1 /* 2 * Copyright (c) 1997, 2019, 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/rootnode.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 #if INCLUDE_SHENANDOAHGC 56 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 57 #endif 58 59 // Portions of code courtesy of Clifford Click 60 61 // Optimization - Graph Style 62 63 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 64 65 //============================================================================= 66 uint MemNode::size_of() const { return sizeof(*this); } 67 68 const TypePtr *MemNode::adr_type() const { 69 Node* adr = in(Address); 70 if (adr == NULL) return NULL; // node is dead 71 const TypePtr* cross_check = NULL; 72 DEBUG_ONLY(cross_check = _adr_type); 73 return calculate_adr_type(adr->bottom_type(), cross_check); 74 } 75 76 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 77 if (adr != NULL) { 78 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 79 return true; 80 } 81 } 82 return false; 83 } 84 85 #ifndef PRODUCT 86 void MemNode::dump_spec(outputStream *st) const { 87 if (in(Address) == NULL) return; // node is dead 88 #ifndef ASSERT 89 // fake the missing field 90 const TypePtr* _adr_type = NULL; 91 if (in(Address) != NULL) 92 _adr_type = in(Address)->bottom_type()->isa_ptr(); 93 #endif 94 dump_adr_type(this, _adr_type, st); 95 96 Compile* C = Compile::current(); 97 if (C->alias_type(_adr_type)->is_volatile()) { 98 st->print(" Volatile!"); 99 } 100 if (_unaligned_access) { 101 st->print(" unaligned"); 102 } 103 if (_mismatched_access) { 104 st->print(" mismatched"); 105 } 106 if (_unsafe_access) { 107 st->print(" unsafe"); 108 } 109 } 110 111 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 112 st->print(" @"); 113 if (adr_type == NULL) { 114 st->print("NULL"); 115 } else { 116 adr_type->dump_on(st); 117 Compile* C = Compile::current(); 118 Compile::AliasType* atp = NULL; 119 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 120 if (atp == NULL) 121 st->print(", idx=?\?;"); 122 else if (atp->index() == Compile::AliasIdxBot) 123 st->print(", idx=Bot;"); 124 else if (atp->index() == Compile::AliasIdxTop) 125 st->print(", idx=Top;"); 126 else if (atp->index() == Compile::AliasIdxRaw) 127 st->print(", idx=Raw;"); 128 else { 129 ciField* field = atp->field(); 130 if (field) { 131 st->print(", name="); 132 field->print_name_on(st); 133 } 134 st->print(", idx=%d;", atp->index()); 135 } 136 } 137 } 138 139 extern void print_alias_types(); 140 141 #endif 142 143 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 144 assert((t_oop != NULL), "sanity"); 145 bool is_instance = t_oop->is_known_instance_field(); 146 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 147 (load != NULL) && load->is_Load() && 148 (phase->is_IterGVN() != NULL); 149 if (!(is_instance || is_boxed_value_load)) 150 return mchain; // don't try to optimize non-instance types 151 uint instance_id = t_oop->instance_id(); 152 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory); 153 Node *prev = NULL; 154 Node *result = mchain; 155 while (prev != result) { 156 prev = result; 157 if (result == start_mem) 158 break; // hit one of our sentinels 159 // skip over a call which does not affect this memory slice 160 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 161 Node *proj_in = result->in(0); 162 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 163 break; // hit one of our sentinels 164 } else if (proj_in->is_Call()) { 165 // ArrayCopyNodes processed here as well 166 CallNode *call = proj_in->as_Call(); 167 if (!call->may_modify(t_oop, phase)) { // returns false for instances 168 result = call->in(TypeFunc::Memory); 169 } 170 } else if (proj_in->is_Initialize()) { 171 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 172 // Stop if this is the initialization for the object instance which 173 // contains this memory slice, otherwise skip over it. 174 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 175 break; 176 } 177 if (is_instance) { 178 result = proj_in->in(TypeFunc::Memory); 179 } else if (is_boxed_value_load) { 180 Node* klass = alloc->in(AllocateNode::KlassNode); 181 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 182 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 183 result = proj_in->in(TypeFunc::Memory); // not related allocation 184 } 185 } 186 } else if (proj_in->is_MemBar()) { 187 ArrayCopyNode* ac = NULL; 188 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 189 break; 190 } 191 result = proj_in->in(TypeFunc::Memory); 192 } else { 193 assert(false, "unexpected projection"); 194 } 195 } else if (result->is_ClearArray()) { 196 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 197 // Can not bypass initialization of the instance 198 // we are looking for. 199 break; 200 } 201 // Otherwise skip it (the call updated 'result' value). 202 } else if (result->is_MergeMem()) { 203 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 204 } 205 } 206 return result; 207 } 208 209 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 210 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 211 if (t_oop == NULL) 212 return mchain; // don't try to optimize non-oop types 213 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 214 bool is_instance = t_oop->is_known_instance_field(); 215 PhaseIterGVN *igvn = phase->is_IterGVN(); 216 if (is_instance && igvn != NULL && result->is_Phi()) { 217 PhiNode *mphi = result->as_Phi(); 218 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 219 const TypePtr *t = mphi->adr_type(); 220 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 221 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 222 t->is_oopptr()->cast_to_exactness(true) 223 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 224 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) { 225 // clone the Phi with our address type 226 result = mphi->split_out_instance(t_adr, igvn); 227 } else { 228 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 229 } 230 } 231 return result; 232 } 233 234 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 235 uint alias_idx = phase->C->get_alias_index(tp); 236 Node *mem = mmem; 237 #ifdef ASSERT 238 { 239 // Check that current type is consistent with the alias index used during graph construction 240 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 241 bool consistent = adr_check == NULL || adr_check->empty() || 242 phase->C->must_alias(adr_check, alias_idx ); 243 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 244 if( !consistent && adr_check != NULL && !adr_check->empty() && 245 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 246 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 247 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 248 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 249 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 250 // don't assert if it is dead code. 251 consistent = true; 252 } 253 if( !consistent ) { 254 st->print("alias_idx==%d, adr_check==", alias_idx); 255 if( adr_check == NULL ) { 256 st->print("NULL"); 257 } else { 258 adr_check->dump(); 259 } 260 st->cr(); 261 print_alias_types(); 262 assert(consistent, "adr_check must match alias idx"); 263 } 264 } 265 #endif 266 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 267 // means an array I have not precisely typed yet. Do not do any 268 // alias stuff with it any time soon. 269 const TypeOopPtr *toop = tp->isa_oopptr(); 270 if( tp->base() != Type::AnyPtr && 271 !(toop && 272 toop->klass() != NULL && 273 toop->klass()->is_java_lang_Object() && 274 toop->offset() == Type::OffsetBot) ) { 275 // compress paths and change unreachable cycles to TOP 276 // If not, we can update the input infinitely along a MergeMem cycle 277 // Equivalent code in PhiNode::Ideal 278 Node* m = phase->transform(mmem); 279 // If transformed to a MergeMem, get the desired slice 280 // Otherwise the returned node represents memory for every slice 281 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 282 // Update input if it is progress over what we have now 283 } 284 return mem; 285 } 286 287 //--------------------------Ideal_common--------------------------------------- 288 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 289 // Unhook non-raw memories from complete (macro-expanded) initializations. 290 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 291 // If our control input is a dead region, kill all below the region 292 Node *ctl = in(MemNode::Control); 293 if (ctl && remove_dead_region(phase, can_reshape)) 294 return this; 295 ctl = in(MemNode::Control); 296 // Don't bother trying to transform a dead node 297 if (ctl && ctl->is_top()) return NodeSentinel; 298 299 PhaseIterGVN *igvn = phase->is_IterGVN(); 300 // Wait if control on the worklist. 301 if (ctl && can_reshape && igvn != NULL) { 302 Node* bol = NULL; 303 Node* cmp = NULL; 304 if (ctl->in(0)->is_If()) { 305 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 306 bol = ctl->in(0)->in(1); 307 if (bol->is_Bool()) 308 cmp = ctl->in(0)->in(1)->in(1); 309 } 310 if (igvn->_worklist.member(ctl) || 311 (bol != NULL && igvn->_worklist.member(bol)) || 312 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 313 // This control path may be dead. 314 // Delay this memory node transformation until the control is processed. 315 phase->is_IterGVN()->_worklist.push(this); 316 return NodeSentinel; // caller will return NULL 317 } 318 } 319 // Ignore if memory is dead, or self-loop 320 Node *mem = in(MemNode::Memory); 321 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 322 assert(mem != this, "dead loop in MemNode::Ideal"); 323 324 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 325 // This memory slice may be dead. 326 // Delay this mem node transformation until the memory is processed. 327 phase->is_IterGVN()->_worklist.push(this); 328 return NodeSentinel; // caller will return NULL 329 } 330 331 Node *address = in(MemNode::Address); 332 const Type *t_adr = phase->type(address); 333 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 334 335 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) { 336 // Unsafe off-heap access with zero address. Remove access and other control users 337 // to not confuse optimizations and add a HaltNode to fail if this is ever executed. 338 assert(ctl != NULL, "unsafe accesses should be control dependent"); 339 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) { 340 Node* u = ctl->fast_out(i); 341 if (u != ctl) { 342 igvn->rehash_node_delayed(u); 343 int nb = u->replace_edge(ctl, phase->C->top()); 344 --i, imax -= nb; 345 } 346 } 347 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr)); 348 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address")); 349 phase->C->root()->add_req(halt); 350 return this; 351 } 352 353 if (can_reshape && igvn != NULL && 354 (igvn->_worklist.member(address) || 355 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 356 // The address's base and type may change when the address is processed. 357 // Delay this mem node transformation until the address is processed. 358 phase->is_IterGVN()->_worklist.push(this); 359 return NodeSentinel; // caller will return NULL 360 } 361 362 // Do NOT remove or optimize the next lines: ensure a new alias index 363 // is allocated for an oop pointer type before Escape Analysis. 364 // Note: C++ will not remove it since the call has side effect. 365 if (t_adr->isa_oopptr()) { 366 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 367 } 368 369 Node* base = NULL; 370 if (address->is_AddP()) { 371 base = address->in(AddPNode::Base); 372 } 373 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 374 !t_adr->isa_rawptr()) { 375 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 376 // Skip this node optimization if its address has TOP base. 377 return NodeSentinel; // caller will return NULL 378 } 379 380 // Avoid independent memory operations 381 Node* old_mem = mem; 382 383 // The code which unhooks non-raw memories from complete (macro-expanded) 384 // initializations was removed. After macro-expansion all stores catched 385 // by Initialize node became raw stores and there is no information 386 // which memory slices they modify. So it is unsafe to move any memory 387 // operation above these stores. Also in most cases hooked non-raw memories 388 // were already unhooked by using information from detect_ptr_independence() 389 // and find_previous_store(). 390 391 if (mem->is_MergeMem()) { 392 MergeMemNode* mmem = mem->as_MergeMem(); 393 const TypePtr *tp = t_adr->is_ptr(); 394 395 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 396 } 397 398 if (mem != old_mem) { 399 set_req(MemNode::Memory, mem); 400 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) { 401 igvn->_worklist.push(old_mem); 402 } 403 if (phase->type(mem) == Type::TOP) return NodeSentinel; 404 return this; 405 } 406 407 // let the subclass continue analyzing... 408 return NULL; 409 } 410 411 // Helper function for proving some simple control dominations. 412 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 413 // Already assumes that 'dom' is available at 'sub', and that 'sub' 414 // is not a constant (dominated by the method's StartNode). 415 // Used by MemNode::find_previous_store to prove that the 416 // control input of a memory operation predates (dominates) 417 // an allocation it wants to look past. 418 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 419 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 420 return false; // Conservative answer for dead code 421 422 // Check 'dom'. Skip Proj and CatchProj nodes. 423 dom = dom->find_exact_control(dom); 424 if (dom == NULL || dom->is_top()) 425 return false; // Conservative answer for dead code 426 427 if (dom == sub) { 428 // For the case when, for example, 'sub' is Initialize and the original 429 // 'dom' is Proj node of the 'sub'. 430 return false; 431 } 432 433 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 434 return true; 435 436 // 'dom' dominates 'sub' if its control edge and control edges 437 // of all its inputs dominate or equal to sub's control edge. 438 439 // Currently 'sub' is either Allocate, Initialize or Start nodes. 440 // Or Region for the check in LoadNode::Ideal(); 441 // 'sub' should have sub->in(0) != NULL. 442 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 443 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 444 445 // Get control edge of 'sub'. 446 Node* orig_sub = sub; 447 sub = sub->find_exact_control(sub->in(0)); 448 if (sub == NULL || sub->is_top()) 449 return false; // Conservative answer for dead code 450 451 assert(sub->is_CFG(), "expecting control"); 452 453 if (sub == dom) 454 return true; 455 456 if (sub->is_Start() || sub->is_Root()) 457 return false; 458 459 { 460 // Check all control edges of 'dom'. 461 462 ResourceMark rm; 463 Arena* arena = Thread::current()->resource_area(); 464 Node_List nlist(arena); 465 Unique_Node_List dom_list(arena); 466 467 dom_list.push(dom); 468 bool only_dominating_controls = false; 469 470 for (uint next = 0; next < dom_list.size(); next++) { 471 Node* n = dom_list.at(next); 472 if (n == orig_sub) 473 return false; // One of dom's inputs dominated by sub. 474 if (!n->is_CFG() && n->pinned()) { 475 // Check only own control edge for pinned non-control nodes. 476 n = n->find_exact_control(n->in(0)); 477 if (n == NULL || n->is_top()) 478 return false; // Conservative answer for dead code 479 assert(n->is_CFG(), "expecting control"); 480 dom_list.push(n); 481 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 482 only_dominating_controls = true; 483 } else if (n->is_CFG()) { 484 if (n->dominates(sub, nlist)) 485 only_dominating_controls = true; 486 else 487 return false; 488 } else { 489 // First, own control edge. 490 Node* m = n->find_exact_control(n->in(0)); 491 if (m != NULL) { 492 if (m->is_top()) 493 return false; // Conservative answer for dead code 494 dom_list.push(m); 495 } 496 // Now, the rest of edges. 497 uint cnt = n->req(); 498 for (uint i = 1; i < cnt; i++) { 499 m = n->find_exact_control(n->in(i)); 500 if (m == NULL || m->is_top()) 501 continue; 502 dom_list.push(m); 503 } 504 } 505 } 506 return only_dominating_controls; 507 } 508 } 509 510 //---------------------detect_ptr_independence--------------------------------- 511 // Used by MemNode::find_previous_store to prove that two base 512 // pointers are never equal. 513 // The pointers are accompanied by their associated allocations, 514 // if any, which have been previously discovered by the caller. 515 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 516 Node* p2, AllocateNode* a2, 517 PhaseTransform* phase) { 518 // Attempt to prove that these two pointers cannot be aliased. 519 // They may both manifestly be allocations, and they should differ. 520 // Or, if they are not both allocations, they can be distinct constants. 521 // Otherwise, one is an allocation and the other a pre-existing value. 522 if (a1 == NULL && a2 == NULL) { // neither an allocation 523 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 524 } else if (a1 != NULL && a2 != NULL) { // both allocations 525 return (a1 != a2); 526 } else if (a1 != NULL) { // one allocation a1 527 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 528 return all_controls_dominate(p2, a1); 529 } else { //(a2 != NULL) // one allocation a2 530 return all_controls_dominate(p1, a2); 531 } 532 return false; 533 } 534 535 536 // Find an arraycopy that must have set (can_see_stored_value=true) or 537 // could have set (can_see_stored_value=false) the value for this load 538 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 539 ArrayCopyNode* ac = find_array_copy_clone(phase, ld_alloc, mem); 540 if (ac != NULL) { 541 return ac; 542 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 543 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 544 545 if (ac->is_arraycopy_validated() || 546 ac->is_copyof_validated() || 547 ac->is_copyofrange_validated()) { 548 Node* ld_addp = in(MemNode::Address); 549 if (ld_addp->is_AddP()) { 550 Node* ld_base = ld_addp->in(AddPNode::Address); 551 Node* ld_offs = ld_addp->in(AddPNode::Offset); 552 553 Node* dest = ac->in(ArrayCopyNode::Dest); 554 555 if (dest == ld_base) { 556 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 557 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 558 return ac; 559 } 560 if (!can_see_stored_value) { 561 mem = ac->in(TypeFunc::Memory); 562 } 563 } 564 } 565 } 566 } 567 return NULL; 568 } 569 570 ArrayCopyNode* MemNode::find_array_copy_clone(PhaseTransform* phase, Node* ld_alloc, Node* mem) const { 571 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 572 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 573 if (ld_alloc != NULL) { 574 // Check if there is an array copy for a clone 575 Node* mb = mem->in(0); 576 ArrayCopyNode* ac = NULL; 577 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 578 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 579 ac = mb->in(0)->in(0)->as_ArrayCopy(); 580 } else { 581 // Step over GC barrier when ReduceInitialCardMarks is disabled 582 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 583 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0)); 584 585 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) { 586 ac = control_proj_ac->in(0)->as_ArrayCopy(); 587 } 588 } 589 590 if (ac != NULL && ac->is_clonebasic()) { 591 intptr_t offset; 592 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 593 if (alloc != NULL && alloc == ld_alloc) { 594 return ac; 595 } 596 } 597 } 598 } 599 return NULL; 600 } 601 602 // The logic for reordering loads and stores uses four steps: 603 // (a) Walk carefully past stores and initializations which we 604 // can prove are independent of this load. 605 // (b) Observe that the next memory state makes an exact match 606 // with self (load or store), and locate the relevant store. 607 // (c) Ensure that, if we were to wire self directly to the store, 608 // the optimizer would fold it up somehow. 609 // (d) Do the rewiring, and return, depending on some other part of 610 // the optimizer to fold up the load. 611 // This routine handles steps (a) and (b). Steps (c) and (d) are 612 // specific to loads and stores, so they are handled by the callers. 613 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 614 // 615 Node* MemNode::find_previous_store(PhaseTransform* phase) { 616 Node* ctrl = in(MemNode::Control); 617 Node* adr = in(MemNode::Address); 618 intptr_t offset = 0; 619 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 620 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 621 622 if (offset == Type::OffsetBot) 623 return NULL; // cannot unalias unless there are precise offsets 624 625 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 626 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 627 628 intptr_t size_in_bytes = memory_size(); 629 630 Node* mem = in(MemNode::Memory); // start searching here... 631 632 int cnt = 50; // Cycle limiter 633 for (;;) { // While we can dance past unrelated stores... 634 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 635 636 Node* prev = mem; 637 if (mem->is_Store()) { 638 Node* st_adr = mem->in(MemNode::Address); 639 intptr_t st_offset = 0; 640 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 641 if (st_base == NULL) 642 break; // inscrutable pointer 643 644 // For raw accesses it's not enough to prove that constant offsets don't intersect. 645 // We need the bases to be the equal in order for the offset check to make sense. 646 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 647 break; 648 } 649 650 if (st_offset != offset && st_offset != Type::OffsetBot) { 651 const int MAX_STORE = BytesPerLong; 652 if (st_offset >= offset + size_in_bytes || 653 st_offset <= offset - MAX_STORE || 654 st_offset <= offset - mem->as_Store()->memory_size()) { 655 // Success: The offsets are provably independent. 656 // (You may ask, why not just test st_offset != offset and be done? 657 // The answer is that stores of different sizes can co-exist 658 // in the same sequence of RawMem effects. We sometimes initialize 659 // a whole 'tile' of array elements with a single jint or jlong.) 660 mem = mem->in(MemNode::Memory); 661 continue; // (a) advance through independent store memory 662 } 663 } 664 if (st_base != base && 665 detect_ptr_independence(base, alloc, 666 st_base, 667 AllocateNode::Ideal_allocation(st_base, phase), 668 phase)) { 669 // Success: The bases are provably independent. 670 mem = mem->in(MemNode::Memory); 671 continue; // (a) advance through independent store memory 672 } 673 674 // (b) At this point, if the bases or offsets do not agree, we lose, 675 // since we have not managed to prove 'this' and 'mem' independent. 676 if (st_base == base && st_offset == offset) { 677 return mem; // let caller handle steps (c), (d) 678 } 679 680 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 681 InitializeNode* st_init = mem->in(0)->as_Initialize(); 682 AllocateNode* st_alloc = st_init->allocation(); 683 if (st_alloc == NULL) 684 break; // something degenerated 685 bool known_identical = false; 686 bool known_independent = false; 687 if (alloc == st_alloc) 688 known_identical = true; 689 else if (alloc != NULL) 690 known_independent = true; 691 else if (all_controls_dominate(this, st_alloc)) 692 known_independent = true; 693 694 if (known_independent) { 695 // The bases are provably independent: Either they are 696 // manifestly distinct allocations, or else the control 697 // of this load dominates the store's allocation. 