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