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 "opto/valuetypenode.hpp" 49 #include "utilities/align.hpp" 50 #include "utilities/copy.hpp" 51 #include "utilities/macros.hpp" 52 #include "utilities/vmError.hpp" 53 54 // Portions of code courtesy of Clifford Click 55 56 // Optimization - Graph Style 57 58 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 59 60 //============================================================================= 61 uint MemNode::size_of() const { return sizeof(*this); } 62 63 const TypePtr *MemNode::adr_type() const { 64 Node* adr = in(Address); 65 if (adr == NULL) return NULL; // node is dead 66 const TypePtr* cross_check = NULL; 67 DEBUG_ONLY(cross_check = _adr_type); 68 return calculate_adr_type(adr->bottom_type(), cross_check); 69 } 70 71 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 72 if (adr != NULL) { 73 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 74 return true; 75 } 76 } 77 return false; 78 } 79 80 #ifndef PRODUCT 81 void MemNode::dump_spec(outputStream *st) const { 82 if (in(Address) == NULL) return; // node is dead 83 #ifndef ASSERT 84 // fake the missing field 85 const TypePtr* _adr_type = NULL; 86 if (in(Address) != NULL) 87 _adr_type = in(Address)->bottom_type()->isa_ptr(); 88 #endif 89 dump_adr_type(this, _adr_type, st); 90 91 Compile* C = Compile::current(); 92 if (C->alias_type(_adr_type)->is_volatile()) { 93 st->print(" Volatile!"); 94 } 95 if (_unaligned_access) { 96 st->print(" unaligned"); 97 } 98 if (_mismatched_access) { 99 st->print(" mismatched"); 100 } 101 if (_unsafe_access) { 102 st->print(" unsafe"); 103 } 104 } 105 106 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 107 st->print(" @"); 108 if (adr_type == NULL) { 109 st->print("NULL"); 110 } else { 111 adr_type->dump_on(st); 112 Compile* C = Compile::current(); 113 Compile::AliasType* atp = NULL; 114 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 115 if (atp == NULL) 116 st->print(", idx=?\?;"); 117 else if (atp->index() == Compile::AliasIdxBot) 118 st->print(", idx=Bot;"); 119 else if (atp->index() == Compile::AliasIdxTop) 120 st->print(", idx=Top;"); 121 else if (atp->index() == Compile::AliasIdxRaw) 122 st->print(", idx=Raw;"); 123 else { 124 ciField* field = atp->field(); 125 if (field) { 126 st->print(", name="); 127 field->print_name_on(st); 128 } 129 st->print(", idx=%d;", atp->index()); 130 } 131 } 132 } 133 134 extern void print_alias_types(); 135 136 #endif 137 138 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 139 assert((t_oop != NULL), "sanity"); 140 bool is_instance = t_oop->is_known_instance_field(); 141 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 142 (load != NULL) && load->is_Load() && 143 (phase->is_IterGVN() != NULL); 144 if (!(is_instance || is_boxed_value_load)) 145 return mchain; // don't try to optimize non-instance types 146 uint instance_id = t_oop->instance_id(); 147 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory); 148 Node *prev = NULL; 149 Node *result = mchain; 150 while (prev != result) { 151 prev = result; 152 if (result == start_mem) 153 break; // hit one of our sentinels 154 // skip over a call which does not affect this memory slice 155 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 156 Node *proj_in = result->in(0); 157 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 158 break; // hit one of our sentinels 159 } else if (proj_in->is_Call()) { 160 // ArrayCopyNodes processed here as well 161 CallNode *call = proj_in->as_Call(); 162 if (!call->may_modify(t_oop, phase)) { // returns false for instances 163 result = call->in(TypeFunc::Memory); 164 } 165 } else if (proj_in->is_Initialize()) { 166 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 167 // Stop if this is the initialization for the object instance which 168 // contains this memory slice, otherwise skip over it. 169 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 170 break; 171 } 172 if (is_instance) { 173 result = proj_in->in(TypeFunc::Memory); 174 } else if (is_boxed_value_load) { 175 Node* klass = alloc->in(AllocateNode::KlassNode); 176 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 177 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 178 result = proj_in->in(TypeFunc::Memory); // not related allocation 179 } 180 } 181 } else if (proj_in->is_MemBar()) { 182 ArrayCopyNode* ac = NULL; 183 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 184 break; 185 } 186 result = proj_in->in(TypeFunc::Memory); 187 } else { 188 assert(false, "unexpected projection"); 189 } 190 } else if (result->is_ClearArray()) { 191 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 192 // Can not bypass initialization of the instance 193 // we are looking for. 194 break; 195 } 196 // Otherwise skip it (the call updated 'result' value). 197 } else if (result->is_MergeMem()) { 198 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 199 } 200 } 201 return result; 202 } 203 204 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 205 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 206 if (t_oop == NULL) 207 return mchain; // don't try to optimize non-oop types 208 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 209 bool is_instance = t_oop->is_known_instance_field(); 210 PhaseIterGVN *igvn = phase->is_IterGVN(); 211 if (is_instance && igvn != NULL && result->is_Phi()) { 212 PhiNode *mphi = result->as_Phi(); 213 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 214 const TypePtr *t = mphi->adr_type(); 215 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 216 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 217 t->is_oopptr()->cast_to_exactness(true) 218 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 219 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) { 220 // clone the Phi with our address type 221 result = mphi->split_out_instance(t_adr, igvn); 222 } else { 223 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 224 } 225 } 226 return result; 227 } 228 229 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 230 uint alias_idx = phase->C->get_alias_index(tp); 231 Node *mem = mmem; 232 #ifdef ASSERT 233 { 234 // Check that current type is consistent with the alias index used during graph construction 235 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 236 bool consistent = adr_check == NULL || adr_check->empty() || 237 phase->C->must_alias(adr_check, alias_idx ); 238 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 239 if( !consistent && adr_check != NULL && !adr_check->empty() && 240 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 241 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 242 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 243 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 244 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 245 // don't assert if it is dead code. 246 consistent = true; 247 } 248 if( !consistent ) { 249 st->print("alias_idx==%d, adr_check==", alias_idx); 250 if( adr_check == NULL ) { 251 st->print("NULL"); 252 } else { 253 adr_check->dump(); 254 } 255 st->cr(); 256 print_alias_types(); 257 assert(consistent, "adr_check must match alias idx"); 258 } 259 } 260 #endif 261 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 262 // means an array I have not precisely typed yet. Do not do any 263 // alias stuff with it any time soon. 264 const TypeOopPtr *toop = tp->isa_oopptr(); 265 if( tp->base() != Type::AnyPtr && 266 !(toop && 267 toop->klass() != NULL && 268 toop->klass()->is_java_lang_Object() && 269 toop->offset() == Type::OffsetBot) ) { 270 // compress paths and change unreachable cycles to TOP 271 // If not, we can update the input infinitely along a MergeMem cycle 272 // Equivalent code in PhiNode::Ideal 273 Node* m = phase->transform(mmem); 274 // If transformed to a MergeMem, get the desired slice 275 // Otherwise the returned node represents memory for every slice 276 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 277 // Update input if it is progress over what we have now 278 } 279 return mem; 280 } 281 282 //--------------------------Ideal_common--------------------------------------- 283 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 284 // Unhook non-raw memories from complete (macro-expanded) initializations. 285 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 286 // If our control input is a dead region, kill all below the region 287 Node *ctl = in(MemNode::Control); 288 if (ctl && remove_dead_region(phase, can_reshape)) 289 return this; 290 ctl = in(MemNode::Control); 291 // Don't bother trying to transform a dead node 292 if (ctl && ctl->is_top()) return NodeSentinel; 293 294 PhaseIterGVN *igvn = phase->is_IterGVN(); 295 // Wait if control on the worklist. 296 if (ctl && can_reshape && igvn != NULL) { 297 Node* bol = NULL; 298 Node* cmp = NULL; 299 if (ctl->in(0)->is_If()) { 300 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 301 bol = ctl->in(0)->in(1); 302 if (bol->is_Bool()) 303 cmp = ctl->in(0)->in(1)->in(1); 304 } 305 if (igvn->_worklist.member(ctl) || 306 (bol != NULL && igvn->_worklist.member(bol)) || 307 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 308 // This control path may be dead. 309 // Delay this memory node transformation until the control is processed. 310 phase->is_IterGVN()->_worklist.push(this); 311 return NodeSentinel; // caller will return NULL 312 } 313 } 314 // Ignore if memory is dead, or self-loop 315 Node *mem = in(MemNode::Memory); 316 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 317 assert(mem != this, "dead loop in MemNode::Ideal"); 318 319 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 320 // This memory slice may be dead. 321 // Delay this mem node transformation until the memory is processed. 322 phase->is_IterGVN()->_worklist.push(this); 323 return NodeSentinel; // caller will return NULL 324 } 325 326 Node *address = in(MemNode::Address); 327 const Type *t_adr = phase->type(address); 328 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 329 330 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) { 331 // Unsafe off-heap access with zero address. Remove access and other control users 332 // to not confuse optimizations and add a HaltNode to fail if this is ever executed. 333 assert(ctl != NULL, "unsafe accesses should be control dependent"); 334 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) { 335 Node* u = ctl->fast_out(i); 336 if (u != ctl) { 337 igvn->rehash_node_delayed(u); 338 int nb = u->replace_edge(ctl, phase->C->top()); 339 --i, imax -= nb; 340 } 341 } 342 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr)); 343 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address")); 344 phase->C->root()->add_req(halt); 345 return this; 346 } 347 348 if (can_reshape && igvn != NULL && 349 (igvn->_worklist.member(address) || 350 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 351 // The address's base and type may change when the address is processed. 352 // Delay this mem node transformation until the address is processed. 353 phase->is_IterGVN()->_worklist.push(this); 354 return NodeSentinel; // caller will return NULL 355 } 356 357 // Do NOT remove or optimize the next lines: ensure a new alias index 358 // is allocated for an oop pointer type before Escape Analysis. 359 // Note: C++ will not remove it since the call has side effect. 360 if (t_adr->isa_oopptr()) { 361 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 362 } 363 364 Node* base = NULL; 365 if (address->is_AddP()) { 366 base = address->in(AddPNode::Base); 367 } 368 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 369 !t_adr->isa_rawptr()) { 370 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 371 // Skip this node optimization if its address has TOP base. 372 return NodeSentinel; // caller will return NULL 373 } 374 375 // Avoid independent memory operations 376 Node* old_mem = mem; 377 378 // The code which unhooks non-raw memories from complete (macro-expanded) 379 // initializations was removed. After macro-expansion all stores catched 380 // by Initialize node became raw stores and there is no information 381 // which memory slices they modify. So it is unsafe to move any memory 382 // operation above these stores. Also in most cases hooked non-raw memories 383 // were already unhooked by using information from detect_ptr_independence() 384 // and find_previous_store(). 385 386 if (mem->is_MergeMem()) { 387 MergeMemNode* mmem = mem->as_MergeMem(); 388 const TypePtr *tp = t_adr->is_ptr(); 389 390 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 391 } 392 393 if (mem != old_mem) { 394 set_req(MemNode::Memory, mem); 395 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) { 396 igvn->_worklist.push(old_mem); 397 } 398 if (phase->type(mem) == Type::TOP) return NodeSentinel; 399 return this; 400 } 401 402 // let the subclass continue analyzing... 403 return NULL; 404 } 405 406 // Helper function for proving some simple control dominations. 407 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 408 // Already assumes that 'dom' is available at 'sub', and that 'sub' 409 // is not a constant (dominated by the method's StartNode). 410 // Used by MemNode::find_previous_store to prove that the 411 // control input of a memory operation predates (dominates) 412 // an allocation it wants to look past. 413 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 414 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 415 return false; // Conservative answer for dead code 416 417 // Check 'dom'. Skip Proj and CatchProj nodes. 418 dom = dom->find_exact_control(dom); 419 if (dom == NULL || dom->is_top()) 420 return false; // Conservative answer for dead code 421 422 if (dom == sub) { 423 // For the case when, for example, 'sub' is Initialize and the original 424 // 'dom' is Proj node of the 'sub'. 425 return false; 426 } 427 428 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 429 return true; 430 431 // 'dom' dominates 'sub' if its control edge and control edges 432 // of all its inputs dominate or equal to sub's control edge. 433 434 // Currently 'sub' is either Allocate, Initialize or Start nodes. 435 // Or Region for the check in LoadNode::Ideal(); 436 // 'sub' should have sub->in(0) != NULL. 437 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 438 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 439 440 // Get control edge of 'sub'. 441 Node* orig_sub = sub; 442 sub = sub->find_exact_control(sub->in(0)); 443 if (sub == NULL || sub->is_top()) 444 return false; // Conservative answer for dead code 445 446 assert(sub->is_CFG(), "expecting control"); 447 448 if (sub == dom) 449 return true; 450 451 if (sub->is_Start() || sub->is_Root()) 452 return false; 453 454 { 455 // Check all control edges of 'dom'. 456 457 ResourceMark rm; 458 Arena* arena = Thread::current()->resource_area(); 459 Node_List nlist(arena); 460 Unique_Node_List dom_list(arena); 461 462 dom_list.push(dom); 463 bool only_dominating_controls = false; 464 465 for (uint next = 0; next < dom_list.size(); next++) { 466 Node* n = dom_list.at(next); 467 if (n == orig_sub) 468 return false; // One of dom's inputs dominated by sub. 469 if (!n->is_CFG() && n->pinned()) { 470 // Check only own control edge for pinned non-control nodes. 471 n = n->find_exact_control(n->in(0)); 472 if (n == NULL || n->is_top()) 473 return false; // Conservative answer for dead code 474 assert(n->is_CFG(), "expecting control"); 475 dom_list.push(n); 476 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 477 only_dominating_controls = true; 478 } else if (n->is_CFG()) { 479 if (n->dominates(sub, nlist)) 480 only_dominating_controls = true; 481 else 482 return false; 483 } else { 484 // First, own control edge. 485 Node* m = n->find_exact_control(n->in(0)); 486 if (m != NULL) { 487 if (m->is_top()) 488 return false; // Conservative answer for dead code 489 dom_list.push(m); 490 } 491 // Now, the rest of edges. 492 uint cnt = n->req(); 493 for (uint i = 1; i < cnt; i++) { 494 m = n->find_exact_control(n->in(i)); 495 if (m == NULL || m->is_top()) 496 continue; 497 dom_list.push(m); 498 } 499 } 500 } 501 return only_dominating_controls; 502 } 503 } 504 505 //---------------------detect_ptr_independence--------------------------------- 506 // Used by MemNode::find_previous_store to prove that two base 507 // pointers are never equal. 508 // The pointers are accompanied by their associated allocations, 509 // if any, which have been previously discovered by the caller. 510 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 511 Node* p2, AllocateNode* a2, 512 PhaseTransform* phase) { 513 // Attempt to prove that these two pointers cannot be aliased. 514 // They may both manifestly be allocations, and they should differ. 515 // Or, if they are not both allocations, they can be distinct constants. 516 // Otherwise, one is an allocation and the other a pre-existing value. 517 if (a1 == NULL && a2 == NULL) { // neither an allocation 518 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 519 } else if (a1 != NULL && a2 != NULL) { // both allocations 520 return (a1 != a2); 521 } else if (a1 != NULL) { // one allocation a1 522 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 523 return all_controls_dominate(p2, a1); 524 } else { //(a2 != NULL) // one allocation a2 525 return all_controls_dominate(p1, a2); 526 } 527 return false; 528 } 529 530 531 // Find an arraycopy that must have set (can_see_stored_value=true) or 532 // could have set (can_see_stored_value=false) the value for this load 533 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 534 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 535 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 536 Node* mb = mem->in(0); 537 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 538 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 539 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 540 if (ac->is_clonebasic()) { 541 intptr_t offset; 542 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 543 if (alloc != NULL && alloc == ld_alloc) { 544 return ac; 545 } 546 } 547 } 548 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 549 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 550 551 if (ac->is_arraycopy_validated() || 552 ac->is_copyof_validated() || 553 ac->is_copyofrange_validated()) { 554 Node* ld_addp = in(MemNode::Address); 555 if (ld_addp->is_AddP()) { 556 Node* ld_base = ld_addp->in(AddPNode::Address); 557 Node* ld_offs = ld_addp->in(AddPNode::Offset); 558 559 Node* dest = ac->in(ArrayCopyNode::Dest); 560 561 if (dest == ld_base) { 562 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 563 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 564 return ac; 565 } 566 if (!can_see_stored_value) { 567 mem = ac->in(TypeFunc::Memory); 568 } 569 } 570 } 571 } 572 } 573 return NULL; 574 } 575 576 // The logic for reordering loads and stores uses four steps: 577 // (a) Walk carefully past stores and initializations which we 578 // can prove are independent of this load. 579 // (b) Observe that the next memory state makes an exact match 580 // with self (load or store), and locate the relevant store. 581 // (c) Ensure that, if we were to wire self directly to the store, 582 // the optimizer would fold it up somehow. 583 // (d) Do the rewiring, and return, depending on some other part of 584 // the optimizer to fold up the load. 585 // This routine handles steps (a) and (b). Steps (c) and (d) are 586 // specific to loads and stores, so they are handled by the callers. 587 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 588 // 589 Node* MemNode::find_previous_store(PhaseTransform* phase) { 590 Node* ctrl = in(MemNode::Control); 591 Node* adr = in(MemNode::Address); 592 intptr_t offset = 0; 593 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 594 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 595 596 if (offset == Type::OffsetBot) 597 return NULL; // cannot unalias unless there are precise offsets 598 599 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 600 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 601 602 intptr_t size_in_bytes = memory_size(); 603 604 Node* mem = in(MemNode::Memory); // start searching here... 605 606 int cnt = 50; // Cycle limiter 607 for (;;) { // While we can dance past unrelated stores... 608 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 609 610 Node* prev = mem; 611 if (mem->is_Store()) { 612 Node* st_adr = mem->in(MemNode::Address); 613 intptr_t st_offset = 0; 614 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 615 if (st_base == NULL) 616 break; // inscrutable pointer 617 618 // For raw accesses it's not enough to prove that constant offsets don't intersect. 619 // We need the bases to be the equal in order for the offset check to make sense. 620 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 621 break; 622 } 623 624 if (st_offset != offset && st_offset != Type::OffsetBot) { 625 const int MAX_STORE = BytesPerLong; 626 if (st_offset >= offset + size_in_bytes || 627 st_offset <= offset - MAX_STORE || 628 st_offset <= offset - mem->as_Store()->memory_size()) { 629 // Success: The offsets are provably independent. 630 // (You may ask, why not just test st_offset != offset and be done? 631 // The answer is that stores of different sizes can co-exist 632 // in the same sequence of RawMem effects. We sometimes initialize 633 // a whole 'tile' of array elements with a single jint or jlong.) 634 mem = mem->in(MemNode::Memory); 635 continue; // (a) advance through independent store memory 636 } 637 } 638 if (st_base != base && 639 detect_ptr_independence(base, alloc, 640 st_base, 641 AllocateNode::Ideal_allocation(st_base, phase), 642 phase)) { 643 // Success: The bases are provably independent. 644 mem = mem->in(MemNode::Memory); 645 continue; // (a) advance through independent store memory 646 } 647 648 // (b) At this point, if the bases or offsets do not agree, we lose, 649 // since we have not managed to prove 'this' and 'mem' independent. 650 if (st_base == base && st_offset == offset) { 651 return mem; // let caller handle steps (c), (d) 652 } 653 654 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 655 InitializeNode* st_init = mem->in(0)->as_Initialize(); 656 AllocateNode* st_alloc = st_init->allocation(); 657 if (st_alloc == NULL) 658 break; // something degenerated 659 bool known_identical = false; 660 bool known_independent = false; 661 if (alloc == st_alloc) 662 known_identical = true; 663 else if (alloc != NULL) 664 known_independent = true; 665 else if (all_controls_dominate(this, st_alloc)) 666 known_independent = true; 667 668 if (known_independent) { 669 // The bases are provably independent: Either they are 670 // manifestly distinct allocations, or else the control 671 // of this load dominates the store's allocation. 672 int alias_idx = phase->C->get_alias_index(adr_type()); 673 if (alias_idx == Compile::AliasIdxRaw) { 674 mem = st_alloc->in(TypeFunc::Memory); 675 } else { 676 mem = st_init->memory(alias_idx); 677 } 678 continue; // (a) advance through independent store memory 679 } 680 681 // (b) at this point, if we are not looking at a store initializing 682 // the same allocation we are loading from, we lose. 683 if (known_identical) { 684 // From caller, can_see_stored_value will consult find_captured_store. 685 return mem; // let caller handle steps (c), (d) 686 } 687 688 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 689 if (prev != mem) { 690 // Found an arraycopy but it doesn't affect that load 691 continue; 692 } 693 // Found an arraycopy that may affect that load 694 return mem; 695 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 696 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 697 if (mem->is_Proj() && mem->in(0)->is_Call()) { 698 // ArrayCopyNodes processed here as well. 699 CallNode *call = mem->in(0)->as_Call(); 700 if (!call->may_modify(addr_t, phase)) { 701 mem = call->in(TypeFunc::Memory); 702 continue; // (a) advance through independent call memory 703 } 704 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 705 ArrayCopyNode* ac = NULL; 706 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 707 break; 708 } 709 mem = mem->in(0)->in(TypeFunc::Memory); 710 continue; // (a) advance through independent MemBar memory 711 } else if (mem->is_ClearArray()) { 712 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 713 // (the call updated 'mem' value) 714 continue; // (a) advance through independent allocation memory 715 } else { 716 // Can not bypass initialization of the instance 717 // we are looking for. 718 return mem; 719 } 720 } else if (mem->is_MergeMem()) { 721 int alias_idx = phase->C->get_alias_index(adr_type()); 722 mem = mem->as_MergeMem()->memory_at(alias_idx); 723 continue; // (a) advance through independent MergeMem memory 724 } 725 } 726 727 // Unless there is an explicit 'continue', we must bail out here, 728 // because 'mem' is an inscrutable memory state (e.g., a call). 