memnode.cpp revision 1472:c18cbe5936b8
1/* 2 * Copyright (c) 1997, 2009, 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// Portions of code courtesy of Clifford Click 26 27// Optimization - Graph Style 28 29#include "incls/_precompiled.incl" 30#include "incls/_memnode.cpp.incl" 31 32static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 33 34//============================================================================= 35uint MemNode::size_of() const { return sizeof(*this); } 36 37const TypePtr *MemNode::adr_type() const { 38 Node* adr = in(Address); 39 const TypePtr* cross_check = NULL; 40 DEBUG_ONLY(cross_check = _adr_type); 41 return calculate_adr_type(adr->bottom_type(), cross_check); 42} 43 44#ifndef PRODUCT 45void MemNode::dump_spec(outputStream *st) const { 46 if (in(Address) == NULL) return; // node is dead 47#ifndef ASSERT 48 // fake the missing field 49 const TypePtr* _adr_type = NULL; 50 if (in(Address) != NULL) 51 _adr_type = in(Address)->bottom_type()->isa_ptr(); 52#endif 53 dump_adr_type(this, _adr_type, st); 54 55 Compile* C = Compile::current(); 56 if( C->alias_type(_adr_type)->is_volatile() ) 57 st->print(" Volatile!"); 58} 59 60void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 61 st->print(" @"); 62 if (adr_type == NULL) { 63 st->print("NULL"); 64 } else { 65 adr_type->dump_on(st); 66 Compile* C = Compile::current(); 67 Compile::AliasType* atp = NULL; 68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 69 if (atp == NULL) 70 st->print(", idx=?\?;"); 71 else if (atp->index() == Compile::AliasIdxBot) 72 st->print(", idx=Bot;"); 73 else if (atp->index() == Compile::AliasIdxTop) 74 st->print(", idx=Top;"); 75 else if (atp->index() == Compile::AliasIdxRaw) 76 st->print(", idx=Raw;"); 77 else { 78 ciField* field = atp->field(); 79 if (field) { 80 st->print(", name="); 81 field->print_name_on(st); 82 } 83 st->print(", idx=%d;", atp->index()); 84 } 85 } 86} 87 88extern void print_alias_types(); 89 90#endif 91 92Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 93 const TypeOopPtr *tinst = t_adr->isa_oopptr(); 94 if (tinst == NULL || !tinst->is_known_instance_field()) 95 return mchain; // don't try to optimize non-instance types 96 uint instance_id = tinst->instance_id(); 97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 98 Node *prev = NULL; 99 Node *result = mchain; 100 while (prev != result) { 101 prev = result; 102 if (result == start_mem) 103 break; // hit one of our sentinels 104 // skip over a call which does not affect this memory slice 105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 106 Node *proj_in = result->in(0); 107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 108 break; // hit one of our sentinels 109 } else if (proj_in->is_Call()) { 110 CallNode *call = proj_in->as_Call(); 111 if (!call->may_modify(t_adr, phase)) { 112 result = call->in(TypeFunc::Memory); 113 } 114 } else if (proj_in->is_Initialize()) { 115 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 116 // Stop if this is the initialization for the object instance which 117 // which contains this memory slice, otherwise skip over it. 118 if (alloc != NULL && alloc->_idx != instance_id) { 119 result = proj_in->in(TypeFunc::Memory); 120 } 121 } else if (proj_in->is_MemBar()) { 122 result = proj_in->in(TypeFunc::Memory); 123 } else { 124 assert(false, "unexpected projection"); 125 } 126 } else if (result->is_ClearArray()) { 127 if (!ClearArrayNode::step_through(&result, instance_id, phase)) { 128 // Can not bypass initialization of the instance 129 // we are looking for. 130 break; 131 } 132 // Otherwise skip it (the call updated 'result' value). 133 } else if (result->is_MergeMem()) { 134 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 135 } 136 } 137 return result; 138} 139 140Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 141 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 142 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); 143 PhaseIterGVN *igvn = phase->is_IterGVN(); 144 Node *result = mchain; 145 result = optimize_simple_memory_chain(result, t_adr, phase); 146 if (is_instance && igvn != NULL && result->is_Phi()) { 147 PhiNode *mphi = result->as_Phi(); 148 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 149 const TypePtr *t = mphi->adr_type(); 150 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 151 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 152 t->is_oopptr()->cast_to_exactness(true) 153 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 154 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 155 // clone the Phi with our address type 156 result = mphi->split_out_instance(t_adr, igvn); 157 } else { 158 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 159 } 160 } 161 return result; 162} 163 164static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 165 uint alias_idx = phase->C->get_alias_index(tp); 166 Node *mem = mmem; 167#ifdef ASSERT 168 { 169 // Check that current type is consistent with the alias index used during graph construction 170 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 171 bool consistent = adr_check == NULL || adr_check->empty() || 172 phase->C->must_alias(adr_check, alias_idx ); 173 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 174 if( !consistent && adr_check != NULL && !adr_check->empty() && 175 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 176 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 177 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 178 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 179 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 180 // don't assert if it is dead code. 181 consistent = true; 182 } 183 if( !consistent ) { 184 st->print("alias_idx==%d, adr_check==", alias_idx); 185 if( adr_check == NULL ) { 186 st->print("NULL"); 187 } else { 188 adr_check->dump(); 189 } 190 st->cr(); 191 print_alias_types(); 192 assert(consistent, "adr_check must match alias idx"); 193 } 194 } 195#endif 196 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 197 // means an array I have not precisely typed yet. Do not do any 198 // alias stuff with it any time soon. 199 const TypeOopPtr *tinst = tp->isa_oopptr(); 200 if( tp->base() != Type::AnyPtr && 201 !