1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "oops/objArrayKlass.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/compile.hpp" 37 #include "opto/connode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/loopnode.hpp" 40 #include "opto/machnode.hpp" 41 #include "opto/matcher.hpp" 42 #include "opto/memnode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/phaseX.hpp" 46 #include "opto/regmask.hpp" 47 #include "opto/valuetypenode.hpp" 48 #include "utilities/align.hpp" 49 #include "utilities/copy.hpp" 50 #include "utilities/macros.hpp" 51 #include "utilities/vmError.hpp" 52 #if INCLUDE_ZGC 53 #include "gc/z/c2/zBarrierSetC2.hpp" 54 #endif 55 56 // Portions of code courtesy of Clifford Click 57 58 // Optimization - Graph Style 59 60 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 61 62 //============================================================================= 63 uint MemNode::size_of() const { return sizeof(*this); } 64 65 const TypePtr *MemNode::adr_type() const { 66 Node* adr = in(Address); 67 if (adr == NULL) return NULL; // node is dead 68 const TypePtr* cross_check = NULL; 69 DEBUG_ONLY(cross_check = _adr_type); 70 return calculate_adr_type(adr->bottom_type(), cross_check); 71 } 72 73 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 74 if (adr != NULL) { 75 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 76 return true; 77 } 78 } 79 return false; 80 } 81 82 #ifndef PRODUCT 83 void MemNode::dump_spec(outputStream *st) const { 84 if (in(Address) == NULL) return; // node is dead 85 #ifndef ASSERT 86 // fake the missing field 87 const TypePtr* _adr_type = NULL; 88 if (in(Address) != NULL) 89 _adr_type = in(Address)->bottom_type()->isa_ptr(); 90 #endif 91 dump_adr_type(this, _adr_type, st); 92 93 Compile* C = Compile::current(); 94 if (C->alias_type(_adr_type)->is_volatile()) { 95 st->print(" Volatile!"); 96 } 97 if (_unaligned_access) { 98 st->print(" unaligned"); 99 } 100 if (_mismatched_access) { 101 st->print(" mismatched"); 102 } 103 if (_unsafe_access) { 104 st->print(" unsafe"); 105 } 106 } 107 108 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 109 st->print(" @"); 110 if (adr_type == NULL) { 111 st->print("NULL"); 112 } else { 113 adr_type->dump_on(st); 114 Compile* C = Compile::current(); 115 Compile::AliasType* atp = NULL; 116 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 117 if (atp == NULL) 118 st->print(", idx=?\?;"); 119 else if (atp->index() == Compile::AliasIdxBot) 120 st->print(", idx=Bot;"); 121 else if (atp->index() == Compile::AliasIdxTop) 122 st->print(", idx=Top;"); 123 else if (atp->index() == Compile::AliasIdxRaw) 124 st->print(", idx=Raw;"); 125 else { 126 ciField* field = atp->field(); 127 if (field) { 128 st->print(", name="); 129 field->print_name_on(st); 130 } 131 st->print(", idx=%d;", atp->index()); 132 } 133 } 134 } 135 136 extern void print_alias_types(); 137 138 #endif 139 140 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 141 assert((t_oop != NULL), "sanity"); 142 bool is_instance = t_oop->is_known_instance_field(); 143 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 144 (load != NULL) && load->is_Load() && 145 (phase->is_IterGVN() != NULL); 146 if (!(is_instance || is_boxed_value_load)) 147 return mchain; // don't try to optimize non-instance types 148 uint instance_id = t_oop->instance_id(); 149 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory); 150 Node *prev = NULL; 151 Node *result = mchain; 152 while (prev != result) { 153 prev = result; 154 if (result == start_mem) 155 break; // hit one of our sentinels 156 // skip over a call which does not affect this memory slice 157 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 158 Node *proj_in = result->in(0); 159 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 160 break; // hit one of our sentinels 161 } else if (proj_in->is_Call()) { 162 // ArrayCopyNodes processed here as well 163 CallNode *call = proj_in->as_Call(); 164 if (!call->may_modify(t_oop, phase)) { // returns false for instances 165 result = call->in(TypeFunc::Memory); 166 } 167 } else if (proj_in->is_Initialize()) { 168 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 169 // Stop if this is the initialization for the object instance which 170 // contains this memory slice, otherwise skip over it. 171 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 172 break; 173 } 174 if (is_instance) { 175 result = proj_in->in(TypeFunc::Memory); 176 } else if (is_boxed_value_load) { 177 Node* klass = alloc->in(AllocateNode::KlassNode); 178 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 179 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 180 result = proj_in->in(TypeFunc::Memory); // not related allocation 181 } 182 } 183 } else if (proj_in->is_MemBar()) { 184 ArrayCopyNode* ac = NULL; 185 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 186 break; 187 } 188 result = proj_in->in(TypeFunc::Memory); 189 } else { 190 assert(false, "unexpected projection"); 191 } 192 } else if (result->is_ClearArray()) { 193 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 194 // Can not bypass initialization of the instance 195 // we are looking for. 196 break; 197 } 198 // Otherwise skip it (the call updated 'result' value). 199 } else if (result->is_MergeMem()) { 200 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 201 } 202 } 203 return result; 204 } 205 206 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 207 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 208 if (t_oop == NULL) 209 return mchain; // don't try to optimize non-oop types 210 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 211 bool is_instance = t_oop->is_known_instance_field(); 212 PhaseIterGVN *igvn = phase->is_IterGVN(); 213 if (is_instance && igvn != NULL && result->is_Phi()) { 214 PhiNode *mphi = result->as_Phi(); 215 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 216 const TypePtr *t = mphi->adr_type(); 217 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 218 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 219 t->is_oopptr()->cast_to_exactness(true) 220 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 221 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) { 222 // clone the Phi with our address type 223 result = mphi->split_out_instance(t_adr, igvn); 224 } else { 225 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 226 } 227 } 228 return result; 229 } 230 231 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 232 uint alias_idx = phase->C->get_alias_index(tp); 233 Node *mem = mmem; 234 #ifdef ASSERT 235 { 236 // Check that current type is consistent with the alias index used during graph construction 237 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 238 bool consistent = adr_check == NULL || adr_check->empty() || 239 phase->C->must_alias(adr_check, alias_idx ); 240 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 241 if( !consistent && adr_check != NULL && !adr_check->empty() && 242 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 243 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 244 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 245 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 246 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 247 // don't assert if it is dead code. 248 consistent = true; 249 } 250 if( !consistent ) { 251 st->print("alias_idx==%d, adr_check==", alias_idx); 252 if( adr_check == NULL ) { 253 st->print("NULL"); 254 } else { 255 adr_check->dump(); 256 } 257 st->cr(); 258 print_alias_types(); 259 assert(consistent, "adr_check must match alias idx"); 260 } 261 } 262 #endif 263 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 264 // means an array I have not precisely typed yet. Do not do any 265 // alias stuff with it any time soon. 266 const TypeOopPtr *toop = tp->isa_oopptr(); 267 if( tp->base() != Type::AnyPtr && 268 !(toop && 269 toop->klass() != NULL && 270 toop->klass()->is_java_lang_Object() && 271 toop->offset() == Type::OffsetBot) ) { 272 // compress paths and change unreachable cycles to TOP 273 // If not, we can update the input infinitely along a MergeMem cycle 274 // Equivalent code in PhiNode::Ideal 275 Node* m = phase->transform(mmem); 276 // If transformed to a MergeMem, get the desired slice 277 // Otherwise the returned node represents memory for every slice 278 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 279 // Update input if it is progress over what we have now 280 } 281 return mem; 282 } 283 284 //--------------------------Ideal_common--------------------------------------- 285 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 286 // Unhook non-raw memories from complete (macro-expanded) initializations. 287 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 288 // If our control input is a dead region, kill all below the region 289 Node *ctl = in(MemNode::Control); 290 if (ctl && remove_dead_region(phase, can_reshape)) 291 return this; 292 ctl = in(MemNode::Control); 293 // Don't bother trying to transform a dead node 294 if (ctl && ctl->is_top()) return NodeSentinel; 295 296 PhaseIterGVN *igvn = phase->is_IterGVN(); 297 // Wait if control on the worklist. 298 if (ctl && can_reshape && igvn != NULL) { 299 Node* bol = NULL; 300 Node* cmp = NULL; 301 if (ctl->in(0)->is_If()) { 302 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 303 bol = ctl->in(0)->in(1); 304 if (bol->is_Bool()) 305 cmp = ctl->in(0)->in(1)->in(1); 306 } 307 if (igvn->_worklist.member(ctl) || 308 (bol != NULL && igvn->_worklist.member(bol)) || 309 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 310 // This control path may be dead. 311 // Delay this memory node transformation until the control is processed. 312 phase->is_IterGVN()->_worklist.push(this); 313 return NodeSentinel; // caller will return NULL 314 } 315 } 316 // Ignore if memory is dead, or self-loop 317 Node *mem = in(MemNode::Memory); 318 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 319 assert(mem != this, "dead loop in MemNode::Ideal"); 320 321 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 322 // This memory slice may be dead. 323 // Delay this mem node transformation until the memory is processed. 324 phase->is_IterGVN()->_worklist.push(this); 325 return NodeSentinel; // caller will return NULL 326 } 327 328 Node *address = in(MemNode::Address); 329 const Type *t_adr = phase->type(address); 330 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 331 332 if (can_reshape && igvn != NULL && 333 (igvn->_worklist.member(address) || 334 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 335 // The address's base and type may change when the address is processed. 336 // Delay this mem node transformation until the address is processed. 337 phase->is_IterGVN()->_worklist.push(this); 338 return NodeSentinel; // caller will return NULL 339 } 340 341 // Do NOT remove or optimize the next lines: ensure a new alias index 342 // is allocated for an oop pointer type before Escape Analysis. 343 // Note: C++ will not remove it since the call has side effect. 344 if (t_adr->isa_oopptr()) { 345 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 346 } 347 348 Node* base = NULL; 349 if (address->is_AddP()) { 350 base = address->in(AddPNode::Base); 351 } 352 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 353 !t_adr->isa_rawptr()) { 354 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 355 // Skip this node optimization if its address has TOP base. 356 return NodeSentinel; // caller will return NULL 357 } 358 359 // Avoid independent memory operations 360 Node* old_mem = mem; 361 362 // The code which unhooks non-raw memories from complete (macro-expanded) 363 // initializations was removed. After macro-expansion all stores catched 364 // by Initialize node became raw stores and there is no information 365 // which memory slices they modify. So it is unsafe to move any memory 366 // operation above these stores. Also in most cases hooked non-raw memories 367 // were already unhooked by using information from detect_ptr_independence() 368 // and find_previous_store(). 369 370 if (mem->is_MergeMem()) { 371 MergeMemNode* mmem = mem->as_MergeMem(); 372 const TypePtr *tp = t_adr->is_ptr(); 373 374 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 375 } 376 377 if (mem != old_mem) { 378 set_req(MemNode::Memory, mem); 379 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) { 380 igvn->_worklist.push(old_mem); 381 } 382 if (phase->type(mem) == Type::TOP) return NodeSentinel; 383 return this; 384 } 385 386 // let the subclass continue analyzing... 387 return NULL; 388 } 389 390 // Helper function for proving some simple control dominations. 391 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 392 // Already assumes that 'dom' is available at 'sub', and that 'sub' 393 // is not a constant (dominated by the method's StartNode). 394 // Used by MemNode::find_previous_store to prove that the 395 // control input of a memory operation predates (dominates) 396 // an allocation it wants to look past. 397 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 398 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 399 return false; // Conservative answer for dead code 400 401 // Check 'dom'. Skip Proj and CatchProj nodes. 402 dom = dom->find_exact_control(dom); 403 if (dom == NULL || dom->is_top()) 404 return false; // Conservative answer for dead code 405 406 if (dom == sub) { 407 // For the case when, for example, 'sub' is Initialize and the original 408 // 'dom' is Proj node of the 'sub'. 409 return false; 410 } 411 412 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 413 return true; 414 415 // 'dom' dominates 'sub' if its control edge and control edges 416 // of all its inputs dominate or equal to sub's control edge. 417 418 // Currently 'sub' is either Allocate, Initialize or Start nodes. 419 // Or Region for the check in LoadNode::Ideal(); 420 // 'sub' should have sub->in(0) != NULL. 421 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 422 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 423 424 // Get control edge of 'sub'. 425 Node* orig_sub = sub; 426 sub = sub->find_exact_control(sub->in(0)); 427 if (sub == NULL || sub->is_top()) 428 return false; // Conservative answer for dead code 429 430 assert(sub->is_CFG(), "expecting control"); 431 432 if (sub == dom) 433 return true; 434 435 if (sub->is_Start() || sub->is_Root()) 436 return false; 437 438 { 439 // Check all control edges of 'dom'. 440 441 ResourceMark rm; 442 Arena* arena = Thread::current()->resource_area(); 443 Node_List nlist(arena); 444 Unique_Node_List dom_list(arena); 445 446 dom_list.push(dom); 447 bool only_dominating_controls = false; 448 449 for (uint next = 0; next < dom_list.size(); next++) { 450 Node* n = dom_list.at(next); 451 if (n == orig_sub) 452 return false; // One of dom's inputs dominated by sub. 453 if (!n->is_CFG() && n->pinned()) { 454 // Check only own control edge for pinned non-control nodes. 455 n = n->find_exact_control(n->in(0)); 456 if (n == NULL || n->is_top()) 457 return false; // Conservative answer for dead code 458 assert(n->is_CFG(), "expecting control"); 459 dom_list.push(n); 460 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 461 only_dominating_controls = true; 462 } else if (n->is_CFG()) { 463 if (n->dominates(sub, nlist)) 464 only_dominating_controls = true; 465 else 466 return false; 467 } else { 468 // First, own control edge. 469 Node* m = n->find_exact_control(n->in(0)); 470 if (m != NULL) { 471 if (m->is_top()) 472 return false; // Conservative answer for dead code 473 dom_list.push(m); 474 } 475 // Now, the rest of edges. 476 uint cnt = n->req(); 477 for (uint i = 1; i < cnt; i++) { 478 m = n->find_exact_control(n->in(i)); 479 if (m == NULL || m->is_top()) 480 continue; 481 dom_list.push(m); 482 } 483 } 484 } 485 return only_dominating_controls; 486 } 487 } 488 489 //---------------------detect_ptr_independence--------------------------------- 490 // Used by MemNode::find_previous_store to prove that two base 491 // pointers are never equal. 492 // The pointers are accompanied by their associated allocations, 493 // if any, which have been previously discovered by the caller. 494 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 495 Node* p2, AllocateNode* a2, 496 PhaseTransform* phase) { 497 // Attempt to prove that these two pointers cannot be aliased. 498 // They may both manifestly be allocations, and they should differ. 499 // Or, if they are not both allocations, they can be distinct constants. 500 // Otherwise, one is an allocation and the other a pre-existing value. 501 if (a1 == NULL && a2 == NULL) { // neither an allocation 502 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 503 } else if (a1 != NULL && a2 != NULL) { // both allocations 504 return (a1 != a2); 505 } else if (a1 != NULL) { // one allocation a1 506 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 507 return all_controls_dominate(p2, a1); 508 } else { //(a2 != NULL) // one allocation a2 509 return all_controls_dominate(p1, a2); 510 } 511 return false; 512 } 513 514 515 // Find an arraycopy that must have set (can_see_stored_value=true) or 516 // could have set (can_see_stored_value=false) the value for this load 517 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 518 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 519 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 520 Node* mb = mem->in(0); 521 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 522 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 523 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 524 if (ac->is_clonebasic()) { 525 intptr_t offset; 526 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 527 if (alloc != NULL && alloc == ld_alloc) { 528 return ac; 529 } 530 } 531 } 532 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 533 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 534 535 if (ac->is_arraycopy_validated() || 536 ac->is_copyof_validated() || 537 ac->is_copyofrange_validated()) { 538 Node* ld_addp = in(MemNode::Address); 539 if (ld_addp->is_AddP()) { 540 Node* ld_base = ld_addp->in(AddPNode::Address); 541 Node* ld_offs = ld_addp->in(AddPNode::Offset); 542 543 Node* dest = ac->in(ArrayCopyNode::Dest); 544 545 if (dest == ld_base) { 546 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 547 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 548 return ac; 549 } 550 if (!can_see_stored_value) { 551 mem = ac->in(TypeFunc::Memory); 552 } 553 } 554 } 555 } 556 } 557 return NULL; 558 } 559 560 // The logic for reordering loads and stores uses four steps: 561 // (a) Walk carefully past stores and initializations which we 562 // can prove are independent of this load. 563 // (b) Observe that the next memory state makes an exact match 564 // with self (load or store), and locate the relevant store. 565 // (c) Ensure that, if we were to wire self directly to the store, 566 // the optimizer would fold it up somehow. 567 // (d) Do the rewiring, and return, depending on some other part of 568 // the optimizer to fold up the load. 569 // This routine handles steps (a) and (b). Steps (c) and (d) are 570 // specific to loads and stores, so they are handled by the callers. 571 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 572 // 573 Node* MemNode::find_previous_store(PhaseTransform* phase) { 574 Node* ctrl = in(MemNode::Control); 575 Node* adr = in(MemNode::Address); 576 intptr_t offset = 0; 577 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 578 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 579 580 if (offset == Type::OffsetBot) 581 return NULL; // cannot unalias unless there are precise offsets 582 583 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 584 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 585 586 intptr_t size_in_bytes = memory_size(); 587 588 Node* mem = in(MemNode::Memory); // start searching here... 589 590 int cnt = 50; // Cycle limiter 591 for (;;) { // While we can dance past unrelated stores... 592 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 593 594 Node* prev = mem; 595 if (mem->is_Store()) { 596 Node* st_adr = mem->in(MemNode::Address); 597 intptr_t st_offset = 0; 598 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 599 if (st_base == NULL) 600 break; // inscrutable pointer 601 602 // For raw accesses it's not enough to prove that constant offsets don't intersect. 603 // We need the bases to be the equal in order for the offset check to make sense. 604 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 605 break; 606 } 607 608 if (st_offset != offset && st_offset != Type::OffsetBot) { 609 const int MAX_STORE = BytesPerLong; 610 if (st_offset >= offset + size_in_bytes || 611 st_offset <= offset - MAX_STORE || 612 st_offset <= offset - mem->as_Store()->memory_size()) { 613 // Success: The offsets are provably independent. 614 // (You may ask, why not just test st_offset != offset and be done? 615 // The answer is that stores of different sizes can co-exist 616 // in the same sequence of RawMem effects. We sometimes initialize 617 // a whole 'tile' of array elements with a single jint or jlong.) 618 mem = mem->in(MemNode::Memory); 619 continue; // (a) advance through independent store memory 620 } 621 } 622 if (st_base != base && 623 detect_ptr_independence(base, alloc, 624 st_base, 625 AllocateNode::Ideal_allocation(st_base, phase), 626 phase)) { 627 // Success: The bases are provably independent. 628 mem = mem->in(MemNode::Memory); 629 continue; // (a) advance through independent store memory 630 } 631 632 // (b) At this point, if the bases or offsets do not agree, we lose, 633 // since we have not managed to prove 'this' and 'mem' independent. 634 if (st_base == base && st_offset == offset) { 635 return mem; // let caller handle steps (c), (d) 636 } 637 638 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 639 InitializeNode* st_init = mem->in(0)->as_Initialize(); 640 AllocateNode* st_alloc = st_init->allocation(); 641 if (st_alloc == NULL) 642 break; // something degenerated 643 bool known_identical = false; 644 bool known_independent = false; 645 if (alloc == st_alloc) 646 known_identical = true; 647 else if (alloc != NULL) 648 known_independent = true; 649 else if (all_controls_dominate(this, st_alloc)) 650 known_independent = true; 651 652 if (known_independent) { 653 // The bases are provably independent: Either they are 654 // manifestly distinct allocations, or else the control 655 // of this load dominates the store's allocation. 656 int alias_idx = phase->C->get_alias_index(adr_type()); 657 if (alias_idx == Compile::AliasIdxRaw) { 658 mem = st_alloc->in(TypeFunc::Memory); 659 } else { 660 mem = st_init->memory(alias_idx); 661 } 662 continue; // (a) advance through independent store memory 663 } 664 665 // (b) at this point, if we are not looking at a store initializing 666 // the same allocation we are loading from, we lose. 667 if (known_identical) { 668 // From caller, can_see_stored_value will consult find_captured_store. 