1 /* 2 * Copyright (c) 1997, 2018, 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 "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "oops/objArrayKlass.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/compile.hpp" 37 #include "opto/connode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/loopnode.hpp" 40 #include "opto/machnode.hpp" 41 #include "opto/matcher.hpp" 42 #include "opto/memnode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/phaseX.hpp" 46 #include "opto/regmask.hpp" 47 #include "opto/valuetypenode.hpp" 48 #include "utilities/align.hpp" 49 #include "utilities/copy.hpp" 50 #include "utilities/macros.hpp" 51 #include "utilities/vmError.hpp" 52 #if INCLUDE_ZGC 53 #include "gc/z/c2/zBarrierSetC2.hpp" 54 #endif 55 56 // Portions of code courtesy of Clifford Click 57 58 // Optimization - Graph Style 59 60 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 61 62 //============================================================================= 63 uint MemNode::size_of() const { return sizeof(*this); } 64 65 const TypePtr *MemNode::adr_type() const { 66 Node* adr = in(Address); 67 if (adr == NULL) return NULL; // node is dead 68 const TypePtr* cross_check = NULL; 69 DEBUG_ONLY(cross_check = _adr_type); 70 return calculate_adr_type(adr->bottom_type(), cross_check); 71 } 72 73 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 74 if (adr != NULL) { 75 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 76 return true; 77 } 78 } 79 return false; 80 } 81 82 #ifndef PRODUCT 83 void MemNode::dump_spec(outputStream *st) const { 84 if (in(Address) == NULL) return; // node is dead 85 #ifndef ASSERT 86 // fake the missing field 87 const TypePtr* _adr_type = NULL; 88 if (in(Address) != NULL) 89 _adr_type = in(Address)->bottom_type()->isa_ptr(); 90 #endif 91 dump_adr_type(this, _adr_type, st); 92 93 Compile* C = Compile::current(); 94 if (C->alias_type(_adr_type)->is_volatile()) { 95 st->print(" Volatile!"); 96 } 97 if (_unaligned_access) { 98 st->print(" unaligned"); 99 } 100 if (_mismatched_access) { 101 st->print(" mismatched"); 102 } 103 if (_unsafe_access) { 104 st->print(" unsafe"); 105 } 106 } 107 108 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 109 st->print(" @"); 110 if (adr_type == NULL) { 111 st->print("NULL"); 112 } else { 113 adr_type->dump_on(st); 114 Compile* C = Compile::current(); 115 Compile::AliasType* atp = NULL; 116 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 117 if (atp == NULL) 118 st->print(", idx=?\?;"); 119 else if (atp->index() == Compile::AliasIdxBot) 120 st->print(", idx=Bot;"); 121 else if (atp->index() == Compile::AliasIdxTop) 122 st->print(", idx=Top;"); 123 else if (atp->index() == Compile::AliasIdxRaw) 124 st->print(", idx=Raw;"); 125 else { 126 ciField* field = atp->field(); 127 if (field) { 128 st->print(", name="); 129 field->print_name_on(st); 130 } 131 st->print(", idx=%d;", atp->index()); 132 } 133 } 134 } 135 136 extern void print_alias_types(); 137 138 #endif 139 140 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 141 assert((t_oop != NULL), "sanity"); 142 bool is_instance = t_oop->is_known_instance_field(); 143 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 144 (load != NULL) && load->is_Load() && 145 (phase->is_IterGVN() != NULL); 146 if (!(is_instance || is_boxed_value_load)) 147 return mchain; // don't try to optimize non-instance types 148 uint instance_id = t_oop->instance_id(); 149 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory); 150 Node *prev = NULL; 151 Node *result = mchain; 152 while (prev != result) { 153 prev = result; 154 if (result == start_mem) 155 break; // hit one of our sentinels 156 // skip over a call which does not affect this memory slice 157 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 158 Node *proj_in = result->in(0); 159 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 160 break; // hit one of our sentinels 161 } else if (proj_in->is_Call()) { 162 // ArrayCopyNodes processed here as well 163 CallNode *call = proj_in->as_Call(); 164 if (!call->may_modify(t_oop, phase)) { // returns false for instances 165 result = call->in(TypeFunc::Memory); 166 } 167 } else if (proj_in->is_Initialize()) { 168 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 169 // Stop if this is the initialization for the object instance which 170 // contains this memory slice, otherwise skip over it. 171 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 172 break; 173 } 174 if (is_instance) { 175 result = proj_in->in(TypeFunc::Memory); 176 } else if (is_boxed_value_load) { 177 Node* klass = alloc->in(AllocateNode::KlassNode); 178 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 179 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 180 result = proj_in->in(TypeFunc::Memory); // not related allocation 181 } 182 } 183 } else if (proj_in->is_MemBar()) { 184 ArrayCopyNode* ac = NULL; 185 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 186 break; 187 } 188 result = proj_in->in(TypeFunc::Memory); 189 } else { 190 assert(false, "unexpected projection"); 191 } 192 } else if (result->is_ClearArray()) { 193 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 194 // Can not bypass initialization of the instance 195 // we are looking for. 196 break; 197 } 198 // Otherwise skip it (the call updated 'result' value). 199 } else if (result->is_MergeMem()) { 200 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 201 } 202 } 203 return result; 204 } 205 206 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 207 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 208 if (t_oop == NULL) 209 return mchain; // don't try to optimize non-oop types 210 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 211 bool is_instance = t_oop->is_known_instance_field(); 212 PhaseIterGVN *igvn = phase->is_IterGVN(); 213 if (is_instance && igvn != NULL && result->is_Phi()) { 214 PhiNode *mphi = result->as_Phi(); 215 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 216 const TypePtr *t = mphi->adr_type(); 217 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 218 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 219 t->is_oopptr()->cast_to_exactness(true) 220 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 221 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) { 222 // clone the Phi with our address type 223 result = mphi->split_out_instance(t_adr, igvn); 224 } else { 225 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 226 } 227 } 228 return result; 229 } 230 231 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 232 uint alias_idx = phase->C->get_alias_index(tp); 233 Node *mem = mmem; 234 #ifdef ASSERT 235 { 236 // Check that current type is consistent with the alias index used during graph construction 237 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 238 bool consistent = adr_check == NULL || adr_check->empty() || 239 phase->C->must_alias(adr_check, alias_idx ); 240 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 241 if( !consistent && adr_check != NULL && !adr_check->empty() && 242 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 243 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 244 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 245 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 246 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 247 // don't assert if it is dead code. 248 consistent = true; 249 } 250 if( !consistent ) { 251 st->print("alias_idx==%d, adr_check==", alias_idx); 252 if( adr_check == NULL ) { 253 st->print("NULL"); 254 } else { 255 adr_check->dump(); 256 } 257 st->cr(); 258 print_alias_types(); 259 assert(consistent, "adr_check must match alias idx"); 260 } 261 } 262 #endif 263 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 264 // means an array I have not precisely typed yet. Do not do any 265 // alias stuff with it any time soon. 266 const TypeOopPtr *toop = tp->isa_oopptr(); 267 if( tp->base() != Type::AnyPtr && 268 !(toop && 269 toop->klass() != NULL && 270 toop->klass()->is_java_lang_Object() && 271 toop->offset() == Type::OffsetBot) ) { 272 // compress paths and change unreachable cycles to TOP 273 // If not, we can update the input infinitely along a MergeMem cycle 274 // Equivalent code in PhiNode::Ideal 275 Node* m = phase->transform(mmem); 276 // If transformed to a MergeMem, get the desired slice 277 // Otherwise the returned node represents memory for every slice 278 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 279 // Update input if it is progress over what we have now 280 } 281 return mem; 282 } 283 284 //--------------------------Ideal_common--------------------------------------- 285 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 286 // Unhook non-raw memories from complete (macro-expanded) initializations. 287 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 288 // If our control input is a dead region, kill all below the region 289 Node *ctl = in(MemNode::Control); 290 if (ctl && remove_dead_region(phase, can_reshape)) 291 return this; 292 ctl = in(MemNode::Control); 293 // Don't bother trying to transform a dead node 294 if (ctl && ctl->is_top()) return NodeSentinel; 295 296 PhaseIterGVN *igvn = phase->is_IterGVN(); 297 // Wait if control on the worklist. 298 if (ctl && can_reshape && igvn != NULL) { 299 Node* bol = NULL; 300 Node* cmp = NULL; 301 if (ctl->in(0)->is_If()) { 302 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 303 bol = ctl->in(0)->in(1); 304 if (bol->is_Bool()) 305 cmp = ctl->in(0)->in(1)->in(1); 306 } 307 if (igvn->_worklist.member(ctl) || 308 (bol != NULL && igvn->_worklist.member(bol)) || 309 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 310 // This control path may be dead. 311 // Delay this memory node transformation until the control is processed. 312 phase->is_IterGVN()->_worklist.push(this); 313 return NodeSentinel; // caller will return NULL 314 } 315 } 316 // Ignore if memory is dead, or self-loop 317 Node *mem = in(MemNode::Memory); 318 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 319 assert(mem != this, "dead loop in MemNode::Ideal"); 320 321 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 322 // This memory slice may be dead. 323 // Delay this mem node transformation until the memory is processed. 324 phase->is_IterGVN()->_worklist.push(this); 325 return NodeSentinel; // caller will return NULL 326 } 327 328 Node *address = in(MemNode::Address); 329 const Type *t_adr = phase->type(address); 330 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 331 332 if (can_reshape && igvn != NULL && 333 (igvn->_worklist.member(address) || 334 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 335 // The address's base and type may change when the address is processed. 336 // Delay this mem node transformation until the address is processed. 337 phase->is_IterGVN()->_worklist.push(this); 338 return NodeSentinel; // caller will return NULL 339 } 340 341 // Do NOT remove or optimize the next lines: ensure a new alias index 342 // is allocated for an oop pointer type before Escape Analysis. 343 // Note: C++ will not remove it since the call has side effect. 344 if (t_adr->isa_oopptr()) { 345 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 346 } 347 348 Node* base = NULL; 349 if (address->is_AddP()) { 350 base = address->in(AddPNode::Base); 351 } 352 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 353 !t_adr->isa_rawptr()) { 354 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 355 // Skip this node optimization if its address has TOP base. 356 return NodeSentinel; // caller will return NULL 357 } 358 359 // Avoid independent memory operations 360 Node* old_mem = mem; 361 362 // The code which unhooks non-raw memories from complete (macro-expanded) 363 // initializations was removed. After macro-expansion all stores catched 364 // by Initialize node became raw stores and there is no information 365 // which memory slices they modify. So it is unsafe to move any memory 366 // operation above these stores. Also in most cases hooked non-raw memories 367 // were already unhooked by using information from detect_ptr_independence() 368 // and find_previous_store(). 369 370 if (mem->is_MergeMem()) { 371 MergeMemNode* mmem = mem->as_MergeMem(); 372 const TypePtr *tp = t_adr->is_ptr(); 373 374 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 375 } 376 377 if (mem != old_mem) { 378 set_req(MemNode::Memory, mem); 379 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) { 380 igvn->_worklist.push(old_mem); 381 } 382 if (phase->type(mem) == Type::TOP) return NodeSentinel; 383 return this; 384 } 385 386 // let the subclass continue analyzing... 387 return NULL; 388 } 389 390 // Helper function for proving some simple control dominations. 391 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 392 // Already assumes that 'dom' is available at 'sub', and that 'sub' 393 // is not a constant (dominated by the method's StartNode). 394 // Used by MemNode::find_previous_store to prove that the 395 // control input of a memory operation predates (dominates) 396 // an allocation it wants to look past. 397 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 398 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 399 return false; // Conservative answer for dead code 400 401 // Check 'dom'. Skip Proj and CatchProj nodes. 402 dom = dom->find_exact_control(dom); 403 if (dom == NULL || dom->is_top()) 404 return false; // Conservative answer for dead code 405 406 if (dom == sub) { 407 // For the case when, for example, 'sub' is Initialize and the original 408 // 'dom' is Proj node of the 'sub'. 409 return false; 410 } 411 412 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 413 return true; 414 415 // 'dom' dominates 'sub' if its control edge and control edges 416 // of all its inputs dominate or equal to sub's control edge. 417 418 // Currently 'sub' is either Allocate, Initialize or Start nodes. 419 // Or Region for the check in LoadNode::Ideal(); 420 // 'sub' should have sub->in(0) != NULL. 421 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 422 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 423 424 // Get control edge of 'sub'. 425 Node* orig_sub = sub; 426 sub = sub->find_exact_control(sub->in(0)); 427 if (sub == NULL || sub->is_top()) 428 return false; // Conservative answer for dead code 429 430 assert(sub->is_CFG(), "expecting control"); 431 432 if (sub == dom) 433 return true; 434 435 if (sub->is_Start() || sub->is_Root()) 436 return false; 437 438 { 439 // Check all control edges of 'dom'. 440 441 ResourceMark rm; 442 Arena* arena = Thread::current()->resource_area(); 443 Node_List nlist(arena); 444 Unique_Node_List dom_list(arena); 445 446 dom_list.push(dom); 447 bool only_dominating_controls = false; 448 449 for (uint next = 0; next < dom_list.size(); next++) { 450 Node* n = dom_list.at(next); 451 if (n == orig_sub) 452 return false; // One of dom's inputs dominated by sub. 453 if (!n->is_CFG() && n->pinned()) { 454 // Check only own control edge for pinned non-control nodes. 455 n = n->find_exact_control(n->in(0)); 456 if (n == NULL || n->is_top()) 457 return false; // Conservative answer for dead code 458 assert(n->is_CFG(), "expecting control"); 459 dom_list.push(n); 460 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 461 only_dominating_controls = true; 462 } else if (n->is_CFG()) { 463 if (n->dominates(sub, nlist)) 464 only_dominating_controls = true; 465 else 466 return false; 467 } else { 468 // First, own control edge. 469 Node* m = n->find_exact_control(n->in(0)); 470 if (m != NULL) { 471 if (m->is_top()) 472 return false; // Conservative answer for dead code 473 dom_list.push(m); 474 } 475 // Now, the rest of edges. 476 uint cnt = n->req(); 477 for (uint i = 1; i < cnt; i++) { 478 m = n->find_exact_control(n->in(i)); 479 if (m == NULL || m->is_top()) 480 continue; 481 dom_list.push(m); 482 } 483 } 484 } 485 return only_dominating_controls; 486 } 487 } 488 489 //---------------------detect_ptr_independence--------------------------------- 490 // Used by MemNode::find_previous_store to prove that two base 491 // pointers are never equal. 492 // The pointers are accompanied by their associated allocations, 493 // if any, which have been previously discovered by the caller. 494 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 495 Node* p2, AllocateNode* a2, 496 PhaseTransform* phase) { 497 // Attempt to prove that these two pointers cannot be aliased. 498 // They may both manifestly be allocations, and they should differ. 499 // Or, if they are not both allocations, they can be distinct constants. 500 // Otherwise, one is an allocation and the other a pre-existing value. 501 if (a1 == NULL && a2 == NULL) { // neither an allocation 502 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 503 } else if (a1 != NULL && a2 != NULL) { // both allocations 504 return (a1 != a2); 505 } else if (a1 != NULL) { // one allocation a1 506 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 507 return all_controls_dominate(p2, a1); 508 } else { //(a2 != NULL) // one allocation a2 509 return all_controls_dominate(p1, a2); 510 } 511 return false; 512 } 513 514 515 // Find an arraycopy that must have set (can_see_stored_value=true) or 516 // could have set (can_see_stored_value=false) the value for this load 517 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 518 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 519 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 520 Node* mb = mem->in(0); 521 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 522 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 523 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 524 if (ac->is_clonebasic()) { 525 intptr_t offset; 526 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 527 if (alloc != NULL && alloc == ld_alloc) { 528 return ac; 529 } 530 } 531 } 532 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 533 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 534 535 if (ac->is_arraycopy_validated() || 536 ac->is_copyof_validated() || 537 ac->is_copyofrange_validated()) { 538 Node* ld_addp = in(MemNode::Address); 539 if (ld_addp->is_AddP()) { 540 Node* ld_base = ld_addp->in(AddPNode::Address); 541 Node* ld_offs = ld_addp->in(AddPNode::Offset); 542 543 Node* dest = ac->in(ArrayCopyNode::Dest); 544 545 if (dest == ld_base) { 546 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 547 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 548 return ac; 549 } 550 if (!can_see_stored_value) { 551 mem = ac->in(TypeFunc::Memory); 552 } 553 } 554 } 555 } 556 } 557 return NULL; 558 } 559 560 // The logic for reordering loads and stores uses four steps: 561 // (a) Walk carefully past stores and initializations which we 562 // can prove are independent of this load. 563 // (b) Observe that the next memory state makes an exact match 564 // with self (load or store), and locate the relevant store. 565 // (c) Ensure that, if we were to wire self directly to the store, 566 // the optimizer would fold it up somehow. 567 // (d) Do the rewiring, and return, depending on some other part of 568 // the optimizer to fold up the load. 569 // This routine handles steps (a) and (b). Steps (c) and (d) are 570 // specific to loads and stores, so they are handled by the callers. 571 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 572 // 573 Node* MemNode::find_previous_store(PhaseTransform* phase) { 574 Node* ctrl = in(MemNode::Control); 575 Node* adr = in(MemNode::Address); 576 intptr_t offset = 0; 577 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 578 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 579 580 if (offset == Type::OffsetBot) 581 return NULL; // cannot unalias unless there are precise offsets 582 583 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 584 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 585 586 intptr_t size_in_bytes = memory_size(); 587 588 Node* mem = in(MemNode::Memory); // start searching here... 589 590 int cnt = 50; // Cycle limiter 591 for (;;) { // While we can dance past unrelated stores... 592 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 593 594 Node* prev = mem; 595 if (mem->is_Store()) { 596 Node* st_adr = mem->in(MemNode::Address); 597 intptr_t st_offset = 0; 598 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 599 if (st_base == NULL) 600 break; // inscrutable pointer 601 602 // For raw accesses it's not enough to prove that constant offsets don't intersect. 603 // We need the bases to be the equal in order for the offset check to make sense. 604 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 605 break; 606 } 607 608 if (st_offset != offset && st_offset != Type::OffsetBot) { 609 const int MAX_STORE = BytesPerLong; 610 if (st_offset >= offset + size_in_bytes || 611 st_offset <= offset - MAX_STORE || 612 st_offset <= offset - mem->as_Store()->memory_size()) { 613 // Success: The offsets are provably independent. 614 // (You may ask, why not just test st_offset != offset and be done? 615 // The answer is that stores of different sizes can co-exist 616 // in the same sequence of RawMem effects. We sometimes initialize 617 // a whole 'tile' of array elements with a single jint or jlong.) 618 mem = mem->in(MemNode::Memory); 619 continue; // (a) advance through independent store memory 620 } 621 } 622 if (st_base != base && 623 detect_ptr_independence(base, alloc, 624 st_base, 625 AllocateNode::Ideal_allocation(st_base, phase), 626 phase)) { 627 // Success: The bases are provably independent. 628 mem = mem->in(MemNode::Memory); 629 continue; // (a) advance through independent store memory 630 } 631 632 // (b) At this point, if the bases or offsets do not agree, we lose, 633 // since we have not managed to prove 'this' and 'mem' independent. 634 if (st_base == base && st_offset == offset) { 635 return mem; // let caller handle steps (c), (d) 636 } 637 638 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 639 InitializeNode* st_init = mem->in(0)->as_Initialize(); 640 AllocateNode* st_alloc = st_init->allocation(); 641 if (st_alloc == NULL) 642 break; // something degenerated 643 bool known_identical = false; 644 bool known_independent = false; 645 if (alloc == st_alloc) 646 known_identical = true; 647 else if (alloc != NULL) 648 known_independent = true; 649 else if (all_controls_dominate(this, st_alloc)) 650 known_independent = true; 651 652 if (known_independent) { 653 // The bases are provably independent: Either they are 654 // manifestly distinct allocations, or else the control 655 // of this load dominates the store's allocation. 656 int alias_idx = phase->C->get_alias_index(adr_type()); 657 if (alias_idx == Compile::AliasIdxRaw) { 658 mem = st_alloc->in(TypeFunc::Memory); 659 } else { 660 mem = st_init->memory(alias_idx); 661 } 662 continue; // (a) advance through independent store memory 663 } 664 665 // (b) at this point, if we are not looking at a store initializing 666 // the same allocation we are loading from, we lose. 667 if (known_identical) { 668 // From caller, can_see_stored_value will consult find_captured_store. 