() == empty_mem, ""); 4188 4189 if( new_base != NULL && new_base->is_MergeMem() ) { 4190 MergeMemNode* mdef = new_base->as_MergeMem(); 4191 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4192 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4193 mms.set_memory(mms.memory2()); 4194 } 4195 assert(base_memory() == mdef->base_memory(), ""); 4196 } else { 4197 set_base_memory(new_base); 4198 } 4199 } 4200 4201 // Make a new, untransformed MergeMem with the same base as 'mem'. 4202 // If mem is itself a MergeMem, populate the result with the same edges. 4203 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 4204 return new(C) MergeMemNode(mem); 4205 } 4206 4207 //------------------------------cmp-------------------------------------------- 4208 uint MergeMemNode::hash() const { return NO_HASH; } 4209 uint MergeMemNode::cmp( const Node &n ) const { 4210 return (&n == this); // Always fail except on self 4211 } 4212 4213 //------------------------------Identity--------------------------------------- 4214 Node* MergeMemNode::Identity(PhaseTransform *phase) { 4215 // Identity if this merge point does not record any interesting memory 4216 // disambiguations. 4217 Node* base_mem = base_memory(); 4218 Node* empty_mem = empty_memory(); 4219 if (base_mem != empty_mem) { // Memory path is not dead? 4220 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4221 Node* mem = in(i); 4222 if (mem != empty_mem && mem != base_mem) { 4223 return this; // Many memory splits; no change 4224 } 4225 } 4226 } 4227 return base_mem; // No memory splits; ID on the one true input 4228 } 4229 4230 //------------------------------Ideal------------------------------------------ 4231 // This method is invoked recursively on chains of MergeMem nodes 4232 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4233 // Remove chain'd MergeMems 4234 // 4235 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4236 // relative to the "in(Bot)". Since we are patching both at the same time, 4237 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4238 // but rewrite each "in(i)" relative to the new "in(Bot)". 4239 Node *progress = NULL; 4240 4241 4242 Node* old_base = base_memory(); 4243 Node* empty_mem = empty_memory(); 4244 if (old_base == empty_mem) 4245 return NULL; // Dead memory path. 4246 4247 MergeMemNode* old_mbase; 4248 if (old_base != NULL && old_base->is_MergeMem()) 4249 old_mbase = old_base->as_MergeMem(); 4250 else 4251 old_mbase = NULL; 4252 Node* new_base = old_base; 4253 4254 // simplify stacked MergeMems in base memory 4255 if (old_mbase) new_base = old_mbase->base_memory(); 4256 4257 // the base memory might contribute new slices beyond my req() 4258 if (old_mbase) grow_to_match(old_mbase); 4259 4260 // Look carefully at the base node if it is a phi. 4261 PhiNode* phi_base; 4262 if (new_base != NULL && new_base->is_Phi()) 4263 phi_base = new_base->as_Phi(); 4264 else 4265 phi_base = NULL; 4266 4267 Node* phi_reg = NULL; 4268 uint phi_len = (uint)-1; 4269 if (phi_base != NULL && !phi_base->is_copy()) { 4270 // do not examine phi if degraded to a copy 4271 phi_reg = phi_base->region(); 4272 phi_len = phi_base->req(); 4273 // see if the phi is unfinished 4274 for (uint i = 1; i < phi_len; i++) { 4275 if (phi_base->in(i) == NULL) { 4276 // incomplete phi; do not look at it yet! 4277 phi_reg = NULL; 4278 phi_len = (uint)-1; 4279 break; 4280 } 4281 } 4282 } 4283 4284 // Note: We do not call verify_sparse on entry, because inputs 4285 // can normalize to the base_memory via subsume_node or similar 4286 // mechanisms. This method repairs that damage. 4287 4288 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4289 4290 // Look at each slice. 4291 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4292 Node* old_in = in(i); 4293 // calculate the old memory value 4294 Node* old_mem = old_in; 4295 if (old_mem == empty_mem) old_mem = old_base; 4296 assert(old_mem == memory_at(i), ""); 4297 4298 // maybe update (reslice) the old memory value 4299 4300 // simplify stacked MergeMems 4301 Node* new_mem = old_mem; 4302 MergeMemNode* old_mmem; 4303 if (old_mem != NULL && old_mem->is_MergeMem()) 4304 old_mmem = old_mem->as_MergeMem(); 4305 else 4306 old_mmem = NULL; 4307 if (old_mmem == this) { 4308 // This can happen if loops break up and safepoints disappear. 4309 // A merge of BotPtr (default) with a RawPtr memory derived from a 4310 // safepoint can be rewritten to a merge of the same BotPtr with 4311 // the BotPtr phi coming into the loop. If that phi disappears 4312 // also, we can end up with a self-loop of the mergemem. 4313 // In general, if loops degenerate and memory effects disappear, 4314 // a mergemem can be left looking at itself. This simply means 4315 // that the mergemem's default should be used, since there is 4316 // no longer any apparent effect on this slice. 