669 return mem; // let caller handle steps (c), (d) 670 } 671 672 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 673 if (prev != mem) { 674 // Found an arraycopy but it doesn't affect that load 675 continue; 676 } 677 // Found an arraycopy that may affect that load 678 return mem; 679 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 680 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 681 if (mem->is_Proj() && mem->in(0)->is_Call()) { 682 // ArrayCopyNodes processed here as well. 683 CallNode *call = mem->in(0)->as_Call(); 684 if (!call->may_modify(addr_t, phase)) { 685 mem = call->in(TypeFunc::Memory); 686 continue; // (a) advance through independent call memory 687 } 688 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 689 ArrayCopyNode* ac = NULL; 690 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 691 break; 692 } 693 mem = mem->in(0)->in(TypeFunc::Memory); 694 continue; // (a) advance through independent MemBar memory 695 } else if (mem->is_ClearArray()) { 696 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 697 // (the call updated 'mem' value) 698 continue; // (a) advance through independent allocation memory 699 } else { 700 // Can not bypass initialization of the instance 701 // we are looking for. 702 return mem; 703 } 704 } else if (mem->is_MergeMem()) { 705 int alias_idx = phase->C->get_alias_index(adr_type()); 706 mem = mem->as_MergeMem()->memory_at(alias_idx); 707 continue; // (a) advance through independent MergeMem memory 708 } 709 } 710 711 // Unless there is an explicit 'continue', we must bail out here, 712 // because 'mem' is an inscrutable memory state (e.g., a call). 713 break; 714 } 715 716 return NULL; // bail out 717 } 718 719 //----------------------calculate_adr_type------------------------------------- 720 // Helper function. Notices when the given type of address hits top or bottom. 721 // Also, asserts a cross-check of the type against the expected address type. 722 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 723 if (t == Type::TOP) return NULL; // does not touch memory any more? 724 #ifdef PRODUCT 725 cross_check = NULL; 726 #else 727 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL; 728 #endif 729 const TypePtr* tp = t->isa_ptr(); 730 if (tp == NULL) { 731 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 732 return TypePtr::BOTTOM; // touches lots of memory 733 } else { 734 #ifdef ASSERT 735 // %%%% [phh] We don't check the alias index if cross_check is 736 // TypeRawPtr::BOTTOM. Needs to be investigated. 737 if (cross_check != NULL && 738 cross_check != TypePtr::BOTTOM && 739 cross_check != TypeRawPtr::BOTTOM) { 740 // Recheck the alias index, to see if it has changed (due to a bug). 741 Compile* C = Compile::current(); 742 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 743 "must stay in the original alias category"); 744 // The type of the address must be contained in the adr_type, 745 // disregarding "null"-ness. 746 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 747 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 748 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 749 "real address must not escape from expected memory type"); 750 } 751 #endif 752 return tp; 753 } 754 } 755 756 //============================================================================= 757 // Should LoadNode::Ideal() attempt to remove control edges? 758 bool LoadNode::can_remove_control() const { 759 return true; 760 } 761 uint LoadNode::size_of() const { return sizeof(*this); } 762 bool LoadNode::cmp( const Node &n ) const 763 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 764 const Type *LoadNode::bottom_type() const { return _type; } 765 uint LoadNode::ideal_reg() const { 766 return _type->ideal_reg(); 767 } 768 769 #ifndef PRODUCT 770 void LoadNode::dump_spec(outputStream *st) const { 771 MemNode::dump_spec(st); 772 if( !Verbose && !WizardMode ) { 773 // standard dump does this in Verbose and WizardMode 774 st->print(" #"); _type->dump_on(st); 775 } 776 if (!depends_only_on_test()) { 777 st->print(" (does not depend only on test)"); 778 } 779 } 780 #endif 781 782 #ifdef ASSERT 783 //----------------------------is_immutable_value------------------------------- 784 // Helper function to allow a raw load without control edge for some cases 785 bool LoadNode::is_immutable_value(Node* adr) { 786 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 787 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 788 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 789 in_bytes(JavaThread::osthread_offset()))); 790 } 791 #endif 792 793 //----------------------------LoadNode::make----------------------------------- 794 // Polymorphic factory method: 795 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 796 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 797 Compile* C = gvn.C; 798 799 // sanity check the alias category against the created node type 800 assert(!(adr_type->isa_oopptr() && 801 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 802 "use LoadKlassNode instead"); 803 assert(!(adr_type->isa_aryptr() && 804 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 805 "use LoadRangeNode instead"); 806 // Check control edge of raw loads 807 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 808 // oop will be recorded in oop map if load crosses safepoint 809 rt->isa_oopptr() || is_immutable_value(adr), 810 "raw memory operations should have control edge"); 811 LoadNode* load = NULL; 812 switch (bt) { 813 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 814 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 815 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 816 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 817 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 818 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 819 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 820 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 821 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 822 case T_VALUETYPE: 823 case T_OBJECT: 824 #ifdef _LP64 825 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 826 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 827 } else 828 #endif 829 { 830 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 831 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 832 } 833 break; 834 default: 835 ShouldNotReachHere(); 836 break; 837 } 838 assert(load != NULL, "LoadNode should have been created"); 839 if (unaligned) { 840 load->set_unaligned_access(); 841 } 842 if (mismatched) { 843 load->set_mismatched_access(); 844 } 845 if (unsafe) { 846 load->set_unsafe_access(); 847 } 848 if (load->Opcode() == Op_LoadN) { 849 Node* ld = gvn.transform(load); 850 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 851 } 852 853 return load; 854 } 855 856 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 857 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 858 bool require_atomic = true; 859 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 860 if (unaligned) { 861 load->set_unaligned_access(); 862 } 863 if (mismatched) { 864 load->set_mismatched_access(); 865 } 866 if (unsafe) { 867 load->set_unsafe_access(); 868 } 869 return load; 870 } 871 872 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 873 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 874 bool require_atomic = true; 875 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 876 if (unaligned) { 877 load->set_unaligned_access(); 878 } 879 if (mismatched) { 880 load->set_mismatched_access(); 881 } 882 if (unsafe) { 883 load->set_unsafe_access(); 884 } 885 return load; 886 } 887 888 889 890 //------------------------------hash------------------------------------------- 891 uint LoadNode::hash() const { 892 // unroll addition of interesting fields 893 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 894 } 895 896 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 897 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 898 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 899 bool is_stable_ary = FoldStableValues && 900 (tp != NULL) && (tp->isa_aryptr() != NULL) && 901 tp->isa_aryptr()->is_stable(); 902 903 return (eliminate_boxing && non_volatile) || is_stable_ary; 904 } 905 906 return false; 907 } 908 909 // Is the value loaded previously stored by an arraycopy? If so return 910 // a load node that reads from the source array so we may be able to 911 // optimize out the ArrayCopy node later. 912 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 913 #if INCLUDE_ZGC 914 if (UseZGC) { 915 if (bottom_type()->make_oopptr() != NULL) { 916 return NULL; 917 } 918 } 919 #endif 920 921 Node* ld_adr = in(MemNode::Address); 922 intptr_t ld_off = 0; 923 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 924 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 925 if (ac != NULL) { 926 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 927 928 Node* mem = ac->in(TypeFunc::Memory); 929 Node* ctl = ac->in(0); 930 Node* src = ac->in(ArrayCopyNode::Src); 931 932 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 933 return NULL; 934 } 935 936 LoadNode* ld = clone()->as_Load(); 937 Node* addp = in(MemNode::Address)->clone(); 938 if (ac->as_ArrayCopy()->is_clonebasic()) { 939 assert(ld_alloc != NULL, "need an alloc"); 940 assert(addp->is_AddP(), "address must be addp"); 941 assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address"); 942 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 943 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern"); 944 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern"); 945 addp->set_req(AddPNode::Base, src->in(AddPNode::Base)); 946 addp->set_req(AddPNode::Address, src->in(AddPNode::Address)); 947 } else { 948 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 949 ac->as_ArrayCopy()->is_copyof_validated() || 950 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 951 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 952 addp->set_req(AddPNode::Base, src); 953 addp->set_req(AddPNode::Address, src); 954 955 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 956 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 957 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 958 uint shift = exact_log2(type2aelembytes(ary_elem)); 959 960 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 961 #ifdef _LP64 962 diff = phase->transform(new ConvI2LNode(diff)); 963 #endif 964 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 965 966 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 967 addp->set_req(AddPNode::Offset, offset); 968 } 969 addp = phase->transform(addp); 970 #ifdef ASSERT 971 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 972 ld->_adr_type = adr_type; 973 #endif 974 ld->set_req(MemNode::Address, addp); 975 ld->set_req(0, ctl); 976 ld->set_req(MemNode::Memory, mem); 977 // load depends on the tests that validate the arraycopy 978 ld->_control_dependency = Pinned; 979 return ld; 980 } 981 return NULL; 982 } 983 984 985 //---------------------------can_see_stored_value------------------------------ 986 // This routine exists to make sure this set of tests is done the same 987 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 988 // will change the graph shape in a way which makes memory alive twice at the 989 // same time (uses the Oracle model of aliasing), then some 990 // LoadXNode::Identity will fold things back to the equivalence-class model 991 // of aliasing. 992 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 993 Node* ld_adr = in(MemNode::Address); 994 intptr_t ld_off = 0; 995 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off); 996 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase); 997 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 998 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 999 // This is more general than load from boxing objects. 1000 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1001 uint alias_idx = atp->index(); 1002 bool final = !atp->is_rewritable(); 1003 Node* result = NULL; 1004 Node* current = st; 1005 // Skip through chains of MemBarNodes checking the MergeMems for 1006 // new states for the slice of this load. Stop once any other 1007 // kind of node is encountered. Loads from final memory can skip 1008 // through any kind of MemBar but normal loads shouldn't skip 1009 // through MemBarAcquire since the could allow them to move out of 1010 // a synchronized region. 1011 while (current->is_Proj()) { 1012 int opc = current->in(0)->Opcode(); 1013 if ((final && (opc == Op_MemBarAcquire || 1014 opc == Op_MemBarAcquireLock || 1015 opc == Op_LoadFence)) || 1016 opc == Op_MemBarRelease || 1017 opc == Op_StoreFence || 1018 opc == Op_MemBarReleaseLock || 1019 opc == Op_MemBarStoreStore || 1020 opc == Op_MemBarCPUOrder) { 1021 Node* mem = current->in(0)->in(TypeFunc::Memory); 1022 if (mem->is_MergeMem()) { 1023 MergeMemNode* merge = mem->as_MergeMem(); 1024 Node* new_st = merge->memory_at(alias_idx); 1025 if (new_st == merge->base_memory()) { 1026 // Keep searching 1027 current = new_st; 1028 continue; 1029 } 1030 // Save the new memory state for the slice and fall through 1031 // to exit. 1032 result = new_st; 1033 } 1034 } 1035 break; 1036 } 1037 if (result != NULL) { 1038 st = result; 1039 } 1040 } 1041 1042 // Loop around twice in the case Load -> Initialize -> Store. 1043 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1044 for (int trip = 0; trip <= 1; trip++) { 1045 1046 if (st->is_Store()) { 1047 Node* st_adr = st->in(MemNode::Address); 1048 if (!phase->eqv(st_adr, ld_adr)) { 1049 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers). 1050 intptr_t st_off = 0; 1051 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off); 1052 if (ld_base == NULL) return NULL; 1053 if (st_base == NULL) return NULL; 1054 if (ld_base->uncast() != st_base->uncast()) return NULL; 1055 if (ld_off != st_off) return NULL; 1056 if (ld_off == Type::OffsetBot) return NULL; 1057 // Same base, same offset. 1058 // Possible improvement for arrays: check index value instead of absolute offset. 1059 1060 // At this point we have proven something like this setup: 1061 // B = << base >> 1062 // L = LoadQ(AddP(Check/CastPP(B), #Off)) 1063 // S = StoreQ(AddP( B , #Off), V) 1064 // (Actually, we haven't yet proven the Q's are the same.) 1065 // In other words, we are loading from a casted version of 1066 // the same pointer-and-offset that we stored to. 1067 // Thus, we are able to replace L by V. 1068 } 1069 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1070 if (store_Opcode() != st->Opcode()) 1071 return NULL; 1072 return st->in(MemNode::ValueIn); 1073 } 1074 1075 // A load from a freshly-created object always returns zero. 1076 // (This can happen after LoadNode::Ideal resets the load's memory input 1077 // to find_captured_store, which returned InitializeNode::zero_memory.) 1078 if (st->is_Proj() && st->in(0)->is_Allocate() && 1079 (st->in(0) == ld_alloc) && 1080 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1081 // return a zero value for the load's basic type 1082 // (This is one of the few places where a generic PhaseTransform 1083 // can create new nodes. Think of it as lazily manifesting 1084 // virtually pre-existing constants.) 1085 assert(memory_type() != T_VALUETYPE, "should not be used for value types"); 1086 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue); 1087 if (default_value != NULL) { 1088 return default_value; 1089 } 1090 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 1091 return phase->zerocon(memory_type()); 1092 } 1093 1094 // A load from an initialization barrier can match a captured store. 1095 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1096 InitializeNode* init = st->in(0)->as_Initialize(); 1097 AllocateNode* alloc = init->allocation(); 1098 if ((alloc != NULL) && (alloc == ld_alloc)) { 1099 // examine a captured store value 1100 st = init->find_captured_store(ld_off, memory_size(), phase); 1101 if (st != NULL) { 1102 continue; // take one more trip around 1103 } 1104 } 1105 } 1106 1107 // Load boxed value from result of valueOf() call is input parameter. 1108 if (this->is_Load() && ld_adr->is_AddP() && 1109 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1110 intptr_t ignore = 0; 1111 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1112 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1113 base = bs->step_over_gc_barrier(base); 1114 if (base != NULL && base->is_Proj() && 1115 base->as_Proj()->_con == TypeFunc::Parms && 1116 base->in(0)->is_CallStaticJava() && 1117 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1118 return base->in(0)->in(TypeFunc::Parms); 1119 } 1120 } 1121 1122 break; 1123 } 1124 1125 return NULL; 1126 } 1127 1128 //----------------------is_instance_field_load_with_local_phi------------------ 1129 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1130 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1131 in(Address)->is_AddP() ) { 1132 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1133 // Only instances and boxed values. 1134 if( t_oop != NULL && 1135 (t_oop->is_ptr_to_boxed_value() || 1136 t_oop->is_known_instance_field()) && 1137 t_oop->offset() != Type::OffsetBot && 1138 t_oop->offset() != Type::OffsetTop) { 1139 return true; 1140 } 1141 } 1142 return false; 1143 } 1144 1145 //------------------------------Identity--------------------------------------- 1146 // Loads are identity if previous store is to same address 1147 Node* LoadNode::Identity(PhaseGVN* phase) { 1148 // Loading from a ValueTypePtr? The ValueTypePtr has the values of 1149 // all fields as input. Look for the field with matching offset. 1150 Node* addr = in(Address); 1151 intptr_t offset; 1152 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset); 1153 if (base != NULL && base->is_ValueTypePtr() && offset > oopDesc::klass_offset_in_bytes()) { 1154 Node* value = base->as_ValueTypePtr()->field_value_by_offset((int)offset, true); 1155 if (value->is_ValueType()) { 1156 // Non-flattened value type field 1157 ValueTypeNode* vt = value->as_ValueType(); 1158 if (vt->is_allocated(phase)) { 1159 value = vt->get_oop(); 1160 } else { 1161 // Not yet allocated, bail out 1162 value = NULL; 1163 } 1164 } 1165 if (value != NULL) { 1166 if (Opcode() == Op_LoadN) { 1167 // Encode oop value if we are loading a narrow oop 1168 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded"); 1169 value = phase->transform(new EncodePNode(value, bottom_type())); 1170 } 1171 return value; 1172 } 1173 } 1174 1175 // If the previous store-maker is the right kind of Store, and the store is 1176 // to the same address, then we are equal to the value stored. 1177 Node* mem = in(Memory); 1178 Node* value = can_see_stored_value(mem, phase); 1179 if( value ) { 1180 // byte, short & char stores truncate naturally. 1181 // A load has to load the truncated value which requires 1182 // some sort of masking operation and that requires an 1183 // Ideal call instead of an Identity call. 1184 if (memory_size() < BytesPerInt) { 1185 // If the input to the store does not fit with the load's result type, 1186 // it must be truncated via an Ideal call. 1187 if (!phase->type(value)->higher_equal(phase->type(this))) 1188 return this; 1189 } 1190 // (This works even when value is a Con, but LoadNode::Value 1191 // usually runs first, producing the singleton type of the Con.) 1192 return value; 1193 } 1194 1195 // Search for an existing data phi which was generated before for the same 1196 // instance's field to avoid infinite generation of phis in a loop. 1197 Node *region = mem->in(0); 1198 if (is_instance_field_load_with_local_phi(region)) { 1199 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1200 int this_index = phase->C->get_alias_index(addr_t); 1201 int this_offset = addr_t->offset(); 1202 int this_iid = addr_t->instance_id(); 1203 if (!addr_t->is_known_instance() && 1204 addr_t->is_ptr_to_boxed_value()) { 1205 // Use _idx of address base (could be Phi node) for boxed values. 1206 intptr_t ignore = 0; 1207 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1208 if (base == NULL) { 1209 return this; 1210 } 1211 this_iid = base->_idx; 1212 } 1213 const Type* this_type = bottom_type(); 1214 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1215 Node* phi = region->fast_out(i); 1216 if (phi->is_Phi() && phi != mem && 1217 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1218 return phi; 1219 } 1220 } 1221 } 1222 1223 return this; 1224 } 1225 1226 // Construct an equivalent unsigned load. 1227 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1228 BasicType bt = T_ILLEGAL; 1229 const Type* rt = NULL; 1230 switch (Opcode()) { 1231 case Op_LoadUB: return this; 1232 case Op_LoadUS: return this; 1233 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1234 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1235 default: 1236 assert(false, "no unsigned variant: %s", Name()); 1237 return NULL; 1238 } 1239 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1240 raw_adr_type(), rt, bt, _mo, _control_dependency, 1241 is_unaligned_access(), is_mismatched_access()); 1242 } 1243 1244 // Construct an equivalent signed load. 1245 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1246 BasicType bt = T_ILLEGAL; 1247 const Type* rt = NULL; 1248 switch (Opcode()) { 1249 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1250 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1251 case Op_LoadB: // fall through 1252 case Op_LoadS: // fall through 1253 case Op_LoadI: // fall through 1254 case Op_LoadL: return this; 1255 default: 1256 assert(false, "no signed variant: %s", Name()); 1257 return NULL; 1258 } 1259 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1260 raw_adr_type(), rt, bt, _mo, _control_dependency, 1261 is_unaligned_access(), is_mismatched_access()); 1262 } 1263 1264 // We're loading from an object which has autobox behaviour. 1265 // If this object is result of a valueOf call we'll have a phi 1266 // merging a newly allocated object and a load from the cache. 1267 // We want to replace this load with the original incoming 1268 // argument to the valueOf call. 1269 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1270 assert(phase->C->eliminate_boxing(), "sanity"); 1271 intptr_t ignore = 0; 1272 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1273 if ((base == NULL) || base->is_Phi()) { 1274 // Push the loads from the phi that comes from valueOf up 1275 // through it to allow elimination of the loads and the recovery 1276 // of the original value. It is done in split_through_phi(). 1277 return NULL; 1278 } else if (base->is_Load() || 1279 (base->is_DecodeN() && base->in(1)->is_Load())) { 1280 // Eliminate the load of boxed value for integer types from the cache 1281 // array by deriving the value from the index into the array. 1282 // Capture the offset of the load and then reverse the computation. 1283 1284 // Get LoadN node which loads a boxing object from 'cache' array. 1285 if (base->is_DecodeN()) { 1286 base = base->in(1); 1287 } 1288 if (!base->in(Address)->is_AddP()) { 1289 return NULL; // Complex address 1290 } 1291 AddPNode* address = base->in(Address)->as_AddP(); 1292 Node* cache_base = address->in(AddPNode::Base); 1293 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1294 // Get ConP node which is static 'cache' field. 1295 cache_base = cache_base->in(1); 1296 } 1297 if ((cache_base != NULL) && cache_base->is_Con()) { 1298 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1299 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1300 Node* elements[4]; 1301 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1302 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1303 if (count > 0 && elements[0]->is_Con() && 1304 (count == 1 || 1305 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1306 elements[1]->in(2) == phase->intcon(shift)))) { 1307 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1308 // Fetch the box object cache[0] at the base of the array and get its value 1309 ciInstance* box = array->obj_at(0)->as_instance(); 1310 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1311 assert(ik->is_box_klass(), "sanity"); 1312 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1313 if (ik->nof_nonstatic_fields() == 1) { 1314 // This should be true nonstatic_field_at requires calling 1315 // nof_nonstatic_fields so check it anyway 1316 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1317 BasicType bt = c.basic_type(); 1318 // Only integer types have boxing cache. 