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->eqv_uncast(st_base, /*keep_deps=*/true)) 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 // Casted version may carry a dependency and it is respected. 1068 // Thus, we are able to replace L by V. 1069 } 1070 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1071 if (store_Opcode() != st->Opcode()) 1072 return NULL; 1073 return st->in(MemNode::ValueIn); 1074 } 1075 1076 // A load from a freshly-created object always returns zero. 1077 // (This can happen after LoadNode::Ideal resets the load's memory input 1078 // to find_captured_store, which returned InitializeNode::zero_memory.) 1079 if (st->is_Proj() && st->in(0)->is_Allocate() && 1080 (st->in(0) == ld_alloc) && 1081 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1082 // return a zero value for the load's basic type 1083 // (This is one of the few places where a generic PhaseTransform 1084 // can create new nodes. Think of it as lazily manifesting 1085 // virtually pre-existing constants.) 1086 assert(memory_type() != T_VALUETYPE, "should not be used for value types"); 1087 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue); 1088 if (default_value != NULL) { 1089 return default_value; 1090 } 1091 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 1092 return phase->zerocon(memory_type()); 1093 } 1094 1095 // A load from an initialization barrier can match a captured store. 1096 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1097 InitializeNode* init = st->in(0)->as_Initialize(); 1098 AllocateNode* alloc = init->allocation(); 1099 if ((alloc != NULL) && (alloc == ld_alloc)) { 1100 // examine a captured store value 1101 st = init->find_captured_store(ld_off, memory_size(), phase); 1102 if (st != NULL) { 1103 continue; // take one more trip around 1104 } 1105 } 1106 } 1107 1108 // Load boxed value from result of valueOf() call is input parameter. 1109 if (this->is_Load() && ld_adr->is_AddP() && 1110 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1111 intptr_t ignore = 0; 1112 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1113 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1114 base = bs->step_over_gc_barrier(base); 1115 if (base != NULL && base->is_Proj() && 1116 base->as_Proj()->_con == TypeFunc::Parms && 1117 base->in(0)->is_CallStaticJava() && 1118 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1119 return base->in(0)->in(TypeFunc::Parms); 1120 } 1121 } 1122 1123 break; 1124 } 1125 1126 return NULL; 1127 } 1128 1129 //----------------------is_instance_field_load_with_local_phi------------------ 1130 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1131 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1132 in(Address)->is_AddP() ) { 1133 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1134 // Only instances and boxed values. 1135 if( t_oop != NULL && 1136 (t_oop->is_ptr_to_boxed_value() || 1137 t_oop->is_known_instance_field()) && 1138 t_oop->offset() != Type::OffsetBot && 1139 t_oop->offset() != Type::OffsetTop) { 1140 return true; 1141 } 1142 } 1143 return false; 1144 } 1145 1146 //------------------------------Identity--------------------------------------- 1147 // Loads are identity if previous store is to same address 1148 Node* LoadNode::Identity(PhaseGVN* phase) { 1149 // Loading from a ValueTypePtr? The ValueTypePtr has the values of 1150 // all fields as input. Look for the field with matching offset. 1151 Node* addr = in(Address); 1152 intptr_t offset; 1153 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset); 1154 if (base != NULL && base->is_ValueTypePtr() && offset > oopDesc::klass_offset_in_bytes()) { 1155 Node* value = base->as_ValueTypePtr()->field_value_by_offset((int)offset, true); 1156 if (value->is_ValueType()) { 1157 // Non-flattened value type field 1158 ValueTypeNode* vt = value->as_ValueType(); 1159 if (vt->is_allocated(phase)) { 1160 value = vt->get_oop(); 1161 } else { 1162 // Not yet allocated, bail out 1163 value = NULL; 1164 } 1165 } 1166 if (value != NULL) { 1167 if (Opcode() == Op_LoadN) { 1168 // Encode oop value if we are loading a narrow oop 1169 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded"); 1170 value = phase->transform(new EncodePNode(value, bottom_type())); 1171 } 1172 return value; 1173 } 1174 } 1175 1176 // If the previous store-maker is the right kind of Store, and the store is 1177 // to the same address, then we are equal to the value stored. 1178 Node* mem = in(Memory); 1179 Node* value = can_see_stored_value(mem, phase); 1180 if( value ) { 1181 // byte, short & char stores truncate naturally. 1182 // A load has to load the truncated value which requires 1183 // some sort of masking operation and that requires an 1184 // Ideal call instead of an Identity call. 1185 if (memory_size() < BytesPerInt) { 1186 // If the input to the store does not fit with the load's result type, 1187 // it must be truncated via an Ideal call. 1188 if (!phase->type(value)->higher_equal(phase->type(this))) 1189 return this; 1190 } 1191 // (This works even when value is a Con, but LoadNode::Value 1192 // usually runs first, producing the singleton type of the Con.) 1193 return value; 1194 } 1195 1196 // Search for an existing data phi which was generated before for the same 1197 // instance's field to avoid infinite generation of phis in a loop. 1198 Node *region = mem->in(0); 1199 if (is_instance_field_load_with_local_phi(region)) { 1200 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1201 int this_index = phase->C->get_alias_index(addr_t); 1202 int this_offset = addr_t->offset(); 1203 int this_iid = addr_t->instance_id(); 1204 if (!addr_t->is_known_instance() && 1205 addr_t->is_ptr_to_boxed_value()) { 1206 // Use _idx of address base (could be Phi node) for boxed values. 1207 intptr_t ignore = 0; 1208 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1209 if (base == NULL) { 1210 return this; 1211 } 1212 this_iid = base->_idx; 1213 } 1214 const Type* this_type = bottom_type(); 1215 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1216 Node* phi = region->fast_out(i); 1217 if (phi->is_Phi() && phi != mem && 1218 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1219 return phi; 1220 } 1221 } 1222 } 1223 1224 return this; 1225 } 1226 1227 // Construct an equivalent unsigned load. 1228 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1229 BasicType bt = T_ILLEGAL; 1230 const Type* rt = NULL; 1231 switch (Opcode()) { 1232 case Op_LoadUB: return this; 1233 case Op_LoadUS: return this; 1234 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1235 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1236 default: 1237 assert(false, "no unsigned variant: %s", Name()); 1238 return NULL; 1239 } 1240 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1241 raw_adr_type(), rt, bt, _mo, _control_dependency, 1242 is_unaligned_access(), is_mismatched_access()); 1243 } 1244 1245 // Construct an equivalent signed load. 1246 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1247 BasicType bt = T_ILLEGAL; 1248 const Type* rt = NULL; 1249 switch (Opcode()) { 1250 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1251 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1252 case Op_LoadB: // fall through 1253 case Op_LoadS: // fall through 1254 case Op_LoadI: // fall through 1255 case Op_LoadL: return this; 1256 default: 1257 assert(false, "no signed variant: %s", Name()); 1258 return NULL; 1259 } 1260 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1261 raw_adr_type(), rt, bt, _mo, _control_dependency, 1262 is_unaligned_access(), is_mismatched_access()); 1263 } 1264 1265 // We're loading from an object which has autobox behaviour. 1266 // If this object is result of a valueOf call we'll have a phi 1267 // merging a newly allocated object and a load from the cache. 1268 // We want to replace this load with the original incoming 1269 // argument to the valueOf call. 1270 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1271 assert(phase->C->eliminate_boxing(), "sanity"); 1272 intptr_t ignore = 0; 1273 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1274 if ((base == NULL) || base->is_Phi()) { 1275 // Push the loads from the phi that comes from valueOf up 1276 // through it to allow elimination of the loads and the recovery 1277 // of the original value. It is done in split_through_phi(). 1278 return NULL; 1279 } else if (base->is_Load() || 1280 (base->is_DecodeN() && base->in(1)->is_Load())) { 1281 // Eliminate the load of boxed value for integer types from the cache 1282 // array by deriving the value from the index into the array. 1283 // Capture the offset of the load and then reverse the computation. 1284 1285 // Get LoadN node which loads a boxing object from 'cache' array. 1286 if (base->is_DecodeN()) { 1287 base = base->in(1); 1288 } 1289 if (!base->in(Address)->is_AddP()) { 1290 return NULL; // Complex address 1291 } 1292 AddPNode* address = base->in(Address)->as_AddP(); 1293 Node* cache_base = address->in(AddPNode::Base); 1294 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1295 // Get ConP node which is static 'cache' field. 1296 cache_base = cache_base->in(1); 1297 } 1298 if ((cache_base != NULL) && cache_base->is_Con()) { 1299 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1300 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1301 Node* elements[4]; 1302 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1303 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1304 if (count > 0 && elements[0]->is_Con() && 1305 (count == 1 || 1306 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1307 elements[1]->in(2) == phase->intcon(shift)))) { 1308 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1309 // Fetch the box object cache[0] at the base of the array and get its value 1310 ciInstance* box = array->obj_at(0)->as_instance(); 1311 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1312 assert(ik->is_box_klass(), "sanity"); 1313 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1314 if (ik->nof_nonstatic_fields() == 1) { 1315 // This should be true nonstatic_field_at requires calling 1316 // nof_nonstatic_fields so check it anyway 1317 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1318 BasicType bt = c.basic_type(); 1319 // Only integer types have boxing cache. 1320 assert(bt == T_BOOLEAN || bt == T_CHAR || 1321 bt == T_BYTE || bt == T_SHORT || 1322 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1323 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1324 if (cache_low != (int)cache_low) { 1325 return NULL; // should not happen since cache is array indexed by value 1326 } 1327 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1328 if (offset != (int)offset) { 1329 return NULL; // should not happen since cache is array indexed by value 1330 } 1331 // Add up all the offsets making of the address of the load 1332 Node* result = elements[0]; 1333 for (int i = 1; i < count; i++) { 1334 result = phase->transform(new AddXNode(result, elements[i])); 1335 } 1336 // Remove the constant offset from the address and then 1337 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1338 // remove the scaling of the offset to recover the original index. 1339 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1340 // Peel the shift off directly but wrap it in a dummy node 1341 // since Ideal can't return existing nodes 1342 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1343 } else if (result->is_Add() && result->in(2)->is_Con() && 1344 result->in(1)->Opcode() == Op_LShiftX && 1345 result->in(1)->in(2) == phase->intcon(shift)) { 1346 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1347 // but for boxing cache access we know that X<<Z will not overflow 1348 // (there is range check) so we do this optimizatrion by hand here. 1349 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1350 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1351 } else { 1352 result = new RShiftXNode(result, phase->intcon(shift)); 1353 } 1354 #ifdef _LP64 1355 if (bt != T_LONG) { 1356 result = new ConvL2INode(phase->transform(result)); 1357 } 1358 #else 1359 if (bt == T_LONG) { 1360 result = new ConvI2LNode(phase->transform(result)); 1361 } 1362 #endif 1363 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1364 // Need to preserve unboxing load type if it is unsigned. 1365 switch(this->Opcode()) { 1366 case Op_LoadUB: 1367 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1368 break; 1369 case Op_LoadUS: 1370 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1371 break; 1372 } 1373 return result; 1374 } 1375 } 1376 } 1377 } 1378 } 1379 return NULL; 1380 } 1381 1382 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1383 Node* region = phi->in(0); 1384 if (region == NULL) { 1385 return false; // Wait stable graph 1386 } 1387 uint cnt = phi->req(); 1388 for (uint i = 1; i < cnt; i++) { 1389 Node* rc = region->in(i); 1390 if (rc == NULL || phase->type(rc) == Type::TOP) 1391 return false; // Wait stable graph 1392 Node* in = phi->in(i); 1393 if (in == NULL || phase->type(in) == Type::TOP) 1394 return false; // Wait stable graph 1395 } 1396 return true; 1397 } 1398 //------------------------------split_through_phi------------------------------ 1399 // Split instance or boxed field load through Phi. 1400 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1401 Node* mem = in(Memory); 1402 Node* address = in(Address); 1403 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1404 1405 assert((t_oop != NULL) && 1406 (t_oop->is_known_instance_field() || 1407 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1408 1409 Compile* C = phase->C; 1410 intptr_t ignore = 0; 1411 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1412 bool base_is_phi = (base != NULL) && base->is_Phi(); 1413 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1414 (base != NULL) && (base == address->in(AddPNode::Base)) && 1415 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1416 1417 if (!((mem->is_Phi() || base_is_phi) && 1418 (load_boxed_values || t_oop->is_known_instance_field()))) { 1419 return NULL; // memory is not Phi 1420 } 1421 1422 if (mem->is_Phi()) { 1423 if (!stable_phi(mem->as_Phi(), phase)) { 1424 return NULL; // Wait stable graph 1425 } 1426 uint cnt = mem->req(); 1427 // Check for loop invariant memory. 1428 if (cnt == 3) { 1429 for (uint i = 1; i < cnt; i++) { 1430 Node* in = mem->in(i); 1431 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1432 if (m == mem) { 1433 if (i == 1) { 1434 // if the first edge was a loop, check second edge too. 1435 // If both are replaceable - we are in an infinite loop 1436 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1437 if (n == mem) { 1438 break; 1439 } 1440 } 1441 set_req(Memory, mem->in(cnt - i)); 1442 return this; // made change 1443 } 1444 } 1445 } 1446 } 1447 if (base_is_phi) { 1448 if (!stable_phi(base->as_Phi(), phase)) { 1449 return NULL; // Wait stable graph 1450 } 1451 uint cnt = base->req(); 1452 // Check for loop invariant memory. 1453 if (cnt == 3) { 1454 for (uint i = 1; i < cnt; i++) { 1455 if (base->in(i) == base) { 1456 return NULL; // Wait stable graph 1457 } 1458 } 1459 } 1460 } 1461 1462 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1463 1464 // Split through Phi (see original code in loopopts.cpp). 1465 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1466 1467 // Do nothing here if Identity will find a value 1468 // (to avoid infinite chain of value phis generation). 1469 if (!phase->eqv(this, phase->apply_identity(this))) 1470 return NULL; 1471 1472 // Select Region to split through. 1473 Node* region; 1474 if (!base_is_phi) { 1475 assert(mem->is_Phi(), "sanity"); 1476 region = mem->in(0); 1477 // Skip if the region dominates some control edge of the address. 1478 if (!MemNode::all_controls_dominate(address, region)) 1479 return NULL; 1480 } else if (!mem->is_Phi()) { 1481 assert(base_is_phi, "sanity"); 1482 region = base->in(0); 1483 // Skip if the region dominates some control edge of the memory. 1484 if (!MemNode::all_controls_dominate(mem, region)) 1485 return NULL; 1486 } else if (base->in(0) != mem->in(0)) { 1487 assert(base_is_phi && mem->is_Phi(), "sanity"); 1488 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1489 region = base->in(0); 1490 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1491 region = mem->in(0); 1492 } else { 1493 return NULL; // complex graph 1494 } 1495 } else { 1496 assert(base->in(0) == mem->in(0), "sanity"); 1497 region = mem->in(0); 1498 } 1499 1500 const Type* this_type = this->bottom_type(); 1501 int this_index = C->get_alias_index(t_oop); 1502 int this_offset = t_oop->offset(); 1503 int this_iid = t_oop->instance_id(); 1504 if (!t_oop->is_known_instance() && load_boxed_values) { 1505 // Use _idx of address base for boxed values. 1506 this_iid = base->_idx; 1507 } 1508 PhaseIterGVN* igvn = phase->is_IterGVN(); 1509 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1510 for (uint i = 1; i < region->req(); i++) { 1511 Node* x; 1512 Node* the_clone = NULL; 1513 if (region->in(i) == C->top()) { 1514 x = C->top(); // Dead path? Use a dead data op 1515 } else { 1516 x = this->clone(); // Else clone up the data op 1517 the_clone = x; // Remember for possible deletion. 1518 // Alter data node to use pre-phi inputs 1519 if (this->in(0) == region) { 1520 x->set_req(0, region->in(i)); 1521 } else { 1522 x->set_req(0, NULL); 1523 } 1524 if (mem->is_Phi() && (mem->in(0) == region)) { 1525 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1526 } 1527 if (address->is_Phi() && address->in(0) == region) { 1528 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1529 } 1530 if (base_is_phi && (base->in(0) == region)) { 1531 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1532 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1533 x->set_req(Address, adr_x); 1534 } 1535 } 1536 // Check for a 'win' on some paths 1537 const Type *t = x->Value(igvn); 1538 1539 bool singleton = t->singleton(); 1540 1541 // See comments in PhaseIdealLoop::split_thru_phi(). 1542 if (singleton && t == Type::TOP) { 1543 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1544 } 1545 1546 if (singleton) { 1547 x = igvn->makecon(t); 1548 } else { 1549 // We now call Identity to try to simplify the cloned node. 1550 // Note that some Identity methods call phase->type(this). 1551 // Make sure that the type array is big enough for 1552 // our new node, even though we may throw the node away. 1553 // (This tweaking with igvn only works because x is a new node.) 1554 igvn->set_type(x, t); 1555 // If x is a TypeNode, capture any more-precise type permanently into Node 1556 // otherwise it will be not updated during igvn->transform since 1557 // igvn->type(x) is set to x->Value() already. 1558 x->raise_bottom_type(t); 1559 Node *y = igvn->apply_identity(x); 1560 if (y != x) { 1561 x = y; 1562 } else { 1563 y = igvn->hash_find_insert(x); 1564 if (y) { 1565 x = y; 1566 } else { 1567 // Else x is a new node we are keeping 1568 // We do not need register_new_node_with_optimizer 1569 // because set_type has already been called. 