698 int alias_idx = phase->C->get_alias_index(adr_type()); 699 if (alias_idx == Compile::AliasIdxRaw) { 700 mem = st_alloc->in(TypeFunc::Memory); 701 } else { 702 mem = st_init->memory(alias_idx); 703 } 704 continue; // (a) advance through independent store memory 705 } 706 707 // (b) at this point, if we are not looking at a store initializing 708 // the same allocation we are loading from, we lose. 709 if (known_identical) { 710 // From caller, can_see_stored_value will consult find_captured_store. 711 return mem; // let caller handle steps (c), (d) 712 } 713 714 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 715 if (prev != mem) { 716 // Found an arraycopy but it doesn't affect that load 717 continue; 718 } 719 // Found an arraycopy that may affect that load 720 return mem; 721 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 722 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 723 if (mem->is_Proj() && mem->in(0)->is_Call()) { 724 // ArrayCopyNodes processed here as well. 725 CallNode *call = mem->in(0)->as_Call(); 726 if (!call->may_modify(addr_t, phase)) { 727 mem = call->in(TypeFunc::Memory); 728 continue; // (a) advance through independent call memory 729 } 730 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 731 ArrayCopyNode* ac = NULL; 732 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 733 break; 734 } 735 mem = mem->in(0)->in(TypeFunc::Memory); 736 continue; // (a) advance through independent MemBar memory 737 } else if (mem->is_ClearArray()) { 738 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 739 // (the call updated 'mem' value) 740 continue; // (a) advance through independent allocation memory 741 } else { 742 // Can not bypass initialization of the instance 743 // we are looking for. 744 return mem; 745 } 746 } else if (mem->is_MergeMem()) { 747 int alias_idx = phase->C->get_alias_index(adr_type()); 748 mem = mem->as_MergeMem()->memory_at(alias_idx); 749 continue; // (a) advance through independent MergeMem memory 750 } 751 } 752 753 // Unless there is an explicit 'continue', we must bail out here, 754 // because 'mem' is an inscrutable memory state (e.g., a call). 755 break; 756 } 757 758 return NULL; // bail out 759 } 760 761 //----------------------calculate_adr_type------------------------------------- 762 // Helper function. Notices when the given type of address hits top or bottom. 763 // Also, asserts a cross-check of the type against the expected address type. 764 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 765 if (t == Type::TOP) return NULL; // does not touch memory any more? 766 #ifdef PRODUCT 767 cross_check = NULL; 768 #else 769 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL; 770 #endif 771 const TypePtr* tp = t->isa_ptr(); 772 if (tp == NULL) { 773 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 774 return TypePtr::BOTTOM; // touches lots of memory 775 } else { 776 #ifdef ASSERT 777 // %%%% [phh] We don't check the alias index if cross_check is 778 // TypeRawPtr::BOTTOM. Needs to be investigated. 779 if (cross_check != NULL && 780 cross_check != TypePtr::BOTTOM && 781 cross_check != TypeRawPtr::BOTTOM) { 782 // Recheck the alias index, to see if it has changed (due to a bug). 783 Compile* C = Compile::current(); 784 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 785 "must stay in the original alias category"); 786 // The type of the address must be contained in the adr_type, 787 // disregarding "null"-ness. 788 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 789 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 790 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 791 "real address must not escape from expected memory type"); 792 } 793 #endif 794 return tp; 795 } 796 } 797 798 //============================================================================= 799 // Should LoadNode::Ideal() attempt to remove control edges? 800 bool LoadNode::can_remove_control() const { 801 return true; 802 } 803 uint LoadNode::size_of() const { return sizeof(*this); } 804 uint LoadNode::cmp( const Node &n ) const 805 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 806 const Type *LoadNode::bottom_type() const { return _type; } 807 uint LoadNode::ideal_reg() const { 808 return _type->ideal_reg(); 809 } 810 811 #ifndef PRODUCT 812 void LoadNode::dump_spec(outputStream *st) const { 813 MemNode::dump_spec(st); 814 if( !Verbose && !WizardMode ) { 815 // standard dump does this in Verbose and WizardMode 816 st->print(" #"); _type->dump_on(st); 817 } 818 if (!depends_only_on_test()) { 819 st->print(" (does not depend only on test)"); 820 } 821 } 822 #endif 823 824 #ifdef ASSERT 825 //----------------------------is_immutable_value------------------------------- 826 // Helper function to allow a raw load without control edge for some cases 827 bool LoadNode::is_immutable_value(Node* adr) { 828 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 829 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 830 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 831 in_bytes(JavaThread::osthread_offset()))); 832 } 833 #endif 834 835 //----------------------------LoadNode::make----------------------------------- 836 // Polymorphic factory method: 837 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 838 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 839 Compile* C = gvn.C; 840 841 // sanity check the alias category against the created node type 842 assert(!(adr_type->isa_oopptr() && 843 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 844 "use LoadKlassNode instead"); 845 assert(!(adr_type->isa_aryptr() && 846 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 847 "use LoadRangeNode instead"); 848 // Check control edge of raw loads 849 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 850 // oop will be recorded in oop map if load crosses safepoint 851 rt->isa_oopptr() || is_immutable_value(adr), 852 "raw memory operations should have control edge"); 853 LoadNode* load = NULL; 854 switch (bt) { 855 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 856 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 857 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 858 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 859 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 860 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 861 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 862 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 863 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 864 case T_OBJECT: 865 #ifdef _LP64 866 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 867 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 868 } else 869 #endif 870 { 871 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 872 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 873 } 874 break; 875 default: 876 ShouldNotReachHere(); 877 break; 878 } 879 assert(load != NULL, "LoadNode should have been created"); 880 if (unaligned) { 881 load->set_unaligned_access(); 882 } 883 if (mismatched) { 884 load->set_mismatched_access(); 885 } 886 if (unsafe) { 887 load->set_unsafe_access(); 888 } 889 if (load->Opcode() == Op_LoadN) { 890 Node* ld = gvn.transform(load); 891 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 892 } 893 894 return load; 895 } 896 897 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 898 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 899 bool require_atomic = true; 900 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 901 if (unaligned) { 902 load->set_unaligned_access(); 903 } 904 if (mismatched) { 905 load->set_mismatched_access(); 906 } 907 if (unsafe) { 908 load->set_unsafe_access(); 909 } 910 return load; 911 } 912 913 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 914 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 915 bool require_atomic = true; 916 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 917 if (unaligned) { 918 load->set_unaligned_access(); 919 } 920 if (mismatched) { 921 load->set_mismatched_access(); 922 } 923 if (unsafe) { 924 load->set_unsafe_access(); 925 } 926 return load; 927 } 928 929 930 931 //------------------------------hash------------------------------------------- 932 uint LoadNode::hash() const { 933 // unroll addition of interesting fields 934 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 935 } 936 937 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 938 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 939 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 940 bool is_stable_ary = FoldStableValues && 941 (tp != NULL) && (tp->isa_aryptr() != NULL) && 942 tp->isa_aryptr()->is_stable(); 943 944 return (eliminate_boxing && non_volatile) || is_stable_ary; 945 } 946 947 return false; 948 } 949 950 // Is the value loaded previously stored by an arraycopy? If so return 951 // a load node that reads from the source array so we may be able to 952 // optimize out the ArrayCopy node later. 953 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 954 #if INCLUDE_ZGC 955 if (UseZGC) { 956 if (bottom_type()->make_oopptr() != NULL) { 957 return NULL; 958 } 959 } 960 #endif 961 962 Node* ld_adr = in(MemNode::Address); 963 intptr_t ld_off = 0; 964 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 965 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 966 if (ac != NULL) { 967 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 968 969 Node* mem = ac->in(TypeFunc::Memory); 970 Node* ctl = ac->in(0); 971 Node* src = ac->in(ArrayCopyNode::Src); 972 973 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 974 return NULL; 975 } 976 977 LoadNode* ld = clone()->as_Load(); 978 Node* addp = in(MemNode::Address)->clone(); 979 if (ac->as_ArrayCopy()->is_clonebasic()) { 980 assert(ld_alloc != NULL, "need an alloc"); 981 assert(addp->is_AddP(), "address must be addp"); 982 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern"); 983 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern"); 984 addp->set_req(AddPNode::Base, src->in(AddPNode::Base)); 985 addp->set_req(AddPNode::Address, src->in(AddPNode::Address)); 986 } else { 987 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 988 ac->as_ArrayCopy()->is_copyof_validated() || 989 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 990 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 991 addp->set_req(AddPNode::Base, src); 992 addp->set_req(AddPNode::Address, src); 993 994 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 995 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 996 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 997 uint shift = exact_log2(type2aelembytes(ary_elem)); 998 999 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 1000 #ifdef _LP64 1001 diff = phase->transform(new ConvI2LNode(diff)); 1002 #endif 1003 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 1004 1005 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 1006 addp->set_req(AddPNode::Offset, offset); 1007 } 1008 addp = phase->transform(addp); 1009 #ifdef ASSERT 1010 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 1011 ld->_adr_type = adr_type; 1012 #endif 1013 ld->set_req(MemNode::Address, addp); 1014 ld->set_req(0, ctl); 1015 ld->set_req(MemNode::Memory, mem); 1016 // load depends on the tests that validate the arraycopy 1017 ld->_control_dependency = Pinned; 1018 return ld; 1019 } 1020 return NULL; 1021 } 1022 1023 1024 //---------------------------can_see_stored_value------------------------------ 1025 // This routine exists to make sure this set of tests is done the same 1026 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 1027 // will change the graph shape in a way which makes memory alive twice at the 1028 // same time (uses the Oracle model of aliasing), then some 1029 // LoadXNode::Identity will fold things back to the equivalence-class model 1030 // of aliasing. 1031 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 1032 Node* ld_adr = in(MemNode::Address); 1033 intptr_t ld_off = 0; 1034 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 1035 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 1036 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 1037 // This is more general than load from boxing objects. 1038 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1039 uint alias_idx = atp->index(); 1040 bool final = !atp->is_rewritable(); 1041 Node* result = NULL; 1042 Node* current = st; 1043 // Skip through chains of MemBarNodes checking the MergeMems for 1044 // new states for the slice of this load. Stop once any other 1045 // kind of node is encountered. Loads from final memory can skip 1046 // through any kind of MemBar but normal loads shouldn't skip 1047 // through MemBarAcquire since the could allow them to move out of 1048 // a synchronized region. 1049 while (current->is_Proj()) { 1050 int opc = current->in(0)->Opcode(); 1051 if ((final && (opc == Op_MemBarAcquire || 1052 opc == Op_MemBarAcquireLock || 1053 opc == Op_LoadFence)) || 1054 opc == Op_MemBarRelease || 1055 opc == Op_StoreFence || 1056 opc == Op_MemBarReleaseLock || 1057 opc == Op_MemBarStoreStore || 1058 opc == Op_MemBarCPUOrder) { 1059 Node* mem = current->in(0)->in(TypeFunc::Memory); 1060 if (mem->is_MergeMem()) { 1061 MergeMemNode* merge = mem->as_MergeMem(); 1062 Node* new_st = merge->memory_at(alias_idx); 1063 if (new_st == merge->base_memory()) { 1064 // Keep searching 1065 current = new_st; 1066 continue; 1067 } 1068 // Save the new memory state for the slice and fall through 1069 // to exit. 1070 result = new_st; 1071 } 1072 } 1073 break; 1074 } 1075 if (result != NULL) { 1076 st = result; 1077 } 1078 } 1079 1080 // Loop around twice in the case Load -> Initialize -> Store. 1081 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1082 for (int trip = 0; trip <= 1; trip++) { 1083 1084 if (st->is_Store()) { 1085 Node* st_adr = st->in(MemNode::Address); 1086 if (!phase->eqv(st_adr, ld_adr)) { 1087 // Try harder before giving up... Match raw and non-raw pointers. 1088 intptr_t st_off = 0; 1089 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 1090 if (alloc == NULL) return NULL; 1091 if (alloc != ld_alloc) return NULL; 1092 if (ld_off != st_off) return NULL; 1093 // At this point we have proven something like this setup: 1094 // A = Allocate(...) 1095 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 1096 // S = StoreQ(, AddP(, A.Parm , #Off), V) 1097 // (Actually, we haven't yet proven the Q's are the same.) 1098 // In other words, we are loading from a casted version of 1099 // the same pointer-and-offset that we stored to. 1100 // Thus, we are able to replace L by V. 1101 } 1102 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1103 if (store_Opcode() != st->Opcode()) 1104 return NULL; 1105 return st->in(MemNode::ValueIn); 1106 } 1107 1108 // A load from a freshly-created object always returns zero. 1109 // (This can happen after LoadNode::Ideal resets the load's memory input 1110 // to find_captured_store, which returned InitializeNode::zero_memory.) 1111 if (st->is_Proj() && st->in(0)->is_Allocate() && 1112 (st->in(0) == ld_alloc) && 1113 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1114 // return a zero value for the load's basic type 1115 // (This is one of the few places where a generic PhaseTransform 1116 // can create new nodes. Think of it as lazily manifesting 1117 // virtually pre-existing constants.) 1118 if (ReduceBulkZeroing || find_array_copy_clone(phase, ld_alloc, in(MemNode::Memory)) == NULL) { 1119 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an 1120 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done 1121 // by the ArrayCopyNode. 1122 return phase->zerocon(memory_type()); 1123 } 1124 } 1125 1126 // A load from an initialization barrier can match a captured store. 1127 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1128 InitializeNode* init = st->in(0)->as_Initialize(); 1129 AllocateNode* alloc = init->allocation(); 1130 if ((alloc != NULL) && (alloc == ld_alloc)) { 1131 // examine a captured store value 1132 st = init->find_captured_store(ld_off, memory_size(), phase); 1133 if (st != NULL) { 1134 continue; // take one more trip around 1135 } 1136 } 1137 } 1138 1139 // Load boxed value from result of valueOf() call is input parameter. 1140 if (this->is_Load() && ld_adr->is_AddP() && 1141 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1142 intptr_t ignore = 0; 1143 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1144 #if INCLUDE_SHENANDOAHGC 1145 if (UseShenandoahGC) { 1146 base = ((ShenandoahBarrierSetC2*) BarrierSet::barrier_set()->barrier_set_c2())->step_over_gc_barrier(base); 1147 } 1148 #endif 1149 if (base != NULL && base->is_Proj() && 1150 base->as_Proj()->_con == TypeFunc::Parms && 1151 base->in(0)->is_CallStaticJava() && 1152 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1153 return base->in(0)->in(TypeFunc::Parms); 1154 } 1155 } 1156 1157 break; 1158 } 1159 1160 return NULL; 1161 } 1162 1163 //----------------------is_instance_field_load_with_local_phi------------------ 1164 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1165 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1166 in(Address)->is_AddP() ) { 1167 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1168 // Only instances and boxed values. 1169 if( t_oop != NULL && 1170 (t_oop->is_ptr_to_boxed_value() || 1171 t_oop->is_known_instance_field()) && 1172 t_oop->offset() != Type::OffsetBot && 1173 t_oop->offset() != Type::OffsetTop) { 1174 return true; 1175 } 1176 } 1177 return false; 1178 } 1179 1180 //------------------------------Identity--------------------------------------- 1181 // Loads are identity if previous store is to same address 1182 Node* LoadNode::Identity(PhaseGVN* phase) { 1183 // If the previous store-maker is the right kind of Store, and the store is 1184 // to the same address, then we are equal to the value stored. 1185 Node* mem = in(Memory); 1186 Node* value = can_see_stored_value(mem, phase); 1187 if( value ) { 1188 // byte, short & char stores truncate naturally. 1189 // A load has to load the truncated value which requires 1190 // some sort of masking operation and that requires an 1191 // Ideal call instead of an Identity call. 1192 if (memory_size() < BytesPerInt) { 1193 // If the input to the store does not fit with the load's result type, 1194 // it must be truncated via an Ideal call. 1195 if (!phase->type(value)->higher_equal(phase->type(this))) 1196 return this; 1197 } 1198 // (This works even when value is a Con, but LoadNode::Value 1199 // usually runs first, producing the singleton type of the Con.) 1200 return value; 1201 } 1202 1203 // Search for an existing data phi which was generated before for the same 1204 // instance's field to avoid infinite generation of phis in a loop. 1205 Node *region = mem->in(0); 1206 if (is_instance_field_load_with_local_phi(region)) { 1207 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1208 int this_index = phase->C->get_alias_index(addr_t); 1209 int this_offset = addr_t->offset(); 1210 int this_iid = addr_t->instance_id(); 1211 if (!addr_t->is_known_instance() && 1212 addr_t->is_ptr_to_boxed_value()) { 1213 // Use _idx of address base (could be Phi node) for boxed values. 1214 intptr_t ignore = 0; 1215 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1216 if (base == NULL) { 1217 return this; 1218 } 1219 this_iid = base->_idx; 1220 } 1221 const Type* this_type = bottom_type(); 1222 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1223 Node* phi = region->fast_out(i); 1224 if (phi->is_Phi() && phi != mem && 1225 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1226 return phi; 1227 } 1228 } 1229 } 1230 1231 return this; 1232 } 1233 1234 // Construct an equivalent unsigned load. 1235 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1236 BasicType bt = T_ILLEGAL; 1237 const Type* rt = NULL; 1238 switch (Opcode()) { 1239 case Op_LoadUB: return this; 1240 case Op_LoadUS: return this; 1241 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1242 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1243 default: 1244 assert(false, "no unsigned variant: %s", Name()); 1245 return NULL; 1246 } 1247 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1248 raw_adr_type(), rt, bt, _mo, _control_dependency, 1249 is_unaligned_access(), is_mismatched_access()); 1250 } 1251 1252 // Construct an equivalent signed load. 1253 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1254 BasicType bt = T_ILLEGAL; 1255 const Type* rt = NULL; 1256 switch (Opcode()) { 1257 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1258 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1259 case Op_LoadB: // fall through 1260 case Op_LoadS: // fall through 1261 case Op_LoadI: // fall through 1262 case Op_LoadL: return this; 1263 default: 1264 assert(false, "no signed variant: %s", Name()); 1265 return NULL; 1266 } 1267 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1268 raw_adr_type(), rt, bt, _mo, _control_dependency, 1269 is_unaligned_access(), is_mismatched_access()); 1270 } 1271 1272 // We're loading from an object which has autobox behaviour. 1273 // If this object is result of a valueOf call we'll have a phi 1274 // merging a newly allocated object and a load from the cache. 1275 // We want to replace this load with the original incoming 1276 // argument to the valueOf call. 1277 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1278 assert(phase->C->eliminate_boxing(), "sanity"); 1279 intptr_t ignore = 0; 1280 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1281 if ((base == NULL) || base->is_Phi()) { 1282 // Push the loads from the phi that comes from valueOf up 1283 // through it to allow elimination of the loads and the recovery 1284 // of the original value. It is done in split_through_phi(). 1285 return NULL; 1286 } else if (base->is_Load() || 1287 (base->is_DecodeN() && base->in(1)->is_Load())) { 1288 // Eliminate the load of boxed value for integer types from the cache 1289 // array by deriving the value from the index into the array. 1290 // Capture the offset of the load and then reverse the computation. 1291 1292 // Get LoadN node which loads a boxing object from 'cache' array. 1293 if (base->is_DecodeN()) { 1294 base = base->in(1); 1295 } 1296 if (!base->in(Address)->is_AddP()) { 1297 return NULL; // Complex address 1298 } 1299 AddPNode* address = base->in(Address)->as_AddP(); 1300 Node* cache_base = address->in(AddPNode::Base); 1301 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1302 // Get ConP node which is static 'cache' field. 