1319 assert(bt == T_BOOLEAN || bt == T_CHAR || 1320 bt == T_BYTE || bt == T_SHORT || 1321 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1322 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1323 if (cache_low != (int)cache_low) { 1324 return NULL; // should not happen since cache is array indexed by value 1325 } 1326 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1327 if (offset != (int)offset) { 1328 return NULL; // should not happen since cache is array indexed by value 1329 } 1330 // Add up all the offsets making of the address of the load 1331 Node* result = elements[0]; 1332 for (int i = 1; i < count; i++) { 1333 result = phase->transform(new AddXNode(result, elements[i])); 1334 } 1335 // Remove the constant offset from the address and then 1336 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1337 // remove the scaling of the offset to recover the original index. 1338 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1339 // Peel the shift off directly but wrap it in a dummy node 1340 // since Ideal can't return existing nodes 1341 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1342 } else if (result->is_Add() && result->in(2)->is_Con() && 1343 result->in(1)->Opcode() == Op_LShiftX && 1344 result->in(1)->in(2) == phase->intcon(shift)) { 1345 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1346 // but for boxing cache access we know that X<<Z will not overflow 1347 // (there is range check) so we do this optimizatrion by hand here. 1348 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1349 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1350 } else { 1351 result = new RShiftXNode(result, phase->intcon(shift)); 1352 } 1353 #ifdef _LP64 1354 if (bt != T_LONG) { 1355 result = new ConvL2INode(phase->transform(result)); 1356 } 1357 #else 1358 if (bt == T_LONG) { 1359 result = new ConvI2LNode(phase->transform(result)); 1360 } 1361 #endif 1362 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1363 // Need to preserve unboxing load type if it is unsigned. 1364 switch(this->Opcode()) { 1365 case Op_LoadUB: 1366 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1367 break; 1368 case Op_LoadUS: 1369 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1370 break; 1371 } 1372 return result; 1373 } 1374 } 1375 } 1376 } 1377 } 1378 return NULL; 1379 } 1380 1381 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1382 Node* region = phi->in(0); 1383 if (region == NULL) { 1384 return false; // Wait stable graph 1385 } 1386 uint cnt = phi->req(); 1387 for (uint i = 1; i < cnt; i++) { 1388 Node* rc = region->in(i); 1389 if (rc == NULL || phase->type(rc) == Type::TOP) 1390 return false; // Wait stable graph 1391 Node* in = phi->in(i); 1392 if (in == NULL || phase->type(in) == Type::TOP) 1393 return false; // Wait stable graph 1394 } 1395 return true; 1396 } 1397 //------------------------------split_through_phi------------------------------ 1398 // Split instance or boxed field load through Phi. 1399 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1400 Node* mem = in(Memory); 1401 Node* address = in(Address); 1402 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1403 1404 assert((t_oop != NULL) && 1405 (t_oop->is_known_instance_field() || 1406 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1407 1408 Compile* C = phase->C; 1409 intptr_t ignore = 0; 1410 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1411 bool base_is_phi = (base != NULL) && base->is_Phi(); 1412 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1413 (base != NULL) && (base == address->in(AddPNode::Base)) && 1414 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1415 1416 if (!((mem->is_Phi() || base_is_phi) && 1417 (load_boxed_values || t_oop->is_known_instance_field()))) { 1418 return NULL; // memory is not Phi 1419 } 1420 1421 if (mem->is_Phi()) { 1422 if (!stable_phi(mem->as_Phi(), phase)) { 1423 return NULL; // Wait stable graph 1424 } 1425 uint cnt = mem->req(); 1426 // Check for loop invariant memory. 1427 if (cnt == 3) { 1428 for (uint i = 1; i < cnt; i++) { 1429 Node* in = mem->in(i); 1430 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1431 if (m == mem) { 1432 if (i == 1) { 1433 // if the first edge was a loop, check second edge too. 1434 // If both are replaceable - we are in an infinite loop 1435 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1436 if (n == mem) { 1437 break; 1438 } 1439 } 1440 set_req(Memory, mem->in(cnt - i)); 1441 return this; // made change 1442 } 1443 } 1444 } 1445 } 1446 if (base_is_phi) { 1447 if (!stable_phi(base->as_Phi(), phase)) { 1448 return NULL; // Wait stable graph 1449 } 1450 uint cnt = base->req(); 1451 // Check for loop invariant memory. 1452 if (cnt == 3) { 1453 for (uint i = 1; i < cnt; i++) { 1454 if (base->in(i) == base) { 1455 return NULL; // Wait stable graph 1456 } 1457 } 1458 } 1459 } 1460 1461 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1462 1463 // Split through Phi (see original code in loopopts.cpp). 1464 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1465 1466 // Do nothing here if Identity will find a value 1467 // (to avoid infinite chain of value phis generation). 1468 if (!phase->eqv(this, phase->apply_identity(this))) 1469 return NULL; 1470 1471 // Select Region to split through. 1472 Node* region; 1473 if (!base_is_phi) { 1474 assert(mem->is_Phi(), "sanity"); 1475 region = mem->in(0); 1476 // Skip if the region dominates some control edge of the address. 1477 if (!MemNode::all_controls_dominate(address, region)) 1478 return NULL; 1479 } else if (!mem->is_Phi()) { 1480 assert(base_is_phi, "sanity"); 1481 region = base->in(0); 1482 // Skip if the region dominates some control edge of the memory. 1483 if (!MemNode::all_controls_dominate(mem, region)) 1484 return NULL; 1485 } else if (base->in(0) != mem->in(0)) { 1486 assert(base_is_phi && mem->is_Phi(), "sanity"); 1487 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1488 region = base->in(0); 1489 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1490 region = mem->in(0); 1491 } else { 1492 return NULL; // complex graph 1493 } 1494 } else { 1495 assert(base->in(0) == mem->in(0), "sanity"); 1496 region = mem->in(0); 1497 } 1498 1499 const Type* this_type = this->bottom_type(); 1500 int this_index = C->get_alias_index(t_oop); 1501 int this_offset = t_oop->offset(); 1502 int this_iid = t_oop->instance_id(); 1503 if (!t_oop->is_known_instance() && load_boxed_values) { 1504 // Use _idx of address base for boxed values. 1505 this_iid = base->_idx; 1506 } 1507 PhaseIterGVN* igvn = phase->is_IterGVN(); 1508 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1509 for (uint i = 1; i < region->req(); i++) { 1510 Node* x; 1511 Node* the_clone = NULL; 1512 if (region->in(i) == C->top()) { 1513 x = C->top(); // Dead path? Use a dead data op 1514 } else { 1515 x = this->clone(); // Else clone up the data op 1516 the_clone = x; // Remember for possible deletion. 1517 // Alter data node to use pre-phi inputs 1518 if (this->in(0) == region) { 1519 x->set_req(0, region->in(i)); 1520 } else { 1521 x->set_req(0, NULL); 1522 } 1523 if (mem->is_Phi() && (mem->in(0) == region)) { 1524 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1525 } 1526 if (address->is_Phi() && address->in(0) == region) { 1527 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1528 } 1529 if (base_is_phi && (base->in(0) == region)) { 1530 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1531 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1532 x->set_req(Address, adr_x); 1533 } 1534 } 1535 // Check for a 'win' on some paths 1536 const Type *t = x->Value(igvn); 1537 1538 bool singleton = t->singleton(); 1539 1540 // See comments in PhaseIdealLoop::split_thru_phi(). 1541 if (singleton && t == Type::TOP) { 1542 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1543 } 1544 1545 if (singleton) { 1546 x = igvn->makecon(t); 1547 } else { 1548 // We now call Identity to try to simplify the cloned node. 1549 // Note that some Identity methods call phase->type(this). 1550 // Make sure that the type array is big enough for 1551 // our new node, even though we may throw the node away. 1552 // (This tweaking with igvn only works because x is a new node.) 1553 igvn->set_type(x, t); 1554 // If x is a TypeNode, capture any more-precise type permanently into Node 1555 // otherwise it will be not updated during igvn->transform since 1556 // igvn->type(x) is set to x->Value() already. 1557 x->raise_bottom_type(t); 1558 Node *y = igvn->apply_identity(x); 1559 if (y != x) { 1560 x = y; 1561 } else { 1562 y = igvn->hash_find_insert(x); 1563 if (y) { 1564 x = y; 1565 } else { 1566 // Else x is a new node we are keeping 1567 // We do not need register_new_node_with_optimizer 1568 // because set_type has already been called. 1569 igvn->_worklist.push(x); 1570 } 1571 } 1572 } 1573 if (x != the_clone && the_clone != NULL) { 1574 igvn->remove_dead_node(the_clone); 1575 } 1576 phi->set_req(i, x); 1577 } 1578 // Record Phi 1579 igvn->register_new_node_with_optimizer(phi); 1580 return phi; 1581 } 1582 1583 //------------------------------Ideal------------------------------------------ 1584 // If the load is from Field memory and the pointer is non-null, it might be possible to 1585 // zero out the control input. 1586 // If the offset is constant and the base is an object allocation, 1587 // try to hook me up to the exact initializing store. 1588 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1589 Node* p = MemNode::Ideal_common(phase, can_reshape); 1590 if (p) return (p == NodeSentinel) ? NULL : p; 1591 1592 Node* ctrl = in(MemNode::Control); 1593 Node* address = in(MemNode::Address); 1594 bool progress = false; 1595 1596 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1597 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1598 1599 // Skip up past a SafePoint control. Cannot do this for Stores because 1600 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1601 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1602 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1603 !addr_mark ) { 1604 ctrl = ctrl->in(0); 1605 set_req(MemNode::Control,ctrl); 1606 progress = true; 1607 } 1608 1609 intptr_t ignore = 0; 1610 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1611 if (base != NULL 1612 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1613 // Check for useless control edge in some common special cases 1614 if (in(MemNode::Control) != NULL 1615 && can_remove_control() 1616 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1617 && all_controls_dominate(base, phase->C->start())) { 1618 // A method-invariant, non-null address (constant or 'this' argument). 1619 set_req(MemNode::Control, NULL); 1620 progress = true; 1621 } 1622 } 1623 1624 Node* mem = in(MemNode::Memory); 1625 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1626 1627 if (can_reshape && (addr_t != NULL)) { 1628 // try to optimize our memory input 1629 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1630 if (opt_mem != mem) { 1631 set_req(MemNode::Memory, opt_mem); 1632 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1633 return this; 1634 } 1635 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1636 if ((t_oop != NULL) && 1637 (t_oop->is_known_instance_field() || 1638 t_oop->is_ptr_to_boxed_value())) { 1639 PhaseIterGVN *igvn = phase->is_IterGVN(); 1640 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1641 // Delay this transformation until memory Phi is processed. 1642 phase->is_IterGVN()->_worklist.push(this); 1643 return NULL; 1644 } 1645 // Split instance field load through Phi. 1646 Node* result = split_through_phi(phase); 1647 if (result != NULL) return result; 1648 1649 if (t_oop->is_ptr_to_boxed_value()) { 1650 Node* result = eliminate_autobox(phase); 1651 if (result != NULL) return result; 1652 } 1653 } 1654 } 1655 1656 // Is there a dominating load that loads the same value? Leave 1657 // anything that is not a load of a field/array element (like 1658 // barriers etc.) alone 1659 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) { 1660 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1661 Node *use = mem->fast_out(i); 1662 if (use != this && 1663 use->Opcode() == Opcode() && 1664 use->in(0) != NULL && 1665 use->in(0) != in(0) && 1666 use->in(Address) == in(Address)) { 1667 Node* ctl = in(0); 1668 for (int i = 0; i < 10 && ctl != NULL; i++) { 1669 ctl = IfNode::up_one_dom(ctl); 1670 if (ctl == use->in(0)) { 1671 set_req(0, use->in(0)); 1672 return this; 1673 } 1674 } 1675 } 1676 } 1677 } 1678 1679 // Check for prior store with a different base or offset; make Load 1680 // independent. Skip through any number of them. Bail out if the stores 1681 // are in an endless dead cycle and report no progress. This is a key 1682 // transform for Reflection. However, if after skipping through the Stores 1683 // we can't then fold up against a prior store do NOT do the transform as 1684 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1685 // array memory alive twice: once for the hoisted Load and again after the 1686 // bypassed Store. This situation only works if EVERYBODY who does 1687 // anti-dependence work knows how to bypass. I.e. we need all 1688 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1689 // the alias index stuff. So instead, peek through Stores and IFF we can 1690 // fold up, do so. 1691 Node* prev_mem = find_previous_store(phase); 1692 if (prev_mem != NULL) { 1693 Node* value = can_see_arraycopy_value(prev_mem, phase); 1694 if (value != NULL) { 1695 return value; 1696 } 1697 } 1698 // Steps (a), (b): Walk past independent stores to find an exact match. 1699 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1700 // (c) See if we can fold up on the spot, but don't fold up here. 1701 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1702 // just return a prior value, which is done by Identity calls. 1703 if (can_see_stored_value(prev_mem, phase)) { 1704 // Make ready for step (d): 1705 set_req(MemNode::Memory, prev_mem); 1706 return this; 1707 } 1708 } 1709 1710 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase); 1711 if (alloc != NULL && mem->is_Proj() && 1712 mem->in(0) != NULL && 1713 mem->in(0) == alloc->initialization() && 1714 Opcode() == Op_LoadX && 1715 alloc->initialization()->proj_out_or_null(0) != NULL) { 1716 InitializeNode* init = alloc->initialization(); 1717 Node* control = init->proj_out(0); 1718 return alloc->make_ideal_mark(phase, address, control, mem, NULL); 1719 } 1720 1721 return progress ? this : NULL; 1722 } 1723 1724 // Helper to recognize certain Klass fields which are invariant across 1725 // some group of array types (e.g., int[] or all T[] where T < Object). 1726 const Type* 1727 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1728 ciKlass* klass) const { 1729 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1730 // The field is Klass::_modifier_flags. Return its (constant) value. 1731 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1732 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1733 return TypeInt::make(klass->modifier_flags()); 1734 } 1735 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1736 // The field is Klass::_access_flags. Return its (constant) value. 1737 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1738 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1739 return TypeInt::make(klass->access_flags()); 1740 } 1741 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1742 // The field is Klass::_layout_helper. Return its constant value if known. 1743 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1744 return TypeInt::make(klass->layout_helper()); 1745 } 1746 1747 // No match. 1748 return NULL; 1749 } 1750 1751 //------------------------------Value----------------------------------------- 1752 const Type* LoadNode::Value(PhaseGVN* phase) const { 1753 // Either input is TOP ==> the result is TOP 1754 Node* mem = in(MemNode::Memory); 1755 const Type *t1 = phase->type(mem); 1756 if (t1 == Type::TOP) return Type::TOP; 1757 Node* adr = in(MemNode::Address); 1758 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1759 if (tp == NULL || tp->empty()) return Type::TOP; 1760 int off = tp->offset(); 1761 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1762 Compile* C = phase->C; 1763 1764 // Try to guess loaded type from pointer type 1765 if (tp->isa_aryptr()) { 1766 const TypeAryPtr* ary = tp->is_aryptr(); 1767 const Type* t = ary->elem(); 1768 1769 // Determine whether the reference is beyond the header or not, by comparing 1770 // the offset against the offset of the start of the array's data. 1771 // Different array types begin at slightly different offsets (12 vs. 16). 1772 // We choose T_BYTE as an example base type that is least restrictive 1773 // as to alignment, which will therefore produce the smallest 1774 // possible base offset. 1775 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1776 const bool off_beyond_header = (off >= min_base_off); 1777 1778 // Try to constant-fold a stable array element. 1779 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1780 // Make sure the reference is not into the header and the offset is constant 1781 ciObject* aobj = ary->const_oop(); 1782 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1783 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1784 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1785 stable_dimension, 1786 memory_type(), is_unsigned()); 1787 if (con_type != NULL) { 1788 return con_type; 1789 } 1790 } 1791 } 1792 1793 // Don't do this for integer types. There is only potential profit if 1794 // the element type t is lower than _type; that is, for int types, if _type is 1795 // more restrictive than t. This only happens here if one is short and the other 1796 // char (both 16 bits), and in those cases we've made an intentional decision 1797 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1798 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1799 // 1800 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1801 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1802 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1803 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1804 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1805 // In fact, that could have been the original type of p1, and p1 could have 1806 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1807 // expression (LShiftL quux 3) independently optimized to the constant 8. 1808 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1809 && (_type->isa_vect() == NULL) 1810 && t->isa_valuetype() == NULL 1811 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1812 // t might actually be lower than _type, if _type is a unique 1813 // concrete subclass of abstract class t. 1814 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header? 1815 const Type* jt = t->join_speculative(_type); 1816 // In any case, do not allow the join, per se, to empty out the type. 1817 if (jt->empty() && !t->empty()) { 1818 // This can happen if a interface-typed array narrows to a class type. 1819 jt = _type; 1820 } 1821 #ifdef ASSERT 1822 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1823 // The pointers in the autobox arrays are always non-null 1824 Node* base = adr->in(AddPNode::Base); 1825 if ((base != NULL) && base->is_DecodeN()) { 1826 // Get LoadN node which loads IntegerCache.cache field 1827 base = base->in(1); 1828 } 1829 if ((base != NULL) && base->is_Con()) { 1830 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1831 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1832 // It could be narrow oop 1833 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1834 } 1835 } 1836 } 1837 #endif 1838 return jt; 1839 } 1840 } 1841 } else if (tp->base() == Type::InstPtr) { 1842 assert( off != Type::OffsetBot || 1843 // arrays can be cast to Objects 1844 tp->is_oopptr()->klass()->is_java_lang_Object() || 1845 tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() || 1846 // unsafe field access may not have a constant offset 1847 C->has_unsafe_access(), 1848 "Field accesses must be precise" ); 1849 // For oop loads, we expect the _type to be precise. 1850 1851 const TypeInstPtr* tinst = tp->is_instptr(); 1852 BasicType bt = memory_type(); 1853 1854 // Fold component and value mirror loads 1855 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1856 if (ik == phase->C->env()->Class_klass() && (off == java_lang_Class::component_mirror_offset_in_bytes() || 1857 off == java_lang_Class::value_mirror_offset_in_bytes())) { 1858 ciType* mirror_type = tinst->java_mirror_type(); 1859 if (mirror_type != NULL) { 1860 const Type* const_oop = TypePtr::NULL_PTR; 1861 if (mirror_type->is_array_klass()) { 1862 const_oop = TypeInstPtr::make(mirror_type->as_array_klass()->component_mirror_instance()); 1863 } else if (mirror_type->is_valuetype()) { 1864 const_oop = TypeInstPtr::make(mirror_type->as_value_klass()->value_mirror_instance()); 1865 } 1866 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop; 1867 } 1868 } 1869 1870 // Optimize loads from constant fields. 1871 ciObject* const_oop = tinst->const_oop(); 1872 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1873 ciType* mirror_type = const_oop->as_instance()->java_mirror_type(); 1874 if (mirror_type != NULL && mirror_type->is_valuetype()) { 1875 ciValueKlass* vk = mirror_type->as_value_klass(); 1876 if (off == vk->default_value_offset()) { 1877 // Loading a special hidden field that contains the oop of the default value type 1878 const Type* const_oop = TypeInstPtr::make(vk->default_value_instance()); 1879 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop; 1880 } 1881 } 1882 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt); 1883 if (con_type != NULL) { 1884 return con_type; 1885 } 1886 } 1887 } else if (tp->base() == Type::KlassPtr) { 1888 assert( off != Type::OffsetBot || 1889 // arrays can be cast to Objects 1890 tp->is_klassptr()->klass() == NULL || 1891 tp->is_klassptr()->klass()->is_java_lang_Object() || 1892 // also allow array-loading from the primary supertype 1893 // array during subtype checks 1894 Opcode() == Op_LoadKlass, 1895 "Field accesses must be precise" ); 1896 // For klass/static loads, we expect the _type to be precise 1897 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) { 1898 if (adr->is_Load() && off == 0) { 1899 /* With mirrors being an indirect in the Klass* 1900 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 1901 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 1902 * 1903 * So check the type and klass of the node before the LoadP. 1904 */ 1905 Node* adr2 = adr->in(MemNode::Address); 1906 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1907 if (tkls != NULL) { 1908 ciKlass* klass = tkls->klass(); 1909 if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1910 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1911 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1912 return TypeInstPtr::make(klass->java_mirror()); 1913 } 1914 } 1915 } else { 1916 // Check for a load of the default value offset from the ValueKlassFixedBlock: 1917 // LoadI(LoadP(value_klass, adr_valueklass_fixed_block_offset), default_value_offset_offset) 1918 intptr_t offset = 0; 1919 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1920 if (base != NULL && base->is_Load() && offset == in_bytes(ValueKlass::default_value_offset_offset())) { 1921 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr(); 1922 if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_valuetype() && 1923 tkls->offset() == in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())) { 1924 assert(base->Opcode() == Op_LoadP, "must load an oop from klass"); 1925 assert(Opcode() == Op_LoadI, "must load an int from fixed block"); 1926 return TypeInt::make(tkls->klass()->as_value_klass()->default_value_offset()); 1927 } 1928 } 1929 } 1930 } 1931 1932 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1933 if (tkls != NULL && !StressReflectiveCode) { 1934 ciKlass* klass = tkls->klass(); 1935 if (tkls->is_loaded() && tkls->klass_is_exact()) { 1936 // We are loading a field from a Klass metaobject whose identity 1937 // is known at compile time (the type is "exact" or "precise"). 