1570 igvn->_worklist.push(x); 1571 } 1572 } 1573 } 1574 if (x != the_clone && the_clone != NULL) { 1575 igvn->remove_dead_node(the_clone); 1576 } 1577 phi->set_req(i, x); 1578 } 1579 // Record Phi 1580 igvn->register_new_node_with_optimizer(phi); 1581 return phi; 1582 } 1583 1584 //------------------------------Ideal------------------------------------------ 1585 // If the load is from Field memory and the pointer is non-null, it might be possible to 1586 // zero out the control input. 1587 // If the offset is constant and the base is an object allocation, 1588 // try to hook me up to the exact initializing store. 1589 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1590 Node* p = MemNode::Ideal_common(phase, can_reshape); 1591 if (p) return (p == NodeSentinel) ? NULL : p; 1592 1593 Node* ctrl = in(MemNode::Control); 1594 Node* address = in(MemNode::Address); 1595 bool progress = false; 1596 1597 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1598 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1599 1600 // Skip up past a SafePoint control. Cannot do this for Stores because 1601 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1602 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1603 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1604 !addr_mark ) { 1605 ctrl = ctrl->in(0); 1606 set_req(MemNode::Control,ctrl); 1607 progress = true; 1608 } 1609 1610 intptr_t ignore = 0; 1611 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1612 if (base != NULL 1613 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1614 // Check for useless control edge in some common special cases 1615 if (in(MemNode::Control) != NULL 1616 && can_remove_control() 1617 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1618 && all_controls_dominate(base, phase->C->start())) { 1619 // A method-invariant, non-null address (constant or 'this' argument). 1620 set_req(MemNode::Control, NULL); 1621 progress = true; 1622 } 1623 } 1624 1625 Node* mem = in(MemNode::Memory); 1626 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1627 1628 if (can_reshape && (addr_t != NULL)) { 1629 // try to optimize our memory input 1630 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1631 if (opt_mem != mem) { 1632 set_req(MemNode::Memory, opt_mem); 1633 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1634 return this; 1635 } 1636 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1637 if ((t_oop != NULL) && 1638 (t_oop->is_known_instance_field() || 1639 t_oop->is_ptr_to_boxed_value())) { 1640 PhaseIterGVN *igvn = phase->is_IterGVN(); 1641 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1642 // Delay this transformation until memory Phi is processed. 1643 phase->is_IterGVN()->_worklist.push(this); 1644 return NULL; 1645 } 1646 // Split instance field load through Phi. 1647 Node* result = split_through_phi(phase); 1648 if (result != NULL) return result; 1649 1650 if (t_oop->is_ptr_to_boxed_value()) { 1651 Node* result = eliminate_autobox(phase); 1652 if (result != NULL) return result; 1653 } 1654 } 1655 } 1656 1657 // Is there a dominating load that loads the same value? Leave 1658 // anything that is not a load of a field/array element (like 1659 // barriers etc.) alone 1660 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) { 1661 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1662 Node *use = mem->fast_out(i); 1663 if (use != this && 1664 use->Opcode() == Opcode() && 1665 use->in(0) != NULL && 1666 use->in(0) != in(0) && 1667 use->in(Address) == in(Address)) { 1668 Node* ctl = in(0); 1669 for (int i = 0; i < 10 && ctl != NULL; i++) { 1670 ctl = IfNode::up_one_dom(ctl); 1671 if (ctl == use->in(0)) { 1672 set_req(0, use->in(0)); 1673 return this; 1674 } 1675 } 1676 } 1677 } 1678 } 1679 1680 // Check for prior store with a different base or offset; make Load 1681 // independent. Skip through any number of them. Bail out if the stores 1682 // are in an endless dead cycle and report no progress. This is a key 1683 // transform for Reflection. However, if after skipping through the Stores 1684 // we can't then fold up against a prior store do NOT do the transform as 1685 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1686 // array memory alive twice: once for the hoisted Load and again after the 1687 // bypassed Store. This situation only works if EVERYBODY who does 1688 // anti-dependence work knows how to bypass. I.e. we need all 1689 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1690 // the alias index stuff. So instead, peek through Stores and IFF we can 1691 // fold up, do so. 1692 Node* prev_mem = find_previous_store(phase); 1693 if (prev_mem != NULL) { 1694 Node* value = can_see_arraycopy_value(prev_mem, phase); 1695 if (value != NULL) { 1696 return value; 1697 } 1698 } 1699 // Steps (a), (b): Walk past independent stores to find an exact match. 1700 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1701 // (c) See if we can fold up on the spot, but don't fold up here. 1702 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1703 // just return a prior value, which is done by Identity calls. 1704 if (can_see_stored_value(prev_mem, phase)) { 1705 // Make ready for step (d): 1706 set_req(MemNode::Memory, prev_mem); 1707 return this; 1708 } 1709 } 1710 1711 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase); 1712 if (alloc != NULL && mem->is_Proj() && 1713 mem->in(0) != NULL && 1714 mem->in(0) == alloc->initialization() && 1715 Opcode() == Op_LoadX && 1716 alloc->initialization()->proj_out_or_null(0) != NULL) { 1717 InitializeNode* init = alloc->initialization(); 1718 Node* control = init->proj_out(0); 1719 return alloc->make_ideal_mark(phase, address, control, mem, NULL); 1720 } 1721 1722 return progress ? this : NULL; 1723 } 1724 1725 // Helper to recognize certain Klass fields which are invariant across 1726 // some group of array types (e.g., int[] or all T[] where T < Object). 1727 const Type* 1728 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1729 ciKlass* klass) const { 1730 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1731 // The field is Klass::_modifier_flags. Return its (constant) value. 1732 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1733 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1734 return TypeInt::make(klass->modifier_flags()); 1735 } 1736 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1737 // The field is Klass::_access_flags. Return its (constant) value. 1738 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1739 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1740 return TypeInt::make(klass->access_flags()); 1741 } 1742 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1743 // The field is Klass::_layout_helper. Return its constant value if known. 1744 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1745 return TypeInt::make(klass->layout_helper()); 1746 } 1747 1748 // No match. 1749 return NULL; 1750 } 1751 1752 //------------------------------Value----------------------------------------- 1753 const Type* LoadNode::Value(PhaseGVN* phase) const { 1754 // Either input is TOP ==> the result is TOP 1755 Node* mem = in(MemNode::Memory); 1756 const Type *t1 = phase->type(mem); 1757 if (t1 == Type::TOP) return Type::TOP; 1758 Node* adr = in(MemNode::Address); 1759 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1760 if (tp == NULL || tp->empty()) return Type::TOP; 1761 int off = tp->offset(); 1762 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1763 Compile* C = phase->C; 1764 1765 // Try to guess loaded type from pointer type 1766 if (tp->isa_aryptr()) { 1767 const TypeAryPtr* ary = tp->is_aryptr(); 1768 const Type* t = ary->elem(); 1769 1770 // Determine whether the reference is beyond the header or not, by comparing 1771 // the offset against the offset of the start of the array's data. 1772 // Different array types begin at slightly different offsets (12 vs. 16). 1773 // We choose T_BYTE as an example base type that is least restrictive 1774 // as to alignment, which will therefore produce the smallest 1775 // possible base offset. 1776 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1777 const bool off_beyond_header = (off >= min_base_off); 1778 1779 // Try to constant-fold a stable array element. 1780 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1781 // Make sure the reference is not into the header and the offset is constant 1782 ciObject* aobj = ary->const_oop(); 1783 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1784 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1785 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1786 stable_dimension, 1787 memory_type(), is_unsigned()); 1788 if (con_type != NULL) { 1789 return con_type; 1790 } 1791 } 1792 } 1793 1794 // Don't do this for integer types. There is only potential profit if 1795 // the element type t is lower than _type; that is, for int types, if _type is 1796 // more restrictive than t. This only happens here if one is short and the other 1797 // char (both 16 bits), and in those cases we've made an intentional decision 1798 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1799 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1800 // 1801 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1802 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1803 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1804 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1805 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1806 // In fact, that could have been the original type of p1, and p1 could have 1807 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1808 // expression (LShiftL quux 3) independently optimized to the constant 8. 1809 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1810 && (_type->isa_vect() == NULL) 1811 && t->isa_valuetype() == NULL 1812 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1813 // t might actually be lower than _type, if _type is a unique 1814 // concrete subclass of abstract class t. 1815 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header? 1816 const Type* jt = t->join_speculative(_type); 1817 // In any case, do not allow the join, per se, to empty out the type. 1818 if (jt->empty() && !t->empty()) { 1819 // This can happen if a interface-typed array narrows to a class type. 1820 jt = _type; 1821 } 1822 #ifdef ASSERT 1823 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1824 // The pointers in the autobox arrays are always non-null 1825 Node* base = adr->in(AddPNode::Base); 1826 if ((base != NULL) && base->is_DecodeN()) { 1827 // Get LoadN node which loads IntegerCache.cache field 1828 base = base->in(1); 1829 } 1830 if ((base != NULL) && base->is_Con()) { 1831 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1832 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1833 // It could be narrow oop 1834 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1835 } 1836 } 1837 } 1838 #endif 1839 return jt; 1840 } 1841 } 1842 } else if (tp->base() == Type::InstPtr) { 1843 assert( off != Type::OffsetBot || 1844 // arrays can be cast to Objects 1845 tp->is_oopptr()->klass()->is_java_lang_Object() || 1846 tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() || 1847 // unsafe field access may not have a constant offset 1848 C->has_unsafe_access(), 1849 "Field accesses must be precise" ); 1850 // For oop loads, we expect the _type to be precise. 1851 1852 const TypeInstPtr* tinst = tp->is_instptr(); 1853 BasicType bt = memory_type(); 1854 1855 // Fold component and value mirror loads 1856 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1857 if (ik == phase->C->env()->Class_klass() && (off == java_lang_Class::component_mirror_offset_in_bytes() || 1858 off == java_lang_Class::value_mirror_offset_in_bytes())) { 1859 ciType* mirror_type = tinst->java_mirror_type(); 1860 if (mirror_type != NULL) { 1861 const Type* const_oop = TypePtr::NULL_PTR; 1862 if (mirror_type->is_array_klass()) { 1863 const_oop = TypeInstPtr::make(mirror_type->as_array_klass()->component_mirror_instance()); 1864 } else if (mirror_type->is_valuetype()) { 1865 const_oop = TypeInstPtr::make(mirror_type->as_value_klass()->value_mirror_instance()); 1866 } 1867 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop; 1868 } 1869 } 1870 1871 // Optimize loads from constant fields. 1872 ciObject* const_oop = tinst->const_oop(); 1873 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1874 ciType* mirror_type = const_oop->as_instance()->java_mirror_type(); 1875 if (mirror_type != NULL && mirror_type->is_valuetype()) { 1876 ciValueKlass* vk = mirror_type->as_value_klass(); 1877 if (off == vk->default_value_offset()) { 1878 // Loading a special hidden field that contains the oop of the default value type 1879 const Type* const_oop = TypeInstPtr::make(vk->default_value_instance()); 1880 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop; 1881 } 1882 } 1883 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt); 1884 if (con_type != NULL) { 1885 return con_type; 1886 } 1887 } 1888 } else if (tp->base() == Type::KlassPtr) { 1889 assert( off != Type::OffsetBot || 1890 // arrays can be cast to Objects 1891 tp->is_klassptr()->klass() == NULL || 1892 tp->is_klassptr()->klass()->is_java_lang_Object() || 1893 // also allow array-loading from the primary supertype 1894 // array during subtype checks 1895 Opcode() == Op_LoadKlass, 1896 "Field accesses must be precise" ); 1897 // For klass/static loads, we expect the _type to be precise 1898 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) { 1899 if (adr->is_Load() && off == 0) { 1900 /* With mirrors being an indirect in the Klass* 1901 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 1902 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 1903 * 1904 * So check the type and klass of the node before the LoadP. 1905 */ 1906 Node* adr2 = adr->in(MemNode::Address); 1907 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1908 if (tkls != NULL) { 1909 ciKlass* klass = tkls->klass(); 1910 if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1911 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1912 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1913 return TypeInstPtr::make(klass->java_mirror()); 1914 } 1915 } 1916 } else { 1917 // Check for a load of the default value offset from the ValueKlassFixedBlock: 1918 // LoadI(LoadP(value_klass, adr_valueklass_fixed_block_offset), default_value_offset_offset) 1919 intptr_t offset = 0; 1920 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1921 if (base != NULL && base->is_Load() && offset == in_bytes(ValueKlass::default_value_offset_offset())) { 1922 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr(); 1923 if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_valuetype() && 1924 tkls->offset() == in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())) { 1925 assert(base->Opcode() == Op_LoadP, "must load an oop from klass"); 1926 assert(Opcode() == Op_LoadI, "must load an int from fixed block"); 1927 return TypeInt::make(tkls->klass()->as_value_klass()->default_value_offset()); 1928 } 1929 } 1930 } 1931 } 1932 1933 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1934 if (tkls != NULL && !StressReflectiveCode) { 1935 ciKlass* klass = tkls->klass(); 1936 if (tkls->is_loaded() && tkls->klass_is_exact()) { 1937 // We are loading a field from a Klass metaobject whose identity 1938 // is known at compile time (the type is "exact" or "precise"). 1939 // Check for fields we know are maintained as constants by the VM. 1940 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1941 // The field is Klass::_super_check_offset. Return its (constant) value. 1942 // (Folds up type checking code.) 1943 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1944 return TypeInt::make(klass->super_check_offset()); 1945 } 1946 // Compute index into primary_supers array 1947 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1948 // Check for overflowing; use unsigned compare to handle the negative case. 1949 if( depth < ciKlass::primary_super_limit() ) { 1950 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1951 // (Folds up type checking code.) 1952 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1953 ciKlass *ss = klass->super_of_depth(depth); 1954 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1955 } 1956 const Type* aift = load_array_final_field(tkls, klass); 1957 if (aift != NULL) return aift; 1958 } 1959 1960 // We can still check if we are loading from the primary_supers array at a 1961 // shallow enough depth. Even though the klass is not exact, entries less 1962 // than or equal to its super depth are correct. 1963 if (tkls->is_loaded()) { 1964 ciType *inner = klass; 1965 while( inner->is_obj_array_klass() ) 1966 inner = inner->as_obj_array_klass()->base_element_type(); 1967 if( inner->is_instance_klass() && 1968 !inner->as_instance_klass()->flags().is_interface() ) { 1969 // Compute index into primary_supers array 1970 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1971 // Check for overflowing; use unsigned compare to handle the negative case. 1972 if( depth < ciKlass::primary_super_limit() && 1973 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1974 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1975 // (Folds up type checking code.) 1976 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1977 ciKlass *ss = klass->super_of_depth(depth); 1978 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1979 } 1980 } 1981 } 1982 1983 // If the type is enough to determine that the thing is not an array, 1984 // we can give the layout_helper a positive interval type. 1985 // This will help short-circuit some reflective code. 1986 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1987 && !klass->is_array_klass() // not directly typed as an array 1988 && !klass->is_interface() // specifically not Serializable & Cloneable 1989 && !klass->is_java_lang_Object() // not the supertype of all T[] 1990 ) { 1991 // Note: When interfaces are reliable, we can narrow the interface 1992 // test to (klass != Serializable && klass != Cloneable). 1993 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1994 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1995 // The key property of this type is that it folds up tests 1996 // for array-ness, since it proves that the layout_helper is positive. 1997 // Thus, a generic value like the basic object layout helper works fine. 1998 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1999 } 2000 } 2001 2002 // If we are loading from a freshly-allocated object, produce a zero, 2003 // if the load is provably beyond the header of the object. 2004 // (Also allow a variable load from a fresh array to produce zero.) 2005 const TypeOopPtr *tinst = tp->isa_oopptr(); 2006 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 2007 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 2008 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 2009 Node* value = can_see_stored_value(mem,phase); 2010 if (value != NULL && value->is_Con()) { 2011 assert(value->bottom_type()->higher_equal(_type),"sanity"); 2012 return value->bottom_type(); 2013 } 2014 } 2015 2016 if (is_instance) { 2017 // If we have an instance type and our memory input is the 2018 // programs's initial memory state, there is no matching store, 2019 // so just return a zero of the appropriate type 2020 Node *mem = in(MemNode::Memory); 2021 if (mem->is_Parm() && mem->in(0)->is_Start()) { 2022 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 2023 return Type::get_zero_type(_type->basic_type()); 2024 } 2025 } 2026 return _type; 2027 } 2028 2029 //------------------------------match_edge------------------------------------- 2030 // Do we Match on this edge index or not? Match only the address. 