729 break; 730 } 731 732 return NULL; // bail out 733 } 734 735 //----------------------calculate_adr_type------------------------------------- 736 // Helper function. Notices when the given type of address hits top or bottom. 737 // Also, asserts a cross-check of the type against the expected address type. 738 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 739 if (t == Type::TOP) return NULL; // does not touch memory any more? 740 #ifdef ASSERT 741 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL; 742 #endif 743 const TypePtr* tp = t->isa_ptr(); 744 if (tp == NULL) { 745 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 746 return TypePtr::BOTTOM; // touches lots of memory 747 } else { 748 #ifdef ASSERT 749 // %%%% [phh] We don't check the alias index if cross_check is 750 // TypeRawPtr::BOTTOM. Needs to be investigated. 751 if (cross_check != NULL && 752 cross_check != TypePtr::BOTTOM && 753 cross_check != TypeRawPtr::BOTTOM) { 754 // Recheck the alias index, to see if it has changed (due to a bug). 755 Compile* C = Compile::current(); 756 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 757 "must stay in the original alias category"); 758 // The type of the address must be contained in the adr_type, 759 // disregarding "null"-ness. 760 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 761 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 762 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 763 "real address must not escape from expected memory type"); 764 } 765 #endif 766 return tp; 767 } 768 } 769 770 //============================================================================= 771 // Should LoadNode::Ideal() attempt to remove control edges? 772 bool LoadNode::can_remove_control() const { 773 return true; 774 } 775 uint LoadNode::size_of() const { return sizeof(*this); } 776 bool LoadNode::cmp( const Node &n ) const 777 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 778 const Type *LoadNode::bottom_type() const { return _type; } 779 uint LoadNode::ideal_reg() const { 780 return _type->ideal_reg(); 781 } 782 783 #ifndef PRODUCT 784 void LoadNode::dump_spec(outputStream *st) const { 785 MemNode::dump_spec(st); 786 if( !Verbose && !WizardMode ) { 787 // standard dump does this in Verbose and WizardMode 788 st->print(" #"); _type->dump_on(st); 789 } 790 if (!depends_only_on_test()) { 791 st->print(" (does not depend only on test)"); 792 } 793 } 794 #endif 795 796 #ifdef ASSERT 797 //----------------------------is_immutable_value------------------------------- 798 // Helper function to allow a raw load without control edge for some cases 799 bool LoadNode::is_immutable_value(Node* adr) { 800 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 801 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 802 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 803 in_bytes(JavaThread::osthread_offset()))); 804 } 805 #endif 806 807 //----------------------------LoadNode::make----------------------------------- 808 // Polymorphic factory method: 809 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 810 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) { 811 Compile* C = gvn.C; 812 813 // sanity check the alias category against the created node type 814 assert(!(adr_type->isa_oopptr() && 815 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 816 "use LoadKlassNode instead"); 817 assert(!(adr_type->isa_aryptr() && 818 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 819 "use LoadRangeNode instead"); 820 // Check control edge of raw loads 821 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 822 // oop will be recorded in oop map if load crosses safepoint 823 rt->isa_oopptr() || is_immutable_value(adr), 824 "raw memory operations should have control edge"); 825 LoadNode* load = NULL; 826 switch (bt) { 827 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 828 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 829 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 830 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 831 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 832 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 833 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 834 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 835 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 836 case T_VALUETYPE: 837 case T_OBJECT: 838 #ifdef _LP64 839 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 840 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 841 } else 842 #endif 843 { 844 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 845 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 846 } 847 break; 848 default: 849 ShouldNotReachHere(); 850 break; 851 } 852 assert(load != NULL, "LoadNode should have been created"); 853 if (unaligned) { 854 load->set_unaligned_access(); 855 } 856 if (mismatched) { 857 load->set_mismatched_access(); 858 } 859 if (unsafe) { 860 load->set_unsafe_access(); 861 } 862 load->set_barrier_data(barrier_data); 863 if (load->Opcode() == Op_LoadN) { 864 Node* ld = gvn.transform(load); 865 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 866 } 867 868 return load; 869 } 870 871 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 872 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) { 873 bool require_atomic = true; 874 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 875 if (unaligned) { 876 load->set_unaligned_access(); 877 } 878 if (mismatched) { 879 load->set_mismatched_access(); 880 } 881 if (unsafe) { 882 load->set_unsafe_access(); 883 } 884 load->set_barrier_data(barrier_data); 885 return load; 886 } 887 888 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 889 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) { 890 bool require_atomic = true; 891 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 892 if (unaligned) { 893 load->set_unaligned_access(); 894 } 895 if (mismatched) { 896 load->set_mismatched_access(); 897 } 898 if (unsafe) { 899 load->set_unsafe_access(); 900 } 901 load->set_barrier_data(barrier_data); 902 return load; 903 } 904 905 906 907 //------------------------------hash------------------------------------------- 908 uint LoadNode::hash() const { 909 // unroll addition of interesting fields 910 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 911 } 912 913 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 914 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 915 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 916 bool is_stable_ary = FoldStableValues && 917 (tp != NULL) && (tp->isa_aryptr() != NULL) && 918 tp->isa_aryptr()->is_stable(); 919 920 return (eliminate_boxing && non_volatile) || is_stable_ary; 921 } 922 923 return false; 924 } 925 926 // Is the value loaded previously stored by an arraycopy? If so return 927 // a load node that reads from the source array so we may be able to 928 // optimize out the ArrayCopy node later. 929 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 930 Node* ld_adr = in(MemNode::Address); 931 intptr_t ld_off = 0; 932 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 933 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 934 if (ac != NULL) { 935 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 936 937 Node* mem = ac->in(TypeFunc::Memory); 938 Node* ctl = ac->in(0); 939 Node* src = ac->in(ArrayCopyNode::Src); 940 941 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 942 return NULL; 943 } 944 945 LoadNode* ld = clone()->as_Load(); 946 Node* addp = in(MemNode::Address)->clone(); 947 if (ac->as_ArrayCopy()->is_clonebasic()) { 948 assert(ld_alloc != NULL, "need an alloc"); 949 assert(addp->is_AddP(), "address must be addp"); 950 assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address"); 951 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 952 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern"); 953 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern"); 954 addp->set_req(AddPNode::Base, src->in(AddPNode::Base)); 955 addp->set_req(AddPNode::Address, src->in(AddPNode::Address)); 956 } else { 957 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 958 ac->as_ArrayCopy()->is_copyof_validated() || 959 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 960 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 961 addp->set_req(AddPNode::Base, src); 962 addp->set_req(AddPNode::Address, src); 963 964 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 965 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 966 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 967 uint shift = exact_log2(type2aelembytes(ary_elem)); 968 if (ary_t->klass()->is_value_array_klass()) { 969 ciValueArrayKlass* vak = ary_t->klass()->as_value_array_klass(); 970 shift = vak->log2_element_size(); 971 } 972 973 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 974 #ifdef _LP64 975 diff = phase->transform(new ConvI2LNode(diff)); 976 #endif 977 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 978 979 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 980 addp->set_req(AddPNode::Offset, offset); 981 } 982 addp = phase->transform(addp); 983 #ifdef ASSERT 984 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 985 ld->_adr_type = adr_type; 986 #endif 987 ld->set_req(MemNode::Address, addp); 988 ld->set_req(0, ctl); 989 ld->set_req(MemNode::Memory, mem); 990 // load depends on the tests that validate the arraycopy 991 ld->_control_dependency = UnknownControl; 992 return ld; 993 } 994 return NULL; 995 } 996 997 998 //---------------------------can_see_stored_value------------------------------ 999 // This routine exists to make sure this set of tests is done the same 1000 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 1001 // will change the graph shape in a way which makes memory alive twice at the 1002 // same time (uses the Oracle model of aliasing), then some 1003 // LoadXNode::Identity will fold things back to the equivalence-class model 1004 // of aliasing. 1005 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 1006 Node* ld_adr = in(MemNode::Address); 1007 intptr_t ld_off = 0; 1008 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off); 1009 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase); 1010 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 1011 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 1012 // This is more general than load from boxing objects. 1013 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1014 uint alias_idx = atp->index(); 1015 bool final = !atp->is_rewritable(); 1016 Node* result = NULL; 1017 Node* current = st; 1018 // Skip through chains of MemBarNodes checking the MergeMems for 1019 // new states for the slice of this load. Stop once any other 1020 // kind of node is encountered. Loads from final memory can skip 1021 // through any kind of MemBar but normal loads shouldn't skip 1022 // through MemBarAcquire since the could allow them to move out of 1023 // a synchronized region. 1024 while (current->is_Proj()) { 1025 int opc = current->in(0)->Opcode(); 1026 if ((final && (opc == Op_MemBarAcquire || 1027 opc == Op_MemBarAcquireLock || 1028 opc == Op_LoadFence)) || 1029 opc == Op_MemBarRelease || 1030 opc == Op_StoreFence || 1031 opc == Op_MemBarReleaseLock || 1032 opc == Op_MemBarStoreStore || 1033 opc == Op_MemBarCPUOrder) { 1034 Node* mem = current->in(0)->in(TypeFunc::Memory); 1035 if (mem->is_MergeMem()) { 1036 MergeMemNode* merge = mem->as_MergeMem(); 1037 Node* new_st = merge->memory_at(alias_idx); 1038 if (new_st == merge->base_memory()) { 1039 // Keep searching 1040 current = new_st; 1041 continue; 1042 } 1043 // Save the new memory state for the slice and fall through 1044 // to exit. 1045 result = new_st; 1046 } 1047 } 1048 break; 1049 } 1050 if (result != NULL) { 1051 st = result; 1052 } 1053 } 1054 1055 // Loop around twice in the case Load -> Initialize -> Store. 1056 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1057 for (int trip = 0; trip <= 1; trip++) { 1058 1059 if (st->is_Store()) { 1060 Node* st_adr = st->in(MemNode::Address); 1061 if (!phase->eqv(st_adr, ld_adr)) { 1062 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers). 1063 intptr_t st_off = 0; 1064 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off); 1065 if (ld_base == NULL) return NULL; 1066 if (st_base == NULL) return NULL; 1067 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL; 1068 if (ld_off != st_off) return NULL; 1069 if (ld_off == Type::OffsetBot) return NULL; 1070 // Same base, same offset. 1071 // Possible improvement for arrays: check index value instead of absolute offset. 1072 1073 // At this point we have proven something like this setup: 1074 // B = << base >> 1075 // L = LoadQ(AddP(Check/CastPP(B), #Off)) 1076 // S = StoreQ(AddP( B , #Off), V) 1077 // (Actually, we haven't yet proven the Q's are the same.) 1078 // In other words, we are loading from a casted version of 1079 // the same pointer-and-offset that we stored to. 1080 // Casted version may carry a dependency and it is respected. 1081 // Thus, we are able to replace L by V. 1082 } 1083 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1084 if (store_Opcode() != st->Opcode()) 1085 return NULL; 1086 return st->in(MemNode::ValueIn); 1087 } 1088 1089 // A load from a freshly-created object always returns zero. 1090 // (This can happen after LoadNode::Ideal resets the load's memory input 1091 // to find_captured_store, which returned InitializeNode::zero_memory.) 1092 if (st->is_Proj() && st->in(0)->is_Allocate() && 1093 (st->in(0) == ld_alloc) && 1094 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1095 // return a zero value for the load's basic type 1096 // (This is one of the few places where a generic PhaseTransform 1097 // can create new nodes. Think of it as lazily manifesting 1098 // virtually pre-existing constants.) 1099 assert(memory_type() != T_VALUETYPE, "should not be used for value types"); 1100 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue); 1101 if (default_value != NULL) { 1102 return default_value; 1103 } 1104 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 1105 return phase->zerocon(memory_type()); 1106 } 1107 1108 // A load from an initialization barrier can match a captured store. 1109 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1110 InitializeNode* init = st->in(0)->as_Initialize(); 1111 AllocateNode* alloc = init->allocation(); 1112 if ((alloc != NULL) && (alloc == ld_alloc)) { 1113 // examine a captured store value 1114 st = init->find_captured_store(ld_off, memory_size(), phase); 1115 if (st != NULL) { 1116 continue; // take one more trip around 1117 } 1118 } 1119 } 1120 1121 // Load boxed value from result of valueOf() call is input parameter. 1122 if (this->is_Load() && ld_adr->is_AddP() && 1123 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1124 intptr_t ignore = 0; 1125 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1126 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1127 base = bs->step_over_gc_barrier(base); 1128 if (base != NULL && base->is_Proj() && 1129 base->as_Proj()->_con == TypeFunc::Parms && 1130 base->in(0)->is_CallStaticJava() && 1131 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1132 return base->in(0)->in(TypeFunc::Parms); 1133 } 1134 } 1135 1136 break; 1137 } 1138 1139 return NULL; 1140 } 1141 1142 //----------------------is_instance_field_load_with_local_phi------------------ 1143 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1144 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1145 in(Address)->is_AddP() ) { 1146 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1147 // Only instances and boxed values. 1148 if( t_oop != NULL && 1149 (t_oop->is_ptr_to_boxed_value() || 1150 t_oop->is_known_instance_field()) && 1151 t_oop->offset() != Type::OffsetBot && 1152 t_oop->offset() != Type::OffsetTop) { 1153 return true; 1154 } 1155 } 1156 return false; 1157 } 1158 1159 //------------------------------Identity--------------------------------------- 1160 // Loads are identity if previous store is to same address 1161 Node* LoadNode::Identity(PhaseGVN* phase) { 1162 // Loading from a ValueTypePtr? The ValueTypePtr has the values of 1163 // all fields as input. Look for the field with matching offset. 1164 Node* addr = in(Address); 1165 intptr_t offset; 1166 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset); 1167 if (base != NULL && base->is_ValueTypePtr() && offset > oopDesc::klass_offset_in_bytes()) { 1168 Node* value = base->as_ValueTypePtr()->field_value_by_offset((int)offset, true); 1169 if (value->is_ValueType()) { 1170 // Non-flattened value type field 1171 ValueTypeNode* vt = value->as_ValueType(); 1172 if (vt->is_allocated(phase)) { 1173 value = vt->get_oop(); 1174 } else { 1175 // Not yet allocated, bail out 1176 value = NULL; 1177 } 1178 } 1179 if (value != NULL) { 1180 if (Opcode() == Op_LoadN) { 1181 // Encode oop value if we are loading a narrow oop 1182 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded"); 1183 value = phase->transform(new EncodePNode(value, bottom_type())); 1184 } 1185 return value; 1186 } 1187 } 1188 1189 // If the previous store-maker is the right kind of Store, and the store is 1190 // to the same address, then we are equal to the value stored. 1191 Node* mem = in(Memory); 1192 Node* value = can_see_stored_value(mem, phase); 1193 if( value ) { 1194 // byte, short & char stores truncate naturally. 1195 // A load has to load the truncated value which requires 1196 // some sort of masking operation and that requires an 1197 // Ideal call instead of an Identity call. 1198 if (memory_size() < BytesPerInt) { 1199 // If the input to the store does not fit with the load's result type, 1200 // it must be truncated via an Ideal call. 1201 if (!phase->type(value)->higher_equal(phase->type(this))) 1202 return this; 1203 } 1204 // (This works even when value is a Con, but LoadNode::Value 1205 // usually runs first, producing the singleton type of the Con.) 1206 return value; 1207 } 1208 1209 // Search for an existing data phi which was generated before for the same 1210 // instance's field to avoid infinite generation of phis in a loop. 1211 Node *region = mem->in(0); 1212 if (is_instance_field_load_with_local_phi(region)) { 1213 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1214 int this_index = phase->C->get_alias_index(addr_t); 1215 int this_offset = addr_t->offset(); 1216 int this_iid = addr_t->instance_id(); 1217 if (!addr_t->is_known_instance() && 1218 addr_t->is_ptr_to_boxed_value()) { 1219 // Use _idx of address base (could be Phi node) for boxed values. 1220 intptr_t ignore = 0; 1221 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1222 if (base == NULL) { 1223 return this; 1224 } 1225 this_iid = base->_idx; 1226 } 1227 const Type* this_type = bottom_type(); 1228 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1229 Node* phi = region->fast_out(i); 1230 if (phi->is_Phi() && phi != mem && 1231 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1232 return phi; 1233 } 1234 } 1235 } 1236 1237 return this; 1238 } 1239 1240 // Construct an equivalent unsigned load. 1241 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1242 BasicType bt = T_ILLEGAL; 1243 const Type* rt = NULL; 1244 switch (Opcode()) { 1245 case Op_LoadUB: return this; 1246 case Op_LoadUS: return this; 1247 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1248 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1249 default: 1250 assert(false, "no unsigned variant: %s", Name()); 1251 return NULL; 1252 } 1253 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1254 raw_adr_type(), rt, bt, _mo, _control_dependency, 1255 is_unaligned_access(), is_mismatched_access()); 1256 } 1257 1258 // Construct an equivalent signed load. 1259 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1260 BasicType bt = T_ILLEGAL; 1261 const Type* rt = NULL; 1262 switch (Opcode()) { 1263 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1264 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1265 case Op_LoadB: // fall through 1266 case Op_LoadS: // fall through 1267 case Op_LoadI: // fall through 1268 case Op_LoadL: return this; 1269 default: 1270 assert(false, "no signed variant: %s", Name()); 1271 return NULL; 1272 } 1273 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1274 raw_adr_type(), rt, bt, _mo, _control_dependency, 1275 is_unaligned_access(), is_mismatched_access()); 1276 } 1277 1278 // We're loading from an object which has autobox behaviour. 1279 // If this object is result of a valueOf call we'll have a phi 1280 // merging a newly allocated object and a load from the cache. 1281 // We want to replace this load with the original incoming 1282 // argument to the valueOf call. 1283 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1284 assert(phase->C->eliminate_boxing(), "sanity"); 1285 intptr_t ignore = 0; 1286 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1287 if ((base == NULL) || base->is_Phi()) { 1288 // Push the loads from the phi that comes from valueOf up 1289 // through it to allow elimination of the loads and the recovery 1290 // of the original value. It is done in split_through_phi(). 1291 return NULL; 1292 } else if (base->is_Load() || 1293 (base->is_DecodeN() && base->in(1)->is_Load())) { 1294 // Eliminate the load of boxed value for integer types from the cache 1295 // array by deriving the value from the index into the array. 1296 // Capture the offset of the load and then reverse the computation. 1297 1298 // Get LoadN node which loads a boxing object from 'cache' array. 1299 if (base->is_DecodeN()) { 1300 base = base->in(1); 1301 } 1302 if (!base->in(Address)->is_AddP()) { 1303 return NULL; // Complex address 1304 } 1305 AddPNode* address = base->in(Address)->as_AddP(); 1306 Node* cache_base = address->in(AddPNode::Base); 1307 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1308 // Get ConP node which is static 'cache' field. 1309 cache_base = cache_base->in(1); 1310 } 1311 if ((cache_base != NULL) && cache_base->is_Con()) { 1312 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1313 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1314 Node* elements[4]; 1315 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1316 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1317 if (count > 0 && elements[0]->is_Con() && 1318 (count == 1 || 1319 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1320 elements[1]->in(2) == phase->intcon(shift)))) { 1321 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1322 // Fetch the box object cache[0] at the base of the array and get its value 1323 ciInstance* box = array->obj_at(0)->as_instance(); 1324 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1325 assert(ik->is_box_klass(), "sanity"); 1326 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1327 if (ik->nof_nonstatic_fields() == 1) { 1328 // This should be true nonstatic_field_at requires calling 1329 // nof_nonstatic_fields so check it anyway 1330 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1331 BasicType bt = c.basic_type(); 1332 // Only integer types have boxing cache. 