1938 // Check for fields we know are maintained as constants by the VM. 1939 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1940 // The field is Klass::_super_check_offset. Return its (constant) value. 1941 // (Folds up type checking code.) 1942 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1943 return TypeInt::make(klass->super_check_offset()); 1944 } 1945 // Compute index into primary_supers array 1946 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1947 // Check for overflowing; use unsigned compare to handle the negative case. 1948 if( depth < ciKlass::primary_super_limit() ) { 1949 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1950 // (Folds up type checking code.) 1951 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1952 ciKlass *ss = klass->super_of_depth(depth); 1953 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1954 } 1955 const Type* aift = load_array_final_field(tkls, klass); 1956 if (aift != NULL) return aift; 1957 } 1958 1959 // We can still check if we are loading from the primary_supers array at a 1960 // shallow enough depth. Even though the klass is not exact, entries less 1961 // than or equal to its super depth are correct. 1962 if (tkls->is_loaded()) { 1963 ciType *inner = klass; 1964 while( inner->is_obj_array_klass() ) 1965 inner = inner->as_obj_array_klass()->base_element_type(); 1966 if( inner->is_instance_klass() && 1967 !inner->as_instance_klass()->flags().is_interface() ) { 1968 // Compute index into primary_supers array 1969 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1970 // Check for overflowing; use unsigned compare to handle the negative case. 1971 if( depth < ciKlass::primary_super_limit() && 1972 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1973 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1974 // (Folds up type checking code.) 1975 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1976 ciKlass *ss = klass->super_of_depth(depth); 1977 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1978 } 1979 } 1980 } 1981 1982 // If the type is enough to determine that the thing is not an array, 1983 // we can give the layout_helper a positive interval type. 1984 // This will help short-circuit some reflective code. 1985 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1986 && !klass->is_array_klass() // not directly typed as an array 1987 && !klass->is_interface() // specifically not Serializable & Cloneable 1988 && !klass->is_java_lang_Object() // not the supertype of all T[] 1989 ) { 1990 // Note: When interfaces are reliable, we can narrow the interface 1991 // test to (klass != Serializable && klass != Cloneable). 1992 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1993 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1994 // The key property of this type is that it folds up tests 1995 // for array-ness, since it proves that the layout_helper is positive. 1996 // Thus, a generic value like the basic object layout helper works fine. 1997 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1998 } 1999 } 2000 2001 // If we are loading from a freshly-allocated object, produce a zero, 2002 // if the load is provably beyond the header of the object. 2003 // (Also allow a variable load from a fresh array to produce zero.) 2004 const TypeOopPtr *tinst = tp->isa_oopptr(); 2005 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 2006 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 2007 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 2008 Node* value = can_see_stored_value(mem,phase); 2009 if (value != NULL && value->is_Con()) { 2010 assert(value->bottom_type()->higher_equal(_type),"sanity"); 2011 return value->bottom_type(); 2012 } 2013 } 2014 2015 if (is_instance) { 2016 // If we have an instance type and our memory input is the 2017 // programs's initial memory state, there is no matching store, 2018 // so just return a zero of the appropriate type 2019 Node *mem = in(MemNode::Memory); 2020 if (mem->is_Parm() && mem->in(0)->is_Start()) { 2021 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 2022 return Type::get_zero_type(_type->basic_type()); 2023 } 2024 } 2025 return _type; 2026 } 2027 2028 //------------------------------match_edge------------------------------------- 2029 // Do we Match on this edge index or not? Match only the address. 2030 uint LoadNode::match_edge(uint idx) const { 2031 return idx == MemNode::Address; 2032 } 2033 2034 //--------------------------LoadBNode::Ideal-------------------------------------- 2035 // 2036 // If the previous store is to the same address as this load, 2037 // and the value stored was larger than a byte, replace this load 2038 // with the value stored truncated to a byte. If no truncation is 2039 // needed, the replacement is done in LoadNode::Identity(). 2040 // 2041 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2042 Node* mem = in(MemNode::Memory); 2043 Node* value = can_see_stored_value(mem,phase); 2044 if( value && !phase->type(value)->higher_equal( _type ) ) { 2045 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 2046 return new RShiftINode(result, phase->intcon(24)); 2047 } 2048 // Identity call will handle the case where truncation is not needed. 2049 return LoadNode::Ideal(phase, can_reshape); 2050 } 2051 2052 const Type* LoadBNode::Value(PhaseGVN* phase) const { 2053 Node* mem = in(MemNode::Memory); 2054 Node* value = can_see_stored_value(mem,phase); 2055 if (value != NULL && value->is_Con() && 2056 !value->bottom_type()->higher_equal(_type)) { 2057 // If the input to the store does not fit with the load's result type, 2058 // it must be truncated. We can't delay until Ideal call since 2059 // a singleton Value is needed for split_thru_phi optimization. 2060 int con = value->get_int(); 2061 return TypeInt::make((con << 24) >> 24); 2062 } 2063 return LoadNode::Value(phase); 2064 } 2065 2066 //--------------------------LoadUBNode::Ideal------------------------------------- 2067 // 2068 // If the previous store is to the same address as this load, 2069 // and the value stored was larger than a byte, replace this load 2070 // with the value stored truncated to a byte. If no truncation is 2071 // needed, the replacement is done in LoadNode::Identity(). 2072 // 2073 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2074 Node* mem = in(MemNode::Memory); 2075 Node* value = can_see_stored_value(mem, phase); 2076 if (value && !phase->type(value)->higher_equal(_type)) 2077 return new AndINode(value, phase->intcon(0xFF)); 2078 // Identity call will handle the case where truncation is not needed. 2079 return LoadNode::Ideal(phase, can_reshape); 2080 } 2081 2082 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2083 Node* mem = in(MemNode::Memory); 2084 Node* value = can_see_stored_value(mem,phase); 2085 if (value != NULL && value->is_Con() && 2086 !value->bottom_type()->higher_equal(_type)) { 2087 // If the input to the store does not fit with the load's result type, 2088 // it must be truncated. We can't delay until Ideal call since 2089 // a singleton Value is needed for split_thru_phi optimization. 2090 int con = value->get_int(); 2091 return TypeInt::make(con & 0xFF); 2092 } 2093 return LoadNode::Value(phase); 2094 } 2095 2096 //--------------------------LoadUSNode::Ideal------------------------------------- 2097 // 2098 // If the previous store is to the same address as this load, 2099 // and the value stored was larger than a char, replace this load 2100 // with the value stored truncated to a char. If no truncation is 2101 // needed, the replacement is done in LoadNode::Identity(). 2102 // 2103 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2104 Node* mem = in(MemNode::Memory); 2105 Node* value = can_see_stored_value(mem,phase); 2106 if( value && !phase->type(value)->higher_equal( _type ) ) 2107 return new AndINode(value,phase->intcon(0xFFFF)); 2108 // Identity call will handle the case where truncation is not needed. 2109 return LoadNode::Ideal(phase, can_reshape); 2110 } 2111 2112 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2113 Node* mem = in(MemNode::Memory); 2114 Node* value = can_see_stored_value(mem,phase); 2115 if (value != NULL && value->is_Con() && 2116 !value->bottom_type()->higher_equal(_type)) { 2117 // If the input to the store does not fit with the load's result type, 2118 // it must be truncated. We can't delay until Ideal call since 2119 // a singleton Value is needed for split_thru_phi optimization. 2120 int con = value->get_int(); 2121 return TypeInt::make(con & 0xFFFF); 2122 } 2123 return LoadNode::Value(phase); 2124 } 2125 2126 //--------------------------LoadSNode::Ideal-------------------------------------- 2127 // 2128 // If the previous store is to the same address as this load, 2129 // and the value stored was larger than a short, replace this load 2130 // with the value stored truncated to a short. If no truncation is 2131 // needed, the replacement is done in LoadNode::Identity(). 2132 // 2133 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2134 Node* mem = in(MemNode::Memory); 2135 Node* value = can_see_stored_value(mem,phase); 2136 if( value && !phase->type(value)->higher_equal( _type ) ) { 2137 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 2138 return new RShiftINode(result, phase->intcon(16)); 2139 } 2140 // Identity call will handle the case where truncation is not needed. 2141 return LoadNode::Ideal(phase, can_reshape); 2142 } 2143 2144 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2145 Node* mem = in(MemNode::Memory); 2146 Node* value = can_see_stored_value(mem,phase); 2147 if (value != NULL && value->is_Con() && 2148 !value->bottom_type()->higher_equal(_type)) { 2149 // If the input to the store does not fit with the load's result type, 2150 // it must be truncated. We can't delay until Ideal call since 2151 // a singleton Value is needed for split_thru_phi optimization. 2152 int con = value->get_int(); 2153 return TypeInt::make((con << 16) >> 16); 2154 } 2155 return LoadNode::Value(phase); 2156 } 2157 2158 //============================================================================= 2159 //----------------------------LoadKlassNode::make------------------------------ 2160 // Polymorphic factory method: 2161 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2162 // sanity check the alias category against the created node type 2163 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2164 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2165 #ifdef _LP64 2166 if (adr_type->is_ptr_to_narrowklass()) { 2167 assert(UseCompressedClassPointers, "no compressed klasses"); 2168 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2169 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2170 } 2171 #endif 2172 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2173 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2174 } 2175 2176 //------------------------------Value------------------------------------------ 2177 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2178 return klass_value_common(phase); 2179 } 2180 2181 // In most cases, LoadKlassNode does not have the control input set. If the control 2182 // input is set, it must not be removed (by LoadNode::Ideal()). 2183 bool LoadKlassNode::can_remove_control() const { 2184 return false; 2185 } 2186 2187 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2188 // Either input is TOP ==> the result is TOP 2189 const Type *t1 = phase->type( in(MemNode::Memory) ); 2190 if (t1 == Type::TOP) return Type::TOP; 2191 Node *adr = in(MemNode::Address); 2192 const Type *t2 = phase->type( adr ); 2193 if (t2 == Type::TOP) return Type::TOP; 2194 const TypePtr *tp = t2->is_ptr(); 2195 if (TypePtr::above_centerline(tp->ptr()) || 2196 tp->ptr() == TypePtr::Null) return Type::TOP; 2197 2198 // Return a more precise klass, if possible 2199 const TypeInstPtr *tinst = tp->isa_instptr(); 2200 if (tinst != NULL) { 2201 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2202 int offset = tinst->offset(); 2203 if (ik == phase->C->env()->Class_klass() 2204 && (offset == java_lang_Class::klass_offset_in_bytes() || 2205 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2206 // We are loading a special hidden field from a Class mirror object, 2207 // the field which points to the VM's Klass metaobject. 2208 bool is_val_type = false; 2209 ciType* t = tinst->java_mirror_type(&is_val_type); 2210 // java_mirror_type returns non-null for compile-time Class constants. 2211 if (t != NULL) { 2212 // constant oop => constant klass 2213 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2214 if (t->is_void()) { 2215 // We cannot create a void array. Since void is a primitive type return null 2216 // klass. Users of this result need to do a null check on the returned klass. 2217 return TypePtr::NULL_PTR; 2218 } 2219 return TypeKlassPtr::make(ciArrayKlass::make(t, /* never_null */ is_val_type)); 2220 } 2221 if (!t->is_klass()) { 2222 // a primitive Class (e.g., int.class) has NULL for a klass field 2223 return TypePtr::NULL_PTR; 2224 } 2225 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2226 return TypeKlassPtr::make(t->as_klass()); 2227 } 2228 // non-constant mirror, so we can't tell what's going on 2229 } 2230 if( !ik->is_loaded() ) 2231 return _type; // Bail out if not loaded 2232 if (offset == oopDesc::klass_offset_in_bytes()) { 2233 if (tinst->klass_is_exact()) { 2234 return TypeKlassPtr::make(ik); 2235 } 2236 // See if we can become precise: no subklasses and no interface 2237 // (Note: We need to support verified interfaces.) 2238 if (!ik->is_interface() && !ik->has_subklass()) { 2239 //assert(!UseExactTypes, "this code should be useless with exact types"); 2240 // Add a dependence; if any subclass added we need to recompile 2241 if (!ik->is_final()) { 2242 // %%% should use stronger assert_unique_concrete_subtype instead 2243 phase->C->dependencies()->assert_leaf_type(ik); 2244 } 2245 // Return precise klass 2246 return TypeKlassPtr::make(ik); 2247 } 2248 2249 // Return root of possible klass 2250 return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0)); 2251 } 2252 } 2253 2254 // Check for loading klass from an array 2255 const TypeAryPtr *tary = tp->isa_aryptr(); 2256 if (tary != NULL) { 2257 ciKlass *tary_klass = tary->klass(); 2258 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2259 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2260 ciArrayKlass* ak = tary_klass->as_array_klass(); 2261 // Do not fold klass loads from [V? because the runtime type might be [V due to [V <: [V? 2262 bool can_be_null_free = !tary->is_known_instance() && ak->is_obj_array_klass() && !ak->storage_properties().is_null_free() && ak->element_klass()->is_valuetype(); 2263 2264 if (tary->klass_is_exact() && !can_be_null_free) { 2265 return TypeKlassPtr::make(tary_klass); 2266 } 2267 2268 // If the klass is an object array, we defer the question to the 2269 // array component klass. 2270 if (ak->is_obj_array_klass() && !can_be_null_free) { 2271 assert(ak->is_loaded(), ""); 2272 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2273 if (base_k->is_loaded() && base_k->is_instance_klass()) { 2274 ciInstanceKlass *ik = base_k->as_instance_klass(); 2275 // See if we can become precise: no subklasses and no interface 2276 if (!ik->is_interface() && !ik->has_subklass()) { 2277 //assert(!UseExactTypes, "this code should be useless with exact types"); 2278 // Add a dependence; if any subclass added we need to recompile 2279 if (!ik->is_final()) { 2280 phase->C->dependencies()->assert_leaf_type(ik); 2281 } 2282 // Return precise array klass 2283 return TypeKlassPtr::make(ak); 2284 } 2285 } 2286 return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0)); 2287 } else if (ak->is_type_array_klass()) { 2288 //assert(!UseExactTypes, "this code should be useless with exact types"); 2289 return TypeKlassPtr::make(ak); // These are always precise 2290 } 2291 } 2292 } 2293 2294 // Check for loading klass from an array klass 2295 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2296 if (tkls != NULL && !StressReflectiveCode) { 2297 if (!tkls->is_loaded()) { 2298 return _type; // Bail out if not loaded 2299 } 2300 ciKlass* klass = tkls->klass(); 2301 if( klass->is_obj_array_klass() && 2302 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2303 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2304 // // Always returning precise element type is incorrect, 2305 // // e.g., element type could be object and array may contain strings 2306 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2307 2308 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2309 // according to the element type's subclassing. 2310 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0)); 2311 } else if (klass->is_value_array_klass() && 2312 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2313 ciKlass* elem = klass->as_value_array_klass()->element_klass(); 2314 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0)); 2315 } 2316 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2317 tkls->offset() == in_bytes(Klass::super_offset())) { 2318 ciKlass* sup = klass->as_instance_klass()->super(); 2319 // The field is Klass::_super. Return its (constant) value. 2320 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2321 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2322 } 2323 } 2324 2325 // Bailout case 2326 return LoadNode::Value(phase); 2327 } 2328 2329 //------------------------------Identity--------------------------------------- 2330 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2331 // Also feed through the klass in Allocate(...klass...)._klass. 2332 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2333 return klass_identity_common(phase); 2334 } 2335 2336 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2337 Node* x = LoadNode::Identity(phase); 2338 if (x != this) return x; 2339 2340 // Take apart the address into an oop and and offset. 2341 // Return 'this' if we cannot. 2342 Node* adr = in(MemNode::Address); 2343 intptr_t offset = 0; 2344 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2345 if (base == NULL) return this; 2346 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2347 if (toop == NULL) return this; 2348 2349 // Step over potential GC barrier for OopHandle resolve 2350 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2351 if (bs->is_gc_barrier_node(base)) { 2352 base = bs->step_over_gc_barrier(base); 2353 } 2354 2355 // We can fetch the klass directly through an AllocateNode. 2356 // This works even if the klass is not constant (clone or newArray). 2357 if (offset == oopDesc::klass_offset_in_bytes()) { 2358 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2359 if (allocated_klass != NULL) { 2360 return allocated_klass; 2361 } 2362 } 2363 2364 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2365 // See inline_native_Class_query for occurrences of these patterns. 2366 // Java Example: x.getClass().isAssignableFrom(y) 2367 // 2368 // This improves reflective code, often making the Class 2369 // mirror go completely dead. (Current exception: Class 2370 // mirrors may appear in debug info, but we could clean them out by 2371 // introducing a new debug info operator for Klass.java_mirror). 2372 2373 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2374 && offset == java_lang_Class::klass_offset_in_bytes()) { 2375 if (base->is_Load()) { 2376 Node* base2 = base->in(MemNode::Address); 2377 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2378 Node* adr2 = base2->in(MemNode::Address); 2379 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2380 if (tkls != NULL && !tkls->empty() 2381 && (tkls->klass()->is_instance_klass() || 2382 tkls->klass()->is_array_klass()) 2383 && adr2->is_AddP() 2384 ) { 2385 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2386 if (tkls->offset() == mirror_field) { 2387 return adr2->in(AddPNode::Base); 2388 } 2389 } 2390 } 2391 } 2392 } 2393 2394 return this; 2395 } 2396 2397 2398 //------------------------------Value------------------------------------------ 2399 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2400 const Type *t = klass_value_common(phase); 2401 if (t == Type::TOP) 2402 return t; 2403 2404 return t->make_narrowklass(); 2405 } 2406 2407 //------------------------------Identity--------------------------------------- 2408 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2409 // Also feed through the klass in Allocate(...klass...)._klass. 2410 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2411 Node *x = klass_identity_common(phase); 2412 2413 const Type *t = phase->type( x ); 2414 if( t == Type::TOP ) return x; 2415 if( t->isa_narrowklass()) return x; 2416 assert (!t->isa_narrowoop(), "no narrow oop here"); 2417 2418 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2419 } 2420 2421 //------------------------------Value----------------------------------------- 2422 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2423 // Either input is TOP ==> the result is TOP 2424 const Type *t1 = phase->type( in(MemNode::Memory) ); 2425 if( t1 == Type::TOP ) return Type::TOP; 2426 Node *adr = in(MemNode::Address); 2427 const Type *t2 = phase->type( adr ); 2428 if( t2 == Type::TOP ) return Type::TOP; 2429 const TypePtr *tp = t2->is_ptr(); 2430 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2431 const TypeAryPtr *tap = tp->isa_aryptr(); 2432 if( !tap ) return _type; 2433 return tap->size(); 2434 } 2435 2436 //-------------------------------Ideal--------------------------------------- 2437 // Feed through the length in AllocateArray(...length...)._length. 2438 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2439 Node* p = MemNode::Ideal_common(phase, can_reshape); 2440 if (p) return (p == NodeSentinel) ? NULL : p; 2441 2442 // Take apart the address into an oop and and offset. 2443 // Return 'this' if we cannot. 2444 Node* adr = in(MemNode::Address); 2445 intptr_t offset = 0; 2446 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2447 if (base == NULL) return NULL; 2448 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2449 if (tary == NULL) return NULL; 2450 2451 // We can fetch the length directly through an AllocateArrayNode. 2452 // This works even if the length is not constant (clone or newArray). 2453 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2454 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2455 if (alloc != NULL) { 2456 Node* allocated_length = alloc->Ideal_length(); 2457 Node* len = alloc->make_ideal_length(tary, phase); 2458 if (allocated_length != len) { 2459 // New CastII improves on this. 2460 return len; 2461 } 2462 } 2463 } 2464 2465 return NULL; 2466 } 2467 2468 //------------------------------Identity--------------------------------------- 2469 // Feed through the length in AllocateArray(...length...)._length. 2470 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2471 Node* x = LoadINode::Identity(phase); 2472 if (x != this) return x; 2473 2474 // Take apart the address into an oop and and offset. 2475 // Return 'this' if we cannot. 2476 Node* adr = in(MemNode::Address); 2477 intptr_t offset = 0; 2478 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2479 if (base == NULL) return this; 2480 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2481 if (tary == NULL) return this; 2482 2483 // We can fetch the length directly through an AllocateArrayNode. 2484 // This works even if the length is not constant (clone or newArray). 