(tinst && 202 tinst->klass()->is_java_lang_Object() && 203 tinst->offset() == Type::OffsetBot) ) { 204 // compress paths and change unreachable cycles to TOP 205 // If not, we can update the input infinitely along a MergeMem cycle 206 // Equivalent code in PhiNode::Ideal 207 Node* m = phase->transform(mmem); 208 // If transformed to a MergeMem, get the desired slice 209 // Otherwise the returned node represents memory for every slice 210 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 211 // Update input if it is progress over what we have now 212 } 213 return mem; 214} 215 216//--------------------------Ideal_common--------------------------------------- 217// Look for degenerate control and memory inputs. Bypass MergeMem inputs. 218// Unhook non-raw memories from complete (macro-expanded) initializations. 219Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 220 // If our control input is a dead region, kill all below the region 221 Node *ctl = in(MemNode::Control); 222 if (ctl && remove_dead_region(phase, can_reshape)) 223 return this; 224 ctl = in(MemNode::Control); 225 // Don't bother trying to transform a dead node 226 if( ctl && ctl->is_top() ) return NodeSentinel; 227 228 PhaseIterGVN *igvn = phase->is_IterGVN(); 229 // Wait if control on the worklist. 230 if (ctl && can_reshape && igvn != NULL) { 231 Node* bol = NULL; 232 Node* cmp = NULL; 233 if (ctl->in(0)->is_If()) { 234 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 235 bol = ctl->in(0)->in(1); 236 if (bol->is_Bool()) 237 cmp = ctl->in(0)->in(1)->in(1); 238 } 239 if (igvn->_worklist.member(ctl) || 240 (bol != NULL && igvn->_worklist.member(bol)) || 241 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 242 // This control path may be dead. 243 // Delay this memory node transformation until the control is processed. 244 phase->is_IterGVN()->_worklist.push(this); 245 return NodeSentinel; // caller will return NULL 246 } 247 } 248 // Ignore if memory is dead, or self-loop 249 Node *mem = in(MemNode::Memory); 250 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 251 assert( mem != this, "dead loop in MemNode::Ideal" ); 252 253 Node *address = in(MemNode::Address); 254 const Type *t_adr = phase->type( address ); 255 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 256 257 if( can_reshape && igvn != NULL && 258 (igvn->_worklist.member(address) || phase->type(address) != adr_type()) ) { 259 // The address's base and type may change when the address is processed. 260 // Delay this mem node transformation until the address is processed. 261 phase->is_IterGVN()->_worklist.push(this); 262 return NodeSentinel; // caller will return NULL 263 } 264 265 // Do NOT remove or optimize the next lines: ensure a new alias index 266 // is allocated for an oop pointer type before Escape Analysis. 267 // Note: C++ will not remove it since the call has side effect. 268 if ( t_adr->isa_oopptr() ) { 269 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 270 } 271 272#ifdef ASSERT 273 Node* base = NULL; 274 if (address->is_AddP()) 275 base = address->in(AddPNode::Base); 276 assert(base == NULL || t_adr->isa_rawptr() || 277 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); 278#endif 279 280 // Avoid independent memory operations 281 Node* old_mem = mem; 282 283 // The code which unhooks non-raw memories from complete (macro-expanded) 284 // initializations was removed. After macro-expansion all stores catched 285 // by Initialize node became raw stores and there is no information 286 // which memory slices they modify. So it is unsafe to move any memory 287 // operation above these stores. Also in most cases hooked non-raw memories 288 // were already unhooked by using information from detect_ptr_independence() 289 // and find_previous_store(). 290 291 if (mem->is_MergeMem()) { 292 MergeMemNode* mmem = mem->as_MergeMem(); 293 const TypePtr *tp = t_adr->is_ptr(); 294 295 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 296 } 297 298 if (mem != old_mem) { 299 set_req(MemNode::Memory, mem); 300 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 301 return this; 302 } 303 304 // let the subclass continue analyzing... 305 return NULL; 306} 307 308// Helper function for proving some simple control dominations. 309// Attempt to prove that all control inputs of 'dom' dominate 'sub'. 310// Already assumes that 'dom' is available at 'sub', and that 'sub' 311// is not a constant (dominated by the method's StartNode). 312// Used by MemNode::find_previous_store to prove that the 313// control input of a memory operation predates (dominates) 314// an allocation it wants to look past. 315bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 316 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 317 return false; // Conservative answer for dead code 318 319 // Check 'dom'. Skip Proj and CatchProj nodes. 320 dom = dom->find_exact_control(dom); 321 if (dom == NULL || dom->is_top()) 322 return false; // Conservative answer for dead code 323 324 if (dom == sub) { 325 // For the case when, for example, 'sub' is Initialize and the original 326 // 'dom' is Proj node of the 'sub'. 327 return false; 328 } 329 330 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 331 return true; 332 333 // 'dom' dominates 'sub' if its control edge and control edges 334 // of all its inputs dominate or equal to sub's control edge. 335 336 // Currently 'sub' is either Allocate, Initialize or Start nodes. 337 // Or Region for the check in LoadNode::Ideal(); 338 // 'sub' should have sub->in(0) != NULL. 339 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 340 sub->is_Region(), "expecting only these nodes"); 341 342 // Get control edge of 'sub'. 343 Node* orig_sub = sub; 344 sub = sub->find_exact_control(sub->in(0)); 345 if (sub == NULL || sub->is_top()) 346 return false; // Conservative answer for dead code 347 348 assert(sub->is_CFG(), "expecting control"); 349 350 if (sub == dom) 351 return true; 352 353 if (sub->is_Start() || sub->is_Root()) 354 return false; 355 356 { 357 // Check all control edges of 'dom'. 358 359 ResourceMark rm; 360 Arena* arena = Thread::current()->resource_area(); 361 Node_List nlist(arena); 362 Unique_Node_List dom_list(arena); 363 364 dom_list.push(dom); 365 bool only_dominating_controls = false; 366 367 for (uint next = 0; next < dom_list.size(); next++) { 368 Node* n = dom_list.at(next); 369 if (n == orig_sub) 370 return false; // One of dom's inputs dominated by sub. 371 if (!n->is_CFG() && n->pinned()) { 372 // Check only own control edge for pinned non-control nodes. 