669 return mem; // let caller handle steps (c), (d) 670 } 671 672 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 673 if (prev != mem) { 674 // Found an arraycopy but it doesn't affect that load 675 continue; 676 } 677 // Found an arraycopy that may affect that load 678 return mem; 679 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 680 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 681 if (mem->is_Proj() && mem->in(0)->is_Call()) { 682 // ArrayCopyNodes processed here as well. 683 CallNode *call = mem->in(0)->as_Call(); 684 if (!call->may_modify(addr_t, phase)) { 685 mem = call->in(TypeFunc::Memory); 686 continue; // (a) advance through independent call memory 687 } 688 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 689 ArrayCopyNode* ac = NULL; 690 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 691 break; 692 } 693 mem = mem->in(0)->in(TypeFunc::Memory); 694 continue; // (a) advance through independent MemBar memory 695 } else if (mem->is_ClearArray()) { 696 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 697 // (the call updated 'mem' value) 698 continue; // (a) advance through independent allocation memory 699 } else { 700 // Can not bypass initialization of the instance 701 // we are looking for. 702 return mem; 703 } 704 } else if (mem->is_MergeMem()) { 705 int alias_idx = phase->C->get_alias_index(adr_type()); 706 mem = mem->as_MergeMem()->memory_at(alias_idx); 707 continue; // (a) advance through independent MergeMem memory 708 } 709 } 710 711 // Unless there is an explicit 'continue', we must bail out here, 712 // because 'mem' is an inscrutable memory state (e.g., a call). 713 break; 714 } 715 716 return NULL; // bail out 717 } 718 719 //----------------------calculate_adr_type------------------------------------- 720 // Helper function. Notices when the given type of address hits top or bottom. 721 // Also, asserts a cross-check of the type against the expected address type. 722 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 723 if (t == Type::TOP) return NULL; // does not touch memory any more? 724 #ifdef PRODUCT 725 cross_check = NULL; 726 #else 727 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL; 728 #endif 729 const TypePtr* tp = t->isa_ptr(); 730 if (tp == NULL) { 731 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 732 return TypePtr::BOTTOM; // touches lots of memory 733 } else { 734 #ifdef ASSERT 735 // %%%% [phh] We don't check the alias index if cross_check is 736 // TypeRawPtr::BOTTOM. Needs to be investigated. 737 if (cross_check != NULL && 738 cross_check != TypePtr::BOTTOM && 739 cross_check != TypeRawPtr::BOTTOM) { 740 // Recheck the alias index, to see if it has changed (due to a bug). 741 Compile* C = Compile::current(); 742 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 743 "must stay in the original alias category"); 744 // The type of the address must be contained in the adr_type, 745 // disregarding "null"-ness. 746 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 747 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 748 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 749 "real address must not escape from expected memory type"); 750 } 751 #endif 752 return tp; 753 } 754 } 755 756 //============================================================================= 757 // Should LoadNode::Ideal() attempt to remove control edges? 758 bool LoadNode::can_remove_control() const { 759 return true; 760 } 761 uint LoadNode::size_of() const { return sizeof(*this); } 762 bool LoadNode::cmp( const Node &n ) const 763 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 764 const Type *LoadNode::bottom_type() const { return _type; } 765 uint LoadNode::ideal_reg() const { 766 return _type->ideal_reg(); 767 } 768 769 #ifndef PRODUCT 770 void LoadNode::dump_spec(outputStream *st) const { 771 MemNode::dump_spec(st); 772 if( !Verbose && !WizardMode ) { 773 // standard dump does this in Verbose and WizardMode 774 st->print(" #"); _type->dump_on(st); 775 } 776 if (!depends_only_on_test()) { 777 st->print(" (does not depend only on test)"); 778 } 779 } 780 #endif 781 782 #ifdef ASSERT 783 //----------------------------is_immutable_value------------------------------- 784 // Helper function to allow a raw load without control edge for some cases 785 bool LoadNode::is_immutable_value(Node* adr) { 786 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 787 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 788 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 789 in_bytes(JavaThread::osthread_offset()))); 790 } 791 #endif 792 793 //----------------------------LoadNode::make----------------------------------- 794 // Polymorphic factory method: 795 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 796 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 797 Compile* C = gvn.C; 798 799 // sanity check the alias category against the created node type 800 assert(!(adr_type->isa_oopptr() && 801 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 802 "use LoadKlassNode instead"); 803 assert(!(adr_type->isa_aryptr() && 804 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 805 "use LoadRangeNode instead"); 806 // Check control edge of raw loads 807 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 808 // oop will be recorded in oop map if load crosses safepoint 809 rt->isa_oopptr() || is_immutable_value(adr), 810 "raw memory operations should have control edge"); 811 LoadNode* load = NULL; 812 switch (bt) { 813 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 814 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 815 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 816 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 817 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 818 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 819 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 820 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 821 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 822 case T_VALUETYPE: 823 case T_OBJECT: 824 #ifdef _LP64 825 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 826 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 827 } else 828 #endif 829 { 830 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 831 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 832 } 833 break; 834 default: 835 ShouldNotReachHere(); 836 break; 837 } 838 assert(load != NULL, "LoadNode should have been created"); 839 if (unaligned) { 840 load->set_unaligned_access(); 841 } 842 if (mismatched) { 843 load->set_mismatched_access(); 844 } 845 if (unsafe) { 846 load->set_unsafe_access(); 847 } 848 if (load->Opcode() == Op_LoadN) { 849 Node* ld = gvn.transform(load); 850 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 851 } 852 853 return load; 854 } 855 856 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 857 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 858 bool require_atomic = true; 859 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 860 if (unaligned) { 861 load->set_unaligned_access(); 862 } 863 if (mismatched) { 864 load->set_mismatched_access(); 865 } 866 if (unsafe) { 867 load->set_unsafe_access(); 868 } 869 return load; 870 } 871 872 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 873 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 874 bool require_atomic = true; 875 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 876 if (unaligned) { 877 load->set_unaligned_access(); 878 } 879 if (mismatched) { 880 load->set_mismatched_access(); 881 } 882 if (unsafe) { 883 load->set_unsafe_access(); 884 } 885 return load; 886 } 887 888 889 890 //------------------------------hash------------------------------------------- 891 uint LoadNode::hash() const { 892 // unroll addition of interesting fields 893 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 894 } 895 896 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 897 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 898 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 899 bool is_stable_ary = FoldStableValues && 900 (tp != NULL) && (tp->isa_aryptr() != NULL) && 901 tp->isa_aryptr()->is_stable(); 902 903 return (eliminate_boxing && non_volatile) || is_stable_ary; 904 } 905 906 return false; 907 } 908 909 // Is the value loaded previously stored by an arraycopy? If so return 910 // a load node that reads from the source array so we may be able to 911 // optimize out the ArrayCopy node later. 912 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 913 #if INCLUDE_ZGC 914 if (UseZGC) { 915 if (bottom_type()->make_oopptr() != NULL) { 916 return NULL; 917 } 918 } 919 #endif 920 921 Node* ld_adr = in(MemNode::Address); 922 intptr_t ld_off = 0; 923 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 924 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 925 if (ac != NULL) { 926 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 927 928 Node* mem = ac->in(TypeFunc::Memory); 929 Node* ctl = ac->in(0); 930 Node* src = ac->in(ArrayCopyNode::Src); 931 932 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 933 return NULL; 934 } 935 936 LoadNode* ld = clone()->as_Load(); 937 Node* addp = in(MemNode::Address)->clone(); 938 if (ac->as_ArrayCopy()->is_clonebasic()) { 939 assert(ld_alloc != NULL, "need an alloc"); 940 assert(addp->is_AddP(), "address must be addp"); 941 assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address"); 942 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 943 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern"); 944 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern"); 945 addp->set_req(AddPNode::Base, src->in(AddPNode::Base)); 946 addp->set_req(AddPNode::Address, src->in(AddPNode::Address)); 947 } else { 948 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 949 ac->as_ArrayCopy()->is_copyof_validated() || 950 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 951 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 952 addp->set_req(AddPNode::Base, src); 953 addp->set_req(AddPNode::Address, src); 954 955 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 956 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 957 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 958 uint shift = exact_log2(type2aelembytes(ary_elem)); 959 960 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 961 #ifdef _LP64 962 diff = phase->transform(new ConvI2LNode(diff)); 963 #endif 964 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 965 966 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 967 addp->set_req(AddPNode::Offset, offset); 968 } 969 addp = phase->transform(addp); 970 #ifdef ASSERT 971 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 972 ld->_adr_type = adr_type; 973 #endif 974 ld->set_req(MemNode::Address, addp); 975 ld->set_req(0, ctl); 976 ld->set_req(MemNode::Memory, mem); 977 // load depends on the tests that validate the arraycopy 978 ld->_control_dependency = Pinned; 979 return ld; 980 } 981 return NULL; 982 } 983 984 985 //---------------------------can_see_stored_value------------------------------ 986 // This routine exists to make sure this set of tests is done the same 987 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 988 // will change the graph shape in a way which makes memory alive twice at the 989 // same time (uses the Oracle model of aliasing), then some 990 // LoadXNode::Identity will fold things back to the equivalence-class model 991 // of aliasing. 992 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 993 Node* ld_adr = in(MemNode::Address); 994 intptr_t ld_off = 0; 995 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off); 996 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase); 997 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 998 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 999 // This is more general than load from boxing objects. 1000 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1001 uint alias_idx = atp->index(); 1002 bool final = !atp->is_rewritable(); 1003 Node* result = NULL; 1004 Node* current = st; 1005 // Skip through chains of MemBarNodes checking the MergeMems for 1006 // new states for the slice of this load. Stop once any other 1007 // kind of node is encountered. Loads from final memory can skip 1008 // through any kind of MemBar but normal loads shouldn't skip 1009 // through MemBarAcquire since the could allow them to move out of 1010 // a synchronized region. 1011 while (current->is_Proj()) { 1012 int opc = current->in(0)->Opcode(); 1013 if ((final && (opc == Op_MemBarAcquire || 1014 opc == Op_MemBarAcquireLock || 1015 opc == Op_LoadFence)) || 1016 opc == Op_MemBarRelease || 1017 opc == Op_StoreFence || 1018 opc == Op_MemBarReleaseLock || 1019 opc == Op_MemBarStoreStore || 1020 opc == Op_MemBarCPUOrder) { 1021 Node* mem = current->in(0)->in(TypeFunc::Memory); 1022 if (mem->is_MergeMem()) { 1023 MergeMemNode* merge = mem->as_MergeMem(); 1024 Node* new_st = merge->memory_at(alias_idx); 1025 if (new_st == merge->base_memory()) { 1026 // Keep searching 1027 current = new_st; 1028 continue; 1029 } 1030 // Save the new memory state for the slice and fall through 1031 // to exit. 1032 result = new_st; 1033 } 1034 } 1035 break; 1036 } 1037 if (result != NULL) { 1038 st = result; 1039 } 1040 } 1041 1042 // Loop around twice in the case Load -> Initialize -> Store. 1043 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1044 for (int trip = 0; trip <= 1; trip++) { 1045 1046 if (st->is_Store()) { 1047 Node* st_adr = st->in(MemNode::Address); 1048 if (!phase->eqv(st_adr, ld_adr)) { 1049 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers). 1050 intptr_t st_off = 0; 1051 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off); 1052 if (ld_base == NULL) return NULL; 1053 if (st_base == NULL) return NULL; 1054 if (ld_base->uncast() != st_base->uncast()) return NULL; 1055 if (ld_off != st_off) return NULL; 1056 if (ld_off == Type::OffsetBot) return NULL; 1057 // Same base, same offset. 1058 // Possible improvement for arrays: check index value instead of absolute offset. 1059 1060 // At this point we have proven something like this setup: 1061 // B = << base >> 1062 // L = LoadQ(AddP(Check/CastPP(B), #Off)) 1063 // S = StoreQ(AddP( B , #Off), V) 1064 // (Actually, we haven't yet proven the Q's are the same.) 1065 // In other words, we are loading from a casted version of 1066 // the same pointer-and-offset that we stored to. 1067 // Thus, we are able to replace L by V. 1068 } 1069 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1070 if (store_Opcode() != st->Opcode()) 1071 return NULL; 1072 return st->in(MemNode::ValueIn); 1073 } 1074 1075 // A load from a freshly-created object always returns zero. 1076 // (This can happen after LoadNode::Ideal resets the load's memory input 1077 // to find_captured_store, which returned InitializeNode::zero_memory.) 1078 if (st->is_Proj() && st->in(0)->is_Allocate() && 1079 (st->in(0) == ld_alloc) && 1080 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1081 // return a zero value for the load's basic type 1082 // (This is one of the few places where a generic PhaseTransform 1083 // can create new nodes. Think of it as lazily manifesting 1084 // virtually pre-existing constants.) 1085 assert(memory_type() != T_VALUETYPE, "should not be used for value types"); 1086 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue); 1087 if (default_value != NULL) { 1088 return default_value; 1089 } 1090 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 1091 return phase->zerocon(memory_type()); 1092 } 1093 1094 // A load from an initialization barrier can match a captured store. 1095 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1096 InitializeNode* init = st->in(0)->as_Initialize(); 1097 AllocateNode* alloc = init->allocation(); 1098 if ((alloc != NULL) && (alloc == ld_alloc)) { 1099 // examine a captured store value 1100 st = init->find_captured_store(ld_off, memory_size(), phase); 1101 if (st != NULL) { 1102 continue; // take one more trip around 1103 } 1104 } 1105 } 1106 1107 // Load boxed value from result of valueOf() call is input parameter. 1108 if (this->is_Load() && ld_adr->is_AddP() && 1109 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1110 intptr_t ignore = 0; 1111 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1112 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1113 base = bs->step_over_gc_barrier(base); 1114 if (base != NULL && base->is_Proj() && 1115 base->as_Proj()->_con == TypeFunc::Parms && 1116 base->in(0)->is_CallStaticJava() && 1117 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1118 return base->in(0)->in(TypeFunc::Parms); 1119 } 1120 } 1121 1122 break; 1123 } 1124 1125 return NULL; 1126 } 1127 1128 //----------------------is_instance_field_load_with_local_phi------------------ 1129 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1130 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1131 in(Address)->is_AddP() ) { 1132 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1133 // Only instances and boxed values. 1134 if( t_oop != NULL && 1135 (t_oop->is_ptr_to_boxed_value() || 1136 t_oop->is_known_instance_field()) && 1137 t_oop->offset() != Type::OffsetBot && 1138 t_oop->offset() != Type::OffsetTop) { 1139 return true; 1140 } 1141 } 1142 return false; 1143 } 1144 1145 //------------------------------Identity--------------------------------------- 1146 // Loads are identity if previous store is to same address 1147 Node* LoadNode::Identity(PhaseGVN* phase) { 1148 // Loading from a ValueTypePtr? The ValueTypePtr has the values of 1149 // all fields as input. Look for the field with matching offset. 1150 Node* addr = in(Address); 1151 intptr_t offset; 1152 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset); 1153 if (base != NULL && base->is_ValueTypePtr() && offset > oopDesc::klass_offset_in_bytes()) { 1154 Node* value = base->as_ValueTypePtr()->field_value_by_offset((int)offset, true); 1155 if (value->is_ValueType()) { 1156 // Non-flattened value type field 1157 ValueTypeNode* vt = value->as_ValueType(); 1158 if (vt->is_allocated(phase)) { 1159 value = vt->get_oop(); 1160 } else { 1161 // Not yet allocated, bail out 1162 value = NULL; 1163 } 1164 } 1165 if (value != NULL) { 1166 if (Opcode() == Op_LoadN) { 1167 // Encode oop value if we are loading a narrow oop 1168 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded"); 1169 value = phase->transform(new EncodePNode(value, bottom_type())); 1170 } 1171 return value; 1172 } 1173 } 1174 1175 // If the previous store-maker is the right kind of Store, and the store is 1176 // to the same address, then we are equal to the value stored. 1177 Node* mem = in(Memory); 1178 Node* value = can_see_stored_value(mem, phase); 1179 if( value ) { 1180 // byte, short & char stores truncate naturally. 1181 // A load has to load the truncated value which requires 1182 // some sort of masking operation and that requires an 1183 // Ideal call instead of an Identity call. 1184 if (memory_size() < BytesPerInt) { 1185 // If the input to the store does not fit with the load's result type, 1186 // it must be truncated via an Ideal call. 1187 if (!phase->type(value)->higher_equal(phase->type(this))) 1188 return this; 1189 } 1190 // (This works even when value is a Con, but LoadNode::Value 1191 // usually runs first, producing the singleton type of the Con.) 1192 return value; 1193 } 1194 1195 // Search for an existing data phi which was generated before for the same 1196 // instance's field to avoid infinite generation of phis in a loop. 1197 Node *region = mem->in(0); 1198 if (is_instance_field_load_with_local_phi(region)) { 1199 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1200 int this_index = phase->C->get_alias_index(addr_t); 1201 int this_offset = addr_t->offset(); 1202 int this_iid = addr_t->instance_id(); 1203 if (!addr_t->is_known_instance() && 1204 addr_t->is_ptr_to_boxed_value()) { 1205 // Use _idx of address base (could be Phi node) for boxed values. 1206 intptr_t ignore = 0; 1207 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1208 if (base == NULL) { 1209 return this; 1210 } 1211 this_iid = base->_idx; 1212 } 1213 const Type* this_type = bottom_type(); 1214 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1215 Node* phi = region->fast_out(i); 1216 if (phi->is_Phi() && phi != mem && 1217 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1218 return phi; 1219 } 1220 } 1221 } 1222 1223 return this; 1224 } 1225 1226 // Construct an equivalent unsigned load. 1227 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1228 BasicType bt = T_ILLEGAL; 1229 const Type* rt = NULL; 1230 switch (Opcode()) { 1231 case Op_LoadUB: return this; 1232 case Op_LoadUS: return this; 1233 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1234 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1235 default: 1236 assert(false, "no unsigned variant: %s", Name()); 1237 return NULL; 1238 } 1239 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1240 raw_adr_type(), rt, bt, _mo, _control_dependency, 1241 is_unaligned_access(), is_mismatched_access()); 1242 } 1243 1244 // Construct an equivalent signed load. 1245 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1246 BasicType bt = T_ILLEGAL; 1247 const Type* rt = NULL; 1248 switch (Opcode()) { 1249 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1250 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1251 case Op_LoadB: // fall through 1252 case Op_LoadS: // fall through 1253 case Op_LoadI: // fall through 1254 case Op_LoadL: return this; 1255 default: 1256 assert(false, "no signed variant: %s", Name()); 1257 return NULL; 1258 } 1259 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1260 raw_adr_type(), rt, bt, _mo, _control_dependency, 1261 is_unaligned_access(), is_mismatched_access()); 1262 } 1263 1264 // We're loading from an object which has autobox behaviour. 1265 // If this object is result of a valueOf call we'll have a phi 1266 // merging a newly allocated object and a load from the cache. 1267 // We want to replace this load with the original incoming 1268 // argument to the valueOf call. 1269 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1270 assert(phase->C->eliminate_boxing(), "sanity"); 1271 intptr_t ignore = 0; 1272 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1273 if ((base == NULL) || base->is_Phi()) { 1274 // Push the loads from the phi that comes from valueOf up 1275 // through it to allow elimination of the loads and the recovery 1276 // of the original value. It is done in split_through_phi(). 1277 return NULL; 1278 } else if (base->is_Load() || 1279 (base->is_DecodeN() && base->in(1)->is_Load())) { 1280 // Eliminate the load of boxed value for integer types from the cache 1281 // array by deriving the value from the index into the array. 1282 // Capture the offset of the load and then reverse the computation. 1283 1284 // Get LoadN node which loads a boxing object from 'cache' array. 1285 if (base->is_DecodeN()) { 1286 base = base->in(1); 1287 } 1288 if (!base->in(Address)->is_AddP()) { 1289 return NULL; // Complex address 1290 } 1291 AddPNode* address = base->in(Address)->as_AddP(); 1292 Node* cache_base = address->in(AddPNode::Base); 1293 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1294 // Get ConP node which is static 'cache' field. 1295 cache_base = cache_base->in(1); 1296 } 1297 if ((cache_base != NULL) && cache_base->is_Con()) { 1298 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1299 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1300 Node* elements[4]; 1301 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1302 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1303 if (count > 0 && elements[0]->is_Con() && 1304 (count == 1 || 1305 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1306 elements[1]->in(2) == phase->intcon(shift)))) { 1307 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1308 // Fetch the box object cache[0] at the base of the array and get its value 1309 ciInstance* box = array->obj_at(0)->as_instance(); 1310 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1311 assert(ik->is_box_klass(), "sanity"); 1312 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1313 if (ik->nof_nonstatic_fields() == 1) { 1314 // This should be true nonstatic_field_at requires calling 1315 // nof_nonstatic_fields so check it anyway 1316 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1317 BasicType bt = c.basic_type(); 1318 // Only integer types have boxing cache. 