2485 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2486 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2487 if (alloc != NULL) { 2488 Node* allocated_length = alloc->Ideal_length(); 2489 // Do not allow make_ideal_length to allocate a CastII node. 2490 Node* len = alloc->make_ideal_length(tary, phase, false); 2491 if (allocated_length == len) { 2492 // Return allocated_length only if it would not be improved by a CastII. 2493 return allocated_length; 2494 } 2495 } 2496 } 2497 2498 return this; 2499 2500 } 2501 2502 //============================================================================= 2503 //---------------------------StoreNode::make----------------------------------- 2504 // Polymorphic factory method: 2505 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2506 assert((mo == unordered || mo == release), "unexpected"); 2507 Compile* C = gvn.C; 2508 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2509 ctl != NULL, "raw memory operations should have control edge"); 2510 2511 switch (bt) { 2512 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2513 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2514 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2515 case T_CHAR: 2516 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2517 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2518 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2519 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2520 case T_METADATA: 2521 case T_ADDRESS: 2522 case T_VALUETYPE: 2523 case T_OBJECT: 2524 #ifdef _LP64 2525 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2526 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2527 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2528 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2529 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2530 adr->bottom_type()->isa_rawptr())) { 2531 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2532 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2533 } 2534 #endif 2535 { 2536 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2537 } 2538 default: 2539 ShouldNotReachHere(); 2540 return (StoreNode*)NULL; 2541 } 2542 } 2543 2544 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2545 bool require_atomic = true; 2546 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2547 } 2548 2549 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2550 bool require_atomic = true; 2551 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2552 } 2553 2554 2555 //--------------------------bottom_type---------------------------------------- 2556 const Type *StoreNode::bottom_type() const { 2557 return Type::MEMORY; 2558 } 2559 2560 //------------------------------hash------------------------------------------- 2561 uint StoreNode::hash() const { 2562 // unroll addition of interesting fields 2563 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2564 2565 // Since they are not commoned, do not hash them: 2566 return NO_HASH; 2567 } 2568 2569 //------------------------------Ideal------------------------------------------ 2570 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2571 // When a store immediately follows a relevant allocation/initialization, 2572 // try to capture it into the initialization, or hoist it above. 2573 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2574 Node* p = MemNode::Ideal_common(phase, can_reshape); 2575 if (p) return (p == NodeSentinel) ? NULL : p; 2576 2577 Node* mem = in(MemNode::Memory); 2578 Node* address = in(MemNode::Address); 2579 // Back-to-back stores to same address? Fold em up. Generally 2580 // unsafe if I have intervening uses... Also disallowed for StoreCM 2581 // since they must follow each StoreP operation. Redundant StoreCMs 2582 // are eliminated just before matching in final_graph_reshape. 2583 { 2584 Node* st = mem; 2585 // If Store 'st' has more than one use, we cannot fold 'st' away. 2586 // For example, 'st' might be the final state at a conditional 2587 // return. Or, 'st' might be used by some node which is live at 2588 // the same time 'st' is live, which might be unschedulable. So, 2589 // require exactly ONE user until such time as we clone 'mem' for 2590 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2591 // true). 2592 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2593 // Looking at a dead closed cycle of memory? 2594 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2595 assert(Opcode() == st->Opcode() || 2596 st->Opcode() == Op_StoreVector || 2597 Opcode() == Op_StoreVector || 2598 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2599 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2600 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2601 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) || 2602 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2603 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2604 2605 if (st->in(MemNode::Address)->eqv_uncast(address) && 2606 st->as_Store()->memory_size() <= this->memory_size()) { 2607 Node* use = st->raw_out(0); 2608 phase->igvn_rehash_node_delayed(use); 2609 if (can_reshape) { 2610 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2611 } else { 2612 // It's OK to do this in the parser, since DU info is always accurate, 2613 // and the parser always refers to nodes via SafePointNode maps. 2614 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2615 } 2616 return this; 2617 } 2618 st = st->in(MemNode::Memory); 2619 } 2620 } 2621 2622 2623 // Capture an unaliased, unconditional, simple store into an initializer. 2624 // Or, if it is independent of the allocation, hoist it above the allocation. 2625 if (ReduceFieldZeroing && /*can_reshape &&*/ 2626 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2627 InitializeNode* init = mem->in(0)->as_Initialize(); 2628 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2629 if (offset > 0) { 2630 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2631 // If the InitializeNode captured me, it made a raw copy of me, 2632 // and I need to disappear. 2633 if (moved != NULL) { 2634 // %%% hack to ensure that Ideal returns a new node: 2635 mem = MergeMemNode::make(mem); 2636 return mem; // fold me away 2637 } 2638 } 2639 } 2640 2641 return NULL; // No further progress 2642 } 2643 2644 //------------------------------Value----------------------------------------- 2645 const Type* StoreNode::Value(PhaseGVN* phase) const { 2646 // Either input is TOP ==> the result is TOP 2647 const Type *t1 = phase->type( in(MemNode::Memory) ); 2648 if( t1 == Type::TOP ) return Type::TOP; 2649 const Type *t2 = phase->type( in(MemNode::Address) ); 2650 if( t2 == Type::TOP ) return Type::TOP; 2651 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2652 if( t3 == Type::TOP ) return Type::TOP; 2653 return Type::MEMORY; 2654 } 2655 2656 //------------------------------Identity--------------------------------------- 2657 // Remove redundant stores: 2658 // Store(m, p, Load(m, p)) changes to m. 2659 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2660 Node* StoreNode::Identity(PhaseGVN* phase) { 2661 Node* mem = in(MemNode::Memory); 2662 Node* adr = in(MemNode::Address); 2663 Node* val = in(MemNode::ValueIn); 2664 2665 Node* result = this; 2666 2667 // Load then Store? Then the Store is useless 2668 if (val->is_Load() && 2669 val->in(MemNode::Address)->eqv_uncast(adr) && 2670 val->in(MemNode::Memory )->eqv_uncast(mem) && 2671 val->as_Load()->store_Opcode() == Opcode()) { 2672 result = mem; 2673 } 2674 2675 // Two stores in a row of the same value? 2676 if (result == this && 2677 mem->is_Store() && 2678 mem->in(MemNode::Address)->eqv_uncast(adr) && 2679 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2680 mem->Opcode() == Opcode()) { 2681 result = mem; 2682 } 2683 2684 // Store of zero anywhere into a freshly-allocated object? 2685 // Then the store is useless. 2686 // (It must already have been captured by the InitializeNode.) 2687 if (result == this && ReduceFieldZeroing) { 2688 // a newly allocated object is already all-zeroes everywhere 2689 if (mem->is_Proj() && mem->in(0)->is_Allocate() && 2690 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) { 2691 assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to value array is forbidden"); 2692 result = mem; 2693 } 2694 2695 if (result == this) { 2696 // the store may also apply to zero-bits in an earlier object 2697 Node* prev_mem = find_previous_store(phase); 2698 // Steps (a), (b): Walk past independent stores to find an exact match. 2699 if (prev_mem != NULL) { 2700 Node* prev_val = can_see_stored_value(prev_mem, phase); 2701 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2702 // prev_val and val might differ by a cast; it would be good 2703 // to keep the more informative of the two. 2704 if (phase->type(val)->is_zero_type()) { 2705 result = mem; 2706 } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) { 2707 InitializeNode* init = prev_mem->in(0)->as_Initialize(); 2708 AllocateNode* alloc = init->allocation(); 2709 if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) { 2710 result = mem; 2711 } 2712 } 2713 } 2714 } 2715 } 2716 } 2717 2718 if (result != this && phase->is_IterGVN() != NULL) { 2719 MemBarNode* trailing = trailing_membar(); 2720 if (trailing != NULL) { 2721 #ifdef ASSERT 2722 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2723 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects"); 2724 #endif 2725 PhaseIterGVN* igvn = phase->is_IterGVN(); 2726 trailing->remove(igvn); 2727 } 2728 } 2729 2730 return result; 2731 } 2732 2733 //------------------------------match_edge------------------------------------- 2734 // Do we Match on this edge index or not? Match only memory & value 2735 uint StoreNode::match_edge(uint idx) const { 2736 return idx == MemNode::Address || idx == MemNode::ValueIn; 2737 } 2738 2739 //------------------------------cmp-------------------------------------------- 2740 // Do not common stores up together. They generally have to be split 2741 // back up anyways, so do not bother. 2742 bool StoreNode::cmp( const Node &n ) const { 2743 return (&n == this); // Always fail except on self 2744 } 2745 2746 //------------------------------Ideal_masked_input----------------------------- 2747 // Check for a useless mask before a partial-word store 2748 // (StoreB ... (AndI valIn conIa) ) 2749 // If (conIa & mask == mask) this simplifies to 2750 // (StoreB ... (valIn) ) 2751 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2752 Node *val = in(MemNode::ValueIn); 2753 if( val->Opcode() == Op_AndI ) { 2754 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2755 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2756 set_req(MemNode::ValueIn, val->in(1)); 2757 return this; 2758 } 2759 } 2760 return NULL; 2761 } 2762 2763 2764 //------------------------------Ideal_sign_extended_input---------------------- 2765 // Check for useless sign-extension before a partial-word store 2766 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2767 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2768 // (StoreB ... (valIn) ) 2769 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2770 Node *val = in(MemNode::ValueIn); 2771 if( val->Opcode() == Op_RShiftI ) { 2772 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2773 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2774 Node *shl = val->in(1); 2775 if( shl->Opcode() == Op_LShiftI ) { 2776 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2777 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2778 set_req(MemNode::ValueIn, shl->in(1)); 2779 return this; 2780 } 2781 } 2782 } 2783 } 2784 return NULL; 2785 } 2786 2787 //------------------------------value_never_loaded----------------------------------- 2788 // Determine whether there are any possible loads of the value stored. 2789 // For simplicity, we actually check if there are any loads from the 2790 // address stored to, not just for loads of the value stored by this node. 2791 // 2792 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2793 Node *adr = in(Address); 2794 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2795 if (adr_oop == NULL) 2796 return false; 2797 if (!adr_oop->is_known_instance_field()) 2798 return false; // if not a distinct instance, there may be aliases of the address 2799 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2800 Node *use = adr->fast_out(i); 2801 if (use->is_Load() || use->is_LoadStore()) { 2802 return false; 2803 } 2804 } 2805 return true; 2806 } 2807 2808 MemBarNode* StoreNode::trailing_membar() const { 2809 if (is_release()) { 2810 MemBarNode* trailing_mb = NULL; 2811 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2812 Node* u = fast_out(i); 2813 if (u->is_MemBar()) { 2814 if (u->as_MemBar()->trailing_store()) { 2815 assert(u->Opcode() == Op_MemBarVolatile, ""); 2816 assert(trailing_mb == NULL, "only one"); 2817 trailing_mb = u->as_MemBar(); 2818 #ifdef ASSERT 2819 Node* leading = u->as_MemBar()->leading_membar(); 2820 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2821 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2822 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2823 #endif 2824 } else { 2825 assert(u->as_MemBar()->standalone(), ""); 2826 } 2827 } 2828 } 2829 return trailing_mb; 2830 } 2831 return NULL; 2832 } 2833 2834 2835 //============================================================================= 2836 //------------------------------Ideal------------------------------------------ 2837 // If the store is from an AND mask that leaves the low bits untouched, then 2838 // we can skip the AND operation. If the store is from a sign-extension 2839 // (a left shift, then right shift) we can skip both. 2840 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2841 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2842 if( progress != NULL ) return progress; 2843 2844 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2845 if( progress != NULL ) return progress; 2846 2847 // Finally check the default case 2848 return StoreNode::Ideal(phase, can_reshape); 2849 } 2850 2851 //============================================================================= 2852 //------------------------------Ideal------------------------------------------ 2853 // If the store is from an AND mask that leaves the low bits untouched, then 2854 // we can skip the AND operation 2855 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2856 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2857 if( progress != NULL ) return progress; 2858 2859 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2860 if( progress != NULL ) return progress; 2861 2862 // Finally check the default case 2863 return StoreNode::Ideal(phase, can_reshape); 2864 } 2865 2866 //============================================================================= 2867 //------------------------------Identity--------------------------------------- 2868 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2869 // No need to card mark when storing a null ptr 2870 Node* my_store = in(MemNode::OopStore); 2871 if (my_store->is_Store()) { 2872 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2873 if( t1 == TypePtr::NULL_PTR ) { 2874 return in(MemNode::Memory); 2875 } 2876 } 2877 return this; 2878 } 2879 2880 //============================================================================= 2881 //------------------------------Ideal--------------------------------------- 2882 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2883 Node* progress = StoreNode::Ideal(phase, can_reshape); 2884 if (progress != NULL) return progress; 2885 2886 Node* my_store = in(MemNode::OopStore); 2887 if (my_store->is_MergeMem()) { 2888 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2889 set_req(MemNode::OopStore, mem); 2890 return this; 2891 } 2892 2893 return NULL; 2894 } 2895 2896 //------------------------------Value----------------------------------------- 2897 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2898 // Either input is TOP ==> the result is TOP 2899 const Type *t = phase->type( in(MemNode::Memory) ); 2900 if( t == Type::TOP ) return Type::TOP; 2901 t = phase->type( in(MemNode::Address) ); 2902 if( t == Type::TOP ) return Type::TOP; 2903 t = phase->type( in(MemNode::ValueIn) ); 2904 if( t == Type::TOP ) return Type::TOP; 2905 // If extra input is TOP ==> the result is TOP 2906 t = phase->type( in(MemNode::OopStore) ); 2907 if( t == Type::TOP ) return Type::TOP; 2908 2909 return StoreNode::Value( phase ); 2910 } 2911 2912 2913 //============================================================================= 2914 //----------------------------------SCMemProjNode------------------------------ 2915 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2916 { 2917 return bottom_type(); 2918 } 2919 2920 //============================================================================= 2921 //----------------------------------LoadStoreNode------------------------------ 2922 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2923 : Node(required), 2924 _type(rt), 2925 _adr_type(at) 2926 { 2927 init_req(MemNode::Control, c ); 2928 init_req(MemNode::Memory , mem); 2929 init_req(MemNode::Address, adr); 2930 init_req(MemNode::ValueIn, val); 2931 init_class_id(Class_LoadStore); 2932 } 2933 2934 uint LoadStoreNode::ideal_reg() const { 2935 return _type->ideal_reg(); 2936 } 2937 2938 bool LoadStoreNode::result_not_used() const { 2939 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2940 Node *x = fast_out(i); 2941 if (x->Opcode() == Op_SCMemProj) continue; 2942 return false; 2943 } 2944 return true; 2945 } 2946 2947 MemBarNode* LoadStoreNode::trailing_membar() const { 2948 MemBarNode* trailing = NULL; 2949 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2950 Node* u = fast_out(i); 2951 if (u->is_MemBar()) { 2952 if (u->as_MemBar()->trailing_load_store()) { 2953 assert(u->Opcode() == Op_MemBarAcquire, ""); 2954 assert(trailing == NULL, "only one"); 2955 trailing = u->as_MemBar(); 2956 #ifdef ASSERT 2957 Node* leading = trailing->leading_membar(); 2958 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2959 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 2960 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 2961 #endif 2962 } else { 2963 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 2964 } 2965 } 2966 } 2967 2968 return trailing; 2969 } 2970 2971 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2972 2973 //============================================================================= 2974 //----------------------------------LoadStoreConditionalNode-------------------- 2975 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2976 init_req(ExpectedIn, ex ); 2977 } 2978 2979 //============================================================================= 2980 //-------------------------------adr_type-------------------------------------- 2981 const TypePtr* ClearArrayNode::adr_type() const { 2982 Node *adr = in(3); 2983 if (adr == NULL) return NULL; // node is dead 2984 return MemNode::calculate_adr_type(adr->bottom_type()); 2985 } 2986 2987 //------------------------------match_edge------------------------------------- 2988 // Do we Match on this edge index or not? Do not match memory 2989 uint ClearArrayNode::match_edge(uint idx) const { 2990 return idx > 1; 2991 } 2992 2993 //------------------------------Identity--------------------------------------- 2994 // Clearing a zero length array does nothing 2995 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2996 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2997 } 2998 2999 //------------------------------Idealize--------------------------------------- 3000 // Clearing a short array is faster with stores 3001 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3002 // Already know this is a large node, do not try to ideal it 3003 if (!IdealizeClearArrayNode || _is_large) return NULL; 3004 3005 const int unit = BytesPerLong; 3006 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 3007 if (!t) return NULL; 3008 if (!t->is_con()) return NULL; 3009 intptr_t raw_count = t->get_con(); 3010 intptr_t size = raw_count; 3011 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 3012 // Clearing nothing uses the Identity call. 3013 // Negative clears are possible on dead ClearArrays 3014 // (see jck test stmt114.stmt11402.val). 3015 if (size <= 0 || size % unit != 0) return NULL; 3016 intptr_t count = size / unit; 3017 // Length too long; communicate this to matchers and assemblers. 3018 // Assemblers are responsible to produce fast hardware clears for it. 3019 if (size > InitArrayShortSize) { 3020 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true); 3021 } 3022 Node *mem = in(1); 3023 if( phase->type(mem)==Type::TOP ) return NULL; 3024 Node *adr = in(3); 3025 const Type* at = phase->type(adr); 3026 if( at==Type::TOP ) return NULL; 3027 const TypePtr* atp = at->isa_ptr(); 3028 // adjust atp to be the correct array element address type 3029 if (atp == NULL) atp = TypePtr::BOTTOM; 3030 else atp = atp->add_offset(Type::OffsetBot); 3031 // Get base for derived pointer purposes 3032 if( adr->Opcode() != Op_AddP ) Unimplemented(); 3033 Node *base = adr->in(1); 3034 3035 Node *val = in(4); 3036 Node *off = phase->MakeConX(BytesPerLong); 3037 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3038 count--; 3039 while( count-- ) { 3040 mem = phase->transform(mem); 3041 adr = phase->transform(new AddPNode(base,adr,off)); 3042 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3043 } 3044 return mem; 3045 } 3046 3047 //----------------------------step_through---------------------------------- 3048 // Return allocation input memory edge if it is different instance 3049 // or itself if it is the one we are looking for. 3050 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 3051 Node* n = *np; 3052 assert(n->is_ClearArray(), "sanity"); 3053 intptr_t offset; 3054 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 3055 // This method is called only before Allocate nodes are expanded 3056 // during macro nodes expansion. Before that ClearArray nodes are 3057 // only generated in PhaseMacroExpand::generate_arraycopy() (before 3058 // Allocate nodes are expanded) which follows allocations. 3059 assert(alloc != NULL, "should have allocation"); 3060 if (alloc->_idx == instance_id) { 3061 // Can not bypass initialization of the instance we are looking for. 3062 return false; 3063 } 3064 // Otherwise skip it. 3065 InitializeNode* init = alloc->initialization(); 3066 if (init != NULL) 3067 *np = init->in(TypeFunc::Memory); 3068 else 3069 *np = alloc->in(TypeFunc::Memory); 3070 return true; 3071 } 3072 3073 //----------------------------clear_memory------------------------------------- 3074 // Generate code to initialize object storage to zero. 3075 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3076 Node* val, 3077 Node* raw_val, 3078 intptr_t start_offset, 3079 Node* end_offset, 3080 PhaseGVN* phase) { 3081 intptr_t offset = start_offset; 3082 3083 int unit = BytesPerLong; 3084 if ((offset % unit) != 0) { 3085 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 3086 adr = phase->transform(adr); 3087 const TypePtr* atp = TypeRawPtr::BOTTOM; 3088 if (val != NULL) { 3089 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3090 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3091 } else { 3092 assert(raw_val == NULL, "val may not be null"); 3093 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3094 } 3095 mem = phase->transform(mem); 3096 offset += BytesPerInt; 3097 } 3098 assert((offset % unit) == 0, ""); 3099 3100 // Initialize the remaining stuff, if any, with a ClearArray. 