1 /* 2 * Copyright (c) 1997, 2017, 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 "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/arraycopynode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/compile.hpp" 35 #include "opto/connode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/loopnode.hpp" 38 #include "opto/machnode.hpp" 39 #include "opto/matcher.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/mulnode.hpp" 42 #include "opto/narrowptrnode.hpp" 43 #include "opto/phaseX.hpp" 44 #include "opto/regmask.hpp" 45 #include "utilities/copy.hpp" 46 47 // Portions of code courtesy of Clifford Click 48 49 // Optimization - Graph Style 50 51 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 52 53 //============================================================================= 54 uint MemNode::size_of() const { return sizeof(*this); } 55 56 const TypePtr *MemNode::adr_type() const { 57 Node* adr = in(Address); 58 if (adr == NULL) return NULL; // node is dead 59 const TypePtr* cross_check = NULL; 60 DEBUG_ONLY(cross_check = _adr_type); 61 return calculate_adr_type(adr->bottom_type(), cross_check); 62 } 63 64 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 65 if (adr != NULL) { 66 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 67 return true; 68 } 69 } 70 return false; 71 } 72 73 #ifndef PRODUCT 74 void MemNode::dump_spec(outputStream *st) const { 75 if (in(Address) == NULL) return; // node is dead 76 #ifndef ASSERT 77 // fake the missing field 78 const TypePtr* _adr_type = NULL; 79 if (in(Address) != NULL) 80 _adr_type = in(Address)->bottom_type()->isa_ptr(); 81 #endif 82 dump_adr_type(this, _adr_type, st); 83 84 Compile* C = Compile::current(); 85 if (C->alias_type(_adr_type)->is_volatile()) { 86 st->print(" Volatile!"); 87 } 88 if (_unaligned_access) { 89 st->print(" unaligned"); 90 } 91 if (_mismatched_access) { 92 st->print(" mismatched"); 93 } 94 } 95 96 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 97 st->print(" @"); 98 if (adr_type == NULL) { 99 st->print("NULL"); 100 } else { 101 adr_type->dump_on(st); 102 Compile* C = Compile::current(); 103 Compile::AliasType* atp = NULL; 104 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 105 if (atp == NULL) 106 st->print(", idx=?\?;"); 107 else if (atp->index() == Compile::AliasIdxBot) 108 st->print(", idx=Bot;"); 109 else if (atp->index() == Compile::AliasIdxTop) 110 st->print(", idx=Top;"); 111 else if (atp->index() == Compile::AliasIdxRaw) 112 st->print(", idx=Raw;"); 113 else { 114 ciField* field = atp->field(); 115 if (field) { 116 st->print(", name="); 117 field->print_name_on(st); 118 } 119 st->print(", idx=%d;", atp->index()); 120 } 121 } 122 } 123 124 extern void print_alias_types(); 125 126 #endif 127 128 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 129 assert((t_oop != NULL), "sanity"); 130 bool is_instance = t_oop->is_known_instance_field(); 131 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 132 (load != NULL) && load->is_Load() && 133 (phase->is_IterGVN() != NULL); 134 if (!(is_instance || is_boxed_value_load)) 135 return mchain; // don't try to optimize non-instance types 136 uint instance_id = t_oop->instance_id(); 137 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 138 Node *prev = NULL; 139 Node *result = mchain; 140 while (prev != result) { 141 prev = result; 142 if (result == start_mem) 143 break; // hit one of our sentinels 144 // skip over a call which does not affect this memory slice 145 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 146 Node *proj_in = result->in(0); 147 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 148 break; // hit one of our sentinels 149 } else if (proj_in->is_Call()) { 150 // ArrayCopyNodes processed here as well 151 CallNode *call = proj_in->as_Call(); 152 if (!call->may_modify(t_oop, phase)) { // returns false for instances 153 result = call->in(TypeFunc::Memory); 154 } 155 } else if (proj_in->is_Initialize()) { 156 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 157 // Stop if this is the initialization for the object instance which 158 // contains this memory slice, otherwise skip over it. 159 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 160 break; 161 } 162 if (is_instance) { 163 result = proj_in->in(TypeFunc::Memory); 164 } else if (is_boxed_value_load) { 165 Node* klass = alloc->in(AllocateNode::KlassNode); 166 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 167 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 168 result = proj_in->in(TypeFunc::Memory); // not related allocation 169 } 170 } 171 } else if (proj_in->is_MemBar()) { 172 ArrayCopyNode* ac = NULL; 173 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 174 break; 175 } 176 result = proj_in->in(TypeFunc::Memory); 177 } else { 178 assert(false, "unexpected projection"); 179 } 180 } else if (result->is_ClearArray()) { 181 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 182 // Can not bypass initialization of the instance 183 // we are looking for. 184 break; 185 } 186 // Otherwise skip it (the call updated 'result' value). 187 } else if (result->is_MergeMem()) { 188 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 189 } 190 } 191 return result; 192 } 193 194 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 195 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 196 if (t_oop == NULL) 197 return mchain; // don't try to optimize non-oop types 198 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 199 bool is_instance = t_oop->is_known_instance_field(); 200 PhaseIterGVN *igvn = phase->is_IterGVN(); 201 if (is_instance && igvn != NULL && result->is_Phi()) { 202 PhiNode *mphi = result->as_Phi(); 203 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 204 const TypePtr *t = mphi->adr_type(); 205 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 206 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 207 t->is_oopptr()->cast_to_exactness(true) 208 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 209 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) { 210 // clone the Phi with our address type 211 result = mphi->split_out_instance(t_adr, igvn); 212 } else { 213 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 214 } 215 } 216 return result; 217 } 218 219 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 220 uint alias_idx = phase->C->get_alias_index(tp); 221 Node *mem = mmem; 222 #ifdef ASSERT 223 { 224 // Check that current type is consistent with the alias index used during graph construction 225 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 226 bool consistent = adr_check == NULL || adr_check->empty() || 227 phase->C->must_alias(adr_check, alias_idx ); 228 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 229 if( !consistent && adr_check != NULL && !adr_check->empty() && 230 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 231 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 232 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 233 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 234 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 235 // don't assert if it is dead code. 236 consistent = true; 237 } 238 if( !consistent ) { 239 st->print("alias_idx==%d, adr_check==", alias_idx); 240 if( adr_check == NULL ) { 241 st->print("NULL"); 242 } else { 243 adr_check->dump(); 244 } 245 st->cr(); 246 print_alias_types(); 247 assert(consistent, "adr_check must match alias idx"); 248 } 249 } 250 #endif 251 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 252 // means an array I have not precisely typed yet. Do not do any 253 // alias stuff with it any time soon. 254 const TypeOopPtr *toop = tp->isa_oopptr(); 255 if( tp->base() != Type::AnyPtr && 256 !(toop && 257 toop->klass() != NULL && 258 toop->klass()->is_java_lang_Object() && 259 toop->offset() == Type::OffsetBot) ) { 260 // compress paths and change unreachable cycles to TOP 261 // If not, we can update the input infinitely along a MergeMem cycle 262 // Equivalent code in PhiNode::Ideal 263 Node* m = phase->transform(mmem); 264 // If transformed to a MergeMem, get the desired slice 265 // Otherwise the returned node represents memory for every slice 266 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 267 // Update input if it is progress over what we have now 268 } 269 return mem; 270 } 271 272 //--------------------------Ideal_common--------------------------------------- 273 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 274 // Unhook non-raw memories from complete (macro-expanded) initializations. 275 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 276 // If our control input is a dead region, kill all below the region 277 Node *ctl = in(MemNode::Control); 278 if (ctl && remove_dead_region(phase, can_reshape)) 279 return this; 280 ctl = in(MemNode::Control); 281 // Don't bother trying to transform a dead node 282 if (ctl && ctl->is_top()) return NodeSentinel; 283 284 PhaseIterGVN *igvn = phase->is_IterGVN(); 285 // Wait if control on the worklist. 286 if (ctl && can_reshape && igvn != NULL) { 287 Node* bol = NULL; 288 Node* cmp = NULL; 289 if (ctl->in(0)->is_If()) { 290 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 291 bol = ctl->in(0)->in(1); 292 if (bol->is_Bool()) 293 cmp = ctl->in(0)->in(1)->in(1); 294 } 295 if (igvn->_worklist.member(ctl) || 296 (bol != NULL && igvn->_worklist.member(bol)) || 297 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 298 // This control path may be dead. 299 // Delay this memory node transformation until the control is processed. 300 phase->is_IterGVN()->_worklist.push(this); 301 return NodeSentinel; // caller will return NULL 302 } 303 } 304 // Ignore if memory is dead, or self-loop 305 Node *mem = in(MemNode::Memory); 306 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 307 assert(mem != this, "dead loop in MemNode::Ideal"); 308 309 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 310 // This memory slice may be dead. 311 // Delay this mem node transformation until the memory is processed. 312 phase->is_IterGVN()->_worklist.push(this); 313 return NodeSentinel; // caller will return NULL 314 } 315 316 Node *address = in(MemNode::Address); 317 const Type *t_adr = phase->type(address); 318 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 319 320 if (can_reshape && igvn != NULL && 321 (igvn->_worklist.member(address) || 322 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 323 // The address's base and type may change when the address is processed. 324 // Delay this mem node transformation until the address is processed. 325 phase->is_IterGVN()->_worklist.push(this); 326 return NodeSentinel; // caller will return NULL 327 } 328 329 // Do NOT remove or optimize the next lines: ensure a new alias index 330 // is allocated for an oop pointer type before Escape Analysis. 331 // Note: C++ will not remove it since the call has side effect. 332 if (t_adr->isa_oopptr()) { 333 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 334 } 335 336 Node* base = NULL; 337 if (address->is_AddP()) { 338 base = address->in(AddPNode::Base); 339 } 340 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 341 !t_adr->isa_rawptr()) { 342 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 343 // Skip this node optimization if its address has TOP base. 344 return NodeSentinel; // caller will return NULL 345 } 346 347 // Avoid independent memory operations 348 Node* old_mem = mem; 349 350 // The code which unhooks non-raw memories from complete (macro-expanded) 351 // initializations was removed. After macro-expansion all stores catched 352 // by Initialize node became raw stores and there is no information 353 // which memory slices they modify. So it is unsafe to move any memory 354 // operation above these stores. Also in most cases hooked non-raw memories 355 // were already unhooked by using information from detect_ptr_independence() 356 // and find_previous_store(). 357 358 if (mem->is_MergeMem()) { 359 MergeMemNode* mmem = mem->as_MergeMem(); 360 const TypePtr *tp = t_adr->is_ptr(); 361 362 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 363 } 364 365 if (mem != old_mem) { 366 set_req(MemNode::Memory, mem); 367 if (can_reshape && old_mem->outcnt() == 0) { 368 igvn->_worklist.push(old_mem); 369 } 370 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 371 return this; 372 } 373 374 // let the subclass continue analyzing... 375 return NULL; 376 } 377 378 // Helper function for proving some simple control dominations. 379 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 380 // Already assumes that 'dom' is available at 'sub', and that 'sub' 381 // is not a constant (dominated by the method's StartNode). 382 // Used by MemNode::find_previous_store to prove that the 383 // control input of a memory operation predates (dominates) 384 // an allocation it wants to look past. 385 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 386 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 387 return false; // Conservative answer for dead code 388 389 // Check 'dom'. Skip Proj and CatchProj nodes. 390 dom = dom->find_exact_control(dom); 391 if (dom == NULL || dom->is_top()) 392 return false; // Conservative answer for dead code 393 394 if (dom == sub) { 395 // For the case when, for example, 'sub' is Initialize and the original 396 // 'dom' is Proj node of the 'sub'. 397 return false; 398 } 399 400 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 401 return true; 402 403 // 'dom' dominates 'sub' if its control edge and control edges 404 // of all its inputs dominate or equal to sub's control edge. 405 406 // Currently 'sub' is either Allocate, Initialize or Start nodes. 407 // Or Region for the check in LoadNode::Ideal(); 408 // 'sub' should have sub->in(0) != NULL. 409 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 410 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 411 412 // Get control edge of 'sub'. 413 Node* orig_sub = sub; 414 sub = sub->find_exact_control(sub->in(0)); 415 if (sub == NULL || sub->is_top()) 416 return false; // Conservative answer for dead code 417 418 assert(sub->is_CFG(), "expecting control"); 419 420 if (sub == dom) 421 return true; 422 423 if (sub->is_Start() || sub->is_Root()) 424 return false; 425 426 { 427 // Check all control edges of 'dom'. 428 429 ResourceMark rm; 430 Arena* arena = Thread::current()->resource_area(); 431 Node_List nlist(arena); 432 Unique_Node_List dom_list(arena); 433 434 dom_list.push(dom); 435 bool only_dominating_controls = false; 436 437 for (uint next = 0; next < dom_list.size(); next++) { 438 Node* n = dom_list.at(next); 439 if (n == orig_sub) 440 return false; // One of dom's inputs dominated by sub. 441 if (!n->is_CFG() && n->pinned()) { 442 // Check only own control edge for pinned non-control nodes. 443 n = n->find_exact_control(n->in(0)); 444 if (n == NULL || n->is_top()) 445 return false; // Conservative answer for dead code 446 assert(n->is_CFG(), "expecting control"); 447 dom_list.push(n); 448 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 449 only_dominating_controls = true; 450 } else if (n->is_CFG()) { 451 if (n->dominates(sub, nlist)) 452 only_dominating_controls = true; 453 else 454 return false; 455 } else { 456 // First, own control edge. 457 Node* m = n->find_exact_control(n->in(0)); 458 if (m != NULL) { 459 if (m->is_top()) 460 return false; // Conservative answer for dead code 461 dom_list.push(m); 462 } 463 // Now, the rest of edges. 464 uint cnt = n->req(); 465 for (uint i = 1; i < cnt; i++) { 466 m = n->find_exact_control(n->in(i)); 467 if (m == NULL || m->is_top()) 468 continue; 469 dom_list.push(m); 470 } 471 } 472 } 473 return only_dominating_controls; 474 } 475 } 476 477 //---------------------detect_ptr_independence--------------------------------- 478 // Used by MemNode::find_previous_store to prove that two base 479 // pointers are never equal. 480 // The pointers are accompanied by their associated allocations, 481 // if any, which have been previously discovered by the caller. 482 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 483 Node* p2, AllocateNode* a2, 484 PhaseTransform* phase) { 485 // Attempt to prove that these two pointers cannot be aliased. 486 // They may both manifestly be allocations, and they should differ. 487 // Or, if they are not both allocations, they can be distinct constants. 488 // Otherwise, one is an allocation and the other a pre-existing value. 489 if (a1 == NULL && a2 == NULL) { // neither an allocation 490 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 491 } else if (a1 != NULL && a2 != NULL) { // both allocations 492 return (a1 != a2); 493 } else if (a1 != NULL) { // one allocation a1 494 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 495 return all_controls_dominate(p2, a1); 496 } else { //(a2 != NULL) // one allocation a2 497 return all_controls_dominate(p1, a2); 498 } 499 return false; 500 } 501 502 503 // Find an arraycopy that must have set (can_see_stored_value=true) or 504 // could have set (can_see_stored_value=false) the value for this load 505 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 506 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 507 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 508 Node* mb = mem->in(0); 509 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 510 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 511 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 512 if (ac->is_clonebasic()) { 513 intptr_t offset; 514 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 515 assert(alloc != NULL && (!ReduceBulkZeroing || alloc->initialization()->is_complete_with_arraycopy()), "broken allocation"); 516 if (alloc == ld_alloc) { 517 return ac; 518 } 519 } 520 } 521 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 522 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 523 524 if (ac->is_arraycopy_validated() || 525 ac->is_copyof_validated() || 526 ac->is_copyofrange_validated()) { 527 Node* ld_addp = in(MemNode::Address); 528 if (ld_addp->is_AddP()) { 529 Node* ld_base = ld_addp->in(AddPNode::Address); 530 Node* ld_offs = ld_addp->in(AddPNode::Offset); 531 532 Node* dest = ac->in(ArrayCopyNode::Dest); 533 534 if (dest == ld_base) { 535 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 536 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 537 return ac; 538 } 539 if (!can_see_stored_value) { 540 mem = ac->in(TypeFunc::Memory); 541 } 542 } 543 } 544 } 545 } 546 return NULL; 547 } 548 549 // The logic for reordering loads and stores uses four steps: 550 // (a) Walk carefully past stores and initializations which we 551 // can prove are independent of this load. 552 // (b) Observe that the next memory state makes an exact match 553 // with self (load or store), and locate the relevant store. 554 // (c) Ensure that, if we were to wire self directly to the store, 555 // the optimizer would fold it up somehow. 556 // (d) Do the rewiring, and return, depending on some other part of 557 // the optimizer to fold up the load. 558 // This routine handles steps (a) and (b). Steps (c) and (d) are 559 // specific to loads and stores, so they are handled by the callers. 560 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 561 // 562 Node* MemNode::find_previous_store(PhaseTransform* phase) { 563 Node* ctrl = in(MemNode::Control); 564 Node* adr = in(MemNode::Address); 565 intptr_t offset = 0; 566 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 567 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 568 569 if (offset == Type::OffsetBot) 570 return NULL; // cannot unalias unless there are precise offsets 571 572 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 573 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 574 575 intptr_t size_in_bytes = memory_size(); 576 577 Node* mem = in(MemNode::Memory); // start searching here... 578 579 int cnt = 50; // Cycle limiter 580 for (;;) { // While we can dance past unrelated stores... 581 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 582 583 Node* prev = mem; 584 if (mem->is_Store()) { 585 Node* st_adr = mem->in(MemNode::Address); 586 intptr_t st_offset = 0; 587 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 588 if (st_base == NULL) 589 break; // inscrutable pointer 590 591 // For raw accesses it's not enough to prove that constant offsets don't intersect. 592 // We need the bases to be the equal in order for the offset check to make sense. 593 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 594 break; 595 } 596 597 if (st_offset != offset && st_offset != Type::OffsetBot) { 598 const int MAX_STORE = BytesPerLong; 599 if (st_offset >= offset + size_in_bytes || 600 st_offset <= offset - MAX_STORE || 601 st_offset <= offset - mem->as_Store()->memory_size()) { 602 // Success: The offsets are provably independent. 603 // (You may ask, why not just test st_offset != offset and be done? 604 // The answer is that stores of different sizes can co-exist 605 // in the same sequence of RawMem effects. We sometimes initialize 606 // a whole 'tile' of array elements with a single jint or jlong.) 607 mem = mem->in(MemNode::Memory); 608 continue; // (a) advance through independent store memory 609 } 610 } 611 if (st_base != base && 612 detect_ptr_independence(base, alloc, 613 st_base, 614 AllocateNode::Ideal_allocation(st_base, phase), 615 phase)) { 616 // Success: The bases are provably independent. 617 mem = mem->in(MemNode::Memory); 618 continue; // (a) advance through independent store memory 619 } 620 621 // (b) At this point, if the bases or offsets do not agree, we lose, 622 // since we have not managed to prove 'this' and 'mem' independent. 623 if (st_base == base && st_offset == offset) { 624 return mem; // let caller handle steps (c), (d) 625 } 626 627 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 628 InitializeNode* st_init = mem->in(0)->as_Initialize(); 629 AllocateNode* st_alloc = st_init->allocation(); 630 if (st_alloc == NULL) 631 break; // something degenerated 632 bool known_identical = false; 633 bool known_independent = false; 634 if (alloc == st_alloc) 635 known_identical = true; 636 else if (alloc != NULL) 637 known_independent = true; 638 else if (all_controls_dominate(this, st_alloc)) 639 known_independent = true; 640 641 if (known_independent) { 642 // The bases are provably independent: Either they are 643 // manifestly distinct allocations, or else the control 644 // of this load dominates the store's allocation. 