1333 assert(bt == T_BOOLEAN || bt == T_CHAR || 1334 bt == T_BYTE || bt == T_SHORT || 1335 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1336 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1337 if (cache_low != (int)cache_low) { 1338 return NULL; // should not happen since cache is array indexed by value 1339 } 1340 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1341 if (offset != (int)offset) { 1342 return NULL; // should not happen since cache is array indexed by value 1343 } 1344 // Add up all the offsets making of the address of the load 1345 Node* result = elements[0]; 1346 for (int i = 1; i < count; i++) { 1347 result = phase->transform(new AddXNode(result, elements[i])); 1348 } 1349 // Remove the constant offset from the address and then 1350 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1351 // remove the scaling of the offset to recover the original index. 1352 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1353 // Peel the shift off directly but wrap it in a dummy node 1354 // since Ideal can't return existing nodes 1355 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1356 } else if (result->is_Add() && result->in(2)->is_Con() && 1357 result->in(1)->Opcode() == Op_LShiftX && 1358 result->in(1)->in(2) == phase->intcon(shift)) { 1359 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1360 // but for boxing cache access we know that X<<Z will not overflow 1361 // (there is range check) so we do this optimizatrion by hand here. 1362 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1363 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1364 } else { 1365 result = new RShiftXNode(result, phase->intcon(shift)); 1366 } 1367 #ifdef _LP64 1368 if (bt != T_LONG) { 1369 result = new ConvL2INode(phase->transform(result)); 1370 } 1371 #else 1372 if (bt == T_LONG) { 1373 result = new ConvI2LNode(phase->transform(result)); 1374 } 1375 #endif 1376 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1377 // Need to preserve unboxing load type if it is unsigned. 1378 switch(this->Opcode()) { 1379 case Op_LoadUB: 1380 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1381 break; 1382 case Op_LoadUS: 1383 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1384 break; 1385 } 1386 return result; 1387 } 1388 } 1389 } 1390 } 1391 } 1392 return NULL; 1393 } 1394 1395 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1396 Node* region = phi->in(0); 1397 if (region == NULL) { 1398 return false; // Wait stable graph 1399 } 1400 uint cnt = phi->req(); 1401 for (uint i = 1; i < cnt; i++) { 1402 Node* rc = region->in(i); 1403 if (rc == NULL || phase->type(rc) == Type::TOP) 1404 return false; // Wait stable graph 1405 Node* in = phi->in(i); 1406 if (in == NULL || phase->type(in) == Type::TOP) 1407 return false; // Wait stable graph 1408 } 1409 return true; 1410 } 1411 //------------------------------split_through_phi------------------------------ 1412 // Split instance or boxed field load through Phi. 1413 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1414 Node* mem = in(Memory); 1415 Node* address = in(Address); 1416 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1417 1418 assert((t_oop != NULL) && 1419 (t_oop->is_known_instance_field() || 1420 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1421 1422 Compile* C = phase->C; 1423 intptr_t ignore = 0; 1424 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1425 bool base_is_phi = (base != NULL) && base->is_Phi(); 1426 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1427 (base != NULL) && (base == address->in(AddPNode::Base)) && 1428 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1429 1430 if (!((mem->is_Phi() || base_is_phi) && 1431 (load_boxed_values || t_oop->is_known_instance_field()))) { 1432 return NULL; // memory is not Phi 1433 } 1434 1435 if (mem->is_Phi()) { 1436 if (!stable_phi(mem->as_Phi(), phase)) { 1437 return NULL; // Wait stable graph 1438 } 1439 uint cnt = mem->req(); 1440 // Check for loop invariant memory. 1441 if (cnt == 3) { 1442 for (uint i = 1; i < cnt; i++) { 1443 Node* in = mem->in(i); 1444 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1445 if (m == mem) { 1446 if (i == 1) { 1447 // if the first edge was a loop, check second edge too. 1448 // If both are replaceable - we are in an infinite loop 1449 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1450 if (n == mem) { 1451 break; 1452 } 1453 } 1454 set_req(Memory, mem->in(cnt - i)); 1455 return this; // made change 1456 } 1457 } 1458 } 1459 } 1460 if (base_is_phi) { 1461 if (!stable_phi(base->as_Phi(), phase)) { 1462 return NULL; // Wait stable graph 1463 } 1464 uint cnt = base->req(); 1465 // Check for loop invariant memory. 1466 if (cnt == 3) { 1467 for (uint i = 1; i < cnt; i++) { 1468 if (base->in(i) == base) { 1469 return NULL; // Wait stable graph 1470 } 1471 } 1472 } 1473 } 1474 1475 // Split through Phi (see original code in loopopts.cpp). 1476 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1477 1478 // Do nothing here if Identity will find a value 1479 // (to avoid infinite chain of value phis generation). 1480 if (!phase->eqv(this, phase->apply_identity(this))) 1481 return NULL; 1482 1483 // Select Region to split through. 1484 Node* region; 1485 if (!base_is_phi) { 1486 assert(mem->is_Phi(), "sanity"); 1487 region = mem->in(0); 1488 // Skip if the region dominates some control edge of the address. 1489 if (!MemNode::all_controls_dominate(address, region)) 1490 return NULL; 1491 } else if (!mem->is_Phi()) { 1492 assert(base_is_phi, "sanity"); 1493 region = base->in(0); 1494 // Skip if the region dominates some control edge of the memory. 1495 if (!MemNode::all_controls_dominate(mem, region)) 1496 return NULL; 1497 } else if (base->in(0) != mem->in(0)) { 1498 assert(base_is_phi && mem->is_Phi(), "sanity"); 1499 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1500 region = base->in(0); 1501 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1502 region = mem->in(0); 1503 } else { 1504 return NULL; // complex graph 1505 } 1506 } else { 1507 assert(base->in(0) == mem->in(0), "sanity"); 1508 region = mem->in(0); 1509 } 1510 1511 const Type* this_type = this->bottom_type(); 1512 int this_index = C->get_alias_index(t_oop); 1513 int this_offset = t_oop->offset(); 1514 int this_iid = t_oop->instance_id(); 1515 if (!t_oop->is_known_instance() && load_boxed_values) { 1516 // Use _idx of address base for boxed values. 1517 this_iid = base->_idx; 1518 } 1519 PhaseIterGVN* igvn = phase->is_IterGVN(); 1520 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1521 for (uint i = 1; i < region->req(); i++) { 1522 Node* x; 1523 Node* the_clone = NULL; 1524 if (region->in(i) == C->top()) { 1525 x = C->top(); // Dead path? Use a dead data op 1526 } else { 1527 x = this->clone(); // Else clone up the data op 1528 the_clone = x; // Remember for possible deletion. 1529 // Alter data node to use pre-phi inputs 1530 if (this->in(0) == region) { 1531 x->set_req(0, region->in(i)); 1532 } else { 1533 x->set_req(0, NULL); 1534 } 1535 if (mem->is_Phi() && (mem->in(0) == region)) { 1536 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1537 } 1538 if (address->is_Phi() && address->in(0) == region) { 1539 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1540 } 1541 if (base_is_phi && (base->in(0) == region)) { 1542 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1543 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1544 x->set_req(Address, adr_x); 1545 } 1546 } 1547 // Check for a 'win' on some paths 1548 const Type *t = x->Value(igvn); 1549 1550 bool singleton = t->singleton(); 1551 1552 // See comments in PhaseIdealLoop::split_thru_phi(). 1553 if (singleton && t == Type::TOP) { 1554 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1555 } 1556 1557 if (singleton) { 1558 x = igvn->makecon(t); 1559 } else { 1560 // We now call Identity to try to simplify the cloned node. 1561 // Note that some Identity methods call phase->type(this). 1562 // Make sure that the type array is big enough for 1563 // our new node, even though we may throw the node away. 1564 // (This tweaking with igvn only works because x is a new node.) 1565 igvn->set_type(x, t); 1566 // If x is a TypeNode, capture any more-precise type permanently into Node 1567 // otherwise it will be not updated during igvn->transform since 1568 // igvn->type(x) is set to x->Value() already. 1569 x->raise_bottom_type(t); 1570 Node *y = igvn->apply_identity(x); 1571 if (y != x) { 1572 x = y; 1573 } else { 1574 y = igvn->hash_find_insert(x); 1575 if (y) { 1576 x = y; 1577 } else { 1578 // Else x is a new node we are keeping 1579 // We do not need register_new_node_with_optimizer 1580 // because set_type has already been called. 1581 igvn->_worklist.push(x); 1582 } 1583 } 1584 } 1585 if (x != the_clone && the_clone != NULL) { 1586 igvn->remove_dead_node(the_clone); 1587 } 1588 phi->set_req(i, x); 1589 } 1590 // Record Phi 1591 igvn->register_new_node_with_optimizer(phi); 1592 return phi; 1593 } 1594 1595 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const { 1596 if (Opcode() == Op_LoadX) { 1597 Node* address = in(MemNode::Address); 1598 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase); 1599 Node* mem = in(MemNode::Memory); 1600 if (alloc != NULL && mem->is_Proj() && 1601 mem->in(0) != NULL && 1602 mem->in(0) == alloc->initialization() && 1603 alloc->initialization()->proj_out_or_null(0) != NULL) { 1604 return alloc; 1605 } 1606 } 1607 return NULL; 1608 } 1609 1610 1611 //------------------------------Ideal------------------------------------------ 1612 // If the load is from Field memory and the pointer is non-null, it might be possible to 1613 // zero out the control input. 1614 // If the offset is constant and the base is an object allocation, 1615 // try to hook me up to the exact initializing store. 1616 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1617 Node* p = MemNode::Ideal_common(phase, can_reshape); 1618 if (p) return (p == NodeSentinel) ? NULL : p; 1619 1620 Node* ctrl = in(MemNode::Control); 1621 Node* address = in(MemNode::Address); 1622 bool progress = false; 1623 1624 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1625 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1626 1627 // Skip up past a SafePoint control. Cannot do this for Stores because 1628 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1629 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1630 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1631 !addr_mark && 1632 (depends_only_on_test() || has_unknown_control_dependency())) { 1633 ctrl = ctrl->in(0); 1634 set_req(MemNode::Control,ctrl); 1635 progress = true; 1636 } 1637 1638 intptr_t ignore = 0; 1639 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1640 if (base != NULL 1641 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1642 // Check for useless control edge in some common special cases 1643 if (in(MemNode::Control) != NULL 1644 && can_remove_control() 1645 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1646 && all_controls_dominate(base, phase->C->start())) { 1647 // A method-invariant, non-null address (constant or 'this' argument). 1648 set_req(MemNode::Control, NULL); 1649 progress = true; 1650 } 1651 } 1652 1653 Node* mem = in(MemNode::Memory); 1654 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1655 1656 if (can_reshape && (addr_t != NULL)) { 1657 // try to optimize our memory input 1658 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1659 if (opt_mem != mem) { 1660 set_req(MemNode::Memory, opt_mem); 1661 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1662 return this; 1663 } 1664 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1665 if ((t_oop != NULL) && 1666 (t_oop->is_known_instance_field() || 1667 t_oop->is_ptr_to_boxed_value())) { 1668 PhaseIterGVN *igvn = phase->is_IterGVN(); 1669 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1670 // Delay this transformation until memory Phi is processed. 1671 phase->is_IterGVN()->_worklist.push(this); 1672 return NULL; 1673 } 1674 // Split instance field load through Phi. 1675 Node* result = split_through_phi(phase); 1676 if (result != NULL) return result; 1677 1678 if (t_oop->is_ptr_to_boxed_value()) { 1679 Node* result = eliminate_autobox(phase); 1680 if (result != NULL) return result; 1681 } 1682 } 1683 } 1684 1685 // Is there a dominating load that loads the same value? Leave 1686 // anything that is not a load of a field/array element (like 1687 // barriers etc.) alone 1688 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) { 1689 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1690 Node *use = mem->fast_out(i); 1691 if (use != this && 1692 use->Opcode() == Opcode() && 1693 use->in(0) != NULL && 1694 use->in(0) != in(0) && 1695 use->in(Address) == in(Address)) { 1696 Node* ctl = in(0); 1697 for (int i = 0; i < 10 && ctl != NULL; i++) { 1698 ctl = IfNode::up_one_dom(ctl); 1699 if (ctl == use->in(0)) { 1700 set_req(0, use->in(0)); 1701 return this; 1702 } 1703 } 1704 } 1705 } 1706 } 1707 1708 // Check for prior store with a different base or offset; make Load 1709 // independent. Skip through any number of them. Bail out if the stores 1710 // are in an endless dead cycle and report no progress. This is a key 1711 // transform for Reflection. However, if after skipping through the Stores 1712 // we can't then fold up against a prior store do NOT do the transform as 1713 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1714 // array memory alive twice: once for the hoisted Load and again after the 1715 // bypassed Store. This situation only works if EVERYBODY who does 1716 // anti-dependence work knows how to bypass. I.e. we need all 1717 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1718 // the alias index stuff. So instead, peek through Stores and IFF we can 1719 // fold up, do so. 1720 Node* prev_mem = find_previous_store(phase); 1721 if (prev_mem != NULL) { 1722 Node* value = can_see_arraycopy_value(prev_mem, phase); 1723 if (value != NULL) { 1724 return value; 1725 } 1726 } 1727 // Steps (a), (b): Walk past independent stores to find an exact match. 1728 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1729 // (c) See if we can fold up on the spot, but don't fold up here. 1730 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1731 // just return a prior value, which is done by Identity calls. 1732 if (can_see_stored_value(prev_mem, phase)) { 1733 // Make ready for step (d): 1734 set_req(MemNode::Memory, prev_mem); 1735 return this; 1736 } 1737 } 1738 1739 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase); 1740 if (alloc != NULL && mem->is_Proj() && 1741 mem->in(0) != NULL && 1742 mem->in(0) == alloc->initialization() && 1743 Opcode() == Op_LoadX && 1744 alloc->initialization()->proj_out_or_null(0) != NULL) { 1745 InitializeNode* init = alloc->initialization(); 1746 Node* control = init->proj_out(0); 1747 return alloc->make_ideal_mark(phase, address, control, mem); 1748 } 1749 1750 return progress ? this : NULL; 1751 } 1752 1753 // Helper to recognize certain Klass fields which are invariant across 1754 // some group of array types (e.g., int[] or all T[] where T < Object). 1755 const Type* 1756 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1757 ciKlass* klass) const { 1758 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1759 // The field is Klass::_modifier_flags. Return its (constant) value. 1760 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1761 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1762 return TypeInt::make(klass->modifier_flags()); 1763 } 1764 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1765 // The field is Klass::_access_flags. Return its (constant) value. 1766 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1767 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1768 return TypeInt::make(klass->access_flags()); 1769 } 1770 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1771 // The field is Klass::_layout_helper. Return its constant value if known. 1772 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1773 return TypeInt::make(klass->layout_helper()); 1774 } 1775 1776 // No match. 1777 return NULL; 1778 } 1779 1780 //------------------------------Value----------------------------------------- 1781 const Type* LoadNode::Value(PhaseGVN* phase) const { 1782 // Either input is TOP ==> the result is TOP 1783 Node* mem = in(MemNode::Memory); 1784 const Type *t1 = phase->type(mem); 1785 if (t1 == Type::TOP) return Type::TOP; 1786 Node* adr = in(MemNode::Address); 1787 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1788 if (tp == NULL || tp->empty()) return Type::TOP; 1789 int off = tp->offset(); 1790 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1791 Compile* C = phase->C; 1792 1793 // Try to guess loaded type from pointer type 1794 if (tp->isa_aryptr()) { 1795 const TypeAryPtr* ary = tp->is_aryptr(); 1796 const Type* t = ary->elem(); 1797 1798 // Determine whether the reference is beyond the header or not, by comparing 1799 // the offset against the offset of the start of the array's data. 1800 // Different array types begin at slightly different offsets (12 vs. 16). 1801 // We choose T_BYTE as an example base type that is least restrictive 1802 // as to alignment, which will therefore produce the smallest 1803 // possible base offset. 1804 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1805 const bool off_beyond_header = (off >= min_base_off); 1806 1807 // Try to constant-fold a stable array element. 1808 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1809 // Make sure the reference is not into the header and the offset is constant 1810 ciObject* aobj = ary->const_oop(); 1811 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1812 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1813 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1814 stable_dimension, 1815 memory_type(), is_unsigned()); 1816 if (con_type != NULL) { 1817 return con_type; 1818 } 1819 } 1820 } 1821 1822 // Don't do this for integer types. There is only potential profit if 1823 // the element type t is lower than _type; that is, for int types, if _type is 1824 // more restrictive than t. This only happens here if one is short and the other 1825 // char (both 16 bits), and in those cases we've made an intentional decision 1826 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1827 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1828 // 1829 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1830 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1831 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1832 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1833 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1834 // In fact, that could have been the original type of p1, and p1 could have 1835 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1836 // expression (LShiftL quux 3) independently optimized to the constant 8. 1837 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1838 && (_type->isa_vect() == NULL) 1839 && t->isa_valuetype() == NULL 1840 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1841 // t might actually be lower than _type, if _type is a unique 1842 // concrete subclass of abstract class t. 1843 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header? 1844 const Type* jt = t->join_speculative(_type); 1845 // In any case, do not allow the join, per se, to empty out the type. 1846 if (jt->empty() && !t->empty()) { 1847 // This can happen if a interface-typed array narrows to a class type. 1848 jt = _type; 1849 } 1850 #ifdef ASSERT 1851 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1852 // The pointers in the autobox arrays are always non-null 1853 Node* base = adr->in(AddPNode::Base); 1854 if ((base != NULL) && base->is_DecodeN()) { 1855 // Get LoadN node which loads IntegerCache.cache field 1856 base = base->in(1); 1857 } 1858 if ((base != NULL) && base->is_Con()) { 1859 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1860 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1861 // It could be narrow oop 1862 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1863 } 1864 } 1865 } 1866 #endif 1867 return jt; 1868 } 1869 } 1870 } else if (tp->base() == Type::InstPtr) { 1871 assert( off != Type::OffsetBot || 1872 // arrays can be cast to Objects 1873 tp->is_oopptr()->klass()->is_java_lang_Object() || 1874 tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() || 1875 // unsafe field access may not have a constant offset 1876 C->has_unsafe_access(), 1877 "Field accesses must be precise" ); 1878 // For oop loads, we expect the _type to be precise. 1879 1880 const TypeInstPtr* tinst = tp->is_instptr(); 1881 BasicType bt = memory_type(); 1882 1883 // Fold component and value mirror loads 1884 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1885 if (ik == phase->C->env()->Class_klass() && (off == java_lang_Class::component_mirror_offset_in_bytes() || 1886 off == java_lang_Class::inline_mirror_offset_in_bytes())) { 1887 ciType* mirror_type = tinst->java_mirror_type(); 1888 if (mirror_type != NULL) { 1889 const Type* const_oop = TypePtr::NULL_PTR; 1890 if (mirror_type->is_array_klass()) { 1891 const_oop = TypeInstPtr::make(mirror_type->as_array_klass()->component_mirror_instance()); 1892 } else if (mirror_type->is_valuetype()) { 1893 const_oop = TypeInstPtr::make(mirror_type->as_value_klass()->inline_mirror_instance()); 1894 } 1895 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop; 1896 } 1897 } 1898 1899 // Optimize loads from constant fields. 1900 ciObject* const_oop = tinst->const_oop(); 1901 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1902 ciType* mirror_type = const_oop->as_instance()->java_mirror_type(); 1903 if (mirror_type != NULL && mirror_type->is_valuetype()) { 1904 ciValueKlass* vk = mirror_type->as_value_klass(); 1905 if (off == vk->default_value_offset()) { 1906 // Loading a special hidden field that contains the oop of the default value type 1907 const Type* const_oop = TypeInstPtr::make(vk->default_value_instance()); 1908 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop; 1909 } 1910 } 1911 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt); 1912 if (con_type != NULL) { 1913 return con_type; 1914 } 1915 } 1916 } else if (tp->base() == Type::KlassPtr) { 1917 assert( off != Type::OffsetBot || 1918 // arrays can be cast to Objects 1919 tp->is_klassptr()->klass() == NULL || 1920 tp->is_klassptr()->klass()->is_java_lang_Object() || 1921 // also allow array-loading from the primary supertype 1922 // array during subtype checks 1923 Opcode() == Op_LoadKlass, 1924 "Field accesses must be precise" ); 1925 // For klass/static loads, we expect the _type to be precise 1926 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) { 1927 if (adr->is_Load() && off == 0) { 1928 /* With mirrors being an indirect in the Klass* 1929 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 1930 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 1931 * 1932 * So check the type and klass of the node before the LoadP. 1933 */ 1934 Node* adr2 = adr->in(MemNode::Address); 1935 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1936 if (tkls != NULL) { 1937 ciKlass* klass = tkls->klass(); 1938 if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1939 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1940 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1941 return TypeInstPtr::make(klass->java_mirror()); 1942 } 1943 } 1944 } else { 1945 // Check for a load of the default value offset from the ValueKlassFixedBlock: 1946 // LoadI(LoadP(value_klass, adr_valueklass_fixed_block_offset), default_value_offset_offset) 1947 intptr_t offset = 0; 1948 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1949 if (base != NULL && base->is_Load() && offset == in_bytes(ValueKlass::default_value_offset_offset())) { 1950 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr(); 1951 if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_valuetype() && 1952 tkls->offset() == in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())) { 1953 assert(base->Opcode() == Op_LoadP, "must load an oop from klass"); 1954 assert(Opcode() == Op_LoadI, "must load an int from fixed block"); 1955 return TypeInt::make(tkls->klass()->as_value_klass()->default_value_offset()); 1956 } 1957 } 1958 } 1959 } 1960 1961 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1962 if (tkls != NULL && !StressReflectiveCode) { 1963 ciKlass* klass = tkls->klass(); 1964 if (tkls->is_loaded() && tkls->klass_is_exact()) { 1965 // We are loading a field from a Klass metaobject whose identity 1966 // is known at compile time (the type is "exact" or "precise"). 