3101 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase); 3102 } 3103 3104 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3105 Node* raw_val, 3106 Node* start_offset, 3107 Node* end_offset, 3108 PhaseGVN* phase) { 3109 if (start_offset == end_offset) { 3110 // nothing to do 3111 return mem; 3112 } 3113 3114 int unit = BytesPerLong; 3115 Node* zbase = start_offset; 3116 Node* zend = end_offset; 3117 3118 // Scale to the unit required by the CPU: 3119 if (!Matcher::init_array_count_is_in_bytes) { 3120 Node* shift = phase->intcon(exact_log2(unit)); 3121 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3122 zend = phase->transform(new URShiftXNode(zend, shift) ); 3123 } 3124 3125 // Bulk clear double-words 3126 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3127 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3128 if (raw_val == NULL) { 3129 raw_val = phase->MakeConX(0); 3130 } 3131 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false); 3132 return phase->transform(mem); 3133 } 3134 3135 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3136 Node* val, 3137 Node* raw_val, 3138 intptr_t start_offset, 3139 intptr_t end_offset, 3140 PhaseGVN* phase) { 3141 if (start_offset == end_offset) { 3142 // nothing to do 3143 return mem; 3144 } 3145 3146 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3147 intptr_t done_offset = end_offset; 3148 if ((done_offset % BytesPerLong) != 0) { 3149 done_offset -= BytesPerInt; 3150 } 3151 if (done_offset > start_offset) { 3152 mem = clear_memory(ctl, mem, dest, val, raw_val, 3153 start_offset, phase->MakeConX(done_offset), phase); 3154 } 3155 if (done_offset < end_offset) { // emit the final 32-bit store 3156 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3157 adr = phase->transform(adr); 3158 const TypePtr* atp = TypeRawPtr::BOTTOM; 3159 if (val != NULL) { 3160 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3161 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3162 } else { 3163 assert(raw_val == NULL, "val may not be null"); 3164 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3165 } 3166 mem = phase->transform(mem); 3167 done_offset += BytesPerInt; 3168 } 3169 assert(done_offset == end_offset, ""); 3170 return mem; 3171 } 3172 3173 //============================================================================= 3174 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3175 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 3176 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3177 #ifdef ASSERT 3178 , _pair_idx(0) 3179 #endif 3180 { 3181 init_class_id(Class_MemBar); 3182 Node* top = C->top(); 3183 init_req(TypeFunc::I_O,top); 3184 init_req(TypeFunc::FramePtr,top); 3185 init_req(TypeFunc::ReturnAdr,top); 3186 if (precedent != NULL) 3187 init_req(TypeFunc::Parms, precedent); 3188 } 3189 3190 //------------------------------cmp-------------------------------------------- 3191 uint MemBarNode::hash() const { return NO_HASH; } 3192 bool MemBarNode::cmp( const Node &n ) const { 3193 return (&n == this); // Always fail except on self 3194 } 3195 3196 //------------------------------make------------------------------------------- 3197 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3198 switch (opcode) { 3199 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3200 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3201 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3202 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3203 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3204 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3205 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3206 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3207 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3208 case Op_Initialize: return new InitializeNode(C, atp, pn); 3209 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3210 default: ShouldNotReachHere(); return NULL; 3211 } 3212 } 3213 3214 void MemBarNode::remove(PhaseIterGVN *igvn) { 3215 if (outcnt() != 2) { 3216 return; 3217 } 3218 if (trailing_store() || trailing_load_store()) { 3219 MemBarNode* leading = leading_membar(); 3220 if (leading != NULL) { 3221 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3222 leading->remove(igvn); 3223 } 3224 } 3225 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3226 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3227 } 3228 3229 //------------------------------Ideal------------------------------------------ 3230 // Return a node which is more "ideal" than the current node. Strip out 3231 // control copies 3232 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3233 if (remove_dead_region(phase, can_reshape)) return this; 3234 // Don't bother trying to transform a dead node 3235 if (in(0) && in(0)->is_top()) { 3236 return NULL; 3237 } 3238 3239 #if INCLUDE_ZGC 3240 if (UseZGC) { 3241 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) { 3242 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop); 3243 set_req(MemBarNode::Precedent, load_node); 3244 return this; 3245 } 3246 } 3247 #endif 3248 3249 bool progress = false; 3250 // Eliminate volatile MemBars for scalar replaced objects. 3251 if (can_reshape && req() == (Precedent+1)) { 3252 bool eliminate = false; 3253 int opc = Opcode(); 3254 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3255 // Volatile field loads and stores. 3256 Node* my_mem = in(MemBarNode::Precedent); 3257 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3258 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3259 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3260 // replace this Precedent (decodeN) with the Load instead. 3261 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3262 Node* load_node = my_mem->in(1); 3263 set_req(MemBarNode::Precedent, load_node); 3264 phase->is_IterGVN()->_worklist.push(my_mem); 3265 my_mem = load_node; 3266 } else { 3267 assert(my_mem->unique_out() == this, "sanity"); 3268 del_req(Precedent); 3269 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3270 my_mem = NULL; 3271 } 3272 progress = true; 3273 } 3274 if (my_mem != NULL && my_mem->is_Mem()) { 3275 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3276 // Check for scalar replaced object reference. 3277 if( t_oop != NULL && t_oop->is_known_instance_field() && 3278 t_oop->offset() != Type::OffsetBot && 3279 t_oop->offset() != Type::OffsetTop) { 3280 eliminate = true; 3281 } 3282 } 3283 } else if (opc == Op_MemBarRelease) { 3284 // Final field stores. 3285 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3286 if ((alloc != NULL) && alloc->is_Allocate() && 3287 alloc->as_Allocate()->does_not_escape_thread()) { 3288 // The allocated object does not escape. 3289 eliminate = true; 3290 } 3291 } 3292 if (eliminate) { 3293 // Replace MemBar projections by its inputs. 3294 PhaseIterGVN* igvn = phase->is_IterGVN(); 3295 remove(igvn); 3296 // Must return either the original node (now dead) or a new node 3297 // (Do not return a top here, since that would break the uniqueness of top.) 3298 return new ConINode(TypeInt::ZERO); 3299 } 3300 } 3301 return progress ? this : NULL; 3302 } 3303 3304 //------------------------------Value------------------------------------------ 3305 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3306 if( !in(0) ) return Type::TOP; 3307 if( phase->type(in(0)) == Type::TOP ) 3308 return Type::TOP; 3309 return TypeTuple::MEMBAR; 3310 } 3311 3312 //------------------------------match------------------------------------------ 3313 // Construct projections for memory. 3314 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) { 3315 switch (proj->_con) { 3316 case TypeFunc::Control: 3317 case TypeFunc::Memory: 3318 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3319 } 3320 ShouldNotReachHere(); 3321 return NULL; 3322 } 3323 3324 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3325 trailing->_kind = TrailingStore; 3326 leading->_kind = LeadingStore; 3327 #ifdef ASSERT 3328 trailing->_pair_idx = leading->_idx; 3329 leading->_pair_idx = leading->_idx; 3330 #endif 3331 } 3332 3333 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3334 trailing->_kind = TrailingLoadStore; 3335 leading->_kind = LeadingLoadStore; 3336 #ifdef ASSERT 3337 trailing->_pair_idx = leading->_idx; 3338 leading->_pair_idx = leading->_idx; 3339 #endif 3340 } 3341 3342 MemBarNode* MemBarNode::trailing_membar() const { 3343 ResourceMark rm; 3344 Node* trailing = (Node*)this; 3345 VectorSet seen(Thread::current()->resource_area()); 3346 Node_Stack multis(0); 3347 do { 3348 Node* c = trailing; 3349 uint i = 0; 3350 do { 3351 trailing = NULL; 3352 for (; i < c->outcnt(); i++) { 3353 Node* next = c->raw_out(i); 3354 if (next != c && next->is_CFG()) { 3355 if (c->is_MultiBranch()) { 3356 if (multis.node() == c) { 3357 multis.set_index(i+1); 3358 } else { 3359 multis.push(c, i+1); 3360 } 3361 } 3362 trailing = next; 3363 break; 3364 } 3365 } 3366 if (trailing != NULL && !seen.test_set(trailing->_idx)) { 3367 break; 3368 } 3369 while (multis.size() > 0) { 3370 c = multis.node(); 3371 i = multis.index(); 3372 if (i < c->req()) { 3373 break; 3374 } 3375 multis.pop(); 3376 } 3377 } while (multis.size() > 0); 3378 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3379 3380 MemBarNode* mb = trailing->as_MemBar(); 3381 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3382 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3383 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3384 return mb; 3385 } 3386 3387 MemBarNode* MemBarNode::leading_membar() const { 3388 ResourceMark rm; 3389 VectorSet seen(Thread::current()->resource_area()); 3390 Node_Stack regions(0); 3391 Node* leading = in(0); 3392 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3393 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) { 3394 leading = NULL; 3395 while (regions.size() > 0 && leading == NULL) { 3396 Node* r = regions.node(); 3397 uint i = regions.index(); 3398 if (i < r->req()) { 3399 leading = r->in(i); 3400 regions.set_index(i+1); 3401 } else { 3402 regions.pop(); 3403 } 3404 } 3405 if (leading == NULL) { 3406 assert(regions.size() == 0, "all paths should have been tried"); 3407 return NULL; 3408 } 3409 } 3410 if (leading->is_Region()) { 3411 regions.push(leading, 2); 3412 leading = leading->in(1); 3413 } else { 3414 leading = leading->in(0); 3415 } 3416 } 3417 #ifdef ASSERT 3418 Unique_Node_List wq; 3419 wq.push((Node*)this); 3420 uint found = 0; 3421 for (uint i = 0; i < wq.size(); i++) { 3422 Node* n = wq.at(i); 3423 if (n->is_Region()) { 3424 for (uint j = 1; j < n->req(); j++) { 3425 Node* in = n->in(j); 3426 if (in != NULL && !in->is_top()) { 3427 wq.push(in); 3428 } 3429 } 3430 } else { 3431 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3432 assert(n == leading, "consistency check failed"); 3433 found++; 3434 } else { 3435 Node* in = n->in(0); 3436 if (in != NULL && !in->is_top()) { 3437 wq.push(in); 3438 } 3439 } 3440 } 3441 } 3442 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed"); 3443 #endif 3444 if (leading == NULL) { 3445 return NULL; 3446 } 3447 MemBarNode* mb = leading->as_MemBar(); 3448 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3449 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3450 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3451 return mb; 3452 } 3453 3454 //===========================InitializeNode==================================== 3455 // SUMMARY: 3456 // This node acts as a memory barrier on raw memory, after some raw stores. 3457 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3458 // The Initialize can 'capture' suitably constrained stores as raw inits. 3459 // It can coalesce related raw stores into larger units (called 'tiles'). 3460 // It can avoid zeroing new storage for memory units which have raw inits. 3461 // At macro-expansion, it is marked 'complete', and does not optimize further. 3462 // 3463 // EXAMPLE: 3464 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3465 // ctl = incoming control; mem* = incoming memory 3466 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3467 // First allocate uninitialized memory and fill in the header: 3468 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3469 // ctl := alloc.Control; mem* := alloc.Memory* 3470 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3471 // Then initialize to zero the non-header parts of the raw memory block: 3472 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3473 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3474 // After the initialize node executes, the object is ready for service: 3475 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3476 // Suppose its body is immediately initialized as {1,2}: 3477 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3478 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3479 // mem.SLICE(#short[*]) := store2 3480 // 3481 // DETAILS: 3482 // An InitializeNode collects and isolates object initialization after 3483 // an AllocateNode and before the next possible safepoint. As a 3484 // memory barrier (MemBarNode), it keeps critical stores from drifting 3485 // down past any safepoint or any publication of the allocation. 3486 // Before this barrier, a newly-allocated object may have uninitialized bits. 3487 // After this barrier, it may be treated as a real oop, and GC is allowed. 3488 // 3489 // The semantics of the InitializeNode include an implicit zeroing of 3490 // the new object from object header to the end of the object. 3491 // (The object header and end are determined by the AllocateNode.) 3492 // 3493 // Certain stores may be added as direct inputs to the InitializeNode. 3494 // These stores must update raw memory, and they must be to addresses 3495 // derived from the raw address produced by AllocateNode, and with 3496 // a constant offset. They must be ordered by increasing offset. 3497 // The first one is at in(RawStores), the last at in(req()-1). 3498 // Unlike most memory operations, they are not linked in a chain, 3499 // but are displayed in parallel as users of the rawmem output of 3500 // the allocation. 3501 // 3502 // (See comments in InitializeNode::capture_store, which continue 3503 // the example given above.) 3504 // 3505 // When the associated Allocate is macro-expanded, the InitializeNode 3506 // may be rewritten to optimize collected stores. A ClearArrayNode 3507 // may also be created at that point to represent any required zeroing. 3508 // The InitializeNode is then marked 'complete', prohibiting further 3509 // capturing of nearby memory operations. 3510 // 3511 // During macro-expansion, all captured initializations which store 3512 // constant values of 32 bits or smaller are coalesced (if advantageous) 3513 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3514 // initialized in fewer memory operations. Memory words which are 3515 // covered by neither tiles nor non-constant stores are pre-zeroed 3516 // by explicit stores of zero. (The code shape happens to do all 3517 // zeroing first, then all other stores, with both sequences occurring 3518 // in order of ascending offsets.) 3519 // 3520 // Alternatively, code may be inserted between an AllocateNode and its 3521 // InitializeNode, to perform arbitrary initialization of the new object. 3522 // E.g., the object copying intrinsics insert complex data transfers here. 3523 // The initialization must then be marked as 'complete' disable the 3524 // built-in zeroing semantics and the collection of initializing stores. 3525 // 3526 // While an InitializeNode is incomplete, reads from the memory state 3527 // produced by it are optimizable if they match the control edge and 3528 // new oop address associated with the allocation/initialization. 3529 // They return a stored value (if the offset matches) or else zero. 3530 // A write to the memory state, if it matches control and address, 3531 // and if it is to a constant offset, may be 'captured' by the 3532 // InitializeNode. It is cloned as a raw memory operation and rewired 3533 // inside the initialization, to the raw oop produced by the allocation. 3534 // Operations on addresses which are provably distinct (e.g., to 3535 // other AllocateNodes) are allowed to bypass the initialization. 3536 // 3537 // The effect of all this is to consolidate object initialization 3538 // (both arrays and non-arrays, both piecewise and bulk) into a 3539 // single location, where it can be optimized as a unit. 3540 // 3541 // Only stores with an offset less than TrackedInitializationLimit words 3542 // will be considered for capture by an InitializeNode. This puts a 3543 // reasonable limit on the complexity of optimized initializations. 3544 3545 //---------------------------InitializeNode------------------------------------ 3546 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3547 : MemBarNode(C, adr_type, rawoop), 3548 _is_complete(Incomplete), _does_not_escape(false) 3549 { 3550 init_class_id(Class_Initialize); 3551 3552 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3553 assert(in(RawAddress) == rawoop, "proper init"); 3554 // Note: allocation() can be NULL, for secondary initialization barriers 3555 } 3556 3557 // Since this node is not matched, it will be processed by the 3558 // register allocator. Declare that there are no constraints 3559 // on the allocation of the RawAddress edge. 3560 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3561 // This edge should be set to top, by the set_complete. But be conservative. 3562 if (idx == InitializeNode::RawAddress) 3563 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3564 return RegMask::Empty; 3565 } 3566 3567 Node* InitializeNode::memory(uint alias_idx) { 3568 Node* mem = in(Memory); 3569 if (mem->is_MergeMem()) { 3570 return mem->as_MergeMem()->memory_at(alias_idx); 3571 } else { 3572 // incoming raw memory is not split 3573 return mem; 3574 } 3575 } 3576 3577 bool InitializeNode::is_non_zero() { 3578 if (is_complete()) return false; 3579 remove_extra_zeroes(); 3580 return (req() > RawStores); 3581 } 3582 3583 void InitializeNode::set_complete(PhaseGVN* phase) { 3584 assert(!is_complete(), "caller responsibility"); 3585 _is_complete = Complete; 3586 3587 // After this node is complete, it contains a bunch of 3588 // raw-memory initializations. There is no need for 3589 // it to have anything to do with non-raw memory effects. 3590 // Therefore, tell all non-raw users to re-optimize themselves, 3591 // after skipping the memory effects of this initialization. 3592 PhaseIterGVN* igvn = phase->is_IterGVN(); 3593 if (igvn) igvn->add_users_to_worklist(this); 3594 } 3595 3596 // convenience function 3597 // return false if the init contains any stores already 3598 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3599 InitializeNode* init = initialization(); 3600 if (init == NULL || init->is_complete()) { 3601 return false; 3602 } 3603 init->remove_extra_zeroes(); 3604 // for now, if this allocation has already collected any inits, bail: 3605 if (init->is_non_zero()) return false; 3606 init->set_complete(phase); 3607 return true; 3608 } 3609 3610 void InitializeNode::remove_extra_zeroes() { 3611 if (req() == RawStores) return; 3612 Node* zmem = zero_memory(); 3613 uint fill = RawStores; 3614 for (uint i = fill; i < req(); i++) { 3615 Node* n = in(i); 3616 if (n->is_top() || n == zmem) continue; // skip 3617 if (fill < i) set_req(fill, n); // compact 3618 ++fill; 3619 } 3620 // delete any empty spaces created: 3621 while (fill < req()) { 3622 del_req(fill); 3623 } 3624 } 3625 3626 // Helper for remembering which stores go with which offsets. 3627 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3628 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3629 intptr_t offset = -1; 3630 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3631 phase, offset); 3632 if (base == NULL) return -1; // something is dead, 3633 if (offset < 0) return -1; // dead, dead 3634 return offset; 3635 } 3636 3637 // Helper for proving that an initialization expression is 3638 // "simple enough" to be folded into an object initialization. 3639 // Attempts to prove that a store's initial value 'n' can be captured 3640 // within the initialization without creating a vicious cycle, such as: 3641 // { Foo p = new Foo(); p.next = p; } 3642 // True for constants and parameters and small combinations thereof. 3643 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3644 if (n == NULL) return true; // (can this really happen?) 3645 if (n->is_Proj()) n = n->in(0); 3646 if (n == this) return false; // found a cycle 3647 if (n->is_Con()) return true; 3648 if (n->is_Start()) return true; // params, etc., are OK 3649 if (n->is_Root()) return true; // even better 3650 3651 Node* ctl = n->in(0); 3652 if (ctl != NULL && !ctl->is_top()) { 3653 if (ctl->is_Proj()) ctl = ctl->in(0); 3654 if (ctl == this) return false; 3655 3656 // If we already know that the enclosing memory op is pinned right after 3657 // the init, then any control flow that the store has picked up 3658 // must have preceded the init, or else be equal to the init. 3659 // Even after loop optimizations (which might change control edges) 3660 // a store is never pinned *before* the availability of its inputs. 3661 if (!MemNode::all_controls_dominate(n, this)) 3662 return false; // failed to prove a good control 3663 } 3664 3665 // Check data edges for possible dependencies on 'this'. 3666 if ((count += 1) > 20) return false; // complexity limit 3667 for (uint i = 1; i < n->req(); i++) { 3668 Node* m = n->in(i); 3669 if (m == NULL || m == n || m->is_top()) continue; 3670 uint first_i = n->find_edge(m); 3671 if (i != first_i) continue; // process duplicate edge just once 3672 if (!detect_init_independence(m, count)) { 3673 return false; 3674 } 3675 } 3676 3677 return true; 3678 } 3679 3680 // Here are all the checks a Store must pass before it can be moved into 3681 // an initialization. Returns zero if a check fails. 3682 // On success, returns the (constant) offset to which the store applies, 3683 // within the initialized memory. 3684 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3685 const int FAIL = 0; 3686 if (st->is_unaligned_access()) { 3687 return FAIL; 3688 } 3689 if (st->req() != MemNode::ValueIn + 1) 3690 return FAIL; // an inscrutable StoreNode (card mark?) 3691 Node* ctl = st->in(MemNode::Control); 3692 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3693 return FAIL; // must be unconditional after the initialization 3694 Node* mem = st->in(MemNode::Memory); 3695 if (!(mem->is_Proj() && mem->in(0) == this)) 3696 return FAIL; // must not be preceded by other stores 3697 Node* adr = st->in(MemNode::Address); 3698 intptr_t offset; 3699 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3700 if (alloc == NULL) 3701 return FAIL; // inscrutable address 3702 if (alloc != allocation()) 3703 return FAIL; // wrong allocation! (store needs to float up) 3704 Node* val = st->in(MemNode::ValueIn); 3705 int complexity_count = 0; 3706 if (!detect_init_independence(val, complexity_count)) 3707 return FAIL; // stored value must be 'simple enough' 3708 3709 // The Store can be captured only if nothing after the allocation 3710 // and before the Store is using the memory location that the store 3711 // overwrites. 3712 bool failed = false; 3713 // If is_complete_with_arraycopy() is true the shape of the graph is 3714 // well defined and is safe so no need for extra checks. 3715 if (!is_complete_with_arraycopy()) { 3716 // We are going to look at each use of the memory state following 3717 // the allocation to make sure nothing reads the memory that the 3718 // Store writes. 3719 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3720 int alias_idx = phase->C->get_alias_index(t_adr); 3721 ResourceMark rm; 3722 Unique_Node_List mems; 3723 mems.push(mem); 3724 Node* unique_merge = NULL; 3725 for (uint next = 0; next < mems.size(); ++next) { 3726 Node *m = mems.at(next); 3727 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3728 Node *n = m->fast_out(j); 3729 if (n->outcnt() == 0) { 3730 continue; 3731 } 3732 if (n == st) { 3733 continue; 3734 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3735 // If the control of this use is different from the control 3736 // of the Store which is right after the InitializeNode then 3737 // this node cannot be between the InitializeNode and the 3738 // Store. 