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