1967 // Check for fields we know are maintained as constants by the VM. 1968 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1969 // The field is Klass::_super_check_offset. Return its (constant) value. 1970 // (Folds up type checking code.) 1971 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1972 return TypeInt::make(klass->super_check_offset()); 1973 } 1974 // Compute index into primary_supers array 1975 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1976 // Check for overflowing; use unsigned compare to handle the negative case. 1977 if( depth < ciKlass::primary_super_limit() ) { 1978 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1979 // (Folds up type checking code.) 1980 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1981 ciKlass *ss = klass->super_of_depth(depth); 1982 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1983 } 1984 const Type* aift = load_array_final_field(tkls, klass); 1985 if (aift != NULL) return aift; 1986 } 1987 1988 // We can still check if we are loading from the primary_supers array at a 1989 // shallow enough depth. Even though the klass is not exact, entries less 1990 // than or equal to its super depth are correct. 1991 if (tkls->is_loaded()) { 1992 ciType *inner = klass; 1993 while( inner->is_obj_array_klass() ) 1994 inner = inner->as_obj_array_klass()->base_element_type(); 1995 if( inner->is_instance_klass() && 1996 !inner->as_instance_klass()->flags().is_interface() ) { 1997 // Compute index into primary_supers array 1998 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1999 // Check for overflowing; use unsigned compare to handle the negative case. 2000 if( depth < ciKlass::primary_super_limit() && 2001 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 2002 // The field is an element of Klass::_primary_supers. Return its (constant) value. 2003 // (Folds up type checking code.) 2004 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 2005 ciKlass *ss = klass->super_of_depth(depth); 2006 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 2007 } 2008 } 2009 } 2010 2011 // If the type is enough to determine that the thing is not an array, 2012 // we can give the layout_helper a positive interval type. 2013 // This will help short-circuit some reflective code. 2014 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 2015 && !klass->is_array_klass() // not directly typed as an array 2016 && !klass->is_interface() // specifically not Serializable & Cloneable 2017 && !klass->is_java_lang_Object() // not the supertype of all T[] 2018 ) { 2019 // Note: When interfaces are reliable, we can narrow the interface 2020 // test to (klass != Serializable && klass != Cloneable). 2021 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 2022 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 2023 // The key property of this type is that it folds up tests 2024 // for array-ness, since it proves that the layout_helper is positive. 2025 // Thus, a generic value like the basic object layout helper works fine. 2026 return TypeInt::make(min_size, max_jint, Type::WidenMin); 2027 } 2028 } 2029 2030 // If we are loading from a freshly-allocated object, produce a zero, 2031 // if the load is provably beyond the header of the object. 2032 // (Also allow a variable load from a fresh array to produce zero.) 2033 const TypeOopPtr *tinst = tp->isa_oopptr(); 2034 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 2035 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 2036 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 2037 Node* value = can_see_stored_value(mem,phase); 2038 if (value != NULL && value->is_Con()) { 2039 assert(value->bottom_type()->higher_equal(_type),"sanity"); 2040 return value->bottom_type(); 2041 } 2042 } 2043 2044 if (is_instance) { 2045 // If we have an instance type and our memory input is the 2046 // programs's initial memory state, there is no matching store, 2047 // so just return a zero of the appropriate type 2048 Node *mem = in(MemNode::Memory); 2049 if (mem->is_Parm() && mem->in(0)->is_Start()) { 2050 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 2051 return Type::get_zero_type(_type->basic_type()); 2052 } 2053 } 2054 2055 Node* alloc = is_new_object_mark_load(phase); 2056 if (alloc != NULL && !(alloc->Opcode() == Op_Allocate && UseBiasedLocking)) { 2057 return TypeX::make(markWord::prototype().value()); 2058 } 2059 2060 return _type; 2061 } 2062 2063 //------------------------------match_edge------------------------------------- 2064 // Do we Match on this edge index or not? Match only the address. 2065 uint LoadNode::match_edge(uint idx) const { 2066 return idx == MemNode::Address; 2067 } 2068 2069 //--------------------------LoadBNode::Ideal-------------------------------------- 2070 // 2071 // If the previous store is to the same address as this load, 2072 // and the value stored was larger than a byte, replace this load 2073 // with the value stored truncated to a byte. If no truncation is 2074 // needed, the replacement is done in LoadNode::Identity(). 2075 // 2076 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2077 Node* mem = in(MemNode::Memory); 2078 Node* value = can_see_stored_value(mem,phase); 2079 if( value && !phase->type(value)->higher_equal( _type ) ) { 2080 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 2081 return new RShiftINode(result, phase->intcon(24)); 2082 } 2083 // Identity call will handle the case where truncation is not needed. 2084 return LoadNode::Ideal(phase, can_reshape); 2085 } 2086 2087 const Type* LoadBNode::Value(PhaseGVN* phase) const { 2088 Node* mem = in(MemNode::Memory); 2089 Node* value = can_see_stored_value(mem,phase); 2090 if (value != NULL && value->is_Con() && 2091 !value->bottom_type()->higher_equal(_type)) { 2092 // If the input to the store does not fit with the load's result type, 2093 // it must be truncated. We can't delay until Ideal call since 2094 // a singleton Value is needed for split_thru_phi optimization. 2095 int con = value->get_int(); 2096 return TypeInt::make((con << 24) >> 24); 2097 } 2098 return LoadNode::Value(phase); 2099 } 2100 2101 //--------------------------LoadUBNode::Ideal------------------------------------- 2102 // 2103 // If the previous store is to the same address as this load, 2104 // and the value stored was larger than a byte, replace this load 2105 // with the value stored truncated to a byte. If no truncation is 2106 // needed, the replacement is done in LoadNode::Identity(). 2107 // 2108 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2109 Node* mem = in(MemNode::Memory); 2110 Node* value = can_see_stored_value(mem, phase); 2111 if (value && !phase->type(value)->higher_equal(_type)) 2112 return new AndINode(value, phase->intcon(0xFF)); 2113 // Identity call will handle the case where truncation is not needed. 2114 return LoadNode::Ideal(phase, can_reshape); 2115 } 2116 2117 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2118 Node* mem = in(MemNode::Memory); 2119 Node* value = can_see_stored_value(mem,phase); 2120 if (value != NULL && value->is_Con() && 2121 !value->bottom_type()->higher_equal(_type)) { 2122 // If the input to the store does not fit with the load's result type, 2123 // it must be truncated. We can't delay until Ideal call since 2124 // a singleton Value is needed for split_thru_phi optimization. 2125 int con = value->get_int(); 2126 return TypeInt::make(con & 0xFF); 2127 } 2128 return LoadNode::Value(phase); 2129 } 2130 2131 //--------------------------LoadUSNode::Ideal------------------------------------- 2132 // 2133 // If the previous store is to the same address as this load, 2134 // and the value stored was larger than a char, replace this load 2135 // with the value stored truncated to a char. If no truncation is 2136 // needed, the replacement is done in LoadNode::Identity(). 2137 // 2138 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2139 Node* mem = in(MemNode::Memory); 2140 Node* value = can_see_stored_value(mem,phase); 2141 if( value && !phase->type(value)->higher_equal( _type ) ) 2142 return new AndINode(value,phase->intcon(0xFFFF)); 2143 // Identity call will handle the case where truncation is not needed. 2144 return LoadNode::Ideal(phase, can_reshape); 2145 } 2146 2147 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2148 Node* mem = in(MemNode::Memory); 2149 Node* value = can_see_stored_value(mem,phase); 2150 if (value != NULL && value->is_Con() && 2151 !value->bottom_type()->higher_equal(_type)) { 2152 // If the input to the store does not fit with the load's result type, 2153 // it must be truncated. We can't delay until Ideal call since 2154 // a singleton Value is needed for split_thru_phi optimization. 2155 int con = value->get_int(); 2156 return TypeInt::make(con & 0xFFFF); 2157 } 2158 return LoadNode::Value(phase); 2159 } 2160 2161 //--------------------------LoadSNode::Ideal-------------------------------------- 2162 // 2163 // If the previous store is to the same address as this load, 2164 // and the value stored was larger than a short, replace this load 2165 // with the value stored truncated to a short. If no truncation is 2166 // needed, the replacement is done in LoadNode::Identity(). 2167 // 2168 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2169 Node* mem = in(MemNode::Memory); 2170 Node* value = can_see_stored_value(mem,phase); 2171 if( value && !phase->type(value)->higher_equal( _type ) ) { 2172 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 2173 return new RShiftINode(result, phase->intcon(16)); 2174 } 2175 // Identity call will handle the case where truncation is not needed. 2176 return LoadNode::Ideal(phase, can_reshape); 2177 } 2178 2179 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2180 Node* mem = in(MemNode::Memory); 2181 Node* value = can_see_stored_value(mem,phase); 2182 if (value != NULL && value->is_Con() && 2183 !value->bottom_type()->higher_equal(_type)) { 2184 // If the input to the store does not fit with the load's result type, 2185 // it must be truncated. We can't delay until Ideal call since 2186 // a singleton Value is needed for split_thru_phi optimization. 2187 int con = value->get_int(); 2188 return TypeInt::make((con << 16) >> 16); 2189 } 2190 return LoadNode::Value(phase); 2191 } 2192 2193 //============================================================================= 2194 //----------------------------LoadKlassNode::make------------------------------ 2195 // Polymorphic factory method: 2196 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, 2197 const TypeKlassPtr* tk, bool clear_prop_bits) { 2198 // sanity check the alias category against the created node type 2199 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2200 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2201 #ifdef _LP64 2202 if (adr_type->is_ptr_to_narrowklass()) { 2203 assert(UseCompressedClassPointers, "no compressed klasses"); 2204 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered, clear_prop_bits)); 2205 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2206 } 2207 #endif 2208 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2209 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered, clear_prop_bits); 2210 } 2211 2212 //------------------------------Value------------------------------------------ 2213 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2214 return klass_value_common(phase); 2215 } 2216 2217 // In most cases, LoadKlassNode does not have the control input set. If the control 2218 // input is set, it must not be removed (by LoadNode::Ideal()). 2219 bool LoadKlassNode::can_remove_control() const { 2220 return false; 2221 } 2222 2223 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2224 // Either input is TOP ==> the result is TOP 2225 const Type *t1 = phase->type( in(MemNode::Memory) ); 2226 if (t1 == Type::TOP) return Type::TOP; 2227 Node *adr = in(MemNode::Address); 2228 const Type *t2 = phase->type( adr ); 2229 if (t2 == Type::TOP) return Type::TOP; 2230 const TypePtr *tp = t2->is_ptr(); 2231 if (TypePtr::above_centerline(tp->ptr()) || 2232 tp->ptr() == TypePtr::Null) return Type::TOP; 2233 2234 // Return a more precise klass, if possible 2235 const TypeInstPtr *tinst = tp->isa_instptr(); 2236 if (tinst != NULL) { 2237 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2238 int offset = tinst->offset(); 2239 if (ik == phase->C->env()->Class_klass() 2240 && (offset == java_lang_Class::klass_offset_in_bytes() || 2241 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2242 // We are loading a special hidden field from a Class mirror object, 2243 // the field which points to the VM's Klass metaobject. 2244 bool is_indirect_type = true; 2245 ciType* t = tinst->java_mirror_type(&is_indirect_type); 2246 // java_mirror_type returns non-null for compile-time Class constants. 2247 if (t != NULL) { 2248 // constant oop => constant klass 2249 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2250 if (t->is_void()) { 2251 // We cannot create a void array. Since void is a primitive type return null 2252 // klass. Users of this result need to do a null check on the returned klass. 2253 return TypePtr::NULL_PTR; 2254 } 2255 return TypeKlassPtr::make(ciArrayKlass::make(t, /* never_null */ !is_indirect_type)); 2256 } 2257 if (!t->is_klass()) { 2258 // a primitive Class (e.g., int.class) has NULL for a klass field 2259 return TypePtr::NULL_PTR; 2260 } 2261 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2262 return TypeKlassPtr::make(t->as_klass()); 2263 } 2264 // non-constant mirror, so we can't tell what's going on 2265 } 2266 if( !ik->is_loaded() ) 2267 return _type; // Bail out if not loaded 2268 if (offset == oopDesc::klass_offset_in_bytes()) { 2269 if (tinst->klass_is_exact()) { 2270 return TypeKlassPtr::make(ik); 2271 } 2272 // See if we can become precise: no subklasses and no interface 2273 // (Note: We need to support verified interfaces.) 2274 if (!ik->is_interface() && !ik->has_subklass()) { 2275 //assert(!UseExactTypes, "this code should be useless with exact types"); 2276 // Add a dependence; if any subclass added we need to recompile 2277 if (!ik->is_final()) { 2278 // %%% should use stronger assert_unique_concrete_subtype instead 2279 phase->C->dependencies()->assert_leaf_type(ik); 2280 } 2281 // Return precise klass 2282 return TypeKlassPtr::make(ik); 2283 } 2284 2285 // Return root of possible klass 2286 return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0), tinst->flat_array()); 2287 } 2288 } 2289 2290 // Check for loading klass from an array 2291 const TypeAryPtr *tary = tp->isa_aryptr(); 2292 if (tary != NULL) { 2293 ciKlass *tary_klass = tary->klass(); 2294 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2295 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2296 ciArrayKlass* ak = tary_klass->as_array_klass(); 2297 // Do not fold klass loads from [V?. The runtime type might be [V due to [V <: [V? 2298 // and the klass for [V is not equal to the klass for [V?. 2299 if (tary->klass_is_exact()) { 2300 return TypeKlassPtr::make(tary_klass); 2301 } 2302 2303 // If the klass is an object array, we defer the question to the 2304 // array component klass. 2305 if (ak->is_obj_array_klass()) { 2306 assert(ak->is_loaded(), ""); 2307 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2308 if (base_k->is_loaded() && base_k->is_instance_klass()) { 2309 ciInstanceKlass *ik = base_k->as_instance_klass(); 2310 // See if we can become precise: no subklasses and no interface 2311 if (!ik->is_interface() && !ik->has_subklass() && (!ik->is_valuetype() || ak->storage_properties().is_null_free())) { 2312 //assert(!UseExactTypes, "this code should be useless with exact types"); 2313 // Add a dependence; if any subclass added we need to recompile 2314 if (!ik->is_final()) { 2315 phase->C->dependencies()->assert_leaf_type(ik); 2316 } 2317 // Return precise array klass 2318 return TypeKlassPtr::make(ak); 2319 } 2320 } 2321 return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0), false); 2322 } else if (ak->is_type_array_klass()) { 2323 //assert(!UseExactTypes, "this code should be useless with exact types"); 2324 return TypeKlassPtr::make(ak); // These are always precise 2325 } 2326 } 2327 } 2328 2329 // Check for loading klass from an array klass 2330 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2331 if (tkls != NULL && !StressReflectiveCode) { 2332 if (!tkls->is_loaded()) { 2333 return _type; // Bail out if not loaded 2334 } 2335 ciKlass* klass = tkls->klass(); 2336 if( klass->is_obj_array_klass() && 2337 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2338 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2339 // // Always returning precise element type is incorrect, 2340 // // e.g., element type could be object and array may contain strings 2341 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2342 2343 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2344 // according to the element type's subclassing. 2345 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0), elem->flatten_array()); 2346 } else if (klass->is_value_array_klass() && 2347 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2348 ciKlass* elem = klass->as_value_array_klass()->element_klass(); 2349 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0), true); 2350 } 2351 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2352 tkls->offset() == in_bytes(Klass::super_offset())) { 2353 ciKlass* sup = klass->as_instance_klass()->super(); 2354 // The field is Klass::_super. Return its (constant) value. 2355 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2356 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2357 } 2358 } 2359 2360 // Bailout case 2361 return LoadNode::Value(phase); 2362 } 2363 2364 //------------------------------Identity--------------------------------------- 2365 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2366 // Also feed through the klass in Allocate(...klass...)._klass. 2367 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2368 return klass_identity_common(phase); 2369 } 2370 2371 const Type* GetStoragePropertyNode::Value(PhaseGVN* phase) const { 2372 if (in(1) != NULL) { 2373 const Type* in1_t = phase->type(in(1)); 2374 if (in1_t == Type::TOP) { 2375 return Type::TOP; 2376 } 2377 const TypeKlassPtr* tk = in1_t->make_ptr()->is_klassptr(); 2378 ciArrayKlass* ak = tk->klass()->as_array_klass(); 2379 ciKlass* elem = ak->element_klass(); 2380 if (tk->klass_is_exact() || (!elem->is_java_lang_Object() && !elem->is_interface() && !elem->is_valuetype())) { 2381 int props_shift = in1_t->isa_narrowklass() ? oopDesc::narrow_storage_props_shift : oopDesc::wide_storage_props_shift; 2382 ArrayStorageProperties props = ak->storage_properties(); 2383 intptr_t storage_properties = 0; 2384 if ((Opcode() == Op_GetFlattenedProperty && props.is_flattened()) || 2385 (Opcode() == Op_GetNullFreeProperty && props.is_null_free())) { 2386 storage_properties = 1; 2387 } 2388 if (in1_t->isa_narrowklass()) { 2389 return TypeInt::make((int)storage_properties); 2390 } 2391 return TypeX::make(storage_properties); 2392 } 2393 } 2394 return bottom_type(); 2395 } 2396 2397 Node* GetStoragePropertyNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2398 if (!can_reshape) { 2399 return NULL; 2400 } 2401 if (in(1) != NULL && in(1)->is_Phi()) { 2402 Node* phi = in(1); 2403 Node* r = phi->in(0); 2404 Node* new_phi = new PhiNode(r, bottom_type()); 2405 for (uint i = 1; i < r->req(); i++) { 2406 Node* in = phi->in(i); 2407 if (in == NULL) continue; 2408 Node* n = clone(); 2409 n->set_req(1, in); 2410 new_phi->init_req(i, phase->transform(n)); 2411 } 2412 return new_phi; 2413 } 2414 return NULL; 2415 } 2416 2417 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2418 Node* x = LoadNode::Identity(phase); 2419 if (x != this) return x; 2420 2421 // Take apart the address into an oop and and offset. 2422 // Return 'this' if we cannot. 2423 Node* adr = in(MemNode::Address); 2424 intptr_t offset = 0; 2425 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2426 if (base == NULL) return this; 2427 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2428 if (toop == NULL) return this; 2429 2430 // Step over potential GC barrier for OopHandle resolve 2431 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2432 if (bs->is_gc_barrier_node(base)) { 2433 base = bs->step_over_gc_barrier(base); 2434 } 2435 2436 // We can fetch the klass directly through an AllocateNode. 2437 // This works even if the klass is not constant (clone or newArray). 2438 if (offset == oopDesc::klass_offset_in_bytes()) { 2439 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2440 if (allocated_klass != NULL) { 2441 return allocated_klass; 2442 } 2443 } 2444 2445 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2446 // See inline_native_Class_query for occurrences of these patterns. 2447 // Java Example: x.getClass().isAssignableFrom(y) 2448 // 2449 // This improves reflective code, often making the Class 2450 // mirror go completely dead. (Current exception: Class 2451 // mirrors may appear in debug info, but we could clean them out by 2452 // introducing a new debug info operator for Klass.java_mirror). 2453 2454 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2455 && offset == java_lang_Class::klass_offset_in_bytes()) { 2456 if (base->is_Load()) { 2457 Node* base2 = base->in(MemNode::Address); 2458 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2459 Node* adr2 = base2->in(MemNode::Address); 2460 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2461 if (tkls != NULL && !tkls->empty() 2462 && (tkls->klass()->is_instance_klass() || 2463 tkls->klass()->is_array_klass()) 2464 && adr2->is_AddP() 2465 ) { 2466 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2467 if (tkls->offset() == mirror_field) { 2468 return adr2->in(AddPNode::Base); 2469 } 2470 } 2471 } 2472 } 2473 } 2474 2475 return this; 2476 } 2477 2478 2479 //------------------------------Value------------------------------------------ 2480 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2481 const Type *t = klass_value_common(phase); 2482 if (t == Type::TOP) 2483 return t; 2484 2485 return t->make_narrowklass(); 2486 } 2487 2488 //------------------------------Identity--------------------------------------- 2489 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2490 // Also feed through the klass in Allocate(...klass...)._klass. 2491 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2492 Node *x = klass_identity_common(phase); 2493 2494 const Type *t = phase->type( x ); 2495 if( t == Type::TOP ) return x; 2496 if( t->isa_narrowklass()) return x; 2497 assert (!t->isa_narrowoop(), "no narrow oop here"); 2498 2499 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2500 } 2501 2502 //------------------------------Value----------------------------------------- 2503 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2504 // Either input is TOP ==> the result is TOP 2505 const Type *t1 = phase->type( in(MemNode::Memory) ); 2506 if( t1 == Type::TOP ) return Type::TOP; 2507 Node *adr = in(MemNode::Address); 2508 const Type *t2 = phase->type( adr ); 2509 if( t2 == Type::TOP ) return Type::TOP; 2510 const TypePtr *tp = t2->is_ptr(); 2511 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2512 const TypeAryPtr *tap = tp->isa_aryptr(); 2513 if( !tap ) return _type; 2514 return tap->size(); 2515 } 2516 2517 //-------------------------------Ideal--------------------------------------- 2518 // Feed through the length in AllocateArray(...length...)._length. 2519 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2520 Node* p = MemNode::Ideal_common(phase, can_reshape); 2521 if (p) return (p == NodeSentinel) ? NULL : p; 2522 2523 // Take apart the address into an oop and and offset. 2524 // Return 'this' if we cannot. 2525 Node* adr = in(MemNode::Address); 2526 intptr_t offset = 0; 2527 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2528 if (base == NULL) return NULL; 2529 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2530 if (tary == NULL) return NULL; 2531 2532 // We can fetch the length directly through an AllocateArrayNode. 