1303 cache_base = cache_base->in(1); 1304 } 1305 if ((cache_base != NULL) && cache_base->is_Con()) { 1306 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1307 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1308 Node* elements[4]; 1309 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1310 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1311 if (count > 0 && elements[0]->is_Con() && 1312 (count == 1 || 1313 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1314 elements[1]->in(2) == phase->intcon(shift)))) { 1315 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1316 // Fetch the box object cache[0] at the base of the array and get its value 1317 ciInstance* box = array->obj_at(0)->as_instance(); 1318 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1319 assert(ik->is_box_klass(), "sanity"); 1320 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1321 if (ik->nof_nonstatic_fields() == 1) { 1322 // This should be true nonstatic_field_at requires calling 1323 // nof_nonstatic_fields so check it anyway 1324 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1325 BasicType bt = c.basic_type(); 1326 // Only integer types have boxing cache. 1327 assert(bt == T_BOOLEAN || bt == T_CHAR || 1328 bt == T_BYTE || bt == T_SHORT || 1329 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1330 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1331 if (cache_low != (int)cache_low) { 1332 return NULL; // should not happen since cache is array indexed by value 1333 } 1334 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1335 if (offset != (int)offset) { 1336 return NULL; // should not happen since cache is array indexed by value 1337 } 1338 // Add up all the offsets making of the address of the load 1339 Node* result = elements[0]; 1340 for (int i = 1; i < count; i++) { 1341 result = phase->transform(new AddXNode(result, elements[i])); 1342 } 1343 // Remove the constant offset from the address and then 1344 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1345 // remove the scaling of the offset to recover the original index. 1346 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1347 // Peel the shift off directly but wrap it in a dummy node 1348 // since Ideal can't return existing nodes 1349 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1350 } else if (result->is_Add() && result->in(2)->is_Con() && 1351 result->in(1)->Opcode() == Op_LShiftX && 1352 result->in(1)->in(2) == phase->intcon(shift)) { 1353 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1354 // but for boxing cache access we know that X<<Z will not overflow 1355 // (there is range check) so we do this optimizatrion by hand here. 1356 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1357 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1358 } else { 1359 result = new RShiftXNode(result, phase->intcon(shift)); 1360 } 1361 #ifdef _LP64 1362 if (bt != T_LONG) { 1363 result = new ConvL2INode(phase->transform(result)); 1364 } 1365 #else 1366 if (bt == T_LONG) { 1367 result = new ConvI2LNode(phase->transform(result)); 1368 } 1369 #endif 1370 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1371 // Need to preserve unboxing load type if it is unsigned. 1372 switch(this->Opcode()) { 1373 case Op_LoadUB: 1374 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1375 break; 1376 case Op_LoadUS: 1377 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1378 break; 1379 } 1380 return result; 1381 } 1382 } 1383 } 1384 } 1385 } 1386 return NULL; 1387 } 1388 1389 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1390 Node* region = phi->in(0); 1391 if (region == NULL) { 1392 return false; // Wait stable graph 1393 } 1394 uint cnt = phi->req(); 1395 for (uint i = 1; i < cnt; i++) { 1396 Node* rc = region->in(i); 1397 if (rc == NULL || phase->type(rc) == Type::TOP) 1398 return false; // Wait stable graph 1399 Node* in = phi->in(i); 1400 if (in == NULL || phase->type(in) == Type::TOP) 1401 return false; // Wait stable graph 1402 } 1403 return true; 1404 } 1405 //------------------------------split_through_phi------------------------------ 1406 // Split instance or boxed field load through Phi. 1407 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1408 Node* mem = in(Memory); 1409 Node* address = in(Address); 1410 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1411 1412 assert((t_oop != NULL) && 1413 (t_oop->is_known_instance_field() || 1414 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1415 1416 Compile* C = phase->C; 1417 intptr_t ignore = 0; 1418 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1419 bool base_is_phi = (base != NULL) && base->is_Phi(); 1420 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1421 (base != NULL) && (base == address->in(AddPNode::Base)) && 1422 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1423 1424 if (!((mem->is_Phi() || base_is_phi) && 1425 (load_boxed_values || t_oop->is_known_instance_field()))) { 1426 return NULL; // memory is not Phi 1427 } 1428 1429 if (mem->is_Phi()) { 1430 if (!stable_phi(mem->as_Phi(), phase)) { 1431 return NULL; // Wait stable graph 1432 } 1433 uint cnt = mem->req(); 1434 // Check for loop invariant memory. 1435 if (cnt == 3) { 1436 for (uint i = 1; i < cnt; i++) { 1437 Node* in = mem->in(i); 1438 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1439 if (m == mem) { 1440 if (i == 1) { 1441 // if the first edge was a loop, check second edge too. 1442 // If both are replaceable - we are in an infinite loop 1443 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1444 if (n == mem) { 1445 break; 1446 } 1447 } 1448 set_req(Memory, mem->in(cnt - i)); 1449 return this; // made change 1450 } 1451 } 1452 } 1453 } 1454 if (base_is_phi) { 1455 if (!stable_phi(base->as_Phi(), phase)) { 1456 return NULL; // Wait stable graph 1457 } 1458 uint cnt = base->req(); 1459 // Check for loop invariant memory. 1460 if (cnt == 3) { 1461 for (uint i = 1; i < cnt; i++) { 1462 if (base->in(i) == base) { 1463 return NULL; // Wait stable graph 1464 } 1465 } 1466 } 1467 } 1468 1469 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1470 1471 // Split through Phi (see original code in loopopts.cpp). 1472 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1473 1474 // Do nothing here if Identity will find a value 1475 // (to avoid infinite chain of value phis generation). 1476 if (!phase->eqv(this, phase->apply_identity(this))) 1477 return NULL; 1478 1479 // Select Region to split through. 1480 Node* region; 1481 if (!base_is_phi) { 1482 assert(mem->is_Phi(), "sanity"); 1483 region = mem->in(0); 1484 // Skip if the region dominates some control edge of the address. 1485 if (!MemNode::all_controls_dominate(address, region)) 1486 return NULL; 1487 } else if (!mem->is_Phi()) { 1488 assert(base_is_phi, "sanity"); 1489 region = base->in(0); 1490 // Skip if the region dominates some control edge of the memory. 1491 if (!MemNode::all_controls_dominate(mem, region)) 1492 return NULL; 1493 } else if (base->in(0) != mem->in(0)) { 1494 assert(base_is_phi && mem->is_Phi(), "sanity"); 1495 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1496 region = base->in(0); 1497 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1498 region = mem->in(0); 1499 } else { 1500 return NULL; // complex graph 1501 } 1502 } else { 1503 assert(base->in(0) == mem->in(0), "sanity"); 1504 region = mem->in(0); 1505 } 1506 1507 const Type* this_type = this->bottom_type(); 1508 int this_index = C->get_alias_index(t_oop); 1509 int this_offset = t_oop->offset(); 1510 int this_iid = t_oop->instance_id(); 1511 if (!t_oop->is_known_instance() && load_boxed_values) { 1512 // Use _idx of address base for boxed values. 1513 this_iid = base->_idx; 1514 } 1515 PhaseIterGVN* igvn = phase->is_IterGVN(); 1516 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1517 for (uint i = 1; i < region->req(); i++) { 1518 Node* x; 1519 Node* the_clone = NULL; 1520 Node* in = region->in(i); 1521 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl && 1522 in != NULL && in->is_OuterStripMinedLoop()) { 1523 // No node should go in the outer strip mined loop 1524 in = in->in(LoopNode::EntryControl); 1525 } 1526 if (in == NULL || in == C->top()) { 1527 x = C->top(); // Dead path? Use a dead data op 1528 } else { 1529 x = this->clone(); // Else clone up the data op 1530 the_clone = x; // Remember for possible deletion. 1531 // Alter data node to use pre-phi inputs 1532 if (this->in(0) == region) { 1533 x->set_req(0, in); 1534 } else { 1535 x->set_req(0, NULL); 1536 } 1537 if (mem->is_Phi() && (mem->in(0) == region)) { 1538 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1539 } 1540 if (address->is_Phi() && address->in(0) == region) { 1541 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1542 } 1543 if (base_is_phi && (base->in(0) == region)) { 1544 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1545 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1546 x->set_req(Address, adr_x); 1547 } 1548 } 1549 // Check for a 'win' on some paths 1550 const Type *t = x->Value(igvn); 1551 1552 bool singleton = t->singleton(); 1553 1554 // See comments in PhaseIdealLoop::split_thru_phi(). 1555 if (singleton && t == Type::TOP) { 1556 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1557 } 1558 1559 if (singleton) { 1560 x = igvn->makecon(t); 1561 } else { 1562 // We now call Identity to try to simplify the cloned node. 1563 // Note that some Identity methods call phase->type(this). 1564 // Make sure that the type array is big enough for 1565 // our new node, even though we may throw the node away. 1566 // (This tweaking with igvn only works because x is a new node.) 1567 igvn->set_type(x, t); 1568 // If x is a TypeNode, capture any more-precise type permanently into Node 1569 // otherwise it will be not updated during igvn->transform since 1570 // igvn->type(x) is set to x->Value() already. 1571 x->raise_bottom_type(t); 1572 Node *y = igvn->apply_identity(x); 1573 if (y != x) { 1574 x = y; 1575 } else { 1576 y = igvn->hash_find_insert(x); 1577 if (y) { 1578 x = y; 1579 } else { 1580 // Else x is a new node we are keeping 1581 // We do not need register_new_node_with_optimizer 1582 // because set_type has already been called. 1583 igvn->_worklist.push(x); 1584 } 1585 } 1586 } 1587 if (x != the_clone && the_clone != NULL) { 1588 igvn->remove_dead_node(the_clone); 1589 } 1590 phi->set_req(i, x); 1591 } 1592 // Record Phi 1593 igvn->register_new_node_with_optimizer(phi); 1594 return phi; 1595 } 1596 1597 //------------------------------Ideal------------------------------------------ 1598 // If the load is from Field memory and the pointer is non-null, it might be possible to 1599 // zero out the control input. 1600 // If the offset is constant and the base is an object allocation, 1601 // try to hook me up to the exact initializing store. 1602 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1603 Node* p = MemNode::Ideal_common(phase, can_reshape); 1604 if (p) return (p == NodeSentinel) ? NULL : p; 1605 1606 Node* ctrl = in(MemNode::Control); 1607 Node* address = in(MemNode::Address); 1608 bool progress = false; 1609 1610 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1611 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1612 1613 // Skip up past a SafePoint control. Cannot do this for Stores because 1614 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1615 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1616 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1617 !addr_mark ) { 1618 ctrl = ctrl->in(0); 1619 set_req(MemNode::Control,ctrl); 1620 progress = true; 1621 } 1622 1623 intptr_t ignore = 0; 1624 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1625 if (base != NULL 1626 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1627 // Check for useless control edge in some common special cases 1628 if (in(MemNode::Control) != NULL 1629 && can_remove_control() 1630 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1631 && all_controls_dominate(base, phase->C->start())) { 1632 // A method-invariant, non-null address (constant or 'this' argument). 1633 set_req(MemNode::Control, NULL); 1634 progress = true; 1635 } 1636 } 1637 1638 Node* mem = in(MemNode::Memory); 1639 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1640 1641 if (can_reshape && (addr_t != NULL)) { 1642 // try to optimize our memory input 1643 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1644 if (opt_mem != mem) { 1645 set_req(MemNode::Memory, opt_mem); 1646 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1647 return this; 1648 } 1649 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1650 if ((t_oop != NULL) && 1651 (t_oop->is_known_instance_field() || 1652 t_oop->is_ptr_to_boxed_value())) { 1653 PhaseIterGVN *igvn = phase->is_IterGVN(); 1654 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1655 // Delay this transformation until memory Phi is processed. 1656 phase->is_IterGVN()->_worklist.push(this); 1657 return NULL; 1658 } 1659 // Split instance field load through Phi. 1660 Node* result = split_through_phi(phase); 1661 if (result != NULL) return result; 1662 1663 if (t_oop->is_ptr_to_boxed_value()) { 1664 Node* result = eliminate_autobox(phase); 1665 if (result != NULL) return result; 1666 } 1667 } 1668 } 1669 1670 // Is there a dominating load that loads the same value? Leave 1671 // anything that is not a load of a field/array element (like 1672 // barriers etc.) alone 1673 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) { 1674 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1675 Node *use = mem->fast_out(i); 1676 if (use != this && 1677 use->Opcode() == Opcode() && 1678 use->in(0) != NULL && 1679 use->in(0) != in(0) && 1680 use->in(Address) == in(Address)) { 1681 Node* ctl = in(0); 1682 for (int i = 0; i < 10 && ctl != NULL; i++) { 1683 ctl = IfNode::up_one_dom(ctl); 1684 if (ctl == use->in(0)) { 1685 set_req(0, use->in(0)); 1686 return this; 1687 } 1688 } 1689 } 1690 } 1691 } 1692 1693 // Check for prior store with a different base or offset; make Load 1694 // independent. Skip through any number of them. Bail out if the stores 1695 // are in an endless dead cycle and report no progress. This is a key 1696 // transform for Reflection. However, if after skipping through the Stores 1697 // we can't then fold up against a prior store do NOT do the transform as 1698 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1699 // array memory alive twice: once for the hoisted Load and again after the 1700 // bypassed Store. This situation only works if EVERYBODY who does 1701 // anti-dependence work knows how to bypass. I.e. we need all 1702 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1703 // the alias index stuff. So instead, peek through Stores and IFF we can 1704 // fold up, do so. 1705 Node* prev_mem = find_previous_store(phase); 1706 if (prev_mem != NULL) { 1707 Node* value = can_see_arraycopy_value(prev_mem, phase); 1708 if (value != NULL) { 1709 return value; 1710 } 1711 } 1712 // Steps (a), (b): Walk past independent stores to find an exact match. 1713 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1714 // (c) See if we can fold up on the spot, but don't fold up here. 1715 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1716 // just return a prior value, which is done by Identity calls. 1717 if (can_see_stored_value(prev_mem, phase)) { 1718 // Make ready for step (d): 1719 set_req(MemNode::Memory, prev_mem); 1720 return this; 1721 } 1722 } 1723 1724 return progress ? this : NULL; 1725 } 1726 1727 // Helper to recognize certain Klass fields which are invariant across 1728 // some group of array types (e.g., int[] or all T[] where T < Object). 1729 const Type* 1730 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1731 ciKlass* klass) const { 1732 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1733 // The field is Klass::_modifier_flags. Return its (constant) value. 1734 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1735 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1736 return TypeInt::make(klass->modifier_flags()); 1737 } 1738 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1739 // The field is Klass::_access_flags. Return its (constant) value. 1740 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1741 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1742 return TypeInt::make(klass->access_flags()); 1743 } 1744 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1745 // The field is Klass::_layout_helper. Return its constant value if known. 1746 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1747 return TypeInt::make(klass->layout_helper()); 1748 } 1749 1750 // No match. 1751 return NULL; 1752 } 1753 1754 //------------------------------Value----------------------------------------- 1755 const Type* LoadNode::Value(PhaseGVN* phase) const { 1756 // Either input is TOP ==> the result is TOP 1757 Node* mem = in(MemNode::Memory); 1758 const Type *t1 = phase->type(mem); 1759 if (t1 == Type::TOP) return Type::TOP; 1760 Node* adr = in(MemNode::Address); 1761 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1762 if (tp == NULL || tp->empty()) return Type::TOP; 1763 int off = tp->offset(); 1764 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1765 Compile* C = phase->C; 1766 1767 // Try to guess loaded type from pointer type 1768 if (tp->isa_aryptr()) { 1769 const TypeAryPtr* ary = tp->is_aryptr(); 1770 const Type* t = ary->elem(); 1771 1772 // Determine whether the reference is beyond the header or not, by comparing 1773 // the offset against the offset of the start of the array's data. 1774 // Different array types begin at slightly different offsets (12 vs. 16). 1775 // We choose T_BYTE as an example base type that is least restrictive 1776 // as to alignment, which will therefore produce the smallest 1777 // possible base offset. 1778 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1779 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1780 1781 // Try to constant-fold a stable array element. 1782 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1783 // Make sure the reference is not into the header and the offset is constant 1784 ciObject* aobj = ary->const_oop(); 1785 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1786 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1787 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1788 stable_dimension, 1789 memory_type(), is_unsigned()); 1790 if (con_type != NULL) { 1791 return con_type; 1792 } 1793 } 1794 } 1795 1796 // Don't do this for integer types. There is only potential profit if 1797 // the element type t is lower than _type; that is, for int types, if _type is 1798 // more restrictive than t. This only happens here if one is short and the other 1799 // char (both 16 bits), and in those cases we've made an intentional decision 1800 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1801 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1802 // 1803 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1804 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1805 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1806 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1807 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1808 // In fact, that could have been the original type of p1, and p1 could have 1809 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1810 // expression (LShiftL quux 3) independently optimized to the constant 8. 1811 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1812 && (_type->isa_vect() == NULL) 1813 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1814 // t might actually be lower than _type, if _type is a unique 1815 // concrete subclass of abstract class t. 1816 if (off_beyond_header) { // is the offset beyond the header? 1817 const Type* jt = t->join_speculative(_type); 1818 // In any case, do not allow the join, per se, to empty out the type. 1819 if (jt->empty() && !t->empty()) { 1820 // This can happen if a interface-typed array narrows to a class type. 1821 jt = _type; 1822 } 1823 #ifdef ASSERT 1824 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1825 // The pointers in the autobox arrays are always non-null 1826 Node* base = adr->in(AddPNode::Base); 1827 if ((base != NULL) && base->is_DecodeN()) { 1828 // Get LoadN node which loads IntegerCache.cache field 1829 base = base->in(1); 1830 } 1831 if ((base != NULL) && base->is_Con()) { 1832 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1833 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1834 // It could be narrow oop 1835 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1836 } 1837 } 1838 } 1839 #endif 1840 return jt; 1841 } 1842 } 1843 } else if (tp->base() == Type::InstPtr) { 1844 assert( off != Type::OffsetBot || 1845 // arrays can be cast to Objects 1846 tp->is_oopptr()->klass()->is_java_lang_Object() || 1847 // unsafe field access may not have a constant offset 1848 C->has_unsafe_access(), 1849 "Field accesses must be precise" ); 1850 // For oop loads, we expect the _type to be precise. 1851 1852 // Optimize loads from constant fields. 1853 const TypeInstPtr* tinst = tp->is_instptr(); 1854 ciObject* const_oop = tinst->const_oop(); 1855 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1856 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type()); 1857 if (con_type != NULL) { 1858 return con_type; 1859 } 1860 } 1861 } else if (tp->base() == Type::KlassPtr) { 1862 assert( off != Type::OffsetBot || 1863 // arrays can be cast to Objects 1864 tp->is_klassptr()->klass()->is_java_lang_Object() || 1865 // also allow array-loading from the primary supertype 1866 // array during subtype checks 1867 Opcode() == Op_LoadKlass, 1868 "Field accesses must be precise" ); 1869 // For klass/static loads, we expect the _type to be precise 1870 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) { 1871 /* With mirrors being an indirect in the Klass* 1872 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 1873 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 1874 * 1875 * So check the type and klass of the node before the LoadP. 1876 */ 1877 Node* adr2 = adr->in(MemNode::Address); 1878 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1879 if (tkls != NULL && !StressReflectiveCode) { 1880 ciKlass* klass = tkls->klass(); 1881 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1882 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1883 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1884 return TypeInstPtr::make(klass->java_mirror()); 1885 } 1886 } 1887 } 1888 1889 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1890 if (tkls != NULL && !StressReflectiveCode) { 1891 ciKlass* klass = tkls->klass(); 1892 if (klass->is_loaded() && tkls->klass_is_exact()) { 1893 // We are loading a field from a Klass metaobject whose identity 1894 // is known at compile time (the type is "exact" or "precise"). 1895 // Check for fields we know are maintained as constants by the VM. 1896 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1897 // The field is Klass::_super_check_offset. Return its (constant) value. 1898 // (Folds up type checking code.) 1899 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1900 return TypeInt::make(klass->super_check_offset()); 1901 } 1902 // Compute index into primary_supers array 1903 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1904 // Check for overflowing; use unsigned compare to handle the negative case. 1905 if( depth < ciKlass::primary_super_limit() ) { 1906 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1907 // (Folds up type checking code.) 1908 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1909 ciKlass *ss = klass->super_of_depth(depth); 1910 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1911 } 1912 const Type* aift = load_array_final_field(tkls, klass); 1913 if (aift != NULL) return aift; 1914 } 1915 1916 // We can still check if we are loading from the primary_supers array at a 1917 // shallow enough depth. Even though the klass is not exact, entries less 1918 // than or equal to its super depth are correct. 