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