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