4317 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4318 // from start. Update the input to TOP. 4319 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4320 } 4321 else if (old_mmem != NULL) { 4322 new_mem = old_mmem->memory_at(i); 4323 } 4324 // else preceding memory was not a MergeMem 4325 4326 // replace equivalent phis (unfortunately, they do not GVN together) 4327 if (new_mem != NULL && new_mem != new_base && 4328 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4329 if (new_mem->is_Phi()) { 4330 PhiNode* phi_mem = new_mem->as_Phi(); 4331 for (uint i = 1; i < phi_len; i++) { 4332 if (phi_base->in(i) != phi_mem->in(i)) { 4333 phi_mem = NULL; 4334 break; 4335 } 4336 } 4337 if (phi_mem != NULL) { 4338 // equivalent phi nodes; revert to the def 4339 new_mem = new_base; 4340 } 4341 } 4342 } 4343 4344 // maybe store down a new value 4345 Node* new_in = new_mem; 4346 if (new_in == new_base) new_in = empty_mem; 4347 4348 if (new_in != old_in) { 4349 // Warning: Do not combine this "if" with the previous "if" 4350 // A memory slice might have be be rewritten even if it is semantically 4351 // unchanged, if the base_memory value has changed. 4352 set_req(i, new_in); 4353 progress = this; // Report progress 4354 } 4355 } 4356 4357 if (new_base != old_base) { 4358 set_req(Compile::AliasIdxBot, new_base); 4359 // Don't use set_base_memory(new_base), because we need to update du. 4360 assert(base_memory() == new_base, ""); 4361 progress = this; 4362 } 4363 4364 if( base_memory() == this ) { 4365 // a self cycle indicates this memory path is dead 4366 set_req(Compile::AliasIdxBot, empty_mem); 4367 } 4368 4369 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4370 // Recursion must occur after the self cycle check above 4371 if( base_memory()->is_MergeMem() ) { 4372 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4373 Node *m = phase->transform(new_mbase); // Rollup any cycles 4374 if( m != NULL && (m->is_top() || 4375 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4376 // propagate rollup of dead cycle to self 4377 set_req(Compile::AliasIdxBot, empty_mem); 4378 } 4379 } 4380 4381 if( base_memory() == empty_mem ) { 4382 progress = this; 4383 // Cut inputs during Parse phase only. 4384 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4385 if( !can_reshape ) { 4386 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4387 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4388 } 4389 } 4390 } 4391 4392 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4393 // Check if PhiNode::Ideal's "Split phis through memory merges" 4394 // transform should be attempted. Look for this->phi->this cycle. 4395 uint merge_width = req(); 4396 if (merge_width > Compile::AliasIdxRaw) { 4397 PhiNode* phi = base_memory()->as_Phi(); 4398 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4399 if (phi->in(i) == this) { 4400 phase->is_IterGVN()->_worklist.push(phi); 4401 break; 4402 } 4403 } 4404 } 4405 } 4406 4407 assert(progress || verify_sparse(), "please, no dups of base"); 4408 return progress; 4409 } 4410 4411 //-------------------------set_base_memory------------------------------------- 4412 void MergeMemNode::set_base_memory(Node *new_base) { 4413 Node* empty_mem = empty_memory(); 4414 set_req(Compile::AliasIdxBot, new_base); 4415 assert(memory_at(req()) == new_base, "must set default memory"); 4416 // Clear out other occurrences of new_base: 4417 if (new_base != empty_mem) { 4418 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4419 if (in(i) == new_base) set_req(i, empty_mem); 4420 } 4421 } 4422 } 4423 4424 //------------------------------out_RegMask------------------------------------ 4425 const RegMask &MergeMemNode::out_RegMask() const { 4426 return RegMask::Empty; 4427 } 4428 4429 //------------------------------dump_spec-------------------------------------- 4430 #ifndef PRODUCT 4431 void MergeMemNode::dump_spec(outputStream *st) const { 4432 st->print(" {"); 4433 Node* base_mem = base_memory(); 4434 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4435 Node* mem = memory_at(i); 4436 if (mem == base_mem) { st->print(" -"); continue; } 4437 st->print( " N%d:", mem->_idx ); 4438 Compile::current()->get_adr_type(i)->dump_on(st); 4439 } 4440 st->print(" }"); 4441 } 4442 #endif // !PRODUCT 4443 4444 4445 #ifdef ASSERT 4446 static bool might_be_same(Node* a, Node* b) { 4447 if (a == b) return true; 4448 if (!(a->is_Phi() || b->is_Phi())) return false; 4449 // phis shift around during optimization 4450 return true; // pretty stupid... 4451 } 4452 4453 // verify a narrow slice (either incoming or outgoing) 4454 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4455 if (!VerifyAliases) return; // don't bother to verify unless requested 4456 if (is_error_reported()) return; // muzzle asserts when debugging an error 4457 if (Node::in_dump()) return; // muzzle asserts when printing 4458 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4459 assert(n != NULL, ""); 4460 // Elide intervening MergeMem's 4461 while (n->is_MergeMem()) { 4462 n = n->as_MergeMem()->memory_at(alias_idx); 4463 } 4464 Compile* C = Compile::current(); 4465 const TypePtr* n_adr_type = n->adr_type(); 4466 if (n == m->empty_memory()) { 4467 // Implicit copy of base_memory() 4468 } else if (n_adr_type != TypePtr::BOTTOM) { 4469 