1919 if (klass->is_loaded() ) { 1920 ciType *inner = klass; 1921 while( inner->is_obj_array_klass() ) 1922 inner = inner->as_obj_array_klass()->base_element_type(); 1923 if( inner->is_instance_klass() && 1924 !inner->as_instance_klass()->flags().is_interface() ) { 1925 // Compute index into primary_supers array 1926 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1927 // Check for overflowing; use unsigned compare to handle the negative case. 1928 if( depth < ciKlass::primary_super_limit() && 1929 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1930 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1931 // (Folds up type checking code.) 1932 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1933 ciKlass *ss = klass->super_of_depth(depth); 1934 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1935 } 1936 } 1937 } 1938 1939 // If the type is enough to determine that the thing is not an array, 1940 // we can give the layout_helper a positive interval type. 1941 // This will help short-circuit some reflective code. 1942 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1943 && !klass->is_array_klass() // not directly typed as an array 1944 && !klass->is_interface() // specifically not Serializable & Cloneable 1945 && !klass->is_java_lang_Object() // not the supertype of all T[] 1946 ) { 1947 // Note: When interfaces are reliable, we can narrow the interface 1948 // test to (klass != Serializable && klass != Cloneable). 1949 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1950 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1951 // The key property of this type is that it folds up tests 1952 // for array-ness, since it proves that the layout_helper is positive. 1953 // Thus, a generic value like the basic object layout helper works fine. 1954 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1955 } 1956 } 1957 1958 // If we are loading from a freshly-allocated object, produce a zero, 1959 // if the load is provably beyond the header of the object. 1960 // (Also allow a variable load from a fresh array to produce zero.) 1961 const TypeOopPtr *tinst = tp->isa_oopptr(); 1962 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1963 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1964 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1965 Node* value = can_see_stored_value(mem,phase); 1966 if (value != NULL && value->is_Con()) { 1967 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1968 return value->bottom_type(); 1969 } 1970 } 1971 1972 if (is_instance) { 1973 // If we have an instance type and our memory input is the 1974 // programs's initial memory state, there is no matching store, 1975 // so just return a zero of the appropriate type 1976 Node *mem = in(MemNode::Memory); 1977 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1978 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1979 return Type::get_zero_type(_type->basic_type()); 1980 } 1981 } 1982 return _type; 1983 } 1984 1985 //------------------------------match_edge------------------------------------- 1986 // Do we Match on this edge index or not? Match only the address. 1987 uint LoadNode::match_edge(uint idx) const { 1988 return idx == MemNode::Address; 1989 } 1990 1991 //--------------------------LoadBNode::Ideal-------------------------------------- 1992 // 1993 // If the previous store is to the same address as this load, 1994 // and the value stored was larger than a byte, replace this load 1995 // with the value stored truncated to a byte. If no truncation is 1996 // needed, the replacement is done in LoadNode::Identity(). 1997 // 1998 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1999 Node* mem = in(MemNode::Memory); 2000 Node* value = can_see_stored_value(mem,phase); 2001 if (value != NULL) { 2002 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false); 2003 if (narrow != value) { 2004 return narrow; 2005 } 2006 } 2007 // Identity call will handle the case where truncation is not needed. 2008 return LoadNode::Ideal(phase, can_reshape); 2009 } 2010 2011 const Type* LoadBNode::Value(PhaseGVN* phase) const { 2012 Node* mem = in(MemNode::Memory); 2013 Node* value = can_see_stored_value(mem,phase); 2014 if (value != NULL && value->is_Con() && 2015 !value->bottom_type()->higher_equal(_type)) { 2016 // If the input to the store does not fit with the load's result type, 2017 // it must be truncated. We can't delay until Ideal call since 2018 // a singleton Value is needed for split_thru_phi optimization. 2019 int con = value->get_int(); 2020 return TypeInt::make((con << 24) >> 24); 2021 } 2022 return LoadNode::Value(phase); 2023 } 2024 2025 //--------------------------LoadUBNode::Ideal------------------------------------- 2026 // 2027 // If the previous store is to the same address as this load, 2028 // and the value stored was larger than a byte, replace this load 2029 // with the value stored truncated to a byte. If no truncation is 2030 // needed, the replacement is done in LoadNode::Identity(). 2031 // 2032 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2033 Node* mem = in(MemNode::Memory); 2034 Node* value = can_see_stored_value(mem, phase); 2035 if (value != NULL) { 2036 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false); 2037 if (narrow != value) { 2038 return narrow; 2039 } 2040 } 2041 // Identity call will handle the case where truncation is not needed. 2042 return LoadNode::Ideal(phase, can_reshape); 2043 } 2044 2045 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2046 Node* mem = in(MemNode::Memory); 2047 Node* value = can_see_stored_value(mem,phase); 2048 if (value != NULL && value->is_Con() && 2049 !value->bottom_type()->higher_equal(_type)) { 2050 // If the input to the store does not fit with the load's result type, 2051 // it must be truncated. We can't delay until Ideal call since 2052 // a singleton Value is needed for split_thru_phi optimization. 2053 int con = value->get_int(); 2054 return TypeInt::make(con & 0xFF); 2055 } 2056 return LoadNode::Value(phase); 2057 } 2058 2059 //--------------------------LoadUSNode::Ideal------------------------------------- 2060 // 2061 // If the previous store is to the same address as this load, 2062 // and the value stored was larger than a char, replace this load 2063 // with the value stored truncated to a char. If no truncation is 2064 // needed, the replacement is done in LoadNode::Identity(). 2065 // 2066 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2067 Node* mem = in(MemNode::Memory); 2068 Node* value = can_see_stored_value(mem,phase); 2069 if (value != NULL) { 2070 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false); 2071 if (narrow != value) { 2072 return narrow; 2073 } 2074 } 2075 // Identity call will handle the case where truncation is not needed. 2076 return LoadNode::Ideal(phase, can_reshape); 2077 } 2078 2079 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2080 Node* mem = in(MemNode::Memory); 2081 Node* value = can_see_stored_value(mem,phase); 2082 if (value != NULL && value->is_Con() && 2083 !value->bottom_type()->higher_equal(_type)) { 2084 // If the input to the store does not fit with the load's result type, 2085 // it must be truncated. We can't delay until Ideal call since 2086 // a singleton Value is needed for split_thru_phi optimization. 2087 int con = value->get_int(); 2088 return TypeInt::make(con & 0xFFFF); 2089 } 2090 return LoadNode::Value(phase); 2091 } 2092 2093 //--------------------------LoadSNode::Ideal-------------------------------------- 2094 // 2095 // If the previous store is to the same address as this load, 2096 // and the value stored was larger than a short, replace this load 2097 // with the value stored truncated to a short. If no truncation is 2098 // needed, the replacement is done in LoadNode::Identity(). 2099 // 2100 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2101 Node* mem = in(MemNode::Memory); 2102 Node* value = can_see_stored_value(mem,phase); 2103 if (value != NULL) { 2104 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false); 2105 if (narrow != value) { 2106 return narrow; 2107 } 2108 } 2109 // Identity call will handle the case where truncation is not needed. 2110 return LoadNode::Ideal(phase, can_reshape); 2111 } 2112 2113 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2114 Node* mem = in(MemNode::Memory); 2115 Node* value = can_see_stored_value(mem,phase); 2116 if (value != NULL && value->is_Con() && 2117 !value->bottom_type()->higher_equal(_type)) { 2118 // If the input to the store does not fit with the load's result type, 2119 // it must be truncated. We can't delay until Ideal call since 2120 // a singleton Value is needed for split_thru_phi optimization. 2121 int con = value->get_int(); 2122 return TypeInt::make((con << 16) >> 16); 2123 } 2124 return LoadNode::Value(phase); 2125 } 2126 2127 //============================================================================= 2128 //----------------------------LoadKlassNode::make------------------------------ 2129 // Polymorphic factory method: 2130 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2131 // sanity check the alias category against the created node type 2132 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2133 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2134 #ifdef _LP64 2135 if (adr_type->is_ptr_to_narrowklass()) { 2136 assert(UseCompressedClassPointers, "no compressed klasses"); 2137 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2138 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2139 } 2140 #endif 2141 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2142 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2143 } 2144 2145 //------------------------------Value------------------------------------------ 2146 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2147 return klass_value_common(phase); 2148 } 2149 2150 // In most cases, LoadKlassNode does not have the control input set. If the control 2151 // input is set, it must not be removed (by LoadNode::Ideal()). 2152 bool LoadKlassNode::can_remove_control() const { 2153 return false; 2154 } 2155 2156 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2157 // Either input is TOP ==> the result is TOP 2158 const Type *t1 = phase->type( in(MemNode::Memory) ); 2159 if (t1 == Type::TOP) return Type::TOP; 2160 Node *adr = in(MemNode::Address); 2161 const Type *t2 = phase->type( adr ); 2162 if (t2 == Type::TOP) return Type::TOP; 2163 const TypePtr *tp = t2->is_ptr(); 2164 if (TypePtr::above_centerline(tp->ptr()) || 2165 tp->ptr() == TypePtr::Null) return Type::TOP; 2166 2167 // Return a more precise klass, if possible 2168 const TypeInstPtr *tinst = tp->isa_instptr(); 2169 if (tinst != NULL) { 2170 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2171 int offset = tinst->offset(); 2172 if (ik == phase->C->env()->Class_klass() 2173 && (offset == java_lang_Class::klass_offset_in_bytes() || 2174 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2175 // We are loading a special hidden field from a Class mirror object, 2176 // the field which points to the VM's Klass metaobject. 2177 ciType* t = tinst->java_mirror_type(); 2178 // java_mirror_type returns non-null for compile-time Class constants. 2179 if (t != NULL) { 2180 // constant oop => constant klass 2181 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2182 if (t->is_void()) { 2183 // We cannot create a void array. Since void is a primitive type return null 2184 // klass. Users of this result need to do a null check on the returned klass. 2185 return TypePtr::NULL_PTR; 2186 } 2187 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2188 } 2189 if (!t->is_klass()) { 2190 // a primitive Class (e.g., int.class) has NULL for a klass field 2191 return TypePtr::NULL_PTR; 2192 } 2193 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2194 return TypeKlassPtr::make(t->as_klass()); 2195 } 2196 // non-constant mirror, so we can't tell what's going on 2197 } 2198 if( !ik->is_loaded() ) 2199 return _type; // Bail out if not loaded 2200 if (offset == oopDesc::klass_offset_in_bytes()) { 2201 if (tinst->klass_is_exact()) { 2202 return TypeKlassPtr::make(ik); 2203 } 2204 // See if we can become precise: no subklasses and no interface 2205 // (Note: We need to support verified interfaces.) 2206 if (!ik->is_interface() && !ik->has_subklass()) { 2207 //assert(!UseExactTypes, "this code should be useless with exact types"); 2208 // Add a dependence; if any subclass added we need to recompile 2209 if (!ik->is_final()) { 2210 // %%% should use stronger assert_unique_concrete_subtype instead 2211 phase->C->dependencies()->assert_leaf_type(ik); 2212 } 2213 // Return precise klass 2214 return TypeKlassPtr::make(ik); 2215 } 2216 2217 // Return root of possible klass 2218 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2219 } 2220 } 2221 2222 // Check for loading klass from an array 2223 const TypeAryPtr *tary = tp->isa_aryptr(); 2224 if( tary != NULL ) { 2225 ciKlass *tary_klass = tary->klass(); 2226 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2227 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2228 if (tary->klass_is_exact()) { 2229 return TypeKlassPtr::make(tary_klass); 2230 } 2231 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2232 // If the klass is an object array, we defer the question to the 2233 // array component klass. 2234 if( ak->is_obj_array_klass() ) { 2235 assert( ak->is_loaded(), "" ); 2236 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2237 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2238 ciInstanceKlass* ik = base_k->as_instance_klass(); 2239 // See if we can become precise: no subklasses and no interface 2240 if (!ik->is_interface() && !ik->has_subklass()) { 2241 //assert(!UseExactTypes, "this code should be useless with exact types"); 2242 // Add a dependence; if any subclass added we need to recompile 2243 if (!ik->is_final()) { 2244 phase->C->dependencies()->assert_leaf_type(ik); 2245 } 2246 // Return precise array klass 2247 return TypeKlassPtr::make(ak); 2248 } 2249 } 2250 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2251 } else { // Found a type-array? 2252 //assert(!UseExactTypes, "this code should be useless with exact types"); 2253 assert( ak->is_type_array_klass(), "" ); 2254 return TypeKlassPtr::make(ak); // These are always precise 2255 } 2256 } 2257 } 2258 2259 // Check for loading klass from an array klass 2260 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2261 if (tkls != NULL && !StressReflectiveCode) { 2262 ciKlass* klass = tkls->klass(); 2263 if( !klass->is_loaded() ) 2264 return _type; // Bail out if not loaded 2265 if( klass->is_obj_array_klass() && 2266 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2267 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2268 // // Always returning precise element type is incorrect, 2269 // // e.g., element type could be object and array may contain strings 2270 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2271 2272 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2273 // according to the element type's subclassing. 2274 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2275 } 2276 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2277 tkls->offset() == in_bytes(Klass::super_offset())) { 2278 ciKlass* sup = klass->as_instance_klass()->super(); 2279 // The field is Klass::_super. Return its (constant) value. 2280 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2281 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2282 } 2283 } 2284 2285 // Bailout case 2286 return LoadNode::Value(phase); 2287 } 2288 2289 //------------------------------Identity--------------------------------------- 2290 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2291 // Also feed through the klass in Allocate(...klass...)._klass. 2292 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2293 return klass_identity_common(phase); 2294 } 2295 2296 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2297 Node* x = LoadNode::Identity(phase); 2298 if (x != this) return x; 2299 2300 // Take apart the address into an oop and and offset. 2301 // Return 'this' if we cannot. 2302 Node* adr = in(MemNode::Address); 2303 intptr_t offset = 0; 2304 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2305 if (base == NULL) return this; 2306 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2307 if (toop == NULL) return this; 2308 2309 // Step over potential GC barrier for OopHandle resolve 2310 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2311 if (bs->is_gc_barrier_node(base)) { 2312 base = bs->step_over_gc_barrier(base); 2313 } 2314 2315 // We can fetch the klass directly through an AllocateNode. 2316 // This works even if the klass is not constant (clone or newArray). 2317 if (offset == oopDesc::klass_offset_in_bytes()) { 2318 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2319 if (allocated_klass != NULL) { 2320 return allocated_klass; 2321 } 2322 } 2323 2324 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2325 // See inline_native_Class_query for occurrences of these patterns. 2326 // Java Example: x.getClass().isAssignableFrom(y) 2327 // 2328 // This improves reflective code, often making the Class 2329 // mirror go completely dead. (Current exception: Class 2330 // mirrors may appear in debug info, but we could clean them out by 2331 // introducing a new debug info operator for Klass.java_mirror). 2332 2333 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2334 && offset == java_lang_Class::klass_offset_in_bytes()) { 2335 if (base->is_Load()) { 2336 Node* base2 = base->in(MemNode::Address); 2337 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2338 Node* adr2 = base2->in(MemNode::Address); 2339 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2340 if (tkls != NULL && !tkls->empty() 2341 && (tkls->klass()->is_instance_klass() || 2342 tkls->klass()->is_array_klass()) 2343 && adr2->is_AddP() 2344 ) { 2345 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2346 if (tkls->offset() == mirror_field) { 2347 return adr2->in(AddPNode::Base); 2348 } 2349 } 2350 } 2351 } 2352 } 2353 2354 return this; 2355 } 2356 2357 2358 //------------------------------Value------------------------------------------ 2359 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2360 const Type *t = klass_value_common(phase); 2361 if (t == Type::TOP) 2362 return t; 2363 2364 return t->make_narrowklass(); 2365 } 2366 2367 //------------------------------Identity--------------------------------------- 2368 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2369 // Also feed through the klass in Allocate(...klass...)._klass. 2370 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2371 Node *x = klass_identity_common(phase); 2372 2373 const Type *t = phase->type( x ); 2374 if( t == Type::TOP ) return x; 2375 if( t->isa_narrowklass()) return x; 2376 assert (!t->isa_narrowoop(), "no narrow oop here"); 2377 2378 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2379 } 2380 2381 //------------------------------Value----------------------------------------- 2382 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2383 // Either input is TOP ==> the result is TOP 2384 const Type *t1 = phase->type( in(MemNode::Memory) ); 2385 if( t1 == Type::TOP ) return Type::TOP; 2386 Node *adr = in(MemNode::Address); 2387 const Type *t2 = phase->type( adr ); 2388 if( t2 == Type::TOP ) return Type::TOP; 2389 const TypePtr *tp = t2->is_ptr(); 2390 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2391 const TypeAryPtr *tap = tp->isa_aryptr(); 2392 if( !tap ) return _type; 2393 return tap->size(); 2394 } 2395 2396 //-------------------------------Ideal--------------------------------------- 2397 // Feed through the length in AllocateArray(...length...)._length. 2398 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2399 Node* p = MemNode::Ideal_common(phase, can_reshape); 2400 if (p) return (p == NodeSentinel) ? NULL : p; 2401 2402 // Take apart the address into an oop and and offset. 2403 // Return 'this' if we cannot. 2404 Node* adr = in(MemNode::Address); 2405 intptr_t offset = 0; 2406 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2407 if (base == NULL) return NULL; 2408 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2409 if (tary == NULL) return NULL; 2410 2411 // We can fetch the length directly through an AllocateArrayNode. 2412 // This works even if the length is not constant (clone or newArray). 2413 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2414 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2415 if (alloc != NULL) { 2416 Node* allocated_length = alloc->Ideal_length(); 2417 Node* len = alloc->make_ideal_length(tary, phase); 2418 if (allocated_length != len) { 2419 // New CastII improves on this. 2420 return len; 2421 } 2422 } 2423 } 2424 2425 return NULL; 2426 } 2427 2428 //------------------------------Identity--------------------------------------- 2429 // Feed through the length in AllocateArray(...length...)._length. 2430 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2431 Node* x = LoadINode::Identity(phase); 2432 if (x != this) return x; 2433 2434 // Take apart the address into an oop and and offset. 2435 // Return 'this' if we cannot. 2436 Node* adr = in(MemNode::Address); 2437 intptr_t offset = 0; 2438 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2439 if (base == NULL) return this; 2440 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2441 if (tary == NULL) return this; 2442 2443 // We can fetch the length directly through an AllocateArrayNode. 2444 // This works even if the length is not constant (clone or newArray). 2445 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2446 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2447 if (alloc != NULL) { 2448 Node* allocated_length = alloc->Ideal_length(); 2449 // Do not allow make_ideal_length to allocate a CastII node. 2450 Node* len = alloc->make_ideal_length(tary, phase, false); 2451 if (allocated_length == len) { 2452 // Return allocated_length only if it would not be improved by a CastII. 