645 int alias_idx = phase->C->get_alias_index(adr_type()); 646 if (alias_idx == Compile::AliasIdxRaw) { 647 mem = st_alloc->in(TypeFunc::Memory); 648 } else { 649 mem = st_init->memory(alias_idx); 650 } 651 continue; // (a) advance through independent store memory 652 } 653 654 // (b) at this point, if we are not looking at a store initializing 655 // the same allocation we are loading from, we lose. 656 if (known_identical) { 657 // From caller, can_see_stored_value will consult find_captured_store. 658 return mem; // let caller handle steps (c), (d) 659 } 660 661 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 662 if (prev != mem) { 663 // Found an arraycopy but it doesn't affect that load 664 continue; 665 } 666 // Found an arraycopy that may affect that load 667 return mem; 668 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 669 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 670 if (mem->is_Proj() && mem->in(0)->is_Call()) { 671 // ArrayCopyNodes processed here as well. 672 CallNode *call = mem->in(0)->as_Call(); 673 if (!call->may_modify(addr_t, phase)) { 674 mem = call->in(TypeFunc::Memory); 675 continue; // (a) advance through independent call memory 676 } 677 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 678 ArrayCopyNode* ac = NULL; 679 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 680 break; 681 } 682 mem = mem->in(0)->in(TypeFunc::Memory); 683 continue; // (a) advance through independent MemBar memory 684 } else if (mem->is_ClearArray()) { 685 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 686 // (the call updated 'mem' value) 687 continue; // (a) advance through independent allocation memory 688 } else { 689 // Can not bypass initialization of the instance 690 // we are looking for. 691 return mem; 692 } 693 } else if (mem->is_MergeMem()) { 694 int alias_idx = phase->C->get_alias_index(adr_type()); 695 mem = mem->as_MergeMem()->memory_at(alias_idx); 696 continue; // (a) advance through independent MergeMem memory 697 } 698 } 699 700 // Unless there is an explicit 'continue', we must bail out here, 701 // because 'mem' is an inscrutable memory state (e.g., a call). 702 break; 703 } 704 705 return NULL; // bail out 706 } 707 708 //----------------------calculate_adr_type------------------------------------- 709 // Helper function. Notices when the given type of address hits top or bottom. 710 // Also, asserts a cross-check of the type against the expected address type. 711 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 712 if (t == Type::TOP) return NULL; // does not touch memory any more? 713 #ifdef PRODUCT 714 cross_check = NULL; 715 #else 716 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 717 #endif 718 const TypePtr* tp = t->isa_ptr(); 719 if (tp == NULL) { 720 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 721 return TypePtr::BOTTOM; // touches lots of memory 722 } else { 723 #ifdef ASSERT 724 // %%%% [phh] We don't check the alias index if cross_check is 725 // TypeRawPtr::BOTTOM. Needs to be investigated. 726 if (cross_check != NULL && 727 cross_check != TypePtr::BOTTOM && 728 cross_check != TypeRawPtr::BOTTOM) { 729 // Recheck the alias index, to see if it has changed (due to a bug). 730 Compile* C = Compile::current(); 731 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 732 "must stay in the original alias category"); 733 // The type of the address must be contained in the adr_type, 734 // disregarding "null"-ness. 735 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 736 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 737 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 738 "real address must not escape from expected memory type"); 739 } 740 #endif 741 return tp; 742 } 743 } 744 745 //============================================================================= 746 // Should LoadNode::Ideal() attempt to remove control edges? 747 bool LoadNode::can_remove_control() const { 748 return true; 749 } 750 uint LoadNode::size_of() const { return sizeof(*this); } 751 uint LoadNode::cmp( const Node &n ) const 752 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 753 const Type *LoadNode::bottom_type() const { return _type; } 754 uint LoadNode::ideal_reg() const { 755 return _type->ideal_reg(); 756 } 757 758 #ifndef PRODUCT 759 void LoadNode::dump_spec(outputStream *st) const { 760 MemNode::dump_spec(st); 761 if( !Verbose && !WizardMode ) { 762 // standard dump does this in Verbose and WizardMode 763 st->print(" #"); _type->dump_on(st); 764 } 765 if (!depends_only_on_test()) { 766 st->print(" (does not depend only on test)"); 767 } 768 } 769 #endif 770 771 #ifdef ASSERT 772 //----------------------------is_immutable_value------------------------------- 773 // Helper function to allow a raw load without control edge for some cases 774 bool LoadNode::is_immutable_value(Node* adr) { 775 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 776 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 777 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 778 in_bytes(JavaThread::osthread_offset()))); 779 } 780 #endif 781 782 //----------------------------LoadNode::make----------------------------------- 783 // Polymorphic factory method: 784 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 785 ControlDependency control_dependency, bool unaligned, bool mismatched) { 786 Compile* C = gvn.C; 787 788 // sanity check the alias category against the created node type 789 assert(!(adr_type->isa_oopptr() && 790 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 791 "use LoadKlassNode instead"); 792 assert(!(adr_type->isa_aryptr() && 793 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 794 "use LoadRangeNode instead"); 795 // Check control edge of raw loads 796 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 797 // oop will be recorded in oop map if load crosses safepoint 798 rt->isa_oopptr() || is_immutable_value(adr), 799 "raw memory operations should have control edge"); 800 LoadNode* load = NULL; 801 switch (bt) { 802 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 803 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 804 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 805 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 806 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 807 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 808 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 809 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 810 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 811 case T_OBJECT: 812 #ifdef _LP64 813 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 814 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 815 } else 816 #endif 817 { 818 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 819 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 820 } 821 break; 822 default: 823 // ShouldNotReachHere(); ??? 824 break; 825 } 826 assert(load != NULL, "LoadNode should have been created"); 827 if (unaligned) { 828 load->set_unaligned_access(); 829 } 830 if (mismatched) { 831 load->set_mismatched_access(); 832 } 833 if (load->Opcode() == Op_LoadN) { 834 Node* ld = gvn.transform(load); 835 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 836 } 837 838 return load; 839 } 840 841 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 842 ControlDependency control_dependency, bool unaligned, bool mismatched) { 843 bool require_atomic = true; 844 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 845 if (unaligned) { 846 load->set_unaligned_access(); 847 } 848 if (mismatched) { 849 load->set_mismatched_access(); 850 } 851 return load; 852 } 853 854 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 855 ControlDependency control_dependency, bool unaligned, bool mismatched) { 856 bool require_atomic = true; 857 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 858 if (unaligned) { 859 load->set_unaligned_access(); 860 } 861 if (mismatched) { 862 load->set_mismatched_access(); 863 } 864 return load; 865 } 866 867 868 869 //------------------------------hash------------------------------------------- 870 uint LoadNode::hash() const { 871 // unroll addition of interesting fields 872 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 873 } 874 875 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 876 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 877 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 878 bool is_stable_ary = FoldStableValues && 879 (tp != NULL) && (tp->isa_aryptr() != NULL) && 880 tp->isa_aryptr()->is_stable(); 881 882 return (eliminate_boxing && non_volatile) || is_stable_ary; 883 } 884 885 return false; 886 } 887 888 // Is the value loaded previously stored by an arraycopy? If so return 889 // a load node that reads from the source array so we may be able to 890 // optimize out the ArrayCopy node later. 891 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const { 892 Node* ld_adr = in(MemNode::Address); 893 intptr_t ld_off = 0; 894 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 895 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 896 if (ac != NULL) { 897 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 898 899 Node* ld = clone(); 900 if (ac->as_ArrayCopy()->is_clonebasic()) { 901 assert(ld_alloc != NULL, "need an alloc"); 902 Node* addp = in(MemNode::Address)->clone(); 903 assert(addp->is_AddP(), "address must be addp"); 904 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern"); 905 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern"); 906 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base)); 907 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address)); 908 ld->set_req(MemNode::Address, phase->transform(addp)); 909 if (in(0) != NULL) { 910 assert(ld_alloc->in(0) != NULL, "alloc must have control"); 911 ld->set_req(0, ld_alloc->in(0)); 912 } 913 } else { 914 Node* src = ac->in(ArrayCopyNode::Src); 915 Node* addp = in(MemNode::Address)->clone(); 916 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 917 addp->set_req(AddPNode::Base, src); 918 addp->set_req(AddPNode::Address, src); 919 920 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 921 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 922 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 923 uint shift = exact_log2(type2aelembytes(ary_elem)); 924 925 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 926 #ifdef _LP64 927 diff = phase->transform(new ConvI2LNode(diff)); 928 #endif 929 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 930 931 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 932 addp->set_req(AddPNode::Offset, offset); 933 ld->set_req(MemNode::Address, phase->transform(addp)); 934 935 const TypeX *ld_offs_t = phase->type(offset)->isa_intptr_t(); 936 937 if (!ac->as_ArrayCopy()->can_replace_dest_load_with_src_load(ld_offs_t->_lo, ld_offs_t->_hi, phase)) { 938 return NULL; 939 } 940 941 if (in(0) != NULL) { 942 assert(ac->in(0) != NULL, "alloc must have control"); 943 ld->set_req(0, ac->in(0)); 944 } 945 } 946 // load depends on the tests that validate the arraycopy 947 ld->as_Load()->_control_dependency = Pinned; 948 return ld; 949 } 950 return NULL; 951 } 952 953 954 //---------------------------can_see_stored_value------------------------------ 955 // This routine exists to make sure this set of tests is done the same 956 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 957 // will change the graph shape in a way which makes memory alive twice at the 958 // same time (uses the Oracle model of aliasing), then some 959 // LoadXNode::Identity will fold things back to the equivalence-class model 960 // of aliasing. 961 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 962 Node* ld_adr = in(MemNode::Address); 963 intptr_t ld_off = 0; 964 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 965 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 966 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 967 // This is more general than load from boxing objects. 968 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 969 uint alias_idx = atp->index(); 970 bool final = !atp->is_rewritable(); 971 Node* result = NULL; 972 Node* current = st; 973 // Skip through chains of MemBarNodes checking the MergeMems for 974 // new states for the slice of this load. Stop once any other 975 // kind of node is encountered. Loads from final memory can skip 976 // through any kind of MemBar but normal loads shouldn't skip 977 // through MemBarAcquire since the could allow them to move out of 978 // a synchronized region. 979 while (current->is_Proj()) { 980 int opc = current->in(0)->Opcode(); 981 if ((final && (opc == Op_MemBarAcquire || 982 opc == Op_MemBarAcquireLock || 983 opc == Op_LoadFence)) || 984 opc == Op_MemBarRelease || 985 opc == Op_StoreFence || 986 opc == Op_MemBarReleaseLock || 987 opc == Op_MemBarStoreStore || 988 opc == Op_MemBarCPUOrder) { 989 Node* mem = current->in(0)->in(TypeFunc::Memory); 990 if (mem->is_MergeMem()) { 991 MergeMemNode* merge = mem->as_MergeMem(); 992 Node* new_st = merge->memory_at(alias_idx); 993 if (new_st == merge->base_memory()) { 994 // Keep searching 995 current = new_st; 996 continue; 997 } 998 // Save the new memory state for the slice and fall through 999 // to exit. 1000 result = new_st; 1001 } 1002 } 1003 break; 1004 } 1005 if (result != NULL) { 1006 st = result; 1007 } 1008 } 1009 1010 // Loop around twice in the case Load -> Initialize -> Store. 1011 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1012 for (int trip = 0; trip <= 1; trip++) { 1013 1014 if (st->is_Store()) { 1015 Node* st_adr = st->in(MemNode::Address); 1016 if (!phase->eqv(st_adr, ld_adr)) { 1017 // Try harder before giving up... Match raw and non-raw pointers. 1018 intptr_t st_off = 0; 1019 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 1020 if (alloc == NULL) return NULL; 1021 if (alloc != ld_alloc) return NULL; 1022 if (ld_off != st_off) return NULL; 1023 // At this point we have proven something like this setup: 1024 // A = Allocate(...) 1025 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 1026 // S = StoreQ(, AddP(, A.Parm , #Off), V) 1027 // (Actually, we haven't yet proven the Q's are the same.) 1028 // In other words, we are loading from a casted version of 1029 // the same pointer-and-offset that we stored to. 1030 // Thus, we are able to replace L by V. 1031 } 1032 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1033 if (store_Opcode() != st->Opcode()) 1034 return NULL; 1035 return st->in(MemNode::ValueIn); 1036 } 1037 1038 // A load from a freshly-created object always returns zero. 1039 // (This can happen after LoadNode::Ideal resets the load's memory input 1040 // to find_captured_store, which returned InitializeNode::zero_memory.) 1041 if (st->is_Proj() && st->in(0)->is_Allocate() && 1042 (st->in(0) == ld_alloc) && 1043 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1044 // return a zero value for the load's basic type 1045 // (This is one of the few places where a generic PhaseTransform 1046 // can create new nodes. Think of it as lazily manifesting 1047 // virtually pre-existing constants.) 1048 return phase->zerocon(memory_type()); 1049 } 1050 1051 // A load from an initialization barrier can match a captured store. 1052 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1053 InitializeNode* init = st->in(0)->as_Initialize(); 1054 AllocateNode* alloc = init->allocation(); 1055 if ((alloc != NULL) && (alloc == ld_alloc)) { 1056 // examine a captured store value 1057 st = init->find_captured_store(ld_off, memory_size(), phase); 1058 if (st != NULL) { 1059 continue; // take one more trip around 1060 } 1061 } 1062 } 1063 1064 // Load boxed value from result of valueOf() call is input parameter. 1065 if (this->is_Load() && ld_adr->is_AddP() && 1066 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1067 intptr_t ignore = 0; 1068 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1069 if (base != NULL && base->is_Proj() && 1070 base->as_Proj()->_con == TypeFunc::Parms && 1071 base->in(0)->is_CallStaticJava() && 1072 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1073 return base->in(0)->in(TypeFunc::Parms); 1074 } 1075 } 1076 1077 break; 1078 } 1079 1080 return NULL; 1081 } 1082 1083 //----------------------is_instance_field_load_with_local_phi------------------ 1084 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1085 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1086 in(Address)->is_AddP() ) { 1087 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1088 // Only instances and boxed values. 1089 if( t_oop != NULL && 1090 (t_oop->is_ptr_to_boxed_value() || 1091 t_oop->is_known_instance_field()) && 1092 t_oop->offset() != Type::OffsetBot && 1093 t_oop->offset() != Type::OffsetTop) { 1094 return true; 1095 } 1096 } 1097 return false; 1098 } 1099 1100 //------------------------------Identity--------------------------------------- 1101 // Loads are identity if previous store is to same address 1102 Node* LoadNode::Identity(PhaseGVN* phase) { 1103 // If the previous store-maker is the right kind of Store, and the store is 1104 // to the same address, then we are equal to the value stored. 1105 Node* mem = in(Memory); 1106 Node* value = can_see_stored_value(mem, phase); 1107 if( value ) { 1108 // byte, short & char stores truncate naturally. 1109 // A load has to load the truncated value which requires 1110 // some sort of masking operation and that requires an 1111 // Ideal call instead of an Identity call. 1112 if (memory_size() < BytesPerInt) { 1113 // If the input to the store does not fit with the load's result type, 1114 // it must be truncated via an Ideal call. 1115 if (!phase->type(value)->higher_equal(phase->type(this))) 1116 return this; 1117 } 1118 // (This works even when value is a Con, but LoadNode::Value 1119 // usually runs first, producing the singleton type of the Con.) 1120 return value; 1121 } 1122 1123 // Search for an existing data phi which was generated before for the same 1124 // instance's field to avoid infinite generation of phis in a loop. 1125 Node *region = mem->in(0); 1126 if (is_instance_field_load_with_local_phi(region)) { 1127 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1128 int this_index = phase->C->get_alias_index(addr_t); 1129 int this_offset = addr_t->offset(); 1130 int this_iid = addr_t->instance_id(); 1131 if (!addr_t->is_known_instance() && 1132 addr_t->is_ptr_to_boxed_value()) { 1133 // Use _idx of address base (could be Phi node) for boxed values. 1134 intptr_t ignore = 0; 1135 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1136 if (base == NULL) { 1137 return this; 1138 } 1139 this_iid = base->_idx; 1140 } 1141 const Type* this_type = bottom_type(); 1142 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1143 Node* phi = region->fast_out(i); 1144 if (phi->is_Phi() && phi != mem && 1145 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1146 return phi; 1147 } 1148 } 1149 } 1150 1151 return this; 1152 } 1153 1154 // Construct an equivalent unsigned load. 1155 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1156 BasicType bt = T_ILLEGAL; 1157 const Type* rt = NULL; 1158 switch (Opcode()) { 1159 case Op_LoadUB: return this; 1160 case Op_LoadUS: return this; 1161 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1162 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1163 default: 1164 assert(false, "no unsigned variant: %s", Name()); 1165 return NULL; 1166 } 1167 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1168 raw_adr_type(), rt, bt, _mo, _control_dependency, 1169 is_unaligned_access(), is_mismatched_access()); 1170 } 1171 1172 // Construct an equivalent signed load. 1173 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1174 BasicType bt = T_ILLEGAL; 1175 const Type* rt = NULL; 1176 switch (Opcode()) { 1177 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1178 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1179 case Op_LoadB: // fall through 1180 case Op_LoadS: // fall through 1181 case Op_LoadI: // fall through 1182 case Op_LoadL: return this; 1183 default: 1184 assert(false, "no signed variant: %s", Name()); 1185 return NULL; 1186 } 1187 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1188 raw_adr_type(), rt, bt, _mo, _control_dependency, 1189 is_unaligned_access(), is_mismatched_access()); 1190 } 1191 1192 // We're loading from an object which has autobox behaviour. 1193 // If this object is result of a valueOf call we'll have a phi 1194 // merging a newly allocated object and a load from the cache. 1195 // We want to replace this load with the original incoming 1196 // argument to the valueOf call. 1197 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1198 assert(phase->C->eliminate_boxing(), "sanity"); 1199 intptr_t ignore = 0; 1200 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1201 if ((base == NULL) || base->is_Phi()) { 1202 // Push the loads from the phi that comes from valueOf up 1203 // through it to allow elimination of the loads and the recovery 1204 // of the original value. It is done in split_through_phi(). 1205 return NULL; 1206 } else if (base->is_Load() || 1207 (base->is_DecodeN() && base->in(1)->is_Load())) { 1208 // Eliminate the load of boxed value for integer types from the cache 1209 // array by deriving the value from the index into the array. 1210 // Capture the offset of the load and then reverse the computation. 1211 1212 // Get LoadN node which loads a boxing object from 'cache' array. 