1319 assert(bt == T_BOOLEAN || bt == T_CHAR || 1320 bt == T_BYTE || bt == T_SHORT || 1321 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1322 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1323 if (cache_low != (int)cache_low) { 1324 return NULL; // should not happen since cache is array indexed by value 1325 } 1326 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1327 if (offset != (int)offset) { 1328 return NULL; // should not happen since cache is array indexed by value 1329 } 1330 // Add up all the offsets making of the address of the load 1331 Node* result = elements[0]; 1332 for (int i = 1; i < count; i++) { 1333 result = phase->transform(new AddXNode(result, elements[i])); 1334 } 1335 // Remove the constant offset from the address and then 1336 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1337 // remove the scaling of the offset to recover the original index. 1338 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1339 // Peel the shift off directly but wrap it in a dummy node 1340 // since Ideal can't return existing nodes 1341 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1342 } else if (result->is_Add() && result->in(2)->is_Con() && 1343 result->in(1)->Opcode() == Op_LShiftX && 1344 result->in(1)->in(2) == phase->intcon(shift)) { 1345 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1346 // but for boxing cache access we know that X<<Z will not overflow 1347 // (there is range check) so we do this optimizatrion by hand here. 1348 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1349 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1350 } else { 1351 result = new RShiftXNode(result, phase->intcon(shift)); 1352 } 1353 #ifdef _LP64 1354 if (bt != T_LONG) { 1355 result = new ConvL2INode(phase->transform(result)); 1356 } 1357 #else 1358 if (bt == T_LONG) { 1359 result = new ConvI2LNode(phase->transform(result)); 1360 } 1361 #endif 1362 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1363 // Need to preserve unboxing load type if it is unsigned. 1364 switch(this->Opcode()) { 1365 case Op_LoadUB: 1366 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1367 break; 1368 case Op_LoadUS: 1369 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1370 break; 1371 } 1372 return result; 1373 } 1374 } 1375 } 1376 } 1377 } 1378 return NULL; 1379 } 1380 1381 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1382 Node* region = phi->in(0); 1383 if (region == NULL) { 1384 return false; // Wait stable graph 1385 } 1386 uint cnt = phi->req(); 1387 for (uint i = 1; i < cnt; i++) { 1388 Node* rc = region->in(i); 1389 if (rc == NULL || phase->type(rc) == Type::TOP) 1390 return false; // Wait stable graph 1391 Node* in = phi->in(i); 1392 if (in == NULL || phase->type(in) == Type::TOP) 1393 return false; // Wait stable graph 1394 } 1395 return true; 1396 } 1397 //------------------------------split_through_phi------------------------------ 1398 // Split instance or boxed field load through Phi. 1399 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1400 Node* mem = in(Memory); 1401 Node* address = in(Address); 1402 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1403 1404 assert((t_oop != NULL) && 1405 (t_oop->is_known_instance_field() || 1406 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1407 1408 Compile* C = phase->C; 1409 intptr_t ignore = 0; 1410 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1411 bool base_is_phi = (base != NULL) && base->is_Phi(); 1412 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1413 (base != NULL) && (base == address->in(AddPNode::Base)) && 1414 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1415 1416 if (!((mem->is_Phi() || base_is_phi) && 1417 (load_boxed_values || t_oop->is_known_instance_field()))) { 1418 return NULL; // memory is not Phi 1419 } 1420 1421 if (mem->is_Phi()) { 1422 if (!stable_phi(mem->as_Phi(), phase)) { 1423 return NULL; // Wait stable graph 1424 } 1425 uint cnt = mem->req(); 1426 // Check for loop invariant memory. 1427 if (cnt == 3) { 1428 for (uint i = 1; i < cnt; i++) { 1429 Node* in = mem->in(i); 1430 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1431 if (m == mem) { 1432 if (i == 1) { 1433 // if the first edge was a loop, check second edge too. 1434 // If both are replaceable - we are in an infinite loop 1435 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1436 if (n == mem) { 1437 break; 1438 } 1439 } 1440 set_req(Memory, mem->in(cnt - i)); 1441 return this; // made change 1442 } 1443 } 1444 } 1445 } 1446 if (base_is_phi) { 1447 if (!stable_phi(base->as_Phi(), phase)) { 1448 return NULL; // Wait stable graph 1449 } 1450 uint cnt = base->req(); 1451 // Check for loop invariant memory. 1452 if (cnt == 3) { 1453 for (uint i = 1; i < cnt; i++) { 1454 if (base->in(i) == base) { 1455 return NULL; // Wait stable graph 1456 } 1457 } 1458 } 1459 } 1460 1461 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1462 1463 // Split through Phi (see original code in loopopts.cpp). 1464 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1465 1466 // Do nothing here if Identity will find a value 1467 // (to avoid infinite chain of value phis generation). 1468 if (!phase->eqv(this, phase->apply_identity(this))) 1469 return NULL; 1470 1471 // Select Region to split through. 1472 Node* region; 1473 if (!base_is_phi) { 1474 assert(mem->is_Phi(), "sanity"); 1475 region = mem->in(0); 1476 // Skip if the region dominates some control edge of the address. 1477 if (!MemNode::all_controls_dominate(address, region)) 1478 return NULL; 1479 } else if (!mem->is_Phi()) { 1480 assert(base_is_phi, "sanity"); 1481 region = base->in(0); 1482 // Skip if the region dominates some control edge of the memory. 1483 if (!MemNode::all_controls_dominate(mem, region)) 1484 return NULL; 1485 } else if (base->in(0) != mem->in(0)) { 1486 assert(base_is_phi && mem->is_Phi(), "sanity"); 1487 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1488 region = base->in(0); 1489 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1490 region = mem->in(0); 1491 } else { 1492 return NULL; // complex graph 1493 } 1494 } else { 1495 assert(base->in(0) == mem->in(0), "sanity"); 1496 region = mem->in(0); 1497 } 1498 1499 const Type* this_type = this->bottom_type(); 1500 int this_index = C->get_alias_index(t_oop); 1501 int this_offset = t_oop->offset(); 1502 int this_iid = t_oop->instance_id(); 1503 if (!t_oop->is_known_instance() && load_boxed_values) { 1504 // Use _idx of address base for boxed values. 1505 this_iid = base->_idx; 1506 } 1507 PhaseIterGVN* igvn = phase->is_IterGVN(); 1508 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1509 for (uint i = 1; i < region->req(); i++) { 1510 Node* x; 1511 Node* the_clone = NULL; 1512 if (region->in(i) == C->top()) { 1513 x = C->top(); // Dead path? Use a dead data op 1514 } else { 1515 x = this->clone(); // Else clone up the data op 1516 the_clone = x; // Remember for possible deletion. 1517 // Alter data node to use pre-phi inputs 1518 if (this->in(0) == region) { 1519 x->set_req(0, region->in(i)); 1520 } else { 1521 x->set_req(0, NULL); 1522 } 1523 if (mem->is_Phi() && (mem->in(0) == region)) { 1524 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1525 } 1526 if (address->is_Phi() && address->in(0) == region) { 1527 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1528 } 1529 if (base_is_phi && (base->in(0) == region)) { 1530 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1531 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1532 x->set_req(Address, adr_x); 1533 } 1534 } 1535 // Check for a 'win' on some paths 1536 const Type *t = x->Value(igvn); 1537 1538 bool singleton = t->singleton(); 1539 1540 // See comments in PhaseIdealLoop::split_thru_phi(). 1541 if (singleton && t == Type::TOP) { 1542 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1543 } 1544 1545 if (singleton) { 1546 x = igvn->makecon(t); 1547 } else { 1548 // We now call Identity to try to simplify the cloned node. 1549 // Note that some Identity methods call phase->type(this). 1550 // Make sure that the type array is big enough for 1551 // our new node, even though we may throw the node away. 1552 // (This tweaking with igvn only works because x is a new node.) 1553 igvn->set_type(x, t); 1554 // If x is a TypeNode, capture any more-precise type permanently into Node 1555 // otherwise it will be not updated during igvn->transform since 1556 // igvn->type(x) is set to x->Value() already. 1557 x->raise_bottom_type(t); 1558 Node *y = igvn->apply_identity(x); 1559 if (y != x) { 1560 x = y; 1561 } else { 1562 y = igvn->hash_find_insert(x); 1563 if (y) { 1564 x = y; 1565 } else { 1566 // Else x is a new node we are keeping 1567 // We do not need register_new_node_with_optimizer 1568 // because set_type has already been called. 1569 igvn->_worklist.push(x); 1570 } 1571 } 1572 } 1573 if (x != the_clone && the_clone != NULL) { 1574 igvn->remove_dead_node(the_clone); 1575 } 1576 phi->set_req(i, x); 1577 } 1578 // Record Phi 1579 igvn->register_new_node_with_optimizer(phi); 1580 return phi; 1581 } 1582 1583 //------------------------------Ideal------------------------------------------ 1584 // If the load is from Field memory and the pointer is non-null, it might be possible to 1585 // zero out the control input. 1586 // If the offset is constant and the base is an object allocation, 1587 // try to hook me up to the exact initializing store. 1588 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1589 Node* p = MemNode::Ideal_common(phase, can_reshape); 1590 if (p) return (p == NodeSentinel) ? NULL : p; 1591 1592 Node* ctrl = in(MemNode::Control); 1593 Node* address = in(MemNode::Address); 1594 bool progress = false; 1595 1596 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1597 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1598 1599 // Skip up past a SafePoint control. Cannot do this for Stores because 1600 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1601 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1602 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1603 !addr_mark ) { 1604 ctrl = ctrl->in(0); 1605 set_req(MemNode::Control,ctrl); 1606 progress = true; 1607 } 1608 1609 intptr_t ignore = 0; 1610 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1611 if (base != NULL 1612 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1613 // Check for useless control edge in some common special cases 1614 if (in(MemNode::Control) != NULL 1615 && can_remove_control() 1616 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1617 && all_controls_dominate(base, phase->C->start())) { 1618 // A method-invariant, non-null address (constant or 'this' argument). 1619 set_req(MemNode::Control, NULL); 1620 progress = true; 1621 } 1622 } 1623 1624 Node* mem = in(MemNode::Memory); 1625 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1626 1627 if (can_reshape && (addr_t != NULL)) { 1628 // try to optimize our memory input 1629 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1630 if (opt_mem != mem) { 1631 set_req(MemNode::Memory, opt_mem); 1632 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1633 return this; 1634 } 1635 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1636 if ((t_oop != NULL) && 1637 (t_oop->is_known_instance_field() || 1638 t_oop->is_ptr_to_boxed_value())) { 1639 PhaseIterGVN *igvn = phase->is_IterGVN(); 1640 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1641 // Delay this transformation until memory Phi is processed. 1642 phase->is_IterGVN()->_worklist.push(this); 1643 return NULL; 1644 } 1645 // Split instance field load through Phi. 1646 Node* result = split_through_phi(phase); 1647 if (result != NULL) return result; 1648 1649 if (t_oop->is_ptr_to_boxed_value()) { 1650 Node* result = eliminate_autobox(phase); 1651 if (result != NULL) return result; 1652 } 1653 } 1654 } 1655 1656 // Is there a dominating load that loads the same value? Leave 1657 // anything that is not a load of a field/array element (like 1658 // barriers etc.) alone 1659 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) { 1660 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1661 Node *use = mem->fast_out(i); 1662 if (use != this && 1663 use->Opcode() == Opcode() && 1664 use->in(0) != NULL && 1665 use->in(0) != in(0) && 1666 use->in(Address) == in(Address)) { 1667 Node* ctl = in(0); 1668 for (int i = 0; i < 10 && ctl != NULL; i++) { 1669 ctl = IfNode::up_one_dom(ctl); 1670 if (ctl == use->in(0)) { 1671 set_req(0, use->in(0)); 1672 return this; 1673 } 1674 } 1675 } 1676 } 1677 } 1678 1679 // Check for prior store with a different base or offset; make Load 1680 // independent. Skip through any number of them. Bail out if the stores 1681 // are in an endless dead cycle and report no progress. This is a key 1682 // transform for Reflection. However, if after skipping through the Stores 1683 // we can't then fold up against a prior store do NOT do the transform as 1684 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1685 // array memory alive twice: once for the hoisted Load and again after the 1686 // bypassed Store. This situation only works if EVERYBODY who does 1687 // anti-dependence work knows how to bypass. I.e. we need all 1688 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1689 // the alias index stuff. So instead, peek through Stores and IFF we can 1690 // fold up, do so. 1691 Node* prev_mem = find_previous_store(phase); 1692 if (prev_mem != NULL) { 1693 Node* value = can_see_arraycopy_value(prev_mem, phase); 1694 if (value != NULL) { 1695 return value; 1696 } 1697 } 1698 // Steps (a), (b): Walk past independent stores to find an exact match. 1699 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1700 // (c) See if we can fold up on the spot, but don't fold up here. 1701 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1702 // just return a prior value, which is done by Identity calls. 1703 if (can_see_stored_value(prev_mem, phase)) { 1704 // Make ready for step (d): 1705 set_req(MemNode::Memory, prev_mem); 1706 return this; 1707 } 1708 } 1709 1710 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase); 1711 if (alloc != NULL && mem->is_Proj() && 1712 mem->in(0) != NULL && 1713 mem->in(0) == alloc->initialization() && 1714 Opcode() == Op_LoadX && 1715 alloc->initialization()->proj_out_or_null(0) != NULL) { 1716 InitializeNode* init = alloc->initialization(); 1717 Node* control = init->proj_out(0); 1718 return alloc->make_ideal_mark(phase, address, control, mem, NULL); 1719 } 1720 1721 return progress ? this : NULL; 1722 } 1723 1724 // Helper to recognize certain Klass fields which are invariant across 1725 // some group of array types (e.g., int[] or all T[] where T < Object). 1726 const Type* 1727 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1728 ciKlass* klass) const { 1729 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1730 // The field is Klass::_modifier_flags. Return its (constant) value. 1731 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1732 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1733 return TypeInt::make(klass->modifier_flags()); 1734 } 1735 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1736 // The field is Klass::_access_flags. Return its (constant) value. 1737 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1738 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1739 return TypeInt::make(klass->access_flags()); 1740 } 1741 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1742 // The field is Klass::_layout_helper. Return its constant value if known. 1743 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1744 return TypeInt::make(klass->layout_helper()); 1745 } 1746 1747 // No match. 1748 return NULL; 1749 } 1750 1751 //------------------------------Value----------------------------------------- 1752 const Type* LoadNode::Value(PhaseGVN* phase) const { 1753 // Either input is TOP ==> the result is TOP 1754 Node* mem = in(MemNode::Memory); 1755 const Type *t1 = phase->type(mem); 1756 if (t1 == Type::TOP) return Type::TOP; 1757 Node* adr = in(MemNode::Address); 1758 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1759 if (tp == NULL || tp->empty()) return Type::TOP; 1760 int off = tp->offset(); 1761 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1762 Compile* C = phase->C; 1763 1764 // Try to guess loaded type from pointer type 1765 if (tp->isa_aryptr()) { 1766 const TypeAryPtr* ary = tp->is_aryptr(); 1767 const Type* t = ary->elem(); 1768 1769 // Determine whether the reference is beyond the header or not, by comparing 1770 // the offset against the offset of the start of the array's data. 1771 // Different array types begin at slightly different offsets (12 vs. 16). 1772 // We choose T_BYTE as an example base type that is least restrictive 1773 // as to alignment, which will therefore produce the smallest 1774 // possible base offset. 1775 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1776 const bool off_beyond_header = (off >= min_base_off); 1777 1778 // Try to constant-fold a stable array element. 1779 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1780 // Make sure the reference is not into the header and the offset is constant 1781 ciObject* aobj = ary->const_oop(); 1782 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1783 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1784 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1785 stable_dimension, 1786 memory_type(), is_unsigned()); 1787 if (con_type != NULL) { 1788 return con_type; 1789 } 1790 } 1791 } 1792 1793 // Don't do this for integer types. There is only potential profit if 1794 // the element type t is lower than _type; that is, for int types, if _type is 1795 // more restrictive than t. This only happens here if one is short and the other 1796 // char (both 16 bits), and in those cases we've made an intentional decision 1797 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1798 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1799 // 1800 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1801 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1802 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1803 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1804 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1805 // In fact, that could have been the original type of p1, and p1 could have 1806 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1807 // expression (LShiftL quux 3) independently optimized to the constant 8. 1808 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1809 && (_type->isa_vect() == NULL) 1810 && t->isa_valuetype() == NULL 1811 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1812 // t might actually be lower than _type, if _type is a unique 1813 // concrete subclass of abstract class t. 1814 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header? 1815 const Type* jt = t->join_speculative(_type); 1816 // In any case, do not allow the join, per se, to empty out the type. 1817 if (jt->empty() && !t->empty()) { 1818 // This can happen if a interface-typed array narrows to a class type. 1819 jt = _type; 1820 } 1821 #ifdef ASSERT 1822 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1823 // The pointers in the autobox arrays are always non-null 1824 Node* base = adr->in(AddPNode::Base); 1825 if ((base != NULL) && base->is_DecodeN()) { 1826 // Get LoadN node which loads IntegerCache.cache field 1827 base = base->in(1); 1828 } 1829 if ((base != NULL) && base->is_Con()) { 1830 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1831 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1832 // It could be narrow oop 1833 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1834 } 1835 } 1836 } 1837 #endif 1838 return jt; 1839 } 1840 } 1841 } else if (tp->base() == Type::InstPtr) { 1842 assert( off != Type::OffsetBot || 1843 // arrays can be cast to Objects 1844 tp->is_oopptr()->klass()->is_java_lang_Object() || 1845 tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() || 1846 // unsafe field access may not have a constant offset 1847 C->has_unsafe_access(), 1848 "Field accesses must be precise" ); 1849 // For oop loads, we expect the _type to be precise. 1850 1851 // Optimize loads from constant fields. 1852 const TypeInstPtr* tinst = tp->is_instptr(); 1853 ciObject* const_oop = tinst->const_oop(); 1854 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1855 BasicType bt = memory_type(); 1856 ciType* mirror_type = const_oop->as_instance()->java_mirror_type(); 1857 if (mirror_type != NULL && mirror_type->is_valuetype()) { 1858 ciValueKlass* vk = mirror_type->as_value_klass(); 1859 if (off == vk->default_value_offset()) { 1860 // Loading a special hidden field that contains the oop of the default value type 1861 const Type* const_oop = TypeInstPtr::make(vk->default_value_instance()); 1862 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop; 1863 } 1864 } 1865 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt); 1866 if (con_type != NULL) { 1867 return con_type; 1868 } 1869 } 1870 } else if (tp->base() == Type::KlassPtr) { 1871 assert( off != Type::OffsetBot || 1872 // arrays can be cast to Objects 1873 tp->is_klassptr()->klass() == NULL || 1874 tp->is_klassptr()->klass()->is_java_lang_Object() || 1875 // also allow array-loading from the primary supertype 1876 // array during subtype checks 1877 Opcode() == Op_LoadKlass, 1878 "Field accesses must be precise" ); 1879 // For klass/static loads, we expect the _type to be precise 1880 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) { 1881 if (adr->is_Load() && off == 0) { 1882 /* With mirrors being an indirect in the Klass* 1883 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 1884 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 1885 * 1886 * So check the type and klass of the node before the LoadP. 1887 */ 1888 Node* adr2 = adr->in(MemNode::Address); 1889 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1890 if (tkls != NULL) { 1891 ciKlass* klass = tkls->klass(); 1892 if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1893 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1894 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1895 return TypeInstPtr::make(klass->java_mirror()); 1896 } 1897 } 1898 } else { 1899 // Check for a load of the default value offset from the ValueKlassFixedBlock: 1900 // LoadI(LoadP(value_klass, adr_valueklass_fixed_block_offset), default_value_offset_offset) 1901 intptr_t offset = 0; 1902 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1903 if (base != NULL && base->is_Load() && offset == in_bytes(ValueKlass::default_value_offset_offset())) { 1904 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr(); 1905 if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_valuetype() && 1906 tkls->offset() == in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())) { 1907 assert(base->Opcode() == Op_LoadP, "must load an oop from klass"); 1908 assert(Opcode() == Op_LoadI, "must load an int from fixed block"); 1909 return TypeInt::make(tkls->klass()->as_value_klass()->default_value_offset()); 1910 } 1911 } 1912 } 1913 } 1914 1915 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1916 if (tkls != NULL && !StressReflectiveCode) { 1917 ciKlass* klass = tkls->klass(); 1918 if (tkls->is_loaded() && tkls->klass_is_exact()) { 1919 // We are loading a field from a Klass metaobject whose identity 1920 // is known at compile time (the type is "exact" or "precise"). 