2453 return allocated_length; 2454 } 2455 } 2456 } 2457 2458 return this; 2459 2460 } 2461 2462 //============================================================================= 2463 //---------------------------StoreNode::make----------------------------------- 2464 // Polymorphic factory method: 2465 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2466 assert((mo == unordered || mo == release), "unexpected"); 2467 Compile* C = gvn.C; 2468 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2469 ctl != NULL, "raw memory operations should have control edge"); 2470 2471 switch (bt) { 2472 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2473 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2474 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2475 case T_CHAR: 2476 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2477 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2478 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2479 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2480 case T_METADATA: 2481 case T_ADDRESS: 2482 case T_OBJECT: 2483 #ifdef _LP64 2484 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2485 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2486 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2487 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2488 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2489 adr->bottom_type()->isa_rawptr())) { 2490 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2491 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2492 } 2493 #endif 2494 { 2495 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2496 } 2497 default: 2498 ShouldNotReachHere(); 2499 return (StoreNode*)NULL; 2500 } 2501 } 2502 2503 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2504 bool require_atomic = true; 2505 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2506 } 2507 2508 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2509 bool require_atomic = true; 2510 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2511 } 2512 2513 2514 //--------------------------bottom_type---------------------------------------- 2515 const Type *StoreNode::bottom_type() const { 2516 return Type::MEMORY; 2517 } 2518 2519 //------------------------------hash------------------------------------------- 2520 uint StoreNode::hash() const { 2521 // unroll addition of interesting fields 2522 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2523 2524 // Since they are not commoned, do not hash them: 2525 return NO_HASH; 2526 } 2527 2528 //------------------------------Ideal------------------------------------------ 2529 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2530 // When a store immediately follows a relevant allocation/initialization, 2531 // try to capture it into the initialization, or hoist it above. 2532 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2533 Node* p = MemNode::Ideal_common(phase, can_reshape); 2534 if (p) return (p == NodeSentinel) ? NULL : p; 2535 2536 Node* mem = in(MemNode::Memory); 2537 Node* address = in(MemNode::Address); 2538 // Back-to-back stores to same address? Fold em up. Generally 2539 // unsafe if I have intervening uses... Also disallowed for StoreCM 2540 // since they must follow each StoreP operation. Redundant StoreCMs 2541 // are eliminated just before matching in final_graph_reshape. 2542 { 2543 Node* st = mem; 2544 // If Store 'st' has more than one use, we cannot fold 'st' away. 2545 // For example, 'st' might be the final state at a conditional 2546 // return. Or, 'st' might be used by some node which is live at 2547 // the same time 'st' is live, which might be unschedulable. So, 2548 // require exactly ONE user until such time as we clone 'mem' for 2549 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2550 // true). 2551 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2552 // Looking at a dead closed cycle of memory? 2553 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2554 assert(Opcode() == st->Opcode() || 2555 st->Opcode() == Op_StoreVector || 2556 Opcode() == Op_StoreVector || 2557 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2558 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2559 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2560 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2561 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2562 2563 if (st->in(MemNode::Address)->eqv_uncast(address) && 2564 st->as_Store()->memory_size() <= this->memory_size()) { 2565 Node* use = st->raw_out(0); 2566 phase->igvn_rehash_node_delayed(use); 2567 if (can_reshape) { 2568 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2569 } else { 2570 // It's OK to do this in the parser, since DU info is always accurate, 2571 // and the parser always refers to nodes via SafePointNode maps. 2572 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2573 } 2574 return this; 2575 } 2576 st = st->in(MemNode::Memory); 2577 } 2578 } 2579 2580 2581 // Capture an unaliased, unconditional, simple store into an initializer. 2582 // Or, if it is independent of the allocation, hoist it above the allocation. 2583 if (ReduceFieldZeroing && /*can_reshape &&*/ 2584 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2585 InitializeNode* init = mem->in(0)->as_Initialize(); 2586 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2587 if (offset > 0) { 2588 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2589 // If the InitializeNode captured me, it made a raw copy of me, 2590 // and I need to disappear. 2591 if (moved != NULL) { 2592 // %%% hack to ensure that Ideal returns a new node: 2593 mem = MergeMemNode::make(mem); 2594 return mem; // fold me away 2595 } 2596 } 2597 } 2598 2599 return NULL; // No further progress 2600 } 2601 2602 //------------------------------Value----------------------------------------- 2603 const Type* StoreNode::Value(PhaseGVN* phase) const { 2604 // Either input is TOP ==> the result is TOP 2605 const Type *t1 = phase->type( in(MemNode::Memory) ); 2606 if( t1 == Type::TOP ) return Type::TOP; 2607 const Type *t2 = phase->type( in(MemNode::Address) ); 2608 if( t2 == Type::TOP ) return Type::TOP; 2609 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2610 if( t3 == Type::TOP ) return Type::TOP; 2611 return Type::MEMORY; 2612 } 2613 2614 //------------------------------Identity--------------------------------------- 2615 // Remove redundant stores: 2616 // Store(m, p, Load(m, p)) changes to m. 2617 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2618 Node* StoreNode::Identity(PhaseGVN* phase) { 2619 Node* mem = in(MemNode::Memory); 2620 Node* adr = in(MemNode::Address); 2621 Node* val = in(MemNode::ValueIn); 2622 2623 Node* result = this; 2624 2625 // Load then Store? Then the Store is useless 2626 if (val->is_Load() && 2627 val->in(MemNode::Address)->eqv_uncast(adr) && 2628 val->in(MemNode::Memory )->eqv_uncast(mem) && 2629 val->as_Load()->store_Opcode() == Opcode()) { 2630 result = mem; 2631 } 2632 2633 // Two stores in a row of the same value? 2634 if (result == this && 2635 mem->is_Store() && 2636 mem->in(MemNode::Address)->eqv_uncast(adr) && 2637 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2638 mem->Opcode() == Opcode()) { 2639 result = mem; 2640 } 2641 2642 // Store of zero anywhere into a freshly-allocated object? 2643 // Then the store is useless. 2644 // (It must already have been captured by the InitializeNode.) 2645 if (result == this && 2646 ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2647 // a newly allocated object is already all-zeroes everywhere 2648 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2649 result = mem; 2650 } 2651 2652 if (result == this) { 2653 // the store may also apply to zero-bits in an earlier object 2654 Node* prev_mem = find_previous_store(phase); 2655 // Steps (a), (b): Walk past independent stores to find an exact match. 2656 if (prev_mem != NULL) { 2657 Node* prev_val = can_see_stored_value(prev_mem, phase); 2658 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2659 // prev_val and val might differ by a cast; it would be good 2660 // to keep the more informative of the two. 2661 result = mem; 2662 } 2663 } 2664 } 2665 } 2666 2667 if (result != this && phase->is_IterGVN() != NULL) { 2668 MemBarNode* trailing = trailing_membar(); 2669 if (trailing != NULL) { 2670 #ifdef ASSERT 2671 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2672 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects"); 2673 #endif 2674 PhaseIterGVN* igvn = phase->is_IterGVN(); 2675 trailing->remove(igvn); 2676 } 2677 } 2678 2679 return result; 2680 } 2681 2682 //------------------------------match_edge------------------------------------- 2683 // Do we Match on this edge index or not? Match only memory & value 2684 uint StoreNode::match_edge(uint idx) const { 2685 return idx == MemNode::Address || idx == MemNode::ValueIn; 2686 } 2687 2688 //------------------------------cmp-------------------------------------------- 2689 // Do not common stores up together. They generally have to be split 2690 // back up anyways, so do not bother. 2691 uint StoreNode::cmp( const Node &n ) const { 2692 return (&n == this); // Always fail except on self 2693 } 2694 2695 //------------------------------Ideal_masked_input----------------------------- 2696 // Check for a useless mask before a partial-word store 2697 // (StoreB ... (AndI valIn conIa) ) 2698 // If (conIa & mask == mask) this simplifies to 2699 // (StoreB ... (valIn) ) 2700 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2701 Node *val = in(MemNode::ValueIn); 2702 if( val->Opcode() == Op_AndI ) { 2703 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2704 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2705 set_req(MemNode::ValueIn, val->in(1)); 2706 return this; 2707 } 2708 } 2709 return NULL; 2710 } 2711 2712 2713 //------------------------------Ideal_sign_extended_input---------------------- 2714 // Check for useless sign-extension before a partial-word store 2715 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2716 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2717 // (StoreB ... (valIn) ) 2718 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2719 Node *val = in(MemNode::ValueIn); 2720 if( val->Opcode() == Op_RShiftI ) { 2721 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2722 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2723 Node *shl = val->in(1); 2724 if( shl->Opcode() == Op_LShiftI ) { 2725 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2726 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2727 set_req(MemNode::ValueIn, shl->in(1)); 2728 return this; 2729 } 2730 } 2731 } 2732 } 2733 return NULL; 2734 } 2735 2736 //------------------------------value_never_loaded----------------------------------- 2737 // Determine whether there are any possible loads of the value stored. 2738 // For simplicity, we actually check if there are any loads from the 2739 // address stored to, not just for loads of the value stored by this node. 2740 // 2741 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2742 Node *adr = in(Address); 2743 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2744 if (adr_oop == NULL) 2745 return false; 2746 if (!adr_oop->is_known_instance_field()) 2747 return false; // if not a distinct instance, there may be aliases of the address 2748 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2749 Node *use = adr->fast_out(i); 2750 if (use->is_Load() || use->is_LoadStore()) { 2751 return false; 2752 } 2753 } 2754 return true; 2755 } 2756 2757 MemBarNode* StoreNode::trailing_membar() const { 2758 if (is_release()) { 2759 MemBarNode* trailing_mb = NULL; 2760 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2761 Node* u = fast_out(i); 2762 if (u->is_MemBar()) { 2763 if (u->as_MemBar()->trailing_store()) { 2764 assert(u->Opcode() == Op_MemBarVolatile, ""); 2765 assert(trailing_mb == NULL, "only one"); 2766 trailing_mb = u->as_MemBar(); 2767 #ifdef ASSERT 2768 Node* leading = u->as_MemBar()->leading_membar(); 2769 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2770 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2771 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2772 #endif 2773 } else { 2774 assert(u->as_MemBar()->standalone(), ""); 2775 } 2776 } 2777 } 2778 return trailing_mb; 2779 } 2780 return NULL; 2781 } 2782 2783 2784 //============================================================================= 2785 //------------------------------Ideal------------------------------------------ 2786 // If the store is from an AND mask that leaves the low bits untouched, then 2787 // we can skip the AND operation. If the store is from a sign-extension 2788 // (a left shift, then right shift) we can skip both. 2789 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2790 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2791 if( progress != NULL ) return progress; 2792 2793 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2794 if( progress != NULL ) return progress; 2795 2796 // Finally check the default case 2797 return StoreNode::Ideal(phase, can_reshape); 2798 } 2799 2800 //============================================================================= 2801 //------------------------------Ideal------------------------------------------ 2802 // If the store is from an AND mask that leaves the low bits untouched, then 2803 // we can skip the AND operation 2804 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2805 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2806 if( progress != NULL ) return progress; 2807 2808 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2809 if( progress != NULL ) return progress; 2810 2811 // Finally check the default case 2812 return StoreNode::Ideal(phase, can_reshape); 2813 } 2814 2815 //============================================================================= 2816 //------------------------------Identity--------------------------------------- 2817 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2818 // No need to card mark when storing a null ptr 2819 Node* my_store = in(MemNode::OopStore); 2820 if (my_store->is_Store()) { 2821 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2822 if( t1 == TypePtr::NULL_PTR ) { 2823 return in(MemNode::Memory); 2824 } 2825 } 2826 return this; 2827 } 2828 2829 //============================================================================= 2830 //------------------------------Ideal--------------------------------------- 2831 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2832 Node* progress = StoreNode::Ideal(phase, can_reshape); 2833 if (progress != NULL) return progress; 2834 2835 Node* my_store = in(MemNode::OopStore); 2836 if (my_store->is_MergeMem()) { 2837 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2838 set_req(MemNode::OopStore, mem); 2839 return this; 2840 } 2841 2842 return NULL; 2843 } 2844 2845 //------------------------------Value----------------------------------------- 2846 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2847 // Either input is TOP ==> the result is TOP (checked in StoreNode::Value). 2848 // If extra input is TOP ==> the result is TOP 2849 const Type* t = phase->type(in(MemNode::OopStore)); 2850 if (t == Type::TOP) { 2851 return Type::TOP; 2852 } 2853 return StoreNode::Value(phase); 2854 } 2855 2856 2857 //============================================================================= 2858 //----------------------------------SCMemProjNode------------------------------ 2859 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2860 { 2861 if (in(0) == NULL || phase->type(in(0)) == Type::TOP) { 2862 return Type::TOP; 2863 } 2864 return bottom_type(); 2865 } 2866 2867 //============================================================================= 2868 //----------------------------------LoadStoreNode------------------------------ 2869 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2870 : Node(required), 2871 _type(rt), 2872 _adr_type(at) 2873 { 2874 init_req(MemNode::Control, c ); 2875 init_req(MemNode::Memory , mem); 2876 init_req(MemNode::Address, adr); 2877 init_req(MemNode::ValueIn, val); 2878 init_class_id(Class_LoadStore); 2879 } 2880 2881 //------------------------------Value----------------------------------------- 2882 const Type* LoadStoreNode::Value(PhaseGVN* phase) const { 2883 // Either input is TOP ==> the result is TOP 2884 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) { 2885 return Type::TOP; 2886 } 2887 const Type* t = phase->type(in(MemNode::Memory)); 2888 if (t == Type::TOP) { 2889 return Type::TOP; 2890 } 2891 t = phase->type(in(MemNode::Address)); 2892 if (t == Type::TOP) { 2893 return Type::TOP; 2894 } 2895 t = phase->type(in(MemNode::ValueIn)); 2896 if (t == Type::TOP) { 2897 return Type::TOP; 2898 } 2899 return bottom_type(); 2900 } 2901 2902 uint LoadStoreNode::ideal_reg() const { 2903 return _type->ideal_reg(); 2904 } 2905 2906 bool LoadStoreNode::result_not_used() const { 2907 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2908 Node *x = fast_out(i); 2909 if (x->Opcode() == Op_SCMemProj) continue; 2910 return false; 2911 } 2912 return true; 2913 } 2914 2915 MemBarNode* LoadStoreNode::trailing_membar() const { 2916 MemBarNode* trailing = NULL; 2917 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2918 Node* u = fast_out(i); 2919 if (u->is_MemBar()) { 2920 if (u->as_MemBar()->trailing_load_store()) { 2921 assert(u->Opcode() == Op_MemBarAcquire, ""); 2922 assert(trailing == NULL, "only one"); 2923 trailing = u->as_MemBar(); 2924 #ifdef ASSERT 2925 Node* leading = trailing->leading_membar(); 2926 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2927 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 2928 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 2929 #endif 2930 } else { 2931 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 2932 } 2933 } 2934 } 2935 2936 return trailing; 2937 } 2938 2939 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2940 2941 //============================================================================= 2942 //----------------------------------LoadStoreConditionalNode-------------------- 2943 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2944 init_req(ExpectedIn, ex ); 2945 } 2946 2947 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const { 2948 // Either input is TOP ==> the result is TOP 2949 const Type* t = phase->type(in(ExpectedIn)); 2950 if (t == Type::TOP) { 2951 return Type::TOP; 2952 } 2953 return LoadStoreNode::Value(phase); 2954 } 2955 2956 //============================================================================= 2957 //-------------------------------adr_type-------------------------------------- 2958 const TypePtr* ClearArrayNode::adr_type() const { 2959 Node *adr = in(3); 2960 if (adr == NULL) return NULL; // node is dead 2961 return MemNode::calculate_adr_type(adr->bottom_type()); 2962 } 2963 2964 //------------------------------match_edge------------------------------------- 2965 // Do we Match on this edge index or not? Do not match memory 2966 uint ClearArrayNode::match_edge(uint idx) const { 2967 return idx > 1; 2968 } 2969 2970 //------------------------------Identity--------------------------------------- 2971 // Clearing a zero length array does nothing 2972 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2973 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2974 } 2975 2976 //------------------------------Idealize--------------------------------------- 2977 // Clearing a short array is faster with stores 2978 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2979 // Already know this is a large node, do not try to ideal it 2980 if (!IdealizeClearArrayNode || _is_large) return NULL; 2981 2982 const int unit = BytesPerLong; 2983 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2984 if (!t) return NULL; 2985 if (!t->is_con()) return NULL; 2986 intptr_t raw_count = t->get_con(); 2987 intptr_t size = raw_count; 2988 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2989 // Clearing nothing uses the Identity call. 2990 // Negative clears are possible on dead ClearArrays 2991 // (see jck test stmt114.stmt11402.val). 2992 if (size <= 0 || size % unit != 0) return NULL; 2993 intptr_t count = size / unit; 2994 // Length too long; communicate this to matchers and assemblers. 2995 // Assemblers are responsible to produce fast hardware clears for it. 2996 if (size > InitArrayShortSize) { 2997 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 2998 } 2999 Node *mem = in(1); 3000 if( phase->type(mem)==Type::TOP ) return NULL; 3001 Node *adr = in(3); 3002 const Type* at = phase->type(adr); 3003 if( at==Type::TOP ) return NULL; 3004 const TypePtr* atp = at->isa_ptr(); 3005 // adjust atp to be the correct array element address type 3006 if (atp == NULL) atp = TypePtr::BOTTOM; 3007 else atp = atp->add_offset(Type::OffsetBot); 3008 // Get base for derived pointer purposes 3009 if( adr->Opcode() != Op_AddP ) Unimplemented(); 3010 Node *base = adr->in(1); 3011 3012 Node *zero = phase->makecon(TypeLong::ZERO); 3013 Node *off = phase->MakeConX(BytesPerLong); 3014 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 3015 count--; 3016 while( count-- ) { 3017 mem = phase->transform(mem); 3018 adr = phase->transform(new AddPNode(base,adr,off)); 3019 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 3020 } 3021 return mem; 3022 } 3023 3024 //----------------------------step_through---------------------------------- 3025 // Return allocation input memory edge if it is different instance 3026 // or itself if it is the one we are looking for. 3027 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 3028 Node* n = *np; 3029 assert(n->is_ClearArray(), "sanity"); 3030 intptr_t offset; 3031 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 3032 // This method is called only before Allocate nodes are expanded 3033 // during macro nodes expansion. Before that ClearArray nodes are 3034 // only generated in PhaseMacroExpand::generate_arraycopy() (before 3035 // Allocate nodes are expanded) which follows allocations. 3036 assert(alloc != NULL, "should have allocation"); 3037 if (alloc->_idx == instance_id) { 3038 // Can not bypass initialization of the instance we are looking for. 3039 return false; 3040 } 3041 // Otherwise skip it. 3042 InitializeNode* init = alloc->initialization(); 3043 if (init != NULL) 3044 *np = init->in(TypeFunc::Memory); 3045 else 3046 *np = alloc->in(TypeFunc::Memory); 3047 return true; 3048 } 3049 3050 //----------------------------clear_memory------------------------------------- 3051 // Generate code to initialize object storage to zero. 3052 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3053 intptr_t start_offset, 3054 Node* end_offset, 3055 PhaseGVN* phase) { 3056 intptr_t offset = start_offset; 3057 3058 int unit = BytesPerLong; 3059 if ((offset % unit) != 0) { 3060 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 3061 adr = phase->transform(adr); 3062 const TypePtr* atp = TypeRawPtr::BOTTOM; 3063 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3064 mem = phase->transform(mem); 3065 offset += BytesPerInt; 3066 } 3067 assert((offset % unit) == 0, ""); 3068 3069 // Initialize the remaining stuff, if any, with a ClearArray. 