2031 uint LoadNode::match_edge(uint idx) const { 2032 return idx == MemNode::Address; 2033 } 2034 2035 //--------------------------LoadBNode::Ideal-------------------------------------- 2036 // 2037 // If the previous store is to the same address as this load, 2038 // and the value stored was larger than a byte, replace this load 2039 // with the value stored truncated to a byte. If no truncation is 2040 // needed, the replacement is done in LoadNode::Identity(). 2041 // 2042 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2043 Node* mem = in(MemNode::Memory); 2044 Node* value = can_see_stored_value(mem,phase); 2045 if( value && !phase->type(value)->higher_equal( _type ) ) { 2046 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 2047 return new RShiftINode(result, phase->intcon(24)); 2048 } 2049 // Identity call will handle the case where truncation is not needed. 2050 return LoadNode::Ideal(phase, can_reshape); 2051 } 2052 2053 const Type* LoadBNode::Value(PhaseGVN* phase) const { 2054 Node* mem = in(MemNode::Memory); 2055 Node* value = can_see_stored_value(mem,phase); 2056 if (value != NULL && value->is_Con() && 2057 !value->bottom_type()->higher_equal(_type)) { 2058 // If the input to the store does not fit with the load's result type, 2059 // it must be truncated. We can't delay until Ideal call since 2060 // a singleton Value is needed for split_thru_phi optimization. 2061 int con = value->get_int(); 2062 return TypeInt::make((con << 24) >> 24); 2063 } 2064 return LoadNode::Value(phase); 2065 } 2066 2067 //--------------------------LoadUBNode::Ideal------------------------------------- 2068 // 2069 // If the previous store is to the same address as this load, 2070 // and the value stored was larger than a byte, replace this load 2071 // with the value stored truncated to a byte. If no truncation is 2072 // needed, the replacement is done in LoadNode::Identity(). 2073 // 2074 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2075 Node* mem = in(MemNode::Memory); 2076 Node* value = can_see_stored_value(mem, phase); 2077 if (value && !phase->type(value)->higher_equal(_type)) 2078 return new AndINode(value, phase->intcon(0xFF)); 2079 // Identity call will handle the case where truncation is not needed. 2080 return LoadNode::Ideal(phase, can_reshape); 2081 } 2082 2083 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2084 Node* mem = in(MemNode::Memory); 2085 Node* value = can_see_stored_value(mem,phase); 2086 if (value != NULL && value->is_Con() && 2087 !value->bottom_type()->higher_equal(_type)) { 2088 // If the input to the store does not fit with the load's result type, 2089 // it must be truncated. We can't delay until Ideal call since 2090 // a singleton Value is needed for split_thru_phi optimization. 2091 int con = value->get_int(); 2092 return TypeInt::make(con & 0xFF); 2093 } 2094 return LoadNode::Value(phase); 2095 } 2096 2097 //--------------------------LoadUSNode::Ideal------------------------------------- 2098 // 2099 // If the previous store is to the same address as this load, 2100 // and the value stored was larger than a char, replace this load 2101 // with the value stored truncated to a char. If no truncation is 2102 // needed, the replacement is done in LoadNode::Identity(). 2103 // 2104 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2105 Node* mem = in(MemNode::Memory); 2106 Node* value = can_see_stored_value(mem,phase); 2107 if( value && !phase->type(value)->higher_equal( _type ) ) 2108 return new AndINode(value,phase->intcon(0xFFFF)); 2109 // Identity call will handle the case where truncation is not needed. 2110 return LoadNode::Ideal(phase, can_reshape); 2111 } 2112 2113 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2114 Node* mem = in(MemNode::Memory); 2115 Node* value = can_see_stored_value(mem,phase); 2116 if (value != NULL && value->is_Con() && 2117 !value->bottom_type()->higher_equal(_type)) { 2118 // If the input to the store does not fit with the load's result type, 2119 // it must be truncated. We can't delay until Ideal call since 2120 // a singleton Value is needed for split_thru_phi optimization. 2121 int con = value->get_int(); 2122 return TypeInt::make(con & 0xFFFF); 2123 } 2124 return LoadNode::Value(phase); 2125 } 2126 2127 //--------------------------LoadSNode::Ideal-------------------------------------- 2128 // 2129 // If the previous store is to the same address as this load, 2130 // and the value stored was larger than a short, replace this load 2131 // with the value stored truncated to a short. If no truncation is 2132 // needed, the replacement is done in LoadNode::Identity(). 2133 // 2134 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2135 Node* mem = in(MemNode::Memory); 2136 Node* value = can_see_stored_value(mem,phase); 2137 if( value && !phase->type(value)->higher_equal( _type ) ) { 2138 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 2139 return new RShiftINode(result, phase->intcon(16)); 2140 } 2141 // Identity call will handle the case where truncation is not needed. 2142 return LoadNode::Ideal(phase, can_reshape); 2143 } 2144 2145 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2146 Node* mem = in(MemNode::Memory); 2147 Node* value = can_see_stored_value(mem,phase); 2148 if (value != NULL && value->is_Con() && 2149 !value->bottom_type()->higher_equal(_type)) { 2150 // If the input to the store does not fit with the load's result type, 2151 // it must be truncated. We can't delay until Ideal call since 2152 // a singleton Value is needed for split_thru_phi optimization. 2153 int con = value->get_int(); 2154 return TypeInt::make((con << 16) >> 16); 2155 } 2156 return LoadNode::Value(phase); 2157 } 2158 2159 //============================================================================= 2160 //----------------------------LoadKlassNode::make------------------------------ 2161 // Polymorphic factory method: 2162 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2163 // sanity check the alias category against the created node type 2164 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2165 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2166 #ifdef _LP64 2167 if (adr_type->is_ptr_to_narrowklass()) { 2168 assert(UseCompressedClassPointers, "no compressed klasses"); 2169 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2170 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2171 } 2172 #endif 2173 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2174 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2175 } 2176 2177 //------------------------------Value------------------------------------------ 2178 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2179 return klass_value_common(phase); 2180 } 2181 2182 // In most cases, LoadKlassNode does not have the control input set. If the control 2183 // input is set, it must not be removed (by LoadNode::Ideal()). 2184 bool LoadKlassNode::can_remove_control() const { 2185 return false; 2186 } 2187 2188 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2189 // Either input is TOP ==> the result is TOP 2190 const Type *t1 = phase->type( in(MemNode::Memory) ); 2191 if (t1 == Type::TOP) return Type::TOP; 2192 Node *adr = in(MemNode::Address); 2193 const Type *t2 = phase->type( adr ); 2194 if (t2 == Type::TOP) return Type::TOP; 2195 const TypePtr *tp = t2->is_ptr(); 2196 if (TypePtr::above_centerline(tp->ptr()) || 2197 tp->ptr() == TypePtr::Null) return Type::TOP; 2198 2199 // Return a more precise klass, if possible 2200 const TypeInstPtr *tinst = tp->isa_instptr(); 2201 if (tinst != NULL) { 2202 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2203 int offset = tinst->offset(); 2204 if (ik == phase->C->env()->Class_klass() 2205 && (offset == java_lang_Class::klass_offset_in_bytes() || 2206 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2207 // We are loading a special hidden field from a Class mirror object, 2208 // the field which points to the VM's Klass metaobject. 2209 bool is_val_type = false; 2210 ciType* t = tinst->java_mirror_type(&is_val_type); 2211 // java_mirror_type returns non-null for compile-time Class constants. 2212 if (t != NULL) { 2213 // constant oop => constant klass 2214 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2215 if (t->is_void()) { 2216 // We cannot create a void array. Since void is a primitive type return null 2217 // klass. Users of this result need to do a null check on the returned klass. 2218 return TypePtr::NULL_PTR; 2219 } 2220 return TypeKlassPtr::make(ciArrayKlass::make(t, /* never_null */ is_val_type)); 2221 } 2222 if (!t->is_klass()) { 2223 // a primitive Class (e.g., int.class) has NULL for a klass field 2224 return TypePtr::NULL_PTR; 2225 } 2226 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2227 return TypeKlassPtr::make(t->as_klass()); 2228 } 2229 // non-constant mirror, so we can't tell what's going on 2230 } 2231 if( !ik->is_loaded() ) 2232 return _type; // Bail out if not loaded 2233 if (offset == oopDesc::klass_offset_in_bytes()) { 2234 if (tinst->klass_is_exact()) { 2235 return TypeKlassPtr::make(ik); 2236 } 2237 // See if we can become precise: no subklasses and no interface 2238 // (Note: We need to support verified interfaces.) 2239 if (!ik->is_interface() && !ik->has_subklass()) { 2240 //assert(!UseExactTypes, "this code should be useless with exact types"); 2241 // Add a dependence; if any subclass added we need to recompile 2242 if (!ik->is_final()) { 2243 // %%% should use stronger assert_unique_concrete_subtype instead 2244 phase->C->dependencies()->assert_leaf_type(ik); 2245 } 2246 // Return precise klass 2247 return TypeKlassPtr::make(ik); 2248 } 2249 2250 // Return root of possible klass 2251 return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0)); 2252 } 2253 } 2254 2255 // Check for loading klass from an array 2256 const TypeAryPtr *tary = tp->isa_aryptr(); 2257 if (tary != NULL) { 2258 ciKlass *tary_klass = tary->klass(); 2259 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2260 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2261 ciArrayKlass* ak = tary_klass->as_array_klass(); 2262 // Do not fold klass loads from [V?. The runtime type might be [V due to [V <: [V? 2263 // and the klass for [V is not equal to the klass for [V?. 2264 if (!tary->is_known_instance() && ak->is_obj_array_klass() && 2265 !ak->storage_properties().is_null_free() && ak->element_klass()->is_valuetype()) { 2266 // Fall back to Object array 2267 ak = ciArrayKlass::make(phase->C->env()->Object_klass()); 2268 } else if (tary->klass_is_exact()) { 2269 return TypeKlassPtr::make(tary_klass); 2270 } 2271 2272 // If the klass is an object array, we defer the question to the 2273 // array component klass. 2274 if (ak->is_obj_array_klass()) { 2275 assert(ak->is_loaded(), ""); 2276 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2277 if (base_k->is_loaded() && base_k->is_instance_klass()) { 2278 ciInstanceKlass *ik = base_k->as_instance_klass(); 2279 // See if we can become precise: no subklasses and no interface 2280 if (!ik->is_interface() && !ik->has_subklass()) { 2281 //assert(!UseExactTypes, "this code should be useless with exact types"); 2282 // Add a dependence; if any subclass added we need to recompile 2283 if (!ik->is_final()) { 2284 phase->C->dependencies()->assert_leaf_type(ik); 2285 } 2286 // Return precise array klass 2287 return TypeKlassPtr::make(ak); 2288 } 2289 } 2290 return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0)); 2291 } else if (ak->is_type_array_klass()) { 2292 //assert(!UseExactTypes, "this code should be useless with exact types"); 2293 return TypeKlassPtr::make(ak); // These are always precise 2294 } 2295 } 2296 } 2297 2298 // Check for loading klass from an array klass 2299 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2300 if (tkls != NULL && !StressReflectiveCode) { 2301 if (!tkls->is_loaded()) { 2302 return _type; // Bail out if not loaded 2303 } 2304 ciKlass* klass = tkls->klass(); 2305 if( klass->is_obj_array_klass() && 2306 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2307 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2308 // // Always returning precise element type is incorrect, 2309 // // e.g., element type could be object and array may contain strings 2310 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2311 2312 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2313 // according to the element type's subclassing. 2314 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0)); 2315 } else if (klass->is_value_array_klass() && 2316 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2317 ciKlass* elem = klass->as_value_array_klass()->element_klass(); 2318 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0)); 2319 } 2320 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2321 tkls->offset() == in_bytes(Klass::super_offset())) { 2322 ciKlass* sup = klass->as_instance_klass()->super(); 2323 // The field is Klass::_super. Return its (constant) value. 2324 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2325 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2326 } 2327 } 2328 2329 // Bailout case 2330 return LoadNode::Value(phase); 2331 } 2332 2333 //------------------------------Identity--------------------------------------- 2334 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2335 // Also feed through the klass in Allocate(...klass...)._klass. 2336 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2337 return klass_identity_common(phase); 2338 } 2339 2340 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2341 Node* x = LoadNode::Identity(phase); 2342 if (x != this) return x; 2343 2344 // Take apart the address into an oop and and offset. 2345 // Return 'this' if we cannot. 2346 Node* adr = in(MemNode::Address); 2347 intptr_t offset = 0; 2348 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2349 if (base == NULL) return this; 2350 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2351 if (toop == NULL) return this; 2352 2353 // Step over potential GC barrier for OopHandle resolve 2354 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2355 if (bs->is_gc_barrier_node(base)) { 2356 base = bs->step_over_gc_barrier(base); 2357 } 2358 2359 // We can fetch the klass directly through an AllocateNode. 2360 // This works even if the klass is not constant (clone or newArray). 2361 if (offset == oopDesc::klass_offset_in_bytes()) { 2362 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2363 if (allocated_klass != NULL) { 2364 return allocated_klass; 2365 } 2366 } 2367 2368 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2369 // See inline_native_Class_query for occurrences of these patterns. 2370 // Java Example: x.getClass().isAssignableFrom(y) 2371 // 2372 // This improves reflective code, often making the Class 2373 // mirror go completely dead. (Current exception: Class 2374 // mirrors may appear in debug info, but we could clean them out by 2375 // introducing a new debug info operator for Klass.java_mirror). 2376 2377 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2378 && offset == java_lang_Class::klass_offset_in_bytes()) { 2379 if (base->is_Load()) { 2380 Node* base2 = base->in(MemNode::Address); 2381 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2382 Node* adr2 = base2->in(MemNode::Address); 2383 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2384 if (tkls != NULL && !tkls->empty() 2385 && (tkls->klass()->is_instance_klass() || 2386 tkls->klass()->is_array_klass()) 2387 && adr2->is_AddP() 2388 ) { 2389 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2390 if (tkls->offset() == mirror_field) { 2391 return adr2->in(AddPNode::Base); 2392 } 2393 } 2394 } 2395 } 2396 } 2397 2398 return this; 2399 } 2400 2401 2402 //------------------------------Value------------------------------------------ 2403 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2404 const Type *t = klass_value_common(phase); 2405 if (t == Type::TOP) 2406 return t; 2407 2408 return t->make_narrowklass(); 2409 } 2410 2411 //------------------------------Identity--------------------------------------- 2412 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2413 // Also feed through the klass in Allocate(...klass...)._klass. 2414 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2415 Node *x = klass_identity_common(phase); 2416 2417 const Type *t = phase->type( x ); 2418 if( t == Type::TOP ) return x; 2419 if( t->isa_narrowklass()) return x; 2420 assert (!t->isa_narrowoop(), "no narrow oop here"); 2421 2422 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2423 } 2424 2425 //------------------------------Value----------------------------------------- 2426 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2427 // Either input is TOP ==> the result is TOP 2428 const Type *t1 = phase->type( in(MemNode::Memory) ); 2429 if( t1 == Type::TOP ) return Type::TOP; 2430 Node *adr = in(MemNode::Address); 2431 const Type *t2 = phase->type( adr ); 2432 if( t2 == Type::TOP ) return Type::TOP; 2433 const TypePtr *tp = t2->is_ptr(); 2434 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2435 const TypeAryPtr *tap = tp->isa_aryptr(); 2436 if( !tap ) return _type; 2437 return tap->size(); 2438 } 2439 2440 //-------------------------------Ideal--------------------------------------- 2441 // Feed through the length in AllocateArray(...length...)._length. 2442 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2443 Node* p = MemNode::Ideal_common(phase, can_reshape); 2444 if (p) return (p == NodeSentinel) ? NULL : p; 2445 2446 // Take apart the address into an oop and and offset. 2447 // Return 'this' if we cannot. 2448 Node* adr = in(MemNode::Address); 2449 intptr_t offset = 0; 2450 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2451 if (base == NULL) return NULL; 2452 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2453 if (tary == NULL) return NULL; 2454 2455 // We can fetch the length directly through an AllocateArrayNode. 2456 // This works even if the length is not constant (clone or newArray). 2457 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2458 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2459 if (alloc != NULL) { 2460 Node* allocated_length = alloc->Ideal_length(); 2461 Node* len = alloc->make_ideal_length(tary, phase); 2462 if (allocated_length != len) { 2463 // New CastII improves on this. 2464 return len; 2465 } 2466 } 2467 } 2468 2469 return NULL; 2470 } 2471 2472 //------------------------------Identity--------------------------------------- 2473 // Feed through the length in AllocateArray(...length...)._length. 2474 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2475 Node* x = LoadINode::Identity(phase); 2476 if (x != this) return x; 2477 2478 // Take apart the address into an oop and and offset. 