1213 if (base->is_DecodeN()) { 1214 base = base->in(1); 1215 } 1216 if (!base->in(Address)->is_AddP()) { 1217 return NULL; // Complex address 1218 } 1219 AddPNode* address = base->in(Address)->as_AddP(); 1220 Node* cache_base = address->in(AddPNode::Base); 1221 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1222 // Get ConP node which is static 'cache' field. 1223 cache_base = cache_base->in(1); 1224 } 1225 if ((cache_base != NULL) && cache_base->is_Con()) { 1226 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1227 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1228 Node* elements[4]; 1229 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1230 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1231 if (count > 0 && elements[0]->is_Con() && 1232 (count == 1 || 1233 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1234 elements[1]->in(2) == phase->intcon(shift)))) { 1235 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1236 // Fetch the box object cache[0] at the base of the array and get its value 1237 ciInstance* box = array->obj_at(0)->as_instance(); 1238 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1239 assert(ik->is_box_klass(), "sanity"); 1240 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1241 if (ik->nof_nonstatic_fields() == 1) { 1242 // This should be true nonstatic_field_at requires calling 1243 // nof_nonstatic_fields so check it anyway 1244 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1245 BasicType bt = c.basic_type(); 1246 // Only integer types have boxing cache. 1247 assert(bt == T_BOOLEAN || bt == T_CHAR || 1248 bt == T_BYTE || bt == T_SHORT || 1249 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1250 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1251 if (cache_low != (int)cache_low) { 1252 return NULL; // should not happen since cache is array indexed by value 1253 } 1254 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1255 if (offset != (int)offset) { 1256 return NULL; // should not happen since cache is array indexed by value 1257 } 1258 // Add up all the offsets making of the address of the load 1259 Node* result = elements[0]; 1260 for (int i = 1; i < count; i++) { 1261 result = phase->transform(new AddXNode(result, elements[i])); 1262 } 1263 // Remove the constant offset from the address and then 1264 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1265 // remove the scaling of the offset to recover the original index. 1266 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1267 // Peel the shift off directly but wrap it in a dummy node 1268 // since Ideal can't return existing nodes 1269 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1270 } else if (result->is_Add() && result->in(2)->is_Con() && 1271 result->in(1)->Opcode() == Op_LShiftX && 1272 result->in(1)->in(2) == phase->intcon(shift)) { 1273 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1274 // but for boxing cache access we know that X<<Z will not overflow 1275 // (there is range check) so we do this optimizatrion by hand here. 1276 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1277 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1278 } else { 1279 result = new RShiftXNode(result, phase->intcon(shift)); 1280 } 1281 #ifdef _LP64 1282 if (bt != T_LONG) { 1283 result = new ConvL2INode(phase->transform(result)); 1284 } 1285 #else 1286 if (bt == T_LONG) { 1287 result = new ConvI2LNode(phase->transform(result)); 1288 } 1289 #endif 1290 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1291 // Need to preserve unboxing load type if it is unsigned. 1292 switch(this->Opcode()) { 1293 case Op_LoadUB: 1294 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1295 break; 1296 case Op_LoadUS: 1297 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1298 break; 1299 } 1300 return result; 1301 } 1302 } 1303 } 1304 } 1305 } 1306 return NULL; 1307 } 1308 1309 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1310 Node* region = phi->in(0); 1311 if (region == NULL) { 1312 return false; // Wait stable graph 1313 } 1314 uint cnt = phi->req(); 1315 for (uint i = 1; i < cnt; i++) { 1316 Node* rc = region->in(i); 1317 if (rc == NULL || phase->type(rc) == Type::TOP) 1318 return false; // Wait stable graph 1319 Node* in = phi->in(i); 1320 if (in == NULL || phase->type(in) == Type::TOP) 1321 return false; // Wait stable graph 1322 } 1323 return true; 1324 } 1325 //------------------------------split_through_phi------------------------------ 1326 // Split instance or boxed field load through Phi. 1327 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1328 Node* mem = in(Memory); 1329 Node* address = in(Address); 1330 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1331 1332 assert((t_oop != NULL) && 1333 (t_oop->is_known_instance_field() || 1334 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1335 1336 Compile* C = phase->C; 1337 intptr_t ignore = 0; 1338 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1339 bool base_is_phi = (base != NULL) && base->is_Phi(); 1340 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1341 (base != NULL) && (base == address->in(AddPNode::Base)) && 1342 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1343 1344 if (!((mem->is_Phi() || base_is_phi) && 1345 (load_boxed_values || t_oop->is_known_instance_field()))) { 1346 return NULL; // memory is not Phi 1347 } 1348 1349 if (mem->is_Phi()) { 1350 if (!stable_phi(mem->as_Phi(), phase)) { 1351 return NULL; // Wait stable graph 1352 } 1353 uint cnt = mem->req(); 1354 // Check for loop invariant memory. 1355 if (cnt == 3) { 1356 for (uint i = 1; i < cnt; i++) { 1357 Node* in = mem->in(i); 1358 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1359 if (m == mem) { 1360 set_req(Memory, mem->in(cnt - i)); 1361 return this; // made change 1362 } 1363 } 1364 } 1365 } 1366 if (base_is_phi) { 1367 if (!stable_phi(base->as_Phi(), phase)) { 1368 return NULL; // Wait stable graph 1369 } 1370 uint cnt = base->req(); 1371 // Check for loop invariant memory. 1372 if (cnt == 3) { 1373 for (uint i = 1; i < cnt; i++) { 1374 if (base->in(i) == base) { 1375 return NULL; // Wait stable graph 1376 } 1377 } 1378 } 1379 } 1380 1381 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1382 1383 // Split through Phi (see original code in loopopts.cpp). 1384 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1385 1386 // Do nothing here if Identity will find a value 1387 // (to avoid infinite chain of value phis generation). 1388 if (!phase->eqv(this, this->Identity(phase))) 1389 return NULL; 1390 1391 // Select Region to split through. 1392 Node* region; 1393 if (!base_is_phi) { 1394 assert(mem->is_Phi(), "sanity"); 1395 region = mem->in(0); 1396 // Skip if the region dominates some control edge of the address. 1397 if (!MemNode::all_controls_dominate(address, region)) 1398 return NULL; 1399 } else if (!mem->is_Phi()) { 1400 assert(base_is_phi, "sanity"); 1401 region = base->in(0); 1402 // Skip if the region dominates some control edge of the memory. 1403 if (!MemNode::all_controls_dominate(mem, region)) 1404 return NULL; 1405 } else if (base->in(0) != mem->in(0)) { 1406 assert(base_is_phi && mem->is_Phi(), "sanity"); 1407 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1408 region = base->in(0); 1409 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1410 region = mem->in(0); 1411 } else { 1412 return NULL; // complex graph 1413 } 1414 } else { 1415 assert(base->in(0) == mem->in(0), "sanity"); 1416 region = mem->in(0); 1417 } 1418 1419 const Type* this_type = this->bottom_type(); 1420 int this_index = C->get_alias_index(t_oop); 1421 int this_offset = t_oop->offset(); 1422 int this_iid = t_oop->instance_id(); 1423 if (!t_oop->is_known_instance() && load_boxed_values) { 1424 // Use _idx of address base for boxed values. 1425 this_iid = base->_idx; 1426 } 1427 PhaseIterGVN* igvn = phase->is_IterGVN(); 1428 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1429 for (uint i = 1; i < region->req(); i++) { 1430 Node* x; 1431 Node* the_clone = NULL; 1432 if (region->in(i) == C->top()) { 1433 x = C->top(); // Dead path? Use a dead data op 1434 } else { 1435 x = this->clone(); // Else clone up the data op 1436 the_clone = x; // Remember for possible deletion. 1437 // Alter data node to use pre-phi inputs 1438 if (this->in(0) == region) { 1439 x->set_req(0, region->in(i)); 1440 } else { 1441 x->set_req(0, NULL); 1442 } 1443 if (mem->is_Phi() && (mem->in(0) == region)) { 1444 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1445 } 1446 if (address->is_Phi() && address->in(0) == region) { 1447 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1448 } 1449 if (base_is_phi && (base->in(0) == region)) { 1450 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1451 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1452 x->set_req(Address, adr_x); 1453 } 1454 } 1455 // Check for a 'win' on some paths 1456 const Type *t = x->Value(igvn); 1457 1458 bool singleton = t->singleton(); 1459 1460 // See comments in PhaseIdealLoop::split_thru_phi(). 1461 if (singleton && t == Type::TOP) { 1462 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1463 } 1464 1465 if (singleton) { 1466 x = igvn->makecon(t); 1467 } else { 1468 // We now call Identity to try to simplify the cloned node. 1469 // Note that some Identity methods call phase->type(this). 1470 // Make sure that the type array is big enough for 1471 // our new node, even though we may throw the node away. 1472 // (This tweaking with igvn only works because x is a new node.) 1473 igvn->set_type(x, t); 1474 // If x is a TypeNode, capture any more-precise type permanently into Node 1475 // otherwise it will be not updated during igvn->transform since 1476 // igvn->type(x) is set to x->Value() already. 1477 x->raise_bottom_type(t); 1478 Node *y = x->Identity(igvn); 1479 if (y != x) { 1480 x = y; 1481 } else { 1482 y = igvn->hash_find_insert(x); 1483 if (y) { 1484 x = y; 1485 } else { 1486 // Else x is a new node we are keeping 1487 // We do not need register_new_node_with_optimizer 1488 // because set_type has already been called. 1489 igvn->_worklist.push(x); 1490 } 1491 } 1492 } 1493 if (x != the_clone && the_clone != NULL) { 1494 igvn->remove_dead_node(the_clone); 1495 } 1496 phi->set_req(i, x); 1497 } 1498 // Record Phi 1499 igvn->register_new_node_with_optimizer(phi); 1500 return phi; 1501 } 1502 1503 //------------------------------Ideal------------------------------------------ 1504 // If the load is from Field memory and the pointer is non-null, it might be possible to 1505 // zero out the control input. 1506 // If the offset is constant and the base is an object allocation, 1507 // try to hook me up to the exact initializing store. 1508 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1509 Node* p = MemNode::Ideal_common(phase, can_reshape); 1510 if (p) return (p == NodeSentinel) ? NULL : p; 1511 1512 Node* ctrl = in(MemNode::Control); 1513 Node* address = in(MemNode::Address); 1514 bool progress = false; 1515 1516 // Skip up past a SafePoint control. Cannot do this for Stores because 1517 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1518 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1519 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1520 ctrl = ctrl->in(0); 1521 set_req(MemNode::Control,ctrl); 1522 progress = true; 1523 } 1524 1525 intptr_t ignore = 0; 1526 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1527 if (base != NULL 1528 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1529 // Check for useless control edge in some common special cases 1530 if (in(MemNode::Control) != NULL 1531 && can_remove_control() 1532 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1533 && all_controls_dominate(base, phase->C->start())) { 1534 // A method-invariant, non-null address (constant or 'this' argument). 1535 set_req(MemNode::Control, NULL); 1536 progress = true; 1537 } 1538 } 1539 1540 Node* mem = in(MemNode::Memory); 1541 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1542 1543 if (can_reshape && (addr_t != NULL)) { 1544 // try to optimize our memory input 1545 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1546 if (opt_mem != mem) { 1547 set_req(MemNode::Memory, opt_mem); 1548 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1549 return this; 1550 } 1551 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1552 if ((t_oop != NULL) && 1553 (t_oop->is_known_instance_field() || 1554 t_oop->is_ptr_to_boxed_value())) { 1555 PhaseIterGVN *igvn = phase->is_IterGVN(); 1556 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1557 // Delay this transformation until memory Phi is processed. 1558 phase->is_IterGVN()->_worklist.push(this); 1559 return NULL; 1560 } 1561 // Split instance field load through Phi. 1562 Node* result = split_through_phi(phase); 1563 if (result != NULL) return result; 1564 1565 if (t_oop->is_ptr_to_boxed_value()) { 1566 Node* result = eliminate_autobox(phase); 1567 if (result != NULL) return result; 1568 } 1569 } 1570 } 1571 1572 // Is there a dominating load that loads the same value? Leave 1573 // anything that is not a load of a field/array element (like 1574 // barriers etc.) alone 1575 if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) { 1576 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1577 Node *use = mem->fast_out(i); 1578 if (use != this && 1579 use->Opcode() == Opcode() && 1580 use->in(0) != NULL && 1581 use->in(0) != in(0) && 1582 use->in(Address) == in(Address)) { 1583 Node* ctl = in(0); 1584 for (int i = 0; i < 10 && ctl != NULL; i++) { 1585 ctl = IfNode::up_one_dom(ctl); 1586 if (ctl == use->in(0)) { 1587 set_req(0, use->in(0)); 1588 return this; 1589 } 1590 } 1591 } 1592 } 1593 } 1594 1595 // Check for prior store with a different base or offset; make Load 1596 // independent. Skip through any number of them. Bail out if the stores 1597 // are in an endless dead cycle and report no progress. This is a key 1598 // transform for Reflection. However, if after skipping through the Stores 1599 // we can't then fold up against a prior store do NOT do the transform as 1600 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1601 // array memory alive twice: once for the hoisted Load and again after the 1602 // bypassed Store. This situation only works if EVERYBODY who does 1603 // anti-dependence work knows how to bypass. I.e. we need all 1604 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1605 // the alias index stuff. So instead, peek through Stores and IFF we can 1606 // fold up, do so. 1607 Node* prev_mem = find_previous_store(phase); 1608 if (prev_mem != NULL) { 1609 Node* value = can_see_arraycopy_value(prev_mem, phase); 1610 if (value != NULL) { 1611 return value; 1612 } 1613 } 1614 // Steps (a), (b): Walk past independent stores to find an exact match. 1615 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1616 // (c) See if we can fold up on the spot, but don't fold up here. 1617 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1618 // just return a prior value, which is done by Identity calls. 1619 if (can_see_stored_value(prev_mem, phase)) { 1620 // Make ready for step (d): 1621 set_req(MemNode::Memory, prev_mem); 1622 return this; 1623 } 1624 } 1625 1626 return progress ? this : NULL; 1627 } 1628 1629 // Helper to recognize certain Klass fields which are invariant across 1630 // some group of array types (e.g., int[] or all T[] where T < Object). 1631 const Type* 1632 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1633 ciKlass* klass) const { 1634 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1635 // The field is Klass::_modifier_flags. Return its (constant) value. 1636 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1637 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1638 return TypeInt::make(klass->modifier_flags()); 1639 } 1640 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1641 // The field is Klass::_access_flags. Return its (constant) value. 1642 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1643 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1644 return TypeInt::make(klass->access_flags()); 1645 } 1646 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1647 // The field is Klass::_layout_helper. Return its constant value if known. 1648 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1649 return TypeInt::make(klass->layout_helper()); 1650 } 1651 1652 // No match. 1653 return NULL; 1654 } 1655 1656 //------------------------------Value----------------------------------------- 1657 const Type* LoadNode::Value(PhaseGVN* phase) const { 1658 // Either input is TOP ==> the result is TOP 1659 Node* mem = in(MemNode::Memory); 1660 const Type *t1 = phase->type(mem); 1661 if (t1 == Type::TOP) return Type::TOP; 1662 Node* adr = in(MemNode::Address); 1663 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1664 if (tp == NULL || tp->empty()) return Type::TOP; 1665 int off = tp->offset(); 1666 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1667 Compile* C = phase->C; 1668 1669 // Try to guess loaded type from pointer type 1670 if (tp->isa_aryptr()) { 1671 const TypeAryPtr* ary = tp->is_aryptr(); 1672 const Type* t = ary->elem(); 1673 1674 // Determine whether the reference is beyond the header or not, by comparing 1675 // the offset against the offset of the start of the array's data. 1676 // Different array types begin at slightly different offsets (12 vs. 16). 1677 // We choose T_BYTE as an example base type that is least restrictive 1678 // as to alignment, which will therefore produce the smallest 1679 // possible base offset. 1680 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1681 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1682 1683 // Try to constant-fold a stable array element. 1684 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1685 // Make sure the reference is not into the header and the offset is constant 1686 ciObject* aobj = ary->const_oop(); 1687 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1688 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1689 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1690 stable_dimension, 1691 memory_type(), is_unsigned()); 1692 if (con_type != NULL) { 1693 return con_type; 1694 } 1695 } 1696 } 1697 1698 // Don't do this for integer types. There is only potential profit if 1699 // the element type t is lower than _type; that is, for int types, if _type is 1700 // more restrictive than t. This only happens here if one is short and the other 1701 // char (both 16 bits), and in those cases we've made an intentional decision 1702 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1703 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1704 // 1705 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1706 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1707 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1708 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1709 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1710 // In fact, that could have been the original type of p1, and p1 could have 1711 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1712 // expression (LShiftL quux 3) independently optimized to the constant 8. 1713 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1714 && (_type->isa_vect() == NULL) 1715 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1716 // t might actually be lower than _type, if _type is a unique 1717 // concrete subclass of abstract class t. 1718 if (off_beyond_header) { // is the offset beyond the header? 1719 const Type* jt = t->join_speculative(_type); 1720 // In any case, do not allow the join, per se, to empty out the type. 1721 if (jt->empty() && !t->empty()) { 1722 // This can happen if a interface-typed array narrows to a class type. 1723 jt = _type; 1724 } 1725 #ifdef ASSERT 1726 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1727 // The pointers in the autobox arrays are always non-null 1728 Node* base = adr->in(AddPNode::Base); 1729 if ((base != NULL) && base->is_DecodeN()) { 1730 // Get LoadN node which loads IntegerCache.cache field 1731 base = base->in(1); 1732 } 1733 if ((base != NULL) && base->is_Con()) { 1734 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1735 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1736 // It could be narrow oop 1737 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1738 } 1739 } 1740 } 1741 #endif 1742 return jt; 1743 } 1744 } 1745 } else if (tp->base() == Type::InstPtr) { 1746 assert( off != Type::OffsetBot || 1747 // arrays can be cast to Objects 1748 tp->is_oopptr()->klass()->is_java_lang_Object() || 1749 // unsafe field access may not have a constant offset 1750 C->has_unsafe_access(), 1751 "Field accesses must be precise" ); 1752 // For oop loads, we expect the _type to be precise. 1753 1754 // Optimize loads from constant fields. 1755 const TypeInstPtr* tinst = tp->is_instptr(); 1756 ciObject* const_oop = tinst->const_oop(); 1757 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1758 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type()); 1759 if (con_type != NULL) { 1760 return con_type; 1761 } 1762 } 1763 } else if (tp->base() == Type::KlassPtr) { 1764 assert( off != Type::OffsetBot || 1765 // arrays can be cast to Objects 1766 tp->is_klassptr()->klass()->is_java_lang_Object() || 1767 // also allow array-loading from the primary supertype 1768 // array during subtype checks 1769 Opcode() == Op_LoadKlass, 1770 "Field accesses must be precise" ); 1771 // For klass/static loads, we expect the _type to be precise 1772 } 1773 1774 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1775 if (tkls != NULL && !StressReflectiveCode) { 1776 ciKlass* klass = tkls->klass(); 1777 if (klass->is_loaded() && tkls->klass_is_exact()) { 1778 // We are loading a field from a Klass metaobject whose identity 1779 // is known at compile time (the type is "exact" or "precise"). 1780 // Check for fields we know are maintained as constants by the VM. 1781 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1782 // The field is Klass::_super_check_offset. Return its (constant) value. 