3739 continue; 3740 } else if (n->is_MergeMem()) { 3741 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3742 // We can hit a MergeMemNode (that will likely go away 3743 // later) that is a direct use of the memory state 3744 // following the InitializeNode on the same slice as the 3745 // store node that we'd like to capture. We need to check 3746 // the uses of the MergeMemNode. 3747 mems.push(n); 3748 } 3749 } else if (n->is_Mem()) { 3750 Node* other_adr = n->in(MemNode::Address); 3751 if (other_adr == adr) { 3752 failed = true; 3753 break; 3754 } else { 3755 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3756 if (other_t_adr != NULL) { 3757 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3758 if (other_alias_idx == alias_idx) { 3759 // A load from the same memory slice as the store right 3760 // after the InitializeNode. We check the control of the 3761 // object/array that is loaded from. If it's the same as 3762 // the store control then we cannot capture the store. 3763 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3764 Node* base = other_adr; 3765 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3766 base = base->in(AddPNode::Base); 3767 if (base != NULL) { 3768 base = base->uncast(); 3769 if (base->is_Proj() && base->in(0) == alloc) { 3770 failed = true; 3771 break; 3772 } 3773 } 3774 } 3775 } 3776 } 3777 } else { 3778 failed = true; 3779 break; 3780 } 3781 } 3782 } 3783 } 3784 if (failed) { 3785 if (!can_reshape) { 3786 // We decided we couldn't capture the store during parsing. We 3787 // should try again during the next IGVN once the graph is 3788 // cleaner. 3789 phase->C->record_for_igvn(st); 3790 } 3791 return FAIL; 3792 } 3793 3794 return offset; // success 3795 } 3796 3797 // Find the captured store in(i) which corresponds to the range 3798 // [start..start+size) in the initialized object. 3799 // If there is one, return its index i. If there isn't, return the 3800 // negative of the index where it should be inserted. 3801 // Return 0 if the queried range overlaps an initialization boundary 3802 // or if dead code is encountered. 3803 // If size_in_bytes is zero, do not bother with overlap checks. 3804 int InitializeNode::captured_store_insertion_point(intptr_t start, 3805 int size_in_bytes, 3806 PhaseTransform* phase) { 3807 const int FAIL = 0, MAX_STORE = BytesPerLong; 3808 3809 if (is_complete()) 3810 return FAIL; // arraycopy got here first; punt 3811 3812 assert(allocation() != NULL, "must be present"); 3813 3814 // no negatives, no header fields: 3815 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3816 3817 // after a certain size, we bail out on tracking all the stores: 3818 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3819 if (start >= ti_limit) return FAIL; 3820 3821 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3822 if (i >= limit) return -(int)i; // not found; here is where to put it 3823 3824 Node* st = in(i); 3825 intptr_t st_off = get_store_offset(st, phase); 3826 if (st_off < 0) { 3827 if (st != zero_memory()) { 3828 return FAIL; // bail out if there is dead garbage 3829 } 3830 } else if (st_off > start) { 3831 // ...we are done, since stores are ordered 3832 if (st_off < start + size_in_bytes) { 3833 return FAIL; // the next store overlaps 3834 } 3835 return -(int)i; // not found; here is where to put it 3836 } else if (st_off < start) { 3837 if (size_in_bytes != 0 && 3838 start < st_off + MAX_STORE && 3839 start < st_off + st->as_Store()->memory_size()) { 3840 return FAIL; // the previous store overlaps 3841 } 3842 } else { 3843 if (size_in_bytes != 0 && 3844 st->as_Store()->memory_size() != size_in_bytes) { 3845 return FAIL; // mismatched store size 3846 } 3847 return i; 3848 } 3849 3850 ++i; 3851 } 3852 } 3853 3854 // Look for a captured store which initializes at the offset 'start' 3855 // with the given size. If there is no such store, and no other 3856 // initialization interferes, then return zero_memory (the memory 3857 // projection of the AllocateNode). 3858 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3859 PhaseTransform* phase) { 3860 assert(stores_are_sane(phase), ""); 3861 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3862 if (i == 0) { 3863 return NULL; // something is dead 3864 } else if (i < 0) { 3865 return zero_memory(); // just primordial zero bits here 3866 } else { 3867 Node* st = in(i); // here is the store at this position 3868 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3869 return st; 3870 } 3871 } 3872 3873 // Create, as a raw pointer, an address within my new object at 'offset'. 3874 Node* InitializeNode::make_raw_address(intptr_t offset, 3875 PhaseTransform* phase) { 3876 Node* addr = in(RawAddress); 3877 if (offset != 0) { 3878 Compile* C = phase->C; 3879 addr = phase->transform( new AddPNode(C->top(), addr, 3880 phase->MakeConX(offset)) ); 3881 } 3882 return addr; 3883 } 3884 3885 // Clone the given store, converting it into a raw store 3886 // initializing a field or element of my new object. 3887 // Caller is responsible for retiring the original store, 3888 // with subsume_node or the like. 3889 // 3890 // From the example above InitializeNode::InitializeNode, 3891 // here are the old stores to be captured: 3892 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3893 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3894 // 3895 // Here is the changed code; note the extra edges on init: 3896 // alloc = (Allocate ...) 3897 // rawoop = alloc.RawAddress 3898 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3899 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3900 // init = (Initialize alloc.Control alloc.Memory rawoop 3901 // rawstore1 rawstore2) 3902 // 3903 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3904 PhaseTransform* phase, bool can_reshape) { 3905 assert(stores_are_sane(phase), ""); 3906 3907 if (start < 0) return NULL; 3908 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3909 3910 Compile* C = phase->C; 3911 int size_in_bytes = st->memory_size(); 3912 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3913 if (i == 0) return NULL; // bail out 3914 Node* prev_mem = NULL; // raw memory for the captured store 3915 if (i > 0) { 3916 prev_mem = in(i); // there is a pre-existing store under this one 3917 set_req(i, C->top()); // temporarily disconnect it 3918 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3919 } else { 3920 i = -i; // no pre-existing store 3921 prev_mem = zero_memory(); // a slice of the newly allocated object 3922 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3923 set_req(--i, C->top()); // reuse this edge; it has been folded away 3924 else 3925 ins_req(i, C->top()); // build a new edge 3926 } 3927 Node* new_st = st->clone(); 3928 new_st->set_req(MemNode::Control, in(Control)); 3929 new_st->set_req(MemNode::Memory, prev_mem); 3930 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3931 new_st = phase->transform(new_st); 3932 3933 // At this point, new_st might have swallowed a pre-existing store 3934 // at the same offset, or perhaps new_st might have disappeared, 3935 // if it redundantly stored the same value (or zero to fresh memory). 3936 3937 // In any case, wire it in: 3938 phase->igvn_rehash_node_delayed(this); 3939 set_req(i, new_st); 3940 3941 // The caller may now kill the old guy. 3942 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3943 assert(check_st == new_st || check_st == NULL, "must be findable"); 3944 assert(!is_complete(), ""); 3945 return new_st; 3946 } 3947 3948 static bool store_constant(jlong* tiles, int num_tiles, 3949 intptr_t st_off, int st_size, 3950 jlong con) { 3951 if ((st_off & (st_size-1)) != 0) 3952 return false; // strange store offset (assume size==2**N) 3953 address addr = (address)tiles + st_off; 3954 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3955 switch (st_size) { 3956 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3957 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3958 case sizeof(jint): *(jint*) addr = (jint) con; break; 3959 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3960 default: return false; // strange store size (detect size!=2**N here) 3961 } 3962 return true; // return success to caller 3963 } 3964 3965 // Coalesce subword constants into int constants and possibly 3966 // into long constants. The goal, if the CPU permits, 3967 // is to initialize the object with a small number of 64-bit tiles. 3968 // Also, convert floating-point constants to bit patterns. 3969 // Non-constants are not relevant to this pass. 3970 // 3971 // In terms of the running example on InitializeNode::InitializeNode 3972 // and InitializeNode::capture_store, here is the transformation 3973 // of rawstore1 and rawstore2 into rawstore12: 3974 // alloc = (Allocate ...) 3975 // rawoop = alloc.RawAddress 3976 // tile12 = 0x00010002 3977 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3978 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3979 // 3980 void 3981 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3982 Node* size_in_bytes, 3983 PhaseGVN* phase) { 3984 Compile* C = phase->C; 3985 3986 assert(stores_are_sane(phase), ""); 3987 // Note: After this pass, they are not completely sane, 3988 // since there may be some overlaps. 3989 3990 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3991 3992 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3993 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3994 size_limit = MIN2(size_limit, ti_limit); 3995 size_limit = align_up(size_limit, BytesPerLong); 3996 int num_tiles = size_limit / BytesPerLong; 3997 3998 // allocate space for the tile map: 3999 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 4000 jlong tiles_buf[small_len]; 4001 Node* nodes_buf[small_len]; 4002 jlong inits_buf[small_len]; 4003 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 4004 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4005 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 4006 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 4007 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 4008 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4009 // tiles: exact bitwise model of all primitive constants 4010 // nodes: last constant-storing node subsumed into the tiles model 4011 // inits: which bytes (in each tile) are touched by any initializations 4012 4013 //// Pass A: Fill in the tile model with any relevant stores. 4014 4015 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 4016 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 4017 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 4018 Node* zmem = zero_memory(); // initially zero memory state 4019 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4020 Node* st = in(i); 4021 intptr_t st_off = get_store_offset(st, phase); 4022 4023 // Figure out the store's offset and constant value: 4024 if (st_off < header_size) continue; //skip (ignore header) 4025 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 4026 int st_size = st->as_Store()->memory_size(); 4027 if (st_off + st_size > size_limit) break; 4028 4029 // Record which bytes are touched, whether by constant or not. 4030 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 4031 continue; // skip (strange store size) 4032 4033 const Type* val = phase->type(st->in(MemNode::ValueIn)); 4034 if (!val->singleton()) continue; //skip (non-con store) 4035 BasicType type = val->basic_type(); 4036 4037 jlong con = 0; 4038 switch (type) { 4039 case T_INT: con = val->is_int()->get_con(); break; 4040 case T_LONG: con = val->is_long()->get_con(); break; 4041 case T_FLOAT: con = jint_cast(val->getf()); break; 4042 case T_DOUBLE: con = jlong_cast(val->getd()); break; 4043 default: continue; //skip (odd store type) 4044 } 4045 4046 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 4047 st->Opcode() == Op_StoreL) { 4048 continue; // This StoreL is already optimal. 4049 } 4050 4051 // Store down the constant. 4052 store_constant(tiles, num_tiles, st_off, st_size, con); 4053 4054 intptr_t j = st_off >> LogBytesPerLong; 4055 4056 if (type == T_INT && st_size == BytesPerInt 4057 && (st_off & BytesPerInt) == BytesPerInt) { 4058 jlong lcon = tiles[j]; 4059 if (!Matcher::isSimpleConstant64(lcon) && 4060 st->Opcode() == Op_StoreI) { 4061 // This StoreI is already optimal by itself. 4062 jint* intcon = (jint*) &tiles[j]; 4063 intcon[1] = 0; // undo the store_constant() 4064 4065 // If the previous store is also optimal by itself, back up and 4066 // undo the action of the previous loop iteration... if we can. 4067 // But if we can't, just let the previous half take care of itself. 4068 st = nodes[j]; 4069 st_off -= BytesPerInt; 4070 con = intcon[0]; 4071 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 4072 assert(st_off >= header_size, "still ignoring header"); 4073 assert(get_store_offset(st, phase) == st_off, "must be"); 4074 assert(in(i-1) == zmem, "must be"); 4075 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 4076 assert(con == tcon->is_int()->get_con(), "must be"); 4077 // Undo the effects of the previous loop trip, which swallowed st: 4078 intcon[0] = 0; // undo store_constant() 4079 set_req(i-1, st); // undo set_req(i, zmem) 4080 nodes[j] = NULL; // undo nodes[j] = st 4081 --old_subword; // undo ++old_subword 4082 } 4083 continue; // This StoreI is already optimal. 4084 } 4085 } 4086 4087 // This store is not needed. 4088 set_req(i, zmem); 4089 nodes[j] = st; // record for the moment 4090 if (st_size < BytesPerLong) // something has changed 4091 ++old_subword; // includes int/float, but who's counting... 4092 else ++old_long; 4093 } 4094 4095 if ((old_subword + old_long) == 0) 4096 return; // nothing more to do 4097 4098 //// Pass B: Convert any non-zero tiles into optimal constant stores. 4099 // Be sure to insert them before overlapping non-constant stores. 4100 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 4101 for (int j = 0; j < num_tiles; j++) { 4102 jlong con = tiles[j]; 4103 jlong init = inits[j]; 4104 if (con == 0) continue; 4105 jint con0, con1; // split the constant, address-wise 4106 jint init0, init1; // split the init map, address-wise 4107 { union { jlong con; jint intcon[2]; } u; 4108 u.con = con; 4109 con0 = u.intcon[0]; 4110 con1 = u.intcon[1]; 4111 u.con = init; 4112 init0 = u.intcon[0]; 4113 init1 = u.intcon[1]; 4114 } 4115 4116 Node* old = nodes[j]; 4117 assert(old != NULL, "need the prior store"); 4118 intptr_t offset = (j * BytesPerLong); 4119 4120 bool split = !Matcher::isSimpleConstant64(con); 4121 4122 if (offset < header_size) { 4123 assert(offset + BytesPerInt >= header_size, "second int counts"); 4124 assert(*(jint*)&tiles[j] == 0, "junk in header"); 4125 split = true; // only the second word counts 4126 // Example: int a[] = { 42 ... } 4127 } else if (con0 == 0 && init0 == -1) { 4128 split = true; // first word is covered by full inits 4129 // Example: int a[] = { ... foo(), 42 ... } 4130 } else if (con1 == 0 && init1 == -1) { 4131 split = true; // second word is covered by full inits 4132 // Example: int a[] = { ... 42, foo() ... } 4133 } 4134 4135 // Here's a case where init0 is neither 0 nor -1: 4136 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4137 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4138 // In this case the tile is not split; it is (jlong)42. 4139 // The big tile is stored down, and then the foo() value is inserted. 4140 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4141 4142 Node* ctl = old->in(MemNode::Control); 4143 Node* adr = make_raw_address(offset, phase); 4144 const TypePtr* atp = TypeRawPtr::BOTTOM; 4145 4146 // One or two coalesced stores to plop down. 4147 Node* st[2]; 4148 intptr_t off[2]; 4149 int nst = 0; 4150 if (!split) { 4151 ++new_long; 4152 off[nst] = offset; 4153 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4154 phase->longcon(con), T_LONG, MemNode::unordered); 4155 } else { 4156 // Omit either if it is a zero. 4157 if (con0 != 0) { 4158 ++new_int; 4159 off[nst] = offset; 4160 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4161 phase->intcon(con0), T_INT, MemNode::unordered); 4162 } 4163 if (con1 != 0) { 4164 ++new_int; 4165 offset += BytesPerInt; 4166 adr = make_raw_address(offset, phase); 4167 off[nst] = offset; 4168 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4169 phase->intcon(con1), T_INT, MemNode::unordered); 4170 } 4171 } 4172 4173 // Insert second store first, then the first before the second. 4174 // Insert each one just before any overlapping non-constant stores. 4175 while (nst > 0) { 4176 Node* st1 = st[--nst]; 4177 C->copy_node_notes_to(st1, old); 4178 st1 = phase->transform(st1); 4179 offset = off[nst]; 4180 assert(offset >= header_size, "do not smash header"); 4181 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4182 guarantee(ins_idx != 0, "must re-insert constant store"); 4183 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4184 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4185 set_req(--ins_idx, st1); 4186 else 4187 ins_req(ins_idx, st1); 4188 } 4189 } 4190 4191 if (PrintCompilation && WizardMode) 4192 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4193 old_subword, old_long, new_int, new_long); 4194 if (C->log() != NULL) 4195 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4196 old_subword, old_long, new_int, new_long); 4197 4198 // Clean up any remaining occurrences of zmem: 4199 remove_extra_zeroes(); 4200 } 4201 4202 // Explore forward from in(start) to find the first fully initialized 4203 // word, and return its offset. Skip groups of subword stores which 4204 // together initialize full words. If in(start) is itself part of a 4205 // fully initialized word, return the offset of in(start). If there 4206 // are no following full-word stores, or if something is fishy, return 4207 // a negative value. 4208 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4209 int int_map = 0; 4210 intptr_t int_map_off = 0; 4211 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4212 4213 for (uint i = start, limit = req(); i < limit; i++) { 4214 Node* st = in(i); 4215 4216 intptr_t st_off = get_store_offset(st, phase); 4217 if (st_off < 0) break; // return conservative answer 4218 4219 int st_size = st->as_Store()->memory_size(); 4220 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4221 return st_off; // we found a complete word init 4222 } 4223 4224 // update the map: 4225 4226 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4227 if (this_int_off != int_map_off) { 4228 // reset the map: 4229 int_map = 0; 4230 int_map_off = this_int_off; 4231 } 4232 4233 int subword_off = st_off - this_int_off; 4234 int_map |= right_n_bits(st_size) << subword_off; 4235 if ((int_map & FULL_MAP) == FULL_MAP) { 4236 return this_int_off; // we found a complete word init 4237 } 4238 4239 // Did this store hit or cross the word boundary? 4240 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4241 if (next_int_off == this_int_off + BytesPerInt) { 4242 // We passed the current int, without fully initializing it. 4243 int_map_off = next_int_off; 4244 int_map >>= BytesPerInt; 4245 } else if (next_int_off > this_int_off + BytesPerInt) { 4246 // We passed the current and next int. 4247 return this_int_off + BytesPerInt; 4248 } 4249 } 4250 4251 return -1; 4252 } 4253 4254 4255 // Called when the associated AllocateNode is expanded into CFG. 4256 // At this point, we may perform additional optimizations. 4257 // Linearize the stores by ascending offset, to make memory 4258 // activity as coherent as possible. 4259 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4260 intptr_t header_size, 4261 Node* size_in_bytes, 4262 PhaseGVN* phase) { 4263 assert(!is_complete(), "not already complete"); 4264 assert(stores_are_sane(phase), ""); 4265 assert(allocation() != NULL, "must be present"); 4266 4267 remove_extra_zeroes(); 4268 4269 if (ReduceFieldZeroing || ReduceBulkZeroing) 4270 // reduce instruction count for common initialization patterns 4271 coalesce_subword_stores(header_size, size_in_bytes, phase); 4272 4273 Node* zmem = zero_memory(); // initially zero memory state 4274 Node* inits = zmem; // accumulating a linearized chain of inits 4275 #ifdef ASSERT 4276 intptr_t first_offset = allocation()->minimum_header_size(); 4277 intptr_t last_init_off = first_offset; // previous init offset 4278 intptr_t last_init_end = first_offset; // previous init offset+size 4279 intptr_t last_tile_end = first_offset; // previous tile offset+size 4280 #endif 4281 intptr_t zeroes_done = header_size; 4282 4283 bool do_zeroing = true; // we might give up if inits are very sparse 4284 int big_init_gaps = 0; // how many large gaps have we seen? 4285 4286 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4287 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4288 4289 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4290 Node* st = in(i); 4291 intptr_t st_off = get_store_offset(st, phase); 4292 if (st_off < 0) 4293 break; // unknown junk in the inits 4294 if (st->in(MemNode::Memory) != zmem) 4295 break; // complicated store chains somehow in list 4296 4297 int st_size = st->as_Store()->memory_size(); 4298 intptr_t next_init_off = st_off + st_size; 4299 4300 if (do_zeroing && zeroes_done < next_init_off) { 4301 // See if this store needs a zero before it or under it. 4302 intptr_t zeroes_needed = st_off; 4303 4304 if (st_size < BytesPerInt) { 4305 // Look for subword stores which only partially initialize words. 4306 // If we find some, we must lay down some word-level zeroes first, 4307 // underneath the subword stores. 4308 // 4309 // Examples: 4310 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4311 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4312 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4313 // 4314 // Note: coalesce_subword_stores may have already done this, 4315 // if it was prompted by constant non-zero subword initializers. 4316 // But this case can still arise with non-constant stores. 4317 4318 intptr_t next_full_store = find_next_fullword_store(i, phase); 4319 4320 // In the examples above: 4321 // in(i) p q r s x y z 4322 // st_off 12 13 14 15 12 13 14 4323 // st_size 1 1 1 1 1 1 1 4324 // next_full_s. 12 16 16 16 16 16 16 4325 // z's_done 12 16 16 16 12 16 12 4326 // z's_needed 12 16 16 16 16 16 16 4327 // zsize 0 0 0 0 4 0 4 4328 if (next_full_store < 0) { 4329 // Conservative tack: Zero to end of current word. 4330 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4331 } else { 4332 // Zero to beginning of next fully initialized word. 4333 // Or, don't zero at all, if we are already in that word. 4334 assert(next_full_store >= zeroes_needed, "must go forward"); 4335 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4336 zeroes_needed = next_full_store; 4337 } 4338 } 4339 4340 if (zeroes_needed > zeroes_done) { 4341 intptr_t zsize = zeroes_needed - zeroes_done; 4342 // Do some incremental zeroing on rawmem, in parallel with inits. 4343 zeroes_done = align_down(zeroes_done, BytesPerInt); 4344 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4345 allocation()->in(AllocateNode::DefaultValue), 4346 allocation()->in(AllocateNode::RawDefaultValue), 4347 zeroes_done, zeroes_needed, 4348 phase); 4349 zeroes_done = zeroes_needed; 4350 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4351 do_zeroing = false; // leave the hole, next time 4352 } 4353 } 4354 4355 // Collect the store and move on: 4356 st->set_req(MemNode::Memory, inits); 4357 inits = st; // put it on the linearized chain 4358 set_req(i, zmem); // unhook from previous position 4359 4360 if (zeroes_done == st_off) 4361 zeroes_done = next_init_off; 4362 4363 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4364 4365 #ifdef ASSERT 4366 // Various order invariants. Weaker than stores_are_sane because 4367 // a large constant tile can be filled in by smaller non-constant stores. 