3070 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 3071 } 3072 3073 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3074 Node* start_offset, 3075 Node* end_offset, 3076 PhaseGVN* phase) { 3077 if (start_offset == end_offset) { 3078 // nothing to do 3079 return mem; 3080 } 3081 3082 int unit = BytesPerLong; 3083 Node* zbase = start_offset; 3084 Node* zend = end_offset; 3085 3086 // Scale to the unit required by the CPU: 3087 if (!Matcher::init_array_count_is_in_bytes) { 3088 Node* shift = phase->intcon(exact_log2(unit)); 3089 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3090 zend = phase->transform(new URShiftXNode(zend, shift) ); 3091 } 3092 3093 // Bulk clear double-words 3094 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3095 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3096 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 3097 return phase->transform(mem); 3098 } 3099 3100 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3101 intptr_t start_offset, 3102 intptr_t end_offset, 3103 PhaseGVN* phase) { 3104 if (start_offset == end_offset) { 3105 // nothing to do 3106 return mem; 3107 } 3108 3109 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3110 intptr_t done_offset = end_offset; 3111 if ((done_offset % BytesPerLong) != 0) { 3112 done_offset -= BytesPerInt; 3113 } 3114 if (done_offset > start_offset) { 3115 mem = clear_memory(ctl, mem, dest, 3116 start_offset, phase->MakeConX(done_offset), phase); 3117 } 3118 if (done_offset < end_offset) { // emit the final 32-bit store 3119 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3120 adr = phase->transform(adr); 3121 const TypePtr* atp = TypeRawPtr::BOTTOM; 3122 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3123 mem = phase->transform(mem); 3124 done_offset += BytesPerInt; 3125 } 3126 assert(done_offset == end_offset, ""); 3127 return mem; 3128 } 3129 3130 //============================================================================= 3131 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3132 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 3133 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3134 #ifdef ASSERT 3135 , _pair_idx(0) 3136 #endif 3137 { 3138 init_class_id(Class_MemBar); 3139 Node* top = C->top(); 3140 init_req(TypeFunc::I_O,top); 3141 init_req(TypeFunc::FramePtr,top); 3142 init_req(TypeFunc::ReturnAdr,top); 3143 if (precedent != NULL) 3144 init_req(TypeFunc::Parms, precedent); 3145 } 3146 3147 //------------------------------cmp-------------------------------------------- 3148 uint MemBarNode::hash() const { return NO_HASH; } 3149 uint MemBarNode::cmp( const Node &n ) const { 3150 return (&n == this); // Always fail except on self 3151 } 3152 3153 //------------------------------make------------------------------------------- 3154 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3155 switch (opcode) { 3156 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3157 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3158 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3159 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3160 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3161 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3162 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3163 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3164 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3165 case Op_Initialize: return new InitializeNode(C, atp, pn); 3166 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3167 default: ShouldNotReachHere(); return NULL; 3168 } 3169 } 3170 3171 void MemBarNode::remove(PhaseIterGVN *igvn) { 3172 if (outcnt() != 2) { 3173 return; 3174 } 3175 if (trailing_store() || trailing_load_store()) { 3176 MemBarNode* leading = leading_membar(); 3177 if (leading != NULL) { 3178 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3179 leading->remove(igvn); 3180 } 3181 } 3182 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3183 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3184 } 3185 3186 //------------------------------Ideal------------------------------------------ 3187 // Return a node which is more "ideal" than the current node. Strip out 3188 // control copies 3189 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3190 if (remove_dead_region(phase, can_reshape)) return this; 3191 // Don't bother trying to transform a dead node 3192 if (in(0) && in(0)->is_top()) { 3193 return NULL; 3194 } 3195 3196 #if INCLUDE_ZGC 3197 if (UseZGC) { 3198 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) { 3199 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop); 3200 set_req(MemBarNode::Precedent, load_node); 3201 return this; 3202 } 3203 } 3204 #endif 3205 3206 bool progress = false; 3207 // Eliminate volatile MemBars for scalar replaced objects. 3208 if (can_reshape && req() == (Precedent+1)) { 3209 bool eliminate = false; 3210 int opc = Opcode(); 3211 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3212 // Volatile field loads and stores. 3213 Node* my_mem = in(MemBarNode::Precedent); 3214 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3215 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3216 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3217 // replace this Precedent (decodeN) with the Load instead. 3218 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3219 Node* load_node = my_mem->in(1); 3220 set_req(MemBarNode::Precedent, load_node); 3221 phase->is_IterGVN()->_worklist.push(my_mem); 3222 my_mem = load_node; 3223 } else { 3224 assert(my_mem->unique_out() == this, "sanity"); 3225 del_req(Precedent); 3226 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3227 my_mem = NULL; 3228 } 3229 progress = true; 3230 } 3231 if (my_mem != NULL && my_mem->is_Mem()) { 3232 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3233 // Check for scalar replaced object reference. 3234 if( t_oop != NULL && t_oop->is_known_instance_field() && 3235 t_oop->offset() != Type::OffsetBot && 3236 t_oop->offset() != Type::OffsetTop) { 3237 eliminate = true; 3238 } 3239 } 3240 } else if (opc == Op_MemBarRelease) { 3241 // Final field stores. 3242 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3243 if ((alloc != NULL) && alloc->is_Allocate() && 3244 alloc->as_Allocate()->does_not_escape_thread()) { 3245 // The allocated object does not escape. 3246 eliminate = true; 3247 } 3248 } 3249 if (eliminate) { 3250 // Replace MemBar projections by its inputs. 3251 PhaseIterGVN* igvn = phase->is_IterGVN(); 3252 remove(igvn); 3253 // Must return either the original node (now dead) or a new node 3254 // (Do not return a top here, since that would break the uniqueness of top.) 3255 return new ConINode(TypeInt::ZERO); 3256 } 3257 } 3258 return progress ? this : NULL; 3259 } 3260 3261 //------------------------------Value------------------------------------------ 3262 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3263 if( !in(0) ) return Type::TOP; 3264 if( phase->type(in(0)) == Type::TOP ) 3265 return Type::TOP; 3266 return TypeTuple::MEMBAR; 3267 } 3268 3269 //------------------------------match------------------------------------------ 3270 // Construct projections for memory. 3271 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3272 switch (proj->_con) { 3273 case TypeFunc::Control: 3274 case TypeFunc::Memory: 3275 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3276 } 3277 ShouldNotReachHere(); 3278 return NULL; 3279 } 3280 3281 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3282 trailing->_kind = TrailingStore; 3283 leading->_kind = LeadingStore; 3284 #ifdef ASSERT 3285 trailing->_pair_idx = leading->_idx; 3286 leading->_pair_idx = leading->_idx; 3287 #endif 3288 } 3289 3290 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3291 trailing->_kind = TrailingLoadStore; 3292 leading->_kind = LeadingLoadStore; 3293 #ifdef ASSERT 3294 trailing->_pair_idx = leading->_idx; 3295 leading->_pair_idx = leading->_idx; 3296 #endif 3297 } 3298 3299 MemBarNode* MemBarNode::trailing_membar() const { 3300 ResourceMark rm; 3301 Node* trailing = (Node*)this; 3302 VectorSet seen(Thread::current()->resource_area()); 3303 Node_Stack multis(0); 3304 do { 3305 Node* c = trailing; 3306 uint i = 0; 3307 do { 3308 trailing = NULL; 3309 for (; i < c->outcnt(); i++) { 3310 Node* next = c->raw_out(i); 3311 if (next != c && next->is_CFG()) { 3312 if (c->is_MultiBranch()) { 3313 if (multis.node() == c) { 3314 multis.set_index(i+1); 3315 } else { 3316 multis.push(c, i+1); 3317 } 3318 } 3319 trailing = next; 3320 break; 3321 } 3322 } 3323 if (trailing != NULL && !seen.test_set(trailing->_idx)) { 3324 break; 3325 } 3326 while (multis.size() > 0) { 3327 c = multis.node(); 3328 i = multis.index(); 3329 if (i < c->req()) { 3330 break; 3331 } 3332 multis.pop(); 3333 } 3334 } while (multis.size() > 0); 3335 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3336 3337 MemBarNode* mb = trailing->as_MemBar(); 3338 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3339 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3340 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3341 return mb; 3342 } 3343 3344 MemBarNode* MemBarNode::leading_membar() const { 3345 ResourceMark rm; 3346 VectorSet seen(Thread::current()->resource_area()); 3347 Node_Stack regions(0); 3348 Node* leading = in(0); 3349 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3350 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) { 3351 leading = NULL; 3352 while (regions.size() > 0 && leading == NULL) { 3353 Node* r = regions.node(); 3354 uint i = regions.index(); 3355 if (i < r->req()) { 3356 leading = r->in(i); 3357 regions.set_index(i+1); 3358 } else { 3359 regions.pop(); 3360 } 3361 } 3362 if (leading == NULL) { 3363 assert(regions.size() == 0, "all paths should have been tried"); 3364 return NULL; 3365 } 3366 } 3367 if (leading->is_Region()) { 3368 regions.push(leading, 2); 3369 leading = leading->in(1); 3370 } else { 3371 leading = leading->in(0); 3372 } 3373 } 3374 #ifdef ASSERT 3375 Unique_Node_List wq; 3376 wq.push((Node*)this); 3377 uint found = 0; 3378 for (uint i = 0; i < wq.size(); i++) { 3379 Node* n = wq.at(i); 3380 if (n->is_Region()) { 3381 for (uint j = 1; j < n->req(); j++) { 3382 Node* in = n->in(j); 3383 if (in != NULL && !in->is_top()) { 3384 wq.push(in); 3385 } 3386 } 3387 } else { 3388 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3389 assert(n == leading, "consistency check failed"); 3390 found++; 3391 } else { 3392 Node* in = n->in(0); 3393 if (in != NULL && !in->is_top()) { 3394 wq.push(in); 3395 } 3396 } 3397 } 3398 } 3399 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed"); 3400 #endif 3401 if (leading == NULL) { 3402 return NULL; 3403 } 3404 MemBarNode* mb = leading->as_MemBar(); 3405 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3406 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3407 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3408 return mb; 3409 } 3410 3411 //===========================InitializeNode==================================== 3412 // SUMMARY: 3413 // This node acts as a memory barrier on raw memory, after some raw stores. 3414 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3415 // The Initialize can 'capture' suitably constrained stores as raw inits. 3416 // It can coalesce related raw stores into larger units (called 'tiles'). 3417 // It can avoid zeroing new storage for memory units which have raw inits. 3418 // At macro-expansion, it is marked 'complete', and does not optimize further. 3419 // 3420 // EXAMPLE: 3421 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3422 // ctl = incoming control; mem* = incoming memory 3423 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3424 // First allocate uninitialized memory and fill in the header: 3425 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3426 // ctl := alloc.Control; mem* := alloc.Memory* 3427 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3428 // Then initialize to zero the non-header parts of the raw memory block: 3429 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3430 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3431 // After the initialize node executes, the object is ready for service: 3432 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3433 // Suppose its body is immediately initialized as {1,2}: 3434 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3435 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3436 // mem.SLICE(#short[*]) := store2 3437 // 3438 // DETAILS: 3439 // An InitializeNode collects and isolates object initialization after 3440 // an AllocateNode and before the next possible safepoint. As a 3441 // memory barrier (MemBarNode), it keeps critical stores from drifting 3442 // down past any safepoint or any publication of the allocation. 3443 // Before this barrier, a newly-allocated object may have uninitialized bits. 3444 // After this barrier, it may be treated as a real oop, and GC is allowed. 3445 // 3446 // The semantics of the InitializeNode include an implicit zeroing of 3447 // the new object from object header to the end of the object. 3448 // (The object header and end are determined by the AllocateNode.) 3449 // 3450 // Certain stores may be added as direct inputs to the InitializeNode. 3451 // These stores must update raw memory, and they must be to addresses 3452 // derived from the raw address produced by AllocateNode, and with 3453 // a constant offset. They must be ordered by increasing offset. 3454 // The first one is at in(RawStores), the last at in(req()-1). 3455 // Unlike most memory operations, they are not linked in a chain, 3456 // but are displayed in parallel as users of the rawmem output of 3457 // the allocation. 3458 // 3459 // (See comments in InitializeNode::capture_store, which continue 3460 // the example given above.) 3461 // 3462 // When the associated Allocate is macro-expanded, the InitializeNode 3463 // may be rewritten to optimize collected stores. A ClearArrayNode 3464 // may also be created at that point to represent any required zeroing. 3465 // The InitializeNode is then marked 'complete', prohibiting further 3466 // capturing of nearby memory operations. 3467 // 3468 // During macro-expansion, all captured initializations which store 3469 // constant values of 32 bits or smaller are coalesced (if advantageous) 3470 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3471 // initialized in fewer memory operations. Memory words which are 3472 // covered by neither tiles nor non-constant stores are pre-zeroed 3473 // by explicit stores of zero. (The code shape happens to do all 3474 // zeroing first, then all other stores, with both sequences occurring 3475 // in order of ascending offsets.) 3476 // 3477 // Alternatively, code may be inserted between an AllocateNode and its 3478 // InitializeNode, to perform arbitrary initialization of the new object. 3479 // E.g., the object copying intrinsics insert complex data transfers here. 3480 // The initialization must then be marked as 'complete' disable the 3481 // built-in zeroing semantics and the collection of initializing stores. 3482 // 3483 // While an InitializeNode is incomplete, reads from the memory state 3484 // produced by it are optimizable if they match the control edge and 3485 // new oop address associated with the allocation/initialization. 3486 // They return a stored value (if the offset matches) or else zero. 3487 // A write to the memory state, if it matches control and address, 3488 // and if it is to a constant offset, may be 'captured' by the 3489 // InitializeNode. It is cloned as a raw memory operation and rewired 3490 // inside the initialization, to the raw oop produced by the allocation. 3491 // Operations on addresses which are provably distinct (e.g., to 3492 // other AllocateNodes) are allowed to bypass the initialization. 3493 // 3494 // The effect of all this is to consolidate object initialization 3495 // (both arrays and non-arrays, both piecewise and bulk) into a 3496 // single location, where it can be optimized as a unit. 3497 // 3498 // Only stores with an offset less than TrackedInitializationLimit words 3499 // will be considered for capture by an InitializeNode. This puts a 3500 // reasonable limit on the complexity of optimized initializations. 3501 3502 //---------------------------InitializeNode------------------------------------ 3503 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3504 : _is_complete(Incomplete), _does_not_escape(false), 3505 MemBarNode(C, adr_type, rawoop) 3506 { 3507 init_class_id(Class_Initialize); 3508 3509 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3510 assert(in(RawAddress) == rawoop, "proper init"); 3511 // Note: allocation() can be NULL, for secondary initialization barriers 3512 } 3513 3514 // Since this node is not matched, it will be processed by the 3515 // register allocator. Declare that there are no constraints 3516 // on the allocation of the RawAddress edge. 3517 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3518 // This edge should be set to top, by the set_complete. But be conservative. 3519 if (idx == InitializeNode::RawAddress) 3520 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3521 return RegMask::Empty; 3522 } 3523 3524 Node* InitializeNode::memory(uint alias_idx) { 3525 Node* mem = in(Memory); 3526 if (mem->is_MergeMem()) { 3527 return mem->as_MergeMem()->memory_at(alias_idx); 3528 } else { 3529 // incoming raw memory is not split 3530 return mem; 3531 } 3532 } 3533 3534 bool InitializeNode::is_non_zero() { 3535 if (is_complete()) return false; 3536 remove_extra_zeroes(); 3537 return (req() > RawStores); 3538 } 3539 3540 void InitializeNode::set_complete(PhaseGVN* phase) { 3541 assert(!is_complete(), "caller responsibility"); 3542 _is_complete = Complete; 3543 3544 // After this node is complete, it contains a bunch of 3545 // raw-memory initializations. There is no need for 3546 // it to have anything to do with non-raw memory effects. 3547 // Therefore, tell all non-raw users to re-optimize themselves, 3548 // after skipping the memory effects of this initialization. 3549 PhaseIterGVN* igvn = phase->is_IterGVN(); 3550 if (igvn) igvn->add_users_to_worklist(this); 3551 } 3552 3553 // convenience function 3554 // return false if the init contains any stores already 3555 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3556 InitializeNode* init = initialization(); 3557 if (init == NULL || init->is_complete()) return false; 3558 init->remove_extra_zeroes(); 3559 // for now, if this allocation has already collected any inits, bail: 3560 if (init->is_non_zero()) return false; 3561 init->set_complete(phase); 3562 return true; 3563 } 3564 3565 void InitializeNode::remove_extra_zeroes() { 3566 if (req() == RawStores) return; 3567 Node* zmem = zero_memory(); 3568 uint fill = RawStores; 3569 for (uint i = fill; i < req(); i++) { 3570 Node* n = in(i); 3571 if (n->is_top() || n == zmem) continue; // skip 3572 if (fill < i) set_req(fill, n); // compact 3573 ++fill; 3574 } 3575 // delete any empty spaces created: 3576 while (fill < req()) { 3577 del_req(fill); 3578 } 3579 } 3580 3581 // Helper for remembering which stores go with which offsets. 3582 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3583 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3584 intptr_t offset = -1; 3585 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3586 phase, offset); 3587 if (base == NULL) return -1; // something is dead, 3588 if (offset < 0) return -1; // dead, dead 3589 return offset; 3590 } 3591 3592 // Helper for proving that an initialization expression is 3593 // "simple enough" to be folded into an object initialization. 3594 // Attempts to prove that a store's initial value 'n' can be captured 3595 // within the initialization without creating a vicious cycle, such as: 3596 // { Foo p = new Foo(); p.next = p; } 3597 // True for constants and parameters and small combinations thereof. 3598 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3599 if (n == NULL) return true; // (can this really happen?) 3600 if (n->is_Proj()) n = n->in(0); 3601 if (n == this) return false; // found a cycle 3602 if (n->is_Con()) return true; 3603 if (n->is_Start()) return true; // params, etc., are OK 3604 if (n->is_Root()) return true; // even better 3605 3606 Node* ctl = n->in(0); 3607 if (ctl != NULL && !ctl->is_top()) { 3608 if (ctl->is_Proj()) ctl = ctl->in(0); 3609 if (ctl == this) return false; 3610 3611 // If we already know that the enclosing memory op is pinned right after 3612 // the init, then any control flow that the store has picked up 3613 // must have preceded the init, or else be equal to the init. 3614 // Even after loop optimizations (which might change control edges) 3615 // a store is never pinned *before* the availability of its inputs. 3616 if (!MemNode::all_controls_dominate(n, this)) 3617 return false; // failed to prove a good control 3618 } 3619 3620 // Check data edges for possible dependencies on 'this'. 3621 if ((count += 1) > 20) return false; // complexity limit 3622 for (uint i = 1; i < n->req(); i++) { 3623 Node* m = n->in(i); 3624 if (m == NULL || m == n || m->is_top()) continue; 3625 uint first_i = n->find_edge(m); 3626 if (i != first_i) continue; // process duplicate edge just once 3627 if (!detect_init_independence(m, count)) { 3628 return false; 3629 } 3630 } 3631 3632 return true; 3633 } 3634 3635 // Here are all the checks a Store must pass before it can be moved into 3636 // an initialization. Returns zero if a check fails. 3637 // On success, returns the (constant) offset to which the store applies, 3638 // within the initialized memory. 3639 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3640 const int FAIL = 0; 3641 if (st->req() != MemNode::ValueIn + 1) 3642 return FAIL; // an inscrutable StoreNode (card mark?) 3643 Node* ctl = st->in(MemNode::Control); 3644 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3645 return FAIL; // must be unconditional after the initialization 3646 Node* mem = st->in(MemNode::Memory); 3647 if (!(mem->is_Proj() && mem->in(0) == this)) 3648 return FAIL; // must not be preceded by other stores 3649 Node* adr = st->in(MemNode::Address); 3650 intptr_t offset; 3651 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3652 if (alloc == NULL) 3653 return FAIL; // inscrutable address 3654 if (alloc != allocation()) 3655 return FAIL; // wrong allocation! (store needs to float up) 3656 int size_in_bytes = st->memory_size(); 3657 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) { 3658 return FAIL; // mismatched access 3659 } 3660 Node* val = st->in(MemNode::ValueIn); 3661 int complexity_count = 0; 3662 if (!detect_init_independence(val, complexity_count)) 3663 return FAIL; // stored value must be 'simple enough' 3664 3665 // The Store can be captured only if nothing after the allocation 3666 // and before the Store is using the memory location that the store 3667 // overwrites. 3668 bool failed = false; 3669 // If is_complete_with_arraycopy() is true the shape of the graph is 3670 // well defined and is safe so no need for extra checks. 3671 if (!is_complete_with_arraycopy()) { 3672 // We are going to look at each use of the memory state following 3673 // the allocation to make sure nothing reads the memory that the 3674 // Store writes. 3675 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3676 int alias_idx = phase->C->get_alias_index(t_adr); 3677 ResourceMark rm; 3678 Unique_Node_List mems; 3679 mems.push(mem); 3680 Node* unique_merge = NULL; 3681 for (uint next = 0; next < mems.size(); ++next) { 3682 Node *m = mems.at(next); 3683 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3684 Node *n = m->fast_out(j); 3685 if (n->outcnt() == 0) { 3686 continue; 3687 } 3688 if (n == st) { 3689 continue; 3690 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3691 // If the control of this use is different from the control 3692 // of the Store which is right after the InitializeNode then 3693 // this node cannot be between the InitializeNode and the 3694 // Store. 