1783 // (Folds up type checking code.) 1784 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1785 return TypeInt::make(klass->super_check_offset()); 1786 } 1787 // Compute index into primary_supers array 1788 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1789 // Check for overflowing; use unsigned compare to handle the negative case. 1790 if( depth < ciKlass::primary_super_limit() ) { 1791 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1792 // (Folds up type checking code.) 1793 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1794 ciKlass *ss = klass->super_of_depth(depth); 1795 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1796 } 1797 const Type* aift = load_array_final_field(tkls, klass); 1798 if (aift != NULL) return aift; 1799 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1800 // The field is Klass::_java_mirror. Return its (constant) value. 1801 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1802 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1803 return TypeInstPtr::make(klass->java_mirror()); 1804 } 1805 } 1806 1807 // We can still check if we are loading from the primary_supers array at a 1808 // shallow enough depth. Even though the klass is not exact, entries less 1809 // than or equal to its super depth are correct. 1810 if (klass->is_loaded() ) { 1811 ciType *inner = klass; 1812 while( inner->is_obj_array_klass() ) 1813 inner = inner->as_obj_array_klass()->base_element_type(); 1814 if( inner->is_instance_klass() && 1815 !inner->as_instance_klass()->flags().is_interface() ) { 1816 // Compute index into primary_supers array 1817 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1818 // Check for overflowing; use unsigned compare to handle the negative case. 1819 if( depth < ciKlass::primary_super_limit() && 1820 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1821 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1822 // (Folds up type checking code.) 1823 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1824 ciKlass *ss = klass->super_of_depth(depth); 1825 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1826 } 1827 } 1828 } 1829 1830 // If the type is enough to determine that the thing is not an array, 1831 // we can give the layout_helper a positive interval type. 1832 // This will help short-circuit some reflective code. 1833 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1834 && !klass->is_array_klass() // not directly typed as an array 1835 && !klass->is_interface() // specifically not Serializable & Cloneable 1836 && !klass->is_java_lang_Object() // not the supertype of all T[] 1837 ) { 1838 // Note: When interfaces are reliable, we can narrow the interface 1839 // test to (klass != Serializable && klass != Cloneable). 1840 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1841 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1842 // The key property of this type is that it folds up tests 1843 // for array-ness, since it proves that the layout_helper is positive. 1844 // Thus, a generic value like the basic object layout helper works fine. 1845 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1846 } 1847 } 1848 1849 // If we are loading from a freshly-allocated object, produce a zero, 1850 // if the load is provably beyond the header of the object. 1851 // (Also allow a variable load from a fresh array to produce zero.) 1852 const TypeOopPtr *tinst = tp->isa_oopptr(); 1853 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1854 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1855 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1856 Node* value = can_see_stored_value(mem,phase); 1857 if (value != NULL && value->is_Con()) { 1858 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1859 return value->bottom_type(); 1860 } 1861 } 1862 1863 if (is_instance) { 1864 // If we have an instance type and our memory input is the 1865 // programs's initial memory state, there is no matching store, 1866 // so just return a zero of the appropriate type 1867 Node *mem = in(MemNode::Memory); 1868 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1869 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1870 return Type::get_zero_type(_type->basic_type()); 1871 } 1872 } 1873 return _type; 1874 } 1875 1876 //------------------------------match_edge------------------------------------- 1877 // Do we Match on this edge index or not? Match only the address. 1878 uint LoadNode::match_edge(uint idx) const { 1879 return idx == MemNode::Address; 1880 } 1881 1882 //--------------------------LoadBNode::Ideal-------------------------------------- 1883 // 1884 // If the previous store is to the same address as this load, 1885 // and the value stored was larger than a byte, replace this load 1886 // with the value stored truncated to a byte. If no truncation is 1887 // needed, the replacement is done in LoadNode::Identity(). 1888 // 1889 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1890 Node* mem = in(MemNode::Memory); 1891 Node* value = can_see_stored_value(mem,phase); 1892 if( value && !phase->type(value)->higher_equal( _type ) ) { 1893 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1894 return new RShiftINode(result, phase->intcon(24)); 1895 } 1896 // Identity call will handle the case where truncation is not needed. 1897 return LoadNode::Ideal(phase, can_reshape); 1898 } 1899 1900 const Type* LoadBNode::Value(PhaseGVN* phase) const { 1901 Node* mem = in(MemNode::Memory); 1902 Node* value = can_see_stored_value(mem,phase); 1903 if (value != NULL && value->is_Con() && 1904 !value->bottom_type()->higher_equal(_type)) { 1905 // If the input to the store does not fit with the load's result type, 1906 // it must be truncated. We can't delay until Ideal call since 1907 // a singleton Value is needed for split_thru_phi optimization. 1908 int con = value->get_int(); 1909 return TypeInt::make((con << 24) >> 24); 1910 } 1911 return LoadNode::Value(phase); 1912 } 1913 1914 //--------------------------LoadUBNode::Ideal------------------------------------- 1915 // 1916 // If the previous store is to the same address as this load, 1917 // and the value stored was larger than a byte, replace this load 1918 // with the value stored truncated to a byte. If no truncation is 1919 // needed, the replacement is done in LoadNode::Identity(). 1920 // 1921 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1922 Node* mem = in(MemNode::Memory); 1923 Node* value = can_see_stored_value(mem, phase); 1924 if (value && !phase->type(value)->higher_equal(_type)) 1925 return new AndINode(value, phase->intcon(0xFF)); 1926 // Identity call will handle the case where truncation is not needed. 1927 return LoadNode::Ideal(phase, can_reshape); 1928 } 1929 1930 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 1931 Node* mem = in(MemNode::Memory); 1932 Node* value = can_see_stored_value(mem,phase); 1933 if (value != NULL && value->is_Con() && 1934 !value->bottom_type()->higher_equal(_type)) { 1935 // If the input to the store does not fit with the load's result type, 1936 // it must be truncated. We can't delay until Ideal call since 1937 // a singleton Value is needed for split_thru_phi optimization. 1938 int con = value->get_int(); 1939 return TypeInt::make(con & 0xFF); 1940 } 1941 return LoadNode::Value(phase); 1942 } 1943 1944 //--------------------------LoadUSNode::Ideal------------------------------------- 1945 // 1946 // If the previous store is to the same address as this load, 1947 // and the value stored was larger than a char, replace this load 1948 // with the value stored truncated to a char. If no truncation is 1949 // needed, the replacement is done in LoadNode::Identity(). 1950 // 1951 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1952 Node* mem = in(MemNode::Memory); 1953 Node* value = can_see_stored_value(mem,phase); 1954 if( value && !phase->type(value)->higher_equal( _type ) ) 1955 return new AndINode(value,phase->intcon(0xFFFF)); 1956 // Identity call will handle the case where truncation is not needed. 1957 return LoadNode::Ideal(phase, can_reshape); 1958 } 1959 1960 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 1961 Node* mem = in(MemNode::Memory); 1962 Node* value = can_see_stored_value(mem,phase); 1963 if (value != NULL && value->is_Con() && 1964 !value->bottom_type()->higher_equal(_type)) { 1965 // If the input to the store does not fit with the load's result type, 1966 // it must be truncated. We can't delay until Ideal call since 1967 // a singleton Value is needed for split_thru_phi optimization. 1968 int con = value->get_int(); 1969 return TypeInt::make(con & 0xFFFF); 1970 } 1971 return LoadNode::Value(phase); 1972 } 1973 1974 //--------------------------LoadSNode::Ideal-------------------------------------- 1975 // 1976 // If the previous store is to the same address as this load, 1977 // and the value stored was larger than a short, replace this load 1978 // with the value stored truncated to a short. If no truncation is 1979 // needed, the replacement is done in LoadNode::Identity(). 1980 // 1981 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1982 Node* mem = in(MemNode::Memory); 1983 Node* value = can_see_stored_value(mem,phase); 1984 if( value && !phase->type(value)->higher_equal( _type ) ) { 1985 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 1986 return new RShiftINode(result, phase->intcon(16)); 1987 } 1988 // Identity call will handle the case where truncation is not needed. 1989 return LoadNode::Ideal(phase, can_reshape); 1990 } 1991 1992 const Type* LoadSNode::Value(PhaseGVN* phase) const { 1993 Node* mem = in(MemNode::Memory); 1994 Node* value = can_see_stored_value(mem,phase); 1995 if (value != NULL && value->is_Con() && 1996 !value->bottom_type()->higher_equal(_type)) { 1997 // If the input to the store does not fit with the load's result type, 1998 // it must be truncated. We can't delay until Ideal call since 1999 // a singleton Value is needed for split_thru_phi optimization. 2000 int con = value->get_int(); 2001 return TypeInt::make((con << 16) >> 16); 2002 } 2003 return LoadNode::Value(phase); 2004 } 2005 2006 //============================================================================= 2007 //----------------------------LoadKlassNode::make------------------------------ 2008 // Polymorphic factory method: 2009 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2010 // sanity check the alias category against the created node type 2011 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2012 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2013 #ifdef _LP64 2014 if (adr_type->is_ptr_to_narrowklass()) { 2015 assert(UseCompressedClassPointers, "no compressed klasses"); 2016 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2017 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2018 } 2019 #endif 2020 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2021 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2022 } 2023 2024 //------------------------------Value------------------------------------------ 2025 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2026 return klass_value_common(phase); 2027 } 2028 2029 // In most cases, LoadKlassNode does not have the control input set. If the control 2030 // input is set, it must not be removed (by LoadNode::Ideal()). 2031 bool LoadKlassNode::can_remove_control() const { 2032 return false; 2033 } 2034 2035 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2036 // Either input is TOP ==> the result is TOP 2037 const Type *t1 = phase->type( in(MemNode::Memory) ); 2038 if (t1 == Type::TOP) return Type::TOP; 2039 Node *adr = in(MemNode::Address); 2040 const Type *t2 = phase->type( adr ); 2041 if (t2 == Type::TOP) return Type::TOP; 2042 const TypePtr *tp = t2->is_ptr(); 2043 if (TypePtr::above_centerline(tp->ptr()) || 2044 tp->ptr() == TypePtr::Null) return Type::TOP; 2045 2046 // Return a more precise klass, if possible 2047 const TypeInstPtr *tinst = tp->isa_instptr(); 2048 if (tinst != NULL) { 2049 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2050 int offset = tinst->offset(); 2051 if (ik == phase->C->env()->Class_klass() 2052 && (offset == java_lang_Class::klass_offset_in_bytes() || 2053 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2054 // We are loading a special hidden field from a Class mirror object, 2055 // the field which points to the VM's Klass metaobject. 2056 ciType* t = tinst->java_mirror_type(); 2057 // java_mirror_type returns non-null for compile-time Class constants. 2058 if (t != NULL) { 2059 // constant oop => constant klass 2060 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2061 if (t->is_void()) { 2062 // We cannot create a void array. Since void is a primitive type return null 2063 // klass. Users of this result need to do a null check on the returned klass. 2064 return TypePtr::NULL_PTR; 2065 } 2066 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2067 } 2068 if (!t->is_klass()) { 2069 // a primitive Class (e.g., int.class) has NULL for a klass field 2070 return TypePtr::NULL_PTR; 2071 } 2072 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2073 return TypeKlassPtr::make(t->as_klass()); 2074 } 2075 // non-constant mirror, so we can't tell what's going on 2076 } 2077 if( !ik->is_loaded() ) 2078 return _type; // Bail out if not loaded 2079 if (offset == oopDesc::klass_offset_in_bytes()) { 2080 if (tinst->klass_is_exact()) { 2081 return TypeKlassPtr::make(ik); 2082 } 2083 // See if we can become precise: no subklasses and no interface 2084 // (Note: We need to support verified interfaces.) 2085 if (!ik->is_interface() && !ik->has_subklass()) { 2086 //assert(!UseExactTypes, "this code should be useless with exact types"); 2087 // Add a dependence; if any subclass added we need to recompile 2088 if (!ik->is_final()) { 2089 // %%% should use stronger assert_unique_concrete_subtype instead 2090 phase->C->dependencies()->assert_leaf_type(ik); 2091 } 2092 // Return precise klass 2093 return TypeKlassPtr::make(ik); 2094 } 2095 2096 // Return root of possible klass 2097 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2098 } 2099 } 2100 2101 // Check for loading klass from an array 2102 const TypeAryPtr *tary = tp->isa_aryptr(); 2103 if( tary != NULL ) { 2104 ciKlass *tary_klass = tary->klass(); 2105 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2106 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2107 if (tary->klass_is_exact()) { 2108 return TypeKlassPtr::make(tary_klass); 2109 } 2110 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2111 // If the klass is an object array, we defer the question to the 2112 // array component klass. 2113 if( ak->is_obj_array_klass() ) { 2114 assert( ak->is_loaded(), "" ); 2115 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2116 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2117 ciInstanceKlass* ik = base_k->as_instance_klass(); 2118 // See if we can become precise: no subklasses and no interface 2119 if (!ik->is_interface() && !ik->has_subklass()) { 2120 //assert(!UseExactTypes, "this code should be useless with exact types"); 2121 // Add a dependence; if any subclass added we need to recompile 2122 if (!ik->is_final()) { 2123 phase->C->dependencies()->assert_leaf_type(ik); 2124 } 2125 // Return precise array klass 2126 return TypeKlassPtr::make(ak); 2127 } 2128 } 2129 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2130 } else { // Found a type-array? 2131 //assert(!UseExactTypes, "this code should be useless with exact types"); 2132 assert( ak->is_type_array_klass(), "" ); 2133 return TypeKlassPtr::make(ak); // These are always precise 2134 } 2135 } 2136 } 2137 2138 // Check for loading klass from an array klass 2139 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2140 if (tkls != NULL && !StressReflectiveCode) { 2141 ciKlass* klass = tkls->klass(); 2142 if( !klass->is_loaded() ) 2143 return _type; // Bail out if not loaded 2144 if( klass->is_obj_array_klass() && 2145 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2146 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2147 // // Always returning precise element type is incorrect, 2148 // // e.g., element type could be object and array may contain strings 2149 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2150 2151 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2152 // according to the element type's subclassing. 2153 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2154 } 2155 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2156 tkls->offset() == in_bytes(Klass::super_offset())) { 2157 ciKlass* sup = klass->as_instance_klass()->super(); 2158 // The field is Klass::_super. Return its (constant) value. 2159 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2160 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2161 } 2162 } 2163 2164 // Bailout case 2165 return LoadNode::Value(phase); 2166 } 2167 2168 //------------------------------Identity--------------------------------------- 2169 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2170 // Also feed through the klass in Allocate(...klass...)._klass. 2171 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2172 return klass_identity_common(phase); 2173 } 2174 2175 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2176 Node* x = LoadNode::Identity(phase); 2177 if (x != this) return x; 2178 2179 // Take apart the address into an oop and and offset. 2180 // Return 'this' if we cannot. 2181 Node* adr = in(MemNode::Address); 2182 intptr_t offset = 0; 2183 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2184 if (base == NULL) return this; 2185 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2186 if (toop == NULL) return this; 2187 2188 // We can fetch the klass directly through an AllocateNode. 2189 // This works even if the klass is not constant (clone or newArray). 2190 if (offset == oopDesc::klass_offset_in_bytes()) { 2191 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2192 if (allocated_klass != NULL) { 2193 return allocated_klass; 2194 } 2195 } 2196 2197 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2198 // See inline_native_Class_query for occurrences of these patterns. 2199 // Java Example: x.getClass().isAssignableFrom(y) 2200 // 2201 // This improves reflective code, often making the Class 2202 // mirror go completely dead. (Current exception: Class 2203 // mirrors may appear in debug info, but we could clean them out by 2204 // introducing a new debug info operator for Klass*.java_mirror). 2205 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2206 && offset == java_lang_Class::klass_offset_in_bytes()) { 2207 // We are loading a special hidden field from a Class mirror, 2208 // the field which points to its Klass or ArrayKlass metaobject. 2209 if (base->is_Load()) { 2210 Node* adr2 = base->in(MemNode::Address); 2211 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2212 if (tkls != NULL && !tkls->empty() 2213 && (tkls->klass()->is_instance_klass() || 2214 tkls->klass()->is_array_klass()) 2215 && adr2->is_AddP() 2216 ) { 2217 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2218 if (tkls->offset() == mirror_field) { 2219 return adr2->in(AddPNode::Base); 2220 } 2221 } 2222 } 2223 } 2224 2225 return this; 2226 } 2227 2228 2229 //------------------------------Value------------------------------------------ 2230 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2231 const Type *t = klass_value_common(phase); 2232 if (t == Type::TOP) 2233 return t; 2234 2235 return t->make_narrowklass(); 2236 } 2237 2238 //------------------------------Identity--------------------------------------- 2239 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2240 // Also feed through the klass in Allocate(...klass...)._klass. 2241 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2242 Node *x = klass_identity_common(phase); 2243 2244 const Type *t = phase->type( x ); 2245 if( t == Type::TOP ) return x; 2246 if( t->isa_narrowklass()) return x; 2247 assert (!t->isa_narrowoop(), "no narrow oop here"); 2248 2249 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2250 } 2251 2252 //------------------------------Value----------------------------------------- 2253 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2254 // Either input is TOP ==> the result is TOP 2255 const Type *t1 = phase->type( in(MemNode::Memory) ); 2256 if( t1 == Type::TOP ) return Type::TOP; 2257 Node *adr = in(MemNode::Address); 2258 const Type *t2 = phase->type( adr ); 2259 if( t2 == Type::TOP ) return Type::TOP; 2260 const TypePtr *tp = t2->is_ptr(); 2261 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2262 const TypeAryPtr *tap = tp->isa_aryptr(); 2263 if( !tap ) return _type; 2264 return tap->size(); 2265 } 2266 2267 //-------------------------------Ideal--------------------------------------- 2268 // Feed through the length in AllocateArray(...length...)._length. 2269 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2270 Node* p = MemNode::Ideal_common(phase, can_reshape); 2271 if (p) return (p == NodeSentinel) ? NULL : p; 2272 2273 // Take apart the address into an oop and and offset. 2274 // Return 'this' if we cannot. 2275 Node* adr = in(MemNode::Address); 2276 intptr_t offset = 0; 2277 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2278 if (base == NULL) return NULL; 2279 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2280 if (tary == NULL) return NULL; 2281 2282 // We can fetch the length directly through an AllocateArrayNode. 2283 // This works even if the length is not constant (clone or newArray). 2284 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2285 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2286 if (alloc != NULL) { 2287 Node* allocated_length = alloc->Ideal_length(); 2288 Node* len = alloc->make_ideal_length(tary, phase); 2289 if (allocated_length != len) { 2290 // New CastII improves on this. 2291 return len; 2292 } 2293 } 2294 } 2295 2296 return NULL; 2297 } 2298 2299 //------------------------------Identity--------------------------------------- 2300 // Feed through the length in AllocateArray(...length...)._length. 2301 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2302 Node* x = LoadINode::Identity(phase); 2303 if (x != this) return x; 2304 2305 // Take apart the address into an oop and and offset. 