4368 assert(st_off >= last_init_off, "inits do not reverse"); 4369 last_init_off = st_off; 4370 const Type* val = NULL; 4371 if (st_size >= BytesPerInt && 4372 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4373 (int)val->basic_type() < (int)T_OBJECT) { 4374 assert(st_off >= last_tile_end, "tiles do not overlap"); 4375 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4376 last_tile_end = MAX2(last_tile_end, next_init_off); 4377 } else { 4378 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4379 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4380 assert(st_off >= last_init_end, "inits do not overlap"); 4381 last_init_end = next_init_off; // it's a non-tile 4382 } 4383 #endif //ASSERT 4384 } 4385 4386 remove_extra_zeroes(); // clear out all the zmems left over 4387 add_req(inits); 4388 4389 if (!(UseTLAB && ZeroTLAB)) { 4390 // If anything remains to be zeroed, zero it all now. 4391 zeroes_done = align_down(zeroes_done, BytesPerInt); 4392 // if it is the last unused 4 bytes of an instance, forget about it 4393 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4394 if (zeroes_done + BytesPerLong >= size_limit) { 4395 AllocateNode* alloc = allocation(); 4396 assert(alloc != NULL, "must be present"); 4397 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 4398 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4399 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4400 if (zeroes_done == k->layout_helper()) 4401 zeroes_done = size_limit; 4402 } 4403 } 4404 if (zeroes_done < size_limit) { 4405 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4406 allocation()->in(AllocateNode::DefaultValue), 4407 allocation()->in(AllocateNode::RawDefaultValue), 4408 zeroes_done, size_in_bytes, phase); 4409 } 4410 } 4411 4412 set_complete(phase); 4413 return rawmem; 4414 } 4415 4416 4417 #ifdef ASSERT 4418 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4419 if (is_complete()) 4420 return true; // stores could be anything at this point 4421 assert(allocation() != NULL, "must be present"); 4422 intptr_t last_off = allocation()->minimum_header_size(); 4423 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4424 Node* st = in(i); 4425 intptr_t st_off = get_store_offset(st, phase); 4426 if (st_off < 0) continue; // ignore dead garbage 4427 if (last_off > st_off) { 4428 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4429 this->dump(2); 4430 assert(false, "ascending store offsets"); 4431 return false; 4432 } 4433 last_off = st_off + st->as_Store()->memory_size(); 4434 } 4435 return true; 4436 } 4437 #endif //ASSERT 4438 4439 4440 4441 4442 //============================MergeMemNode===================================== 4443 // 4444 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4445 // contributing store or call operations. Each contributor provides the memory 4446 // state for a particular "alias type" (see Compile::alias_type). For example, 4447 // if a MergeMem has an input X for alias category #6, then any memory reference 4448 // to alias category #6 may use X as its memory state input, as an exact equivalent 4449 // to using the MergeMem as a whole. 4450 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4451 // 4452 // (Here, the <N> notation gives the index of the relevant adr_type.) 4453 // 4454 // In one special case (and more cases in the future), alias categories overlap. 4455 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4456 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4457 // it is exactly equivalent to that state W: 4458 // MergeMem(<Bot>: W) <==> W 4459 // 4460 // Usually, the merge has more than one input. In that case, where inputs 4461 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4462 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4463 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4464 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4465 // 4466 // A merge can take a "wide" memory state as one of its narrow inputs. 4467 // This simply means that the merge observes out only the relevant parts of 4468 // the wide input. That is, wide memory states arriving at narrow merge inputs 4469 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4470 // 4471 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4472 // and that memory slices "leak through": 4473 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4474 // 4475 // But, in such a cascade, repeated memory slices can "block the leak": 4476 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4477 // 4478 // In the last example, Y is not part of the combined memory state of the 4479 // outermost MergeMem. The system must, of course, prevent unschedulable 4480 // memory states from arising, so you can be sure that the state Y is somehow 4481 // a precursor to state Y'. 4482 // 4483 // 4484 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4485 // of each MergeMemNode array are exactly the numerical alias indexes, including 4486 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4487 // Compile::alias_type (and kin) produce and manage these indexes. 4488 // 4489 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4490 // (Note that this provides quick access to the top node inside MergeMem methods, 4491 // without the need to reach out via TLS to Compile::current.) 4492 // 4493 // As a consequence of what was just described, a MergeMem that represents a full 4494 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4495 // containing all alias categories. 4496 // 4497 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4498 // 4499 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4500 // a memory state for the alias type <N>, or else the top node, meaning that 4501 // there is no particular input for that alias type. Note that the length of 4502 // a MergeMem is variable, and may be extended at any time to accommodate new 4503 // memory states at larger alias indexes. When merges grow, they are of course 4504 // filled with "top" in the unused in() positions. 4505 // 4506 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4507 // (Top was chosen because it works smoothly with passes like GCM.) 4508 // 4509 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4510 // the type of random VM bits like TLS references.) Since it is always the 4511 // first non-Bot memory slice, some low-level loops use it to initialize an 4512 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4513 // 4514 // 4515 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4516 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4517 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4518 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4519 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4520 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4521 // 4522 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4523 // really that different from the other memory inputs. An abbreviation called 4524 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4525 // 4526 // 4527 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4528 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4529 // that "emerges though" the base memory will be marked as excluding the alias types 4530 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4531 // 4532 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4533 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4534 // 4535 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4536 // (It is currently unimplemented.) As you can see, the resulting merge is 4537 // actually a disjoint union of memory states, rather than an overlay. 4538 // 4539 4540 //------------------------------MergeMemNode----------------------------------- 4541 Node* MergeMemNode::make_empty_memory() { 4542 Node* empty_memory = (Node*) Compile::current()->top(); 4543 assert(empty_memory->is_top(), "correct sentinel identity"); 4544 return empty_memory; 4545 } 4546 4547 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4548 init_class_id(Class_MergeMem); 4549 // all inputs are nullified in Node::Node(int) 4550 // set_input(0, NULL); // no control input 4551 4552 // Initialize the edges uniformly to top, for starters. 4553 Node* empty_mem = make_empty_memory(); 4554 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4555 init_req(i,empty_mem); 4556 } 4557 assert(empty_memory() == empty_mem, ""); 4558 4559 if( new_base != NULL && new_base->is_MergeMem() ) { 4560 MergeMemNode* mdef = new_base->as_MergeMem(); 4561 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4562 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4563 mms.set_memory(mms.memory2()); 4564 } 4565 assert(base_memory() == mdef->base_memory(), ""); 4566 } else { 4567 set_base_memory(new_base); 4568 } 4569 } 4570 4571 // Make a new, untransformed MergeMem with the same base as 'mem'. 4572 // If mem is itself a MergeMem, populate the result with the same edges. 4573 MergeMemNode* MergeMemNode::make(Node* mem) { 4574 return new MergeMemNode(mem); 4575 } 4576 4577 //------------------------------cmp-------------------------------------------- 4578 uint MergeMemNode::hash() const { return NO_HASH; } 4579 bool MergeMemNode::cmp( const Node &n ) const { 4580 return (&n == this); // Always fail except on self 4581 } 4582 4583 //------------------------------Identity--------------------------------------- 4584 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4585 // Identity if this merge point does not record any interesting memory 4586 // disambiguations. 4587 Node* base_mem = base_memory(); 4588 Node* empty_mem = empty_memory(); 4589 if (base_mem != empty_mem) { // Memory path is not dead? 4590 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4591 Node* mem = in(i); 4592 if (mem != empty_mem && mem != base_mem) { 4593 return this; // Many memory splits; no change 4594 } 4595 } 4596 } 4597 return base_mem; // No memory splits; ID on the one true input 4598 } 4599 4600 //------------------------------Ideal------------------------------------------ 4601 // This method is invoked recursively on chains of MergeMem nodes 4602 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4603 // Remove chain'd MergeMems 4604 // 4605 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4606 // relative to the "in(Bot)". Since we are patching both at the same time, 4607 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4608 // but rewrite each "in(i)" relative to the new "in(Bot)". 4609 Node *progress = NULL; 4610 4611 4612 Node* old_base = base_memory(); 4613 Node* empty_mem = empty_memory(); 4614 if (old_base == empty_mem) 4615 return NULL; // Dead memory path. 4616 4617 MergeMemNode* old_mbase; 4618 if (old_base != NULL && old_base->is_MergeMem()) 4619 old_mbase = old_base->as_MergeMem(); 4620 else 4621 old_mbase = NULL; 4622 Node* new_base = old_base; 4623 4624 // simplify stacked MergeMems in base memory 4625 if (old_mbase) new_base = old_mbase->base_memory(); 4626 4627 // the base memory might contribute new slices beyond my req() 4628 if (old_mbase) grow_to_match(old_mbase); 4629 4630 // Look carefully at the base node if it is a phi. 4631 PhiNode* phi_base; 4632 if (new_base != NULL && new_base->is_Phi()) 4633 phi_base = new_base->as_Phi(); 4634 else 4635 phi_base = NULL; 4636 4637 Node* phi_reg = NULL; 4638 uint phi_len = (uint)-1; 4639 if (phi_base != NULL && !phi_base->is_copy()) { 4640 // do not examine phi if degraded to a copy 4641 phi_reg = phi_base->region(); 4642 phi_len = phi_base->req(); 4643 // see if the phi is unfinished 4644 for (uint i = 1; i < phi_len; i++) { 4645 if (phi_base->in(i) == NULL) { 4646 // incomplete phi; do not look at it yet! 4647 phi_reg = NULL; 4648 phi_len = (uint)-1; 4649 break; 4650 } 4651 } 4652 } 4653 4654 // Note: We do not call verify_sparse on entry, because inputs 4655 // can normalize to the base_memory via subsume_node or similar 4656 // mechanisms. This method repairs that damage. 4657 4658 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4659 4660 // Look at each slice. 4661 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4662 Node* old_in = in(i); 4663 // calculate the old memory value 4664 Node* old_mem = old_in; 4665 if (old_mem == empty_mem) old_mem = old_base; 4666 assert(old_mem == memory_at(i), ""); 4667 4668 // maybe update (reslice) the old memory value 4669 4670 // simplify stacked MergeMems 4671 Node* new_mem = old_mem; 4672 MergeMemNode* old_mmem; 4673 if (old_mem != NULL && old_mem->is_MergeMem()) 4674 old_mmem = old_mem->as_MergeMem(); 4675 else 4676 old_mmem = NULL; 4677 if (old_mmem == this) { 4678 // This can happen if loops break up and safepoints disappear. 4679 // A merge of BotPtr (default) with a RawPtr memory derived from a 4680 // safepoint can be rewritten to a merge of the same BotPtr with 4681 // the BotPtr phi coming into the loop. If that phi disappears 4682 // also, we can end up with a self-loop of the mergemem. 4683 // In general, if loops degenerate and memory effects disappear, 4684 // a mergemem can be left looking at itself. This simply means 4685 // that the mergemem's default should be used, since there is 4686 // no longer any apparent effect on this slice. 4687 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4688 // from start. Update the input to TOP. 4689 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4690 } 4691 else if (old_mmem != NULL) { 4692 new_mem = old_mmem->memory_at(i); 4693 } 4694 // else preceding memory was not a MergeMem 4695 4696 // replace equivalent phis (unfortunately, they do not GVN together) 4697 if (new_mem != NULL && new_mem != new_base && 4698 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4699 if (new_mem->is_Phi()) { 4700 PhiNode* phi_mem = new_mem->as_Phi(); 4701 for (uint i = 1; i < phi_len; i++) { 4702 if (phi_base->in(i) != phi_mem->in(i)) { 4703 phi_mem = NULL; 4704 break; 4705 } 4706 } 4707 if (phi_mem != NULL) { 4708 // equivalent phi nodes; revert to the def 4709 new_mem = new_base; 4710 } 4711 } 4712 } 4713 4714 // maybe store down a new value 4715 Node* new_in = new_mem; 4716 if (new_in == new_base) new_in = empty_mem; 4717 4718 if (new_in != old_in) { 4719 // Warning: Do not combine this "if" with the previous "if" 4720 // A memory slice might have be be rewritten even if it is semantically 4721 // unchanged, if the base_memory value has changed. 4722 set_req(i, new_in); 4723 progress = this; // Report progress 4724 } 4725 } 4726 4727 if (new_base != old_base) { 4728 set_req(Compile::AliasIdxBot, new_base); 4729 // Don't use set_base_memory(new_base), because we need to update du. 4730 assert(base_memory() == new_base, ""); 4731 progress = this; 4732 } 4733 4734 if( base_memory() == this ) { 4735 // a self cycle indicates this memory path is dead 4736 set_req(Compile::AliasIdxBot, empty_mem); 4737 } 4738 4739 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4740 // Recursion must occur after the self cycle check above 4741 if( base_memory()->is_MergeMem() ) { 4742 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4743 Node *m = phase->transform(new_mbase); // Rollup any cycles 4744 if( m != NULL && 4745 (m->is_top() || 4746 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4747 // propagate rollup of dead cycle to self 4748 set_req(Compile::AliasIdxBot, empty_mem); 4749 } 4750 } 4751 4752 if( base_memory() == empty_mem ) { 4753 progress = this; 4754 // Cut inputs during Parse phase only. 4755 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4756 if( !can_reshape ) { 4757 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4758 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4759 } 4760 } 4761 } 4762 4763 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4764 // Check if PhiNode::Ideal's "Split phis through memory merges" 4765 // transform should be attempted. Look for this->phi->this cycle. 4766 uint merge_width = req(); 4767 if (merge_width > Compile::AliasIdxRaw) { 4768 PhiNode* phi = base_memory()->as_Phi(); 4769 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4770 if (phi->in(i) == this) { 4771 phase->is_IterGVN()->_worklist.push(phi); 4772 break; 4773 } 4774 } 4775 } 4776 } 4777 4778 assert(progress || verify_sparse(), "please, no dups of base"); 4779 return progress; 4780 } 4781 4782 //-------------------------set_base_memory------------------------------------- 4783 void MergeMemNode::set_base_memory(Node *new_base) { 4784 Node* empty_mem = empty_memory(); 4785 set_req(Compile::AliasIdxBot, new_base); 4786 assert(memory_at(req()) == new_base, "must set default memory"); 4787 // Clear out other occurrences of new_base: 4788 if (new_base != empty_mem) { 4789 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4790 if (in(i) == new_base) set_req(i, empty_mem); 4791 } 4792 } 4793 } 4794 4795 //------------------------------out_RegMask------------------------------------ 4796 const RegMask &MergeMemNode::out_RegMask() const { 4797 return RegMask::Empty; 4798 } 4799 4800 //------------------------------dump_spec-------------------------------------- 4801 #ifndef PRODUCT 4802 void MergeMemNode::dump_spec(outputStream *st) const { 4803 st->print(" {"); 4804 Node* base_mem = base_memory(); 4805 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4806 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4807 if (mem == base_mem) { st->print(" -"); continue; } 4808 st->print( " N%d:", mem->_idx ); 4809 Compile::current()->get_adr_type(i)->dump_on(st); 4810 } 4811 st->print(" }"); 4812 } 4813 #endif // !PRODUCT 4814 4815 4816 #ifdef ASSERT 4817 static bool might_be_same(Node* a, Node* b) { 4818 if (a == b) return true; 4819 if (!(a->is_Phi() || b->is_Phi())) return false; 4820 // phis shift around during optimization 4821 return true; // pretty stupid... 4822 } 4823 4824 // verify a narrow slice (either incoming or outgoing) 4825 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4826 if (!VerifyAliases) return; // don't bother to verify unless requested 4827 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4828 if (Node::in_dump()) return; // muzzle asserts when printing 4829 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4830 assert(n != NULL, ""); 4831 // Elide intervening MergeMem's 4832 while (n->is_MergeMem()) { 4833 n = n->as_MergeMem()->memory_at(alias_idx); 4834 } 4835 Compile* C = Compile::current(); 4836 const TypePtr* n_adr_type = n->adr_type(); 4837 if (n == m->empty_memory()) { 4838 // Implicit copy of base_memory() 4839 } else if (n_adr_type != TypePtr::BOTTOM) { 4840 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4841 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4842 } else { 4843 // A few places like make_runtime_call "know" that VM calls are narrow, 4844 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4845 bool expected_wide_mem = false; 4846 if (n == m->base_memory()) { 4847 expected_wide_mem = true; 4848 } else if (alias_idx == Compile::AliasIdxRaw || 4849 n == m->memory_at(Compile::AliasIdxRaw)) { 4850 expected_wide_mem = true; 4851 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4852 // memory can "leak through" calls on channels that 4853 // are write-once. Allow this also. 4854 expected_wide_mem = true; 4855 } 4856 assert(expected_wide_mem, "expected narrow slice replacement"); 4857 } 4858 } 4859 #else // !ASSERT 4860 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4861 #endif 4862 4863 4864 //-----------------------------memory_at--------------------------------------- 4865 Node* MergeMemNode::memory_at(uint alias_idx) const { 4866 assert(alias_idx >= Compile::AliasIdxRaw || 4867 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4868 "must avoid base_memory and AliasIdxTop"); 4869 4870 // Otherwise, it is a narrow slice. 4871 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4872 Compile *C = Compile::current(); 4873 if (is_empty_memory(n)) { 4874 // the array is sparse; empty slots are the "top" node 4875 n = base_memory(); 4876 assert(Node::in_dump() 4877 || n == NULL || n->bottom_type() == Type::TOP 4878 || n->adr_type() == NULL // address is TOP 4879 || n->adr_type() == TypePtr::BOTTOM 4880 || n->adr_type() == TypeRawPtr::BOTTOM 4881 || Compile::current()->AliasLevel() == 0, 4882 "must be a wide memory"); 4883 // AliasLevel == 0 if we are organizing the memory states manually. 4884 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4885 } else { 4886 // make sure the stored slice is sane 4887 #ifdef ASSERT 4888 if (VMError::is_error_reported() || Node::in_dump()) { 4889 } else if (might_be_same(n, base_memory())) { 4890 // Give it a pass: It is a mostly harmless repetition of the base. 4891 // This can arise normally from node subsumption during optimization. 4892 } else { 4893 verify_memory_slice(this, alias_idx, n); 4894 } 4895 #endif 4896 } 4897 return n; 4898 } 4899 4900 //---------------------------set_memory_at------------------------------------- 4901 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4902 verify_memory_slice(this, alias_idx, n); 4903 Node* empty_mem = empty_memory(); 4904 if (n == base_memory()) n = empty_mem; // collapse default 4905 uint need_req = alias_idx+1; 4906 if (req() < need_req) { 4907 if (n == empty_mem) return; // already the default, so do not grow me 4908 // grow the sparse array 4909 do { 4910 add_req(empty_mem); 4911 } while (req() < need_req); 4912 } 4913 set_req( alias_idx, n ); 4914 } 4915 4916 4917 4918 //--------------------------iteration_setup------------------------------------ 4919 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4920 if (other != NULL) { 4921 grow_to_match(other); 4922 // invariant: the finite support of mm2 is within mm->req() 4923 #ifdef ASSERT 4924 for (uint i = req(); i < other->req(); i++) { 4925 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4926 } 4927 #endif 4928 } 4929 // Replace spurious copies of base_memory by top. 4930 Node* base_mem = base_memory(); 4931 if (base_mem != NULL && !base_mem->is_top()) { 4932 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4933 if (in(i) == base_mem) 4934 set_req(i, empty_memory()); 4935 } 4936 } 4937 } 4938 4939 //---------------------------grow_to_match------------------------------------- 4940 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4941 Node* empty_mem = empty_memory(); 4942 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4943 // look for the finite support of the other memory 4944 for (uint i = other->req(); --i >= req(); ) { 4945 if (other->in(i) != empty_mem) { 4946 uint new_len = i+1; 4947 while (req() < new_len) add_req(empty_mem); 4948 break; 4949 } 4950 } 4951 } 4952 4953 //---------------------------verify_sparse------------------------------------- 4954 #ifndef PRODUCT 4955 bool MergeMemNode::verify_sparse() const { 4956 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4957 Node* base_mem = base_memory(); 4958 // The following can happen in degenerate cases, since empty==top. 4959 if (is_empty_memory(base_mem)) return true; 4960 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4961 assert(in(i) != NULL, "sane slice"); 4962 if (in(i) == base_mem) return false; // should have been the sentinel value! 4963 } 4964 return true; 4965 } 4966 4967 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4968 Node* n; 4969 n = mm->in(idx); 4970 if (mem == n) return true; // might be empty_memory() 4971 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4972 if (mem == n) return true; 4973 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4974 if (mem == n) return true; 4975 if (n == NULL) break; 4976 } 4977 return false; 4978 } 4979 #endif // !PRODUCT