2479 // Return 'this' if we cannot. 2480 Node* adr = in(MemNode::Address); 2481 intptr_t offset = 0; 2482 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2483 if (base == NULL) return this; 2484 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2485 if (tary == NULL) return this; 2486 2487 // We can fetch the length directly through an AllocateArrayNode. 2488 // This works even if the length is not constant (clone or newArray). 2489 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2490 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2491 if (alloc != NULL) { 2492 Node* allocated_length = alloc->Ideal_length(); 2493 // Do not allow make_ideal_length to allocate a CastII node. 2494 Node* len = alloc->make_ideal_length(tary, phase, false); 2495 if (allocated_length == len) { 2496 // Return allocated_length only if it would not be improved by a CastII. 2497 return allocated_length; 2498 } 2499 } 2500 } 2501 2502 return this; 2503 2504 } 2505 2506 //============================================================================= 2507 //---------------------------StoreNode::make----------------------------------- 2508 // Polymorphic factory method: 2509 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2510 assert((mo == unordered || mo == release), "unexpected"); 2511 Compile* C = gvn.C; 2512 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2513 ctl != NULL, "raw memory operations should have control edge"); 2514 2515 switch (bt) { 2516 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2517 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2518 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2519 case T_CHAR: 2520 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2521 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2522 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2523 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2524 case T_METADATA: 2525 case T_ADDRESS: 2526 case T_VALUETYPE: 2527 case T_OBJECT: 2528 #ifdef _LP64 2529 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2530 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2531 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2532 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2533 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2534 adr->bottom_type()->isa_rawptr())) { 2535 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2536 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2537 } 2538 #endif 2539 { 2540 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2541 } 2542 default: 2543 ShouldNotReachHere(); 2544 return (StoreNode*)NULL; 2545 } 2546 } 2547 2548 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2549 bool require_atomic = true; 2550 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2551 } 2552 2553 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2554 bool require_atomic = true; 2555 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2556 } 2557 2558 2559 //--------------------------bottom_type---------------------------------------- 2560 const Type *StoreNode::bottom_type() const { 2561 return Type::MEMORY; 2562 } 2563 2564 //------------------------------hash------------------------------------------- 2565 uint StoreNode::hash() const { 2566 // unroll addition of interesting fields 2567 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2568 2569 // Since they are not commoned, do not hash them: 2570 return NO_HASH; 2571 } 2572 2573 //------------------------------Ideal------------------------------------------ 2574 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2575 // When a store immediately follows a relevant allocation/initialization, 2576 // try to capture it into the initialization, or hoist it above. 2577 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2578 Node* p = MemNode::Ideal_common(phase, can_reshape); 2579 if (p) return (p == NodeSentinel) ? NULL : p; 2580 2581 Node* mem = in(MemNode::Memory); 2582 Node* address = in(MemNode::Address); 2583 // Back-to-back stores to same address? Fold em up. Generally 2584 // unsafe if I have intervening uses... Also disallowed for StoreCM 2585 // since they must follow each StoreP operation. Redundant StoreCMs 2586 // are eliminated just before matching in final_graph_reshape. 2587 { 2588 Node* st = mem; 2589 // If Store 'st' has more than one use, we cannot fold 'st' away. 2590 // For example, 'st' might be the final state at a conditional 2591 // return. Or, 'st' might be used by some node which is live at 2592 // the same time 'st' is live, which might be unschedulable. So, 2593 // require exactly ONE user until such time as we clone 'mem' for 2594 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2595 // true). 2596 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2597 // Looking at a dead closed cycle of memory? 2598 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2599 assert(Opcode() == st->Opcode() || 2600 st->Opcode() == Op_StoreVector || 2601 Opcode() == Op_StoreVector || 2602 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2603 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2604 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2605 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) || 2606 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2607 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2608 2609 if (st->in(MemNode::Address)->eqv_uncast(address) && 2610 st->as_Store()->memory_size() <= this->memory_size()) { 2611 Node* use = st->raw_out(0); 2612 phase->igvn_rehash_node_delayed(use); 2613 if (can_reshape) { 2614 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2615 } else { 2616 // It's OK to do this in the parser, since DU info is always accurate, 2617 // and the parser always refers to nodes via SafePointNode maps. 2618 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2619 } 2620 return this; 2621 } 2622 st = st->in(MemNode::Memory); 2623 } 2624 } 2625 2626 2627 // Capture an unaliased, unconditional, simple store into an initializer. 2628 // Or, if it is independent of the allocation, hoist it above the allocation. 2629 if (ReduceFieldZeroing && /*can_reshape &&*/ 2630 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2631 InitializeNode* init = mem->in(0)->as_Initialize(); 2632 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2633 if (offset > 0) { 2634 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2635 // If the InitializeNode captured me, it made a raw copy of me, 2636 // and I need to disappear. 2637 if (moved != NULL) { 2638 // %%% hack to ensure that Ideal returns a new node: 2639 mem = MergeMemNode::make(mem); 2640 return mem; // fold me away 2641 } 2642 } 2643 } 2644 2645 return NULL; // No further progress 2646 } 2647 2648 //------------------------------Value----------------------------------------- 2649 const Type* StoreNode::Value(PhaseGVN* phase) const { 2650 // Either input is TOP ==> the result is TOP 2651 const Type *t1 = phase->type( in(MemNode::Memory) ); 2652 if( t1 == Type::TOP ) return Type::TOP; 2653 const Type *t2 = phase->type( in(MemNode::Address) ); 2654 if( t2 == Type::TOP ) return Type::TOP; 2655 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2656 if( t3 == Type::TOP ) return Type::TOP; 2657 return Type::MEMORY; 2658 } 2659 2660 //------------------------------Identity--------------------------------------- 2661 // Remove redundant stores: 2662 // Store(m, p, Load(m, p)) changes to m. 2663 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2664 Node* StoreNode::Identity(PhaseGVN* phase) { 2665 Node* mem = in(MemNode::Memory); 2666 Node* adr = in(MemNode::Address); 2667 Node* val = in(MemNode::ValueIn); 2668 2669 Node* result = this; 2670 2671 // Load then Store? Then the Store is useless 2672 if (val->is_Load() && 2673 val->in(MemNode::Address)->eqv_uncast(adr) && 2674 val->in(MemNode::Memory )->eqv_uncast(mem) && 2675 val->as_Load()->store_Opcode() == Opcode()) { 2676 result = mem; 2677 } 2678 2679 // Two stores in a row of the same value? 2680 if (result == this && 2681 mem->is_Store() && 2682 mem->in(MemNode::Address)->eqv_uncast(adr) && 2683 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2684 mem->Opcode() == Opcode()) { 2685 result = mem; 2686 } 2687 2688 // Store of zero anywhere into a freshly-allocated object? 2689 // Then the store is useless. 2690 // (It must already have been captured by the InitializeNode.) 2691 if (result == this && ReduceFieldZeroing) { 2692 // a newly allocated object is already all-zeroes everywhere 2693 if (mem->is_Proj() && mem->in(0)->is_Allocate() && 2694 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) { 2695 assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to value array is forbidden"); 2696 result = mem; 2697 } 2698 2699 if (result == this) { 2700 // the store may also apply to zero-bits in an earlier object 2701 Node* prev_mem = find_previous_store(phase); 2702 // Steps (a), (b): Walk past independent stores to find an exact match. 2703 if (prev_mem != NULL) { 2704 Node* prev_val = can_see_stored_value(prev_mem, phase); 2705 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2706 // prev_val and val might differ by a cast; it would be good 2707 // to keep the more informative of the two. 2708 if (phase->type(val)->is_zero_type()) { 2709 result = mem; 2710 } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) { 2711 InitializeNode* init = prev_mem->in(0)->as_Initialize(); 2712 AllocateNode* alloc = init->allocation(); 2713 if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) { 2714 result = mem; 2715 } 2716 } 2717 } 2718 } 2719 } 2720 } 2721 2722 if (result != this && phase->is_IterGVN() != NULL) { 2723 MemBarNode* trailing = trailing_membar(); 2724 if (trailing != NULL) { 2725 #ifdef ASSERT 2726 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2727 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects"); 2728 #endif 2729 PhaseIterGVN* igvn = phase->is_IterGVN(); 2730 trailing->remove(igvn); 2731 } 2732 } 2733 2734 return result; 2735 } 2736 2737 //------------------------------match_edge------------------------------------- 2738 // Do we Match on this edge index or not? Match only memory & value 2739 uint StoreNode::match_edge(uint idx) const { 2740 return idx == MemNode::Address || idx == MemNode::ValueIn; 2741 } 2742 2743 //------------------------------cmp-------------------------------------------- 2744 // Do not common stores up together. They generally have to be split 2745 // back up anyways, so do not bother. 2746 bool StoreNode::cmp( const Node &n ) const { 2747 return (&n == this); // Always fail except on self 2748 } 2749 2750 //------------------------------Ideal_masked_input----------------------------- 2751 // Check for a useless mask before a partial-word store 2752 // (StoreB ... (AndI valIn conIa) ) 2753 // If (conIa & mask == mask) this simplifies to 2754 // (StoreB ... (valIn) ) 2755 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2756 Node *val = in(MemNode::ValueIn); 2757 if( val->Opcode() == Op_AndI ) { 2758 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2759 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2760 set_req(MemNode::ValueIn, val->in(1)); 2761 return this; 2762 } 2763 } 2764 return NULL; 2765 } 2766 2767 2768 //------------------------------Ideal_sign_extended_input---------------------- 2769 // Check for useless sign-extension before a partial-word store 2770 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2771 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2772 // (StoreB ... (valIn) ) 2773 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2774 Node *val = in(MemNode::ValueIn); 2775 if( val->Opcode() == Op_RShiftI ) { 2776 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2777 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2778 Node *shl = val->in(1); 2779 if( shl->Opcode() == Op_LShiftI ) { 2780 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2781 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2782 set_req(MemNode::ValueIn, shl->in(1)); 2783 return this; 2784 } 2785 } 2786 } 2787 } 2788 return NULL; 2789 } 2790 2791 //------------------------------value_never_loaded----------------------------------- 2792 // Determine whether there are any possible loads of the value stored. 2793 // For simplicity, we actually check if there are any loads from the 2794 // address stored to, not just for loads of the value stored by this node. 2795 // 2796 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2797 Node *adr = in(Address); 2798 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2799 if (adr_oop == NULL) 2800 return false; 2801 if (!adr_oop->is_known_instance_field()) 2802 return false; // if not a distinct instance, there may be aliases of the address 2803 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2804 Node *use = adr->fast_out(i); 2805 if (use->is_Load() || use->is_LoadStore()) { 2806 return false; 2807 } 2808 } 2809 return true; 2810 } 2811 2812 MemBarNode* StoreNode::trailing_membar() const { 2813 if (is_release()) { 2814 MemBarNode* trailing_mb = NULL; 2815 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2816 Node* u = fast_out(i); 2817 if (u->is_MemBar()) { 2818 if (u->as_MemBar()->trailing_store()) { 2819 assert(u->Opcode() == Op_MemBarVolatile, ""); 2820 assert(trailing_mb == NULL, "only one"); 2821 trailing_mb = u->as_MemBar(); 2822 #ifdef ASSERT 2823 Node* leading = u->as_MemBar()->leading_membar(); 2824 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2825 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2826 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2827 #endif 2828 } else { 2829 assert(u->as_MemBar()->standalone(), ""); 2830 } 2831 } 2832 } 2833 return trailing_mb; 2834 } 2835 return NULL; 2836 } 2837 2838 2839 //============================================================================= 2840 //------------------------------Ideal------------------------------------------ 2841 // If the store is from an AND mask that leaves the low bits untouched, then 2842 // we can skip the AND operation. If the store is from a sign-extension 2843 // (a left shift, then right shift) we can skip both. 2844 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2845 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2846 if( progress != NULL ) return progress; 2847 2848 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2849 if( progress != NULL ) return progress; 2850 2851 // Finally check the default case 2852 return StoreNode::Ideal(phase, can_reshape); 2853 } 2854 2855 //============================================================================= 2856 //------------------------------Ideal------------------------------------------ 2857 // If the store is from an AND mask that leaves the low bits untouched, then 2858 // we can skip the AND operation 2859 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2860 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2861 if( progress != NULL ) return progress; 2862 2863 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2864 if( progress != NULL ) return progress; 2865 2866 // Finally check the default case 2867 return StoreNode::Ideal(phase, can_reshape); 2868 } 2869 2870 //============================================================================= 2871 //------------------------------Identity--------------------------------------- 2872 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2873 // No need to card mark when storing a null ptr 2874 Node* my_store = in(MemNode::OopStore); 2875 if (my_store->is_Store()) { 2876 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2877 if( t1 == TypePtr::NULL_PTR ) { 2878 return in(MemNode::Memory); 2879 } 2880 } 2881 return this; 2882 } 2883 2884 //============================================================================= 2885 //------------------------------Ideal--------------------------------------- 2886 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2887 Node* progress = StoreNode::Ideal(phase, can_reshape); 2888 if (progress != NULL) return progress; 2889 2890 Node* my_store = in(MemNode::OopStore); 2891 if (my_store->is_MergeMem()) { 2892 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2893 set_req(MemNode::OopStore, mem); 2894 return this; 2895 } 2896 2897 return NULL; 2898 } 2899 2900 //------------------------------Value----------------------------------------- 2901 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2902 // Either input is TOP ==> the result is TOP 2903 const Type *t = phase->type( in(MemNode::Memory) ); 2904 if( t == Type::TOP ) return Type::TOP; 2905 t = phase->type( in(MemNode::Address) ); 2906 if( t == Type::TOP ) return Type::TOP; 2907 t = phase->type( in(MemNode::ValueIn) ); 2908 if( t == Type::TOP ) return Type::TOP; 2909 // If extra input is TOP ==> the result is TOP 2910 t = phase->type( in(MemNode::OopStore) ); 2911 if( t == Type::TOP ) return Type::TOP; 2912 2913 return StoreNode::Value( phase ); 2914 } 2915 2916 2917 //============================================================================= 2918 //----------------------------------SCMemProjNode------------------------------ 2919 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2920 { 2921 return bottom_type(); 2922 } 2923 2924 //============================================================================= 2925 //----------------------------------LoadStoreNode------------------------------ 2926 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2927 : Node(required), 2928 _type(rt), 2929 _adr_type(at) 2930 { 2931 init_req(MemNode::Control, c ); 2932 init_req(MemNode::Memory , mem); 2933 init_req(MemNode::Address, adr); 2934 init_req(MemNode::ValueIn, val); 2935 init_class_id(Class_LoadStore); 2936 } 2937 2938 uint LoadStoreNode::ideal_reg() const { 2939 return _type->ideal_reg(); 2940 } 2941 2942 bool LoadStoreNode::result_not_used() const { 2943 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2944 Node *x = fast_out(i); 2945 if (x->Opcode() == Op_SCMemProj) continue; 2946 return false; 2947 } 2948 return true; 2949 } 2950 2951 MemBarNode* LoadStoreNode::trailing_membar() const { 2952 MemBarNode* trailing = NULL; 2953 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2954 Node* u = fast_out(i); 2955 if (u->is_MemBar()) { 2956 if (u->as_MemBar()->trailing_load_store()) { 2957 assert(u->Opcode() == Op_MemBarAcquire, ""); 2958 assert(trailing == NULL, "only one"); 2959 trailing = u->as_MemBar(); 2960 #ifdef ASSERT 2961 Node* leading = trailing->leading_membar(); 2962 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2963 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 2964 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 2965 #endif 2966 } else { 2967 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 2968 } 2969 } 2970 } 2971 2972 return trailing; 2973 } 2974 2975 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2976 2977 //============================================================================= 2978 //----------------------------------LoadStoreConditionalNode-------------------- 2979 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2980 init_req(ExpectedIn, ex ); 2981 } 2982 2983 //============================================================================= 2984 //-------------------------------adr_type-------------------------------------- 2985 const TypePtr* ClearArrayNode::adr_type() const { 2986 Node *adr = in(3); 2987 if (adr == NULL) return NULL; // node is dead 2988 return MemNode::calculate_adr_type(adr->bottom_type()); 2989 } 2990 2991 //------------------------------match_edge------------------------------------- 2992 // Do we Match on this edge index or not? Do not match memory 2993 uint ClearArrayNode::match_edge(uint idx) const { 2994 return idx > 1; 2995 } 2996 2997 //------------------------------Identity--------------------------------------- 2998 // Clearing a zero length array does nothing 2999 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 3000 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 3001 } 3002 3003 //------------------------------Idealize--------------------------------------- 3004 // Clearing a short array is faster with stores 3005 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3006 // Already know this is a large node, do not try to ideal it 3007 if (!IdealizeClearArrayNode || _is_large) return NULL; 3008 3009 const int unit = BytesPerLong; 3010 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 3011 if (!t) return NULL; 3012 if (!t->is_con()) return NULL; 3013 intptr_t raw_count = t->get_con(); 3014 intptr_t size = raw_count; 3015 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 3016 // Clearing nothing uses the Identity call. 3017 // Negative clears are possible on dead ClearArrays 3018 // (see jck test stmt114.stmt11402.val). 