1921 // Check for fields we know are maintained as constants by the VM. 1922 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1923 // The field is Klass::_super_check_offset. Return its (constant) value. 1924 // (Folds up type checking code.) 1925 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1926 return TypeInt::make(klass->super_check_offset()); 1927 } 1928 // Compute index into primary_supers array 1929 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1930 // Check for overflowing; use unsigned compare to handle the negative case. 1931 if( depth < ciKlass::primary_super_limit() ) { 1932 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1933 // (Folds up type checking code.) 1934 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1935 ciKlass *ss = klass->super_of_depth(depth); 1936 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1937 } 1938 const Type* aift = load_array_final_field(tkls, klass); 1939 if (aift != NULL) return aift; 1940 } 1941 1942 // We can still check if we are loading from the primary_supers array at a 1943 // shallow enough depth. Even though the klass is not exact, entries less 1944 // than or equal to its super depth are correct. 1945 if (tkls->is_loaded()) { 1946 ciType *inner = klass; 1947 while( inner->is_obj_array_klass() ) 1948 inner = inner->as_obj_array_klass()->base_element_type(); 1949 if( inner->is_instance_klass() && 1950 !inner->as_instance_klass()->flags().is_interface() ) { 1951 // Compute index into primary_supers array 1952 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1953 // Check for overflowing; use unsigned compare to handle the negative case. 1954 if( depth < ciKlass::primary_super_limit() && 1955 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1956 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1957 // (Folds up type checking code.) 1958 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1959 ciKlass *ss = klass->super_of_depth(depth); 1960 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1961 } 1962 } 1963 } 1964 1965 // If the type is enough to determine that the thing is not an array, 1966 // we can give the layout_helper a positive interval type. 1967 // This will help short-circuit some reflective code. 1968 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1969 && !klass->is_array_klass() // not directly typed as an array 1970 && !klass->is_interface() // specifically not Serializable & Cloneable 1971 && !klass->is_java_lang_Object() // not the supertype of all T[] 1972 ) { 1973 // Note: When interfaces are reliable, we can narrow the interface 1974 // test to (klass != Serializable && klass != Cloneable). 1975 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1976 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1977 // The key property of this type is that it folds up tests 1978 // for array-ness, since it proves that the layout_helper is positive. 1979 // Thus, a generic value like the basic object layout helper works fine. 1980 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1981 } 1982 } 1983 1984 // If we are loading from a freshly-allocated object, produce a zero, 1985 // if the load is provably beyond the header of the object. 1986 // (Also allow a variable load from a fresh array to produce zero.) 1987 const TypeOopPtr *tinst = tp->isa_oopptr(); 1988 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1989 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1990 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1991 Node* value = can_see_stored_value(mem,phase); 1992 if (value != NULL && value->is_Con()) { 1993 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1994 return value->bottom_type(); 1995 } 1996 } 1997 1998 if (is_instance) { 1999 // If we have an instance type and our memory input is the 2000 // programs's initial memory state, there is no matching store, 2001 // so just return a zero of the appropriate type 2002 Node *mem = in(MemNode::Memory); 2003 if (mem->is_Parm() && mem->in(0)->is_Start()) { 2004 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 2005 return Type::get_zero_type(_type->basic_type()); 2006 } 2007 } 2008 return _type; 2009 } 2010 2011 //------------------------------match_edge------------------------------------- 2012 // Do we Match on this edge index or not? Match only the address. 2013 uint LoadNode::match_edge(uint idx) const { 2014 return idx == MemNode::Address; 2015 } 2016 2017 //--------------------------LoadBNode::Ideal-------------------------------------- 2018 // 2019 // If the previous store is to the same address as this load, 2020 // and the value stored was larger than a byte, replace this load 2021 // with the value stored truncated to a byte. If no truncation is 2022 // needed, the replacement is done in LoadNode::Identity(). 2023 // 2024 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2025 Node* mem = in(MemNode::Memory); 2026 Node* value = can_see_stored_value(mem,phase); 2027 if( value && !phase->type(value)->higher_equal( _type ) ) { 2028 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 2029 return new RShiftINode(result, phase->intcon(24)); 2030 } 2031 // Identity call will handle the case where truncation is not needed. 2032 return LoadNode::Ideal(phase, can_reshape); 2033 } 2034 2035 const Type* LoadBNode::Value(PhaseGVN* phase) const { 2036 Node* mem = in(MemNode::Memory); 2037 Node* value = can_see_stored_value(mem,phase); 2038 if (value != NULL && value->is_Con() && 2039 !value->bottom_type()->higher_equal(_type)) { 2040 // If the input to the store does not fit with the load's result type, 2041 // it must be truncated. We can't delay until Ideal call since 2042 // a singleton Value is needed for split_thru_phi optimization. 2043 int con = value->get_int(); 2044 return TypeInt::make((con << 24) >> 24); 2045 } 2046 return LoadNode::Value(phase); 2047 } 2048 2049 //--------------------------LoadUBNode::Ideal------------------------------------- 2050 // 2051 // If the previous store is to the same address as this load, 2052 // and the value stored was larger than a byte, replace this load 2053 // with the value stored truncated to a byte. If no truncation is 2054 // needed, the replacement is done in LoadNode::Identity(). 2055 // 2056 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2057 Node* mem = in(MemNode::Memory); 2058 Node* value = can_see_stored_value(mem, phase); 2059 if (value && !phase->type(value)->higher_equal(_type)) 2060 return new AndINode(value, phase->intcon(0xFF)); 2061 // Identity call will handle the case where truncation is not needed. 2062 return LoadNode::Ideal(phase, can_reshape); 2063 } 2064 2065 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2066 Node* mem = in(MemNode::Memory); 2067 Node* value = can_see_stored_value(mem,phase); 2068 if (value != NULL && value->is_Con() && 2069 !value->bottom_type()->higher_equal(_type)) { 2070 // If the input to the store does not fit with the load's result type, 2071 // it must be truncated. We can't delay until Ideal call since 2072 // a singleton Value is needed for split_thru_phi optimization. 2073 int con = value->get_int(); 2074 return TypeInt::make(con & 0xFF); 2075 } 2076 return LoadNode::Value(phase); 2077 } 2078 2079 //--------------------------LoadUSNode::Ideal------------------------------------- 2080 // 2081 // If the previous store is to the same address as this load, 2082 // and the value stored was larger than a char, replace this load 2083 // with the value stored truncated to a char. If no truncation is 2084 // needed, the replacement is done in LoadNode::Identity(). 2085 // 2086 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2087 Node* mem = in(MemNode::Memory); 2088 Node* value = can_see_stored_value(mem,phase); 2089 if( value && !phase->type(value)->higher_equal( _type ) ) 2090 return new AndINode(value,phase->intcon(0xFFFF)); 2091 // Identity call will handle the case where truncation is not needed. 2092 return LoadNode::Ideal(phase, can_reshape); 2093 } 2094 2095 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2096 Node* mem = in(MemNode::Memory); 2097 Node* value = can_see_stored_value(mem,phase); 2098 if (value != NULL && value->is_Con() && 2099 !value->bottom_type()->higher_equal(_type)) { 2100 // If the input to the store does not fit with the load's result type, 2101 // it must be truncated. We can't delay until Ideal call since 2102 // a singleton Value is needed for split_thru_phi optimization. 2103 int con = value->get_int(); 2104 return TypeInt::make(con & 0xFFFF); 2105 } 2106 return LoadNode::Value(phase); 2107 } 2108 2109 //--------------------------LoadSNode::Ideal-------------------------------------- 2110 // 2111 // If the previous store is to the same address as this load, 2112 // and the value stored was larger than a short, replace this load 2113 // with the value stored truncated to a short. If no truncation is 2114 // needed, the replacement is done in LoadNode::Identity(). 2115 // 2116 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2117 Node* mem = in(MemNode::Memory); 2118 Node* value = can_see_stored_value(mem,phase); 2119 if( value && !phase->type(value)->higher_equal( _type ) ) { 2120 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 2121 return new RShiftINode(result, phase->intcon(16)); 2122 } 2123 // Identity call will handle the case where truncation is not needed. 2124 return LoadNode::Ideal(phase, can_reshape); 2125 } 2126 2127 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2128 Node* mem = in(MemNode::Memory); 2129 Node* value = can_see_stored_value(mem,phase); 2130 if (value != NULL && value->is_Con() && 2131 !value->bottom_type()->higher_equal(_type)) { 2132 // If the input to the store does not fit with the load's result type, 2133 // it must be truncated. We can't delay until Ideal call since 2134 // a singleton Value is needed for split_thru_phi optimization. 2135 int con = value->get_int(); 2136 return TypeInt::make((con << 16) >> 16); 2137 } 2138 return LoadNode::Value(phase); 2139 } 2140 2141 //============================================================================= 2142 //----------------------------LoadKlassNode::make------------------------------ 2143 // Polymorphic factory method: 2144 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2145 // sanity check the alias category against the created node type 2146 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2147 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2148 #ifdef _LP64 2149 if (adr_type->is_ptr_to_narrowklass()) { 2150 assert(UseCompressedClassPointers, "no compressed klasses"); 2151 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2152 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2153 } 2154 #endif 2155 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2156 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2157 } 2158 2159 //------------------------------Value------------------------------------------ 2160 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2161 return klass_value_common(phase); 2162 } 2163 2164 // In most cases, LoadKlassNode does not have the control input set. If the control 2165 // input is set, it must not be removed (by LoadNode::Ideal()). 2166 bool LoadKlassNode::can_remove_control() const { 2167 return false; 2168 } 2169 2170 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2171 // Either input is TOP ==> the result is TOP 2172 const Type *t1 = phase->type( in(MemNode::Memory) ); 2173 if (t1 == Type::TOP) return Type::TOP; 2174 Node *adr = in(MemNode::Address); 2175 const Type *t2 = phase->type( adr ); 2176 if (t2 == Type::TOP) return Type::TOP; 2177 const TypePtr *tp = t2->is_ptr(); 2178 if (TypePtr::above_centerline(tp->ptr()) || 2179 tp->ptr() == TypePtr::Null) return Type::TOP; 2180 2181 // Return a more precise klass, if possible 2182 const TypeInstPtr *tinst = tp->isa_instptr(); 2183 if (tinst != NULL) { 2184 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2185 int offset = tinst->offset(); 2186 if (ik == phase->C->env()->Class_klass() 2187 && (offset == java_lang_Class::klass_offset_in_bytes() || 2188 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2189 // We are loading a special hidden field from a Class mirror object, 2190 // the field which points to the VM's Klass metaobject. 2191 bool is_val_type = false; 2192 ciType* t = tinst->java_mirror_type(&is_val_type); 2193 // java_mirror_type returns non-null for compile-time Class constants. 2194 if (t != NULL) { 2195 // constant oop => constant klass 2196 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2197 if (t->is_void()) { 2198 // We cannot create a void array. Since void is a primitive type return null 2199 // klass. Users of this result need to do a null check on the returned klass. 2200 return TypePtr::NULL_PTR; 2201 } 2202 return TypeKlassPtr::make(ciArrayKlass::make(t, /* never_null */ is_val_type)); 2203 } 2204 if (!t->is_klass()) { 2205 // a primitive Class (e.g., int.class) has NULL for a klass field 2206 return TypePtr::NULL_PTR; 2207 } 2208 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2209 return TypeKlassPtr::make(t->as_klass()); 2210 } 2211 // non-constant mirror, so we can't tell what's going on 2212 } 2213 if( !ik->is_loaded() ) 2214 return _type; // Bail out if not loaded 2215 if (offset == oopDesc::klass_offset_in_bytes()) { 2216 if (tinst->klass_is_exact()) { 2217 return TypeKlassPtr::make(ik); 2218 } 2219 // See if we can become precise: no subklasses and no interface 2220 // (Note: We need to support verified interfaces.) 2221 if (!ik->is_interface() && !ik->has_subklass()) { 2222 //assert(!UseExactTypes, "this code should be useless with exact types"); 2223 // Add a dependence; if any subclass added we need to recompile 2224 if (!ik->is_final()) { 2225 // %%% should use stronger assert_unique_concrete_subtype instead 2226 phase->C->dependencies()->assert_leaf_type(ik); 2227 } 2228 // Return precise klass 2229 return TypeKlassPtr::make(ik); 2230 } 2231 2232 // Return root of possible klass 2233 return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0)); 2234 } 2235 } 2236 2237 // Check for loading klass from an array 2238 const TypeAryPtr *tary = tp->isa_aryptr(); 2239 if( tary != NULL ) { 2240 ciKlass *tary_klass = tary->klass(); 2241 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2242 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2243 if (tary->klass_is_exact()) { 2244 return TypeKlassPtr::make(tary_klass); 2245 } 2246 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2247 // If the klass is an object array, we defer the question to the 2248 // array component klass. 2249 if( ak->is_obj_array_klass() ) { 2250 assert( ak->is_loaded(), "" ); 2251 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2252 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2253 ciInstanceKlass* ik = base_k->as_instance_klass(); 2254 // See if we can become precise: no subklasses and no interface 2255 if (!ik->is_interface() && !ik->has_subklass()) { 2256 //assert(!UseExactTypes, "this code should be useless with exact types"); 2257 // Add a dependence; if any subclass added we need to recompile 2258 if (!ik->is_final()) { 2259 phase->C->dependencies()->assert_leaf_type(ik); 2260 } 2261 // Return precise array klass 2262 return TypeKlassPtr::make(ak); 2263 } 2264 } 2265 return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0)); 2266 } else { // Found a type-array? 2267 //assert(!UseExactTypes, "this code should be useless with exact types"); 2268 assert( ak->is_type_array_klass(), "" ); 2269 return TypeKlassPtr::make(ak); // These are always precise 2270 } 2271 } 2272 } 2273 2274 // Check for loading klass from an array klass 2275 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2276 if (tkls != NULL && !StressReflectiveCode) { 2277 if (!tkls->is_loaded()) { 2278 return _type; // Bail out if not loaded 2279 } 2280 ciKlass* klass = tkls->klass(); 2281 if( klass->is_obj_array_klass() && 2282 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2283 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2284 // // Always returning precise element type is incorrect, 2285 // // e.g., element type could be object and array may contain strings 2286 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2287 2288 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2289 // according to the element type's subclassing. 2290 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0)); 2291 } 2292 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2293 tkls->offset() == in_bytes(Klass::super_offset())) { 2294 ciKlass* sup = klass->as_instance_klass()->super(); 2295 // The field is Klass::_super. Return its (constant) value. 2296 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2297 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2298 } 2299 } 2300 2301 // Bailout case 2302 return LoadNode::Value(phase); 2303 } 2304 2305 //------------------------------Identity--------------------------------------- 2306 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2307 // Also feed through the klass in Allocate(...klass...)._klass. 2308 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2309 return klass_identity_common(phase); 2310 } 2311 2312 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2313 Node* x = LoadNode::Identity(phase); 2314 if (x != this) return x; 2315 2316 // Take apart the address into an oop and and offset. 2317 // Return 'this' if we cannot. 2318 Node* adr = in(MemNode::Address); 2319 intptr_t offset = 0; 2320 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2321 if (base == NULL) return this; 2322 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2323 if (toop == NULL) return this; 2324 2325 // Step over potential GC barrier for OopHandle resolve 2326 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2327 if (bs->is_gc_barrier_node(base)) { 2328 base = bs->step_over_gc_barrier(base); 2329 } 2330 2331 // We can fetch the klass directly through an AllocateNode. 2332 // This works even if the klass is not constant (clone or newArray). 2333 if (offset == oopDesc::klass_offset_in_bytes()) { 2334 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2335 if (allocated_klass != NULL) { 2336 return allocated_klass; 2337 } 2338 } 2339 2340 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2341 // See inline_native_Class_query for occurrences of these patterns. 2342 // Java Example: x.getClass().isAssignableFrom(y) 2343 // 2344 // This improves reflective code, often making the Class 2345 // mirror go completely dead. (Current exception: Class 2346 // mirrors may appear in debug info, but we could clean them out by 2347 // introducing a new debug info operator for Klass.java_mirror). 2348 2349 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2350 && offset == java_lang_Class::klass_offset_in_bytes()) { 2351 if (base->is_Load()) { 2352 Node* base2 = base->in(MemNode::Address); 2353 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2354 Node* adr2 = base2->in(MemNode::Address); 2355 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2356 if (tkls != NULL && !tkls->empty() 2357 && (tkls->klass()->is_instance_klass() || 2358 tkls->klass()->is_array_klass()) 2359 && adr2->is_AddP() 2360 ) { 2361 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2362 if (tkls->offset() == mirror_field) { 2363 return adr2->in(AddPNode::Base); 2364 } 2365 } 2366 } 2367 } 2368 } 2369 2370 return this; 2371 } 2372 2373 2374 //------------------------------Value------------------------------------------ 2375 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2376 const Type *t = klass_value_common(phase); 2377 if (t == Type::TOP) 2378 return t; 2379 2380 return t->make_narrowklass(); 2381 } 2382 2383 //------------------------------Identity--------------------------------------- 2384 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2385 // Also feed through the klass in Allocate(...klass...)._klass. 2386 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2387 Node *x = klass_identity_common(phase); 2388 2389 const Type *t = phase->type( x ); 2390 if( t == Type::TOP ) return x; 2391 if( t->isa_narrowklass()) return x; 2392 assert (!t->isa_narrowoop(), "no narrow oop here"); 2393 2394 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2395 } 2396 2397 //------------------------------Value----------------------------------------- 2398 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2399 // Either input is TOP ==> the result is TOP 2400 const Type *t1 = phase->type( in(MemNode::Memory) ); 2401 if( t1 == Type::TOP ) return Type::TOP; 2402 Node *adr = in(MemNode::Address); 2403 const Type *t2 = phase->type( adr ); 2404 if( t2 == Type::TOP ) return Type::TOP; 2405 const TypePtr *tp = t2->is_ptr(); 2406 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2407 const TypeAryPtr *tap = tp->isa_aryptr(); 2408 if( !tap ) return _type; 2409 return tap->size(); 2410 } 2411 2412 //-------------------------------Ideal--------------------------------------- 2413 // Feed through the length in AllocateArray(...length...)._length. 2414 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2415 Node* p = MemNode::Ideal_common(phase, can_reshape); 2416 if (p) return (p == NodeSentinel) ? NULL : p; 2417 2418 // Take apart the address into an oop and and offset. 2419 // Return 'this' if we cannot. 2420 Node* adr = in(MemNode::Address); 2421 intptr_t offset = 0; 2422 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2423 if (base == NULL) return NULL; 2424 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2425 if (tary == NULL) return NULL; 2426 2427 // We can fetch the length directly through an AllocateArrayNode. 2428 // This works even if the length is not constant (clone or newArray). 2429 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2430 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2431 if (alloc != NULL) { 2432 Node* allocated_length = alloc->Ideal_length(); 2433 Node* len = alloc->make_ideal_length(tary, phase); 2434 if (allocated_length != len) { 2435 // New CastII improves on this. 2436 return len; 2437 } 2438 } 2439 } 2440 2441 return NULL; 2442 } 2443 2444 //------------------------------Identity--------------------------------------- 2445 // Feed through the length in AllocateArray(...length...)._length. 2446 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2447 Node* x = LoadINode::Identity(phase); 2448 if (x != this) return x; 2449 2450 // Take apart the address into an oop and and offset. 2451 // Return 'this' if we cannot. 2452 Node* adr = in(MemNode::Address); 2453 intptr_t offset = 0; 2454 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2455 if (base == NULL) return this; 2456 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2457 if (tary == NULL) return this; 2458 2459 // We can fetch the length directly through an AllocateArrayNode. 2460 // This works even if the length is not constant (clone or newArray). 2461 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2462 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2463 if (alloc != NULL) { 2464 Node* allocated_length = alloc->Ideal_length(); 2465 // Do not allow make_ideal_length to allocate a CastII node. 2466 Node* len = alloc->make_ideal_length(tary, phase, false); 2467 if (allocated_length == len) { 2468 // Return allocated_length only if it would not be improved by a CastII. 