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