373 n = n->find_exact_control(n->in(0)); 374 if (n == NULL || n->is_top()) 375 return false; // Conservative answer for dead code 376 assert(n->is_CFG(), "expecting control"); 377 dom_list.push(n); 378 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 379 only_dominating_controls = true; 380 } else if (n->is_CFG()) { 381 if (n->dominates(sub, nlist)) 382 only_dominating_controls = true; 383 else 384 return false; 385 } else { 386 // First, own control edge. 387 Node* m = n->find_exact_control(n->in(0)); 388 if (m != NULL) { 389 if (m->is_top()) 390 return false; // Conservative answer for dead code 391 dom_list.push(m); 392 } 393 // Now, the rest of edges. 394 uint cnt = n->req(); 395 for (uint i = 1; i < cnt; i++) { 396 m = n->find_exact_control(n->in(i)); 397 if (m == NULL || m->is_top()) 398 continue; 399 dom_list.push(m); 400 } 401 } 402 } 403 return only_dominating_controls; 404 } 405} 406 407//---------------------detect_ptr_independence--------------------------------- 408// Used by MemNode::find_previous_store to prove that two base 409// pointers are never equal. 410// The pointers are accompanied by their associated allocations, 411// if any, which have been previously discovered by the caller. 412bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 413 Node* p2, AllocateNode* a2, 414 PhaseTransform* phase) { 415 // Attempt to prove that these two pointers cannot be aliased. 416 // They may both manifestly be allocations, and they should differ. 417 // Or, if they are not both allocations, they can be distinct constants. 418 // Otherwise, one is an allocation and the other a pre-existing value. 419 if (a1 == NULL && a2 == NULL) { // neither an allocation 420 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 421 } else if (a1 != NULL && a2 != NULL) { // both allocations 422 return (a1 != a2); 423 } else if (a1 != NULL) { // one allocation a1 424 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 425 return all_controls_dominate(p2, a1); 426 } else { //(a2 != NULL) // one allocation a2 427 return all_controls_dominate(p1, a2); 428 } 429 return false; 430} 431 432 433// The logic for reordering loads and stores uses four steps: 434// (a) Walk carefully past stores and initializations which we 435// can prove are independent of this load. 436// (b) Observe that the next memory state makes an exact match 437// with self (load or store), and locate the relevant store. 438// (c) Ensure that, if we were to wire self directly to the store, 439// the optimizer would fold it up somehow. 440// (d) Do the rewiring, and return, depending on some other part of 441// the optimizer to fold up the load. 442// This routine handles steps (a) and (b). Steps (c) and (d) are 443// specific to loads and stores, so they are handled by the callers. 444// (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 445// 446Node* MemNode::find_previous_store(PhaseTransform* phase) { 447 Node* ctrl = in(MemNode::Control); 448 Node* adr = in(MemNode::Address); 449 intptr_t offset = 0; 450 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 451 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 452 453 if (offset == Type::OffsetBot) 454 return NULL; // cannot unalias unless there are precise offsets 455 456 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 457 458 intptr_t size_in_bytes = memory_size(); 459 460 Node* mem = in(MemNode::Memory); // start searching here... 461 462 int cnt = 50; // Cycle limiter 463 for (;;) { // While we can dance past unrelated stores... 464 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 465 466 if (mem->is_Store()) { 467 Node* st_adr = mem->in(MemNode::Address); 468 intptr_t st_offset = 0; 469 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 470 if (st_base == NULL) 471 break; // inscrutable pointer 472 if (st_offset != offset && st_offset != Type::OffsetBot) { 473 const int MAX_STORE = BytesPerLong; 474 if (st_offset >= offset + size_in_bytes || 475 st_offset <= offset - MAX_STORE || 476 st_offset <= offset - mem->as_Store()->memory_size()) { 477 // Success: The offsets are provably independent. 478 // (You may ask, why not just test st_offset != offset and be done? 479 // The answer is that stores of different sizes can co-exist 480 // in the same sequence of RawMem effects. We sometimes initialize 481 // a whole 'tile' of array elements with a single jint or jlong.) 482 mem = mem->in(MemNode::Memory); 483 continue; // (a) advance through independent store memory 484 } 485 } 486 if (st_base != base && 487 detect_ptr_independence(base, alloc, 488 st_base, 489 AllocateNode::Ideal_allocation(st_base, phase), 490 phase)) { 491 // Success: The bases are provably independent. 492 mem = mem->in(MemNode::Memory); 493 continue; // (a) advance through independent store memory 494 } 495 496 // (b) At this point, if the bases or offsets do not agree, we lose, 497 // since we have not managed to prove 'this' and 'mem' independent. 498 if (st_base == base && st_offset == offset) { 499 return mem; // let caller handle steps (c), (d) 500 } 501 502 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 503 InitializeNode* st_init = mem->in(0)->as_Initialize(); 504 AllocateNode* st_alloc = st_init->allocation(); 505 if (st_alloc == NULL) 506 break; // something degenerated 507 bool known_identical = false; 508 bool known_independent = false; 509 if (alloc == st_alloc) 510 known_identical = true; 511 else if (alloc != NULL) 512 known_independent = true; 513 else if (all_controls_dominate(this, st_alloc)) 514 known_independent = true; 515 516 if (known_independent) { 517 // The bases are provably independent: Either they are 518 // manifestly distinct allocations, or else the control 519 // of this load dominates the store's allocation. 520 int alias_idx = phase->C->get_alias_index(adr_type()); 521 if (alias_idx == Compile::AliasIdxRaw) { 522 mem = st_alloc->in(TypeFunc::Memory); 523 } else { 524 mem = st_init->memory(alias_idx); 525 } 526 continue; // (a) advance through independent store memory 527 } 528 529 // (b) at this point, if we are not looking at a store initializing 530 // the same allocation we are loading from, we lose. 531 if (known_identical) { 532 // From caller, can_see_stored_value will consult find_captured_store. 