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