2533 // This works even if the length is not constant (clone or newArray). 2534 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2535 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2536 if (alloc != NULL) { 2537 Node* allocated_length = alloc->Ideal_length(); 2538 Node* len = alloc->make_ideal_length(tary, phase); 2539 if (allocated_length != len) { 2540 // New CastII improves on this. 2541 return len; 2542 } 2543 } 2544 } 2545 2546 return NULL; 2547 } 2548 2549 //------------------------------Identity--------------------------------------- 2550 // Feed through the length in AllocateArray(...length...)._length. 2551 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2552 Node* x = LoadINode::Identity(phase); 2553 if (x != this) return x; 2554 2555 // Take apart the address into an oop and and offset. 2556 // Return 'this' if we cannot. 2557 Node* adr = in(MemNode::Address); 2558 intptr_t offset = 0; 2559 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2560 if (base == NULL) return this; 2561 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2562 if (tary == NULL) return this; 2563 2564 // We can fetch the length directly through an AllocateArrayNode. 2565 // This works even if the length is not constant (clone or newArray). 2566 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2567 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2568 if (alloc != NULL) { 2569 Node* allocated_length = alloc->Ideal_length(); 2570 // Do not allow make_ideal_length to allocate a CastII node. 2571 Node* len = alloc->make_ideal_length(tary, phase, false); 2572 if (allocated_length == len) { 2573 // Return allocated_length only if it would not be improved by a CastII. 2574 return allocated_length; 2575 } 2576 } 2577 } 2578 2579 return this; 2580 2581 } 2582 2583 //============================================================================= 2584 //---------------------------StoreNode::make----------------------------------- 2585 // Polymorphic factory method: 2586 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2587 assert((mo == unordered || mo == release), "unexpected"); 2588 Compile* C = gvn.C; 2589 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2590 ctl != NULL, "raw memory operations should have control edge"); 2591 2592 switch (bt) { 2593 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2594 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2595 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2596 case T_CHAR: 2597 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2598 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2599 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2600 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2601 case T_METADATA: 2602 case T_ADDRESS: 2603 case T_VALUETYPE: 2604 case T_OBJECT: 2605 #ifdef _LP64 2606 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2607 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2608 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2609 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2610 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2611 adr->bottom_type()->isa_rawptr())) { 2612 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2613 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2614 } 2615 #endif 2616 { 2617 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2618 } 2619 default: 2620 ShouldNotReachHere(); 2621 return (StoreNode*)NULL; 2622 } 2623 } 2624 2625 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2626 bool require_atomic = true; 2627 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2628 } 2629 2630 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2631 bool require_atomic = true; 2632 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2633 } 2634 2635 2636 //--------------------------bottom_type---------------------------------------- 2637 const Type *StoreNode::bottom_type() const { 2638 return Type::MEMORY; 2639 } 2640 2641 //------------------------------hash------------------------------------------- 2642 uint StoreNode::hash() const { 2643 // unroll addition of interesting fields 2644 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2645 2646 // Since they are not commoned, do not hash them: 2647 return NO_HASH; 2648 } 2649 2650 //------------------------------Ideal------------------------------------------ 2651 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2652 // When a store immediately follows a relevant allocation/initialization, 2653 // try to capture it into the initialization, or hoist it above. 2654 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2655 Node* p = MemNode::Ideal_common(phase, can_reshape); 2656 if (p) return (p == NodeSentinel) ? NULL : p; 2657 2658 Node* mem = in(MemNode::Memory); 2659 Node* address = in(MemNode::Address); 2660 // Back-to-back stores to same address? Fold em up. Generally 2661 // unsafe if I have intervening uses... Also disallowed for StoreCM 2662 // since they must follow each StoreP operation. Redundant StoreCMs 2663 // are eliminated just before matching in final_graph_reshape. 2664 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::VALUES) { 2665 Node* st = mem; 2666 // If Store 'st' has more than one use, we cannot fold 'st' away. 2667 // For example, 'st' might be the final state at a conditional 2668 // return. Or, 'st' might be used by some node which is live at 2669 // the same time 'st' is live, which might be unschedulable. So, 2670 // require exactly ONE user until such time as we clone 'mem' for 2671 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2672 // true). 2673 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2674 // Looking at a dead closed cycle of memory? 2675 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2676 assert(Opcode() == st->Opcode() || 2677 st->Opcode() == Op_StoreVector || 2678 Opcode() == Op_StoreVector || 2679 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2680 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2681 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2682 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) || 2683 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2684 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2685 2686 if (st->in(MemNode::Address)->eqv_uncast(address) && 2687 st->as_Store()->memory_size() <= this->memory_size()) { 2688 Node* use = st->raw_out(0); 2689 phase->igvn_rehash_node_delayed(use); 2690 if (can_reshape) { 2691 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2692 } else { 2693 // It's OK to do this in the parser, since DU info is always accurate, 2694 // and the parser always refers to nodes via SafePointNode maps. 2695 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2696 } 2697 return this; 2698 } 2699 st = st->in(MemNode::Memory); 2700 } 2701 } 2702 2703 2704 // Capture an unaliased, unconditional, simple store into an initializer. 2705 // Or, if it is independent of the allocation, hoist it above the allocation. 2706 if (ReduceFieldZeroing && /*can_reshape &&*/ 2707 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2708 InitializeNode* init = mem->in(0)->as_Initialize(); 2709 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2710 if (offset > 0) { 2711 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2712 // If the InitializeNode captured me, it made a raw copy of me, 2713 // and I need to disappear. 2714 if (moved != NULL) { 2715 // %%% hack to ensure that Ideal returns a new node: 2716 mem = MergeMemNode::make(mem); 2717 return mem; // fold me away 2718 } 2719 } 2720 } 2721 2722 return NULL; // No further progress 2723 } 2724 2725 //------------------------------Value----------------------------------------- 2726 const Type* StoreNode::Value(PhaseGVN* phase) const { 2727 // Either input is TOP ==> the result is TOP 2728 const Type *t1 = phase->type( in(MemNode::Memory) ); 2729 if( t1 == Type::TOP ) return Type::TOP; 2730 const Type *t2 = phase->type( in(MemNode::Address) ); 2731 if( t2 == Type::TOP ) return Type::TOP; 2732 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2733 if( t3 == Type::TOP ) return Type::TOP; 2734 return Type::MEMORY; 2735 } 2736 2737 //------------------------------Identity--------------------------------------- 2738 // Remove redundant stores: 2739 // Store(m, p, Load(m, p)) changes to m. 2740 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2741 Node* StoreNode::Identity(PhaseGVN* phase) { 2742 Node* mem = in(MemNode::Memory); 2743 Node* adr = in(MemNode::Address); 2744 Node* val = in(MemNode::ValueIn); 2745 2746 Node* result = this; 2747 2748 // Load then Store? Then the Store is useless 2749 if (val->is_Load() && 2750 val->in(MemNode::Address)->eqv_uncast(adr) && 2751 val->in(MemNode::Memory )->eqv_uncast(mem) && 2752 val->as_Load()->store_Opcode() == Opcode()) { 2753 result = mem; 2754 } 2755 2756 // Two stores in a row of the same value? 2757 if (result == this && 2758 mem->is_Store() && 2759 mem->in(MemNode::Address)->eqv_uncast(adr) && 2760 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2761 mem->Opcode() == Opcode()) { 2762 result = mem; 2763 } 2764 2765 // Store of zero anywhere into a freshly-allocated object? 2766 // Then the store is useless. 2767 // (It must already have been captured by the InitializeNode.) 2768 if (result == this && ReduceFieldZeroing) { 2769 // a newly allocated object is already all-zeroes everywhere 2770 if (mem->is_Proj() && mem->in(0)->is_Allocate() && 2771 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) { 2772 assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to value array is forbidden"); 2773 result = mem; 2774 } 2775 2776 if (result == this) { 2777 // the store may also apply to zero-bits in an earlier object 2778 Node* prev_mem = find_previous_store(phase); 2779 // Steps (a), (b): Walk past independent stores to find an exact match. 2780 if (prev_mem != NULL) { 2781 Node* prev_val = can_see_stored_value(prev_mem, phase); 2782 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2783 // prev_val and val might differ by a cast; it would be good 2784 // to keep the more informative of the two. 2785 if (phase->type(val)->is_zero_type()) { 2786 result = mem; 2787 } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) { 2788 InitializeNode* init = prev_mem->in(0)->as_Initialize(); 2789 AllocateNode* alloc = init->allocation(); 2790 if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) { 2791 result = mem; 2792 } 2793 } 2794 } 2795 } 2796 } 2797 } 2798 2799 if (result != this && phase->is_IterGVN() != NULL) { 2800 MemBarNode* trailing = trailing_membar(); 2801 if (trailing != NULL) { 2802 #ifdef ASSERT 2803 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2804 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects"); 2805 #endif 2806 PhaseIterGVN* igvn = phase->is_IterGVN(); 2807 trailing->remove(igvn); 2808 } 2809 } 2810 2811 return result; 2812 } 2813 2814 //------------------------------match_edge------------------------------------- 2815 // Do we Match on this edge index or not? Match only memory & value 2816 uint StoreNode::match_edge(uint idx) const { 2817 return idx == MemNode::Address || idx == MemNode::ValueIn; 2818 } 2819 2820 //------------------------------cmp-------------------------------------------- 2821 // Do not common stores up together. They generally have to be split 2822 // back up anyways, so do not bother. 2823 bool StoreNode::cmp( const Node &n ) const { 2824 return (&n == this); // Always fail except on self 2825 } 2826 2827 //------------------------------Ideal_masked_input----------------------------- 2828 // Check for a useless mask before a partial-word store 2829 // (StoreB ... (AndI valIn conIa) ) 2830 // If (conIa & mask == mask) this simplifies to 2831 // (StoreB ... (valIn) ) 2832 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2833 Node *val = in(MemNode::ValueIn); 2834 if( val->Opcode() == Op_AndI ) { 2835 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2836 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2837 set_req(MemNode::ValueIn, val->in(1)); 2838 return this; 2839 } 2840 } 2841 return NULL; 2842 } 2843 2844 2845 //------------------------------Ideal_sign_extended_input---------------------- 2846 // Check for useless sign-extension before a partial-word store 2847 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2848 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2849 // (StoreB ... (valIn) ) 2850 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2851 Node *val = in(MemNode::ValueIn); 2852 if( val->Opcode() == Op_RShiftI ) { 2853 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2854 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2855 Node *shl = val->in(1); 2856 if( shl->Opcode() == Op_LShiftI ) { 2857 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2858 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2859 set_req(MemNode::ValueIn, shl->in(1)); 2860 return this; 2861 } 2862 } 2863 } 2864 } 2865 return NULL; 2866 } 2867 2868 //------------------------------value_never_loaded----------------------------------- 2869 // Determine whether there are any possible loads of the value stored. 2870 // For simplicity, we actually check if there are any loads from the 2871 // address stored to, not just for loads of the value stored by this node. 2872 // 2873 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2874 Node *adr = in(Address); 2875 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2876 if (adr_oop == NULL) 2877 return false; 2878 if (!adr_oop->is_known_instance_field()) 2879 return false; // if not a distinct instance, there may be aliases of the address 2880 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2881 Node *use = adr->fast_out(i); 2882 if (use->is_Load() || use->is_LoadStore()) { 2883 return false; 2884 } 2885 } 2886 return true; 2887 } 2888 2889 MemBarNode* StoreNode::trailing_membar() const { 2890 if (is_release()) { 2891 MemBarNode* trailing_mb = NULL; 2892 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2893 Node* u = fast_out(i); 2894 if (u->is_MemBar()) { 2895 if (u->as_MemBar()->trailing_store()) { 2896 assert(u->Opcode() == Op_MemBarVolatile, ""); 2897 assert(trailing_mb == NULL, "only one"); 2898 trailing_mb = u->as_MemBar(); 2899 #ifdef ASSERT 2900 Node* leading = u->as_MemBar()->leading_membar(); 2901 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2902 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2903 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2904 #endif 2905 } else { 2906 assert(u->as_MemBar()->standalone(), ""); 2907 } 2908 } 2909 } 2910 return trailing_mb; 2911 } 2912 return NULL; 2913 } 2914 2915 2916 //============================================================================= 2917 //------------------------------Ideal------------------------------------------ 2918 // If the store is from an AND mask that leaves the low bits untouched, then 2919 // we can skip the AND operation. If the store is from a sign-extension 2920 // (a left shift, then right shift) we can skip both. 2921 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2922 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2923 if( progress != NULL ) return progress; 2924 2925 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2926 if( progress != NULL ) return progress; 2927 2928 // Finally check the default case 2929 return StoreNode::Ideal(phase, can_reshape); 2930 } 2931 2932 //============================================================================= 2933 //------------------------------Ideal------------------------------------------ 2934 // If the store is from an AND mask that leaves the low bits untouched, then 2935 // we can skip the AND operation 2936 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2937 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2938 if( progress != NULL ) return progress; 2939 2940 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2941 if( progress != NULL ) return progress; 2942 2943 // Finally check the default case 2944 return StoreNode::Ideal(phase, can_reshape); 2945 } 2946 2947 //============================================================================= 2948 //------------------------------Identity--------------------------------------- 2949 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2950 // No need to card mark when storing a null ptr 2951 Node* my_store = in(MemNode::OopStore); 2952 if (my_store->is_Store()) { 2953 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2954 if( t1 == TypePtr::NULL_PTR ) { 2955 return in(MemNode::Memory); 2956 } 2957 } 2958 return this; 2959 } 2960 2961 //============================================================================= 2962 //------------------------------Ideal--------------------------------------- 2963 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2964 Node* progress = StoreNode::Ideal(phase, can_reshape); 2965 if (progress != NULL) return progress; 2966 2967 Node* my_store = in(MemNode::OopStore); 2968 if (my_store->is_MergeMem()) { 2969 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2970 set_req(MemNode::OopStore, mem); 2971 return this; 2972 } 2973 2974 return NULL; 2975 } 2976 2977 //------------------------------Value----------------------------------------- 2978 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2979 // Either input is TOP ==> the result is TOP 2980 const Type *t = phase->type( in(MemNode::Memory) ); 2981 if( t == Type::TOP ) return Type::TOP; 2982 t = phase->type( in(MemNode::Address) ); 2983 if( t == Type::TOP ) return Type::TOP; 2984 t = phase->type( in(MemNode::ValueIn) ); 2985 if( t == Type::TOP ) return Type::TOP; 2986 // If extra input is TOP ==> the result is TOP 2987 t = phase->type( in(MemNode::OopStore) ); 2988 if( t == Type::TOP ) return Type::TOP; 2989 2990 return StoreNode::Value( phase ); 2991 } 2992 2993 2994 //============================================================================= 2995 //----------------------------------SCMemProjNode------------------------------ 2996 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2997 { 2998 return bottom_type(); 2999 } 3000 3001 //============================================================================= 3002 //----------------------------------LoadStoreNode------------------------------ 3003 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 3004 : Node(required), 3005 _type(rt), 3006 _adr_type(at), 3007 _barrier(0) 3008 { 3009 init_req(MemNode::Control, c ); 3010 init_req(MemNode::Memory , mem); 3011 init_req(MemNode::Address, adr); 3012 init_req(MemNode::ValueIn, val); 3013 init_class_id(Class_LoadStore); 3014 } 3015 3016 uint LoadStoreNode::ideal_reg() const { 3017 return _type->ideal_reg(); 3018 } 3019 3020 bool LoadStoreNode::result_not_used() const { 3021 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 3022 Node *x = fast_out(i); 3023 if (x->Opcode() == Op_SCMemProj) continue; 3024 return false; 3025 } 3026 return true; 3027 } 3028 3029 MemBarNode* LoadStoreNode::trailing_membar() const { 3030 MemBarNode* trailing = NULL; 3031 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3032 Node* u = fast_out(i); 3033 if (u->is_MemBar()) { 3034 if (u->as_MemBar()->trailing_load_store()) { 3035 assert(u->Opcode() == Op_MemBarAcquire, ""); 3036 assert(trailing == NULL, "only one"); 3037 trailing = u->as_MemBar(); 3038 #ifdef ASSERT 3039 Node* leading = trailing->leading_membar(); 3040 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 3041 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 3042 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 3043 #endif 3044 } else { 3045 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 3046 } 3047 } 3048 } 3049 3050 return trailing; 3051 } 3052 3053 uint LoadStoreNode::size_of() const { return sizeof(*this); } 3054 3055 //============================================================================= 3056 //----------------------------------LoadStoreConditionalNode-------------------- 3057 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 3058 init_req(ExpectedIn, ex ); 3059 } 3060 3061 //============================================================================= 3062 //-------------------------------adr_type-------------------------------------- 3063 const TypePtr* ClearArrayNode::adr_type() const { 3064 Node *adr = in(3); 3065 if (adr == NULL) return NULL; // node is dead 3066 return MemNode::calculate_adr_type(adr->bottom_type()); 3067 } 3068 3069 //------------------------------match_edge------------------------------------- 3070 // Do we Match on this edge index or not? Do not match memory 3071 uint ClearArrayNode::match_edge(uint idx) const { 3072 return idx > 1; 3073 } 3074 3075 //------------------------------Identity--------------------------------------- 3076 // Clearing a zero length array does nothing 3077 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 3078 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 3079 } 3080 3081 //------------------------------Idealize--------------------------------------- 3082 // Clearing a short array is faster with stores 3083 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3084 // Already know this is a large node, do not try to ideal it 3085 if (!IdealizeClearArrayNode || _is_large) return NULL; 3086 3087 const int unit = BytesPerLong; 3088 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 3089 if (!t) return NULL; 3090 if (!t->is_con()) return NULL; 3091 intptr_t raw_count = t->get_con(); 3092 intptr_t size = raw_count; 3093 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 3094 // Clearing nothing uses the Identity call. 3095 // Negative clears are possible on dead ClearArrays 3096 // (see jck test stmt114.stmt11402.val). 3097 if (size <= 0 || size % unit != 0) return NULL; 3098 intptr_t count = size / unit; 3099 // Length too long; communicate this to matchers and assemblers. 3100 // Assemblers are responsible to produce fast hardware clears for it. 3101 if (size > InitArrayShortSize) { 3102 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true); 3103 } 3104 Node *mem = in(1); 3105 if( phase->type(mem)==Type::TOP ) return NULL; 3106 Node *adr = in(3); 3107 const Type* at = phase->type(adr); 3108 if( at==Type::TOP ) return NULL; 3109 const TypePtr* atp = at->isa_ptr(); 3110 // adjust atp to be the correct array element address type 3111 if (atp == NULL) atp = TypePtr::BOTTOM; 3112 else atp = atp->add_offset(Type::OffsetBot); 3113 // Get base for derived pointer purposes 3114 if( adr->Opcode() != Op_AddP ) Unimplemented(); 3115 Node *base = adr->in(1); 3116 3117 Node *val = in(4); 3118 Node *off = phase->MakeConX(BytesPerLong); 3119 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3120 count--; 3121 while( count-- ) { 3122 mem = phase->transform(mem); 3123 adr = phase->transform(new AddPNode(base,adr,off)); 3124 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3125 } 3126 return mem; 3127 } 3128 3129 //----------------------------step_through---------------------------------- 3130 // Return allocation input memory edge if it is different instance 3131 // or itself if it is the one we are looking for. 3132 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 3133 Node* n = *np; 3134 assert(n->is_ClearArray(), "sanity"); 3135 intptr_t offset; 3136 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 3137 // This method is called only before Allocate nodes are expanded 3138 // during macro nodes expansion. Before that ClearArray nodes are 3139 // only generated in PhaseMacroExpand::generate_arraycopy() (before 3140 // Allocate nodes are expanded) which follows allocations. 