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