2306 // Return 'this' if we cannot. 2307 Node* adr = in(MemNode::Address); 2308 intptr_t offset = 0; 2309 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2310 if (base == NULL) return this; 2311 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2312 if (tary == NULL) return this; 2313 2314 // We can fetch the length directly through an AllocateArrayNode. 2315 // This works even if the length is not constant (clone or newArray). 2316 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2317 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2318 if (alloc != NULL) { 2319 Node* allocated_length = alloc->Ideal_length(); 2320 // Do not allow make_ideal_length to allocate a CastII node. 2321 Node* len = alloc->make_ideal_length(tary, phase, false); 2322 if (allocated_length == len) { 2323 // Return allocated_length only if it would not be improved by a CastII. 2324 return allocated_length; 2325 } 2326 } 2327 } 2328 2329 return this; 2330 2331 } 2332 2333 //============================================================================= 2334 //---------------------------StoreNode::make----------------------------------- 2335 // Polymorphic factory method: 2336 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2337 assert((mo == unordered || mo == release), "unexpected"); 2338 Compile* C = gvn.C; 2339 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2340 ctl != NULL, "raw memory operations should have control edge"); 2341 2342 switch (bt) { 2343 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2344 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2345 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2346 case T_CHAR: 2347 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2348 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2349 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2350 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2351 case T_METADATA: 2352 case T_ADDRESS: 2353 case T_OBJECT: 2354 #ifdef _LP64 2355 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2356 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2357 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2358 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2359 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2360 adr->bottom_type()->isa_rawptr())) { 2361 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2362 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2363 } 2364 #endif 2365 { 2366 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2367 } 2368 default: 2369 ShouldNotReachHere(); 2370 return (StoreNode*)NULL; 2371 } 2372 } 2373 2374 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2375 bool require_atomic = true; 2376 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2377 } 2378 2379 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2380 bool require_atomic = true; 2381 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2382 } 2383 2384 2385 //--------------------------bottom_type---------------------------------------- 2386 const Type *StoreNode::bottom_type() const { 2387 return Type::MEMORY; 2388 } 2389 2390 //------------------------------hash------------------------------------------- 2391 uint StoreNode::hash() const { 2392 // unroll addition of interesting fields 2393 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2394 2395 // Since they are not commoned, do not hash them: 2396 return NO_HASH; 2397 } 2398 2399 //------------------------------Ideal------------------------------------------ 2400 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2401 // When a store immediately follows a relevant allocation/initialization, 2402 // try to capture it into the initialization, or hoist it above. 2403 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2404 Node* p = MemNode::Ideal_common(phase, can_reshape); 2405 if (p) return (p == NodeSentinel) ? NULL : p; 2406 2407 Node* mem = in(MemNode::Memory); 2408 Node* address = in(MemNode::Address); 2409 // Back-to-back stores to same address? Fold em up. Generally 2410 // unsafe if I have intervening uses... Also disallowed for StoreCM 2411 // since they must follow each StoreP operation. Redundant StoreCMs 2412 // are eliminated just before matching in final_graph_reshape. 2413 { 2414 Node* st = mem; 2415 // If Store 'st' has more than one use, we cannot fold 'st' away. 2416 // For example, 'st' might be the final state at a conditional 2417 // return. Or, 'st' might be used by some node which is live at 2418 // the same time 'st' is live, which might be unschedulable. So, 2419 // require exactly ONE user until such time as we clone 'mem' for 2420 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2421 // true). 2422 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2423 // Looking at a dead closed cycle of memory? 2424 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2425 assert(Opcode() == st->Opcode() || 2426 st->Opcode() == Op_StoreVector || 2427 Opcode() == Op_StoreVector || 2428 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2429 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2430 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2431 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2432 2433 if (st->in(MemNode::Address)->eqv_uncast(address) && 2434 st->as_Store()->memory_size() <= this->memory_size()) { 2435 Node* use = st->raw_out(0); 2436 phase->igvn_rehash_node_delayed(use); 2437 if (can_reshape) { 2438 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2439 } else { 2440 // It's OK to do this in the parser, since DU info is always accurate, 2441 // and the parser always refers to nodes via SafePointNode maps. 2442 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2443 } 2444 return this; 2445 } 2446 st = st->in(MemNode::Memory); 2447 } 2448 } 2449 2450 2451 // Capture an unaliased, unconditional, simple store into an initializer. 2452 // Or, if it is independent of the allocation, hoist it above the allocation. 2453 if (ReduceFieldZeroing && /*can_reshape &&*/ 2454 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2455 InitializeNode* init = mem->in(0)->as_Initialize(); 2456 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2457 if (offset > 0) { 2458 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2459 // If the InitializeNode captured me, it made a raw copy of me, 2460 // and I need to disappear. 2461 if (moved != NULL) { 2462 // %%% hack to ensure that Ideal returns a new node: 2463 mem = MergeMemNode::make(mem); 2464 return mem; // fold me away 2465 } 2466 } 2467 } 2468 2469 return NULL; // No further progress 2470 } 2471 2472 //------------------------------Value----------------------------------------- 2473 const Type* StoreNode::Value(PhaseGVN* phase) const { 2474 // Either input is TOP ==> the result is TOP 2475 const Type *t1 = phase->type( in(MemNode::Memory) ); 2476 if( t1 == Type::TOP ) return Type::TOP; 2477 const Type *t2 = phase->type( in(MemNode::Address) ); 2478 if( t2 == Type::TOP ) return Type::TOP; 2479 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2480 if( t3 == Type::TOP ) return Type::TOP; 2481 return Type::MEMORY; 2482 } 2483 2484 //------------------------------Identity--------------------------------------- 2485 // Remove redundant stores: 2486 // Store(m, p, Load(m, p)) changes to m. 2487 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2488 Node* StoreNode::Identity(PhaseGVN* phase) { 2489 Node* mem = in(MemNode::Memory); 2490 Node* adr = in(MemNode::Address); 2491 Node* val = in(MemNode::ValueIn); 2492 2493 // Load then Store? Then the Store is useless 2494 if (val->is_Load() && 2495 val->in(MemNode::Address)->eqv_uncast(adr) && 2496 val->in(MemNode::Memory )->eqv_uncast(mem) && 2497 val->as_Load()->store_Opcode() == Opcode()) { 2498 return mem; 2499 } 2500 2501 // Two stores in a row of the same value? 2502 if (mem->is_Store() && 2503 mem->in(MemNode::Address)->eqv_uncast(adr) && 2504 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2505 mem->Opcode() == Opcode()) { 2506 return mem; 2507 } 2508 2509 // Store of zero anywhere into a freshly-allocated object? 2510 // Then the store is useless. 2511 // (It must already have been captured by the InitializeNode.) 2512 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2513 // a newly allocated object is already all-zeroes everywhere 2514 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2515 return mem; 2516 } 2517 2518 // the store may also apply to zero-bits in an earlier object 2519 Node* prev_mem = find_previous_store(phase); 2520 // Steps (a), (b): Walk past independent stores to find an exact match. 2521 if (prev_mem != NULL) { 2522 Node* prev_val = can_see_stored_value(prev_mem, phase); 2523 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2524 // prev_val and val might differ by a cast; it would be good 2525 // to keep the more informative of the two. 2526 return mem; 2527 } 2528 } 2529 } 2530 2531 return this; 2532 } 2533 2534 //------------------------------match_edge------------------------------------- 2535 // Do we Match on this edge index or not? Match only memory & value 2536 uint StoreNode::match_edge(uint idx) const { 2537 return idx == MemNode::Address || idx == MemNode::ValueIn; 2538 } 2539 2540 //------------------------------cmp-------------------------------------------- 2541 // Do not common stores up together. They generally have to be split 2542 // back up anyways, so do not bother. 2543 uint StoreNode::cmp( const Node &n ) const { 2544 return (&n == this); // Always fail except on self 2545 } 2546 2547 //------------------------------Ideal_masked_input----------------------------- 2548 // Check for a useless mask before a partial-word store 2549 // (StoreB ... (AndI valIn conIa) ) 2550 // If (conIa & mask == mask) this simplifies to 2551 // (StoreB ... (valIn) ) 2552 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2553 Node *val = in(MemNode::ValueIn); 2554 if( val->Opcode() == Op_AndI ) { 2555 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2556 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2557 set_req(MemNode::ValueIn, val->in(1)); 2558 return this; 2559 } 2560 } 2561 return NULL; 2562 } 2563 2564 2565 //------------------------------Ideal_sign_extended_input---------------------- 2566 // Check for useless sign-extension before a partial-word store 2567 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2568 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2569 // (StoreB ... (valIn) ) 2570 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2571 Node *val = in(MemNode::ValueIn); 2572 if( val->Opcode() == Op_RShiftI ) { 2573 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2574 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2575 Node *shl = val->in(1); 2576 if( shl->Opcode() == Op_LShiftI ) { 2577 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2578 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2579 set_req(MemNode::ValueIn, shl->in(1)); 2580 return this; 2581 } 2582 } 2583 } 2584 } 2585 return NULL; 2586 } 2587 2588 //------------------------------value_never_loaded----------------------------------- 2589 // Determine whether there are any possible loads of the value stored. 2590 // For simplicity, we actually check if there are any loads from the 2591 // address stored to, not just for loads of the value stored by this node. 2592 // 2593 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2594 Node *adr = in(Address); 2595 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2596 if (adr_oop == NULL) 2597 return false; 2598 if (!adr_oop->is_known_instance_field()) 2599 return false; // if not a distinct instance, there may be aliases of the address 2600 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2601 Node *use = adr->fast_out(i); 2602 if (use->is_Load() || use->is_LoadStore()) { 2603 return false; 2604 } 2605 } 2606 return true; 2607 } 2608 2609 //============================================================================= 2610 //------------------------------Ideal------------------------------------------ 2611 // If the store is from an AND mask that leaves the low bits untouched, then 2612 // we can skip the AND operation. If the store is from a sign-extension 2613 // (a left shift, then right shift) we can skip both. 2614 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2615 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2616 if( progress != NULL ) return progress; 2617 2618 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2619 if( progress != NULL ) return progress; 2620 2621 // Finally check the default case 2622 return StoreNode::Ideal(phase, can_reshape); 2623 } 2624 2625 //============================================================================= 2626 //------------------------------Ideal------------------------------------------ 2627 // If the store is from an AND mask that leaves the low bits untouched, then 2628 // we can skip the AND operation 2629 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2630 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2631 if( progress != NULL ) return progress; 2632 2633 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2634 if( progress != NULL ) return progress; 2635 2636 // Finally check the default case 2637 return StoreNode::Ideal(phase, can_reshape); 2638 } 2639 2640 //============================================================================= 2641 //------------------------------Identity--------------------------------------- 2642 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2643 // No need to card mark when storing a null ptr 2644 Node* my_store = in(MemNode::OopStore); 2645 if (my_store->is_Store()) { 2646 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2647 if( t1 == TypePtr::NULL_PTR ) { 2648 return in(MemNode::Memory); 2649 } 2650 } 2651 return this; 2652 } 2653 2654 //============================================================================= 2655 //------------------------------Ideal--------------------------------------- 2656 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2657 Node* progress = StoreNode::Ideal(phase, can_reshape); 2658 if (progress != NULL) return progress; 2659 2660 Node* my_store = in(MemNode::OopStore); 2661 if (my_store->is_MergeMem()) { 2662 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2663 set_req(MemNode::OopStore, mem); 2664 return this; 2665 } 2666 2667 return NULL; 2668 } 2669 2670 //------------------------------Value----------------------------------------- 2671 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2672 // Either input is TOP ==> the result is TOP 2673 const Type *t = phase->type( in(MemNode::Memory) ); 2674 if( t == Type::TOP ) return Type::TOP; 2675 t = phase->type( in(MemNode::Address) ); 2676 if( t == Type::TOP ) return Type::TOP; 2677 t = phase->type( in(MemNode::ValueIn) ); 2678 if( t == Type::TOP ) return Type::TOP; 2679 // If extra input is TOP ==> the result is TOP 2680 t = phase->type( in(MemNode::OopStore) ); 2681 if( t == Type::TOP ) return Type::TOP; 2682 2683 return StoreNode::Value( phase ); 2684 } 2685 2686 2687 //============================================================================= 2688 //----------------------------------SCMemProjNode------------------------------ 2689 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2690 { 2691 return bottom_type(); 2692 } 2693 2694 //============================================================================= 2695 //----------------------------------LoadStoreNode------------------------------ 2696 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2697 : Node(required), 2698 _type(rt), 2699 _adr_type(at) 2700 { 2701 init_req(MemNode::Control, c ); 2702 init_req(MemNode::Memory , mem); 2703 init_req(MemNode::Address, adr); 2704 init_req(MemNode::ValueIn, val); 2705 init_class_id(Class_LoadStore); 2706 } 2707 2708 uint LoadStoreNode::ideal_reg() const { 2709 return _type->ideal_reg(); 2710 } 2711 2712 bool LoadStoreNode::result_not_used() const { 2713 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2714 Node *x = fast_out(i); 2715 if (x->Opcode() == Op_SCMemProj) continue; 2716 return false; 2717 } 2718 return true; 2719 } 2720 2721 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2722 2723 //============================================================================= 2724 //----------------------------------LoadStoreConditionalNode-------------------- 2725 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2726 init_req(ExpectedIn, ex ); 2727 } 2728 2729 //============================================================================= 2730 //-------------------------------adr_type-------------------------------------- 2731 const TypePtr* ClearArrayNode::adr_type() const { 2732 Node *adr = in(3); 2733 if (adr == NULL) return NULL; // node is dead 2734 return MemNode::calculate_adr_type(adr->bottom_type()); 2735 } 2736 2737 //------------------------------match_edge------------------------------------- 2738 // Do we Match on this edge index or not? Do not match memory 2739 uint ClearArrayNode::match_edge(uint idx) const { 2740 return idx > 1; 2741 } 2742 2743 //------------------------------Identity--------------------------------------- 2744 // Clearing a zero length array does nothing 2745 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2746 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2747 } 2748 2749 //------------------------------Idealize--------------------------------------- 2750 // Clearing a short array is faster with stores 2751 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2752 // Already know this is a large node, do not try to ideal it 2753 if (!IdealizeClearArrayNode || _is_large) return NULL; 2754 2755 const int unit = BytesPerLong; 2756 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2757 if (!t) return NULL; 2758 if (!t->is_con()) return NULL; 2759 intptr_t raw_count = t->get_con(); 2760 intptr_t size = raw_count; 2761 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2762 // Clearing nothing uses the Identity call. 2763 // Negative clears are possible on dead ClearArrays 2764 // (see jck test stmt114.stmt11402.val). 2765 if (size <= 0 || size % unit != 0) return NULL; 2766 intptr_t count = size / unit; 2767 // Length too long; communicate this to matchers and assemblers. 2768 // Assemblers are responsible to produce fast hardware clears for it. 2769 if (size > InitArrayShortSize) { 2770 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 2771 } 2772 Node *mem = in(1); 2773 if( phase->type(mem)==Type::TOP ) return NULL; 2774 Node *adr = in(3); 2775 const Type* at = phase->type(adr); 2776 if( at==Type::TOP ) return NULL; 2777 const TypePtr* atp = at->isa_ptr(); 2778 // adjust atp to be the correct array element address type 2779 if (atp == NULL) atp = TypePtr::BOTTOM; 2780 else atp = atp->add_offset(Type::OffsetBot); 2781 // Get base for derived pointer purposes 2782 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2783 Node *base = adr->in(1); 2784 2785 Node *zero = phase->makecon(TypeLong::ZERO); 2786 Node *off = phase->MakeConX(BytesPerLong); 2787 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2788 count--; 2789 while( count-- ) { 2790 mem = phase->transform(mem); 2791 adr = phase->transform(new AddPNode(base,adr,off)); 2792 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2793 } 2794 return mem; 2795 } 2796 2797 //----------------------------step_through---------------------------------- 2798 // Return allocation input memory edge if it is different instance 2799 // or itself if it is the one we are looking for. 2800 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2801 Node* n = *np; 2802 assert(n->is_ClearArray(), "sanity"); 2803 intptr_t offset; 2804 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2805 // This method is called only before Allocate nodes are expanded 2806 // during macro nodes expansion. Before that ClearArray nodes are 2807 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2808 // Allocate nodes are expanded) which follows allocations. 2809 assert(alloc != NULL, "should have allocation"); 2810 if (alloc->_idx == instance_id) { 2811 // Can not bypass initialization of the instance we are looking for. 2812 return false; 2813 } 2814 // Otherwise skip it. 2815 InitializeNode* init = alloc->initialization(); 2816 if (init != NULL) 2817 *np = init->in(TypeFunc::Memory); 2818 else 2819 *np = alloc->in(TypeFunc::Memory); 2820 return true; 2821 } 2822 2823 //----------------------------clear_memory------------------------------------- 2824 // Generate code to initialize object storage to zero. 2825 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2826 intptr_t start_offset, 2827 Node* end_offset, 2828 PhaseGVN* phase) { 2829 intptr_t offset = start_offset; 2830 2831 int unit = BytesPerLong; 2832 if ((offset % unit) != 0) { 2833 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2834 adr = phase->transform(adr); 2835 const TypePtr* atp = TypeRawPtr::BOTTOM; 2836 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2837 mem = phase->transform(mem); 2838 offset += BytesPerInt; 2839 } 2840 assert((offset % unit) == 0, ""); 2841 2842 // Initialize the remaining stuff, if any, with a ClearArray. 