533 return mem; // let caller handle steps (c), (d) 534 } 535 536 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 537 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 538 if (mem->is_Proj() && mem->in(0)->is_Call()) { 539 CallNode *call = mem->in(0)->as_Call(); 540 if (!call->may_modify(addr_t, phase)) { 541 mem = call->in(TypeFunc::Memory); 542 continue; // (a) advance through independent call memory 543 } 544 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 545 mem = mem->in(0)->in(TypeFunc::Memory); 546 continue; // (a) advance through independent MemBar memory 547 } else if (mem->is_ClearArray()) { 548 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 549 // (the call updated 'mem' value) 550 continue; // (a) advance through independent allocation memory 551 } else { 552 // Can not bypass initialization of the instance 553 // we are looking for. 554 return mem; 555 } 556 } else if (mem->is_MergeMem()) { 557 int alias_idx = phase->C->get_alias_index(adr_type()); 558 mem = mem->as_MergeMem()->memory_at(alias_idx); 559 continue; // (a) advance through independent MergeMem memory 560 } 561 } 562 563 // Unless there is an explicit 'continue', we must bail out here, 564 // because 'mem' is an inscrutable memory state (e.g., a call). 565 break; 566 } 567 568 return NULL; // bail out 569} 570 571//----------------------calculate_adr_type------------------------------------- 572// Helper function. Notices when the given type of address hits top or bottom. 573// Also, asserts a cross-check of the type against the expected address type. 574const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 575 if (t == Type::TOP) return NULL; // does not touch memory any more? 576 #ifdef PRODUCT 577 cross_check = NULL; 578 #else 579 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 580 #endif 581 const TypePtr* tp = t->isa_ptr(); 582 if (tp == NULL) { 583 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 584 return TypePtr::BOTTOM; // touches lots of memory 585 } else { 586 #ifdef ASSERT 587 // %%%% [phh] We don't check the alias index if cross_check is 588 // TypeRawPtr::BOTTOM. Needs to be investigated. 589 if (cross_check != NULL && 590 cross_check != TypePtr::BOTTOM && 591 cross_check != TypeRawPtr::BOTTOM) { 592 // Recheck the alias index, to see if it has changed (due to a bug). 593 Compile* C = Compile::current(); 594 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 595 "must stay in the original alias category"); 596 // The type of the address must be contained in the adr_type, 597 // disregarding "null"-ness. 598 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 599 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 600 assert(cross_check->meet(tp_notnull) == cross_check, 601 "real address must not escape from expected memory type"); 602 } 603 #endif 604 return tp; 605 } 606} 607 608//------------------------adr_phi_is_loop_invariant---------------------------- 609// A helper function for Ideal_DU_postCCP to check if a Phi in a counted 610// loop is loop invariant. Make a quick traversal of Phi and associated 611// CastPP nodes, looking to see if they are a closed group within the loop. 612bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 613 // The idea is that the phi-nest must boil down to only CastPP nodes 614 // with the same data. This implies that any path into the loop already 615 // includes such a CastPP, and so the original cast, whatever its input, 616 // must be covered by an equivalent cast, with an earlier control input. 617 ResourceMark rm; 618 619 // The loop entry input of the phi should be the unique dominating 620 // node for every Phi/CastPP in the loop. 621 Unique_Node_List closure; 622 closure.push(adr_phi->in(LoopNode::EntryControl)); 623 624 // Add the phi node and the cast to the worklist. 625 Unique_Node_List worklist; 626 worklist.push(adr_phi); 627 if( cast != NULL ){ 628 if( !cast->is_ConstraintCast() ) return false; 629 worklist.push(cast); 630 } 631 632 // Begin recursive walk of phi nodes. 633 while( worklist.size() ){ 634 // Take a node off the worklist 635 Node *n = worklist.pop(); 636 if( !closure.member(n) ){ 637 // Add it to the closure. 638 closure.push(n); 639 // Make a sanity check to ensure we don't waste too much time here. 640 if( closure.size() > 20) return false; 641 // This node is OK if: 642 // - it is a cast of an identical value 643 // - or it is a phi node (then we add its inputs to the worklist) 644 // Otherwise, the node is not OK, and we presume the cast is not invariant 645 if( n->is_ConstraintCast() ){ 646 worklist.push(n->in(1)); 647 } else if( n->is_Phi() ) { 648 for( uint i = 1; i < n->req(); i++ ) { 649 worklist.push(n->in(i)); 650 } 651 } else { 652 return false; 653 } 654 } 655 } 656 657 // Quit when the worklist is empty, and we've found no offending nodes. 658 return true; 659} 660 661//------------------------------Ideal_DU_postCCP------------------------------- 662// Find any cast-away of null-ness and keep its control. Null cast-aways are 663// going away in this pass and we need to make this memory op depend on the 664// gating null check. 665Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 666 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 667} 668 669// I tried to leave the CastPP's in. This makes the graph more accurate in 670// some sense; we get to keep around the knowledge that an oop is not-null 671// after some test. Alas, the CastPP's interfere with GVN (some values are 672// the regular oop, some are the CastPP of the oop, all merge at Phi's which 673// cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 674// some of the more trivial cases in the optimizer. Removing more useless 675// Phi's started allowing Loads to illegally float above null checks. I gave 676// up on this approach. CNC 10/20/2000 677// This static method may be called not from MemNode (EncodePNode calls it). 678// Only the control edge of the node 'n' might be updated. 679Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 680 Node *skipped_cast = NULL; 681 // Need a null check? Regular static accesses do not because they are 682 // from constant addresses. Array ops are gated by the range check (which 683 // always includes a NULL check). Just check field ops. 684 if( n->in(MemNode::Control) == NULL ) { 685 // Scan upwards for the highest location we can place this memory op. 686 while( true ) { 687 switch( adr->Opcode() ) { 688 689 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 690 adr = adr->in(AddPNode::Base); 691 continue; 692 693 case Op_DecodeN: // No change to NULL-ness, so peek thru 694 adr = adr->in(1); 695 continue; 696 697 case Op_CastPP: 698 // If the CastPP is useless, just peek on through it. 699 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 700 // Remember the cast that we've peeked though. If we peek 701 // through more than one, then we end up remembering the highest 702 // one, that is, if in a loop, the one closest to the top. 703 skipped_cast = adr; 704 adr = adr->in(1); 705 continue; 706 } 707 // CastPP is going away in this pass! We need this memory op to be 708 // control-dependent on the test that is guarding the CastPP. 709 ccp->hash_delete(n); 710 n->set_req(MemNode::Control, adr->in(0)); 711 ccp->hash_insert(n); 712 return n; 713 714 case Op_Phi: 715 // Attempt to float above a Phi to some dominating point. 716 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 717 // If we've already peeked through a Cast (which could have set the 718 // control), we can't float above a Phi, because the skipped Cast 719 // may not be loop invariant. 720 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 721 adr = adr->in(1); 722 continue; 723 } 724 } 725 726 // Intentional fallthrough! 727 728 // No obvious dominating point. The mem op is pinned below the Phi 729 // by the Phi itself. If the Phi goes away (no true value is merged) 730 // then the mem op can float, but not indefinitely. It must be pinned 731 // behind the controls leading to the Phi. 732 case Op_CheckCastPP: 733 // These usually stick around to change address type, however a 734 // useless one can be elided and we still need to pick up a control edge 735 if (adr->in(0) == NULL) { 736 // This CheckCastPP node has NO control and is likely useless. But we 737 // need check further up the ancestor chain for a control input to keep 738 // the node in place. 4959717. 739 skipped_cast = adr; 740 adr = adr->in(1); 741 continue; 742 } 743 ccp->hash_delete(n); 744 n->set_req(MemNode::Control, adr->in(0)); 745 ccp->hash_insert(n); 746 return n; 747 748 // List of "safe" opcodes; those that implicitly block the memory 749 // op below any null check. 750 case Op_CastX2P: // no null checks on native pointers 751 case Op_Parm: // 'this' pointer is not null 752 case Op_LoadP: // Loading from within a klass 753 case Op_LoadN: // Loading from within a klass 754 case Op_LoadKlass: // Loading from within a klass 755 case Op_LoadNKlass: // Loading from within a klass 756 case Op_ConP: // Loading from a klass 757 case Op_ConN: // Loading from a klass 758 case Op_CreateEx: // Sucking up the guts of an exception oop 759 case Op_Con: // Reading from TLS 760 case Op_CMoveP: // CMoveP is pinned 761 case Op_CMoveN: // CMoveN is pinned 762 break; // No progress 763 764 case Op_Proj: // Direct call to an allocation routine 765 case Op_SCMemProj: // Memory state from store conditional ops 766#ifdef ASSERT 767 { 768 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 769 const Node* call = adr->in(0); 770 if (call->is_CallJava()) { 771 const CallJavaNode* call_java = call->as_CallJava(); 772 const TypeTuple *r = call_java->tf()->range(); 773 assert(r->cnt() > TypeFunc::Parms, "must return value"); 774 const Type* ret_type = r->field_at(TypeFunc::Parms); 775 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 776 // We further presume that this is one of 777 // new_instance_Java, new_array_Java, or 778 // the like, but do not assert for this. 779 } else if (call->is_Allocate()) { 780 // similar case to new_instance_Java, etc. 781 } else if (!call->is_CallLeaf()) { 782 // Projections from fetch_oop (OSR) are allowed as well. 783 ShouldNotReachHere(); 784 } 785 } 786#endif 787 break; 788 default: 789 ShouldNotReachHere(); 790 } 791 break; 792 } 793 } 794 795 return NULL; // No progress 796} 797 798 799//============================================================================= 800uint LoadNode::size_of() const { return sizeof(*this); } 801uint LoadNode::cmp( const Node &n ) const 802{ return !Type::cmp( _type, ((LoadNode&)n)._type ); } 803const Type *LoadNode::bottom_type() const { return _type; } 804uint LoadNode::ideal_reg() const { 805 return Matcher::base2reg[_type->base()]; 806} 807 808#ifndef PRODUCT 809void LoadNode::dump_spec(outputStream *st) const { 810 MemNode::dump_spec(st); 811 if( !Verbose && !WizardMode ) { 812 // standard dump does this in Verbose and WizardMode 813 st->print(" #"); _type->dump_on(st); 814 } 815} 816#endif 817 818 819//----------------------------LoadNode::make----------------------------------- 820// Polymorphic factory method: 821Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 822 Compile* C = gvn.C; 823 824 // sanity check the alias category against the created node type 825 assert(!(adr_type->isa_oopptr() && 826 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 827 "use LoadKlassNode instead"); 828 assert(!(adr_type->isa_aryptr() && 829 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 830 "use LoadRangeNode instead"); 831 switch (bt) { 832 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 833 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 834 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 835 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 836 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 837 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 838 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 839 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 840 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 841 case T_OBJECT: 842#ifdef _LP64 843 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 844 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 845 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 846 } else 847#endif 848 { 849 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 850 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 851 } 852 } 853 ShouldNotReachHere(); 854 return (LoadNode*)NULL; 855} 856 857LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 858 bool require_atomic = true; 859 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 860} 861 862 863 864 865//------------------------------hash------------------------------------------- 866uint LoadNode::hash() const { 867 // unroll addition of interesting fields 868 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 869} 870 871//---------------------------can_see_stored_value------------------------------ 872// This routine exists to make sure this set of tests is done the same 873// everywhere. We need to make a coordinated change: first LoadNode::Ideal 874// will change the graph shape in a way which makes memory alive twice at the 875// same time (uses the Oracle model of aliasing), then some 876// LoadXNode::Identity will fold things back to the equivalence-class model 877// of aliasing. 878Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 879 Node* ld_adr = in(MemNode::Address); 880 881 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 882 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 883 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 884 atp->field() != NULL && !atp->field()->is_volatile()) { 885 uint alias_idx = atp->index(); 886 bool final = atp->field()->is_final(); 887 Node* result = NULL; 888 Node* current = st; 889 // Skip through chains of MemBarNodes checking the MergeMems for 890 // new states for the slice of this load. Stop once any other 891 // kind of node is encountered. Loads from final memory can skip 892 // through any kind of MemBar but normal loads shouldn't skip 893 // through MemBarAcquire since the could allow them to move out of 894 // a synchronized region. 895 while (current->is_Proj()) { 896 int opc = current->in(0)->Opcode(); 897 if ((final && opc == Op_MemBarAcquire) || 898 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 899 Node* mem = current->in(0)->in(TypeFunc::Memory); 900 if (mem->is_MergeMem()) { 901 MergeMemNode* merge = mem->as_MergeMem(); 902 Node* new_st = merge->memory_at(alias_idx); 903 if (new_st == merge->base_memory()) { 904 // Keep searching 905 current = merge->base_memory(); 906 continue; 907 } 908 // Save the new memory state for the slice and fall through 909 // to exit. 910 result = new_st; 911 } 912 } 913 break; 914 } 915 if (result != NULL) { 916 st = result; 917 } 918 } 919 920 921 // Loop around twice in the case Load -> Initialize -> Store. 922 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 923 for (int trip = 0; trip <= 1; trip++) { 924 925 if (st->is_Store()) { 926 Node* st_adr = st->in(MemNode::Address); 927 if (!phase->eqv(st_adr, ld_adr)) { 928 // Try harder before giving up... Match raw and non-raw pointers. 929 intptr_t st_off = 0; 930 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 931 if (alloc == NULL) return NULL; 932 intptr_t ld_off = 0; 933 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 934 if (alloc != allo2) return NULL; 935 if (ld_off != st_off) return NULL; 936 // At this point we have proven something like this setup: 937 // A = Allocate(...) 938 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 939 // S = StoreQ(, AddP(, A.Parm , #Off), V) 940 // (Actually, we haven't yet proven the Q's are the same.) 941 // In other words, we are loading from a casted version of 942 // the same pointer-and-offset that we stored to. 943 // Thus, we are able to replace L by V. 944 } 945 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 946 if (store_Opcode() != st->Opcode()) 947 return NULL; 948 return st->in(MemNode::ValueIn); 949 } 950 951 intptr_t offset = 0; // scratch 952 953 // A load from a freshly-created object always returns zero. 954 // (This can happen after LoadNode::Ideal resets the load's memory input 955 // to find_captured_store, which returned InitializeNode::zero_memory.) 956 if (st->is_Proj() && st->in(0)->is_Allocate() && 957 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 958 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 959 // return a zero value for the load's basic type 960 // (This is one of the few places where a generic PhaseTransform 961 // can create new nodes. Think of it as lazily manifesting 962 // virtually pre-existing constants.) 963 return phase->zerocon(memory_type()); 964 } 965 966 // A load from an initialization barrier can match a captured store. 967 if (st->is_Proj() && st->in(0)->is_Initialize()) { 968 InitializeNode* init = st->in(0)->as_Initialize(); 969 AllocateNode* alloc = init->allocation(); 970 if (alloc != NULL && 971 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 972 // examine a captured store value 973 st = init->find_captured_store(offset, memory_size(), phase); 974 if (st != NULL) 975 continue; // take one more trip around 976 } 977 } 978 979 break; 980 } 981 982 return NULL; 983} 984 985//----------------------is_instance_field_load_with_local_phi------------------ 986bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 987 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 988 in(MemNode::Address)->is_AddP() ) { 989 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 990 // Only instances. 991 if( t_oop != NULL && t_oop->is_known_instance_field() && 992 t_oop->offset() != Type::OffsetBot && 993 t_oop->offset() != Type::OffsetTop) { 994 return true; 995 } 996 } 997 return false; 998} 999 1000//------------------------------Identity--------------------------------------- 1001// Loads are identity if previous store is to same address 1002Node *LoadNode::Identity( PhaseTransform *phase ) { 1003 // If the previous store-maker is the right kind of Store, and the store is 1004 // to the same address, then we are equal to the value stored. 1005 Node* mem = in(MemNode::Memory); 1006 Node* value = can_see_stored_value(mem, phase); 1007 if( value ) { 1008 // byte, short & char stores truncate naturally. 