3019 if (size <= 0 || size % unit != 0) return NULL; 3020 intptr_t count = size / unit; 3021 // Length too long; communicate this to matchers and assemblers. 3022 // Assemblers are responsible to produce fast hardware clears for it. 3023 if (size > InitArrayShortSize) { 3024 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true); 3025 } 3026 Node *mem = in(1); 3027 if( phase->type(mem)==Type::TOP ) return NULL; 3028 Node *adr = in(3); 3029 const Type* at = phase->type(adr); 3030 if( at==Type::TOP ) return NULL; 3031 const TypePtr* atp = at->isa_ptr(); 3032 // adjust atp to be the correct array element address type 3033 if (atp == NULL) atp = TypePtr::BOTTOM; 3034 else atp = atp->add_offset(Type::OffsetBot); 3035 // Get base for derived pointer purposes 3036 if( adr->Opcode() != Op_AddP ) Unimplemented(); 3037 Node *base = adr->in(1); 3038 3039 Node *val = in(4); 3040 Node *off = phase->MakeConX(BytesPerLong); 3041 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3042 count--; 3043 while( count-- ) { 3044 mem = phase->transform(mem); 3045 adr = phase->transform(new AddPNode(base,adr,off)); 3046 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3047 } 3048 return mem; 3049 } 3050 3051 //----------------------------step_through---------------------------------- 3052 // Return allocation input memory edge if it is different instance 3053 // or itself if it is the one we are looking for. 3054 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 3055 Node* n = *np; 3056 assert(n->is_ClearArray(), "sanity"); 3057 intptr_t offset; 3058 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 3059 // This method is called only before Allocate nodes are expanded 3060 // during macro nodes expansion. Before that ClearArray nodes are 3061 // only generated in PhaseMacroExpand::generate_arraycopy() (before 3062 // Allocate nodes are expanded) which follows allocations. 3063 assert(alloc != NULL, "should have allocation"); 3064 if (alloc->_idx == instance_id) { 3065 // Can not bypass initialization of the instance we are looking for. 3066 return false; 3067 } 3068 // Otherwise skip it. 3069 InitializeNode* init = alloc->initialization(); 3070 if (init != NULL) 3071 *np = init->in(TypeFunc::Memory); 3072 else 3073 *np = alloc->in(TypeFunc::Memory); 3074 return true; 3075 } 3076 3077 //----------------------------clear_memory------------------------------------- 3078 // Generate code to initialize object storage to zero. 3079 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3080 Node* val, 3081 Node* raw_val, 3082 intptr_t start_offset, 3083 Node* end_offset, 3084 PhaseGVN* phase) { 3085 intptr_t offset = start_offset; 3086 3087 int unit = BytesPerLong; 3088 if ((offset % unit) != 0) { 3089 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 3090 adr = phase->transform(adr); 3091 const TypePtr* atp = TypeRawPtr::BOTTOM; 3092 if (val != NULL) { 3093 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3094 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3095 } else { 3096 assert(raw_val == NULL, "val may not be null"); 3097 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3098 } 3099 mem = phase->transform(mem); 3100 offset += BytesPerInt; 3101 } 3102 assert((offset % unit) == 0, ""); 3103 3104 // Initialize the remaining stuff, if any, with a ClearArray. 3105 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase); 3106 } 3107 3108 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3109 Node* raw_val, 3110 Node* start_offset, 3111 Node* end_offset, 3112 PhaseGVN* phase) { 3113 if (start_offset == end_offset) { 3114 // nothing to do 3115 return mem; 3116 } 3117 3118 int unit = BytesPerLong; 3119 Node* zbase = start_offset; 3120 Node* zend = end_offset; 3121 3122 // Scale to the unit required by the CPU: 3123 if (!Matcher::init_array_count_is_in_bytes) { 3124 Node* shift = phase->intcon(exact_log2(unit)); 3125 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3126 zend = phase->transform(new URShiftXNode(zend, shift) ); 3127 } 3128 3129 // Bulk clear double-words 3130 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3131 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3132 if (raw_val == NULL) { 3133 raw_val = phase->MakeConX(0); 3134 } 3135 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false); 3136 return phase->transform(mem); 3137 } 3138 3139 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3140 Node* val, 3141 Node* raw_val, 3142 intptr_t start_offset, 3143 intptr_t end_offset, 3144 PhaseGVN* phase) { 3145 if (start_offset == end_offset) { 3146 // nothing to do 3147 return mem; 3148 } 3149 3150 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3151 intptr_t done_offset = end_offset; 3152 if ((done_offset % BytesPerLong) != 0) { 3153 done_offset -= BytesPerInt; 3154 } 3155 if (done_offset > start_offset) { 3156 mem = clear_memory(ctl, mem, dest, val, raw_val, 3157 start_offset, phase->MakeConX(done_offset), phase); 3158 } 3159 if (done_offset < end_offset) { // emit the final 32-bit store 3160 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3161 adr = phase->transform(adr); 3162 const TypePtr* atp = TypeRawPtr::BOTTOM; 3163 if (val != NULL) { 3164 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3165 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3166 } else { 3167 assert(raw_val == NULL, "val may not be null"); 3168 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3169 } 3170 mem = phase->transform(mem); 3171 done_offset += BytesPerInt; 3172 } 3173 assert(done_offset == end_offset, ""); 3174 return mem; 3175 } 3176 3177 //============================================================================= 3178 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3179 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 3180 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3181 #ifdef ASSERT 3182 , _pair_idx(0) 3183 #endif 3184 { 3185 init_class_id(Class_MemBar); 3186 Node* top = C->top(); 3187 init_req(TypeFunc::I_O,top); 3188 init_req(TypeFunc::FramePtr,top); 3189 init_req(TypeFunc::ReturnAdr,top); 3190 if (precedent != NULL) 3191 init_req(TypeFunc::Parms, precedent); 3192 } 3193 3194 //------------------------------cmp-------------------------------------------- 3195 uint MemBarNode::hash() const { return NO_HASH; } 3196 bool MemBarNode::cmp( const Node &n ) const { 3197 return (&n == this); // Always fail except on self 3198 } 3199 3200 //------------------------------make------------------------------------------- 3201 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3202 switch (opcode) { 3203 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3204 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3205 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3206 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3207 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3208 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3209 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3210 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3211 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3212 case Op_Initialize: return new InitializeNode(C, atp, pn); 3213 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3214 default: ShouldNotReachHere(); return NULL; 3215 } 3216 } 3217 3218 void MemBarNode::remove(PhaseIterGVN *igvn) { 3219 if (outcnt() != 2) { 3220 return; 3221 } 3222 if (trailing_store() || trailing_load_store()) { 3223 MemBarNode* leading = leading_membar(); 3224 if (leading != NULL) { 3225 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3226 leading->remove(igvn); 3227 } 3228 } 3229 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3230 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3231 } 3232 3233 //------------------------------Ideal------------------------------------------ 3234 // Return a node which is more "ideal" than the current node. Strip out 3235 // control copies 3236 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3237 if (remove_dead_region(phase, can_reshape)) return this; 3238 // Don't bother trying to transform a dead node 3239 if (in(0) && in(0)->is_top()) { 3240 return NULL; 3241 } 3242 3243 #if INCLUDE_ZGC 3244 if (UseZGC) { 3245 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) { 3246 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop); 3247 set_req(MemBarNode::Precedent, load_node); 3248 return this; 3249 } 3250 } 3251 #endif 3252 3253 bool progress = false; 3254 // Eliminate volatile MemBars for scalar replaced objects. 3255 if (can_reshape && req() == (Precedent+1)) { 3256 bool eliminate = false; 3257 int opc = Opcode(); 3258 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3259 // Volatile field loads and stores. 3260 Node* my_mem = in(MemBarNode::Precedent); 3261 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3262 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3263 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3264 // replace this Precedent (decodeN) with the Load instead. 3265 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3266 Node* load_node = my_mem->in(1); 3267 set_req(MemBarNode::Precedent, load_node); 3268 phase->is_IterGVN()->_worklist.push(my_mem); 3269 my_mem = load_node; 3270 } else { 3271 assert(my_mem->unique_out() == this, "sanity"); 3272 del_req(Precedent); 3273 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3274 my_mem = NULL; 3275 } 3276 progress = true; 3277 } 3278 if (my_mem != NULL && my_mem->is_Mem()) { 3279 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3280 // Check for scalar replaced object reference. 3281 if( t_oop != NULL && t_oop->is_known_instance_field() && 3282 t_oop->offset() != Type::OffsetBot && 3283 t_oop->offset() != Type::OffsetTop) { 3284 eliminate = true; 3285 } 3286 } 3287 } else if (opc == Op_MemBarRelease) { 3288 // Final field stores. 3289 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3290 if ((alloc != NULL) && alloc->is_Allocate() && 3291 alloc->as_Allocate()->does_not_escape_thread()) { 3292 // The allocated object does not escape. 3293 eliminate = true; 3294 } 3295 } 3296 if (eliminate) { 3297 // Replace MemBar projections by its inputs. 3298 PhaseIterGVN* igvn = phase->is_IterGVN(); 3299 remove(igvn); 3300 // Must return either the original node (now dead) or a new node 3301 // (Do not return a top here, since that would break the uniqueness of top.) 3302 return new ConINode(TypeInt::ZERO); 3303 } 3304 } 3305 return progress ? this : NULL; 3306 } 3307 3308 //------------------------------Value------------------------------------------ 3309 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3310 if( !in(0) ) return Type::TOP; 3311 if( phase->type(in(0)) == Type::TOP ) 3312 return Type::TOP; 3313 return TypeTuple::MEMBAR; 3314 } 3315 3316 //------------------------------match------------------------------------------ 3317 // Construct projections for memory. 3318 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) { 3319 switch (proj->_con) { 3320 case TypeFunc::Control: 3321 case TypeFunc::Memory: 3322 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3323 } 3324 ShouldNotReachHere(); 3325 return NULL; 3326 } 3327 3328 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3329 trailing->_kind = TrailingStore; 3330 leading->_kind = LeadingStore; 3331 #ifdef ASSERT 3332 trailing->_pair_idx = leading->_idx; 3333 leading->_pair_idx = leading->_idx; 3334 #endif 3335 } 3336 3337 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3338 trailing->_kind = TrailingLoadStore; 3339 leading->_kind = LeadingLoadStore; 3340 #ifdef ASSERT 3341 trailing->_pair_idx = leading->_idx; 3342 leading->_pair_idx = leading->_idx; 3343 #endif 3344 } 3345 3346 MemBarNode* MemBarNode::trailing_membar() const { 3347 ResourceMark rm; 3348 Node* trailing = (Node*)this; 3349 VectorSet seen(Thread::current()->resource_area()); 3350 Node_Stack multis(0); 3351 do { 3352 Node* c = trailing; 3353 uint i = 0; 3354 do { 3355 trailing = NULL; 3356 for (; i < c->outcnt(); i++) { 3357 Node* next = c->raw_out(i); 3358 if (next != c && next->is_CFG()) { 3359 if (c->is_MultiBranch()) { 3360 if (multis.node() == c) { 3361 multis.set_index(i+1); 3362 } else { 3363 multis.push(c, i+1); 3364 } 3365 } 3366 trailing = next; 3367 break; 3368 } 3369 } 3370 if (trailing != NULL && !seen.test_set(trailing->_idx)) { 3371 break; 3372 } 3373 while (multis.size() > 0) { 3374 c = multis.node(); 3375 i = multis.index(); 3376 if (i < c->req()) { 3377 break; 3378 } 3379 multis.pop(); 3380 } 3381 } while (multis.size() > 0); 3382 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3383 3384 MemBarNode* mb = trailing->as_MemBar(); 3385 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3386 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3387 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3388 return mb; 3389 } 3390 3391 MemBarNode* MemBarNode::leading_membar() const { 3392 ResourceMark rm; 3393 VectorSet seen(Thread::current()->resource_area()); 3394 Node_Stack regions(0); 3395 Node* leading = in(0); 3396 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3397 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) { 3398 leading = NULL; 3399 while (regions.size() > 0 && leading == NULL) { 3400 Node* r = regions.node(); 3401 uint i = regions.index(); 3402 if (i < r->req()) { 3403 leading = r->in(i); 3404 regions.set_index(i+1); 3405 } else { 3406 regions.pop(); 3407 } 3408 } 3409 if (leading == NULL) { 3410 assert(regions.size() == 0, "all paths should have been tried"); 3411 return NULL; 3412 } 3413 } 3414 if (leading->is_Region()) { 3415 regions.push(leading, 2); 3416 leading = leading->in(1); 3417 } else { 3418 leading = leading->in(0); 3419 } 3420 } 3421 #ifdef ASSERT 3422 Unique_Node_List wq; 3423 wq.push((Node*)this); 3424 uint found = 0; 3425 for (uint i = 0; i < wq.size(); i++) { 3426 Node* n = wq.at(i); 3427 if (n->is_Region()) { 3428 for (uint j = 1; j < n->req(); j++) { 3429 Node* in = n->in(j); 3430 if (in != NULL && !in->is_top()) { 3431 wq.push(in); 3432 } 3433 } 3434 } else { 3435 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3436 assert(n == leading, "consistency check failed"); 3437 found++; 3438 } else { 3439 Node* in = n->in(0); 3440 if (in != NULL && !in->is_top()) { 3441 wq.push(in); 3442 } 3443 } 3444 } 3445 } 3446 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed"); 3447 #endif 3448 if (leading == NULL) { 3449 return NULL; 3450 } 3451 MemBarNode* mb = leading->as_MemBar(); 3452 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3453 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3454 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3455 return mb; 3456 } 3457 3458 //===========================InitializeNode==================================== 3459 // SUMMARY: 3460 // This node acts as a memory barrier on raw memory, after some raw stores. 3461 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3462 // The Initialize can 'capture' suitably constrained stores as raw inits. 3463 // It can coalesce related raw stores into larger units (called 'tiles'). 3464 // It can avoid zeroing new storage for memory units which have raw inits. 3465 // At macro-expansion, it is marked 'complete', and does not optimize further. 3466 // 3467 // EXAMPLE: 3468 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3469 // ctl = incoming control; mem* = incoming memory 3470 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3471 // First allocate uninitialized memory and fill in the header: 3472 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3473 // ctl := alloc.Control; mem* := alloc.Memory* 3474 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3475 // Then initialize to zero the non-header parts of the raw memory block: 3476 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3477 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3478 // After the initialize node executes, the object is ready for service: 3479 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3480 // Suppose its body is immediately initialized as {1,2}: 3481 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3482 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3483 // mem.SLICE(#short[*]) := store2 3484 // 3485 // DETAILS: 3486 // An InitializeNode collects and isolates object initialization after 3487 // an AllocateNode and before the next possible safepoint. As a 3488 // memory barrier (MemBarNode), it keeps critical stores from drifting 3489 // down past any safepoint or any publication of the allocation. 3490 // Before this barrier, a newly-allocated object may have uninitialized bits. 3491 // After this barrier, it may be treated as a real oop, and GC is allowed. 3492 // 3493 // The semantics of the InitializeNode include an implicit zeroing of 3494 // the new object from object header to the end of the object. 3495 // (The object header and end are determined by the AllocateNode.) 3496 // 3497 // Certain stores may be added as direct inputs to the InitializeNode. 3498 // These stores must update raw memory, and they must be to addresses 3499 // derived from the raw address produced by AllocateNode, and with 3500 // a constant offset. They must be ordered by increasing offset. 3501 // The first one is at in(RawStores), the last at in(req()-1). 3502 // Unlike most memory operations, they are not linked in a chain, 3503 // but are displayed in parallel as users of the rawmem output of 3504 // the allocation. 3505 // 3506 // (See comments in InitializeNode::capture_store, which continue 3507 // the example given above.) 