2843 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2844 } 2845 2846 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2847 Node* start_offset, 2848 Node* end_offset, 2849 PhaseGVN* phase) { 2850 if (start_offset == end_offset) { 2851 // nothing to do 2852 return mem; 2853 } 2854 2855 int unit = BytesPerLong; 2856 Node* zbase = start_offset; 2857 Node* zend = end_offset; 2858 2859 // Scale to the unit required by the CPU: 2860 if (!Matcher::init_array_count_is_in_bytes) { 2861 Node* shift = phase->intcon(exact_log2(unit)); 2862 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2863 zend = phase->transform(new URShiftXNode(zend, shift) ); 2864 } 2865 2866 // Bulk clear double-words 2867 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2868 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2869 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 2870 return phase->transform(mem); 2871 } 2872 2873 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2874 intptr_t start_offset, 2875 intptr_t end_offset, 2876 PhaseGVN* phase) { 2877 if (start_offset == end_offset) { 2878 // nothing to do 2879 return mem; 2880 } 2881 2882 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2883 intptr_t done_offset = end_offset; 2884 if ((done_offset % BytesPerLong) != 0) { 2885 done_offset -= BytesPerInt; 2886 } 2887 if (done_offset > start_offset) { 2888 mem = clear_memory(ctl, mem, dest, 2889 start_offset, phase->MakeConX(done_offset), phase); 2890 } 2891 if (done_offset < end_offset) { // emit the final 32-bit store 2892 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 2893 adr = phase->transform(adr); 2894 const TypePtr* atp = TypeRawPtr::BOTTOM; 2895 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2896 mem = phase->transform(mem); 2897 done_offset += BytesPerInt; 2898 } 2899 assert(done_offset == end_offset, ""); 2900 return mem; 2901 } 2902 2903 //============================================================================= 2904 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2905 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2906 _adr_type(C->get_adr_type(alias_idx)) 2907 { 2908 init_class_id(Class_MemBar); 2909 Node* top = C->top(); 2910 init_req(TypeFunc::I_O,top); 2911 init_req(TypeFunc::FramePtr,top); 2912 init_req(TypeFunc::ReturnAdr,top); 2913 if (precedent != NULL) 2914 init_req(TypeFunc::Parms, precedent); 2915 } 2916 2917 //------------------------------cmp-------------------------------------------- 2918 uint MemBarNode::hash() const { return NO_HASH; } 2919 uint MemBarNode::cmp( const Node &n ) const { 2920 return (&n == this); // Always fail except on self 2921 } 2922 2923 //------------------------------make------------------------------------------- 2924 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2925 switch (opcode) { 2926 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 2927 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 2928 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 2929 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 2930 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 2931 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 2932 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 2933 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 2934 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 2935 case Op_Initialize: return new InitializeNode(C, atp, pn); 2936 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 2937 default: ShouldNotReachHere(); return NULL; 2938 } 2939 } 2940 2941 //------------------------------Ideal------------------------------------------ 2942 // Return a node which is more "ideal" than the current node. Strip out 2943 // control copies 2944 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2945 if (remove_dead_region(phase, can_reshape)) return this; 2946 // Don't bother trying to transform a dead node 2947 if (in(0) && in(0)->is_top()) { 2948 return NULL; 2949 } 2950 2951 bool progress = false; 2952 // Eliminate volatile MemBars for scalar replaced objects. 2953 if (can_reshape && req() == (Precedent+1)) { 2954 bool eliminate = false; 2955 int opc = Opcode(); 2956 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2957 // Volatile field loads and stores. 2958 Node* my_mem = in(MemBarNode::Precedent); 2959 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2960 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2961 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2962 // replace this Precedent (decodeN) with the Load instead. 2963 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2964 Node* load_node = my_mem->in(1); 2965 set_req(MemBarNode::Precedent, load_node); 2966 phase->is_IterGVN()->_worklist.push(my_mem); 2967 my_mem = load_node; 2968 } else { 2969 assert(my_mem->unique_out() == this, "sanity"); 2970 del_req(Precedent); 2971 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2972 my_mem = NULL; 2973 } 2974 progress = true; 2975 } 2976 if (my_mem != NULL && my_mem->is_Mem()) { 2977 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2978 // Check for scalar replaced object reference. 2979 if( t_oop != NULL && t_oop->is_known_instance_field() && 2980 t_oop->offset() != Type::OffsetBot && 2981 t_oop->offset() != Type::OffsetTop) { 2982 eliminate = true; 2983 } 2984 } 2985 } else if (opc == Op_MemBarRelease) { 2986 // Final field stores. 2987 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 2988 if ((alloc != NULL) && alloc->is_Allocate() && 2989 alloc->as_Allocate()->does_not_escape_thread()) { 2990 // The allocated object does not escape. 2991 eliminate = true; 2992 } 2993 } 2994 if (eliminate) { 2995 // Replace MemBar projections by its inputs. 2996 PhaseIterGVN* igvn = phase->is_IterGVN(); 2997 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2998 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2999 // Must return either the original node (now dead) or a new node 3000 // (Do not return a top here, since that would break the uniqueness of top.) 3001 return new ConINode(TypeInt::ZERO); 3002 } 3003 } 3004 return progress ? this : NULL; 3005 } 3006 3007 //------------------------------Value------------------------------------------ 3008 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3009 if( !in(0) ) return Type::TOP; 3010 if( phase->type(in(0)) == Type::TOP ) 3011 return Type::TOP; 3012 return TypeTuple::MEMBAR; 3013 } 3014 3015 //------------------------------match------------------------------------------ 3016 // Construct projections for memory. 3017 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3018 switch (proj->_con) { 3019 case TypeFunc::Control: 3020 case TypeFunc::Memory: 3021 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3022 } 3023 ShouldNotReachHere(); 3024 return NULL; 3025 } 3026 3027 //===========================InitializeNode==================================== 3028 // SUMMARY: 3029 // This node acts as a memory barrier on raw memory, after some raw stores. 3030 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3031 // The Initialize can 'capture' suitably constrained stores as raw inits. 3032 // It can coalesce related raw stores into larger units (called 'tiles'). 3033 // It can avoid zeroing new storage for memory units which have raw inits. 3034 // At macro-expansion, it is marked 'complete', and does not optimize further. 3035 // 3036 // EXAMPLE: 3037 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3038 // ctl = incoming control; mem* = incoming memory 3039 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3040 // First allocate uninitialized memory and fill in the header: 3041 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3042 // ctl := alloc.Control; mem* := alloc.Memory* 3043 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3044 // Then initialize to zero the non-header parts of the raw memory block: 3045 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3046 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3047 // After the initialize node executes, the object is ready for service: 3048 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3049 // Suppose its body is immediately initialized as {1,2}: 3050 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3051 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3052 // mem.SLICE(#short[*]) := store2 3053 // 3054 // DETAILS: 3055 // An InitializeNode collects and isolates object initialization after 3056 // an AllocateNode and before the next possible safepoint. As a 3057 // memory barrier (MemBarNode), it keeps critical stores from drifting 3058 // down past any safepoint or any publication of the allocation. 3059 // Before this barrier, a newly-allocated object may have uninitialized bits. 3060 // After this barrier, it may be treated as a real oop, and GC is allowed. 3061 // 3062 // The semantics of the InitializeNode include an implicit zeroing of 3063 // the new object from object header to the end of the object. 3064 // (The object header and end are determined by the AllocateNode.) 3065 // 3066 // Certain stores may be added as direct inputs to the InitializeNode. 3067 // These stores must update raw memory, and they must be to addresses 3068 // derived from the raw address produced by AllocateNode, and with 3069 // a constant offset. They must be ordered by increasing offset. 3070 // The first one is at in(RawStores), the last at in(req()-1). 3071 // Unlike most memory operations, they are not linked in a chain, 3072 // but are displayed in parallel as users of the rawmem output of 3073 // the allocation. 3074 // 3075 // (See comments in InitializeNode::capture_store, which continue 3076 // the example given above.) 3077 // 3078 // When the associated Allocate is macro-expanded, the InitializeNode 3079 // may be rewritten to optimize collected stores. A ClearArrayNode 3080 // may also be created at that point to represent any required zeroing. 3081 // The InitializeNode is then marked 'complete', prohibiting further 3082 // capturing of nearby memory operations. 3083 // 3084 // During macro-expansion, all captured initializations which store 3085 // constant values of 32 bits or smaller are coalesced (if advantageous) 3086 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3087 // initialized in fewer memory operations. Memory words which are 3088 // covered by neither tiles nor non-constant stores are pre-zeroed 3089 // by explicit stores of zero. (The code shape happens to do all 3090 // zeroing first, then all other stores, with both sequences occurring 3091 // in order of ascending offsets.) 3092 // 3093 // Alternatively, code may be inserted between an AllocateNode and its 3094 // InitializeNode, to perform arbitrary initialization of the new object. 3095 // E.g., the object copying intrinsics insert complex data transfers here. 3096 // The initialization must then be marked as 'complete' disable the 3097 // built-in zeroing semantics and the collection of initializing stores. 3098 // 3099 // While an InitializeNode is incomplete, reads from the memory state 3100 // produced by it are optimizable if they match the control edge and 3101 // new oop address associated with the allocation/initialization. 3102 // They return a stored value (if the offset matches) or else zero. 3103 // A write to the memory state, if it matches control and address, 3104 // and if it is to a constant offset, may be 'captured' by the 3105 // InitializeNode. It is cloned as a raw memory operation and rewired 3106 // inside the initialization, to the raw oop produced by the allocation. 3107 // Operations on addresses which are provably distinct (e.g., to 3108 // other AllocateNodes) are allowed to bypass the initialization. 3109 // 3110 // The effect of all this is to consolidate object initialization 3111 // (both arrays and non-arrays, both piecewise and bulk) into a 3112 // single location, where it can be optimized as a unit. 3113 // 3114 // Only stores with an offset less than TrackedInitializationLimit words 3115 // will be considered for capture by an InitializeNode. This puts a 3116 // reasonable limit on the complexity of optimized initializations. 3117 3118 //---------------------------InitializeNode------------------------------------ 3119 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3120 : _is_complete(Incomplete), _does_not_escape(false), 3121 MemBarNode(C, adr_type, rawoop) 3122 { 3123 init_class_id(Class_Initialize); 3124 3125 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3126 assert(in(RawAddress) == rawoop, "proper init"); 3127 // Note: allocation() can be NULL, for secondary initialization barriers 3128 } 3129 3130 // Since this node is not matched, it will be processed by the 3131 // register allocator. Declare that there are no constraints 3132 // on the allocation of the RawAddress edge. 3133 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3134 // This edge should be set to top, by the set_complete. But be conservative. 3135 if (idx == InitializeNode::RawAddress) 3136 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3137 return RegMask::Empty; 3138 } 3139 3140 Node* InitializeNode::memory(uint alias_idx) { 3141 Node* mem = in(Memory); 3142 if (mem->is_MergeMem()) { 3143 return mem->as_MergeMem()->memory_at(alias_idx); 3144 } else { 3145 // incoming raw memory is not split 3146 return mem; 3147 } 3148 } 3149 3150 bool InitializeNode::is_non_zero() { 3151 if (is_complete()) return false; 3152 remove_extra_zeroes(); 3153 return (req() > RawStores); 3154 } 3155 3156 void InitializeNode::set_complete(PhaseGVN* phase) { 3157 assert(!is_complete(), "caller responsibility"); 3158 _is_complete = Complete; 3159 3160 // After this node is complete, it contains a bunch of 3161 // raw-memory initializations. There is no need for 3162 // it to have anything to do with non-raw memory effects. 3163 // Therefore, tell all non-raw users to re-optimize themselves, 3164 // after skipping the memory effects of this initialization. 3165 PhaseIterGVN* igvn = phase->is_IterGVN(); 3166 if (igvn) igvn->add_users_to_worklist(this); 3167 } 3168 3169 // convenience function 3170 // return false if the init contains any stores already 3171 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3172 InitializeNode* init = initialization(); 3173 if (init == NULL || init->is_complete()) return false; 3174 init->remove_extra_zeroes(); 3175 // for now, if this allocation has already collected any inits, bail: 3176 if (init->is_non_zero()) return false; 3177 init->set_complete(phase); 3178 return true; 3179 } 3180 3181 void InitializeNode::remove_extra_zeroes() { 3182 if (req() == RawStores) return; 3183 Node* zmem = zero_memory(); 3184 uint fill = RawStores; 3185 for (uint i = fill; i < req(); i++) { 3186 Node* n = in(i); 3187 if (n->is_top() || n == zmem) continue; // skip 3188 if (fill < i) set_req(fill, n); // compact 3189 ++fill; 3190 } 3191 // delete any empty spaces created: 3192 while (fill < req()) { 3193 del_req(fill); 3194 } 3195 } 3196 3197 // Helper for remembering which stores go with which offsets. 3198 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3199 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3200 intptr_t offset = -1; 3201 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3202 phase, offset); 3203 if (base == NULL) return -1; // something is dead, 3204 if (offset < 0) return -1; // dead, dead 3205 return offset; 3206 } 3207 3208 // Helper for proving that an initialization expression is 3209 // "simple enough" to be folded into an object initialization. 3210 // Attempts to prove that a store's initial value 'n' can be captured 3211 // within the initialization without creating a vicious cycle, such as: 3212 // { Foo p = new Foo(); p.next = p; } 3213 // True for constants and parameters and small combinations thereof. 3214 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3215 if (n == NULL) return true; // (can this really happen?) 3216 if (n->is_Proj()) n = n->in(0); 3217 if (n == this) return false; // found a cycle 3218 if (n->is_Con()) return true; 3219 if (n->is_Start()) return true; // params, etc., are OK 3220 if (n->is_Root()) return true; // even better 3221 3222 Node* ctl = n->in(0); 3223 if (ctl != NULL && !ctl->is_top()) { 3224 if (ctl->is_Proj()) ctl = ctl->in(0); 3225 if (ctl == this) return false; 3226 3227 // If we already know that the enclosing memory op is pinned right after 3228 // the init, then any control flow that the store has picked up 3229 // must have preceded the init, or else be equal to the init. 3230 // Even after loop optimizations (which might change control edges) 3231 // a store is never pinned *before* the availability of its inputs. 3232 if (!MemNode::all_controls_dominate(n, this)) 3233 return false; // failed to prove a good control 3234 } 3235 3236 // Check data edges for possible dependencies on 'this'. 3237 if ((count += 1) > 20) return false; // complexity limit 3238 for (uint i = 1; i < n->req(); i++) { 3239 Node* m = n->in(i); 3240 if (m == NULL || m == n || m->is_top()) continue; 3241 uint first_i = n->find_edge(m); 3242 if (i != first_i) continue; // process duplicate edge just once 3243 if (!detect_init_independence(m, count)) { 3244 return false; 3245 } 3246 } 3247 3248 return true; 3249 } 3250 3251 // Here are all the checks a Store must pass before it can be moved into 3252 // an initialization. Returns zero if a check fails. 3253 // On success, returns the (constant) offset to which the store applies, 3254 // within the initialized memory. 3255 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3256 const int FAIL = 0; 3257 if (st->is_unaligned_access()) { 3258 return FAIL; 3259 } 3260 if (st->req() != MemNode::ValueIn + 1) 3261 return FAIL; // an inscrutable StoreNode (card mark?) 3262 Node* ctl = st->in(MemNode::Control); 3263 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3264 return FAIL; // must be unconditional after the initialization 3265 Node* mem = st->in(MemNode::Memory); 3266 if (!(mem->is_Proj() && mem->in(0) == this)) 3267 return FAIL; // must not be preceded by other stores 3268 Node* adr = st->in(MemNode::Address); 3269 intptr_t offset; 3270 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3271 if (alloc == NULL) 3272 return FAIL; // inscrutable address 3273 if (alloc != allocation()) 3274 return FAIL; // wrong allocation! (store needs to float up) 3275 Node* val = st->in(MemNode::ValueIn); 3276 int complexity_count = 0; 3277 if (!detect_init_independence(val, complexity_count)) 3278 return FAIL; // stored value must be 'simple enough' 3279 3280 // The Store can be captured only if nothing after the allocation 3281 // and before the Store is using the memory location that the store 3282 // overwrites. 3283 bool failed = false; 3284 // If is_complete_with_arraycopy() is true the shape of the graph is 3285 // well defined and is safe so no need for extra checks. 3286 if (!is_complete_with_arraycopy()) { 3287 // We are going to look at each use of the memory state following 3288 // the allocation to make sure nothing reads the memory that the 3289 // Store writes. 3290 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3291 int alias_idx = phase->C->get_alias_index(t_adr); 3292 ResourceMark rm; 3293 Unique_Node_List mems; 3294 mems.push(mem); 3295 Node* unique_merge = NULL; 3296 for (uint next = 0; next < mems.size(); ++next) { 3297 Node *m = mems.at(next); 3298 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3299 Node *n = m->fast_out(j); 3300 if (n->outcnt() == 0) { 3301 continue; 3302 } 3303 if (n == st) { 3304 continue; 3305 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3306 // If the control of this use is different from the control 3307 // of the Store which is right after the InitializeNode then 3308 // this node cannot be between the InitializeNode and the 3309 // Store. 3310 continue; 3311 } else if (n->is_MergeMem()) { 3312 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3313 // We can hit a MergeMemNode (that will likely go away 3314 // later) that is a direct use of the memory state 3315 // following the InitializeNode on the same slice as the 3316 // store node that we'd like to capture. We need to check 3317 // the uses of the MergeMemNode. 3318 mems.push(n); 3319 } 3320 } else if (n->is_Mem()) { 3321 Node* other_adr = n->in(MemNode::Address); 3322 if (other_adr == adr) { 3323 failed = true; 3324 break; 3325 } else { 3326 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3327 if (other_t_adr != NULL) { 3328 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3329 if (other_alias_idx == alias_idx) { 3330 // A load from the same memory slice as the store right 3331 // after the InitializeNode. We check the control of the 3332 // object/array that is loaded from. If it's the same as 3333 // the store control then we cannot capture the store. 3334 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3335 Node* base = other_adr; 3336 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3337 base = base->in(AddPNode::Base); 3338 if (base != NULL) { 3339 base = base->uncast(); 3340 if (base->is_Proj() && base->in(0) == alloc) { 3341 failed = true; 3342 break; 3343 } 3344 } 3345 } 3346 } 3347 } 3348 } else { 3349 failed = true; 3350 break; 3351 } 3352 } 3353 } 3354 } 3355 if (failed) { 3356 if (!can_reshape) { 3357 // We decided we couldn't capture the store during parsing. We 3358 // should try again during the next IGVN once the graph is 3359 // cleaner. 3360 phase->C->record_for_igvn(st); 3361 } 3362 return FAIL; 3363 } 3364 3365 return offset; // success 3366 } 3367 3368 // Find the captured store in(i) which corresponds to the range 3369 // [start..start+size) in the initialized object. 3370 // If there is one, return its index i. If there isn't, return the 3371 // negative of the index where it should be inserted. 3372 // Return 0 if the queried range overlaps an initialization boundary 3373 // or if dead code is encountered. 3374 // If size_in_bytes is zero, do not bother with overlap checks. 