3141 assert(alloc != NULL, "should have allocation"); 3142 if (alloc->_idx == instance_id) { 3143 // Can not bypass initialization of the instance we are looking for. 3144 return false; 3145 } 3146 // Otherwise skip it. 3147 InitializeNode* init = alloc->initialization(); 3148 if (init != NULL) 3149 *np = init->in(TypeFunc::Memory); 3150 else 3151 *np = alloc->in(TypeFunc::Memory); 3152 return true; 3153 } 3154 3155 //----------------------------clear_memory------------------------------------- 3156 // Generate code to initialize object storage to zero. 3157 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3158 Node* val, 3159 Node* raw_val, 3160 intptr_t start_offset, 3161 Node* end_offset, 3162 PhaseGVN* phase) { 3163 intptr_t offset = start_offset; 3164 3165 int unit = BytesPerLong; 3166 if ((offset % unit) != 0) { 3167 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 3168 adr = phase->transform(adr); 3169 const TypePtr* atp = TypeRawPtr::BOTTOM; 3170 if (val != NULL) { 3171 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3172 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3173 } else { 3174 assert(raw_val == NULL, "val may not be null"); 3175 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3176 } 3177 mem = phase->transform(mem); 3178 offset += BytesPerInt; 3179 } 3180 assert((offset % unit) == 0, ""); 3181 3182 // Initialize the remaining stuff, if any, with a ClearArray. 3183 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase); 3184 } 3185 3186 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3187 Node* raw_val, 3188 Node* start_offset, 3189 Node* end_offset, 3190 PhaseGVN* phase) { 3191 if (start_offset == end_offset) { 3192 // nothing to do 3193 return mem; 3194 } 3195 3196 int unit = BytesPerLong; 3197 Node* zbase = start_offset; 3198 Node* zend = end_offset; 3199 3200 // Scale to the unit required by the CPU: 3201 if (!Matcher::init_array_count_is_in_bytes) { 3202 Node* shift = phase->intcon(exact_log2(unit)); 3203 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3204 zend = phase->transform(new URShiftXNode(zend, shift) ); 3205 } 3206 3207 // Bulk clear double-words 3208 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3209 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3210 if (raw_val == NULL) { 3211 raw_val = phase->MakeConX(0); 3212 } 3213 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false); 3214 return phase->transform(mem); 3215 } 3216 3217 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3218 Node* val, 3219 Node* raw_val, 3220 intptr_t start_offset, 3221 intptr_t end_offset, 3222 PhaseGVN* phase) { 3223 if (start_offset == end_offset) { 3224 // nothing to do 3225 return mem; 3226 } 3227 3228 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3229 intptr_t done_offset = end_offset; 3230 if ((done_offset % BytesPerLong) != 0) { 3231 done_offset -= BytesPerInt; 3232 } 3233 if (done_offset > start_offset) { 3234 mem = clear_memory(ctl, mem, dest, val, raw_val, 3235 start_offset, phase->MakeConX(done_offset), phase); 3236 } 3237 if (done_offset < end_offset) { // emit the final 32-bit store 3238 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3239 adr = phase->transform(adr); 3240 const TypePtr* atp = TypeRawPtr::BOTTOM; 3241 if (val != NULL) { 3242 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3243 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3244 } else { 3245 assert(raw_val == NULL, "val may not be null"); 3246 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3247 } 3248 mem = phase->transform(mem); 3249 done_offset += BytesPerInt; 3250 } 3251 assert(done_offset == end_offset, ""); 3252 return mem; 3253 } 3254 3255 //============================================================================= 3256 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3257 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 3258 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3259 #ifdef ASSERT 3260 , _pair_idx(0) 3261 #endif 3262 { 3263 init_class_id(Class_MemBar); 3264 Node* top = C->top(); 3265 init_req(TypeFunc::I_O,top); 3266 init_req(TypeFunc::FramePtr,top); 3267 init_req(TypeFunc::ReturnAdr,top); 3268 if (precedent != NULL) 3269 init_req(TypeFunc::Parms, precedent); 3270 } 3271 3272 //------------------------------cmp-------------------------------------------- 3273 uint MemBarNode::hash() const { return NO_HASH; } 3274 bool MemBarNode::cmp( const Node &n ) const { 3275 return (&n == this); // Always fail except on self 3276 } 3277 3278 //------------------------------make------------------------------------------- 3279 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3280 switch (opcode) { 3281 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3282 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3283 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3284 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3285 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3286 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3287 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3288 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3289 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3290 case Op_Initialize: return new InitializeNode(C, atp, pn); 3291 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3292 default: ShouldNotReachHere(); return NULL; 3293 } 3294 } 3295 3296 void MemBarNode::remove(PhaseIterGVN *igvn) { 3297 if (outcnt() != 2) { 3298 return; 3299 } 3300 if (trailing_store() || trailing_load_store()) { 3301 MemBarNode* leading = leading_membar(); 3302 if (leading != NULL) { 3303 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3304 leading->remove(igvn); 3305 } 3306 } 3307 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3308 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3309 } 3310 3311 //------------------------------Ideal------------------------------------------ 3312 // Return a node which is more "ideal" than the current node. Strip out 3313 // control copies 3314 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3315 if (remove_dead_region(phase, can_reshape)) return this; 3316 // Don't bother trying to transform a dead node 3317 if (in(0) && in(0)->is_top()) { 3318 return NULL; 3319 } 3320 3321 bool progress = false; 3322 // Eliminate volatile MemBars for scalar replaced objects. 3323 if (can_reshape && req() == (Precedent+1)) { 3324 bool eliminate = false; 3325 int opc = Opcode(); 3326 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3327 // Volatile field loads and stores. 3328 Node* my_mem = in(MemBarNode::Precedent); 3329 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3330 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3331 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3332 // replace this Precedent (decodeN) with the Load instead. 3333 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3334 Node* load_node = my_mem->in(1); 3335 set_req(MemBarNode::Precedent, load_node); 3336 phase->is_IterGVN()->_worklist.push(my_mem); 3337 my_mem = load_node; 3338 } else { 3339 assert(my_mem->unique_out() == this, "sanity"); 3340 del_req(Precedent); 3341 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3342 my_mem = NULL; 3343 } 3344 progress = true; 3345 } 3346 if (my_mem != NULL && my_mem->is_Mem()) { 3347 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3348 // Check for scalar replaced object reference. 3349 if( t_oop != NULL && t_oop->is_known_instance_field() && 3350 t_oop->offset() != Type::OffsetBot && 3351 t_oop->offset() != Type::OffsetTop) { 3352 eliminate = true; 3353 } 3354 } 3355 } else if (opc == Op_MemBarRelease) { 3356 // Final field stores. 3357 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3358 if ((alloc != NULL) && alloc->is_Allocate() && 3359 alloc->as_Allocate()->does_not_escape_thread()) { 3360 // The allocated object does not escape. 3361 eliminate = true; 3362 } 3363 } 3364 if (eliminate) { 3365 // Replace MemBar projections by its inputs. 3366 PhaseIterGVN* igvn = phase->is_IterGVN(); 3367 remove(igvn); 3368 // Must return either the original node (now dead) or a new node 3369 // (Do not return a top here, since that would break the uniqueness of top.) 3370 return new ConINode(TypeInt::ZERO); 3371 } 3372 } 3373 return progress ? this : NULL; 3374 } 3375 3376 //------------------------------Value------------------------------------------ 3377 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3378 if( !in(0) ) return Type::TOP; 3379 if( phase->type(in(0)) == Type::TOP ) 3380 return Type::TOP; 3381 return TypeTuple::MEMBAR; 3382 } 3383 3384 //------------------------------match------------------------------------------ 3385 // Construct projections for memory. 3386 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) { 3387 switch (proj->_con) { 3388 case TypeFunc::Control: 3389 case TypeFunc::Memory: 3390 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3391 } 3392 ShouldNotReachHere(); 3393 return NULL; 3394 } 3395 3396 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3397 trailing->_kind = TrailingStore; 3398 leading->_kind = LeadingStore; 3399 #ifdef ASSERT 3400 trailing->_pair_idx = leading->_idx; 3401 leading->_pair_idx = leading->_idx; 3402 #endif 3403 } 3404 3405 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3406 trailing->_kind = TrailingLoadStore; 3407 leading->_kind = LeadingLoadStore; 3408 #ifdef ASSERT 3409 trailing->_pair_idx = leading->_idx; 3410 leading->_pair_idx = leading->_idx; 3411 #endif 3412 } 3413 3414 MemBarNode* MemBarNode::trailing_membar() const { 3415 ResourceMark rm; 3416 Node* trailing = (Node*)this; 3417 VectorSet seen(Thread::current()->resource_area()); 3418 Node_Stack multis(0); 3419 do { 3420 Node* c = trailing; 3421 uint i = 0; 3422 do { 3423 trailing = NULL; 3424 for (; i < c->outcnt(); i++) { 3425 Node* next = c->raw_out(i); 3426 if (next != c && next->is_CFG()) { 3427 if (c->is_MultiBranch()) { 3428 if (multis.node() == c) { 3429 multis.set_index(i+1); 3430 } else { 3431 multis.push(c, i+1); 3432 } 3433 } 3434 trailing = next; 3435 break; 3436 } 3437 } 3438 if (trailing != NULL && !seen.test_set(trailing->_idx)) { 3439 break; 3440 } 3441 while (multis.size() > 0) { 3442 c = multis.node(); 3443 i = multis.index(); 3444 if (i < c->req()) { 3445 break; 3446 } 3447 multis.pop(); 3448 } 3449 } while (multis.size() > 0); 3450 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3451 3452 MemBarNode* mb = trailing->as_MemBar(); 3453 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3454 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3455 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3456 return mb; 3457 } 3458 3459 MemBarNode* MemBarNode::leading_membar() const { 3460 ResourceMark rm; 3461 VectorSet seen(Thread::current()->resource_area()); 3462 Node_Stack regions(0); 3463 Node* leading = in(0); 3464 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3465 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) { 3466 leading = NULL; 3467 while (regions.size() > 0 && leading == NULL) { 3468 Node* r = regions.node(); 3469 uint i = regions.index(); 3470 if (i < r->req()) { 3471 leading = r->in(i); 3472 regions.set_index(i+1); 3473 } else { 3474 regions.pop(); 3475 } 3476 } 3477 if (leading == NULL) { 3478 assert(regions.size() == 0, "all paths should have been tried"); 3479 return NULL; 3480 } 3481 } 3482 if (leading->is_Region()) { 3483 regions.push(leading, 2); 3484 leading = leading->in(1); 3485 } else { 3486 leading = leading->in(0); 3487 } 3488 } 3489 #ifdef ASSERT 3490 Unique_Node_List wq; 3491 wq.push((Node*)this); 3492 uint found = 0; 3493 for (uint i = 0; i < wq.size(); i++) { 3494 Node* n = wq.at(i); 3495 if (n->is_Region()) { 3496 for (uint j = 1; j < n->req(); j++) { 3497 Node* in = n->in(j); 3498 if (in != NULL && !in->is_top()) { 3499 wq.push(in); 3500 } 3501 } 3502 } else { 3503 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3504 assert(n == leading, "consistency check failed"); 3505 found++; 3506 } else { 3507 Node* in = n->in(0); 3508 if (in != NULL && !in->is_top()) { 3509 wq.push(in); 3510 } 3511 } 3512 } 3513 } 3514 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed"); 3515 #endif 3516 if (leading == NULL) { 3517 return NULL; 3518 } 3519 MemBarNode* mb = leading->as_MemBar(); 3520 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3521 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3522 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3523 return mb; 3524 } 3525 3526 //===========================InitializeNode==================================== 3527 // SUMMARY: 3528 // This node acts as a memory barrier on raw memory, after some raw stores. 3529 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3530 // The Initialize can 'capture' suitably constrained stores as raw inits. 3531 // It can coalesce related raw stores into larger units (called 'tiles'). 3532 // It can avoid zeroing new storage for memory units which have raw inits. 3533 // At macro-expansion, it is marked 'complete', and does not optimize further. 3534 // 3535 // EXAMPLE: 3536 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3537 // ctl = incoming control; mem* = incoming memory 3538 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3539 // First allocate uninitialized memory and fill in the header: 3540 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3541 // ctl := alloc.Control; mem* := alloc.Memory* 3542 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3543 // Then initialize to zero the non-header parts of the raw memory block: 3544 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3545 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3546 // After the initialize node executes, the object is ready for service: 3547 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3548 // Suppose its body is immediately initialized as {1,2}: 3549 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3550 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3551 // mem.SLICE(#short[*]) := store2 3552 // 3553 // DETAILS: 3554 // An InitializeNode collects and isolates object initialization after 3555 // an AllocateNode and before the next possible safepoint. As a 3556 // memory barrier (MemBarNode), it keeps critical stores from drifting 3557 // down past any safepoint or any publication of the allocation. 3558 // Before this barrier, a newly-allocated object may have uninitialized bits. 3559 // After this barrier, it may be treated as a real oop, and GC is allowed. 3560 // 3561 // The semantics of the InitializeNode include an implicit zeroing of 3562 // the new object from object header to the end of the object. 3563 // (The object header and end are determined by the AllocateNode.) 3564 // 3565 // Certain stores may be added as direct inputs to the InitializeNode. 3566 // These stores must update raw memory, and they must be to addresses 3567 // derived from the raw address produced by AllocateNode, and with 3568 // a constant offset. They must be ordered by increasing offset. 3569 // The first one is at in(RawStores), the last at in(req()-1). 3570 // Unlike most memory operations, they are not linked in a chain, 3571 // but are displayed in parallel as users of the rawmem output of 3572 // the allocation. 3573 // 3574 // (See comments in InitializeNode::capture_store, which continue 3575 // the example given above.) 3576 // 3577 // When the associated Allocate is macro-expanded, the InitializeNode 3578 // may be rewritten to optimize collected stores. A ClearArrayNode 3579 // may also be created at that point to represent any required zeroing. 3580 // The InitializeNode is then marked 'complete', prohibiting further 3581 // capturing of nearby memory operations. 3582 // 3583 // During macro-expansion, all captured initializations which store 3584 // constant values of 32 bits or smaller are coalesced (if advantageous) 3585 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3586 // initialized in fewer memory operations. Memory words which are 3587 // covered by neither tiles nor non-constant stores are pre-zeroed 3588 // by explicit stores of zero. (The code shape happens to do all 3589 // zeroing first, then all other stores, with both sequences occurring 3590 // in order of ascending offsets.) 3591 // 3592 // Alternatively, code may be inserted between an AllocateNode and its 3593 // InitializeNode, to perform arbitrary initialization of the new object. 3594 // E.g., the object copying intrinsics insert complex data transfers here. 3595 // The initialization must then be marked as 'complete' disable the 3596 // built-in zeroing semantics and the collection of initializing stores. 3597 // 3598 // While an InitializeNode is incomplete, reads from the memory state 3599 // produced by it are optimizable if they match the control edge and 3600 // new oop address associated with the allocation/initialization. 3601 // They return a stored value (if the offset matches) or else zero. 3602 // A write to the memory state, if it matches control and address, 3603 // and if it is to a constant offset, may be 'captured' by the 3604 // InitializeNode. It is cloned as a raw memory operation and rewired 3605 // inside the initialization, to the raw oop produced by the allocation. 3606 // Operations on addresses which are provably distinct (e.g., to 3607 // other AllocateNodes) are allowed to bypass the initialization. 3608 // 3609 // The effect of all this is to consolidate object initialization 3610 // (both arrays and non-arrays, both piecewise and bulk) into a 3611 // single location, where it can be optimized as a unit. 3612 // 3613 // Only stores with an offset less than TrackedInitializationLimit words 3614 // will be considered for capture by an InitializeNode. This puts a 3615 // reasonable limit on the complexity of optimized initializations. 3616 3617 //---------------------------InitializeNode------------------------------------ 3618 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3619 : MemBarNode(C, adr_type, rawoop), 3620 _is_complete(Incomplete), _does_not_escape(false) 3621 { 3622 init_class_id(Class_Initialize); 3623 3624 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3625 assert(in(RawAddress) == rawoop, "proper init"); 3626 // Note: allocation() can be NULL, for secondary initialization barriers 3627 } 3628 3629 // Since this node is not matched, it will be processed by the 3630 // register allocator. Declare that there are no constraints 3631 // on the allocation of the RawAddress edge. 3632 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3633 // This edge should be set to top, by the set_complete. But be conservative. 3634 if (idx == InitializeNode::RawAddress) 3635 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3636 return RegMask::Empty; 3637 } 3638 3639 Node* InitializeNode::memory(uint alias_idx) { 3640 Node* mem = in(Memory); 3641 if (mem->is_MergeMem()) { 3642 return mem->as_MergeMem()->memory_at(alias_idx); 3643 } else { 3644 // incoming raw memory is not split 3645 return mem; 3646 } 3647 } 3648 3649 bool InitializeNode::is_non_zero() { 3650 if (is_complete()) return false; 3651 remove_extra_zeroes(); 3652 return (req() > RawStores); 3653 } 3654 3655 void InitializeNode::set_complete(PhaseGVN* phase) { 3656 assert(!is_complete(), "caller responsibility"); 3657 _is_complete = Complete; 3658 3659 // After this node is complete, it contains a bunch of 3660 // raw-memory initializations. There is no need for 3661 // it to have anything to do with non-raw memory effects. 3662 // Therefore, tell all non-raw users to re-optimize themselves, 3663 // after skipping the memory effects of this initialization. 3664 PhaseIterGVN* igvn = phase->is_IterGVN(); 3665 if (igvn) igvn->add_users_to_worklist(this); 3666 } 3667 3668 // convenience function 3669 // return false if the init contains any stores already 3670 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3671 InitializeNode* init = initialization(); 3672 if (init == NULL || init->is_complete()) { 3673 return false; 3674 } 3675 init->remove_extra_zeroes(); 3676 // for now, if this allocation has already collected any inits, bail: 3677 if (init->is_non_zero()) return false; 3678 init->set_complete(phase); 3679 return true; 3680 } 3681 3682 void InitializeNode::remove_extra_zeroes() { 3683 if (req() == RawStores) return; 3684 Node* zmem = zero_memory(); 3685 uint fill = RawStores; 3686 for (uint i = fill; i < req(); i++) { 3687 Node* n = in(i); 3688 if (n->is_top() || n == zmem) continue; // skip 3689 if (fill < i) set_req(fill, n); // compact 3690 ++fill; 3691 } 3692 // delete any empty spaces created: 3693 while (fill < req()) { 3694 del_req(fill); 3695 } 3696 } 3697 3698 // Helper for remembering which stores go with which offsets. 3699 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3700 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3701 intptr_t offset = -1; 3702 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3703 phase, offset); 3704 if (base == NULL) return -1; // something is dead, 3705 if (offset < 0) return -1; // dead, dead 3706 return offset; 3707 } 3708 3709 // Helper for proving that an initialization expression is 3710 // "simple enough" to be folded into an object initialization. 3711 // Attempts to prove that a store's initial value 'n' can be captured 3712 // within the initialization without creating a vicious cycle, such as: 3713 // { Foo p = new Foo(); p.next = p; } 3714 // True for constants and parameters and small combinations thereof. 3715 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) { 3716 ResourceMark rm; 3717 Unique_Node_List worklist; 3718 worklist.push(value); 3719 3720 uint complexity_limit = 20; 3721 for (uint j = 0; j < worklist.size(); j++) { 3722 if (j >= complexity_limit) { 3723 return false; // Bail out if processed too many nodes 3724 } 3725 3726 Node* n = worklist.at(j); 3727 if (n == NULL) continue; // (can this really happen?) 3728 if (n->is_Proj()) n = n->in(0); 3729 if (n == this) return false; // found a cycle 3730 if (n->is_Con()) continue; 3731 if (n->is_Start()) continue; // params, etc., are OK 3732 if (n->is_Root()) continue; // even better 3733 3734 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation 3735 if (n->is_CFG() && phase->is_dominator(n, allocation())) { 3736 continue; 3737 } 3738 3739 Node* ctl = n->in(0); 3740 if (ctl != NULL && !ctl->is_top()) { 3741 if (ctl->is_Proj()) ctl = ctl->in(0); 3742 if (ctl == this) return false; 3743 3744 // If we already know that the enclosing memory op is pinned right after 3745 // the init, then any control flow that the store has picked up 3746 // must have preceded the init, or else be equal to the init. 3747 // Even after loop optimizations (which might change control edges) 3748 // a store is never pinned *before* the availability of its inputs. 3749 if (!MemNode::all_controls_dominate(n, this)) 3750 return false; // failed to prove a good control 3751 } 3752 3753 // Check data edges for possible dependencies on 'this'. 3754 for (uint i = 1; i < n->req(); i++) { 3755 Node* m = n->in(i); 3756 if (m == NULL || m == n || m->is_top()) continue; 3757 3758 // Only process data inputs once 3759 worklist.push(m); 3760 } 3761 } 3762 3763 return true; 3764 } 3765 3766 // Here are all the checks a Store must pass before it can be moved into 3767 // an initialization. Returns zero if a check fails. 3768 // On success, returns the (constant) offset to which the store applies, 3769 // within the initialized memory. 3770 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) { 3771 const int FAIL = 0; 3772 if (st->req() != MemNode::ValueIn + 1) 3773 return FAIL; // an inscrutable StoreNode (card mark?) 3774 Node* ctl = st->in(MemNode::Control); 3775 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3776 return FAIL; // must be unconditional after the initialization 3777 Node* mem = st->in(MemNode::Memory); 3778 if (!(mem->is_Proj() && mem->in(0) == this)) 3779 return FAIL; // must not be preceded by other stores 3780 Node* adr = st->in(MemNode::Address); 3781 intptr_t offset; 3782 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3783 if (alloc == NULL) 3784 return FAIL; // inscrutable address 3785 if (alloc != allocation()) 3786 return FAIL; // wrong allocation! (store needs to float up) 3787 int size_in_bytes = st->memory_size(); 3788 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) { 3789 return FAIL; // mismatched access 3790 } 3791 Node* val = st->in(MemNode::ValueIn); 3792 3793 if (!detect_init_independence(val, phase)) 3794 return FAIL; // stored value must be 'simple enough' 3795 3796 // The Store can be captured only if nothing after the allocation 3797 // and before the Store is using the memory location that the store 3798 // overwrites. 