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