2469 return allocated_length; 2470 } 2471 } 2472 } 2473 2474 return this; 2475 2476 } 2477 2478 //============================================================================= 2479 //---------------------------StoreNode::make----------------------------------- 2480 // Polymorphic factory method: 2481 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2482 assert((mo == unordered || mo == release), "unexpected"); 2483 Compile* C = gvn.C; 2484 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2485 ctl != NULL, "raw memory operations should have control edge"); 2486 2487 switch (bt) { 2488 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2489 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2490 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2491 case T_CHAR: 2492 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2493 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2494 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2495 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2496 case T_METADATA: 2497 case T_ADDRESS: 2498 case T_VALUETYPE: 2499 case T_OBJECT: 2500 #ifdef _LP64 2501 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2502 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2503 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2504 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2505 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2506 adr->bottom_type()->isa_rawptr())) { 2507 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2508 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2509 } 2510 #endif 2511 { 2512 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2513 } 2514 default: 2515 ShouldNotReachHere(); 2516 return (StoreNode*)NULL; 2517 } 2518 } 2519 2520 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2521 bool require_atomic = true; 2522 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2523 } 2524 2525 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2526 bool require_atomic = true; 2527 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2528 } 2529 2530 2531 //--------------------------bottom_type---------------------------------------- 2532 const Type *StoreNode::bottom_type() const { 2533 return Type::MEMORY; 2534 } 2535 2536 //------------------------------hash------------------------------------------- 2537 uint StoreNode::hash() const { 2538 // unroll addition of interesting fields 2539 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2540 2541 // Since they are not commoned, do not hash them: 2542 return NO_HASH; 2543 } 2544 2545 //------------------------------Ideal------------------------------------------ 2546 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2547 // When a store immediately follows a relevant allocation/initialization, 2548 // try to capture it into the initialization, or hoist it above. 2549 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2550 Node* p = MemNode::Ideal_common(phase, can_reshape); 2551 if (p) return (p == NodeSentinel) ? NULL : p; 2552 2553 Node* mem = in(MemNode::Memory); 2554 Node* address = in(MemNode::Address); 2555 // Back-to-back stores to same address? Fold em up. Generally 2556 // unsafe if I have intervening uses... Also disallowed for StoreCM 2557 // since they must follow each StoreP operation. Redundant StoreCMs 2558 // are eliminated just before matching in final_graph_reshape. 2559 { 2560 Node* st = mem; 2561 // If Store 'st' has more than one use, we cannot fold 'st' away. 2562 // For example, 'st' might be the final state at a conditional 2563 // return. Or, 'st' might be used by some node which is live at 2564 // the same time 'st' is live, which might be unschedulable. So, 2565 // require exactly ONE user until such time as we clone 'mem' for 2566 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2567 // true). 2568 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2569 // Looking at a dead closed cycle of memory? 2570 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2571 assert(Opcode() == st->Opcode() || 2572 st->Opcode() == Op_StoreVector || 2573 Opcode() == Op_StoreVector || 2574 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2575 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2576 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2577 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) || 2578 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2579 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2580 2581 if (st->in(MemNode::Address)->eqv_uncast(address) && 2582 st->as_Store()->memory_size() <= this->memory_size()) { 2583 Node* use = st->raw_out(0); 2584 phase->igvn_rehash_node_delayed(use); 2585 if (can_reshape) { 2586 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2587 } else { 2588 // It's OK to do this in the parser, since DU info is always accurate, 2589 // and the parser always refers to nodes via SafePointNode maps. 2590 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2591 } 2592 return this; 2593 } 2594 st = st->in(MemNode::Memory); 2595 } 2596 } 2597 2598 2599 // Capture an unaliased, unconditional, simple store into an initializer. 2600 // Or, if it is independent of the allocation, hoist it above the allocation. 2601 if (ReduceFieldZeroing && /*can_reshape &&*/ 2602 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2603 InitializeNode* init = mem->in(0)->as_Initialize(); 2604 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2605 if (offset > 0) { 2606 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2607 // If the InitializeNode captured me, it made a raw copy of me, 2608 // and I need to disappear. 2609 if (moved != NULL) { 2610 // %%% hack to ensure that Ideal returns a new node: 2611 mem = MergeMemNode::make(mem); 2612 return mem; // fold me away 2613 } 2614 } 2615 } 2616 2617 return NULL; // No further progress 2618 } 2619 2620 //------------------------------Value----------------------------------------- 2621 const Type* StoreNode::Value(PhaseGVN* phase) const { 2622 // Either input is TOP ==> the result is TOP 2623 const Type *t1 = phase->type( in(MemNode::Memory) ); 2624 if( t1 == Type::TOP ) return Type::TOP; 2625 const Type *t2 = phase->type( in(MemNode::Address) ); 2626 if( t2 == Type::TOP ) return Type::TOP; 2627 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2628 if( t3 == Type::TOP ) return Type::TOP; 2629 return Type::MEMORY; 2630 } 2631 2632 //------------------------------Identity--------------------------------------- 2633 // Remove redundant stores: 2634 // Store(m, p, Load(m, p)) changes to m. 2635 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2636 Node* StoreNode::Identity(PhaseGVN* phase) { 2637 Node* mem = in(MemNode::Memory); 2638 Node* adr = in(MemNode::Address); 2639 Node* val = in(MemNode::ValueIn); 2640 2641 Node* result = this; 2642 2643 // Load then Store? Then the Store is useless 2644 if (val->is_Load() && 2645 val->in(MemNode::Address)->eqv_uncast(adr) && 2646 val->in(MemNode::Memory )->eqv_uncast(mem) && 2647 val->as_Load()->store_Opcode() == Opcode()) { 2648 result = mem; 2649 } 2650 2651 // Two stores in a row of the same value? 2652 if (result == this && 2653 mem->is_Store() && 2654 mem->in(MemNode::Address)->eqv_uncast(adr) && 2655 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2656 mem->Opcode() == Opcode()) { 2657 result = mem; 2658 } 2659 2660 // Store of zero anywhere into a freshly-allocated object? 2661 // Then the store is useless. 2662 // (It must already have been captured by the InitializeNode.) 2663 if (result == this && ReduceFieldZeroing) { 2664 // a newly allocated object is already all-zeroes everywhere 2665 if (mem->is_Proj() && mem->in(0)->is_Allocate() && 2666 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) { 2667 assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to value array is forbidden"); 2668 result = mem; 2669 } 2670 2671 if (result == this) { 2672 // the store may also apply to zero-bits in an earlier object 2673 Node* prev_mem = find_previous_store(phase); 2674 // Steps (a), (b): Walk past independent stores to find an exact match. 2675 if (prev_mem != NULL) { 2676 Node* prev_val = can_see_stored_value(prev_mem, phase); 2677 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2678 // prev_val and val might differ by a cast; it would be good 2679 // to keep the more informative of the two. 2680 if (phase->type(val)->is_zero_type()) { 2681 result = mem; 2682 } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) { 2683 InitializeNode* init = prev_mem->in(0)->as_Initialize(); 2684 AllocateNode* alloc = init->allocation(); 2685 if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) { 2686 result = mem; 2687 } 2688 } 2689 } 2690 } 2691 } 2692 } 2693 2694 if (result != this && phase->is_IterGVN() != NULL) { 2695 MemBarNode* trailing = trailing_membar(); 2696 if (trailing != NULL) { 2697 #ifdef ASSERT 2698 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2699 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects"); 2700 #endif 2701 PhaseIterGVN* igvn = phase->is_IterGVN(); 2702 trailing->remove(igvn); 2703 } 2704 } 2705 2706 return result; 2707 } 2708 2709 //------------------------------match_edge------------------------------------- 2710 // Do we Match on this edge index or not? Match only memory & value 2711 uint StoreNode::match_edge(uint idx) const { 2712 return idx == MemNode::Address || idx == MemNode::ValueIn; 2713 } 2714 2715 //------------------------------cmp-------------------------------------------- 2716 // Do not common stores up together. They generally have to be split 2717 // back up anyways, so do not bother. 2718 bool StoreNode::cmp( const Node &n ) const { 2719 return (&n == this); // Always fail except on self 2720 } 2721 2722 //------------------------------Ideal_masked_input----------------------------- 2723 // Check for a useless mask before a partial-word store 2724 // (StoreB ... (AndI valIn conIa) ) 2725 // If (conIa & mask == mask) this simplifies to 2726 // (StoreB ... (valIn) ) 2727 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2728 Node *val = in(MemNode::ValueIn); 2729 if( val->Opcode() == Op_AndI ) { 2730 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2731 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2732 set_req(MemNode::ValueIn, val->in(1)); 2733 return this; 2734 } 2735 } 2736 return NULL; 2737 } 2738 2739 2740 //------------------------------Ideal_sign_extended_input---------------------- 2741 // Check for useless sign-extension before a partial-word store 2742 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2743 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2744 // (StoreB ... (valIn) ) 2745 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2746 Node *val = in(MemNode::ValueIn); 2747 if( val->Opcode() == Op_RShiftI ) { 2748 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2749 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2750 Node *shl = val->in(1); 2751 if( shl->Opcode() == Op_LShiftI ) { 2752 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2753 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2754 set_req(MemNode::ValueIn, shl->in(1)); 2755 return this; 2756 } 2757 } 2758 } 2759 } 2760 return NULL; 2761 } 2762 2763 //------------------------------value_never_loaded----------------------------------- 2764 // Determine whether there are any possible loads of the value stored. 2765 // For simplicity, we actually check if there are any loads from the 2766 // address stored to, not just for loads of the value stored by this node. 2767 // 2768 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2769 Node *adr = in(Address); 2770 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2771 if (adr_oop == NULL) 2772 return false; 2773 if (!adr_oop->is_known_instance_field()) 2774 return false; // if not a distinct instance, there may be aliases of the address 2775 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2776 Node *use = adr->fast_out(i); 2777 if (use->is_Load() || use->is_LoadStore()) { 2778 return false; 2779 } 2780 } 2781 return true; 2782 } 2783 2784 MemBarNode* StoreNode::trailing_membar() const { 2785 if (is_release()) { 2786 MemBarNode* trailing_mb = NULL; 2787 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2788 Node* u = fast_out(i); 2789 if (u->is_MemBar()) { 2790 if (u->as_MemBar()->trailing_store()) { 2791 assert(u->Opcode() == Op_MemBarVolatile, ""); 2792 assert(trailing_mb == NULL, "only one"); 2793 trailing_mb = u->as_MemBar(); 2794 #ifdef ASSERT 2795 Node* leading = u->as_MemBar()->leading_membar(); 2796 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2797 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2798 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2799 #endif 2800 } else { 2801 assert(u->as_MemBar()->standalone(), ""); 2802 } 2803 } 2804 } 2805 return trailing_mb; 2806 } 2807 return NULL; 2808 } 2809 2810 2811 //============================================================================= 2812 //------------------------------Ideal------------------------------------------ 2813 // If the store is from an AND mask that leaves the low bits untouched, then 2814 // we can skip the AND operation. If the store is from a sign-extension 2815 // (a left shift, then right shift) we can skip both. 2816 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2817 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2818 if( progress != NULL ) return progress; 2819 2820 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2821 if( progress != NULL ) return progress; 2822 2823 // Finally check the default case 2824 return StoreNode::Ideal(phase, can_reshape); 2825 } 2826 2827 //============================================================================= 2828 //------------------------------Ideal------------------------------------------ 2829 // If the store is from an AND mask that leaves the low bits untouched, then 2830 // we can skip the AND operation 2831 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2832 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2833 if( progress != NULL ) return progress; 2834 2835 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2836 if( progress != NULL ) return progress; 2837 2838 // Finally check the default case 2839 return StoreNode::Ideal(phase, can_reshape); 2840 } 2841 2842 //============================================================================= 2843 //------------------------------Identity--------------------------------------- 2844 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2845 // No need to card mark when storing a null ptr 2846 Node* my_store = in(MemNode::OopStore); 2847 if (my_store->is_Store()) { 2848 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2849 if( t1 == TypePtr::NULL_PTR ) { 2850 return in(MemNode::Memory); 2851 } 2852 } 2853 return this; 2854 } 2855 2856 //============================================================================= 2857 //------------------------------Ideal--------------------------------------- 2858 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2859 Node* progress = StoreNode::Ideal(phase, can_reshape); 2860 if (progress != NULL) return progress; 2861 2862 Node* my_store = in(MemNode::OopStore); 2863 if (my_store->is_MergeMem()) { 2864 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2865 set_req(MemNode::OopStore, mem); 2866 return this; 2867 } 2868 2869 return NULL; 2870 } 2871 2872 //------------------------------Value----------------------------------------- 2873 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2874 // Either input is TOP ==> the result is TOP 2875 const Type *t = phase->type( in(MemNode::Memory) ); 2876 if( t == Type::TOP ) return Type::TOP; 2877 t = phase->type( in(MemNode::Address) ); 2878 if( t == Type::TOP ) return Type::TOP; 2879 t = phase->type( in(MemNode::ValueIn) ); 2880 if( t == Type::TOP ) return Type::TOP; 2881 // If extra input is TOP ==> the result is TOP 2882 t = phase->type( in(MemNode::OopStore) ); 2883 if( t == Type::TOP ) return Type::TOP; 2884 2885 return StoreNode::Value( phase ); 2886 } 2887 2888 2889 //============================================================================= 2890 //----------------------------------SCMemProjNode------------------------------ 2891 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2892 { 2893 return bottom_type(); 2894 } 2895 2896 //============================================================================= 2897 //----------------------------------LoadStoreNode------------------------------ 2898 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2899 : Node(required), 2900 _type(rt), 2901 _adr_type(at) 2902 { 2903 init_req(MemNode::Control, c ); 2904 init_req(MemNode::Memory , mem); 2905 init_req(MemNode::Address, adr); 2906 init_req(MemNode::ValueIn, val); 2907 init_class_id(Class_LoadStore); 2908 } 2909 2910 uint LoadStoreNode::ideal_reg() const { 2911 return _type->ideal_reg(); 2912 } 2913 2914 bool LoadStoreNode::result_not_used() const { 2915 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2916 Node *x = fast_out(i); 2917 if (x->Opcode() == Op_SCMemProj) continue; 2918 return false; 2919 } 2920 return true; 2921 } 2922 2923 MemBarNode* LoadStoreNode::trailing_membar() const { 2924 MemBarNode* trailing = NULL; 2925 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2926 Node* u = fast_out(i); 2927 if (u->is_MemBar()) { 2928 if (u->as_MemBar()->trailing_load_store()) { 2929 assert(u->Opcode() == Op_MemBarAcquire, ""); 2930 assert(trailing == NULL, "only one"); 2931 trailing = u->as_MemBar(); 2932 #ifdef ASSERT 2933 Node* leading = trailing->leading_membar(); 2934 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2935 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 2936 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 2937 #endif 2938 } else { 2939 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 2940 } 2941 } 2942 } 2943 2944 return trailing; 2945 } 2946 2947 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2948 2949 //============================================================================= 2950 //----------------------------------LoadStoreConditionalNode-------------------- 2951 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2952 init_req(ExpectedIn, ex ); 2953 } 2954 2955 //============================================================================= 2956 //-------------------------------adr_type-------------------------------------- 2957 const TypePtr* ClearArrayNode::adr_type() const { 2958 Node *adr = in(3); 2959 if (adr == NULL) return NULL; // node is dead 2960 return MemNode::calculate_adr_type(adr->bottom_type()); 2961 } 2962 2963 //------------------------------match_edge------------------------------------- 2964 // Do we Match on this edge index or not? Do not match memory 2965 uint ClearArrayNode::match_edge(uint idx) const { 2966 return idx > 1; 2967 } 2968 2969 //------------------------------Identity--------------------------------------- 2970 // Clearing a zero length array does nothing 2971 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2972 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2973 } 2974 2975 //------------------------------Idealize--------------------------------------- 2976 // Clearing a short array is faster with stores 2977 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2978 // Already know this is a large node, do not try to ideal it 2979 if (!IdealizeClearArrayNode || _is_large) return NULL; 2980 2981 const int unit = BytesPerLong; 2982 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2983 if (!t) return NULL; 2984 if (!t->is_con()) return NULL; 2985 intptr_t raw_count = t->get_con(); 2986 intptr_t size = raw_count; 2987 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2988 // Clearing nothing uses the Identity call. 2989 // Negative clears are possible on dead ClearArrays 2990 // (see jck test stmt114.stmt11402.val). 2991 if (size <= 0 || size % unit != 0) return NULL; 2992 intptr_t count = size / unit; 2993 // Length too long; communicate this to matchers and assemblers. 2994 // Assemblers are responsible to produce fast hardware clears for it. 2995 if (size > InitArrayShortSize) { 2996 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true); 2997 } 2998 Node *mem = in(1); 2999 if( phase->type(mem)==Type::TOP ) return NULL; 3000 Node *adr = in(3); 3001 const Type* at = phase->type(adr); 3002 if( at==Type::TOP ) return NULL; 3003 const TypePtr* atp = at->isa_ptr(); 3004 // adjust atp to be the correct array element address type 3005 if (atp == NULL) atp = TypePtr::BOTTOM; 3006 else atp = atp->add_offset(Type::OffsetBot); 3007 // Get base for derived pointer purposes 3008 if( adr->Opcode() != Op_AddP ) Unimplemented(); 3009 Node *base = adr->in(1); 3010 3011 Node *val = in(4); 3012 Node *off = phase->MakeConX(BytesPerLong); 3013 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3014 count--; 3015 while( count-- ) { 3016 mem = phase->transform(mem); 3017 adr = phase->transform(new AddPNode(base,adr,off)); 3018 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3019 } 3020 return mem; 3021 } 3022 3023 //----------------------------step_through---------------------------------- 3024 // Return allocation input memory edge if it is different instance 3025 // or itself if it is the one we are looking for. 3026 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 3027 Node* n = *np; 3028 assert(n->is_ClearArray(), "sanity"); 3029 intptr_t offset; 3030 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 3031 // This method is called only before Allocate nodes are expanded 3032 // during macro nodes expansion. Before that ClearArray nodes are 3033 // only generated in PhaseMacroExpand::generate_arraycopy() (before 3034 // Allocate nodes are expanded) which follows allocations. 3035 assert(alloc != NULL, "should have allocation"); 3036 if (alloc->_idx == instance_id) { 3037 // Can not bypass initialization of the instance we are looking for. 3038 return false; 3039 } 3040 // Otherwise skip it. 3041 InitializeNode* init = alloc->initialization(); 3042 if (init != NULL) 3043 *np = init->in(TypeFunc::Memory); 3044 else 3045 *np = alloc->in(TypeFunc::Memory); 3046 return true; 3047 } 3048 3049 //----------------------------clear_memory------------------------------------- 3050 // Generate code to initialize object storage to zero. 3051 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3052 Node* val, 3053 Node* raw_val, 3054 intptr_t start_offset, 3055 Node* end_offset, 3056 PhaseGVN* phase) { 3057 intptr_t offset = start_offset; 3058 3059 int unit = BytesPerLong; 3060 if ((offset % unit) != 0) { 3061 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 3062 adr = phase->transform(adr); 3063 const TypePtr* atp = TypeRawPtr::BOTTOM; 3064 if (val != NULL) { 3065 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3066 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3067 } else { 3068 assert(raw_val == NULL, "val may not be null"); 3069 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3070 } 3071 mem = phase->transform(mem); 3072 offset += BytesPerInt; 3073 } 3074 assert((offset % unit) == 0, ""); 3075 3076 // Initialize the remaining stuff, if any, with a ClearArray. 