3799 bool failed = false; 3800 // If is_complete_with_arraycopy() is true the shape of the graph is 3801 // well defined and is safe so no need for extra checks. 3802 if (!is_complete_with_arraycopy()) { 3803 // We are going to look at each use of the memory state following 3804 // the allocation to make sure nothing reads the memory that the 3805 // Store writes. 3806 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3807 int alias_idx = phase->C->get_alias_index(t_adr); 3808 ResourceMark rm; 3809 Unique_Node_List mems; 3810 mems.push(mem); 3811 Node* unique_merge = NULL; 3812 for (uint next = 0; next < mems.size(); ++next) { 3813 Node *m = mems.at(next); 3814 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3815 Node *n = m->fast_out(j); 3816 if (n->outcnt() == 0) { 3817 continue; 3818 } 3819 if (n == st) { 3820 continue; 3821 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3822 // If the control of this use is different from the control 3823 // of the Store which is right after the InitializeNode then 3824 // this node cannot be between the InitializeNode and the 3825 // Store. 3826 continue; 3827 } else if (n->is_MergeMem()) { 3828 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3829 // We can hit a MergeMemNode (that will likely go away 3830 // later) that is a direct use of the memory state 3831 // following the InitializeNode on the same slice as the 3832 // store node that we'd like to capture. We need to check 3833 // the uses of the MergeMemNode. 3834 mems.push(n); 3835 } 3836 } else if (n->is_Mem()) { 3837 Node* other_adr = n->in(MemNode::Address); 3838 if (other_adr == adr) { 3839 failed = true; 3840 break; 3841 } else { 3842 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3843 if (other_t_adr != NULL) { 3844 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3845 if (other_alias_idx == alias_idx) { 3846 // A load from the same memory slice as the store right 3847 // after the InitializeNode. We check the control of the 3848 // object/array that is loaded from. If it's the same as 3849 // the store control then we cannot capture the store. 3850 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3851 Node* base = other_adr; 3852 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3853 base = base->in(AddPNode::Base); 3854 if (base != NULL) { 3855 base = base->uncast(); 3856 if (base->is_Proj() && base->in(0) == alloc) { 3857 failed = true; 3858 break; 3859 } 3860 } 3861 } 3862 } 3863 } 3864 } else { 3865 failed = true; 3866 break; 3867 } 3868 } 3869 } 3870 } 3871 if (failed) { 3872 if (!can_reshape) { 3873 // We decided we couldn't capture the store during parsing. We 3874 // should try again during the next IGVN once the graph is 3875 // cleaner. 3876 phase->C->record_for_igvn(st); 3877 } 3878 return FAIL; 3879 } 3880 3881 return offset; // success 3882 } 3883 3884 // Find the captured store in(i) which corresponds to the range 3885 // [start..start+size) in the initialized object. 3886 // If there is one, return its index i. If there isn't, return the 3887 // negative of the index where it should be inserted. 3888 // Return 0 if the queried range overlaps an initialization boundary 3889 // or if dead code is encountered. 3890 // If size_in_bytes is zero, do not bother with overlap checks. 3891 int InitializeNode::captured_store_insertion_point(intptr_t start, 3892 int size_in_bytes, 3893 PhaseTransform* phase) { 3894 const int FAIL = 0, MAX_STORE = BytesPerLong; 3895 3896 if (is_complete()) 3897 return FAIL; // arraycopy got here first; punt 3898 3899 assert(allocation() != NULL, "must be present"); 3900 3901 // no negatives, no header fields: 3902 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3903 3904 // after a certain size, we bail out on tracking all the stores: 3905 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3906 if (start >= ti_limit) return FAIL; 3907 3908 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3909 if (i >= limit) return -(int)i; // not found; here is where to put it 3910 3911 Node* st = in(i); 3912 intptr_t st_off = get_store_offset(st, phase); 3913 if (st_off < 0) { 3914 if (st != zero_memory()) { 3915 return FAIL; // bail out if there is dead garbage 3916 } 3917 } else if (st_off > start) { 3918 // ...we are done, since stores are ordered 3919 if (st_off < start + size_in_bytes) { 3920 return FAIL; // the next store overlaps 3921 } 3922 return -(int)i; // not found; here is where to put it 3923 } else if (st_off < start) { 3924 if (size_in_bytes != 0 && 3925 start < st_off + MAX_STORE && 3926 start < st_off + st->as_Store()->memory_size()) { 3927 return FAIL; // the previous store overlaps 3928 } 3929 } else { 3930 if (size_in_bytes != 0 && 3931 st->as_Store()->memory_size() != size_in_bytes) { 3932 return FAIL; // mismatched store size 3933 } 3934 return i; 3935 } 3936 3937 ++i; 3938 } 3939 } 3940 3941 // Look for a captured store which initializes at the offset 'start' 3942 // with the given size. If there is no such store, and no other 3943 // initialization interferes, then return zero_memory (the memory 3944 // projection of the AllocateNode). 3945 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3946 PhaseTransform* phase) { 3947 assert(stores_are_sane(phase), ""); 3948 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3949 if (i == 0) { 3950 return NULL; // something is dead 3951 } else if (i < 0) { 3952 return zero_memory(); // just primordial zero bits here 3953 } else { 3954 Node* st = in(i); // here is the store at this position 3955 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3956 return st; 3957 } 3958 } 3959 3960 // Create, as a raw pointer, an address within my new object at 'offset'. 3961 Node* InitializeNode::make_raw_address(intptr_t offset, 3962 PhaseTransform* phase) { 3963 Node* addr = in(RawAddress); 3964 if (offset != 0) { 3965 Compile* C = phase->C; 3966 addr = phase->transform( new AddPNode(C->top(), addr, 3967 phase->MakeConX(offset)) ); 3968 } 3969 return addr; 3970 } 3971 3972 // Clone the given store, converting it into a raw store 3973 // initializing a field or element of my new object. 3974 // Caller is responsible for retiring the original store, 3975 // with subsume_node or the like. 3976 // 3977 // From the example above InitializeNode::InitializeNode, 3978 // here are the old stores to be captured: 3979 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3980 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3981 // 3982 // Here is the changed code; note the extra edges on init: 3983 // alloc = (Allocate ...) 3984 // rawoop = alloc.RawAddress 3985 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3986 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3987 // init = (Initialize alloc.Control alloc.Memory rawoop 3988 // rawstore1 rawstore2) 3989 // 3990 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3991 PhaseGVN* phase, bool can_reshape) { 3992 assert(stores_are_sane(phase), ""); 3993 3994 if (start < 0) return NULL; 3995 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3996 3997 Compile* C = phase->C; 3998 int size_in_bytes = st->memory_size(); 3999 int i = captured_store_insertion_point(start, size_in_bytes, phase); 4000 if (i == 0) return NULL; // bail out 4001 Node* prev_mem = NULL; // raw memory for the captured store 4002 if (i > 0) { 4003 prev_mem = in(i); // there is a pre-existing store under this one 4004 set_req(i, C->top()); // temporarily disconnect it 4005 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 4006 } else { 4007 i = -i; // no pre-existing store 4008 prev_mem = zero_memory(); // a slice of the newly allocated object 4009 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 4010 set_req(--i, C->top()); // reuse this edge; it has been folded away 4011 else 4012 ins_req(i, C->top()); // build a new edge 4013 } 4014 Node* new_st = st->clone(); 4015 new_st->set_req(MemNode::Control, in(Control)); 4016 new_st->set_req(MemNode::Memory, prev_mem); 4017 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 4018 new_st = phase->transform(new_st); 4019 4020 // At this point, new_st might have swallowed a pre-existing store 4021 // at the same offset, or perhaps new_st might have disappeared, 4022 // if it redundantly stored the same value (or zero to fresh memory). 4023 4024 // In any case, wire it in: 4025 phase->igvn_rehash_node_delayed(this); 4026 set_req(i, new_st); 4027 4028 // The caller may now kill the old guy. 4029 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 4030 assert(check_st == new_st || check_st == NULL, "must be findable"); 4031 assert(!is_complete(), ""); 4032 return new_st; 4033 } 4034 4035 static bool store_constant(jlong* tiles, int num_tiles, 4036 intptr_t st_off, int st_size, 4037 jlong con) { 4038 if ((st_off & (st_size-1)) != 0) 4039 return false; // strange store offset (assume size==2**N) 4040 address addr = (address)tiles + st_off; 4041 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 4042 switch (st_size) { 4043 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 4044 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 4045 case sizeof(jint): *(jint*) addr = (jint) con; break; 4046 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 4047 default: return false; // strange store size (detect size!=2**N here) 4048 } 4049 return true; // return success to caller 4050 } 4051 4052 // Coalesce subword constants into int constants and possibly 4053 // into long constants. The goal, if the CPU permits, 4054 // is to initialize the object with a small number of 64-bit tiles. 4055 // Also, convert floating-point constants to bit patterns. 4056 // Non-constants are not relevant to this pass. 4057 // 4058 // In terms of the running example on InitializeNode::InitializeNode 4059 // and InitializeNode::capture_store, here is the transformation 4060 // of rawstore1 and rawstore2 into rawstore12: 4061 // alloc = (Allocate ...) 4062 // rawoop = alloc.RawAddress 4063 // tile12 = 0x00010002 4064 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 4065 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 4066 // 4067 void 4068 InitializeNode::coalesce_subword_stores(intptr_t header_size, 4069 Node* size_in_bytes, 4070 PhaseGVN* phase) { 4071 Compile* C = phase->C; 4072 4073 assert(stores_are_sane(phase), ""); 4074 // Note: After this pass, they are not completely sane, 4075 // since there may be some overlaps. 4076 4077 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 4078 4079 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 4080 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 4081 size_limit = MIN2(size_limit, ti_limit); 4082 size_limit = align_up(size_limit, BytesPerLong); 4083 int num_tiles = size_limit / BytesPerLong; 4084 4085 // allocate space for the tile map: 4086 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 4087 jlong tiles_buf[small_len]; 4088 Node* nodes_buf[small_len]; 4089 jlong inits_buf[small_len]; 4090 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 4091 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4092 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 4093 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 4094 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 4095 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4096 // tiles: exact bitwise model of all primitive constants 4097 // nodes: last constant-storing node subsumed into the tiles model 4098 // inits: which bytes (in each tile) are touched by any initializations 4099 4100 //// Pass A: Fill in the tile model with any relevant stores. 4101 4102 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 4103 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 4104 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 4105 Node* zmem = zero_memory(); // initially zero memory state 4106 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4107 Node* st = in(i); 4108 intptr_t st_off = get_store_offset(st, phase); 4109 4110 // Figure out the store's offset and constant value: 4111 if (st_off < header_size) continue; //skip (ignore header) 4112 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 4113 int st_size = st->as_Store()->memory_size(); 4114 if (st_off + st_size > size_limit) break; 4115 4116 // Record which bytes are touched, whether by constant or not. 4117 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 4118 continue; // skip (strange store size) 4119 4120 const Type* val = phase->type(st->in(MemNode::ValueIn)); 4121 if (!val->singleton()) continue; //skip (non-con store) 4122 BasicType type = val->basic_type(); 4123 4124 jlong con = 0; 4125 switch (type) { 4126 case T_INT: con = val->is_int()->get_con(); break; 4127 case T_LONG: con = val->is_long()->get_con(); break; 4128 case T_FLOAT: con = jint_cast(val->getf()); break; 4129 case T_DOUBLE: con = jlong_cast(val->getd()); break; 4130 default: continue; //skip (odd store type) 4131 } 4132 4133 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 4134 st->Opcode() == Op_StoreL) { 4135 continue; // This StoreL is already optimal. 4136 } 4137 4138 // Store down the constant. 4139 store_constant(tiles, num_tiles, st_off, st_size, con); 4140 4141 intptr_t j = st_off >> LogBytesPerLong; 4142 4143 if (type == T_INT && st_size == BytesPerInt 4144 && (st_off & BytesPerInt) == BytesPerInt) { 4145 jlong lcon = tiles[j]; 4146 if (!Matcher::isSimpleConstant64(lcon) && 4147 st->Opcode() == Op_StoreI) { 4148 // This StoreI is already optimal by itself. 4149 jint* intcon = (jint*) &tiles[j]; 4150 intcon[1] = 0; // undo the store_constant() 4151 4152 // If the previous store is also optimal by itself, back up and 4153 // undo the action of the previous loop iteration... if we can. 4154 // But if we can't, just let the previous half take care of itself. 4155 st = nodes[j]; 4156 st_off -= BytesPerInt; 4157 con = intcon[0]; 4158 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 4159 assert(st_off >= header_size, "still ignoring header"); 4160 assert(get_store_offset(st, phase) == st_off, "must be"); 4161 assert(in(i-1) == zmem, "must be"); 4162 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 4163 assert(con == tcon->is_int()->get_con(), "must be"); 4164 // Undo the effects of the previous loop trip, which swallowed st: 4165 intcon[0] = 0; // undo store_constant() 4166 set_req(i-1, st); // undo set_req(i, zmem) 4167 nodes[j] = NULL; // undo nodes[j] = st 4168 --old_subword; // undo ++old_subword 4169 } 4170 continue; // This StoreI is already optimal. 4171 } 4172 } 4173 4174 // This store is not needed. 4175 set_req(i, zmem); 4176 nodes[j] = st; // record for the moment 4177 if (st_size < BytesPerLong) // something has changed 4178 ++old_subword; // includes int/float, but who's counting... 4179 else ++old_long; 4180 } 4181 4182 if ((old_subword + old_long) == 0) 4183 return; // nothing more to do 4184 4185 //// Pass B: Convert any non-zero tiles into optimal constant stores. 4186 // Be sure to insert them before overlapping non-constant stores. 4187 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 4188 for (int j = 0; j < num_tiles; j++) { 4189 jlong con = tiles[j]; 4190 jlong init = inits[j]; 4191 if (con == 0) continue; 4192 jint con0, con1; // split the constant, address-wise 4193 jint init0, init1; // split the init map, address-wise 4194 { union { jlong con; jint intcon[2]; } u; 4195 u.con = con; 4196 con0 = u.intcon[0]; 4197 con1 = u.intcon[1]; 4198 u.con = init; 4199 init0 = u.intcon[0]; 4200 init1 = u.intcon[1]; 4201 } 4202 4203 Node* old = nodes[j]; 4204 assert(old != NULL, "need the prior store"); 4205 intptr_t offset = (j * BytesPerLong); 4206 4207 bool split = !Matcher::isSimpleConstant64(con); 4208 4209 if (offset < header_size) { 4210 assert(offset + BytesPerInt >= header_size, "second int counts"); 4211 assert(*(jint*)&tiles[j] == 0, "junk in header"); 4212 split = true; // only the second word counts 4213 // Example: int a[] = { 42 ... } 4214 } else if (con0 == 0 && init0 == -1) { 4215 split = true; // first word is covered by full inits 4216 // Example: int a[] = { ... foo(), 42 ... } 4217 } else if (con1 == 0 && init1 == -1) { 4218 split = true; // second word is covered by full inits 4219 // Example: int a[] = { ... 42, foo() ... } 4220 } 4221 4222 // Here's a case where init0 is neither 0 nor -1: 4223 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4224 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4225 // In this case the tile is not split; it is (jlong)42. 4226 // The big tile is stored down, and then the foo() value is inserted. 4227 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4228 4229 Node* ctl = old->in(MemNode::Control); 4230 Node* adr = make_raw_address(offset, phase); 4231 const TypePtr* atp = TypeRawPtr::BOTTOM; 4232 4233 // One or two coalesced stores to plop down. 4234 Node* st[2]; 4235 intptr_t off[2]; 4236 int nst = 0; 4237 if (!split) { 4238 ++new_long; 4239 off[nst] = offset; 4240 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4241 phase->longcon(con), T_LONG, MemNode::unordered); 4242 } else { 4243 // Omit either if it is a zero. 4244 if (con0 != 0) { 4245 ++new_int; 4246 off[nst] = offset; 4247 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4248 phase->intcon(con0), T_INT, MemNode::unordered); 4249 } 4250 if (con1 != 0) { 4251 ++new_int; 4252 offset += BytesPerInt; 4253 adr = make_raw_address(offset, phase); 4254 off[nst] = offset; 4255 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4256 phase->intcon(con1), T_INT, MemNode::unordered); 4257 } 4258 } 4259 4260 // Insert second store first, then the first before the second. 4261 // Insert each one just before any overlapping non-constant stores. 4262 while (nst > 0) { 4263 Node* st1 = st[--nst]; 4264 C->copy_node_notes_to(st1, old); 4265 st1 = phase->transform(st1); 4266 offset = off[nst]; 4267 assert(offset >= header_size, "do not smash header"); 4268 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4269 guarantee(ins_idx != 0, "must re-insert constant store"); 4270 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4271 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4272 set_req(--ins_idx, st1); 4273 else 4274 ins_req(ins_idx, st1); 4275 } 4276 } 4277 4278 if (PrintCompilation && WizardMode) 4279 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4280 old_subword, old_long, new_int, new_long); 4281 if (C->log() != NULL) 4282 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4283 old_subword, old_long, new_int, new_long); 4284 4285 // Clean up any remaining occurrences of zmem: 4286 remove_extra_zeroes(); 4287 } 4288 4289 // Explore forward from in(start) to find the first fully initialized 4290 // word, and return its offset. Skip groups of subword stores which 4291 // together initialize full words. If in(start) is itself part of a 4292 // fully initialized word, return the offset of in(start). If there 4293 // are no following full-word stores, or if something is fishy, return 4294 // a negative value. 4295 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4296 int int_map = 0; 4297 intptr_t int_map_off = 0; 4298 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4299 4300 for (uint i = start, limit = req(); i < limit; i++) { 4301 Node* st = in(i); 4302 4303 intptr_t st_off = get_store_offset(st, phase); 4304 if (st_off < 0) break; // return conservative answer 4305 4306 int st_size = st->as_Store()->memory_size(); 4307 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4308 return st_off; // we found a complete word init 4309 } 4310 4311 // update the map: 4312 4313 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4314 if (this_int_off != int_map_off) { 4315 // reset the map: 4316 int_map = 0; 4317 int_map_off = this_int_off; 4318 } 4319 4320 int subword_off = st_off - this_int_off; 4321 int_map |= right_n_bits(st_size) << subword_off; 4322 if ((int_map & FULL_MAP) == FULL_MAP) { 4323 return this_int_off; // we found a complete word init 4324 } 4325 4326 // Did this store hit or cross the word boundary? 4327 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4328 if (next_int_off == this_int_off + BytesPerInt) { 4329 // We passed the current int, without fully initializing it. 4330 int_map_off = next_int_off; 4331 int_map >>= BytesPerInt; 4332 } else if (next_int_off > this_int_off + BytesPerInt) { 4333 // We passed the current and next int. 4334 return this_int_off + BytesPerInt; 4335 } 4336 } 4337 4338 return -1; 4339 } 4340 4341 4342 // Called when the associated AllocateNode is expanded into CFG. 4343 // At this point, we may perform additional optimizations. 4344 // Linearize the stores by ascending offset, to make memory 4345 // activity as coherent as possible. 4346 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4347 intptr_t header_size, 4348 Node* size_in_bytes, 4349 PhaseGVN* phase) { 4350 assert(!is_complete(), "not already complete"); 4351 assert(stores_are_sane(phase), ""); 4352 assert(allocation() != NULL, "must be present"); 4353 4354 remove_extra_zeroes(); 4355 4356 if (ReduceFieldZeroing || ReduceBulkZeroing) 4357 // reduce instruction count for common initialization patterns 4358 coalesce_subword_stores(header_size, size_in_bytes, phase); 4359 4360 Node* zmem = zero_memory(); // initially zero memory state 4361 Node* inits = zmem; // accumulating a linearized chain of inits 4362 #ifdef ASSERT 4363 intptr_t first_offset = allocation()->minimum_header_size(); 4364 intptr_t last_init_off = first_offset; // previous init offset 4365 intptr_t last_init_end = first_offset; // previous init offset+size 4366 intptr_t last_tile_end = first_offset; // previous tile offset+size 4367 #endif 4368 intptr_t zeroes_done = header_size; 4369 4370 bool do_zeroing = true; // we might give up if inits are very sparse 4371 int big_init_gaps = 0; // how many large gaps have we seen? 4372 4373 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4374 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4375 4376 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4377 Node* st = in(i); 4378 intptr_t st_off = get_store_offset(st, phase); 4379 if (st_off < 0) 4380 break; // unknown junk in the inits 4381 if (st->in(MemNode::Memory) != zmem) 4382 break; // complicated store chains somehow in list 4383 4384 int st_size = st->as_Store()->memory_size(); 4385 intptr_t next_init_off = st_off + st_size; 4386 4387 if (do_zeroing && zeroes_done < next_init_off) { 4388 // See if this store needs a zero before it or under it. 4389 intptr_t zeroes_needed = st_off; 4390 4391 if (st_size < BytesPerInt) { 4392 // Look for subword stores which only partially initialize words. 4393 // If we find some, we must lay down some word-level zeroes first, 4394 // underneath the subword stores. 4395 // 4396 // Examples: 4397 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4398 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4399 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4400 // 4401 // Note: coalesce_subword_stores may have already done this, 4402 // if it was prompted by constant non-zero subword initializers. 4403 // But this case can still arise with non-constant stores. 4404 4405 intptr_t next_full_store = find_next_fullword_store(i, phase); 4406 4407 // In the examples above: 4408 // in(i) p q r s x y z 4409 // st_off 12 13 14 15 12 13 14 4410 // st_size 1 1 1 1 1 1 1 4411 // next_full_s. 12 16 16 16 16 16 16 4412 // z's_done 12 16 16 16 12 16 12 4413 // z's_needed 12 16 16 16 16 16 16 4414 // zsize 0 0 0 0 4 0 4 4415 if (next_full_store < 0) { 4416 // Conservative tack: Zero to end of current word. 4417 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4418 } else { 4419 // Zero to beginning of next fully initialized word. 4420 // Or, don't zero at all, if we are already in that word. 4421 assert(next_full_store >= zeroes_needed, "must go forward"); 4422 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4423 zeroes_needed = next_full_store; 4424 } 4425 } 4426 4427 if (zeroes_needed > zeroes_done) { 4428 intptr_t zsize = zeroes_needed - zeroes_done; 4429 // Do some incremental zeroing on rawmem, in parallel with inits. 