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