3375 int InitializeNode::captured_store_insertion_point(intptr_t start, 3376 int size_in_bytes, 3377 PhaseTransform* phase) { 3378 const int FAIL = 0, MAX_STORE = BytesPerLong; 3379 3380 if (is_complete()) 3381 return FAIL; // arraycopy got here first; punt 3382 3383 assert(allocation() != NULL, "must be present"); 3384 3385 // no negatives, no header fields: 3386 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3387 3388 // after a certain size, we bail out on tracking all the stores: 3389 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3390 if (start >= ti_limit) return FAIL; 3391 3392 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3393 if (i >= limit) return -(int)i; // not found; here is where to put it 3394 3395 Node* st = in(i); 3396 intptr_t st_off = get_store_offset(st, phase); 3397 if (st_off < 0) { 3398 if (st != zero_memory()) { 3399 return FAIL; // bail out if there is dead garbage 3400 } 3401 } else if (st_off > start) { 3402 // ...we are done, since stores are ordered 3403 if (st_off < start + size_in_bytes) { 3404 return FAIL; // the next store overlaps 3405 } 3406 return -(int)i; // not found; here is where to put it 3407 } else if (st_off < start) { 3408 if (size_in_bytes != 0 && 3409 start < st_off + MAX_STORE && 3410 start < st_off + st->as_Store()->memory_size()) { 3411 return FAIL; // the previous store overlaps 3412 } 3413 } else { 3414 if (size_in_bytes != 0 && 3415 st->as_Store()->memory_size() != size_in_bytes) { 3416 return FAIL; // mismatched store size 3417 } 3418 return i; 3419 } 3420 3421 ++i; 3422 } 3423 } 3424 3425 // Look for a captured store which initializes at the offset 'start' 3426 // with the given size. If there is no such store, and no other 3427 // initialization interferes, then return zero_memory (the memory 3428 // projection of the AllocateNode). 3429 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3430 PhaseTransform* phase) { 3431 assert(stores_are_sane(phase), ""); 3432 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3433 if (i == 0) { 3434 return NULL; // something is dead 3435 } else if (i < 0) { 3436 return zero_memory(); // just primordial zero bits here 3437 } else { 3438 Node* st = in(i); // here is the store at this position 3439 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3440 return st; 3441 } 3442 } 3443 3444 // Create, as a raw pointer, an address within my new object at 'offset'. 3445 Node* InitializeNode::make_raw_address(intptr_t offset, 3446 PhaseTransform* phase) { 3447 Node* addr = in(RawAddress); 3448 if (offset != 0) { 3449 Compile* C = phase->C; 3450 addr = phase->transform( new AddPNode(C->top(), addr, 3451 phase->MakeConX(offset)) ); 3452 } 3453 return addr; 3454 } 3455 3456 // Clone the given store, converting it into a raw store 3457 // initializing a field or element of my new object. 3458 // Caller is responsible for retiring the original store, 3459 // with subsume_node or the like. 3460 // 3461 // From the example above InitializeNode::InitializeNode, 3462 // here are the old stores to be captured: 3463 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3464 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3465 // 3466 // Here is the changed code; note the extra edges on init: 3467 // alloc = (Allocate ...) 3468 // rawoop = alloc.RawAddress 3469 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3470 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3471 // init = (Initialize alloc.Control alloc.Memory rawoop 3472 // rawstore1 rawstore2) 3473 // 3474 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3475 PhaseTransform* phase, bool can_reshape) { 3476 assert(stores_are_sane(phase), ""); 3477 3478 if (start < 0) return NULL; 3479 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3480 3481 Compile* C = phase->C; 3482 int size_in_bytes = st->memory_size(); 3483 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3484 if (i == 0) return NULL; // bail out 3485 Node* prev_mem = NULL; // raw memory for the captured store 3486 if (i > 0) { 3487 prev_mem = in(i); // there is a pre-existing store under this one 3488 set_req(i, C->top()); // temporarily disconnect it 3489 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3490 } else { 3491 i = -i; // no pre-existing store 3492 prev_mem = zero_memory(); // a slice of the newly allocated object 3493 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3494 set_req(--i, C->top()); // reuse this edge; it has been folded away 3495 else 3496 ins_req(i, C->top()); // build a new edge 3497 } 3498 Node* new_st = st->clone(); 3499 new_st->set_req(MemNode::Control, in(Control)); 3500 new_st->set_req(MemNode::Memory, prev_mem); 3501 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3502 new_st = phase->transform(new_st); 3503 3504 // At this point, new_st might have swallowed a pre-existing store 3505 // at the same offset, or perhaps new_st might have disappeared, 3506 // if it redundantly stored the same value (or zero to fresh memory). 3507 3508 // In any case, wire it in: 3509 phase->igvn_rehash_node_delayed(this); 3510 set_req(i, new_st); 3511 3512 // The caller may now kill the old guy. 3513 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3514 assert(check_st == new_st || check_st == NULL, "must be findable"); 3515 assert(!is_complete(), ""); 3516 return new_st; 3517 } 3518 3519 static bool store_constant(jlong* tiles, int num_tiles, 3520 intptr_t st_off, int st_size, 3521 jlong con) { 3522 if ((st_off & (st_size-1)) != 0) 3523 return false; // strange store offset (assume size==2**N) 3524 address addr = (address)tiles + st_off; 3525 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3526 switch (st_size) { 3527 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3528 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3529 case sizeof(jint): *(jint*) addr = (jint) con; break; 3530 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3531 default: return false; // strange store size (detect size!=2**N here) 3532 } 3533 return true; // return success to caller 3534 } 3535 3536 // Coalesce subword constants into int constants and possibly 3537 // into long constants. The goal, if the CPU permits, 3538 // is to initialize the object with a small number of 64-bit tiles. 3539 // Also, convert floating-point constants to bit patterns. 3540 // Non-constants are not relevant to this pass. 3541 // 3542 // In terms of the running example on InitializeNode::InitializeNode 3543 // and InitializeNode::capture_store, here is the transformation 3544 // of rawstore1 and rawstore2 into rawstore12: 3545 // alloc = (Allocate ...) 3546 // rawoop = alloc.RawAddress 3547 // tile12 = 0x00010002 3548 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3549 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3550 // 3551 void 3552 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3553 Node* size_in_bytes, 3554 PhaseGVN* phase) { 3555 Compile* C = phase->C; 3556 3557 assert(stores_are_sane(phase), ""); 3558 // Note: After this pass, they are not completely sane, 3559 // since there may be some overlaps. 3560 3561 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3562 3563 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3564 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3565 size_limit = MIN2(size_limit, ti_limit); 3566 size_limit = align_size_up(size_limit, BytesPerLong); 3567 int num_tiles = size_limit / BytesPerLong; 3568 3569 // allocate space for the tile map: 3570 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3571 jlong tiles_buf[small_len]; 3572 Node* nodes_buf[small_len]; 3573 jlong inits_buf[small_len]; 3574 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3575 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3576 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3577 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3578 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3579 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3580 // tiles: exact bitwise model of all primitive constants 3581 // nodes: last constant-storing node subsumed into the tiles model 3582 // inits: which bytes (in each tile) are touched by any initializations 3583 3584 //// Pass A: Fill in the tile model with any relevant stores. 3585 3586 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3587 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3588 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3589 Node* zmem = zero_memory(); // initially zero memory state 3590 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3591 Node* st = in(i); 3592 intptr_t st_off = get_store_offset(st, phase); 3593 3594 // Figure out the store's offset and constant value: 3595 if (st_off < header_size) continue; //skip (ignore header) 3596 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3597 int st_size = st->as_Store()->memory_size(); 3598 if (st_off + st_size > size_limit) break; 3599 3600 // Record which bytes are touched, whether by constant or not. 3601 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3602 continue; // skip (strange store size) 3603 3604 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3605 if (!val->singleton()) continue; //skip (non-con store) 3606 BasicType type = val->basic_type(); 3607 3608 jlong con = 0; 3609 switch (type) { 3610 case T_INT: con = val->is_int()->get_con(); break; 3611 case T_LONG: con = val->is_long()->get_con(); break; 3612 case T_FLOAT: con = jint_cast(val->getf()); break; 3613 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3614 default: continue; //skip (odd store type) 3615 } 3616 3617 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3618 st->Opcode() == Op_StoreL) { 3619 continue; // This StoreL is already optimal. 3620 } 3621 3622 // Store down the constant. 3623 store_constant(tiles, num_tiles, st_off, st_size, con); 3624 3625 intptr_t j = st_off >> LogBytesPerLong; 3626 3627 if (type == T_INT && st_size == BytesPerInt 3628 && (st_off & BytesPerInt) == BytesPerInt) { 3629 jlong lcon = tiles[j]; 3630 if (!Matcher::isSimpleConstant64(lcon) && 3631 st->Opcode() == Op_StoreI) { 3632 // This StoreI is already optimal by itself. 3633 jint* intcon = (jint*) &tiles[j]; 3634 intcon[1] = 0; // undo the store_constant() 3635 3636 // If the previous store is also optimal by itself, back up and 3637 // undo the action of the previous loop iteration... if we can. 3638 // But if we can't, just let the previous half take care of itself. 3639 st = nodes[j]; 3640 st_off -= BytesPerInt; 3641 con = intcon[0]; 3642 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3643 assert(st_off >= header_size, "still ignoring header"); 3644 assert(get_store_offset(st, phase) == st_off, "must be"); 3645 assert(in(i-1) == zmem, "must be"); 3646 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3647 assert(con == tcon->is_int()->get_con(), "must be"); 3648 // Undo the effects of the previous loop trip, which swallowed st: 3649 intcon[0] = 0; // undo store_constant() 3650 set_req(i-1, st); // undo set_req(i, zmem) 3651 nodes[j] = NULL; // undo nodes[j] = st 3652 --old_subword; // undo ++old_subword 3653 } 3654 continue; // This StoreI is already optimal. 3655 } 3656 } 3657 3658 // This store is not needed. 3659 set_req(i, zmem); 3660 nodes[j] = st; // record for the moment 3661 if (st_size < BytesPerLong) // something has changed 3662 ++old_subword; // includes int/float, but who's counting... 3663 else ++old_long; 3664 } 3665 3666 if ((old_subword + old_long) == 0) 3667 return; // nothing more to do 3668 3669 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3670 // Be sure to insert them before overlapping non-constant stores. 3671 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3672 for (int j = 0; j < num_tiles; j++) { 3673 jlong con = tiles[j]; 3674 jlong init = inits[j]; 3675 if (con == 0) continue; 3676 jint con0, con1; // split the constant, address-wise 3677 jint init0, init1; // split the init map, address-wise 3678 { union { jlong con; jint intcon[2]; } u; 3679 u.con = con; 3680 con0 = u.intcon[0]; 3681 con1 = u.intcon[1]; 3682 u.con = init; 3683 init0 = u.intcon[0]; 3684 init1 = u.intcon[1]; 3685 } 3686 3687 Node* old = nodes[j]; 3688 assert(old != NULL, "need the prior store"); 3689 intptr_t offset = (j * BytesPerLong); 3690 3691 bool split = !Matcher::isSimpleConstant64(con); 3692 3693 if (offset < header_size) { 3694 assert(offset + BytesPerInt >= header_size, "second int counts"); 3695 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3696 split = true; // only the second word counts 3697 // Example: int a[] = { 42 ... } 3698 } else if (con0 == 0 && init0 == -1) { 3699 split = true; // first word is covered by full inits 3700 // Example: int a[] = { ... foo(), 42 ... } 3701 } else if (con1 == 0 && init1 == -1) { 3702 split = true; // second word is covered by full inits 3703 // Example: int a[] = { ... 42, foo() ... } 3704 } 3705 3706 // Here's a case where init0 is neither 0 nor -1: 3707 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3708 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3709 // In this case the tile is not split; it is (jlong)42. 3710 // The big tile is stored down, and then the foo() value is inserted. 3711 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3712 3713 Node* ctl = old->in(MemNode::Control); 3714 Node* adr = make_raw_address(offset, phase); 3715 const TypePtr* atp = TypeRawPtr::BOTTOM; 3716 3717 // One or two coalesced stores to plop down. 3718 Node* st[2]; 3719 intptr_t off[2]; 3720 int nst = 0; 3721 if (!split) { 3722 ++new_long; 3723 off[nst] = offset; 3724 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3725 phase->longcon(con), T_LONG, MemNode::unordered); 3726 } else { 3727 // Omit either if it is a zero. 3728 if (con0 != 0) { 3729 ++new_int; 3730 off[nst] = offset; 3731 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3732 phase->intcon(con0), T_INT, MemNode::unordered); 3733 } 3734 if (con1 != 0) { 3735 ++new_int; 3736 offset += BytesPerInt; 3737 adr = make_raw_address(offset, phase); 3738 off[nst] = offset; 3739 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3740 phase->intcon(con1), T_INT, MemNode::unordered); 3741 } 3742 } 3743 3744 // Insert second store first, then the first before the second. 3745 // Insert each one just before any overlapping non-constant stores. 3746 while (nst > 0) { 3747 Node* st1 = st[--nst]; 3748 C->copy_node_notes_to(st1, old); 3749 st1 = phase->transform(st1); 3750 offset = off[nst]; 3751 assert(offset >= header_size, "do not smash header"); 3752 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3753 guarantee(ins_idx != 0, "must re-insert constant store"); 3754 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3755 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3756 set_req(--ins_idx, st1); 3757 else 3758 ins_req(ins_idx, st1); 3759 } 3760 } 3761 3762 if (PrintCompilation && WizardMode) 3763 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3764 old_subword, old_long, new_int, new_long); 3765 if (C->log() != NULL) 3766 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3767 old_subword, old_long, new_int, new_long); 3768 3769 // Clean up any remaining occurrences of zmem: 3770 remove_extra_zeroes(); 3771 } 3772 3773 // Explore forward from in(start) to find the first fully initialized 3774 // word, and return its offset. Skip groups of subword stores which 3775 // together initialize full words. If in(start) is itself part of a 3776 // fully initialized word, return the offset of in(start). If there 3777 // are no following full-word stores, or if something is fishy, return 3778 // a negative value. 3779 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3780 int int_map = 0; 3781 intptr_t int_map_off = 0; 3782 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3783 3784 for (uint i = start, limit = req(); i < limit; i++) { 3785 Node* st = in(i); 3786 3787 intptr_t st_off = get_store_offset(st, phase); 3788 if (st_off < 0) break; // return conservative answer 3789 3790 int st_size = st->as_Store()->memory_size(); 3791 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3792 return st_off; // we found a complete word init 3793 } 3794 3795 // update the map: 3796 3797 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3798 if (this_int_off != int_map_off) { 3799 // reset the map: 3800 int_map = 0; 3801 int_map_off = this_int_off; 3802 } 3803 3804 int subword_off = st_off - this_int_off; 3805 int_map |= right_n_bits(st_size) << subword_off; 3806 if ((int_map & FULL_MAP) == FULL_MAP) { 3807 return this_int_off; // we found a complete word init 3808 } 3809 3810 // Did this store hit or cross the word boundary? 3811 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3812 if (next_int_off == this_int_off + BytesPerInt) { 3813 // We passed the current int, without fully initializing it. 3814 int_map_off = next_int_off; 3815 int_map >>= BytesPerInt; 3816 } else if (next_int_off > this_int_off + BytesPerInt) { 3817 // We passed the current and next int. 3818 return this_int_off + BytesPerInt; 3819 } 3820 } 3821 3822 return -1; 3823 } 3824 3825 3826 // Called when the associated AllocateNode is expanded into CFG. 3827 // At this point, we may perform additional optimizations. 3828 // Linearize the stores by ascending offset, to make memory 3829 // activity as coherent as possible. 3830 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3831 intptr_t header_size, 3832 Node* size_in_bytes, 3833 PhaseGVN* phase) { 3834 assert(!is_complete(), "not already complete"); 3835 assert(stores_are_sane(phase), ""); 3836 assert(allocation() != NULL, "must be present"); 3837 3838 remove_extra_zeroes(); 3839 3840 if (ReduceFieldZeroing || ReduceBulkZeroing) 3841 // reduce instruction count for common initialization patterns 3842 coalesce_subword_stores(header_size, size_in_bytes, phase); 3843 3844 Node* zmem = zero_memory(); // initially zero memory state 3845 Node* inits = zmem; // accumulating a linearized chain of inits 3846 #ifdef ASSERT 3847 intptr_t first_offset = allocation()->minimum_header_size(); 3848 intptr_t last_init_off = first_offset; // previous init offset 3849 intptr_t last_init_end = first_offset; // previous init offset+size 3850 intptr_t last_tile_end = first_offset; // previous tile offset+size 3851 #endif 3852 intptr_t zeroes_done = header_size; 3853 3854 bool do_zeroing = true; // we might give up if inits are very sparse 3855 int big_init_gaps = 0; // how many large gaps have we seen? 3856 3857 if (UseTLAB && ZeroTLAB) do_zeroing = false; 3858 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3859 3860 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3861 Node* st = in(i); 3862 intptr_t st_off = get_store_offset(st, phase); 3863 if (st_off < 0) 3864 break; // unknown junk in the inits 3865 if (st->in(MemNode::Memory) != zmem) 3866 break; // complicated store chains somehow in list 3867 3868 int st_size = st->as_Store()->memory_size(); 3869 intptr_t next_init_off = st_off + st_size; 3870 3871 if (do_zeroing && zeroes_done < next_init_off) { 3872 // See if this store needs a zero before it or under it. 3873 intptr_t zeroes_needed = st_off; 3874 3875 if (st_size < BytesPerInt) { 3876 // Look for subword stores which only partially initialize words. 3877 // If we find some, we must lay down some word-level zeroes first, 3878 // underneath the subword stores. 3879 // 3880 // Examples: 3881 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3882 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3883 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3884 // 3885 // Note: coalesce_subword_stores may have already done this, 3886 // if it was prompted by constant non-zero subword initializers. 3887 // But this case can still arise with non-constant stores. 3888 3889 intptr_t next_full_store = find_next_fullword_store(i, phase); 3890 3891 // In the examples above: 3892 // in(i) p q r s x y z 3893 // st_off 12 13 14 15 12 13 14 3894 // st_size 1 1 1 1 1 1 1 3895 // next_full_s. 12 16 16 16 16 16 16 3896 // z's_done 12 16 16 16 12 16 12 3897 // z's_needed 12 16 16 16 16 16 16 3898 // zsize 0 0 0 0 4 0 4 3899 if (next_full_store < 0) { 3900 // Conservative tack: Zero to end of current word. 3901 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3902 } else { 3903 // Zero to beginning of next fully initialized word. 3904 // Or, don't zero at all, if we are already in that word. 3905 assert(next_full_store >= zeroes_needed, "must go forward"); 3906 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3907 zeroes_needed = next_full_store; 3908 } 3909 } 3910 3911 if (zeroes_needed > zeroes_done) { 3912 intptr_t zsize = zeroes_needed - zeroes_done; 3913 // Do some incremental zeroing on rawmem, in parallel with inits. 3914 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3915 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3916 zeroes_done, zeroes_needed, 3917 phase); 3918 zeroes_done = zeroes_needed; 3919 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 3920 do_zeroing = false; // leave the hole, next time 3921 } 3922 } 3923 3924 // Collect the store and move on: 3925 st->set_req(MemNode::Memory, inits); 3926 inits = st; // put it on the linearized chain 3927 set_req(i, zmem); // unhook from previous position 3928 3929 if (zeroes_done == st_off) 3930 zeroes_done = next_init_off; 3931 3932 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3933 3934 #ifdef ASSERT 3935 // Various order invariants. Weaker than stores_are_sane because 3936 // a large constant tile can be filled in by smaller non-constant stores. 3937 assert(st_off >= last_init_off, "inits do not reverse"); 3938 last_init_off = st_off; 3939 const Type* val = NULL; 3940 if (st_size >= BytesPerInt && 3941 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3942 (int)val->basic_type() < (int)T_OBJECT) { 3943 assert(st_off >= last_tile_end, "tiles do not overlap"); 3944 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3945 last_tile_end = MAX2(last_tile_end, next_init_off); 3946 } else { 3947 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3948 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3949 assert(st_off >= last_init_end, "inits do not overlap"); 3950 last_init_end = next_init_off; // it's a non-tile 3951 } 3952 #endif //ASSERT 3953 } 3954 3955 remove_extra_zeroes(); // clear out all the zmems left over 3956 add_req(inits); 3957 3958 if (!