1009 // A load has to load the truncated value which requires 1010 // some sort of masking operation and that requires an 1011 // Ideal call instead of an Identity call. 1012 if (memory_size() < BytesPerInt) { 1013 // If the input to the store does not fit with the load's result type, 1014 // it must be truncated via an Ideal call. 1015 if (!phase->type(value)->higher_equal(phase->type(this))) 1016 return this; 1017 } 1018 // (This works even when value is a Con, but LoadNode::Value 1019 // usually runs first, producing the singleton type of the Con.) 1020 return value; 1021 } 1022 1023 // Search for an existing data phi which was generated before for the same 1024 // instance's field to avoid infinite generation of phis in a loop. 1025 Node *region = mem->in(0); 1026 if (is_instance_field_load_with_local_phi(region)) { 1027 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 1028 int this_index = phase->C->get_alias_index(addr_t); 1029 int this_offset = addr_t->offset(); 1030 int this_id = addr_t->is_oopptr()->instance_id(); 1031 const Type* this_type = bottom_type(); 1032 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1033 Node* phi = region->fast_out(i); 1034 if (phi->is_Phi() && phi != mem && 1035 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 1036 return phi; 1037 } 1038 } 1039 } 1040 1041 return this; 1042} 1043 1044 1045// Returns true if the AliasType refers to the field that holds the 1046// cached box array. Currently only handles the IntegerCache case. 1047static bool is_autobox_cache(Compile::AliasType* atp) { 1048 if (atp != NULL && atp->field() != NULL) { 1049 ciField* field = atp->field(); 1050 ciSymbol* klass = field->holder()->name(); 1051 if (field->name() == ciSymbol::cache_field_name() && 1052 field->holder()->uses_default_loader() && 1053 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1054 return true; 1055 } 1056 } 1057 return false; 1058} 1059 1060// Fetch the base value in the autobox array 1061static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1062 if (atp != NULL && atp->field() != NULL) { 1063 ciField* field = atp->field(); 1064 ciSymbol* klass = field->holder()->name(); 1065 if (field->name() == ciSymbol::cache_field_name() && 1066 field->holder()->uses_default_loader() && 1067 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1068 assert(field->is_constant(), "what?"); 1069 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1070 // Fetch the box object at the base of the array and get its value 1071 ciInstance* box = array->obj_at(0)->as_instance(); 1072 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1073 if (ik->nof_nonstatic_fields() == 1) { 1074 // This should be true nonstatic_field_at requires calling 1075 // nof_nonstatic_fields so check it anyway 1076 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1077 cache_offset = c.as_int(); 1078 } 1079 return true; 1080 } 1081 } 1082 return false; 1083} 1084 1085// Returns true if the AliasType refers to the value field of an 1086// autobox object. Currently only handles Integer. 1087static bool is_autobox_object(Compile::AliasType* atp) { 1088 if (atp != NULL && atp->field() != NULL) { 1089 ciField* field = atp->field(); 1090 ciSymbol* klass = field->holder()->name(); 1091 if (field->name() == ciSymbol::value_name() && 1092 field->holder()->uses_default_loader() && 1093 klass == ciSymbol::java_lang_Integer()) { 1094 return true; 1095 } 1096 } 1097 return false; 1098} 1099 1100 1101// We're loading from an object which has autobox behaviour. 1102// If this object is result of a valueOf call we'll have a phi 1103// merging a newly allocated object and a load from the cache. 1104// We want to replace this load with the original incoming 1105// argument to the valueOf call. 1106Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1107 Node* base = in(Address)->in(AddPNode::Base); 1108 if (base->is_Phi() && base->req() == 3) { 1109 AllocateNode* allocation = NULL; 1110 int allocation_index = -1; 1111 int load_index = -1; 1112 for (uint i = 1; i < base->req(); i++) { 1113 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1114 if (allocation != NULL) { 1115 allocation_index = i; 1116 load_index = 3 - allocation_index; 1117 break; 1118 } 1119 } 1120 bool has_load = ( allocation != NULL && 1121 (base->in(load_index)->is_Load() || 1122 base->in(load_index)->is_DecodeN() && 1123 base->in(load_index)->in(1)->is_Load()) ); 1124 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1125 // Push the loads from the phi that comes from valueOf up 1126 // through it to allow elimination of the loads and the recovery 1127 // of the original value. 1128 Node* mem_phi = in(Memory); 1129 Node* offset = in(Address)->in(AddPNode::Offset); 1130 Node* region = base->in(0); 1131 1132 Node* in1 = clone(); 1133 Node* in1_addr = in1->in(Address)->clone(); 1134 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1135 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1136 in1_addr->set_req(AddPNode::Offset, offset); 1137 in1->set_req(0, region->in(allocation_index)); 1138 in1->set_req(Address, in1_addr); 1139 in1->set_req(Memory, mem_phi->in(allocation_index)); 1140 1141 Node* in2 = clone(); 1142 Node* in2_addr = in2->in(Address)->clone(); 1143 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1144 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1145 in2_addr->set_req(AddPNode::Offset, offset); 1146 in2->set_req(0, region->in(load_index)); 1147 in2->set_req(Address, in2_addr); 1148 in2->set_req(Memory, mem_phi->in(load_index)); 1149 1150 in1_addr = phase->transform(in1_addr); 1151 in1 = phase->transform(in1); 1152 in2_addr = phase->transform(in2_addr); 1153 in2 = phase->transform(in2); 1154 1155 PhiNode* result = PhiNode::make_blank(region, this); 1156 result->set_req(allocation_index, in1); 1157 result->set_req(load_index, in2); 1158 return result; 1159 } 1160 } else if (base->is_Load() || 1161 base->is_DecodeN() && base->in(1)->is_Load()) { 1162 if (base->is_DecodeN()) { 1163 // Get LoadN node which loads cached Integer object 1164 base = base->in(1); 1165 } 1166 // Eliminate the load of Integer.value for integers from the cache 1167 // array by deriving the value from the index into the array. 1168 // Capture the offset of the load and then reverse the computation. 1169 Node* load_base = base->in(Address)->in(AddPNode::Base); 1170 if (load_base->is_DecodeN()) { 1171 // Get LoadN node which loads IntegerCache.cache field 1172 load_base = load_base->in(1); 1173 } 1174 if (load_base != NULL) { 1175 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1176 intptr_t cache_offset; 1177 int shift = -1; 1178 Node* cache = NULL; 1179 if (is_autobox_cache(atp)) { 1180 shift = exact_log2(type2aelembytes(T_OBJECT)); 1181 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1182 } 1183 if (cache != NULL && base->in(Address)->is_AddP()) { 1184 Node* elements[4]; 1185 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1186 int cache_low; 1187 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1188 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1189 // Add up all the offsets making of the address of the load 1190 Node* result = elements[0]; 1191 for (int i = 1; i < count; i++) { 1192 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1193 } 1194 // Remove the constant offset from the address and then 1195 // remove the scaling of the offset to recover the original index. 