3077 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase); 3078 } 3079 3080 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3081 Node* raw_val, 3082 Node* start_offset, 3083 Node* end_offset, 3084 PhaseGVN* phase) { 3085 if (start_offset == end_offset) { 3086 // nothing to do 3087 return mem; 3088 } 3089 3090 int unit = BytesPerLong; 3091 Node* zbase = start_offset; 3092 Node* zend = end_offset; 3093 3094 // Scale to the unit required by the CPU: 3095 if (!Matcher::init_array_count_is_in_bytes) { 3096 Node* shift = phase->intcon(exact_log2(unit)); 3097 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3098 zend = phase->transform(new URShiftXNode(zend, shift) ); 3099 } 3100 3101 // Bulk clear double-words 3102 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3103 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3104 if (raw_val == NULL) { 3105 raw_val = phase->MakeConX(0); 3106 } 3107 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false); 3108 return phase->transform(mem); 3109 } 3110 3111 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3112 Node* val, 3113 Node* raw_val, 3114 intptr_t start_offset, 3115 intptr_t end_offset, 3116 PhaseGVN* phase) { 3117 if (start_offset == end_offset) { 3118 // nothing to do 3119 return mem; 3120 } 3121 3122 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3123 intptr_t done_offset = end_offset; 3124 if ((done_offset % BytesPerLong) != 0) { 3125 done_offset -= BytesPerInt; 3126 } 3127 if (done_offset > start_offset) { 3128 mem = clear_memory(ctl, mem, dest, val, raw_val, 3129 start_offset, phase->MakeConX(done_offset), phase); 3130 } 3131 if (done_offset < end_offset) { // emit the final 32-bit store 3132 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3133 adr = phase->transform(adr); 3134 const TypePtr* atp = TypeRawPtr::BOTTOM; 3135 if (val != NULL) { 3136 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3137 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3138 } else { 3139 assert(raw_val == NULL, "val may not be null"); 3140 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3141 } 3142 mem = phase->transform(mem); 3143 done_offset += BytesPerInt; 3144 } 3145 assert(done_offset == end_offset, ""); 3146 return mem; 3147 } 3148 3149 //============================================================================= 3150 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3151 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 3152 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3153 #ifdef ASSERT 3154 , _pair_idx(0) 3155 #endif 3156 { 3157 init_class_id(Class_MemBar); 3158 Node* top = C->top(); 3159 init_req(TypeFunc::I_O,top); 3160 init_req(TypeFunc::FramePtr,top); 3161 init_req(TypeFunc::ReturnAdr,top); 3162 if (precedent != NULL) 3163 init_req(TypeFunc::Parms, precedent); 3164 } 3165 3166 //------------------------------cmp-------------------------------------------- 3167 uint MemBarNode::hash() const { return NO_HASH; } 3168 bool MemBarNode::cmp( const Node &n ) const { 3169 return (&n == this); // Always fail except on self 3170 } 3171 3172 //------------------------------make------------------------------------------- 3173 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3174 switch (opcode) { 3175 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3176 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3177 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3178 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3179 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3180 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3181 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3182 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3183 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3184 case Op_Initialize: return new InitializeNode(C, atp, pn); 3185 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3186 default: ShouldNotReachHere(); return NULL; 3187 } 3188 } 3189 3190 void MemBarNode::remove(PhaseIterGVN *igvn) { 3191 if (outcnt() != 2) { 3192 return; 3193 } 3194 if (trailing_store() || trailing_load_store()) { 3195 MemBarNode* leading = leading_membar(); 3196 if (leading != NULL) { 3197 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3198 leading->remove(igvn); 3199 } 3200 } 3201 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3202 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3203 } 3204 3205 //------------------------------Ideal------------------------------------------ 3206 // Return a node which is more "ideal" than the current node. Strip out 3207 // control copies 3208 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3209 if (remove_dead_region(phase, can_reshape)) return this; 3210 // Don't bother trying to transform a dead node 3211 if (in(0) && in(0)->is_top()) { 3212 return NULL; 3213 } 3214 3215 #if INCLUDE_ZGC 3216 if (UseZGC) { 3217 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) { 3218 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop); 3219 set_req(MemBarNode::Precedent, load_node); 3220 return this; 3221 } 3222 } 3223 #endif 3224 3225 bool progress = false; 3226 // Eliminate volatile MemBars for scalar replaced objects. 3227 if (can_reshape && req() == (Precedent+1)) { 3228 bool eliminate = false; 3229 int opc = Opcode(); 3230 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3231 // Volatile field loads and stores. 3232 Node* my_mem = in(MemBarNode::Precedent); 3233 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3234 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3235 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3236 // replace this Precedent (decodeN) with the Load instead. 3237 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3238 Node* load_node = my_mem->in(1); 3239 set_req(MemBarNode::Precedent, load_node); 3240 phase->is_IterGVN()->_worklist.push(my_mem); 3241 my_mem = load_node; 3242 } else { 3243 assert(my_mem->unique_out() == this, "sanity"); 3244 del_req(Precedent); 3245 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3246 my_mem = NULL; 3247 } 3248 progress = true; 3249 } 3250 if (my_mem != NULL && my_mem->is_Mem()) { 3251 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3252 // Check for scalar replaced object reference. 3253 if( t_oop != NULL && t_oop->is_known_instance_field() && 3254 t_oop->offset() != Type::OffsetBot && 3255 t_oop->offset() != Type::OffsetTop) { 3256 eliminate = true; 3257 } 3258 } 3259 } else if (opc == Op_MemBarRelease) { 3260 // Final field stores. 3261 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3262 if ((alloc != NULL) && alloc->is_Allocate() && 3263 alloc->as_Allocate()->does_not_escape_thread()) { 3264 // The allocated object does not escape. 3265 eliminate = true; 3266 } 3267 } 3268 if (eliminate) { 3269 // Replace MemBar projections by its inputs. 3270 PhaseIterGVN* igvn = phase->is_IterGVN(); 3271 remove(igvn); 3272 // Must return either the original node (now dead) or a new node 3273 // (Do not return a top here, since that would break the uniqueness of top.) 3274 return new ConINode(TypeInt::ZERO); 3275 } 3276 } 3277 return progress ? this : NULL; 3278 } 3279 3280 //------------------------------Value------------------------------------------ 3281 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3282 if( !in(0) ) return Type::TOP; 3283 if( phase->type(in(0)) == Type::TOP ) 3284 return Type::TOP; 3285 return TypeTuple::MEMBAR; 3286 } 3287 3288 //------------------------------match------------------------------------------ 3289 // Construct projections for memory. 3290 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) { 3291 switch (proj->_con) { 3292 case TypeFunc::Control: 3293 case TypeFunc::Memory: 3294 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3295 } 3296 ShouldNotReachHere(); 3297 return NULL; 3298 } 3299 3300 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3301 trailing->_kind = TrailingStore; 3302 leading->_kind = LeadingStore; 3303 #ifdef ASSERT 3304 trailing->_pair_idx = leading->_idx; 3305 leading->_pair_idx = leading->_idx; 3306 #endif 3307 } 3308 3309 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3310 trailing->_kind = TrailingLoadStore; 3311 leading->_kind = LeadingLoadStore; 3312 #ifdef ASSERT 3313 trailing->_pair_idx = leading->_idx; 3314 leading->_pair_idx = leading->_idx; 3315 #endif 3316 } 3317 3318 MemBarNode* MemBarNode::trailing_membar() const { 3319 ResourceMark rm; 3320 Node* trailing = (Node*)this; 3321 VectorSet seen(Thread::current()->resource_area()); 3322 Node_Stack multis(0); 3323 do { 3324 Node* c = trailing; 3325 uint i = 0; 3326 do { 3327 trailing = NULL; 3328 for (; i < c->outcnt(); i++) { 3329 Node* next = c->raw_out(i); 3330 if (next != c && next->is_CFG()) { 3331 if (c->is_MultiBranch()) { 3332 if (multis.node() == c) { 3333 multis.set_index(i+1); 3334 } else { 3335 multis.push(c, i+1); 3336 } 3337 } 3338 trailing = next; 3339 break; 3340 } 3341 } 3342 if (trailing != NULL && !seen.test_set(trailing->_idx)) { 3343 break; 3344 } 3345 while (multis.size() > 0) { 3346 c = multis.node(); 3347 i = multis.index(); 3348 if (i < c->req()) { 3349 break; 3350 } 3351 multis.pop(); 3352 } 3353 } while (multis.size() > 0); 3354 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3355 3356 MemBarNode* mb = trailing->as_MemBar(); 3357 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3358 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3359 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3360 return mb; 3361 } 3362 3363 MemBarNode* MemBarNode::leading_membar() const { 3364 ResourceMark rm; 3365 VectorSet seen(Thread::current()->resource_area()); 3366 Node_Stack regions(0); 3367 Node* leading = in(0); 3368 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3369 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) { 3370 leading = NULL; 3371 while (regions.size() > 0 && leading == NULL) { 3372 Node* r = regions.node(); 3373 uint i = regions.index(); 3374 if (i < r->req()) { 3375 leading = r->in(i); 3376 regions.set_index(i+1); 3377 } else { 3378 regions.pop(); 3379 } 3380 } 3381 if (leading == NULL) { 3382 assert(regions.size() == 0, "all paths should have been tried"); 3383 return NULL; 3384 } 3385 } 3386 if (leading->is_Region()) { 3387 regions.push(leading, 2); 3388 leading = leading->in(1); 3389 } else { 3390 leading = leading->in(0); 3391 } 3392 } 3393 #ifdef ASSERT 3394 Unique_Node_List wq; 3395 wq.push((Node*)this); 3396 uint found = 0; 3397 for (uint i = 0; i < wq.size(); i++) { 3398 Node* n = wq.at(i); 3399 if (n->is_Region()) { 3400 for (uint j = 1; j < n->req(); j++) { 3401 Node* in = n->in(j); 3402 if (in != NULL && !in->is_top()) { 3403 wq.push(in); 3404 } 3405 } 3406 } else { 3407 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3408 assert(n == leading, "consistency check failed"); 3409 found++; 3410 } else { 3411 Node* in = n->in(0); 3412 if (in != NULL && !in->is_top()) { 3413 wq.push(in); 3414 } 3415 } 3416 } 3417 } 3418 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed"); 3419 #endif 3420 if (leading == NULL) { 3421 return NULL; 3422 } 3423 MemBarNode* mb = leading->as_MemBar(); 3424 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3425 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3426 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3427 return mb; 3428 } 3429 3430 //===========================InitializeNode==================================== 3431 // SUMMARY: 3432 // This node acts as a memory barrier on raw memory, after some raw stores. 3433 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3434 // The Initialize can 'capture' suitably constrained stores as raw inits. 3435 // It can coalesce related raw stores into larger units (called 'tiles'). 3436 // It can avoid zeroing new storage for memory units which have raw inits. 3437 // At macro-expansion, it is marked 'complete', and does not optimize further. 3438 // 3439 // EXAMPLE: 3440 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3441 // ctl = incoming control; mem* = incoming memory 3442 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3443 // First allocate uninitialized memory and fill in the header: 3444 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3445 // ctl := alloc.Control; mem* := alloc.Memory* 3446 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3447 // Then initialize to zero the non-header parts of the raw memory block: 3448 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3449 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3450 // After the initialize node executes, the object is ready for service: 3451 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3452 // Suppose its body is immediately initialized as {1,2}: 3453 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3454 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3455 // mem.SLICE(#short[*]) := store2 3456 // 3457 // DETAILS: 3458 // An InitializeNode collects and isolates object initialization after 3459 // an AllocateNode and before the next possible safepoint. As a 3460 // memory barrier (MemBarNode), it keeps critical stores from drifting 3461 // down past any safepoint or any publication of the allocation. 3462 // Before this barrier, a newly-allocated object may have uninitialized bits. 3463 // After this barrier, it may be treated as a real oop, and GC is allowed. 3464 // 3465 // The semantics of the InitializeNode include an implicit zeroing of 3466 // the new object from object header to the end of the object. 3467 // (The object header and end are determined by the AllocateNode.) 3468 // 3469 // Certain stores may be added as direct inputs to the InitializeNode. 3470 // These stores must update raw memory, and they must be to addresses 3471 // derived from the raw address produced by AllocateNode, and with 3472 // a constant offset. They must be ordered by increasing offset. 3473 // The first one is at in(RawStores), the last at in(req()-1). 3474 // Unlike most memory operations, they are not linked in a chain, 3475 // but are displayed in parallel as users of the rawmem output of 3476 // the allocation. 3477 // 3478 // (See comments in InitializeNode::capture_store, which continue 3479 // the example given above.) 3480 // 3481 // When the associated Allocate is macro-expanded, the InitializeNode 3482 // may be rewritten to optimize collected stores. A ClearArrayNode 3483 // may also be created at that point to represent any required zeroing. 3484 // The InitializeNode is then marked 'complete', prohibiting further 3485 // capturing of nearby memory operations. 3486 // 3487 // During macro-expansion, all captured initializations which store 3488 // constant values of 32 bits or smaller are coalesced (if advantageous) 3489 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3490 // initialized in fewer memory operations. Memory words which are 3491 // covered by neither tiles nor non-constant stores are pre-zeroed 3492 // by explicit stores of zero. (The code shape happens to do all 3493 // zeroing first, then all other stores, with both sequences occurring 3494 // in order of ascending offsets.) 3495 // 3496 // Alternatively, code may be inserted between an AllocateNode and its 3497 // InitializeNode, to perform arbitrary initialization of the new object. 3498 // E.g., the object copying intrinsics insert complex data transfers here. 3499 // The initialization must then be marked as 'complete' disable the 3500 // built-in zeroing semantics and the collection of initializing stores. 3501 // 3502 // While an InitializeNode is incomplete, reads from the memory state 3503 // produced by it are optimizable if they match the control edge and 3504 // new oop address associated with the allocation/initialization. 3505 // They return a stored value (if the offset matches) or else zero. 3506 // A write to the memory state, if it matches control and address, 3507 // and if it is to a constant offset, may be 'captured' by the 3508 // InitializeNode. It is cloned as a raw memory operation and rewired 3509 // inside the initialization, to the raw oop produced by the allocation. 3510 // Operations on addresses which are provably distinct (e.g., to 3511 // other AllocateNodes) are allowed to bypass the initialization. 3512 // 3513 // The effect of all this is to consolidate object initialization 3514 // (both arrays and non-arrays, both piecewise and bulk) into a 3515 // single location, where it can be optimized as a unit. 3516 // 3517 // Only stores with an offset less than TrackedInitializationLimit words 3518 // will be considered for capture by an InitializeNode. This puts a 3519 // reasonable limit on the complexity of optimized initializations. 3520 3521 //---------------------------InitializeNode------------------------------------ 3522 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3523 : MemBarNode(C, adr_type, rawoop), 3524 _is_complete(Incomplete), _does_not_escape(false) 3525 { 3526 init_class_id(Class_Initialize); 3527 3528 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3529 assert(in(RawAddress) == rawoop, "proper init"); 3530 // Note: allocation() can be NULL, for secondary initialization barriers 3531 } 3532 3533 // Since this node is not matched, it will be processed by the 3534 // register allocator. Declare that there are no constraints 3535 // on the allocation of the RawAddress edge. 3536 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3537 // This edge should be set to top, by the set_complete. But be conservative. 3538 if (idx == InitializeNode::RawAddress) 3539 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3540 return RegMask::Empty; 3541 } 3542 3543 Node* InitializeNode::memory(uint alias_idx) { 3544 Node* mem = in(Memory); 3545 if (mem->is_MergeMem()) { 3546 return mem->as_MergeMem()->memory_at(alias_idx); 3547 } else { 3548 // incoming raw memory is not split 3549 return mem; 3550 } 3551 } 3552 3553 bool InitializeNode::is_non_zero() { 3554 if (is_complete()) return false; 3555 remove_extra_zeroes(); 3556 return (req() > RawStores); 3557 } 3558 3559 void InitializeNode::set_complete(PhaseGVN* phase) { 3560 assert(!is_complete(), "caller responsibility"); 3561 _is_complete = Complete; 3562 3563 // After this node is complete, it contains a bunch of 3564 // raw-memory initializations. There is no need for 3565 // it to have anything to do with non-raw memory effects. 3566 // Therefore, tell all non-raw users to re-optimize themselves, 3567 // after skipping the memory effects of this initialization. 3568 PhaseIterGVN* igvn = phase->is_IterGVN(); 3569 if (igvn) igvn->add_users_to_worklist(this); 3570 } 3571 3572 // convenience function 3573 // return false if the init contains any stores already 3574 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3575 InitializeNode* init = initialization(); 3576 if (init == NULL || init->is_complete()) { 3577 return false; 3578 } 3579 init->remove_extra_zeroes(); 3580 // for now, if this allocation has already collected any inits, bail: 3581 if (init->is_non_zero()) return false; 3582 init->set_complete(phase); 3583 return true; 3584 } 3585 3586 void InitializeNode::remove_extra_zeroes() { 3587 if (req() == RawStores) return; 3588 Node* zmem = zero_memory(); 3589 uint fill = RawStores; 3590 for (uint i = fill; i < req(); i++) { 3591 Node* n = in(i); 3592 if (n->is_top() || n == zmem) continue; // skip 3593 if (fill < i) set_req(fill, n); // compact 3594 ++fill; 3595 } 3596 // delete any empty spaces created: 3597 while (fill < req()) { 3598 del_req(fill); 3599 } 3600 } 3601 3602 // Helper for remembering which stores go with which offsets. 3603 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3604 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3605 intptr_t offset = -1; 3606 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3607 phase, offset); 3608 if (base == NULL) return -1; // something is dead, 3609 if (offset < 0) return -1; // dead, dead 3610 return offset; 3611 } 3612 3613 // Helper for proving that an initialization expression is 3614 // "simple enough" to be folded into an object initialization. 3615 // Attempts to prove that a store's initial value 'n' can be captured 3616 // within the initialization without creating a vicious cycle, such as: 3617 // { Foo p = new Foo(); p.next = p; } 3618 // True for constants and parameters and small combinations thereof. 3619 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3620 if (n == NULL) return true; // (can this really happen?) 3621 if (n->is_Proj()) n = n->in(0); 3622 if (n == this) return false; // found a cycle 3623 if (n->is_Con()) return true; 3624 if (n->is_Start()) return true; // params, etc., are OK 3625 if (n->is_Root()) return true; // even better 3626 3627 Node* ctl = n->in(0); 3628 if (ctl != NULL && !ctl->is_top()) { 3629 if (ctl->is_Proj()) ctl = ctl->in(0); 3630 if (ctl == this) return false; 3631 3632 // If we already know that the enclosing memory op is pinned right after 3633 // the init, then any control flow that the store has picked up 3634 // must have preceded the init, or else be equal to the init. 3635 // Even after loop optimizations (which might change control edges) 3636 // a store is never pinned *before* the availability of its inputs. 3637 if (!MemNode::all_controls_dominate(n, this)) 3638 return false; // failed to prove a good control 3639 } 3640 3641 // Check data edges for possible dependencies on 'this'. 3642 if ((count += 1) > 20) return false; // complexity limit 3643 for (uint i = 1; i < n->req(); i++) { 3644 Node* m = n->in(i); 3645 if (m == NULL || m == n || m->is_top()) continue; 3646 uint first_i = n->find_edge(m); 3647 if (i != first_i) continue; // process duplicate edge just once 3648 if (!detect_init_independence(m, count)) { 3649 return false; 3650 } 3651 } 3652 3653 return true; 3654 } 3655 3656 // Here are all the checks a Store must pass before it can be moved into 3657 // an initialization. Returns zero if a check fails. 3658 // On success, returns the (constant) offset to which the store applies, 3659 // within the initialized memory. 3660 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3661 const int FAIL = 0; 3662 if (st->is_unaligned_access()) { 3663 return FAIL; 3664 } 3665 if (st->req() != MemNode::ValueIn + 1) 3666 return FAIL; // an inscrutable StoreNode (card mark?) 3667 Node* ctl = st->in(MemNode::Control); 3668 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3669 return FAIL; // must be unconditional after the initialization 3670 Node* mem = st->in(MemNode::Memory); 3671 if (!(mem->is_Proj() && mem->in(0) == this)) 3672 return FAIL; // must not be preceded by other stores 3673 Node* adr = st->in(MemNode::Address); 3674 intptr_t offset; 3675 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3676 if (alloc == NULL) 3677 return FAIL; // inscrutable address 3678 if (alloc != allocation()) 3679 return FAIL; // wrong allocation! (store needs to float up) 3680 Node* val = st->in(MemNode::ValueIn); 3681 int complexity_count = 0; 3682 if (!detect_init_independence(val, complexity_count)) 3683 return FAIL; // stored value must be 'simple enough' 3684 3685 // The Store can be captured only if nothing after the allocation 3686 // and before the Store is using the memory location that the store 3687 // overwrites. 3688 bool failed = false; 3689 // If is_complete_with_arraycopy() is true the shape of the graph is 3690 // well defined and is safe so no need for extra checks. 3691 if (!is_complete_with_arraycopy()) { 3692 // We are going to look at each use of the memory state following 3693 // the allocation to make sure nothing reads the memory that the 3694 // Store writes. 3695 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3696 int alias_idx = phase->C->get_alias_index(t_adr); 3697 ResourceMark rm; 3698 Unique_Node_List mems; 3699 mems.push(mem); 3700 Node* unique_merge = NULL; 3701 for (uint next = 0; next < mems.size(); ++next) { 3702 Node *m = mems.at(next); 3703 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3704 Node *n = m->fast_out(j); 3705 if (n->outcnt() == 0) { 3706 continue; 3707 } 3708 if (n == st) { 3709 continue; 3710 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3711 // If the control of this use is different from the control 3712 // of the Store which is right after the InitializeNode then 3713 // this node cannot be between the InitializeNode and the 3714 // Store. 