(UseTLAB && ZeroTLAB)) { 3959 // If anything remains to be zeroed, zero it all now. 3960 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3961 // if it is the last unused 4 bytes of an instance, forget about it 3962 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3963 if (zeroes_done + BytesPerLong >= size_limit) { 3964 AllocateNode* alloc = allocation(); 3965 assert(alloc != NULL, "must be present"); 3966 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 3967 Node* klass_node = alloc->in(AllocateNode::KlassNode); 3968 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3969 if (zeroes_done == k->layout_helper()) 3970 zeroes_done = size_limit; 3971 } 3972 } 3973 if (zeroes_done < size_limit) { 3974 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3975 zeroes_done, size_in_bytes, phase); 3976 } 3977 } 3978 3979 set_complete(phase); 3980 return rawmem; 3981 } 3982 3983 3984 #ifdef ASSERT 3985 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3986 if (is_complete()) 3987 return true; // stores could be anything at this point 3988 assert(allocation() != NULL, "must be present"); 3989 intptr_t last_off = allocation()->minimum_header_size(); 3990 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3991 Node* st = in(i); 3992 intptr_t st_off = get_store_offset(st, phase); 3993 if (st_off < 0) continue; // ignore dead garbage 3994 if (last_off > st_off) { 3995 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 3996 this->dump(2); 3997 assert(false, "ascending store offsets"); 3998 return false; 3999 } 4000 last_off = st_off + st->as_Store()->memory_size(); 4001 } 4002 return true; 4003 } 4004 #endif //ASSERT 4005 4006 4007 4008 4009 //============================MergeMemNode===================================== 4010 // 4011 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4012 // contributing store or call operations. Each contributor provides the memory 4013 // state for a particular "alias type" (see Compile::alias_type). For example, 4014 // if a MergeMem has an input X for alias category #6, then any memory reference 4015 // to alias category #6 may use X as its memory state input, as an exact equivalent 4016 // to using the MergeMem as a whole. 4017 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4018 // 4019 // (Here, the <N> notation gives the index of the relevant adr_type.) 4020 // 4021 // In one special case (and more cases in the future), alias categories overlap. 4022 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4023 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4024 // it is exactly equivalent to that state W: 4025 // MergeMem(<Bot>: W) <==> W 4026 // 4027 // Usually, the merge has more than one input. In that case, where inputs 4028 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4029 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4030 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4031 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4032 // 4033 // A merge can take a "wide" memory state as one of its narrow inputs. 4034 // This simply means that the merge observes out only the relevant parts of 4035 // the wide input. That is, wide memory states arriving at narrow merge inputs 4036 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4037 // 4038 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4039 // and that memory slices "leak through": 4040 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4041 // 4042 // But, in such a cascade, repeated memory slices can "block the leak": 4043 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4044 // 4045 // In the last example, Y is not part of the combined memory state of the 4046 // outermost MergeMem. The system must, of course, prevent unschedulable 4047 // memory states from arising, so you can be sure that the state Y is somehow 4048 // a precursor to state Y'. 4049 // 4050 // 4051 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4052 // of each MergeMemNode array are exactly the numerical alias indexes, including 4053 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4054 // Compile::alias_type (and kin) produce and manage these indexes. 4055 // 4056 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4057 // (Note that this provides quick access to the top node inside MergeMem methods, 4058 // without the need to reach out via TLS to Compile::current.) 4059 // 4060 // As a consequence of what was just described, a MergeMem that represents a full 4061 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4062 // containing all alias categories. 4063 // 4064 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4065 // 4066 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4067 // a memory state for the alias type <N>, or else the top node, meaning that 4068 // there is no particular input for that alias type. Note that the length of 4069 // a MergeMem is variable, and may be extended at any time to accommodate new 4070 // memory states at larger alias indexes. When merges grow, they are of course 4071 // filled with "top" in the unused in() positions. 4072 // 4073 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4074 // (Top was chosen because it works smoothly with passes like GCM.) 4075 // 4076 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4077 // the type of random VM bits like TLS references.) Since it is always the 4078 // first non-Bot memory slice, some low-level loops use it to initialize an 4079 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4080 // 4081 // 4082 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4083 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4084 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4085 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4086 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4087 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4088 // 4089 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4090 // really that different from the other memory inputs. An abbreviation called 4091 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4092 // 4093 // 4094 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4095 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4096 // that "emerges though" the base memory will be marked as excluding the alias types 4097 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4098 // 4099 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4100 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4101 // 4102 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4103 // (It is currently unimplemented.) As you can see, the resulting merge is 4104 // actually a disjoint union of memory states, rather than an overlay. 4105 // 4106 4107 //------------------------------MergeMemNode----------------------------------- 4108 Node* MergeMemNode::make_empty_memory() { 4109 Node* empty_memory = (Node*) Compile::current()->top(); 4110 assert(empty_memory->is_top(), "correct sentinel identity"); 4111 return empty_memory; 4112 } 4113 4114 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4115 init_class_id(Class_MergeMem); 4116 // all inputs are nullified in Node::Node(int) 4117 // set_input(0, NULL); // no control input 4118 4119 // Initialize the edges uniformly to top, for starters. 4120 Node* empty_mem = make_empty_memory(); 4121 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4122 init_req(i,empty_mem); 4123 } 4124 assert(empty_memory() == empty_mem, ""); 4125 4126 if( new_base != NULL && new_base->is_MergeMem() ) { 4127 MergeMemNode* mdef = new_base->as_MergeMem(); 4128 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4129 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4130 mms.set_memory(mms.memory2()); 4131 } 4132 assert(base_memory() == mdef->base_memory(), ""); 4133 } else { 4134 set_base_memory(new_base); 4135 } 4136 } 4137 4138 // Make a new, untransformed MergeMem with the same base as 'mem'. 4139 // If mem is itself a MergeMem, populate the result with the same edges. 4140 MergeMemNode* MergeMemNode::make(Node* mem) { 4141 return new MergeMemNode(mem); 4142 } 4143 4144 //------------------------------cmp-------------------------------------------- 4145 uint MergeMemNode::hash() const { return NO_HASH; } 4146 uint MergeMemNode::cmp( const Node &n ) const { 4147 return (&n == this); // Always fail except on self 4148 } 4149 4150 //------------------------------Identity--------------------------------------- 4151 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4152 // Identity if this merge point does not record any interesting memory 4153 // disambiguations. 4154 Node* base_mem = base_memory(); 4155 Node* empty_mem = empty_memory(); 4156 if (base_mem != empty_mem) { // Memory path is not dead? 4157 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4158 Node* mem = in(i); 4159 if (mem != empty_mem && mem != base_mem) { 4160 return this; // Many memory splits; no change 4161 } 4162 } 4163 } 4164 return base_mem; // No memory splits; ID on the one true input 4165 } 4166 4167 //------------------------------Ideal------------------------------------------ 4168 // This method is invoked recursively on chains of MergeMem nodes 4169 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4170 // Remove chain'd MergeMems 4171 // 4172 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4173 // relative to the "in(Bot)". Since we are patching both at the same time, 4174 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4175 // but rewrite each "in(i)" relative to the new "in(Bot)". 4176 Node *progress = NULL; 4177 4178 4179 Node* old_base = base_memory(); 4180 Node* empty_mem = empty_memory(); 4181 if (old_base == empty_mem) 4182 return NULL; // Dead memory path. 4183 4184 MergeMemNode* old_mbase; 4185 if (old_base != NULL && old_base->is_MergeMem()) 4186 old_mbase = old_base->as_MergeMem(); 4187 else 4188 old_mbase = NULL; 4189 Node* new_base = old_base; 4190 4191 // simplify stacked MergeMems in base memory 4192 if (old_mbase) new_base = old_mbase->base_memory(); 4193 4194 // the base memory might contribute new slices beyond my req() 4195 if (old_mbase) grow_to_match(old_mbase); 4196 4197 // Look carefully at the base node if it is a phi. 4198 PhiNode* phi_base; 4199 if (new_base != NULL && new_base->is_Phi()) 4200 phi_base = new_base->as_Phi(); 4201 else 4202 phi_base = NULL; 4203 4204 Node* phi_reg = NULL; 4205 uint phi_len = (uint)-1; 4206 if (phi_base != NULL && !phi_base->is_copy()) { 4207 // do not examine phi if degraded to a copy 4208 phi_reg = phi_base->region(); 4209 phi_len = phi_base->req(); 4210 // see if the phi is unfinished 4211 for (uint i = 1; i < phi_len; i++) { 4212 if (phi_base->in(i) == NULL) { 4213 // incomplete phi; do not look at it yet! 4214 phi_reg = NULL; 4215 phi_len = (uint)-1; 4216 break; 4217 } 4218 } 4219 } 4220 4221 // Note: We do not call verify_sparse on entry, because inputs 4222 // can normalize to the base_memory via subsume_node or similar 4223 // mechanisms. This method repairs that damage. 4224 4225 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4226 4227 // Look at each slice. 4228 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4229 Node* old_in = in(i); 4230 // calculate the old memory value 4231 Node* old_mem = old_in; 4232 if (old_mem == empty_mem) old_mem = old_base; 4233 assert(old_mem == memory_at(i), ""); 4234 4235 // maybe update (reslice) the old memory value 4236 4237 // simplify stacked MergeMems 4238 Node* new_mem = old_mem; 4239 MergeMemNode* old_mmem; 4240 if (old_mem != NULL && old_mem->is_MergeMem()) 4241 old_mmem = old_mem->as_MergeMem(); 4242 else 4243 old_mmem = NULL; 4244 if (old_mmem == this) { 4245 // This can happen if loops break up and safepoints disappear. 4246 // A merge of BotPtr (default) with a RawPtr memory derived from a 4247 // safepoint can be rewritten to a merge of the same BotPtr with 4248 // the BotPtr phi coming into the loop. If that phi disappears 4249 // also, we can end up with a self-loop of the mergemem. 4250 // In general, if loops degenerate and memory effects disappear, 4251 // a mergemem can be left looking at itself. This simply means 4252 // that the mergemem's default should be used, since there is 4253 // no longer any apparent effect on this slice. 4254 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4255 // from start. Update the input to TOP. 4256 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4257 } 4258 else if (old_mmem != NULL) { 4259 new_mem = old_mmem->memory_at(i); 4260 } 4261 // else preceding memory was not a MergeMem 4262 4263 // replace equivalent phis (unfortunately, they do not GVN together) 4264 if (new_mem != NULL && new_mem != new_base && 4265 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4266 if (new_mem->is_Phi()) { 4267 PhiNode* phi_mem = new_mem->as_Phi(); 4268 for (uint i = 1; i < phi_len; i++) { 4269 if (phi_base->in(i) != phi_mem->in(i)) { 4270 phi_mem = NULL; 4271 break; 4272 } 4273 } 4274 if (phi_mem != NULL) { 4275 // equivalent phi nodes; revert to the def 4276 new_mem = new_base; 4277 } 4278 } 4279 } 4280 4281 // maybe store down a new value 4282 Node* new_in = new_mem; 4283 if (new_in == new_base) new_in = empty_mem; 4284 4285 if (new_in != old_in) { 4286 // Warning: Do not combine this "if" with the previous "if" 4287 // A memory slice might have be be rewritten even if it is semantically 4288 // unchanged, if the base_memory value has changed. 4289 set_req(i, new_in); 4290 progress = this; // Report progress 4291 } 4292 } 4293 4294 if (new_base != old_base) { 4295 set_req(Compile::AliasIdxBot, new_base); 4296 // Don't use set_base_memory(new_base), because we need to update du. 4297 assert(base_memory() == new_base, ""); 4298 progress = this; 4299 } 4300 4301 if( base_memory() == this ) { 4302 // a self cycle indicates this memory path is dead 4303 set_req(Compile::AliasIdxBot, empty_mem); 4304 } 4305 4306 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4307 // Recursion must occur after the self cycle check above 4308 if( base_memory()->is_MergeMem() ) { 4309 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4310 Node *m = phase->transform(new_mbase); // Rollup any cycles 4311 if( m != NULL && 4312 (m->is_top() || 4313 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4314 // propagate rollup of dead cycle to self 4315 set_req(Compile::AliasIdxBot, empty_mem); 4316 } 4317 } 4318 4319 if( base_memory() == empty_mem ) { 4320 progress = this; 4321 // Cut inputs during Parse phase only. 4322 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4323 if( !can_reshape ) { 4324 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4325 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4326 } 4327 } 4328 } 4329 4330 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4331 // Check if PhiNode::Ideal's "Split phis through memory merges" 4332 // transform should be attempted. Look for this->phi->this cycle. 4333 uint merge_width = req(); 4334 if (merge_width > Compile::AliasIdxRaw) { 4335 PhiNode* phi = base_memory()->as_Phi(); 4336 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4337 if (phi->in(i) == this) { 4338 phase->is_IterGVN()->_worklist.push(phi); 4339 break; 4340 } 4341 } 4342 } 4343 } 4344 4345 assert(progress || verify_sparse(), "please, no dups of base"); 4346 return progress; 4347 } 4348 4349 //-------------------------set_base_memory------------------------------------- 4350 void MergeMemNode::set_base_memory(Node *new_base) { 4351 Node* empty_mem = empty_memory(); 4352 set_req(Compile::AliasIdxBot, new_base); 4353 assert(memory_at(req()) == new_base, "must set default memory"); 4354 // Clear out other occurrences of new_base: 4355 if (new_base != empty_mem) { 4356 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4357 if (in(i) == new_base) set_req(i, empty_mem); 4358 } 4359 } 4360 } 4361 4362 //------------------------------out_RegMask------------------------------------ 4363 const RegMask &MergeMemNode::out_RegMask() const { 4364 return RegMask::Empty; 4365 } 4366 4367 //------------------------------dump_spec-------------------------------------- 4368 #ifndef PRODUCT 4369 void MergeMemNode::dump_spec(outputStream *st) const { 4370 st->print(" {"); 4371 Node* base_mem = base_memory(); 4372 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4373 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4374 if (mem == base_mem) { st->print(" -"); continue; } 4375 st->print( " N%d:", mem->_idx ); 4376 Compile::current()->get_adr_type(i)->dump_on(st); 4377 } 4378 st->print(" }"); 4379 } 4380 #endif // !PRODUCT 4381 4382 4383 #ifdef ASSERT 4384 static bool might_be_same(Node* a, Node* b) { 4385 if (a == b) return true; 4386 if (!(a->is_Phi() || b->is_Phi())) return false; 4387 // phis shift around during optimization 4388 return true; // pretty stupid... 4389 } 4390 4391 // verify a narrow slice (either incoming or outgoing) 4392 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4393 if (!VerifyAliases) return; // don't bother to verify unless requested 4394 if (is_error_reported()) return; // muzzle asserts when debugging an error 4395 if (Node::in_dump()) return; // muzzle asserts when printing 4396 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4397 assert(n != NULL, ""); 4398 // Elide intervening MergeMem's 4399 while (n->is_MergeMem()) { 4400 n = n->as_MergeMem()->memory_at(alias_idx); 4401 } 4402 Compile* C = Compile::current(); 4403 const TypePtr* n_adr_type = n->adr_type(); 4404 if (n == m->empty_memory()) { 4405 // Implicit copy of base_memory() 4406 } else if (n_adr_type != TypePtr::BOTTOM) { 4407 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4408 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4409 } else { 4410 // A few places like make_runtime_call "know" that VM calls are narrow, 4411 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4412 bool expected_wide_mem = false; 4413 if (n == m->base_memory()) { 4414 expected_wide_mem = true; 4415 } else if (alias_idx == Compile::AliasIdxRaw || 4416 n == m->memory_at(Compile::AliasIdxRaw)) { 4417 expected_wide_mem = true; 4418 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4419 // memory can "leak through" calls on channels that 4420 // are write-once. Allow this also. 4421 expected_wide_mem = true; 4422 } 4423 assert(expected_wide_mem, "expected narrow slice replacement"); 4424 } 4425 } 4426 #else // !ASSERT 4427 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4428 #endif 4429 4430 4431 //-----------------------------memory_at--------------------------------------- 4432 Node* MergeMemNode::memory_at(uint alias_idx) const { 4433 assert(alias_idx >= Compile::AliasIdxRaw || 4434 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4435 "must avoid base_memory and AliasIdxTop"); 4436 4437 // Otherwise, it is a narrow slice. 4438 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4439 Compile *C = Compile::current(); 4440 if (is_empty_memory(n)) { 4441 // the array is sparse; empty slots are the "top" node 4442 n = base_memory(); 4443 assert(Node::in_dump() 4444 || n == NULL || n->bottom_type() == Type::TOP 4445 || n->adr_type() == NULL // address is TOP 4446 || n->adr_type() == TypePtr::BOTTOM 4447 || n->adr_type() == TypeRawPtr::BOTTOM 4448 || Compile::current()->AliasLevel() == 0, 4449 "must be a wide memory"); 4450 // AliasLevel == 0 if we are organizing the memory states manually. 4451 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4452 } else { 4453 // make sure the stored slice is sane 4454 #ifdef ASSERT 4455 if (is_error_reported() || Node::in_dump()) { 4456 } else if (might_be_same(n, base_memory())) { 4457 // Give it a pass: It is a mostly harmless repetition of the base. 4458 // This can arise normally from node subsumption during optimization. 4459 } else { 4460 verify_memory_slice(this, alias_idx, n); 4461 } 4462 #endif 4463 } 4464 return n; 4465 } 4466 4467 //---------------------------set_memory_at------------------------------------- 4468 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4469 verify_memory_slice(this, alias_idx, n); 4470 Node* empty_mem = empty_memory(); 4471 if (n == base_memory()) n = empty_mem; // collapse default 4472 uint need_req = alias_idx+1; 4473 if (req() < need_req) { 4474 if (n == empty_mem) return; // already the default, so do not grow me 4475 // grow the sparse array 4476 do { 4477 add_req(empty_mem); 4478 } while (req() < need_req); 4479 } 4480 set_req( alias_idx, n ); 4481 } 4482 4483 4484 4485 //--------------------------iteration_setup------------------------------------ 4486 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4487 if (other != NULL) { 4488 grow_to_match(other); 4489 // invariant: the finite support of mm2 is within mm->req() 4490 #ifdef ASSERT 4491 for (uint i = req(); i < other->req(); i++) { 4492 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4493 } 4494 #endif 4495 } 4496 // Replace spurious copies of base_memory by top. 4497 Node* base_mem = base_memory(); 4498 if (base_mem != NULL && !base_mem->is_top()) { 4499 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4500 if (in(i) == base_mem) 4501 set_req(i, empty_memory()); 4502 } 4503 } 4504 } 4505 4506 //---------------------------grow_to_match------------------------------------- 4507 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4508 Node* empty_mem = empty_memory(); 4509 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4510 // look for the finite support of the other memory 4511 for (uint i = other->req(); --i >= req(); ) { 4512 if (other->in(i) != empty_mem) { 4513 uint new_len = i+1; 4514 while (req() < new_len) add_req(empty_mem); 4515 break; 4516 } 4517 } 4518 } 4519 4520 //---------------------------verify_sparse------------------------------------- 4521 #ifndef PRODUCT 4522 bool MergeMemNode::verify_sparse() const { 4523 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4524 Node* base_mem = base_memory(); 4525 // The following can happen in degenerate cases, since empty==top. 4526 if (is_empty_memory(base_mem)) return true; 4527 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4528 assert(in(i) != NULL, "sane slice"); 4529 if (in(i) == base_mem) return false; // should have been the sentinel value! 4530 } 4531 return true; 4532 } 4533 4534 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4535 Node* n; 4536 n = mm->in(idx); 4537 if (mem == n) return true; // might be empty_memory() 4538 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4539 if (mem == n) return true; 4540 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4541 if (mem == n) return true; 4542 if (n == NULL) break; 4543 } 4544 return false; 4545 } 4546 #endif // !PRODUCT