3508 // 3509 // When the associated Allocate is macro-expanded, the InitializeNode 3510 // may be rewritten to optimize collected stores. A ClearArrayNode 3511 // may also be created at that point to represent any required zeroing. 3512 // The InitializeNode is then marked 'complete', prohibiting further 3513 // capturing of nearby memory operations. 3514 // 3515 // During macro-expansion, all captured initializations which store 3516 // constant values of 32 bits or smaller are coalesced (if advantageous) 3517 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3518 // initialized in fewer memory operations. Memory words which are 3519 // covered by neither tiles nor non-constant stores are pre-zeroed 3520 // by explicit stores of zero. (The code shape happens to do all 3521 // zeroing first, then all other stores, with both sequences occurring 3522 // in order of ascending offsets.) 3523 // 3524 // Alternatively, code may be inserted between an AllocateNode and its 3525 // InitializeNode, to perform arbitrary initialization of the new object. 3526 // E.g., the object copying intrinsics insert complex data transfers here. 3527 // The initialization must then be marked as 'complete' disable the 3528 // built-in zeroing semantics and the collection of initializing stores. 3529 // 3530 // While an InitializeNode is incomplete, reads from the memory state 3531 // produced by it are optimizable if they match the control edge and 3532 // new oop address associated with the allocation/initialization. 3533 // They return a stored value (if the offset matches) or else zero. 3534 // A write to the memory state, if it matches control and address, 3535 // and if it is to a constant offset, may be 'captured' by the 3536 // InitializeNode. It is cloned as a raw memory operation and rewired 3537 // inside the initialization, to the raw oop produced by the allocation. 3538 // Operations on addresses which are provably distinct (e.g., to 3539 // other AllocateNodes) are allowed to bypass the initialization. 3540 // 3541 // The effect of all this is to consolidate object initialization 3542 // (both arrays and non-arrays, both piecewise and bulk) into a 3543 // single location, where it can be optimized as a unit. 3544 // 3545 // Only stores with an offset less than TrackedInitializationLimit words 3546 // will be considered for capture by an InitializeNode. This puts a 3547 // reasonable limit on the complexity of optimized initializations. 3548 3549 //---------------------------InitializeNode------------------------------------ 3550 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3551 : MemBarNode(C, adr_type, rawoop), 3552 _is_complete(Incomplete), _does_not_escape(false) 3553 { 3554 init_class_id(Class_Initialize); 3555 3556 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3557 assert(in(RawAddress) == rawoop, "proper init"); 3558 // Note: allocation() can be NULL, for secondary initialization barriers 3559 } 3560 3561 // Since this node is not matched, it will be processed by the 3562 // register allocator. Declare that there are no constraints 3563 // on the allocation of the RawAddress edge. 3564 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3565 // This edge should be set to top, by the set_complete. But be conservative. 3566 if (idx == InitializeNode::RawAddress) 3567 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3568 return RegMask::Empty; 3569 } 3570 3571 Node* InitializeNode::memory(uint alias_idx) { 3572 Node* mem = in(Memory); 3573 if (mem->is_MergeMem()) { 3574 return mem->as_MergeMem()->memory_at(alias_idx); 3575 } else { 3576 // incoming raw memory is not split 3577 return mem; 3578 } 3579 } 3580 3581 bool InitializeNode::is_non_zero() { 3582 if (is_complete()) return false; 3583 remove_extra_zeroes(); 3584 return (req() > RawStores); 3585 } 3586 3587 void InitializeNode::set_complete(PhaseGVN* phase) { 3588 assert(!is_complete(), "caller responsibility"); 3589 _is_complete = Complete; 3590 3591 // After this node is complete, it contains a bunch of 3592 // raw-memory initializations. There is no need for 3593 // it to have anything to do with non-raw memory effects. 3594 // Therefore, tell all non-raw users to re-optimize themselves, 3595 // after skipping the memory effects of this initialization. 3596 PhaseIterGVN* igvn = phase->is_IterGVN(); 3597 if (igvn) igvn->add_users_to_worklist(this); 3598 } 3599 3600 // convenience function 3601 // return false if the init contains any stores already 3602 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3603 InitializeNode* init = initialization(); 3604 if (init == NULL || init->is_complete()) { 3605 return false; 3606 } 3607 init->remove_extra_zeroes(); 3608 // for now, if this allocation has already collected any inits, bail: 3609 if (init->is_non_zero()) return false; 3610 init->set_complete(phase); 3611 return true; 3612 } 3613 3614 void InitializeNode::remove_extra_zeroes() { 3615 if (req() == RawStores) return; 3616 Node* zmem = zero_memory(); 3617 uint fill = RawStores; 3618 for (uint i = fill; i < req(); i++) { 3619 Node* n = in(i); 3620 if (n->is_top() || n == zmem) continue; // skip 3621 if (fill < i) set_req(fill, n); // compact 3622 ++fill; 3623 } 3624 // delete any empty spaces created: 3625 while (fill < req()) { 3626 del_req(fill); 3627 } 3628 } 3629 3630 // Helper for remembering which stores go with which offsets. 3631 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3632 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3633 intptr_t offset = -1; 3634 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3635 phase, offset); 3636 if (base == NULL) return -1; // something is dead, 3637 if (offset < 0) return -1; // dead, dead 3638 return offset; 3639 } 3640 3641 // Helper for proving that an initialization expression is 3642 // "simple enough" to be folded into an object initialization. 3643 // Attempts to prove that a store's initial value 'n' can be captured 3644 // within the initialization without creating a vicious cycle, such as: 3645 // { Foo p = new Foo(); p.next = p; } 3646 // True for constants and parameters and small combinations thereof. 3647 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3648 if (n == NULL) return true; // (can this really happen?) 3649 if (n->is_Proj()) n = n->in(0); 3650 if (n == this) return false; // found a cycle 3651 if (n->is_Con()) return true; 3652 if (n->is_Start()) return true; // params, etc., are OK 3653 if (n->is_Root()) return true; // even better 3654 3655 Node* ctl = n->in(0); 3656 if (ctl != NULL && !ctl->is_top()) { 3657 if (ctl->is_Proj()) ctl = ctl->in(0); 3658 if (ctl == this) return false; 3659 3660 // If we already know that the enclosing memory op is pinned right after 3661 // the init, then any control flow that the store has picked up 3662 // must have preceded the init, or else be equal to the init. 3663 // Even after loop optimizations (which might change control edges) 3664 // a store is never pinned *before* the availability of its inputs. 3665 if (!MemNode::all_controls_dominate(n, this)) 3666 return false; // failed to prove a good control 3667 } 3668 3669 // Check data edges for possible dependencies on 'this'. 3670 if ((count += 1) > 20) return false; // complexity limit 3671 for (uint i = 1; i < n->req(); i++) { 3672 Node* m = n->in(i); 3673 if (m == NULL || m == n || m->is_top()) continue; 3674 uint first_i = n->find_edge(m); 3675 if (i != first_i) continue; // process duplicate edge just once 3676 if (!detect_init_independence(m, count)) { 3677 return false; 3678 } 3679 } 3680 3681 return true; 3682 } 3683 3684 // Here are all the checks a Store must pass before it can be moved into 3685 // an initialization. Returns zero if a check fails. 3686 // On success, returns the (constant) offset to which the store applies, 3687 // within the initialized memory. 3688 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3689 const int FAIL = 0; 3690 if (st->req() != MemNode::ValueIn + 1) 3691 return FAIL; // an inscrutable StoreNode (card mark?) 3692 Node* ctl = st->in(MemNode::Control); 3693 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3694 return FAIL; // must be unconditional after the initialization 3695 Node* mem = st->in(MemNode::Memory); 3696 if (!(mem->is_Proj() && mem->in(0) == this)) 3697 return FAIL; // must not be preceded by other stores 3698 Node* adr = st->in(MemNode::Address); 3699 intptr_t offset; 3700 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3701 if (alloc == NULL) 3702 return FAIL; // inscrutable address 3703 if (alloc != allocation()) 3704 return FAIL; // wrong allocation! (store needs to float up) 3705 int size_in_bytes = st->memory_size(); 3706 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) { 3707 return FAIL; // mismatched access 3708 } 3709 Node* val = st->in(MemNode::ValueIn); 3710 int complexity_count = 0; 3711 if (!detect_init_independence(val, complexity_count)) 3712 return FAIL; // stored value must be 'simple enough' 3713 3714 // The Store can be captured only if nothing after the allocation 3715 // and before the Store is using the memory location that the store 3716 // overwrites. 3717 bool failed = false; 3718 // If is_complete_with_arraycopy() is true the shape of the graph is 3719 // well defined and is safe so no need for extra checks. 3720 if (!is_complete_with_arraycopy()) { 3721 // We are going to look at each use of the memory state following 3722 // the allocation to make sure nothing reads the memory that the 3723 // Store writes. 3724 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3725 int alias_idx = phase->C->get_alias_index(t_adr); 3726 ResourceMark rm; 3727 Unique_Node_List mems; 3728 mems.push(mem); 3729 Node* unique_merge = NULL; 3730 for (uint next = 0; next < mems.size(); ++next) { 3731 Node *m = mems.at(next); 3732 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3733 Node *n = m->fast_out(j); 3734 if (n->outcnt() == 0) { 3735 continue; 3736 } 3737 if (n == st) { 3738 continue; 3739 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3740 // If the control of this use is different from the control 3741 // of the Store which is right after the InitializeNode then 3742 // this node cannot be between the InitializeNode and the 3743 // Store. 3744 continue; 3745 } else if (n->is_MergeMem()) { 3746 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3747 // We can hit a MergeMemNode (that will likely go away 3748 // later) that is a direct use of the memory state 3749 // following the InitializeNode on the same slice as the 3750 // store node that we'd like to capture. We need to check 3751 // the uses of the MergeMemNode. 3752 mems.push(n); 3753 } 3754 } else if (n->is_Mem()) { 3755 Node* other_adr = n->in(MemNode::Address); 3756 if (other_adr == adr) { 3757 failed = true; 3758 break; 3759 } else { 3760 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3761 if (other_t_adr != NULL) { 3762 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3763 if (other_alias_idx == alias_idx) { 3764 // A load from the same memory slice as the store right 3765 // after the InitializeNode. We check the control of the 3766 // object/array that is loaded from. If it's the same as 3767 // the store control then we cannot capture the store. 3768 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3769 Node* base = other_adr; 3770 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3771 base = base->in(AddPNode::Base); 3772 if (base != NULL) { 3773 base = base->uncast(); 3774 if (base->is_Proj() && base->in(0) == alloc) { 3775 failed = true; 3776 break; 3777 } 3778 } 3779 } 3780 } 3781 } 3782 } else { 3783 failed = true; 3784 break; 3785 } 3786 } 3787 } 3788 } 3789 if (failed) { 3790 if (!can_reshape) { 3791 // We decided we couldn't capture the store during parsing. We 3792 // should try again during the next IGVN once the graph is 3793 // cleaner. 3794 phase->C->record_for_igvn(st); 3795 } 3796 return FAIL; 3797 } 3798 3799 return offset; // success 3800 } 3801 3802 // Find the captured store in(i) which corresponds to the range 3803 // [start..start+size) in the initialized object. 3804 // If there is one, return its index i. If there isn't, return the 3805 // negative of the index where it should be inserted. 3806 // Return 0 if the queried range overlaps an initialization boundary 3807 // or if dead code is encountered. 3808 // If size_in_bytes is zero, do not bother with overlap checks. 3809 int InitializeNode::captured_store_insertion_point(intptr_t start, 3810 int size_in_bytes, 3811 PhaseTransform* phase) { 3812 const int FAIL = 0, MAX_STORE = BytesPerLong; 3813 3814 if (is_complete()) 3815 return FAIL; // arraycopy got here first; punt 3816 3817 assert(allocation() != NULL, "must be present"); 3818 3819 // no negatives, no header fields: 3820 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3821 3822 // after a certain size, we bail out on tracking all the stores: 3823 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3824 if (start >= ti_limit) return FAIL; 3825 3826 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3827 if (i >= limit) return -(int)i; // not found; here is where to put it 3828 3829 Node* st = in(i); 3830 intptr_t st_off = get_store_offset(st, phase); 3831 if (st_off < 0) { 3832 if (st != zero_memory()) { 3833 return FAIL; // bail out if there is dead garbage 3834 } 3835 } else if (st_off > start) { 3836 // ...we are done, since stores are ordered 3837 if (st_off < start + size_in_bytes) { 3838 return FAIL; // the next store overlaps 3839 } 3840 return -(int)i; // not found; here is where to put it 3841 } else if (st_off < start) { 3842 if (size_in_bytes != 0 && 3843 start < st_off + MAX_STORE && 3844 start < st_off + st->as_Store()->memory_size()) { 3845 return FAIL; // the previous store overlaps 3846 } 3847 } else { 3848 if (size_in_bytes != 0 && 3849 st->as_Store()->memory_size() != size_in_bytes) { 3850 return FAIL; // mismatched store size 3851 } 3852 return i; 3853 } 3854 3855 ++i; 3856 } 3857 } 3858 3859 // Look for a captured store which initializes at the offset 'start' 3860 // with the given size. If there is no such store, and no other 3861 // initialization interferes, then return zero_memory (the memory 3862 // projection of the AllocateNode). 3863 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3864 PhaseTransform* phase) { 3865 assert(stores_are_sane(phase), ""); 3866 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3867 if (i == 0) { 3868 return NULL; // something is dead 3869 } else if (i < 0) { 3870 return zero_memory(); // just primordial zero bits here 3871 } else { 3872 Node* st = in(i); // here is the store at this position 3873 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3874 return st; 3875 } 3876 } 3877 3878 // Create, as a raw pointer, an address within my new object at 'offset'. 3879 Node* InitializeNode::make_raw_address(intptr_t offset, 3880 PhaseTransform* phase) { 3881 Node* addr = in(RawAddress); 3882 if (offset != 0) { 3883 Compile* C = phase->C; 3884 addr = phase->transform( new AddPNode(C->top(), addr, 3885 phase->MakeConX(offset)) ); 3886 } 3887 return addr; 3888 } 3889 3890 // Clone the given store, converting it into a raw store 3891 // initializing a field or element of my new object. 3892 // Caller is responsible for retiring the original store, 3893 // with subsume_node or the like. 3894 // 3895 // From the example above InitializeNode::InitializeNode, 3896 // here are the old stores to be captured: 3897 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3898 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3899 // 3900 // Here is the changed code; note the extra edges on init: 3901 // alloc = (Allocate ...) 3902 // rawoop = alloc.RawAddress 3903 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3904 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3905 // init = (Initialize alloc.Control alloc.Memory rawoop 3906 // rawstore1 rawstore2) 3907 // 3908 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3909 PhaseTransform* phase, bool can_reshape) { 3910 assert(stores_are_sane(phase), ""); 3911 3912 if (start < 0) return NULL; 3913 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3914 3915 Compile* C = phase->C; 3916 int size_in_bytes = st->memory_size(); 3917 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3918 if (i == 0) return NULL; // bail out 3919 Node* prev_mem = NULL; // raw memory for the captured store 3920 if (i > 0) { 3921 prev_mem = in(i); // there is a pre-existing store under this one 3922 set_req(i, C->top()); // temporarily disconnect it 3923 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3924 } else { 3925 i = -i; // no pre-existing store 3926 prev_mem = zero_memory(); // a slice of the newly allocated object 3927 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3928 set_req(--i, C->top()); // reuse this edge; it has been folded away 3929 else 3930 ins_req(i, C->top()); // build a new edge 3931 } 3932 Node* new_st = st->clone(); 3933 new_st->set_req(MemNode::Control, in(Control)); 3934 new_st->set_req(MemNode::Memory, prev_mem); 3935 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3936 new_st = phase->transform(new_st); 3937 3938 // At this point, new_st might have swallowed a pre-existing store 3939 // at the same offset, or perhaps new_st might have disappeared, 3940 // if it redundantly stored the same value (or zero to fresh memory). 3941 3942 // In any case, wire it in: 3943 phase->igvn_rehash_node_delayed(this); 3944 set_req(i, new_st); 3945 3946 // The caller may now kill the old guy. 3947 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3948 assert(check_st == new_st || check_st == NULL, "must be findable"); 3949 assert(!is_complete(), ""); 3950 return new_st; 3951 } 3952 3953 static bool store_constant(jlong* tiles, int num_tiles, 3954 intptr_t st_off, int st_size, 3955 jlong con) { 3956 if ((st_off & (st_size-1)) != 0) 3957 return false; // strange store offset (assume size==2**N) 3958 address addr = (address)tiles + st_off; 3959 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3960 switch (st_size) { 3961 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3962 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3963 case sizeof(jint): *(jint*) addr = (jint) con; break; 3964 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3965 default: return false; // strange store size (detect size!=2**N here) 3966 } 3967 return true; // return success to caller 3968 } 3969 3970 // Coalesce subword constants into int constants and possibly 3971 // into long constants. The goal, if the CPU permits, 3972 // is to initialize the object with a small number of 64-bit tiles. 3973 // Also, convert floating-point constants to bit patterns. 3974 // Non-constants are not relevant to this pass. 3975 // 3976 // In terms of the running example on InitializeNode::InitializeNode 3977 // and InitializeNode::capture_store, here is the transformation 3978 // of rawstore1 and rawstore2 into rawstore12: 3979 // alloc = (Allocate ...) 