3715 continue; 3716 } else if (n->is_MergeMem()) { 3717 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3718 // We can hit a MergeMemNode (that will likely go away 3719 // later) that is a direct use of the memory state 3720 // following the InitializeNode on the same slice as the 3721 // store node that we'd like to capture. We need to check 3722 // the uses of the MergeMemNode. 3723 mems.push(n); 3724 } 3725 } else if (n->is_Mem()) { 3726 Node* other_adr = n->in(MemNode::Address); 3727 if (other_adr == adr) { 3728 failed = true; 3729 break; 3730 } else { 3731 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3732 if (other_t_adr != NULL) { 3733 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3734 if (other_alias_idx == alias_idx) { 3735 // A load from the same memory slice as the store right 3736 // after the InitializeNode. We check the control of the 3737 // object/array that is loaded from. If it's the same as 3738 // the store control then we cannot capture the store. 3739 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3740 Node* base = other_adr; 3741 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3742 base = base->in(AddPNode::Base); 3743 if (base != NULL) { 3744 base = base->uncast(); 3745 if (base->is_Proj() && base->in(0) == alloc) { 3746 failed = true; 3747 break; 3748 } 3749 } 3750 } 3751 } 3752 } 3753 } else { 3754 failed = true; 3755 break; 3756 } 3757 } 3758 } 3759 } 3760 if (failed) { 3761 if (!can_reshape) { 3762 // We decided we couldn't capture the store during parsing. We 3763 // should try again during the next IGVN once the graph is 3764 // cleaner. 3765 phase->C->record_for_igvn(st); 3766 } 3767 return FAIL; 3768 } 3769 3770 return offset; // success 3771 } 3772 3773 // Find the captured store in(i) which corresponds to the range 3774 // [start..start+size) in the initialized object. 3775 // If there is one, return its index i. If there isn't, return the 3776 // negative of the index where it should be inserted. 3777 // Return 0 if the queried range overlaps an initialization boundary 3778 // or if dead code is encountered. 3779 // If size_in_bytes is zero, do not bother with overlap checks. 3780 int InitializeNode::captured_store_insertion_point(intptr_t start, 3781 int size_in_bytes, 3782 PhaseTransform* phase) { 3783 const int FAIL = 0, MAX_STORE = BytesPerLong; 3784 3785 if (is_complete()) 3786 return FAIL; // arraycopy got here first; punt 3787 3788 assert(allocation() != NULL, "must be present"); 3789 3790 // no negatives, no header fields: 3791 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3792 3793 // after a certain size, we bail out on tracking all the stores: 3794 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3795 if (start >= ti_limit) return FAIL; 3796 3797 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3798 if (i >= limit) return -(int)i; // not found; here is where to put it 3799 3800 Node* st = in(i); 3801 intptr_t st_off = get_store_offset(st, phase); 3802 if (st_off < 0) { 3803 if (st != zero_memory()) { 3804 return FAIL; // bail out if there is dead garbage 3805 } 3806 } else if (st_off > start) { 3807 // ...we are done, since stores are ordered 3808 if (st_off < start + size_in_bytes) { 3809 return FAIL; // the next store overlaps 3810 } 3811 return -(int)i; // not found; here is where to put it 3812 } else if (st_off < start) { 3813 if (size_in_bytes != 0 && 3814 start < st_off + MAX_STORE && 3815 start < st_off + st->as_Store()->memory_size()) { 3816 return FAIL; // the previous store overlaps 3817 } 3818 } else { 3819 if (size_in_bytes != 0 && 3820 st->as_Store()->memory_size() != size_in_bytes) { 3821 return FAIL; // mismatched store size 3822 } 3823 return i; 3824 } 3825 3826 ++i; 3827 } 3828 } 3829 3830 // Look for a captured store which initializes at the offset 'start' 3831 // with the given size. If there is no such store, and no other 3832 // initialization interferes, then return zero_memory (the memory 3833 // projection of the AllocateNode). 3834 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3835 PhaseTransform* phase) { 3836 assert(stores_are_sane(phase), ""); 3837 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3838 if (i == 0) { 3839 return NULL; // something is dead 3840 } else if (i < 0) { 3841 return zero_memory(); // just primordial zero bits here 3842 } else { 3843 Node* st = in(i); // here is the store at this position 3844 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3845 return st; 3846 } 3847 } 3848 3849 // Create, as a raw pointer, an address within my new object at 'offset'. 3850 Node* InitializeNode::make_raw_address(intptr_t offset, 3851 PhaseTransform* phase) { 3852 Node* addr = in(RawAddress); 3853 if (offset != 0) { 3854 Compile* C = phase->C; 3855 addr = phase->transform( new AddPNode(C->top(), addr, 3856 phase->MakeConX(offset)) ); 3857 } 3858 return addr; 3859 } 3860 3861 // Clone the given store, converting it into a raw store 3862 // initializing a field or element of my new object. 3863 // Caller is responsible for retiring the original store, 3864 // with subsume_node or the like. 3865 // 3866 // From the example above InitializeNode::InitializeNode, 3867 // here are the old stores to be captured: 3868 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3869 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3870 // 3871 // Here is the changed code; note the extra edges on init: 3872 // alloc = (Allocate ...) 3873 // rawoop = alloc.RawAddress 3874 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3875 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3876 // init = (Initialize alloc.Control alloc.Memory rawoop 3877 // rawstore1 rawstore2) 3878 // 3879 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3880 PhaseTransform* phase, bool can_reshape) { 3881 assert(stores_are_sane(phase), ""); 3882 3883 if (start < 0) return NULL; 3884 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3885 3886 Compile* C = phase->C; 3887 int size_in_bytes = st->memory_size(); 3888 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3889 if (i == 0) return NULL; // bail out 3890 Node* prev_mem = NULL; // raw memory for the captured store 3891 if (i > 0) { 3892 prev_mem = in(i); // there is a pre-existing store under this one 3893 set_req(i, C->top()); // temporarily disconnect it 3894 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3895 } else { 3896 i = -i; // no pre-existing store 3897 prev_mem = zero_memory(); // a slice of the newly allocated object 3898 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3899 set_req(--i, C->top()); // reuse this edge; it has been folded away 3900 else 3901 ins_req(i, C->top()); // build a new edge 3902 } 3903 Node* new_st = st->clone(); 3904 new_st->set_req(MemNode::Control, in(Control)); 3905 new_st->set_req(MemNode::Memory, prev_mem); 3906 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3907 new_st = phase->transform(new_st); 3908 3909 // At this point, new_st might have swallowed a pre-existing store 3910 // at the same offset, or perhaps new_st might have disappeared, 3911 // if it redundantly stored the same value (or zero to fresh memory). 3912 3913 // In any case, wire it in: 3914 phase->igvn_rehash_node_delayed(this); 3915 set_req(i, new_st); 3916 3917 // The caller may now kill the old guy. 3918 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3919 assert(check_st == new_st || check_st == NULL, "must be findable"); 3920 assert(!is_complete(), ""); 3921 return new_st; 3922 } 3923 3924 static bool store_constant(jlong* tiles, int num_tiles, 3925 intptr_t st_off, int st_size, 3926 jlong con) { 3927 if ((st_off & (st_size-1)) != 0) 3928 return false; // strange store offset (assume size==2**N) 3929 address addr = (address)tiles + st_off; 3930 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3931 switch (st_size) { 3932 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3933 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3934 case sizeof(jint): *(jint*) addr = (jint) con; break; 3935 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3936 default: return false; // strange store size (detect size!=2**N here) 3937 } 3938 return true; // return success to caller 3939 } 3940 3941 // Coalesce subword constants into int constants and possibly 3942 // into long constants. The goal, if the CPU permits, 3943 // is to initialize the object with a small number of 64-bit tiles. 3944 // Also, convert floating-point constants to bit patterns. 3945 // Non-constants are not relevant to this pass. 3946 // 3947 // In terms of the running example on InitializeNode::InitializeNode 3948 // and InitializeNode::capture_store, here is the transformation 3949 // of rawstore1 and rawstore2 into rawstore12: 3950 // alloc = (Allocate ...) 3951 // rawoop = alloc.RawAddress 3952 // tile12 = 0x00010002 3953 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3954 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3955 // 3956 void 3957 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3958 Node* size_in_bytes, 3959 PhaseGVN* phase) { 3960 Compile* C = phase->C; 3961 3962 assert(stores_are_sane(phase), ""); 3963 // Note: After this pass, they are not completely sane, 3964 // since there may be some overlaps. 3965 3966 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3967 3968 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3969 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3970 size_limit = MIN2(size_limit, ti_limit); 3971 size_limit = align_up(size_limit, BytesPerLong); 3972 int num_tiles = size_limit / BytesPerLong; 3973 3974 // allocate space for the tile map: 3975 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3976 jlong tiles_buf[small_len]; 3977 Node* nodes_buf[small_len]; 3978 jlong inits_buf[small_len]; 3979 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3980 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3981 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3982 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3983 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3984 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3985 // tiles: exact bitwise model of all primitive constants 3986 // nodes: last constant-storing node subsumed into the tiles model 3987 // inits: which bytes (in each tile) are touched by any initializations 3988 3989 //// Pass A: Fill in the tile model with any relevant stores. 3990 3991 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3992 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3993 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3994 Node* zmem = zero_memory(); // initially zero memory state 3995 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3996 Node* st = in(i); 3997 intptr_t st_off = get_store_offset(st, phase); 3998 3999 // Figure out the store's offset and constant value: 4000 if (st_off < header_size) continue; //skip (ignore header) 4001 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 4002 int st_size = st->as_Store()->memory_size(); 4003 if (st_off + st_size > size_limit) break; 4004 4005 // Record which bytes are touched, whether by constant or not. 4006 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 4007 continue; // skip (strange store size) 4008 4009 const Type* val = phase->type(st->in(MemNode::ValueIn)); 4010 if (!val->singleton()) continue; //skip (non-con store) 4011 BasicType type = val->basic_type(); 4012 4013 jlong con = 0; 4014 switch (type) { 4015 case T_INT: con = val->is_int()->get_con(); break; 4016 case T_LONG: con = val->is_long()->get_con(); break; 4017 case T_FLOAT: con = jint_cast(val->getf()); break; 4018 case T_DOUBLE: con = jlong_cast(val->getd()); break; 4019 default: continue; //skip (odd store type) 4020 } 4021 4022 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 4023 st->Opcode() == Op_StoreL) { 4024 continue; // This StoreL is already optimal. 4025 } 4026 4027 // Store down the constant. 4028 store_constant(tiles, num_tiles, st_off, st_size, con); 4029 4030 intptr_t j = st_off >> LogBytesPerLong; 4031 4032 if (type == T_INT && st_size == BytesPerInt 4033 && (st_off & BytesPerInt) == BytesPerInt) { 4034 jlong lcon = tiles[j]; 4035 if (!Matcher::isSimpleConstant64(lcon) && 4036 st->Opcode() == Op_StoreI) { 4037 // This StoreI is already optimal by itself. 4038 jint* intcon = (jint*) &tiles[j]; 4039 intcon[1] = 0; // undo the store_constant() 4040 4041 // If the previous store is also optimal by itself, back up and 4042 // undo the action of the previous loop iteration... if we can. 4043 // But if we can't, just let the previous half take care of itself. 4044 st = nodes[j]; 4045 st_off -= BytesPerInt; 4046 con = intcon[0]; 4047 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 4048 assert(st_off >= header_size, "still ignoring header"); 4049 assert(get_store_offset(st, phase) == st_off, "must be"); 4050 assert(in(i-1) == zmem, "must be"); 4051 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 4052 assert(con == tcon->is_int()->get_con(), "must be"); 4053 // Undo the effects of the previous loop trip, which swallowed st: 4054 intcon[0] = 0; // undo store_constant() 4055 set_req(i-1, st); // undo set_req(i, zmem) 4056 nodes[j] = NULL; // undo nodes[j] = st 4057 --old_subword; // undo ++old_subword 4058 } 4059 continue; // This StoreI is already optimal. 4060 } 4061 } 4062 4063 // This store is not needed. 4064 set_req(i, zmem); 4065 nodes[j] = st; // record for the moment 4066 if (st_size < BytesPerLong) // something has changed 4067 ++old_subword; // includes int/float, but who's counting... 4068 else ++old_long; 4069 } 4070 4071 if ((old_subword + old_long) == 0) 4072 return; // nothing more to do 4073 4074 //// Pass B: Convert any non-zero tiles into optimal constant stores. 4075 // Be sure to insert them before overlapping non-constant stores. 4076 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 4077 for (int j = 0; j < num_tiles; j++) { 4078 jlong con = tiles[j]; 4079 jlong init = inits[j]; 4080 if (con == 0) continue; 4081 jint con0, con1; // split the constant, address-wise 4082 jint init0, init1; // split the init map, address-wise 4083 { union { jlong con; jint intcon[2]; } u; 4084 u.con = con; 4085 con0 = u.intcon[0]; 4086 con1 = u.intcon[1]; 4087 u.con = init; 4088 init0 = u.intcon[0]; 4089 init1 = u.intcon[1]; 4090 } 4091 4092 Node* old = nodes[j]; 4093 assert(old != NULL, "need the prior store"); 4094 intptr_t offset = (j * BytesPerLong); 4095 4096 bool split = !Matcher::isSimpleConstant64(con); 4097 4098 if (offset < header_size) { 4099 assert(offset + BytesPerInt >= header_size, "second int counts"); 4100 assert(*(jint*)&tiles[j] == 0, "junk in header"); 4101 split = true; // only the second word counts 4102 // Example: int a[] = { 42 ... } 4103 } else if (con0 == 0 && init0 == -1) { 4104 split = true; // first word is covered by full inits 4105 // Example: int a[] = { ... foo(), 42 ... } 4106 } else if (con1 == 0 && init1 == -1) { 4107 split = true; // second word is covered by full inits 4108 // Example: int a[] = { ... 42, foo() ... } 4109 } 4110 4111 // Here's a case where init0 is neither 0 nor -1: 4112 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4113 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4114 // In this case the tile is not split; it is (jlong)42. 4115 // The big tile is stored down, and then the foo() value is inserted. 4116 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4117 4118 Node* ctl = old->in(MemNode::Control); 4119 Node* adr = make_raw_address(offset, phase); 4120 const TypePtr* atp = TypeRawPtr::BOTTOM; 4121 4122 // One or two coalesced stores to plop down. 4123 Node* st[2]; 4124 intptr_t off[2]; 4125 int nst = 0; 4126 if (!split) { 4127 ++new_long; 4128 off[nst] = offset; 4129 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4130 phase->longcon(con), T_LONG, MemNode::unordered); 4131 } else { 4132 // Omit either if it is a zero. 4133 if (con0 != 0) { 4134 ++new_int; 4135 off[nst] = offset; 4136 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4137 phase->intcon(con0), T_INT, MemNode::unordered); 4138 } 4139 if (con1 != 0) { 4140 ++new_int; 4141 offset += BytesPerInt; 4142 adr = make_raw_address(offset, phase); 4143 off[nst] = offset; 4144 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4145 phase->intcon(con1), T_INT, MemNode::unordered); 4146 } 4147 } 4148 4149 // Insert second store first, then the first before the second. 4150 // Insert each one just before any overlapping non-constant stores. 4151 while (nst > 0) { 4152 Node* st1 = st[--nst]; 4153 C->copy_node_notes_to(st1, old); 4154 st1 = phase->transform(st1); 4155 offset = off[nst]; 4156 assert(offset >= header_size, "do not smash header"); 4157 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4158 guarantee(ins_idx != 0, "must re-insert constant store"); 4159 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4160 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4161 set_req(--ins_idx, st1); 4162 else 4163 ins_req(ins_idx, st1); 4164 } 4165 } 4166 4167 if (PrintCompilation && WizardMode) 4168 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4169 old_subword, old_long, new_int, new_long); 4170 if (C->log() != NULL) 4171 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4172 old_subword, old_long, new_int, new_long); 4173 4174 // Clean up any remaining occurrences of zmem: 4175 remove_extra_zeroes(); 4176 } 4177 4178 // Explore forward from in(start) to find the first fully initialized 4179 // word, and return its offset. Skip groups of subword stores which 4180 // together initialize full words. If in(start) is itself part of a 4181 // fully initialized word, return the offset of in(start). If there 4182 // are no following full-word stores, or if something is fishy, return 4183 // a negative value. 4184 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4185 int int_map = 0; 4186 intptr_t int_map_off = 0; 4187 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4188 4189 for (uint i = start, limit = req(); i < limit; i++) { 4190 Node* st = in(i); 4191 4192 intptr_t st_off = get_store_offset(st, phase); 4193 if (st_off < 0) break; // return conservative answer 4194 4195 int st_size = st->as_Store()->memory_size(); 4196 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4197 return st_off; // we found a complete word init 4198 } 4199 4200 // update the map: 4201 4202 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4203 if (this_int_off != int_map_off) { 4204 // reset the map: 4205 int_map = 0; 4206 int_map_off = this_int_off; 4207 } 4208 4209 int subword_off = st_off - this_int_off; 4210 int_map |= right_n_bits(st_size) << subword_off; 4211 if ((int_map & FULL_MAP) == FULL_MAP) { 4212 return this_int_off; // we found a complete word init 4213 } 4214 4215 // Did this store hit or cross the word boundary? 4216 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4217 if (next_int_off == this_int_off + BytesPerInt) { 4218 // We passed the current int, without fully initializing it. 4219 int_map_off = next_int_off; 4220 int_map >>= BytesPerInt; 4221 } else if (next_int_off > this_int_off + BytesPerInt) { 4222 // We passed the current and next int. 4223 return this_int_off + BytesPerInt; 4224 } 4225 } 4226 4227 return -1; 4228 } 4229 4230 4231 // Called when the associated AllocateNode is expanded into CFG. 4232 // At this point, we may perform additional optimizations. 4233 // Linearize the stores by ascending offset, to make memory 4234 // activity as coherent as possible. 4235 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4236 intptr_t header_size, 4237 Node* size_in_bytes, 4238 PhaseGVN* phase) { 4239 assert(!is_complete(), "not already complete"); 4240 assert(stores_are_sane(phase), ""); 4241 assert(allocation() != NULL, "must be present"); 4242 4243 remove_extra_zeroes(); 4244 4245 if (ReduceFieldZeroing || ReduceBulkZeroing) 4246 // reduce instruction count for common initialization patterns 4247 coalesce_subword_stores(header_size, size_in_bytes, phase); 4248 4249 Node* zmem = zero_memory(); // initially zero memory state 4250 Node* inits = zmem; // accumulating a linearized chain of inits 4251 #ifdef ASSERT 4252 intptr_t first_offset = allocation()->minimum_header_size(); 4253 intptr_t last_init_off = first_offset; // previous init offset 4254 intptr_t last_init_end = first_offset; // previous init offset+size 4255 intptr_t last_tile_end = first_offset; // previous tile offset+size 4256 #endif 4257 intptr_t zeroes_done = header_size; 4258 4259 bool do_zeroing = true; // we might give up if inits are very sparse 4260 int big_init_gaps = 0; // how many large gaps have we seen? 4261 4262 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4263 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4264 4265 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4266 Node* st = in(i); 4267 intptr_t st_off = get_store_offset(st, phase); 4268 if (st_off < 0) 4269 break; // unknown junk in the inits 4270 if (st->in(MemNode::Memory) != zmem) 4271 break; // complicated store chains somehow in list 4272 4273 int st_size = st->as_Store()->memory_size(); 4274 intptr_t next_init_off = st_off + st_size; 4275 4276 if (do_zeroing && zeroes_done < next_init_off) { 4277 // See if this store needs a zero before it or under it. 4278 intptr_t zeroes_needed = st_off; 4279 4280 if (st_size < BytesPerInt) { 4281 // Look for subword stores which only partially initialize words. 4282 // If we find some, we must lay down some word-level zeroes first, 4283 // underneath the subword stores. 4284 // 4285 // Examples: 4286 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4287 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4288 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4289 // 4290 // Note: coalesce_subword_stores may have already done this, 4291 // if it was prompted by constant non-zero subword initializers. 4292 // But this case can still arise with non-constant stores. 4293 4294 intptr_t next_full_store = find_next_fullword_store(i, phase); 4295 4296 // In the examples above: 4297 // in(i) p q r s x y z 4298 // st_off 12 13 14 15 12 13 14 4299 // st_size 1 1 1 1 1 1 1 4300 // next_full_s. 12 16 16 16 16 16 16 4301 // z's_done 12 16 16 16 12 16 12 4302 // z's_needed 12 16 16 16 16 16 16 4303 // zsize 0 0 0 0 4 0 4 4304 if (next_full_store < 0) { 4305 // Conservative tack: Zero to end of current word. 4306 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4307 } else { 4308 // Zero to beginning of next fully initialized word. 4309 // Or, don't zero at all, if we are already in that word. 4310 assert(next_full_store >= zeroes_needed, "must go forward"); 4311 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4312 zeroes_needed = next_full_store; 4313 } 4314 } 4315 4316 if (zeroes_needed > zeroes_done) { 4317 intptr_t zsize = zeroes_needed - zeroes_done; 4318 // Do some incremental zeroing on rawmem, in parallel with inits. 4319 zeroes_done = align_down(zeroes_done, BytesPerInt); 4320 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4321 allocation()->in(AllocateNode::DefaultValue), 4322 allocation()->in(AllocateNode::RawDefaultValue), 4323 zeroes_done, zeroes_needed, 4324 phase); 4325 zeroes_done = zeroes_needed; 4326 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4327 do_zeroing = false; // leave the hole, next time 4328 } 4329 } 4330 4331 // Collect the store and move on: 4332 st->set_req(MemNode::Memory, inits); 4333 inits = st; // put it on the linearized chain 4334 set_req(i, zmem); // unhook from previous position 4335 4336 if (zeroes_done == st_off) 4337 zeroes_done = next_init_off; 4338 4339 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4340 4341 #ifdef ASSERT 4342 // Various order invariants. Weaker than stores_are_sane because 4343 // a large constant tile can be filled in by smaller non-constant stores. 