3695 continue; 3696 } else if (n->is_MergeMem()) { 3697 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3698 // We can hit a MergeMemNode (that will likely go away 3699 // later) that is a direct use of the memory state 3700 // following the InitializeNode on the same slice as the 3701 // store node that we'd like to capture. We need to check 3702 // the uses of the MergeMemNode. 3703 mems.push(n); 3704 } 3705 } else if (n->is_Mem()) { 3706 Node* other_adr = n->in(MemNode::Address); 3707 if (other_adr == adr) { 3708 failed = true; 3709 break; 3710 } else { 3711 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3712 if (other_t_adr != NULL) { 3713 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3714 if (other_alias_idx == alias_idx) { 3715 // A load from the same memory slice as the store right 3716 // after the InitializeNode. We check the control of the 3717 // object/array that is loaded from. If it's the same as 3718 // the store control then we cannot capture the store. 3719 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3720 Node* base = other_adr; 3721 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3722 base = base->in(AddPNode::Base); 3723 if (base != NULL) { 3724 base = base->uncast(); 3725 if (base->is_Proj() && base->in(0) == alloc) { 3726 failed = true; 3727 break; 3728 } 3729 } 3730 } 3731 } 3732 } 3733 } else { 3734 failed = true; 3735 break; 3736 } 3737 } 3738 } 3739 } 3740 if (failed) { 3741 if (!can_reshape) { 3742 // We decided we couldn't capture the store during parsing. We 3743 // should try again during the next IGVN once the graph is 3744 // cleaner. 3745 phase->C->record_for_igvn(st); 3746 } 3747 return FAIL; 3748 } 3749 3750 return offset; // success 3751 } 3752 3753 // Find the captured store in(i) which corresponds to the range 3754 // [start..start+size) in the initialized object. 3755 // If there is one, return its index i. If there isn't, return the 3756 // negative of the index where it should be inserted. 3757 // Return 0 if the queried range overlaps an initialization boundary 3758 // or if dead code is encountered. 3759 // If size_in_bytes is zero, do not bother with overlap checks. 3760 int InitializeNode::captured_store_insertion_point(intptr_t start, 3761 int size_in_bytes, 3762 PhaseTransform* phase) { 3763 const int FAIL = 0, MAX_STORE = BytesPerLong; 3764 3765 if (is_complete()) 3766 return FAIL; // arraycopy got here first; punt 3767 3768 assert(allocation() != NULL, "must be present"); 3769 3770 // no negatives, no header fields: 3771 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3772 3773 // after a certain size, we bail out on tracking all the stores: 3774 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3775 if (start >= ti_limit) return FAIL; 3776 3777 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3778 if (i >= limit) return -(int)i; // not found; here is where to put it 3779 3780 Node* st = in(i); 3781 intptr_t st_off = get_store_offset(st, phase); 3782 if (st_off < 0) { 3783 if (st != zero_memory()) { 3784 return FAIL; // bail out if there is dead garbage 3785 } 3786 } else if (st_off > start) { 3787 // ...we are done, since stores are ordered 3788 if (st_off < start + size_in_bytes) { 3789 return FAIL; // the next store overlaps 3790 } 3791 return -(int)i; // not found; here is where to put it 3792 } else if (st_off < start) { 3793 if (size_in_bytes != 0 && 3794 start < st_off + MAX_STORE && 3795 start < st_off + st->as_Store()->memory_size()) { 3796 return FAIL; // the previous store overlaps 3797 } 3798 } else { 3799 if (size_in_bytes != 0 && 3800 st->as_Store()->memory_size() != size_in_bytes) { 3801 return FAIL; // mismatched store size 3802 } 3803 return i; 3804 } 3805 3806 ++i; 3807 } 3808 } 3809 3810 // Look for a captured store which initializes at the offset 'start' 3811 // with the given size. If there is no such store, and no other 3812 // initialization interferes, then return zero_memory (the memory 3813 // projection of the AllocateNode). 3814 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3815 PhaseTransform* phase) { 3816 assert(stores_are_sane(phase), ""); 3817 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3818 if (i == 0) { 3819 return NULL; // something is dead 3820 } else if (i < 0) { 3821 return zero_memory(); // just primordial zero bits here 3822 } else { 3823 Node* st = in(i); // here is the store at this position 3824 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3825 return st; 3826 } 3827 } 3828 3829 // Create, as a raw pointer, an address within my new object at 'offset'. 3830 Node* InitializeNode::make_raw_address(intptr_t offset, 3831 PhaseTransform* phase) { 3832 Node* addr = in(RawAddress); 3833 if (offset != 0) { 3834 Compile* C = phase->C; 3835 addr = phase->transform( new AddPNode(C->top(), addr, 3836 phase->MakeConX(offset)) ); 3837 } 3838 return addr; 3839 } 3840 3841 // Clone the given store, converting it into a raw store 3842 // initializing a field or element of my new object. 3843 // Caller is responsible for retiring the original store, 3844 // with subsume_node or the like. 3845 // 3846 // From the example above InitializeNode::InitializeNode, 3847 // here are the old stores to be captured: 3848 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3849 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3850 // 3851 // Here is the changed code; note the extra edges on init: 3852 // alloc = (Allocate ...) 3853 // rawoop = alloc.RawAddress 3854 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3855 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3856 // init = (Initialize alloc.Control alloc.Memory rawoop 3857 // rawstore1 rawstore2) 3858 // 3859 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3860 PhaseTransform* phase, bool can_reshape) { 3861 assert(stores_are_sane(phase), ""); 3862 3863 if (start < 0) return NULL; 3864 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3865 3866 Compile* C = phase->C; 3867 int size_in_bytes = st->memory_size(); 3868 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3869 if (i == 0) return NULL; // bail out 3870 Node* prev_mem = NULL; // raw memory for the captured store 3871 if (i > 0) { 3872 prev_mem = in(i); // there is a pre-existing store under this one 3873 set_req(i, C->top()); // temporarily disconnect it 3874 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3875 } else { 3876 i = -i; // no pre-existing store 3877 prev_mem = zero_memory(); // a slice of the newly allocated object 3878 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3879 set_req(--i, C->top()); // reuse this edge; it has been folded away 3880 else 3881 ins_req(i, C->top()); // build a new edge 3882 } 3883 Node* new_st = st->clone(); 3884 new_st->set_req(MemNode::Control, in(Control)); 3885 new_st->set_req(MemNode::Memory, prev_mem); 3886 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3887 new_st = phase->transform(new_st); 3888 3889 // At this point, new_st might have swallowed a pre-existing store 3890 // at the same offset, or perhaps new_st might have disappeared, 3891 // if it redundantly stored the same value (or zero to fresh memory). 3892 3893 // In any case, wire it in: 3894 phase->igvn_rehash_node_delayed(this); 3895 set_req(i, new_st); 3896 3897 // The caller may now kill the old guy. 3898 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3899 assert(check_st == new_st || check_st == NULL, "must be findable"); 3900 assert(!is_complete(), ""); 3901 return new_st; 3902 } 3903 3904 static bool store_constant(jlong* tiles, int num_tiles, 3905 intptr_t st_off, int st_size, 3906 jlong con) { 3907 if ((st_off & (st_size-1)) != 0) 3908 return false; // strange store offset (assume size==2**N) 3909 address addr = (address)tiles + st_off; 3910 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3911 switch (st_size) { 3912 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3913 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3914 case sizeof(jint): *(jint*) addr = (jint) con; break; 3915 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3916 default: return false; // strange store size (detect size!=2**N here) 3917 } 3918 return true; // return success to caller 3919 } 3920 3921 // Coalesce subword constants into int constants and possibly 3922 // into long constants. The goal, if the CPU permits, 3923 // is to initialize the object with a small number of 64-bit tiles. 3924 // Also, convert floating-point constants to bit patterns. 3925 // Non-constants are not relevant to this pass. 3926 // 3927 // In terms of the running example on InitializeNode::InitializeNode 3928 // and InitializeNode::capture_store, here is the transformation 3929 // of rawstore1 and rawstore2 into rawstore12: 3930 // alloc = (Allocate ...) 3931 // rawoop = alloc.RawAddress 3932 // tile12 = 0x00010002 3933 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3934 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3935 // 3936 void 3937 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3938 Node* size_in_bytes, 3939 PhaseGVN* phase) { 3940 Compile* C = phase->C; 3941 3942 assert(stores_are_sane(phase), ""); 3943 // Note: After this pass, they are not completely sane, 3944 // since there may be some overlaps. 3945 3946 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3947 3948 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3949 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3950 size_limit = MIN2(size_limit, ti_limit); 3951 size_limit = align_up(size_limit, BytesPerLong); 3952 int num_tiles = size_limit / BytesPerLong; 3953 3954 // allocate space for the tile map: 3955 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3956 jlong tiles_buf[small_len]; 3957 Node* nodes_buf[small_len]; 3958 jlong inits_buf[small_len]; 3959 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3960 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3961 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3962 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3963 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3964 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3965 // tiles: exact bitwise model of all primitive constants 3966 // nodes: last constant-storing node subsumed into the tiles model 3967 // inits: which bytes (in each tile) are touched by any initializations 3968 3969 //// Pass A: Fill in the tile model with any relevant stores. 3970 3971 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3972 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3973 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3974 Node* zmem = zero_memory(); // initially zero memory state 3975 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3976 Node* st = in(i); 3977 intptr_t st_off = get_store_offset(st, phase); 3978 3979 // Figure out the store's offset and constant value: 3980 if (st_off < header_size) continue; //skip (ignore header) 3981 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3982 int st_size = st->as_Store()->memory_size(); 3983 if (st_off + st_size > size_limit) break; 3984 3985 // Record which bytes are touched, whether by constant or not. 3986 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3987 continue; // skip (strange store size) 3988 3989 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3990 if (!val->singleton()) continue; //skip (non-con store) 3991 BasicType type = val->basic_type(); 3992 3993 jlong con = 0; 3994 switch (type) { 3995 case T_INT: con = val->is_int()->get_con(); break; 3996 case T_LONG: con = val->is_long()->get_con(); break; 3997 case T_FLOAT: con = jint_cast(val->getf()); break; 3998 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3999 default: continue; //skip (odd store type) 4000 } 4001 4002 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 4003 st->Opcode() == Op_StoreL) { 4004 continue; // This StoreL is already optimal. 4005 } 4006 4007 // Store down the constant. 4008 store_constant(tiles, num_tiles, st_off, st_size, con); 4009 4010 intptr_t j = st_off >> LogBytesPerLong; 4011 4012 if (type == T_INT && st_size == BytesPerInt 4013 && (st_off & BytesPerInt) == BytesPerInt) { 4014 jlong lcon = tiles[j]; 4015 if (!Matcher::isSimpleConstant64(lcon) && 4016 st->Opcode() == Op_StoreI) { 4017 // This StoreI is already optimal by itself. 4018 jint* intcon = (jint*) &tiles[j]; 4019 intcon[1] = 0; // undo the store_constant() 4020 4021 // If the previous store is also optimal by itself, back up and 4022 // undo the action of the previous loop iteration... if we can. 4023 // But if we can't, just let the previous half take care of itself. 4024 st = nodes[j]; 4025 st_off -= BytesPerInt; 4026 con = intcon[0]; 4027 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 4028 assert(st_off >= header_size, "still ignoring header"); 4029 assert(get_store_offset(st, phase) == st_off, "must be"); 4030 assert(in(i-1) == zmem, "must be"); 4031 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 4032 assert(con == tcon->is_int()->get_con(), "must be"); 4033 // Undo the effects of the previous loop trip, which swallowed st: 4034 intcon[0] = 0; // undo store_constant() 4035 set_req(i-1, st); // undo set_req(i, zmem) 4036 nodes[j] = NULL; // undo nodes[j] = st 4037 --old_subword; // undo ++old_subword 4038 } 4039 continue; // This StoreI is already optimal. 4040 } 4041 } 4042 4043 // This store is not needed. 4044 set_req(i, zmem); 4045 nodes[j] = st; // record for the moment 4046 if (st_size < BytesPerLong) // something has changed 4047 ++old_subword; // includes int/float, but who's counting... 4048 else ++old_long; 4049 } 4050 4051 if ((old_subword + old_long) == 0) 4052 return; // nothing more to do 4053 4054 //// Pass B: Convert any non-zero tiles into optimal constant stores. 4055 // Be sure to insert them before overlapping non-constant stores. 4056 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 4057 for (int j = 0; j < num_tiles; j++) { 4058 jlong con = tiles[j]; 4059 jlong init = inits[j]; 4060 if (con == 0) continue; 4061 jint con0, con1; // split the constant, address-wise 4062 jint init0, init1; // split the init map, address-wise 4063 { union { jlong con; jint intcon[2]; } u; 4064 u.con = con; 4065 con0 = u.intcon[0]; 4066 con1 = u.intcon[1]; 4067 u.con = init; 4068 init0 = u.intcon[0]; 4069 init1 = u.intcon[1]; 4070 } 4071 4072 Node* old = nodes[j]; 4073 assert(old != NULL, "need the prior store"); 4074 intptr_t offset = (j * BytesPerLong); 4075 4076 bool split = !Matcher::isSimpleConstant64(con); 4077 4078 if (offset < header_size) { 4079 assert(offset + BytesPerInt >= header_size, "second int counts"); 4080 assert(*(jint*)&tiles[j] == 0, "junk in header"); 4081 split = true; // only the second word counts 4082 // Example: int a[] = { 42 ... } 4083 } else if (con0 == 0 && init0 == -1) { 4084 split = true; // first word is covered by full inits 4085 // Example: int a[] = { ... foo(), 42 ... } 4086 } else if (con1 == 0 && init1 == -1) { 4087 split = true; // second word is covered by full inits 4088 // Example: int a[] = { ... 42, foo() ... } 4089 } 4090 4091 // Here's a case where init0 is neither 0 nor -1: 4092 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4093 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4094 // In this case the tile is not split; it is (jlong)42. 4095 // The big tile is stored down, and then the foo() value is inserted. 4096 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4097 4098 Node* ctl = old->in(MemNode::Control); 4099 Node* adr = make_raw_address(offset, phase); 4100 const TypePtr* atp = TypeRawPtr::BOTTOM; 4101 4102 // One or two coalesced stores to plop down. 4103 Node* st[2]; 4104 intptr_t off[2]; 4105 int nst = 0; 4106 if (!split) { 4107 ++new_long; 4108 off[nst] = offset; 4109 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4110 phase->longcon(con), T_LONG, MemNode::unordered); 4111 } else { 4112 // Omit either if it is a zero. 4113 if (con0 != 0) { 4114 ++new_int; 4115 off[nst] = offset; 4116 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4117 phase->intcon(con0), T_INT, MemNode::unordered); 4118 } 4119 if (con1 != 0) { 4120 ++new_int; 4121 offset += BytesPerInt; 4122 adr = make_raw_address(offset, phase); 4123 off[nst] = offset; 4124 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4125 phase->intcon(con1), T_INT, MemNode::unordered); 4126 } 4127 } 4128 4129 // Insert second store first, then the first before the second. 4130 // Insert each one just before any overlapping non-constant stores. 4131 while (nst > 0) { 4132 Node* st1 = st[--nst]; 4133 C->copy_node_notes_to(st1, old); 4134 st1 = phase->transform(st1); 4135 offset = off[nst]; 4136 assert(offset >= header_size, "do not smash header"); 4137 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4138 guarantee(ins_idx != 0, "must re-insert constant store"); 4139 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4140 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4141 set_req(--ins_idx, st1); 4142 else 4143 ins_req(ins_idx, st1); 4144 } 4145 } 4146 4147 if (PrintCompilation && WizardMode) 4148 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4149 old_subword, old_long, new_int, new_long); 4150 if (C->log() != NULL) 4151 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4152 old_subword, old_long, new_int, new_long); 4153 4154 // Clean up any remaining occurrences of zmem: 4155 remove_extra_zeroes(); 4156 } 4157 4158 // Explore forward from in(start) to find the first fully initialized 4159 // word, and return its offset. Skip groups of subword stores which 4160 // together initialize full words. If in(start) is itself part of a 4161 // fully initialized word, return the offset of in(start). If there 4162 // are no following full-word stores, or if something is fishy, return 4163 // a negative value. 4164 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4165 int int_map = 0; 4166 intptr_t int_map_off = 0; 4167 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4168 4169 for (uint i = start, limit = req(); i < limit; i++) { 4170 Node* st = in(i); 4171 4172 intptr_t st_off = get_store_offset(st, phase); 4173 if (st_off < 0) break; // return conservative answer 4174 4175 int st_size = st->as_Store()->memory_size(); 4176 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4177 return st_off; // we found a complete word init 4178 } 4179 4180 // update the map: 4181 4182 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4183 if (this_int_off != int_map_off) { 4184 // reset the map: 4185 int_map = 0; 4186 int_map_off = this_int_off; 4187 } 4188 4189 int subword_off = st_off - this_int_off; 4190 int_map |= right_n_bits(st_size) << subword_off; 4191 if ((int_map & FULL_MAP) == FULL_MAP) { 4192 return this_int_off; // we found a complete word init 4193 } 4194 4195 // Did this store hit or cross the word boundary? 4196 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4197 if (next_int_off == this_int_off + BytesPerInt) { 4198 // We passed the current int, without fully initializing it. 4199 int_map_off = next_int_off; 4200 int_map >>= BytesPerInt; 4201 } else if (next_int_off > this_int_off + BytesPerInt) { 4202 // We passed the current and next int. 4203 return this_int_off + BytesPerInt; 4204 } 4205 } 4206 4207 return -1; 4208 } 4209 4210 4211 // Called when the associated AllocateNode is expanded into CFG. 4212 // At this point, we may perform additional optimizations. 4213 // Linearize the stores by ascending offset, to make memory 4214 // activity as coherent as possible. 4215 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4216 intptr_t header_size, 4217 Node* size_in_bytes, 4218 PhaseIterGVN* phase) { 4219 assert(!is_complete(), "not already complete"); 4220 assert(stores_are_sane(phase), ""); 4221 assert(allocation() != NULL, "must be present"); 4222 4223 remove_extra_zeroes(); 4224 4225 if (ReduceFieldZeroing || ReduceBulkZeroing) 4226 // reduce instruction count for common initialization patterns 4227 coalesce_subword_stores(header_size, size_in_bytes, phase); 4228 4229 Node* zmem = zero_memory(); // initially zero memory state 4230 Node* inits = zmem; // accumulating a linearized chain of inits 4231 #ifdef ASSERT 4232 intptr_t first_offset = allocation()->minimum_header_size(); 4233 intptr_t last_init_off = first_offset; // previous init offset 4234 intptr_t last_init_end = first_offset; // previous init offset+size 4235 intptr_t last_tile_end = first_offset; // previous tile offset+size 4236 #endif 4237 intptr_t zeroes_done = header_size; 4238 4239 bool do_zeroing = true; // we might give up if inits are very sparse 4240 int big_init_gaps = 0; // how many large gaps have we seen? 4241 4242 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4243 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4244 4245 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4246 Node* st = in(i); 4247 intptr_t st_off = get_store_offset(st, phase); 4248 if (st_off < 0) 4249 break; // unknown junk in the inits 4250 if (st->in(MemNode::Memory) != zmem) 4251 break; // complicated store chains somehow in list 4252 4253 int st_size = st->as_Store()->memory_size(); 4254 intptr_t next_init_off = st_off + st_size; 4255 4256 if (do_zeroing && zeroes_done < next_init_off) { 4257 // See if this store needs a zero before it or under it. 4258 intptr_t zeroes_needed = st_off; 4259 4260 if (st_size < BytesPerInt) { 4261 // Look for subword stores which only partially initialize words. 4262 // If we find some, we must lay down some word-level zeroes first, 4263 // underneath the subword stores. 4264 // 4265 // Examples: 4266 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4267 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4268 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4269 // 4270 // Note: coalesce_subword_stores may have already done this, 4271 // if it was prompted by constant non-zero subword initializers. 4272 // But this case can still arise with non-constant stores. 4273 4274 intptr_t next_full_store = find_next_fullword_store(i, phase); 4275 4276 // In the examples above: 4277 // in(i) p q r s x y z 4278 // st_off 12 13 14 15 12 13 14 4279 // st_size 1 1 1 1 1 1 1 4280 // next_full_s. 12 16 16 16 16 16 16 4281 // z's_done 12 16 16 16 12 16 12 4282 // z's_needed 12 16 16 16 16 16 16 4283 // zsize 0 0 0 0 4 0 4 4284 if (next_full_store < 0) { 4285 // Conservative tack: Zero to end of current word. 4286 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4287 } else { 4288 // Zero to beginning of next fully initialized word. 4289 // Or, don't zero at all, if we are already in that word. 4290 assert(next_full_store >= zeroes_needed, "must go forward"); 4291 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4292 zeroes_needed = next_full_store; 4293 } 4294 } 4295 4296 if (zeroes_needed > zeroes_done) { 4297 intptr_t zsize = zeroes_needed - zeroes_done; 4298 // Do some incremental zeroing on rawmem, in parallel with inits. 4299 zeroes_done = align_down(zeroes_done, BytesPerInt); 4300 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4301 zeroes_done, zeroes_needed, 4302 phase); 4303 zeroes_done = zeroes_needed; 4304 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4305 do_zeroing = false; // leave the hole, next time 4306 } 4307 } 4308 4309 // Collect the store and move on: 4310 phase->replace_input_of(st, MemNode::Memory, inits); 4311 inits = st; // put it on the linearized chain 4312 set_req(i, zmem); // unhook from previous position 4313 4314 if (zeroes_done == st_off) 4315 zeroes_done = next_init_off; 4316 4317 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4318 4319 #ifdef ASSERT 4320 // Various order invariants. Weaker than stores_are_sane because 4321 // a large constant tile can be filled in by smaller non-constant stores. 