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4470 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4471 } else { 4472 // A few places like make_runtime_call "know" that VM calls are narrow, 4473 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4474 bool expected_wide_mem = false; 4475 if (n == m->base_memory()) { 4476 expected_wide_mem = true; 4477 } else if (alias_idx == Compile::AliasIdxRaw || 4478 n == m->memory_at(Compile::AliasIdxRaw)) { 4479 expected_wide_mem = true; 4480 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4481 // memory can "leak through" calls on channels that 4482 // are write-once. Allow this also. 4483 expected_wide_mem = true; 4484 } 4485 assert(expected_wide_mem, "expected narrow slice replacement"); 4486 } 4487 } 4488 #else // !ASSERT 4489 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4490 #endif 4491 4492 4493 //-----------------------------memory_at--------------------------------------- 4494 Node* MergeMemNode::memory_at(uint alias_idx) const { 4495 assert(alias_idx >= Compile::AliasIdxRaw || 4496 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4497 "must avoid base_memory and AliasIdxTop"); 4498 4499 // Otherwise, it is a narrow slice. 4500 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4501 Compile *C = Compile::current(); 4502 if (is_empty_memory(n)) { 4503 // the array is sparse; empty slots are the "top" node 4504 n = base_memory(); 4505 assert(Node::in_dump() 4506 || n == NULL || n->bottom_type() == Type::TOP 4507 || n->adr_type() == NULL // address is TOP 4508 || n->adr_type() == TypePtr::BOTTOM 4509 || n->adr_type() == TypeRawPtr::BOTTOM 4510 || Compile::current()->AliasLevel() == 0, 4511 "must be a wide memory"); 4512 // AliasLevel == 0 if we are organizing the memory states manually. 4513 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4514 } else { 4515 // make sure the stored slice is sane 4516 #ifdef ASSERT 4517 if (is_error_reported() || Node::in_dump()) { 4518 } else if (might_be_same(n, base_memory())) { 4519 // Give it a pass: It is a mostly harmless repetition of the base. 4520 // This can arise normally from node subsumption during optimization. 4521 } else { 4522 verify_memory_slice(this, alias_idx, n); 4523 } 4524 #endif 4525 } 4526 return n; 4527 } 4528 4529 //---------------------------set_memory_at------------------------------------- 4530 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4531 verify_memory_slice(this, alias_idx, n); 4532 Node* empty_mem = empty_memory(); 4533 if (n == base_memory()) n = empty_mem; // collapse default 4534 uint need_req = alias_idx+1; 4535 if (req() < need_req) { 4536 if (n == empty_mem) return; // already the default, so do not grow me 4537 // grow the sparse array 4538 do { 4539 add_req(empty_mem); 4540 } while (req() < need_req); 4541 } 4542 set_req( alias_idx, n ); 4543 } 4544 4545 4546 4547 //--------------------------iteration_setup------------------------------------ 4548 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4549 if (other != NULL) { 4550 grow_to_match(other); 4551 // invariant: the finite support of mm2 is within mm->req() 4552 #ifdef ASSERT 4553 for (uint i = req(); i < other->req(); i++) { 4554 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4555 } 4556 #endif 4557 } 4558 // Replace spurious copies of base_memory by top. 4559 Node* base_mem = base_memory(); 4560 if (base_mem != NULL && !base_mem->is_top()) { 4561 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4562 if (in(i) == base_mem) 4563 set_req(i, empty_memory()); 4564 } 4565 } 4566 } 4567 4568 //---------------------------grow_to_match------------------------------------- 4569 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4570 Node* empty_mem = empty_memory(); 4571 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4572 // look for the finite support of the other memory 4573 for (uint i = other->req(); --i >= req(); ) { 4574 if (other->in(i) != empty_mem) { 4575 uint new_len = i+1; 4576 while (req() < new_len) add_req(empty_mem); 4577 break; 4578 } 4579 } 4580 } 4581 4582 //---------------------------verify_sparse------------------------------------- 4583 #ifndef PRODUCT 4584 bool MergeMemNode::verify_sparse() const { 4585 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4586 Node* base_mem = base_memory(); 4587 // The following can happen in degenerate cases, since empty==top. 4588 if (is_empty_memory(base_mem)) return true; 4589 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4590 assert(in(i) != NULL, "sane slice"); 4591 if (in(i) == base_mem) return false; // should have been the sentinel value! 4592 } 4593 return true; 4594 } 4595 4596 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4597 Node* n; 4598 n = mm->in(idx); 4599 if (mem == n) return true; // might be empty_memory() 4600 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4601 if (mem == n) return true; 4602 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4603 if (mem == n) return true; 4604 if (n == NULL) break; 4605 } 4606 return false; 4607 } 4608 #endif // !PRODUCT