4344 assert(st_off >= last_init_off, "inits do not reverse"); 4345 last_init_off = st_off; 4346 const Type* val = NULL; 4347 if (st_size >= BytesPerInt && 4348 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4349 (int)val->basic_type() < (int)T_OBJECT) { 4350 assert(st_off >= last_tile_end, "tiles do not overlap"); 4351 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4352 last_tile_end = MAX2(last_tile_end, next_init_off); 4353 } else { 4354 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4355 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4356 assert(st_off >= last_init_end, "inits do not overlap"); 4357 last_init_end = next_init_off; // it's a non-tile 4358 } 4359 #endif //ASSERT 4360 } 4361 4362 remove_extra_zeroes(); // clear out all the zmems left over 4363 add_req(inits); 4364 4365 if (!(UseTLAB && ZeroTLAB)) { 4366 // If anything remains to be zeroed, zero it all now. 4367 zeroes_done = align_down(zeroes_done, BytesPerInt); 4368 // if it is the last unused 4 bytes of an instance, forget about it 4369 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4370 if (zeroes_done + BytesPerLong >= size_limit) { 4371 AllocateNode* alloc = allocation(); 4372 assert(alloc != NULL, "must be present"); 4373 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 4374 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4375 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4376 if (zeroes_done == k->layout_helper()) 4377 zeroes_done = size_limit; 4378 } 4379 } 4380 if (zeroes_done < size_limit) { 4381 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4382 allocation()->in(AllocateNode::DefaultValue), 4383 allocation()->in(AllocateNode::RawDefaultValue), 4384 zeroes_done, size_in_bytes, phase); 4385 } 4386 } 4387 4388 set_complete(phase); 4389 return rawmem; 4390 } 4391 4392 4393 #ifdef ASSERT 4394 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4395 if (is_complete()) 4396 return true; // stores could be anything at this point 4397 assert(allocation() != NULL, "must be present"); 4398 intptr_t last_off = allocation()->minimum_header_size(); 4399 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4400 Node* st = in(i); 4401 intptr_t st_off = get_store_offset(st, phase); 4402 if (st_off < 0) continue; // ignore dead garbage 4403 if (last_off > st_off) { 4404 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4405 this->dump(2); 4406 assert(false, "ascending store offsets"); 4407 return false; 4408 } 4409 last_off = st_off + st->as_Store()->memory_size(); 4410 } 4411 return true; 4412 } 4413 #endif //ASSERT 4414 4415 4416 4417 4418 //============================MergeMemNode===================================== 4419 // 4420 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4421 // contributing store or call operations. Each contributor provides the memory 4422 // state for a particular "alias type" (see Compile::alias_type). For example, 4423 // if a MergeMem has an input X for alias category #6, then any memory reference 4424 // to alias category #6 may use X as its memory state input, as an exact equivalent 4425 // to using the MergeMem as a whole. 4426 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4427 // 4428 // (Here, the <N> notation gives the index of the relevant adr_type.) 4429 // 4430 // In one special case (and more cases in the future), alias categories overlap. 4431 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4432 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4433 // it is exactly equivalent to that state W: 4434 // MergeMem(<Bot>: W) <==> W 4435 // 4436 // Usually, the merge has more than one input. In that case, where inputs 4437 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4438 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4439 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4440 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4441 // 4442 // A merge can take a "wide" memory state as one of its narrow inputs. 4443 // This simply means that the merge observes out only the relevant parts of 4444 // the wide input. That is, wide memory states arriving at narrow merge inputs 4445 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4446 // 4447 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4448 // and that memory slices "leak through": 4449 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4450 // 4451 // But, in such a cascade, repeated memory slices can "block the leak": 4452 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4453 // 4454 // In the last example, Y is not part of the combined memory state of the 4455 // outermost MergeMem. The system must, of course, prevent unschedulable 4456 // memory states from arising, so you can be sure that the state Y is somehow 4457 // a precursor to state Y'. 4458 // 4459 // 4460 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4461 // of each MergeMemNode array are exactly the numerical alias indexes, including 4462 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4463 // Compile::alias_type (and kin) produce and manage these indexes. 4464 // 4465 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4466 // (Note that this provides quick access to the top node inside MergeMem methods, 4467 // without the need to reach out via TLS to Compile::current.) 4468 // 4469 // As a consequence of what was just described, a MergeMem that represents a full 4470 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4471 // containing all alias categories. 4472 // 4473 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4474 // 4475 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4476 // a memory state for the alias type <N>, or else the top node, meaning that 4477 // there is no particular input for that alias type. Note that the length of 4478 // a MergeMem is variable, and may be extended at any time to accommodate new 4479 // memory states at larger alias indexes. When merges grow, they are of course 4480 // filled with "top" in the unused in() positions. 4481 // 4482 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4483 // (Top was chosen because it works smoothly with passes like GCM.) 4484 // 4485 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4486 // the type of random VM bits like TLS references.) Since it is always the 4487 // first non-Bot memory slice, some low-level loops use it to initialize an 4488 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4489 // 4490 // 4491 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4492 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4493 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4494 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4495 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4496 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4497 // 4498 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4499 // really that different from the other memory inputs. An abbreviation called 4500 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4501 // 4502 // 4503 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4504 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4505 // that "emerges though" the base memory will be marked as excluding the alias types 4506 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4507 // 4508 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4509 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4510 // 4511 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4512 // (It is currently unimplemented.) As you can see, the resulting merge is 4513 // actually a disjoint union of memory states, rather than an overlay. 4514 // 4515 4516 //------------------------------MergeMemNode----------------------------------- 4517 Node* MergeMemNode::make_empty_memory() { 4518 Node* empty_memory = (Node*) Compile::current()->top(); 4519 assert(empty_memory->is_top(), "correct sentinel identity"); 4520 return empty_memory; 4521 } 4522 4523 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4524 init_class_id(Class_MergeMem); 4525 // all inputs are nullified in Node::Node(int) 4526 // set_input(0, NULL); // no control input 4527 4528 // Initialize the edges uniformly to top, for starters. 4529 Node* empty_mem = make_empty_memory(); 4530 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4531 init_req(i,empty_mem); 4532 } 4533 assert(empty_memory() == empty_mem, ""); 4534 4535 if( new_base != NULL && new_base->is_MergeMem() ) { 4536 MergeMemNode* mdef = new_base->as_MergeMem(); 4537 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4538 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4539 mms.set_memory(mms.memory2()); 4540 } 4541 assert(base_memory() == mdef->base_memory(), ""); 4542 } else { 4543 set_base_memory(new_base); 4544 } 4545 } 4546 4547 // Make a new, untransformed MergeMem with the same base as 'mem'. 4548 // If mem is itself a MergeMem, populate the result with the same edges. 4549 MergeMemNode* MergeMemNode::make(Node* mem) { 4550 return new MergeMemNode(mem); 4551 } 4552 4553 //------------------------------cmp-------------------------------------------- 4554 uint MergeMemNode::hash() const { return NO_HASH; } 4555 bool MergeMemNode::cmp( const Node &n ) const { 4556 return (&n == this); // Always fail except on self 4557 } 4558 4559 //------------------------------Identity--------------------------------------- 4560 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4561 // Identity if this merge point does not record any interesting memory 4562 // disambiguations. 4563 Node* base_mem = base_memory(); 4564 Node* empty_mem = empty_memory(); 4565 if (base_mem != empty_mem) { // Memory path is not dead? 4566 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4567 Node* mem = in(i); 4568 if (mem != empty_mem && mem != base_mem) { 4569 return this; // Many memory splits; no change 4570 } 4571 } 4572 } 4573 return base_mem; // No memory splits; ID on the one true input 4574 } 4575 4576 //------------------------------Ideal------------------------------------------ 4577 // This method is invoked recursively on chains of MergeMem nodes 4578 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4579 // Remove chain'd MergeMems 4580 // 4581 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4582 // relative to the "in(Bot)". Since we are patching both at the same time, 4583 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4584 // but rewrite each "in(i)" relative to the new "in(Bot)". 4585 Node *progress = NULL; 4586 4587 4588 Node* old_base = base_memory(); 4589 Node* empty_mem = empty_memory(); 4590 if (old_base == empty_mem) 4591 return NULL; // Dead memory path. 4592 4593 MergeMemNode* old_mbase; 4594 if (old_base != NULL && old_base->is_MergeMem()) 4595 old_mbase = old_base->as_MergeMem(); 4596 else 4597 old_mbase = NULL; 4598 Node* new_base = old_base; 4599 4600 // simplify stacked MergeMems in base memory 4601 if (old_mbase) new_base = old_mbase->base_memory(); 4602 4603 // the base memory might contribute new slices beyond my req() 4604 if (old_mbase) grow_to_match(old_mbase); 4605 4606 // Look carefully at the base node if it is a phi. 4607 PhiNode* phi_base; 4608 if (new_base != NULL && new_base->is_Phi()) 4609 phi_base = new_base->as_Phi(); 4610 else 4611 phi_base = NULL; 4612 4613 Node* phi_reg = NULL; 4614 uint phi_len = (uint)-1; 4615 if (phi_base != NULL && !phi_base->is_copy()) { 4616 // do not examine phi if degraded to a copy 4617 phi_reg = phi_base->region(); 4618 phi_len = phi_base->req(); 4619 // see if the phi is unfinished 4620 for (uint i = 1; i < phi_len; i++) { 4621 if (phi_base->in(i) == NULL) { 4622 // incomplete phi; do not look at it yet! 4623 phi_reg = NULL; 4624 phi_len = (uint)-1; 4625 break; 4626 } 4627 } 4628 } 4629 4630 // Note: We do not call verify_sparse on entry, because inputs 4631 // can normalize to the base_memory via subsume_node or similar 4632 // mechanisms. This method repairs that damage. 4633 4634 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4635 4636 // Look at each slice. 4637 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4638 Node* old_in = in(i); 4639 // calculate the old memory value 4640 Node* old_mem = old_in; 4641 if (old_mem == empty_mem) old_mem = old_base; 4642 assert(old_mem == memory_at(i), ""); 4643 4644 // maybe update (reslice) the old memory value 4645 4646 // simplify stacked MergeMems 4647 Node* new_mem = old_mem; 4648 MergeMemNode* old_mmem; 4649 if (old_mem != NULL && old_mem->is_MergeMem()) 4650 old_mmem = old_mem->as_MergeMem(); 4651 else 4652 old_mmem = NULL; 4653 if (old_mmem == this) { 4654 // This can happen if loops break up and safepoints disappear. 4655 // A merge of BotPtr (default) with a RawPtr memory derived from a 4656 // safepoint can be rewritten to a merge of the same BotPtr with 4657 // the BotPtr phi coming into the loop. If that phi disappears 4658 // also, we can end up with a self-loop of the mergemem. 4659 // In general, if loops degenerate and memory effects disappear, 4660 // a mergemem can be left looking at itself. This simply means 4661 // that the mergemem's default should be used, since there is 4662 // no longer any apparent effect on this slice. 4663 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4664 // from start. Update the input to TOP. 4665 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4666 } 4667 else if (old_mmem != NULL) { 4668 new_mem = old_mmem->memory_at(i); 4669 } 4670 // else preceding memory was not a MergeMem 4671 4672 // replace equivalent phis (unfortunately, they do not GVN together) 4673 if (new_mem != NULL && new_mem != new_base && 4674 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4675 if (new_mem->is_Phi()) { 4676 PhiNode* phi_mem = new_mem->as_Phi(); 4677 for (uint i = 1; i < phi_len; i++) { 4678 if (phi_base->in(i) != phi_mem->in(i)) { 4679 phi_mem = NULL; 4680 break; 4681 } 4682 } 4683 if (phi_mem != NULL) { 4684 // equivalent phi nodes; revert to the def 4685 new_mem = new_base; 4686 } 4687 } 4688 } 4689 4690 // maybe store down a new value 4691 Node* new_in = new_mem; 4692 if (new_in == new_base) new_in = empty_mem; 4693 4694 if (new_in != old_in) { 4695 // Warning: Do not combine this "if" with the previous "if" 4696 // A memory slice might have be be rewritten even if it is semantically 4697 // unchanged, if the base_memory value has changed. 4698 set_req(i, new_in); 4699 progress = this; // Report progress 4700 } 4701 } 4702 4703 if (new_base != old_base) { 4704 set_req(Compile::AliasIdxBot, new_base); 4705 // Don't use set_base_memory(new_base), because we need to update du. 4706 assert(base_memory() == new_base, ""); 4707 progress = this; 4708 } 4709 4710 if( base_memory() == this ) { 4711 // a self cycle indicates this memory path is dead 4712 set_req(Compile::AliasIdxBot, empty_mem); 4713 } 4714 4715 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4716 // Recursion must occur after the self cycle check above 4717 if( base_memory()->is_MergeMem() ) { 4718 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4719 Node *m = phase->transform(new_mbase); // Rollup any cycles 4720 if( m != NULL && 4721 (m->is_top() || 4722 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4723 // propagate rollup of dead cycle to self 4724 set_req(Compile::AliasIdxBot, empty_mem); 4725 } 4726 } 4727 4728 if( base_memory() == empty_mem ) { 4729 progress = this; 4730 // Cut inputs during Parse phase only. 4731 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4732 if( !can_reshape ) { 4733 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4734 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4735 } 4736 } 4737 } 4738 4739 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4740 // Check if PhiNode::Ideal's "Split phis through memory merges" 4741 // transform should be attempted. Look for this->phi->this cycle. 4742 uint merge_width = req(); 4743 if (merge_width > Compile::AliasIdxRaw) { 4744 PhiNode* phi = base_memory()->as_Phi(); 4745 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4746 if (phi->in(i) == this) { 4747 phase->is_IterGVN()->_worklist.push(phi); 4748 break; 4749 } 4750 } 4751 } 4752 } 4753 4754 assert(progress || verify_sparse(), "please, no dups of base"); 4755 return progress; 4756 } 4757 4758 //-------------------------set_base_memory------------------------------------- 4759 void MergeMemNode::set_base_memory(Node *new_base) { 4760 Node* empty_mem = empty_memory(); 4761 set_req(Compile::AliasIdxBot, new_base); 4762 assert(memory_at(req()) == new_base, "must set default memory"); 4763 // Clear out other occurrences of new_base: 4764 if (new_base != empty_mem) { 4765 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4766 if (in(i) == new_base) set_req(i, empty_mem); 4767 } 4768 } 4769 } 4770 4771 //------------------------------out_RegMask------------------------------------ 4772 const RegMask &MergeMemNode::out_RegMask() const { 4773 return RegMask::Empty; 4774 } 4775 4776 //------------------------------dump_spec-------------------------------------- 4777 #ifndef PRODUCT 4778 void MergeMemNode::dump_spec(outputStream *st) const { 4779 st->print(" {"); 4780 Node* base_mem = base_memory(); 4781 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4782 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4783 if (mem == base_mem) { st->print(" -"); continue; } 4784 st->print( " N%d:", mem->_idx ); 4785 Compile::current()->get_adr_type(i)->dump_on(st); 4786 } 4787 st->print(" }"); 4788 } 4789 #endif // !PRODUCT 4790 4791 4792 #ifdef ASSERT 4793 static bool might_be_same(Node* a, Node* b) { 4794 if (a == b) return true; 4795 if (!(a->is_Phi() || b->is_Phi())) return false; 4796 // phis shift around during optimization 4797 return true; // pretty stupid... 4798 } 4799 4800 // verify a narrow slice (either incoming or outgoing) 4801 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4802 if (!VerifyAliases) return; // don't bother to verify unless requested 4803 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4804 if (Node::in_dump()) return; // muzzle asserts when printing 4805 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4806 assert(n != NULL, ""); 4807 // Elide intervening MergeMem's 4808 while (n->is_MergeMem()) { 4809 n = n->as_MergeMem()->memory_at(alias_idx); 4810 } 4811 Compile* C = Compile::current(); 4812 const TypePtr* n_adr_type = n->adr_type(); 4813 if (n == m->empty_memory()) { 4814 // Implicit copy of base_memory() 4815 } else if (n_adr_type != TypePtr::BOTTOM) { 4816 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4817 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4818 } else { 4819 // A few places like make_runtime_call "know" that VM calls are narrow, 4820 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4821 bool expected_wide_mem = false; 4822 if (n == m->base_memory()) { 4823 expected_wide_mem = true; 4824 } else if (alias_idx == Compile::AliasIdxRaw || 4825 n == m->memory_at(Compile::AliasIdxRaw)) { 4826 expected_wide_mem = true; 4827 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4828 // memory can "leak through" calls on channels that 4829 // are write-once. Allow this also. 4830 expected_wide_mem = true; 4831 } 4832 assert(expected_wide_mem, "expected narrow slice replacement"); 4833 } 4834 } 4835 #else // !ASSERT 4836 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4837 #endif 4838 4839 4840 //-----------------------------memory_at--------------------------------------- 4841 Node* MergeMemNode::memory_at(uint alias_idx) const { 4842 assert(alias_idx >= Compile::AliasIdxRaw || 4843 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4844 "must avoid base_memory and AliasIdxTop"); 4845 4846 // Otherwise, it is a narrow slice. 4847 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4848 Compile *C = Compile::current(); 4849 if (is_empty_memory(n)) { 4850 // the array is sparse; empty slots are the "top" node 4851 n = base_memory(); 4852 assert(Node::in_dump() 4853 || n == NULL || n->bottom_type() == Type::TOP 4854 || n->adr_type() == NULL // address is TOP 4855 || n->adr_type() == TypePtr::BOTTOM 4856 || n->adr_type() == TypeRawPtr::BOTTOM 4857 || Compile::current()->AliasLevel() == 0, 4858 "must be a wide memory"); 4859 // AliasLevel == 0 if we are organizing the memory states manually. 4860 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4861 } else { 4862 // make sure the stored slice is sane 4863 #ifdef ASSERT 4864 if (VMError::is_error_reported() || Node::in_dump()) { 4865 } else if (might_be_same(n, base_memory())) { 4866 // Give it a pass: It is a mostly harmless repetition of the base. 4867 // This can arise normally from node subsumption during optimization. 4868 } else { 4869 verify_memory_slice(this, alias_idx, n); 4870 } 4871 #endif 4872 } 4873 return n; 4874 } 4875 4876 //---------------------------set_memory_at------------------------------------- 4877 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4878 verify_memory_slice(this, alias_idx, n); 4879 Node* empty_mem = empty_memory(); 4880 if (n == base_memory()) n = empty_mem; // collapse default 4881 uint need_req = alias_idx+1; 4882 if (req() < need_req) { 4883 if (n == empty_mem) return; // already the default, so do not grow me 4884 // grow the sparse array 4885 do { 4886 add_req(empty_mem); 4887 } while (req() < need_req); 4888 } 4889 set_req( alias_idx, n ); 4890 } 4891 4892 4893 4894 //--------------------------iteration_setup------------------------------------ 4895 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4896 if (other != NULL) { 4897 grow_to_match(other); 4898 // invariant: the finite support of mm2 is within mm->req() 4899 #ifdef ASSERT 4900 for (uint i = req(); i < other->req(); i++) { 4901 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4902 } 4903 #endif 4904 } 4905 // Replace spurious copies of base_memory by top. 4906 Node* base_mem = base_memory(); 4907 if (base_mem != NULL && !base_mem->is_top()) { 4908 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4909 if (in(i) == base_mem) 4910 set_req(i, empty_memory()); 4911 } 4912 } 4913 } 4914 4915 //---------------------------grow_to_match------------------------------------- 4916 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4917 Node* empty_mem = empty_memory(); 4918 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4919 // look for the finite support of the other memory 4920 for (uint i = other->req(); --i >= req(); ) { 4921 if (other->in(i) != empty_mem) { 4922 uint new_len = i+1; 4923 while (req() < new_len) add_req(empty_mem); 4924 break; 4925 } 4926 } 4927 } 4928 4929 //---------------------------verify_sparse------------------------------------- 4930 #ifndef PRODUCT 4931 bool MergeMemNode::verify_sparse() const { 4932 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4933 Node* base_mem = base_memory(); 4934 // The following can happen in degenerate cases, since empty==top. 4935 if (is_empty_memory(base_mem)) return true; 4936 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4937 assert(in(i) != NULL, "sane slice"); 4938 if (in(i) == base_mem) return false; // should have been the sentinel value! 4939 } 4940 return true; 4941 } 4942 4943 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4944 Node* n; 4945 n = mm->in(idx); 4946 if (mem == n) return true; // might be empty_memory() 4947 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4948 if (mem == n) return true; 4949 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4950 if (mem == n) return true; 4951 if (n == NULL) break; 4952 } 4953 return false; 4954 } 4955 #endif // !PRODUCT