3980 // rawoop = alloc.RawAddress 3981 // tile12 = 0x00010002 3982 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3983 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3984 // 3985 void 3986 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3987 Node* size_in_bytes, 3988 PhaseGVN* phase) { 3989 Compile* C = phase->C; 3990 3991 assert(stores_are_sane(phase), ""); 3992 // Note: After this pass, they are not completely sane, 3993 // since there may be some overlaps. 3994 3995 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3996 3997 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3998 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3999 size_limit = MIN2(size_limit, ti_limit); 4000 size_limit = align_up(size_limit, BytesPerLong); 4001 int num_tiles = size_limit / BytesPerLong; 4002 4003 // allocate space for the tile map: 4004 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 4005 jlong tiles_buf[small_len]; 4006 Node* nodes_buf[small_len]; 4007 jlong inits_buf[small_len]; 4008 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 4009 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4010 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 4011 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 4012 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 4013 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4014 // tiles: exact bitwise model of all primitive constants 4015 // nodes: last constant-storing node subsumed into the tiles model 4016 // inits: which bytes (in each tile) are touched by any initializations 4017 4018 //// Pass A: Fill in the tile model with any relevant stores. 4019 4020 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 4021 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 4022 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 4023 Node* zmem = zero_memory(); // initially zero memory state 4024 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4025 Node* st = in(i); 4026 intptr_t st_off = get_store_offset(st, phase); 4027 4028 // Figure out the store's offset and constant value: 4029 if (st_off < header_size) continue; //skip (ignore header) 4030 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 4031 int st_size = st->as_Store()->memory_size(); 4032 if (st_off + st_size > size_limit) break; 4033 4034 // Record which bytes are touched, whether by constant or not. 4035 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 4036 continue; // skip (strange store size) 4037 4038 const Type* val = phase->type(st->in(MemNode::ValueIn)); 4039 if (!val->singleton()) continue; //skip (non-con store) 4040 BasicType type = val->basic_type(); 4041 4042 jlong con = 0; 4043 switch (type) { 4044 case T_INT: con = val->is_int()->get_con(); break; 4045 case T_LONG: con = val->is_long()->get_con(); break; 4046 case T_FLOAT: con = jint_cast(val->getf()); break; 4047 case T_DOUBLE: con = jlong_cast(val->getd()); break; 4048 default: continue; //skip (odd store type) 4049 } 4050 4051 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 4052 st->Opcode() == Op_StoreL) { 4053 continue; // This StoreL is already optimal. 4054 } 4055 4056 // Store down the constant. 4057 store_constant(tiles, num_tiles, st_off, st_size, con); 4058 4059 intptr_t j = st_off >> LogBytesPerLong; 4060 4061 if (type == T_INT && st_size == BytesPerInt 4062 && (st_off & BytesPerInt) == BytesPerInt) { 4063 jlong lcon = tiles[j]; 4064 if (!Matcher::isSimpleConstant64(lcon) && 4065 st->Opcode() == Op_StoreI) { 4066 // This StoreI is already optimal by itself. 4067 jint* intcon = (jint*) &tiles[j]; 4068 intcon[1] = 0; // undo the store_constant() 4069 4070 // If the previous store is also optimal by itself, back up and 4071 // undo the action of the previous loop iteration... if we can. 4072 // But if we can't, just let the previous half take care of itself. 4073 st = nodes[j]; 4074 st_off -= BytesPerInt; 4075 con = intcon[0]; 4076 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 4077 assert(st_off >= header_size, "still ignoring header"); 4078 assert(get_store_offset(st, phase) == st_off, "must be"); 4079 assert(in(i-1) == zmem, "must be"); 4080 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 4081 assert(con == tcon->is_int()->get_con(), "must be"); 4082 // Undo the effects of the previous loop trip, which swallowed st: 4083 intcon[0] = 0; // undo store_constant() 4084 set_req(i-1, st); // undo set_req(i, zmem) 4085 nodes[j] = NULL; // undo nodes[j] = st 4086 --old_subword; // undo ++old_subword 4087 } 4088 continue; // This StoreI is already optimal. 4089 } 4090 } 4091 4092 // This store is not needed. 4093 set_req(i, zmem); 4094 nodes[j] = st; // record for the moment 4095 if (st_size < BytesPerLong) // something has changed 4096 ++old_subword; // includes int/float, but who's counting... 4097 else ++old_long; 4098 } 4099 4100 if ((old_subword + old_long) == 0) 4101 return; // nothing more to do 4102 4103 //// Pass B: Convert any non-zero tiles into optimal constant stores. 4104 // Be sure to insert them before overlapping non-constant stores. 4105 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 4106 for (int j = 0; j < num_tiles; j++) { 4107 jlong con = tiles[j]; 4108 jlong init = inits[j]; 4109 if (con == 0) continue; 4110 jint con0, con1; // split the constant, address-wise 4111 jint init0, init1; // split the init map, address-wise 4112 { union { jlong con; jint intcon[2]; } u; 4113 u.con = con; 4114 con0 = u.intcon[0]; 4115 con1 = u.intcon[1]; 4116 u.con = init; 4117 init0 = u.intcon[0]; 4118 init1 = u.intcon[1]; 4119 } 4120 4121 Node* old = nodes[j]; 4122 assert(old != NULL, "need the prior store"); 4123 intptr_t offset = (j * BytesPerLong); 4124 4125 bool split = !Matcher::isSimpleConstant64(con); 4126 4127 if (offset < header_size) { 4128 assert(offset + BytesPerInt >= header_size, "second int counts"); 4129 assert(*(jint*)&tiles[j] == 0, "junk in header"); 4130 split = true; // only the second word counts 4131 // Example: int a[] = { 42 ... } 4132 } else if (con0 == 0 && init0 == -1) { 4133 split = true; // first word is covered by full inits 4134 // Example: int a[] = { ... foo(), 42 ... } 4135 } else if (con1 == 0 && init1 == -1) { 4136 split = true; // second word is covered by full inits 4137 // Example: int a[] = { ... 42, foo() ... } 4138 } 4139 4140 // Here's a case where init0 is neither 0 nor -1: 4141 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4142 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4143 // In this case the tile is not split; it is (jlong)42. 4144 // The big tile is stored down, and then the foo() value is inserted. 4145 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4146 4147 Node* ctl = old->in(MemNode::Control); 4148 Node* adr = make_raw_address(offset, phase); 4149 const TypePtr* atp = TypeRawPtr::BOTTOM; 4150 4151 // One or two coalesced stores to plop down. 4152 Node* st[2]; 4153 intptr_t off[2]; 4154 int nst = 0; 4155 if (!split) { 4156 ++new_long; 4157 off[nst] = offset; 4158 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4159 phase->longcon(con), T_LONG, MemNode::unordered); 4160 } else { 4161 // Omit either if it is a zero. 4162 if (con0 != 0) { 4163 ++new_int; 4164 off[nst] = offset; 4165 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4166 phase->intcon(con0), T_INT, MemNode::unordered); 4167 } 4168 if (con1 != 0) { 4169 ++new_int; 4170 offset += BytesPerInt; 4171 adr = make_raw_address(offset, phase); 4172 off[nst] = offset; 4173 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4174 phase->intcon(con1), T_INT, MemNode::unordered); 4175 } 4176 } 4177 4178 // Insert second store first, then the first before the second. 4179 // Insert each one just before any overlapping non-constant stores. 4180 while (nst > 0) { 4181 Node* st1 = st[--nst]; 4182 C->copy_node_notes_to(st1, old); 4183 st1 = phase->transform(st1); 4184 offset = off[nst]; 4185 assert(offset >= header_size, "do not smash header"); 4186 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4187 guarantee(ins_idx != 0, "must re-insert constant store"); 4188 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4189 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4190 set_req(--ins_idx, st1); 4191 else 4192 ins_req(ins_idx, st1); 4193 } 4194 } 4195 4196 if (PrintCompilation && WizardMode) 4197 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4198 old_subword, old_long, new_int, new_long); 4199 if (C->log() != NULL) 4200 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4201 old_subword, old_long, new_int, new_long); 4202 4203 // Clean up any remaining occurrences of zmem: 4204 remove_extra_zeroes(); 4205 } 4206 4207 // Explore forward from in(start) to find the first fully initialized 4208 // word, and return its offset. Skip groups of subword stores which 4209 // together initialize full words. If in(start) is itself part of a 4210 // fully initialized word, return the offset of in(start). If there 4211 // are no following full-word stores, or if something is fishy, return 4212 // a negative value. 4213 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4214 int int_map = 0; 4215 intptr_t int_map_off = 0; 4216 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4217 4218 for (uint i = start, limit = req(); i < limit; i++) { 4219 Node* st = in(i); 4220 4221 intptr_t st_off = get_store_offset(st, phase); 4222 if (st_off < 0) break; // return conservative answer 4223 4224 int st_size = st->as_Store()->memory_size(); 4225 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4226 return st_off; // we found a complete word init 4227 } 4228 4229 // update the map: 4230 4231 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4232 if (this_int_off != int_map_off) { 4233 // reset the map: 4234 int_map = 0; 4235 int_map_off = this_int_off; 4236 } 4237 4238 int subword_off = st_off - this_int_off; 4239 int_map |= right_n_bits(st_size) << subword_off; 4240 if ((int_map & FULL_MAP) == FULL_MAP) { 4241 return this_int_off; // we found a complete word init 4242 } 4243 4244 // Did this store hit or cross the word boundary? 4245 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4246 if (next_int_off == this_int_off + BytesPerInt) { 4247 // We passed the current int, without fully initializing it. 4248 int_map_off = next_int_off; 4249 int_map >>= BytesPerInt; 4250 } else if (next_int_off > this_int_off + BytesPerInt) { 4251 // We passed the current and next int. 4252 return this_int_off + BytesPerInt; 4253 } 4254 } 4255 4256 return -1; 4257 } 4258 4259 4260 // Called when the associated AllocateNode is expanded into CFG. 4261 // At this point, we may perform additional optimizations. 4262 // Linearize the stores by ascending offset, to make memory 4263 // activity as coherent as possible. 4264 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4265 intptr_t header_size, 4266 Node* size_in_bytes, 4267 PhaseGVN* phase) { 4268 assert(!is_complete(), "not already complete"); 4269 assert(stores_are_sane(phase), ""); 4270 assert(allocation() != NULL, "must be present"); 4271 4272 remove_extra_zeroes(); 4273 4274 if (ReduceFieldZeroing || ReduceBulkZeroing) 4275 // reduce instruction count for common initialization patterns 4276 coalesce_subword_stores(header_size, size_in_bytes, phase); 4277 4278 Node* zmem = zero_memory(); // initially zero memory state 4279 Node* inits = zmem; // accumulating a linearized chain of inits 4280 #ifdef ASSERT 4281 intptr_t first_offset = allocation()->minimum_header_size(); 4282 intptr_t last_init_off = first_offset; // previous init offset 4283 intptr_t last_init_end = first_offset; // previous init offset+size 4284 intptr_t last_tile_end = first_offset; // previous tile offset+size 4285 #endif 4286 intptr_t zeroes_done = header_size; 4287 4288 bool do_zeroing = true; // we might give up if inits are very sparse 4289 int big_init_gaps = 0; // how many large gaps have we seen? 4290 4291 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4292 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4293 4294 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4295 Node* st = in(i); 4296 intptr_t st_off = get_store_offset(st, phase); 4297 if (st_off < 0) 4298 break; // unknown junk in the inits 4299 if (st->in(MemNode::Memory) != zmem) 4300 break; // complicated store chains somehow in list 4301 4302 int st_size = st->as_Store()->memory_size(); 4303 intptr_t next_init_off = st_off + st_size; 4304 4305 if (do_zeroing && zeroes_done < next_init_off) { 4306 // See if this store needs a zero before it or under it. 4307 intptr_t zeroes_needed = st_off; 4308 4309 if (st_size < BytesPerInt) { 4310 // Look for subword stores which only partially initialize words. 4311 // If we find some, we must lay down some word-level zeroes first, 4312 // underneath the subword stores. 4313 // 4314 // Examples: 4315 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4316 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4317 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4318 // 4319 // Note: coalesce_subword_stores may have already done this, 4320 // if it was prompted by constant non-zero subword initializers. 4321 // But this case can still arise with non-constant stores. 4322 4323 intptr_t next_full_store = find_next_fullword_store(i, phase); 4324 4325 // In the examples above: 4326 // in(i) p q r s x y z 4327 // st_off 12 13 14 15 12 13 14 4328 // st_size 1 1 1 1 1 1 1 4329 // next_full_s. 12 16 16 16 16 16 16 4330 // z's_done 12 16 16 16 12 16 12 4331 // z's_needed 12 16 16 16 16 16 16 4332 // zsize 0 0 0 0 4 0 4 4333 if (next_full_store < 0) { 4334 // Conservative tack: Zero to end of current word. 4335 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4336 } else { 4337 // Zero to beginning of next fully initialized word. 4338 // Or, don't zero at all, if we are already in that word. 4339 assert(next_full_store >= zeroes_needed, "must go forward"); 4340 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4341 zeroes_needed = next_full_store; 4342 } 4343 } 4344 4345 if (zeroes_needed > zeroes_done) { 4346 intptr_t zsize = zeroes_needed - zeroes_done; 4347 // Do some incremental zeroing on rawmem, in parallel with inits. 4348 zeroes_done = align_down(zeroes_done, BytesPerInt); 4349 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4350 allocation()->in(AllocateNode::DefaultValue), 4351 allocation()->in(AllocateNode::RawDefaultValue), 4352 zeroes_done, zeroes_needed, 4353 phase); 4354 zeroes_done = zeroes_needed; 4355 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4356 do_zeroing = false; // leave the hole, next time 4357 } 4358 } 4359 4360 // Collect the store and move on: 4361 st->set_req(MemNode::Memory, inits); 4362 inits = st; // put it on the linearized chain 4363 set_req(i, zmem); // unhook from previous position 4364 4365 if (zeroes_done == st_off) 4366 zeroes_done = next_init_off; 4367 4368 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4369 4370 #ifdef ASSERT 4371 // Various order invariants. Weaker than stores_are_sane because 4372 // a large constant tile can be filled in by smaller non-constant stores. 4373 assert(st_off >= last_init_off, "inits do not reverse"); 4374 last_init_off = st_off; 4375 const Type* val = NULL; 4376 if (st_size >= BytesPerInt && 4377 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4378 (int)val->basic_type() < (int)T_OBJECT) { 4379 assert(st_off >= last_tile_end, "tiles do not overlap"); 4380 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4381 last_tile_end = MAX2(last_tile_end, next_init_off); 4382 } else { 4383 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4384 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4385 assert(st_off >= last_init_end, "inits do not overlap"); 4386 last_init_end = next_init_off; // it's a non-tile 4387 } 4388 #endif //ASSERT 4389 } 4390 4391 remove_extra_zeroes(); // clear out all the zmems left over 4392 add_req(inits); 4393 4394 if (!(UseTLAB && ZeroTLAB)) { 4395 // If anything remains to be zeroed, zero it all now. 4396 zeroes_done = align_down(zeroes_done, BytesPerInt); 4397 // if it is the last unused 4 bytes of an instance, forget about it 4398 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4399 if (zeroes_done + BytesPerLong >= size_limit) { 4400 AllocateNode* alloc = allocation(); 4401 assert(alloc != NULL, "must be present"); 4402 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 4403 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4404 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4405 if (zeroes_done == k->layout_helper()) 4406 zeroes_done = size_limit; 4407 } 4408 } 4409 if (zeroes_done < size_limit) { 4410 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4411 allocation()->in(AllocateNode::DefaultValue), 4412 allocation()->in(AllocateNode::RawDefaultValue), 4413 zeroes_done, size_in_bytes, phase); 4414 } 4415 } 4416 4417 set_complete(phase); 4418 return rawmem; 4419 } 4420 4421 4422 #ifdef ASSERT 4423 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4424 if (is_complete()) 4425 return true; // stores could be anything at this point 4426 assert(allocation() != NULL, "must be present"); 4427 intptr_t last_off = allocation()->minimum_header_size(); 4428 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4429 Node* st = in(i); 4430 intptr_t st_off = get_store_offset(st, phase); 4431 if (st_off < 0) continue; // ignore dead garbage 4432 if (last_off > st_off) { 4433 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4434 this->dump(2); 4435 assert(false, "ascending store offsets"); 4436 return false; 4437 } 4438 last_off = st_off + st->as_Store()->memory_size(); 4439 } 4440 return true; 4441 } 4442 #endif //ASSERT 4443 4444 4445 4446 4447 //============================MergeMemNode===================================== 4448 // 4449 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4450 // contributing store or call operations. Each contributor provides the memory 4451 // state for a particular "alias type" (see Compile::alias_type). For example, 4452 // if a MergeMem has an input X for alias category #6, then any memory reference 4453 // to alias category #6 may use X as its memory state input, as an exact equivalent 4454 // to using the MergeMem as a whole. 4455 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4456 // 4457 // (Here, the <N> notation gives the index of the relevant adr_type.) 4458 // 4459 // In one special case (and more cases in the future), alias categories overlap. 4460 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4461 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4462 // it is exactly equivalent to that state W: 4463 // MergeMem(<Bot>: W) <==> W 4464 // 4465 // Usually, the merge has more than one input. In that case, where inputs 4466 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4467 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4468 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4469 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4470 // 4471 // A merge can take a "wide" memory state as one of its narrow inputs. 