4430 zeroes_done = align_down(zeroes_done, BytesPerInt); 4431 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4432 allocation()->in(AllocateNode::DefaultValue), 4433 allocation()->in(AllocateNode::RawDefaultValue), 4434 zeroes_done, zeroes_needed, 4435 phase); 4436 zeroes_done = zeroes_needed; 4437 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4438 do_zeroing = false; // leave the hole, next time 4439 } 4440 } 4441 4442 // Collect the store and move on: 4443 st->set_req(MemNode::Memory, inits); 4444 inits = st; // put it on the linearized chain 4445 set_req(i, zmem); // unhook from previous position 4446 4447 if (zeroes_done == st_off) 4448 zeroes_done = next_init_off; 4449 4450 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4451 4452 #ifdef ASSERT 4453 // Various order invariants. Weaker than stores_are_sane because 4454 // a large constant tile can be filled in by smaller non-constant stores. 4455 assert(st_off >= last_init_off, "inits do not reverse"); 4456 last_init_off = st_off; 4457 const Type* val = NULL; 4458 if (st_size >= BytesPerInt && 4459 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4460 (int)val->basic_type() < (int)T_OBJECT) { 4461 assert(st_off >= last_tile_end, "tiles do not overlap"); 4462 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4463 last_tile_end = MAX2(last_tile_end, next_init_off); 4464 } else { 4465 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4466 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4467 assert(st_off >= last_init_end, "inits do not overlap"); 4468 last_init_end = next_init_off; // it's a non-tile 4469 } 4470 #endif //ASSERT 4471 } 4472 4473 remove_extra_zeroes(); // clear out all the zmems left over 4474 add_req(inits); 4475 4476 if (!(UseTLAB && ZeroTLAB)) { 4477 // If anything remains to be zeroed, zero it all now. 4478 zeroes_done = align_down(zeroes_done, BytesPerInt); 4479 // if it is the last unused 4 bytes of an instance, forget about it 4480 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4481 if (zeroes_done + BytesPerLong >= size_limit) { 4482 AllocateNode* alloc = allocation(); 4483 assert(alloc != NULL, "must be present"); 4484 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 4485 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4486 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4487 if (zeroes_done == k->layout_helper()) 4488 zeroes_done = size_limit; 4489 } 4490 } 4491 if (zeroes_done < size_limit) { 4492 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4493 allocation()->in(AllocateNode::DefaultValue), 4494 allocation()->in(AllocateNode::RawDefaultValue), 4495 zeroes_done, size_in_bytes, phase); 4496 } 4497 } 4498 4499 set_complete(phase); 4500 return rawmem; 4501 } 4502 4503 4504 #ifdef ASSERT 4505 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4506 if (is_complete()) 4507 return true; // stores could be anything at this point 4508 assert(allocation() != NULL, "must be present"); 4509 intptr_t last_off = allocation()->minimum_header_size(); 4510 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4511 Node* st = in(i); 4512 intptr_t st_off = get_store_offset(st, phase); 4513 if (st_off < 0) continue; // ignore dead garbage 4514 if (last_off > st_off) { 4515 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4516 this->dump(2); 4517 assert(false, "ascending store offsets"); 4518 return false; 4519 } 4520 last_off = st_off + st->as_Store()->memory_size(); 4521 } 4522 return true; 4523 } 4524 #endif //ASSERT 4525 4526 4527 4528 4529 //============================MergeMemNode===================================== 4530 // 4531 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4532 // contributing store or call operations. Each contributor provides the memory 4533 // state for a particular "alias type" (see Compile::alias_type). For example, 4534 // if a MergeMem has an input X for alias category #6, then any memory reference 4535 // to alias category #6 may use X as its memory state input, as an exact equivalent 4536 // to using the MergeMem as a whole. 4537 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4538 // 4539 // (Here, the <N> notation gives the index of the relevant adr_type.) 4540 // 4541 // In one special case (and more cases in the future), alias categories overlap. 4542 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4543 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4544 // it is exactly equivalent to that state W: 4545 // MergeMem(<Bot>: W) <==> W 4546 // 4547 // Usually, the merge has more than one input. In that case, where inputs 4548 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4549 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4550 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4551 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4552 // 4553 // A merge can take a "wide" memory state as one of its narrow inputs. 4554 // This simply means that the merge observes out only the relevant parts of 4555 // the wide input. That is, wide memory states arriving at narrow merge inputs 4556 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4557 // 4558 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4559 // and that memory slices "leak through": 4560 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4561 // 4562 // But, in such a cascade, repeated memory slices can "block the leak": 4563 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4564 // 4565 // In the last example, Y is not part of the combined memory state of the 4566 // outermost MergeMem. The system must, of course, prevent unschedulable 4567 // memory states from arising, so you can be sure that the state Y is somehow 4568 // a precursor to state Y'. 4569 // 4570 // 4571 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4572 // of each MergeMemNode array are exactly the numerical alias indexes, including 4573 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4574 // Compile::alias_type (and kin) produce and manage these indexes. 4575 // 4576 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4577 // (Note that this provides quick access to the top node inside MergeMem methods, 4578 // without the need to reach out via TLS to Compile::current.) 4579 // 4580 // As a consequence of what was just described, a MergeMem that represents a full 4581 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4582 // containing all alias categories. 4583 // 4584 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4585 // 4586 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4587 // a memory state for the alias type <N>, or else the top node, meaning that 4588 // there is no particular input for that alias type. Note that the length of 4589 // a MergeMem is variable, and may be extended at any time to accommodate new 4590 // memory states at larger alias indexes. When merges grow, they are of course 4591 // filled with "top" in the unused in() positions. 4592 // 4593 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4594 // (Top was chosen because it works smoothly with passes like GCM.) 4595 // 4596 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4597 // the type of random VM bits like TLS references.) Since it is always the 4598 // first non-Bot memory slice, some low-level loops use it to initialize an 4599 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4600 // 4601 // 4602 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4603 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4604 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4605 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4606 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4607 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4608 // 4609 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4610 // really that different from the other memory inputs. An abbreviation called 4611 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4612 // 4613 // 4614 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4615 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4616 // that "emerges though" the base memory will be marked as excluding the alias types 4617 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4618 // 4619 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4620 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4621 // 4622 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4623 // (It is currently unimplemented.) As you can see, the resulting merge is 4624 // actually a disjoint union of memory states, rather than an overlay. 4625 // 4626 4627 //------------------------------MergeMemNode----------------------------------- 4628 Node* MergeMemNode::make_empty_memory() { 4629 Node* empty_memory = (Node*) Compile::current()->top(); 4630 assert(empty_memory->is_top(), "correct sentinel identity"); 4631 return empty_memory; 4632 } 4633 4634 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4635 init_class_id(Class_MergeMem); 4636 // all inputs are nullified in Node::Node(int) 4637 // set_input(0, NULL); // no control input 4638 4639 // Initialize the edges uniformly to top, for starters. 4640 Node* empty_mem = make_empty_memory(); 4641 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4642 init_req(i,empty_mem); 4643 } 4644 assert(empty_memory() == empty_mem, ""); 4645 4646 if( new_base != NULL && new_base->is_MergeMem() ) { 4647 MergeMemNode* mdef = new_base->as_MergeMem(); 4648 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4649 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4650 mms.set_memory(mms.memory2()); 4651 } 4652 assert(base_memory() == mdef->base_memory(), ""); 4653 } else { 4654 set_base_memory(new_base); 4655 } 4656 } 4657 4658 // Make a new, untransformed MergeMem with the same base as 'mem'. 4659 // If mem is itself a MergeMem, populate the result with the same edges. 4660 MergeMemNode* MergeMemNode::make(Node* mem) { 4661 return new MergeMemNode(mem); 4662 } 4663 4664 //------------------------------cmp-------------------------------------------- 4665 uint MergeMemNode::hash() const { return NO_HASH; } 4666 bool MergeMemNode::cmp( const Node &n ) const { 4667 return (&n == this); // Always fail except on self 4668 } 4669 4670 //------------------------------Identity--------------------------------------- 4671 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4672 // Identity if this merge point does not record any interesting memory 4673 // disambiguations. 4674 Node* base_mem = base_memory(); 4675 Node* empty_mem = empty_memory(); 4676 if (base_mem != empty_mem) { // Memory path is not dead? 4677 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4678 Node* mem = in(i); 4679 if (mem != empty_mem && mem != base_mem) { 4680 return this; // Many memory splits; no change 4681 } 4682 } 4683 } 4684 return base_mem; // No memory splits; ID on the one true input 4685 } 4686 4687 //------------------------------Ideal------------------------------------------ 4688 // This method is invoked recursively on chains of MergeMem nodes 4689 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4690 // Remove chain'd MergeMems 4691 // 4692 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4693 // relative to the "in(Bot)". Since we are patching both at the same time, 4694 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4695 // but rewrite each "in(i)" relative to the new "in(Bot)". 4696 Node *progress = NULL; 4697 4698 4699 Node* old_base = base_memory(); 4700 Node* empty_mem = empty_memory(); 4701 if (old_base == empty_mem) 4702 return NULL; // Dead memory path. 4703 4704 MergeMemNode* old_mbase; 4705 if (old_base != NULL && old_base->is_MergeMem()) 4706 old_mbase = old_base->as_MergeMem(); 4707 else 4708 old_mbase = NULL; 4709 Node* new_base = old_base; 4710 4711 // simplify stacked MergeMems in base memory 4712 if (old_mbase) new_base = old_mbase->base_memory(); 4713 4714 // the base memory might contribute new slices beyond my req() 4715 if (old_mbase) grow_to_match(old_mbase); 4716 4717 // Look carefully at the base node if it is a phi. 4718 PhiNode* phi_base; 4719 if (new_base != NULL && new_base->is_Phi()) 4720 phi_base = new_base->as_Phi(); 4721 else 4722 phi_base = NULL; 4723 4724 Node* phi_reg = NULL; 4725 uint phi_len = (uint)-1; 4726 if (phi_base != NULL && !phi_base->is_copy()) { 4727 // do not examine phi if degraded to a copy 4728 phi_reg = phi_base->region(); 4729 phi_len = phi_base->req(); 4730 // see if the phi is unfinished 4731 for (uint i = 1; i < phi_len; i++) { 4732 if (phi_base->in(i) == NULL) { 4733 // incomplete phi; do not look at it yet! 4734 phi_reg = NULL; 4735 phi_len = (uint)-1; 4736 break; 4737 } 4738 } 4739 } 4740 4741 // Note: We do not call verify_sparse on entry, because inputs 4742 // can normalize to the base_memory via subsume_node or similar 4743 // mechanisms. This method repairs that damage. 4744 4745 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4746 4747 // Look at each slice. 4748 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4749 Node* old_in = in(i); 4750 // calculate the old memory value 4751 Node* old_mem = old_in; 4752 if (old_mem == empty_mem) old_mem = old_base; 4753 assert(old_mem == memory_at(i), ""); 4754 4755 // maybe update (reslice) the old memory value 4756 4757 // simplify stacked MergeMems 4758 Node* new_mem = old_mem; 4759 MergeMemNode* old_mmem; 4760 if (old_mem != NULL && old_mem->is_MergeMem()) 4761 old_mmem = old_mem->as_MergeMem(); 4762 else 4763 old_mmem = NULL; 4764 if (old_mmem == this) { 4765 // This can happen if loops break up and safepoints disappear. 4766 // A merge of BotPtr (default) with a RawPtr memory derived from a 4767 // safepoint can be rewritten to a merge of the same BotPtr with 4768 // the BotPtr phi coming into the loop. If that phi disappears 4769 // also, we can end up with a self-loop of the mergemem. 4770 // In general, if loops degenerate and memory effects disappear, 4771 // a mergemem can be left looking at itself. This simply means 4772 // that the mergemem's default should be used, since there is 4773 // no longer any apparent effect on this slice. 4774 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4775 // from start. Update the input to TOP. 4776 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4777 } 4778 else if (old_mmem != NULL) { 4779 new_mem = old_mmem->memory_at(i); 4780 } 4781 // else preceding memory was not a MergeMem 4782 4783 // replace equivalent phis (unfortunately, they do not GVN together) 4784 if (new_mem != NULL && new_mem != new_base && 4785 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4786 if (new_mem->is_Phi()) { 4787 PhiNode* phi_mem = new_mem->as_Phi(); 4788 for (uint i = 1; i < phi_len; i++) { 4789 if (phi_base->in(i) != phi_mem->in(i)) { 4790 phi_mem = NULL; 4791 break; 4792 } 4793 } 4794 if (phi_mem != NULL) { 4795 // equivalent phi nodes; revert to the def 4796 new_mem = new_base; 4797 } 4798 } 4799 } 4800 4801 // maybe store down a new value 4802 Node* new_in = new_mem; 4803 if (new_in == new_base) new_in = empty_mem; 4804 4805 if (new_in != old_in) { 4806 // Warning: Do not combine this "if" with the previous "if" 4807 // A memory slice might have be be rewritten even if it is semantically 4808 // unchanged, if the base_memory value has changed. 4809 set_req(i, new_in); 4810 progress = this; // Report progress 4811 } 4812 } 4813 4814 if (new_base != old_base) { 4815 set_req(Compile::AliasIdxBot, new_base); 4816 // Don't use set_base_memory(new_base), because we need to update du. 4817 assert(base_memory() == new_base, ""); 4818 progress = this; 4819 } 4820 4821 if( base_memory() == this ) { 4822 // a self cycle indicates this memory path is dead 4823 set_req(Compile::AliasIdxBot, empty_mem); 4824 } 4825 4826 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4827 // Recursion must occur after the self cycle check above 4828 if( base_memory()->is_MergeMem() ) { 4829 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4830 Node *m = phase->transform(new_mbase); // Rollup any cycles 4831 if( m != NULL && 4832 (m->is_top() || 4833 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4834 // propagate rollup of dead cycle to self 4835 set_req(Compile::AliasIdxBot, empty_mem); 4836 } 4837 } 4838 4839 if( base_memory() == empty_mem ) { 4840 progress = this; 4841 // Cut inputs during Parse phase only. 4842 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4843 if( !can_reshape ) { 4844 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4845 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4846 } 4847 } 4848 } 4849 4850 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4851 // Check if PhiNode::Ideal's "Split phis through memory merges" 4852 // transform should be attempted. Look for this->phi->this cycle. 4853 uint merge_width = req(); 4854 if (merge_width > Compile::AliasIdxRaw) { 4855 PhiNode* phi = base_memory()->as_Phi(); 4856 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4857 if (phi->in(i) == this) { 4858 phase->is_IterGVN()->_worklist.push(phi); 4859 break; 4860 } 4861 } 4862 } 4863 } 4864 4865 assert(progress || verify_sparse(), "please, no dups of base"); 4866 return progress; 4867 } 4868 4869 //-------------------------set_base_memory------------------------------------- 4870 void MergeMemNode::set_base_memory(Node *new_base) { 4871 Node* empty_mem = empty_memory(); 4872 set_req(Compile::AliasIdxBot, new_base); 4873 assert(memory_at(req()) == new_base, "must set default memory"); 4874 // Clear out other occurrences of new_base: 4875 if (new_base != empty_mem) { 4876 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4877 if (in(i) == new_base) set_req(i, empty_mem); 4878 } 4879 } 4880 } 4881 4882 //------------------------------out_RegMask------------------------------------ 4883 const RegMask &MergeMemNode::out_RegMask() const { 4884 return RegMask::Empty; 4885 } 4886 4887 //------------------------------dump_spec-------------------------------------- 4888 #ifndef PRODUCT 4889 void MergeMemNode::dump_spec(outputStream *st) const { 4890 st->print(" {"); 4891 Node* base_mem = base_memory(); 4892 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4893 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4894 if (mem == base_mem) { st->print(" -"); continue; } 4895 st->print( " N%d:", mem->_idx ); 4896 Compile::current()->get_adr_type(i)->dump_on(st); 4897 } 4898 st->print(" }"); 4899 } 4900 #endif // !PRODUCT 4901 4902 4903 #ifdef ASSERT 4904 static bool might_be_same(Node* a, Node* b) { 4905 if (a == b) return true; 4906 if (!(a->is_Phi() || b->is_Phi())) return false; 4907 // phis shift around during optimization 4908 return true; // pretty stupid... 4909 } 4910 4911 // verify a narrow slice (either incoming or outgoing) 4912 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4913 if (!VerifyAliases) return; // don't bother to verify unless requested 4914 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4915 if (Node::in_dump()) return; // muzzle asserts when printing 4916 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4917 assert(n != NULL, ""); 4918 // Elide intervening MergeMem's 4919 while (n->is_MergeMem()) { 4920 n = n->as_MergeMem()->memory_at(alias_idx); 4921 } 4922 Compile* C = Compile::current(); 4923 const TypePtr* n_adr_type = n->adr_type(); 4924 if (n == m->empty_memory()) { 4925 // Implicit copy of base_memory() 4926 } else if (n_adr_type != TypePtr::BOTTOM) { 4927 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4928 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4929 } else { 4930 // A few places like make_runtime_call "know" that VM calls are narrow, 4931 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4932 bool expected_wide_mem = false; 4933 if (n == m->base_memory()) { 4934 expected_wide_mem = true; 4935 } else if (alias_idx == Compile::AliasIdxRaw || 4936 n == m->memory_at(Compile::AliasIdxRaw)) { 4937 expected_wide_mem = true; 4938 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4939 // memory can "leak through" calls on channels that 4940 // are write-once. Allow this also. 4941 expected_wide_mem = true; 4942 } 4943 assert(expected_wide_mem, "expected narrow slice replacement"); 4944 } 4945 } 4946 #else // !ASSERT 4947 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4948 #endif 4949 4950 4951 //-----------------------------memory_at--------------------------------------- 4952 Node* MergeMemNode::memory_at(uint alias_idx) const { 4953 assert(alias_idx >= Compile::AliasIdxRaw || 4954 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4955 "must avoid base_memory and AliasIdxTop"); 4956 4957 // Otherwise, it is a narrow slice. 4958 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4959 Compile *C = Compile::current(); 4960 if (is_empty_memory(n)) { 4961 // the array is sparse; empty slots are the "top" node 4962 n = base_memory(); 4963 assert(Node::in_dump() 4964 || n == NULL || n->bottom_type() == Type::TOP 4965 || n->adr_type() == NULL // address is TOP 4966 || n->adr_type() == TypePtr::BOTTOM 4967 || n->adr_type() == TypeRawPtr::BOTTOM 4968 || Compile::current()->AliasLevel() == 0, 4969 "must be a wide memory"); 4970 // AliasLevel == 0 if we are organizing the memory states manually. 4971 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4972 } else { 4973 // make sure the stored slice is sane 4974 #ifdef ASSERT 4975 if (VMError::is_error_reported() || Node::in_dump()) { 4976 } else if (might_be_same(n, base_memory())) { 4977 // Give it a pass: It is a mostly harmless repetition of the base. 4978 // This can arise normally from node subsumption during optimization. 4979 } else { 4980 verify_memory_slice(this, alias_idx, n); 4981 } 4982 #endif 4983 } 4984 return n; 4985 } 4986 4987 //---------------------------set_memory_at------------------------------------- 4988 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4989 verify_memory_slice(this, alias_idx, n); 4990 Node* empty_mem = empty_memory(); 4991 if (n == base_memory()) n = empty_mem; // collapse default 4992 uint need_req = alias_idx+1; 4993 if (req() < need_req) { 4994 if (n == empty_mem) return; // already the default, so do not grow me 4995 // grow the sparse array 4996 do { 4997 add_req(empty_mem); 4998 } while (req() < need_req); 4999 } 5000 set_req( alias_idx, n ); 5001 } 5002 5003 5004 5005 //--------------------------iteration_setup------------------------------------ 5006 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 5007 if (other != NULL) { 5008 grow_to_match(other); 5009 // invariant: the finite support of mm2 is within mm->req() 5010 #ifdef ASSERT 5011 for (uint i = req(); i < other->req(); i++) { 5012 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 5013 } 5014 #endif 5015 } 5016 // Replace spurious copies of base_memory by top. 5017 Node* base_mem = base_memory(); 5018 if (base_mem != NULL && !base_mem->is_top()) { 5019 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 5020 if (in(i) == base_mem) 5021 set_req(i, empty_memory()); 5022 } 5023 } 5024 } 5025 5026 //---------------------------grow_to_match------------------------------------- 5027 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 5028 Node* empty_mem = empty_memory(); 5029 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 5030 // look for the finite support of the other memory 5031 for (uint i = other->req(); --i >= req(); ) { 5032 if (other->in(i) != empty_mem) { 5033 uint new_len = i+1; 5034 while (req() < new_len) add_req(empty_mem); 5035 break; 5036 } 5037 } 5038 } 5039 5040 //---------------------------verify_sparse------------------------------------- 5041 #ifndef PRODUCT 5042 bool MergeMemNode::verify_sparse() const { 5043 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 5044 Node* base_mem = base_memory(); 5045 // The following can happen in degenerate cases, since empty==top. 5046 if (is_empty_memory(base_mem)) return true; 5047 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5048 assert(in(i) != NULL, "sane slice"); 5049 if (in(i) == base_mem) return false; // should have been the sentinel value! 5050 } 5051 return true; 5052 } 5053 5054 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 5055 Node* n; 5056 n = mm->in(idx); 5057 if (mem == n) return true; // might be empty_memory() 5058 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 5059 if (mem == n) return true; 5060 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 5061 if (mem == n) return true; 5062 if (n == NULL) break; 5063 } 5064 return false; 5065 } 5066 #endif // !PRODUCT