4322 assert(st_off >= last_init_off, "inits do not reverse"); 4323 last_init_off = st_off; 4324 const Type* val = NULL; 4325 if (st_size >= BytesPerInt && 4326 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4327 (int)val->basic_type() < (int)T_OBJECT) { 4328 assert(st_off >= last_tile_end, "tiles do not overlap"); 4329 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4330 last_tile_end = MAX2(last_tile_end, next_init_off); 4331 } else { 4332 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4333 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4334 assert(st_off >= last_init_end, "inits do not overlap"); 4335 last_init_end = next_init_off; // it's a non-tile 4336 } 4337 #endif //ASSERT 4338 } 4339 4340 remove_extra_zeroes(); // clear out all the zmems left over 4341 add_req(inits); 4342 4343 if (!(UseTLAB && ZeroTLAB)) { 4344 // If anything remains to be zeroed, zero it all now. 4345 zeroes_done = align_down(zeroes_done, BytesPerInt); 4346 // if it is the last unused 4 bytes of an instance, forget about it 4347 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4348 if (zeroes_done + BytesPerLong >= size_limit) { 4349 AllocateNode* alloc = allocation(); 4350 assert(alloc != NULL, "must be present"); 4351 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 4352 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4353 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4354 if (zeroes_done == k->layout_helper()) 4355 zeroes_done = size_limit; 4356 } 4357 } 4358 if (zeroes_done < size_limit) { 4359 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4360 zeroes_done, size_in_bytes, phase); 4361 } 4362 } 4363 4364 set_complete(phase); 4365 return rawmem; 4366 } 4367 4368 4369 #ifdef ASSERT 4370 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4371 if (is_complete()) 4372 return true; // stores could be anything at this point 4373 assert(allocation() != NULL, "must be present"); 4374 intptr_t last_off = allocation()->minimum_header_size(); 4375 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4376 Node* st = in(i); 4377 intptr_t st_off = get_store_offset(st, phase); 4378 if (st_off < 0) continue; // ignore dead garbage 4379 if (last_off > st_off) { 4380 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4381 this->dump(2); 4382 assert(false, "ascending store offsets"); 4383 return false; 4384 } 4385 last_off = st_off + st->as_Store()->memory_size(); 4386 } 4387 return true; 4388 } 4389 #endif //ASSERT 4390 4391 4392 4393 4394 //============================MergeMemNode===================================== 4395 // 4396 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4397 // contributing store or call operations. Each contributor provides the memory 4398 // state for a particular "alias type" (see Compile::alias_type). For example, 4399 // if a MergeMem has an input X for alias category #6, then any memory reference 4400 // to alias category #6 may use X as its memory state input, as an exact equivalent 4401 // to using the MergeMem as a whole. 4402 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4403 // 4404 // (Here, the <N> notation gives the index of the relevant adr_type.) 4405 // 4406 // In one special case (and more cases in the future), alias categories overlap. 4407 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4408 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4409 // it is exactly equivalent to that state W: 4410 // MergeMem(<Bot>: W) <==> W 4411 // 4412 // Usually, the merge has more than one input. In that case, where inputs 4413 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4414 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4415 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4416 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4417 // 4418 // A merge can take a "wide" memory state as one of its narrow inputs. 4419 // This simply means that the merge observes out only the relevant parts of 4420 // the wide input. That is, wide memory states arriving at narrow merge inputs 4421 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4422 // 4423 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4424 // and that memory slices "leak through": 4425 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4426 // 4427 // But, in such a cascade, repeated memory slices can "block the leak": 4428 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4429 // 4430 // In the last example, Y is not part of the combined memory state of the 4431 // outermost MergeMem. The system must, of course, prevent unschedulable 4432 // memory states from arising, so you can be sure that the state Y is somehow 4433 // a precursor to state Y'. 4434 // 4435 // 4436 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4437 // of each MergeMemNode array are exactly the numerical alias indexes, including 4438 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4439 // Compile::alias_type (and kin) produce and manage these indexes. 4440 // 4441 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4442 // (Note that this provides quick access to the top node inside MergeMem methods, 4443 // without the need to reach out via TLS to Compile::current.) 4444 // 4445 // As a consequence of what was just described, a MergeMem that represents a full 4446 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4447 // containing all alias categories. 4448 // 4449 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4450 // 4451 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4452 // a memory state for the alias type <N>, or else the top node, meaning that 4453 // there is no particular input for that alias type. Note that the length of 4454 // a MergeMem is variable, and may be extended at any time to accommodate new 4455 // memory states at larger alias indexes. When merges grow, they are of course 4456 // filled with "top" in the unused in() positions. 4457 // 4458 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4459 // (Top was chosen because it works smoothly with passes like GCM.) 4460 // 4461 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4462 // the type of random VM bits like TLS references.) Since it is always the 4463 // first non-Bot memory slice, some low-level loops use it to initialize an 4464 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4465 // 4466 // 4467 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4468 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4469 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4470 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4471 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4472 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4473 // 4474 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4475 // really that different from the other memory inputs. An abbreviation called 4476 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4477 // 4478 // 4479 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4480 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4481 // that "emerges though" the base memory will be marked as excluding the alias types 4482 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4483 // 4484 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4485 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4486 // 4487 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4488 // (It is currently unimplemented.) As you can see, the resulting merge is 4489 // actually a disjoint union of memory states, rather than an overlay. 4490 // 4491 4492 //------------------------------MergeMemNode----------------------------------- 4493 Node* MergeMemNode::make_empty_memory() { 4494 Node* empty_memory = (Node*) Compile::current()->top(); 4495 assert(empty_memory->is_top(), "correct sentinel identity"); 4496 return empty_memory; 4497 } 4498 4499 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4500 init_class_id(Class_MergeMem); 4501 // all inputs are nullified in Node::Node(int) 4502 // set_input(0, NULL); // no control input 4503 4504 // Initialize the edges uniformly to top, for starters. 4505 Node* empty_mem = make_empty_memory(); 4506 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4507 init_req(i,empty_mem); 4508 } 4509 assert(empty_memory() == empty_mem, ""); 4510 4511 if( new_base != NULL && new_base->is_MergeMem() ) { 4512 MergeMemNode* mdef = new_base->as_MergeMem(); 4513 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4514 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4515 mms.set_memory(mms.memory2()); 4516 } 4517 assert(base_memory() == mdef->base_memory(), ""); 4518 } else { 4519 set_base_memory(new_base); 4520 } 4521 } 4522 4523 // Make a new, untransformed MergeMem with the same base as 'mem'. 4524 // If mem is itself a MergeMem, populate the result with the same edges. 4525 MergeMemNode* MergeMemNode::make(Node* mem) { 4526 return new MergeMemNode(mem); 4527 } 4528 4529 //------------------------------cmp-------------------------------------------- 4530 uint MergeMemNode::hash() const { return NO_HASH; } 4531 uint MergeMemNode::cmp( const Node &n ) const { 4532 return (&n == this); // Always fail except on self 4533 } 4534 4535 //------------------------------Identity--------------------------------------- 4536 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4537 // Identity if this merge point does not record any interesting memory 4538 // disambiguations. 4539 Node* base_mem = base_memory(); 4540 Node* empty_mem = empty_memory(); 4541 if (base_mem != empty_mem) { // Memory path is not dead? 4542 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4543 Node* mem = in(i); 4544 if (mem != empty_mem && mem != base_mem) { 4545 return this; // Many memory splits; no change 4546 } 4547 } 4548 } 4549 return base_mem; // No memory splits; ID on the one true input 4550 } 4551 4552 //------------------------------Ideal------------------------------------------ 4553 // This method is invoked recursively on chains of MergeMem nodes 4554 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4555 // Remove chain'd MergeMems 4556 // 4557 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4558 // relative to the "in(Bot)". Since we are patching both at the same time, 4559 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4560 // but rewrite each "in(i)" relative to the new "in(Bot)". 4561 Node *progress = NULL; 4562 4563 4564 Node* old_base = base_memory(); 4565 Node* empty_mem = empty_memory(); 4566 if (old_base == empty_mem) 4567 return NULL; // Dead memory path. 4568 4569 MergeMemNode* old_mbase; 4570 if (old_base != NULL && old_base->is_MergeMem()) 4571 old_mbase = old_base->as_MergeMem(); 4572 else 4573 old_mbase = NULL; 4574 Node* new_base = old_base; 4575 4576 // simplify stacked MergeMems in base memory 4577 if (old_mbase) new_base = old_mbase->base_memory(); 4578 4579 // the base memory might contribute new slices beyond my req() 4580 if (old_mbase) grow_to_match(old_mbase); 4581 4582 // Look carefully at the base node if it is a phi. 4583 PhiNode* phi_base; 4584 if (new_base != NULL && new_base->is_Phi()) 4585 phi_base = new_base->as_Phi(); 4586 else 4587 phi_base = NULL; 4588 4589 Node* phi_reg = NULL; 4590 uint phi_len = (uint)-1; 4591 if (phi_base != NULL && !phi_base->is_copy()) { 4592 // do not examine phi if degraded to a copy 4593 phi_reg = phi_base->region(); 4594 phi_len = phi_base->req(); 4595 // see if the phi is unfinished 4596 for (uint i = 1; i < phi_len; i++) { 4597 if (phi_base->in(i) == NULL) { 4598 // incomplete phi; do not look at it yet! 4599 phi_reg = NULL; 4600 phi_len = (uint)-1; 4601 break; 4602 } 4603 } 4604 } 4605 4606 // Note: We do not call verify_sparse on entry, because inputs 4607 // can normalize to the base_memory via subsume_node or similar 4608 // mechanisms. This method repairs that damage. 4609 4610 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4611 4612 // Look at each slice. 4613 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4614 Node* old_in = in(i); 4615 // calculate the old memory value 4616 Node* old_mem = old_in; 4617 if (old_mem == empty_mem) old_mem = old_base; 4618 assert(old_mem == memory_at(i), ""); 4619 4620 // maybe update (reslice) the old memory value 4621 4622 // simplify stacked MergeMems 4623 Node* new_mem = old_mem; 4624 MergeMemNode* old_mmem; 4625 if (old_mem != NULL && old_mem->is_MergeMem()) 4626 old_mmem = old_mem->as_MergeMem(); 4627 else 4628 old_mmem = NULL; 4629 if (old_mmem == this) { 4630 // This can happen if loops break up and safepoints disappear. 4631 // A merge of BotPtr (default) with a RawPtr memory derived from a 4632 // safepoint can be rewritten to a merge of the same BotPtr with 4633 // the BotPtr phi coming into the loop. If that phi disappears 4634 // also, we can end up with a self-loop of the mergemem. 4635 // In general, if loops degenerate and memory effects disappear, 4636 // a mergemem can be left looking at itself. This simply means 4637 // that the mergemem's default should be used, since there is 4638 // no longer any apparent effect on this slice. 4639 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4640 // from start. Update the input to TOP. 4641 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4642 } 4643 else if (old_mmem != NULL) { 4644 new_mem = old_mmem->memory_at(i); 4645 } 4646 // else preceding memory was not a MergeMem 4647 4648 // maybe store down a new value 4649 Node* new_in = new_mem; 4650 if (new_in == new_base) new_in = empty_mem; 4651 4652 if (new_in != old_in) { 4653 // Warning: Do not combine this "if" with the previous "if" 4654 // A memory slice might have be be rewritten even if it is semantically 4655 // unchanged, if the base_memory value has changed. 4656 set_req(i, new_in); 4657 progress = this; // Report progress 4658 } 4659 } 4660 4661 if (new_base != old_base) { 4662 set_req(Compile::AliasIdxBot, new_base); 4663 // Don't use set_base_memory(new_base), because we need to update du. 4664 assert(base_memory() == new_base, ""); 4665 progress = this; 4666 } 4667 4668 if( base_memory() == this ) { 4669 // a self cycle indicates this memory path is dead 4670 set_req(Compile::AliasIdxBot, empty_mem); 4671 } 4672 4673 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4674 // Recursion must occur after the self cycle check above 4675 if( base_memory()->is_MergeMem() ) { 4676 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4677 Node *m = phase->transform(new_mbase); // Rollup any cycles 4678 if( m != NULL && 4679 (m->is_top() || 4680 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4681 // propagate rollup of dead cycle to self 4682 set_req(Compile::AliasIdxBot, empty_mem); 4683 } 4684 } 4685 4686 if( base_memory() == empty_mem ) { 4687 progress = this; 4688 // Cut inputs during Parse phase only. 4689 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4690 if( !can_reshape ) { 4691 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4692 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4693 } 4694 } 4695 } 4696 4697 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4698 // Check if PhiNode::Ideal's "Split phis through memory merges" 4699 // transform should be attempted. Look for this->phi->this cycle. 4700 uint merge_width = req(); 4701 if (merge_width > Compile::AliasIdxRaw) { 4702 PhiNode* phi = base_memory()->as_Phi(); 4703 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4704 if (phi->in(i) == this) { 4705 phase->is_IterGVN()->_worklist.push(phi); 4706 break; 4707 } 4708 } 4709 } 4710 } 4711 4712 assert(progress || verify_sparse(), "please, no dups of base"); 4713 return progress; 4714 } 4715 4716 //-------------------------set_base_memory------------------------------------- 4717 void MergeMemNode::set_base_memory(Node *new_base) { 4718 Node* empty_mem = empty_memory(); 4719 set_req(Compile::AliasIdxBot, new_base); 4720 assert(memory_at(req()) == new_base, "must set default memory"); 4721 // Clear out other occurrences of new_base: 4722 if (new_base != empty_mem) { 4723 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4724 if (in(i) == new_base) set_req(i, empty_mem); 4725 } 4726 } 4727 } 4728 4729 //------------------------------out_RegMask------------------------------------ 4730 const RegMask &MergeMemNode::out_RegMask() const { 4731 return RegMask::Empty; 4732 } 4733 4734 //------------------------------dump_spec-------------------------------------- 4735 #ifndef PRODUCT 4736 void MergeMemNode::dump_spec(outputStream *st) const { 4737 st->print(" {"); 4738 Node* base_mem = base_memory(); 4739 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4740 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4741 if (mem == base_mem) { st->print(" -"); continue; } 4742 st->print( " N%d:", mem->_idx ); 4743 Compile::current()->get_adr_type(i)->dump_on(st); 4744 } 4745 st->print(" }"); 4746 } 4747 #endif // !PRODUCT 4748 4749 4750 #ifdef ASSERT 4751 static bool might_be_same(Node* a, Node* b) { 4752 if (a == b) return true; 4753 if (!(a->is_Phi() || b->is_Phi())) return false; 4754 // phis shift around during optimization 4755 return true; // pretty stupid... 4756 } 4757 4758 // verify a narrow slice (either incoming or outgoing) 4759 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4760 if (!VerifyAliases) return; // don't bother to verify unless requested 4761 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4762 if (Node::in_dump()) return; // muzzle asserts when printing 4763 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4764 assert(n != NULL, ""); 4765 // Elide intervening MergeMem's 4766 while (n->is_MergeMem()) { 4767 n = n->as_MergeMem()->memory_at(alias_idx); 4768 } 4769 Compile* C = Compile::current(); 4770 const TypePtr* n_adr_type = n->adr_type(); 4771 if (n == m->empty_memory()) { 4772 // Implicit copy of base_memory() 4773 } else if (n_adr_type != TypePtr::BOTTOM) { 4774 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4775 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4776 } else { 4777 // A few places like make_runtime_call "know" that VM calls are narrow, 4778 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4779 bool expected_wide_mem = false; 4780 if (n == m->base_memory()) { 4781 expected_wide_mem = true; 4782 } else if (alias_idx == Compile::AliasIdxRaw || 4783 n == m->memory_at(Compile::AliasIdxRaw)) { 4784 expected_wide_mem = true; 4785 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4786 // memory can "leak through" calls on channels that 4787 // are write-once. Allow this also. 4788 expected_wide_mem = true; 4789 } 4790 assert(expected_wide_mem, "expected narrow slice replacement"); 4791 } 4792 } 4793 #else // !ASSERT 4794 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4795 #endif 4796 4797 4798 //-----------------------------memory_at--------------------------------------- 4799 Node* MergeMemNode::memory_at(uint alias_idx) const { 4800 assert(alias_idx >= Compile::AliasIdxRaw || 4801 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4802 "must avoid base_memory and AliasIdxTop"); 4803 4804 // Otherwise, it is a narrow slice. 4805 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4806 Compile *C = Compile::current(); 4807 if (is_empty_memory(n)) { 4808 // the array is sparse; empty slots are the "top" node 4809 n = base_memory(); 4810 assert(Node::in_dump() 4811 || n == NULL || n->bottom_type() == Type::TOP 4812 || n->adr_type() == NULL // address is TOP 4813 || n->adr_type() == TypePtr::BOTTOM 4814 || n->adr_type() == TypeRawPtr::BOTTOM 4815 || Compile::current()->AliasLevel() == 0, 4816 "must be a wide memory"); 4817 // AliasLevel == 0 if we are organizing the memory states manually. 4818 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4819 } else { 4820 // make sure the stored slice is sane 4821 #ifdef ASSERT 4822 if (VMError::is_error_reported() || Node::in_dump()) { 4823 } else if (might_be_same(n, base_memory())) { 4824 // Give it a pass: It is a mostly harmless repetition of the base. 4825 // This can arise normally from node subsumption during optimization. 4826 } else { 4827 verify_memory_slice(this, alias_idx, n); 4828 } 4829 #endif 4830 } 4831 return n; 4832 } 4833 4834 //---------------------------set_memory_at------------------------------------- 4835 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4836 verify_memory_slice(this, alias_idx, n); 4837 Node* empty_mem = empty_memory(); 4838 if (n == base_memory()) n = empty_mem; // collapse default 4839 uint need_req = alias_idx+1; 4840 if (req() < need_req) { 4841 if (n == empty_mem) return; // already the default, so do not grow me 4842 // grow the sparse array 4843 do { 4844 add_req(empty_mem); 4845 } while (req() < need_req); 4846 } 4847 set_req( alias_idx, n ); 4848 } 4849 4850 4851 4852 //--------------------------iteration_setup------------------------------------ 4853 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4854 if (other != NULL) { 4855 grow_to_match(other); 4856 // invariant: the finite support of mm2 is within mm->req() 4857 #ifdef ASSERT 4858 for (uint i = req(); i < other->req(); i++) { 4859 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4860 } 4861 #endif 4862 } 4863 // Replace spurious copies of base_memory by top. 4864 Node* base_mem = base_memory(); 4865 if (base_mem != NULL && !base_mem->is_top()) { 4866 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4867 if (in(i) == base_mem) 4868 set_req(i, empty_memory()); 4869 } 4870 } 4871 } 4872 4873 //---------------------------grow_to_match------------------------------------- 4874 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4875 Node* empty_mem = empty_memory(); 4876 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4877 // look for the finite support of the other memory 4878 for (uint i = other->req(); --i >= req(); ) { 4879 if (other->in(i) != empty_mem) { 4880 uint new_len = i+1; 4881 while (req() < new_len) add_req(empty_mem); 4882 break; 4883 } 4884 } 4885 } 4886 4887 //---------------------------verify_sparse------------------------------------- 4888 #ifndef PRODUCT 4889 bool MergeMemNode::verify_sparse() const { 4890 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4891 Node* base_mem = base_memory(); 4892 // The following can happen in degenerate cases, since empty==top. 4893 if (is_empty_memory(base_mem)) return true; 4894 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4895 assert(in(i) != NULL, "sane slice"); 4896 if (in(i) == base_mem) return false; // should have been the sentinel value! 4897 } 4898 return true; 4899 } 4900 4901 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4902 Node* n; 4903 n = mm->in(idx); 4904 if (mem == n) return true; // might be empty_memory() 4905 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4906 if (mem == n) return true; 4907 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4908 if (mem == n) return true; 4909 if (n == NULL) break; 4910 } 4911 return false; 4912 } 4913 #endif // !PRODUCT