4472 // This simply means that the merge observes out only the relevant parts of 4473 // the wide input. That is, wide memory states arriving at narrow merge inputs 4474 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4475 // 4476 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4477 // and that memory slices "leak through": 4478 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4479 // 4480 // But, in such a cascade, repeated memory slices can "block the leak": 4481 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4482 // 4483 // In the last example, Y is not part of the combined memory state of the 4484 // outermost MergeMem. The system must, of course, prevent unschedulable 4485 // memory states from arising, so you can be sure that the state Y is somehow 4486 // a precursor to state Y'. 4487 // 4488 // 4489 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4490 // of each MergeMemNode array are exactly the numerical alias indexes, including 4491 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4492 // Compile::alias_type (and kin) produce and manage these indexes. 4493 // 4494 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4495 // (Note that this provides quick access to the top node inside MergeMem methods, 4496 // without the need to reach out via TLS to Compile::current.) 4497 // 4498 // As a consequence of what was just described, a MergeMem that represents a full 4499 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4500 // containing all alias categories. 4501 // 4502 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4503 // 4504 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4505 // a memory state for the alias type <N>, or else the top node, meaning that 4506 // there is no particular input for that alias type. Note that the length of 4507 // a MergeMem is variable, and may be extended at any time to accommodate new 4508 // memory states at larger alias indexes. When merges grow, they are of course 4509 // filled with "top" in the unused in() positions. 4510 // 4511 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4512 // (Top was chosen because it works smoothly with passes like GCM.) 4513 // 4514 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4515 // the type of random VM bits like TLS references.) Since it is always the 4516 // first non-Bot memory slice, some low-level loops use it to initialize an 4517 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4518 // 4519 // 4520 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4521 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4522 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4523 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4524 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4525 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4526 // 4527 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4528 // really that different from the other memory inputs. An abbreviation called 4529 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4530 // 4531 // 4532 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4533 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4534 // that "emerges though" the base memory will be marked as excluding the alias types 4535 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4536 // 4537 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4538 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4539 // 4540 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4541 // (It is currently unimplemented.) As you can see, the resulting merge is 4542 // actually a disjoint union of memory states, rather than an overlay. 4543 // 4544 4545 //------------------------------MergeMemNode----------------------------------- 4546 Node* MergeMemNode::make_empty_memory() { 4547 Node* empty_memory = (Node*) Compile::current()->top(); 4548 assert(empty_memory->is_top(), "correct sentinel identity"); 4549 return empty_memory; 4550 } 4551 4552 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4553 init_class_id(Class_MergeMem); 4554 // all inputs are nullified in Node::Node(int) 4555 // set_input(0, NULL); // no control input 4556 4557 // Initialize the edges uniformly to top, for starters. 4558 Node* empty_mem = make_empty_memory(); 4559 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4560 init_req(i,empty_mem); 4561 } 4562 assert(empty_memory() == empty_mem, ""); 4563 4564 if( new_base != NULL && new_base->is_MergeMem() ) { 4565 MergeMemNode* mdef = new_base->as_MergeMem(); 4566 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4567 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4568 mms.set_memory(mms.memory2()); 4569 } 4570 assert(base_memory() == mdef->base_memory(), ""); 4571 } else { 4572 set_base_memory(new_base); 4573 } 4574 } 4575 4576 // Make a new, untransformed MergeMem with the same base as 'mem'. 4577 // If mem is itself a MergeMem, populate the result with the same edges. 4578 MergeMemNode* MergeMemNode::make(Node* mem) { 4579 return new MergeMemNode(mem); 4580 } 4581 4582 //------------------------------cmp-------------------------------------------- 4583 uint MergeMemNode::hash() const { return NO_HASH; } 4584 bool MergeMemNode::cmp( const Node &n ) const { 4585 return (&n == this); // Always fail except on self 4586 } 4587 4588 //------------------------------Identity--------------------------------------- 4589 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4590 // Identity if this merge point does not record any interesting memory 4591 // disambiguations. 4592 Node* base_mem = base_memory(); 4593 Node* empty_mem = empty_memory(); 4594 if (base_mem != empty_mem) { // Memory path is not dead? 4595 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4596 Node* mem = in(i); 4597 if (mem != empty_mem && mem != base_mem) { 4598 return this; // Many memory splits; no change 4599 } 4600 } 4601 } 4602 return base_mem; // No memory splits; ID on the one true input 4603 } 4604 4605 //------------------------------Ideal------------------------------------------ 4606 // This method is invoked recursively on chains of MergeMem nodes 4607 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4608 // Remove chain'd MergeMems 4609 // 4610 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4611 // relative to the "in(Bot)". Since we are patching both at the same time, 4612 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4613 // but rewrite each "in(i)" relative to the new "in(Bot)". 4614 Node *progress = NULL; 4615 4616 4617 Node* old_base = base_memory(); 4618 Node* empty_mem = empty_memory(); 4619 if (old_base == empty_mem) 4620 return NULL; // Dead memory path. 4621 4622 MergeMemNode* old_mbase; 4623 if (old_base != NULL && old_base->is_MergeMem()) 4624 old_mbase = old_base->as_MergeMem(); 4625 else 4626 old_mbase = NULL; 4627 Node* new_base = old_base; 4628 4629 // simplify stacked MergeMems in base memory 4630 if (old_mbase) new_base = old_mbase->base_memory(); 4631 4632 // the base memory might contribute new slices beyond my req() 4633 if (old_mbase) grow_to_match(old_mbase); 4634 4635 // Look carefully at the base node if it is a phi. 4636 PhiNode* phi_base; 4637 if (new_base != NULL && new_base->is_Phi()) 4638 phi_base = new_base->as_Phi(); 4639 else 4640 phi_base = NULL; 4641 4642 Node* phi_reg = NULL; 4643 uint phi_len = (uint)-1; 4644 if (phi_base != NULL && !phi_base->is_copy()) { 4645 // do not examine phi if degraded to a copy 4646 phi_reg = phi_base->region(); 4647 phi_len = phi_base->req(); 4648 // see if the phi is unfinished 4649 for (uint i = 1; i < phi_len; i++) { 4650 if (phi_base->in(i) == NULL) { 4651 // incomplete phi; do not look at it yet! 4652 phi_reg = NULL; 4653 phi_len = (uint)-1; 4654 break; 4655 } 4656 } 4657 } 4658 4659 // Note: We do not call verify_sparse on entry, because inputs 4660 // can normalize to the base_memory via subsume_node or similar 4661 // mechanisms. This method repairs that damage. 4662 4663 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4664 4665 // Look at each slice. 4666 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4667 Node* old_in = in(i); 4668 // calculate the old memory value 4669 Node* old_mem = old_in; 4670 if (old_mem == empty_mem) old_mem = old_base; 4671 assert(old_mem == memory_at(i), ""); 4672 4673 // maybe update (reslice) the old memory value 4674 4675 // simplify stacked MergeMems 4676 Node* new_mem = old_mem; 4677 MergeMemNode* old_mmem; 4678 if (old_mem != NULL && old_mem->is_MergeMem()) 4679 old_mmem = old_mem->as_MergeMem(); 4680 else 4681 old_mmem = NULL; 4682 if (old_mmem == this) { 4683 // This can happen if loops break up and safepoints disappear. 4684 // A merge of BotPtr (default) with a RawPtr memory derived from a 4685 // safepoint can be rewritten to a merge of the same BotPtr with 4686 // the BotPtr phi coming into the loop. If that phi disappears 4687 // also, we can end up with a self-loop of the mergemem. 4688 // In general, if loops degenerate and memory effects disappear, 4689 // a mergemem can be left looking at itself. This simply means 4690 // that the mergemem's default should be used, since there is 4691 // no longer any apparent effect on this slice. 4692 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4693 // from start. Update the input to TOP. 4694 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4695 } 4696 else if (old_mmem != NULL) { 4697 new_mem = old_mmem->memory_at(i); 4698 } 4699 // else preceding memory was not a MergeMem 4700 4701 // replace equivalent phis (unfortunately, they do not GVN together) 4702 if (new_mem != NULL && new_mem != new_base && 4703 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4704 if (new_mem->is_Phi()) { 4705 PhiNode* phi_mem = new_mem->as_Phi(); 4706 for (uint i = 1; i < phi_len; i++) { 4707 if (phi_base->in(i) != phi_mem->in(i)) { 4708 phi_mem = NULL; 4709 break; 4710 } 4711 } 4712 if (phi_mem != NULL) { 4713 // equivalent phi nodes; revert to the def 4714 new_mem = new_base; 4715 } 4716 } 4717 } 4718 4719 // maybe store down a new value 4720 Node* new_in = new_mem; 4721 if (new_in == new_base) new_in = empty_mem; 4722 4723 if (new_in != old_in) { 4724 // Warning: Do not combine this "if" with the previous "if" 4725 // A memory slice might have be be rewritten even if it is semantically 4726 // unchanged, if the base_memory value has changed. 4727 set_req(i, new_in); 4728 progress = this; // Report progress 4729 } 4730 } 4731 4732 if (new_base != old_base) { 4733 set_req(Compile::AliasIdxBot, new_base); 4734 // Don't use set_base_memory(new_base), because we need to update du. 4735 assert(base_memory() == new_base, ""); 4736 progress = this; 4737 } 4738 4739 if( base_memory() == this ) { 4740 // a self cycle indicates this memory path is dead 4741 set_req(Compile::AliasIdxBot, empty_mem); 4742 } 4743 4744 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4745 // Recursion must occur after the self cycle check above 4746 if( base_memory()->is_MergeMem() ) { 4747 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4748 Node *m = phase->transform(new_mbase); // Rollup any cycles 4749 if( m != NULL && 4750 (m->is_top() || 4751 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4752 // propagate rollup of dead cycle to self 4753 set_req(Compile::AliasIdxBot, empty_mem); 4754 } 4755 } 4756 4757 if( base_memory() == empty_mem ) { 4758 progress = this; 4759 // Cut inputs during Parse phase only. 4760 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4761 if( !can_reshape ) { 4762 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4763 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4764 } 4765 } 4766 } 4767 4768 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4769 // Check if PhiNode::Ideal's "Split phis through memory merges" 4770 // transform should be attempted. Look for this->phi->this cycle. 4771 uint merge_width = req(); 4772 if (merge_width > Compile::AliasIdxRaw) { 4773 PhiNode* phi = base_memory()->as_Phi(); 4774 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4775 if (phi->in(i) == this) { 4776 phase->is_IterGVN()->_worklist.push(phi); 4777 break; 4778 } 4779 } 4780 } 4781 } 4782 4783 assert(progress || verify_sparse(), "please, no dups of base"); 4784 return progress; 4785 } 4786 4787 //-------------------------set_base_memory------------------------------------- 4788 void MergeMemNode::set_base_memory(Node *new_base) { 4789 Node* empty_mem = empty_memory(); 4790 set_req(Compile::AliasIdxBot, new_base); 4791 assert(memory_at(req()) == new_base, "must set default memory"); 4792 // Clear out other occurrences of new_base: 4793 if (new_base != empty_mem) { 4794 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4795 if (in(i) == new_base) set_req(i, empty_mem); 4796 } 4797 } 4798 } 4799 4800 //------------------------------out_RegMask------------------------------------ 4801 const RegMask &MergeMemNode::out_RegMask() const { 4802 return RegMask::Empty; 4803 } 4804 4805 //------------------------------dump_spec-------------------------------------- 4806 #ifndef PRODUCT 4807 void MergeMemNode::dump_spec(outputStream *st) const { 4808 st->print(" {"); 4809 Node* base_mem = base_memory(); 4810 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4811 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4812 if (mem == base_mem) { st->print(" -"); continue; } 4813 st->print( " N%d:", mem->_idx ); 4814 Compile::current()->get_adr_type(i)->dump_on(st); 4815 } 4816 st->print(" }"); 4817 } 4818 #endif // !PRODUCT 4819 4820 4821 #ifdef ASSERT 4822 static bool might_be_same(Node* a, Node* b) { 4823 if (a == b) return true; 4824 if (!(a->is_Phi() || b->is_Phi())) return false; 4825 // phis shift around during optimization 4826 return true; // pretty stupid... 4827 } 4828 4829 // verify a narrow slice (either incoming or outgoing) 4830 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4831 if (!VerifyAliases) return; // don't bother to verify unless requested 4832 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4833 if (Node::in_dump()) return; // muzzle asserts when printing 4834 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4835 assert(n != NULL, ""); 4836 // Elide intervening MergeMem's 4837 while (n->is_MergeMem()) { 4838 n = n->as_MergeMem()->memory_at(alias_idx); 4839 } 4840 Compile* C = Compile::current(); 4841 const TypePtr* n_adr_type = n->adr_type(); 4842 if (n == m->empty_memory()) { 4843 // Implicit copy of base_memory() 4844 } else if (n_adr_type != TypePtr::BOTTOM) { 4845 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4846 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4847 } else { 4848 // A few places like make_runtime_call "know" that VM calls are narrow, 4849 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4850 bool expected_wide_mem = false; 4851 if (n == m->base_memory()) { 4852 expected_wide_mem = true; 4853 } else if (alias_idx == Compile::AliasIdxRaw || 4854 n == m->memory_at(Compile::AliasIdxRaw)) { 4855 expected_wide_mem = true; 4856 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4857 // memory can "leak through" calls on channels that 4858 // are write-once. Allow this also. 4859 expected_wide_mem = true; 4860 } 4861 assert(expected_wide_mem, "expected narrow slice replacement"); 4862 } 4863 } 4864 #else // !ASSERT 4865 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4866 #endif 4867 4868 4869 //-----------------------------memory_at--------------------------------------- 4870 Node* MergeMemNode::memory_at(uint alias_idx) const { 4871 assert(alias_idx >= Compile::AliasIdxRaw || 4872 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4873 "must avoid base_memory and AliasIdxTop"); 4874 4875 // Otherwise, it is a narrow slice. 4876 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4877 Compile *C = Compile::current(); 4878 if (is_empty_memory(n)) { 4879 // the array is sparse; empty slots are the "top" node 4880 n = base_memory(); 4881 assert(Node::in_dump() 4882 || n == NULL || n->bottom_type() == Type::TOP 4883 || n->adr_type() == NULL // address is TOP 4884 || n->adr_type() == TypePtr::BOTTOM 4885 || n->adr_type() == TypeRawPtr::BOTTOM 4886 || Compile::current()->AliasLevel() == 0, 4887 "must be a wide memory"); 4888 // AliasLevel == 0 if we are organizing the memory states manually. 4889 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4890 } else { 4891 // make sure the stored slice is sane 4892 #ifdef ASSERT 4893 if (VMError::is_error_reported() || Node::in_dump()) { 4894 } else if (might_be_same(n, base_memory())) { 4895 // Give it a pass: It is a mostly harmless repetition of the base. 4896 // This can arise normally from node subsumption during optimization. 4897 } else { 4898 verify_memory_slice(this, alias_idx, n); 4899 } 4900 #endif 4901 } 4902 return n; 4903 } 4904 4905 //---------------------------set_memory_at------------------------------------- 4906 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4907 verify_memory_slice(this, alias_idx, n); 4908 Node* empty_mem = empty_memory(); 4909 if (n == base_memory()) n = empty_mem; // collapse default 4910 uint need_req = alias_idx+1; 4911 if (req() < need_req) { 4912 if (n == empty_mem) return; // already the default, so do not grow me 4913 // grow the sparse array 4914 do { 4915 add_req(empty_mem); 4916 } while (req() < need_req); 4917 } 4918 set_req( alias_idx, n ); 4919 } 4920 4921 4922 4923 //--------------------------iteration_setup------------------------------------ 4924 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4925 if (other != NULL) { 4926 grow_to_match(other); 4927 // invariant: the finite support of mm2 is within mm->req() 4928 #ifdef ASSERT 4929 for (uint i = req(); i < other->req(); i++) { 4930 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4931 } 4932 #endif 4933 } 4934 // Replace spurious copies of base_memory by top. 4935 Node* base_mem = base_memory(); 4936 if (base_mem != NULL && !base_mem->is_top()) { 4937 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4938 if (in(i) == base_mem) 4939 set_req(i, empty_memory()); 4940 } 4941 } 4942 } 4943 4944 //---------------------------grow_to_match------------------------------------- 4945 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4946 Node* empty_mem = empty_memory(); 4947 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4948 // look for the finite support of the other memory 4949 for (uint i = other->req(); --i >= req(); ) { 4950 if (other->in(i) != empty_mem) { 4951 uint new_len = i+1; 4952 while (req() < new_len) add_req(empty_mem); 4953 break; 4954 } 4955 } 4956 } 4957 4958 //---------------------------verify_sparse------------------------------------- 4959 #ifndef PRODUCT 4960 bool MergeMemNode::verify_sparse() const { 4961 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4962 Node* base_mem = base_memory(); 4963 // The following can happen in degenerate cases, since empty==top. 4964 if (is_empty_memory(base_mem)) return true; 4965 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4966 assert(in(i) != NULL, "sane slice"); 4967 if (in(i) == base_mem) return false; // should have been the sentinel value! 4968 } 4969 return true; 4970 } 4971 4972 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4973 Node* n; 4974 n = mm->in(idx); 4975 if (mem == n) return true; // might be empty_memory() 4976 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4977 if (mem == n) return true; 4978 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4979 if (mem == n) return true; 4980 if (n == NULL) break; 4981 } 4982 return false; 4983 } 4984 #endif // !PRODUCT