1196 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1197 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1198 // Peel the shift off directly but wrap it in a dummy node 1199 // since Ideal can't return existing nodes 1200 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1201 } else { 1202 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1203 } 1204#ifdef _LP64 1205 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1206#endif 1207 return result; 1208 } 1209 } 1210 } 1211 } 1212 return NULL; 1213} 1214 1215//------------------------------split_through_phi------------------------------ 1216// Split instance field load through Phi. 1217Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1218 Node* mem = in(MemNode::Memory); 1219 Node* address = in(MemNode::Address); 1220 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1221 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1222 1223 assert(mem->is_Phi() && (t_oop != NULL) && 1224 t_oop->is_known_instance_field(), "invalide conditions"); 1225 1226 Node *region = mem->in(0); 1227 if (region == NULL) { 1228 return NULL; // Wait stable graph 1229 } 1230 uint cnt = mem->req(); 1231 for( uint i = 1; i < cnt; i++ ) { 1232 Node *in = mem->in(i); 1233 if( in == NULL ) { 1234 return NULL; // Wait stable graph 1235 } 1236 } 1237 // Check for loop invariant. 1238 if (cnt == 3) { 1239 for( uint i = 1; i < cnt; i++ ) { 1240 Node *in = mem->in(i); 1241 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1242 if (m == mem) { 1243 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1244 return this; 1245 } 1246 } 1247 } 1248 // Split through Phi (see original code in loopopts.cpp). 1249 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1250 1251 // Do nothing here if Identity will find a value 1252 // (to avoid infinite chain of value phis generation). 1253 if ( !phase->eqv(this, this->Identity(phase)) ) 1254 return NULL; 1255 1256 // Skip the split if the region dominates some control edge of the address. 1257 if (cnt == 3 && !MemNode::all_controls_dominate(address, region)) 1258 return NULL; 1259 1260 const Type* this_type = this->bottom_type(); 1261 int this_index = phase->C->get_alias_index(addr_t); 1262 int this_offset = addr_t->offset(); 1263 int this_iid = addr_t->is_oopptr()->instance_id(); 1264 int wins = 0; 1265 PhaseIterGVN *igvn = phase->is_IterGVN(); 1266 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1267 for( uint i = 1; i < region->req(); i++ ) { 1268 Node *x; 1269 Node* the_clone = NULL; 1270 if( region->in(i) == phase->C->top() ) { 1271 x = phase->C->top(); // Dead path? Use a dead data op 1272 } else { 1273 x = this->clone(); // Else clone up the data op 1274 the_clone = x; // Remember for possible deletion. 1275 // Alter data node to use pre-phi inputs 1276 if( this->in(0) == region ) { 1277 x->set_req( 0, region->in(i) ); 1278 } else { 1279 x->set_req( 0, NULL ); 1280 } 1281 for( uint j = 1; j < this->req(); j++ ) { 1282 Node *in = this->in(j); 1283 if( in->is_Phi() && in->in(0) == region ) 1284 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 1285 } 1286 } 1287 // Check for a 'win' on some paths 1288 const Type *t = x->Value(igvn); 1289 1290 bool singleton = t->singleton(); 1291 1292 // See comments in PhaseIdealLoop::split_thru_phi(). 1293 if( singleton && t == Type::TOP ) { 1294 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1295 } 1296 1297 if( singleton ) { 1298 wins++; 1299 x = igvn->makecon(t); 1300 } else { 1301 // We now call Identity to try to simplify the cloned node. 1302 // Note that some Identity methods call phase->type(this). 1303 // Make sure that the type array is big enough for 1304 // our new node, even though we may throw the node away. 1305 // (This tweaking with igvn only works because x is a new node.) 1306 igvn->set_type(x, t); 1307 // If x is a TypeNode, capture any more-precise type permanently into Node 1308 // otherwise it will be not updated during igvn->transform since 1309 // igvn->type(x) is set to x->Value() already. 1310 x->raise_bottom_type(t); 1311 Node *y = x->Identity(igvn); 1312 if( y != x ) { 1313 wins++; 1314 x = y; 1315 } else { 1316 y = igvn->hash_find(x); 1317 if( y ) { 1318 wins++; 1319 x = y; 1320 } else { 1321 // Else x is a new node we are keeping 1322 // We do not need register_new_node_with_optimizer 1323 // because set_type has already been called. 1324 igvn->_worklist.push(x); 1325 } 1326 } 1327 } 1328 if (x != the_clone && the_clone != NULL) 1329 igvn->remove_dead_node(the_clone); 1330 phi->set_req(i, x); 1331 } 1332 if( wins > 0 ) { 1333 // Record Phi 1334 igvn->register_new_node_with_optimizer(phi); 1335 return phi; 1336 } 1337 igvn->remove_dead_node(phi); 1338 return NULL; 1339} 1340 1341//------------------------------Ideal------------------------------------------ 1342// If the load is from Field memory and the pointer is non-null, we can 1343// zero out the control input. 1344// If the offset is constant and the base is an object allocation, 1345// try to hook me up to the exact initializing store. 1346Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1347 Node* p = MemNode::Ideal_common(phase, can_reshape); 1348 if (p) return (p == NodeSentinel) ? NULL : p; 1349 1350 Node* ctrl = in(MemNode::Control); 1351 Node* address = in(MemNode::Address); 1352 1353 // Skip up past a SafePoint control. Cannot do this for Stores because 1354 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1355 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1356 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1357 ctrl = ctrl->in(0); 1358 set_req(MemNode::Control,ctrl); 1359 } 1360 1361 intptr_t ignore = 0; 1362 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1363 if (base != NULL 1364 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1365 // Check for useless control edge in some common special cases 1366 if (in(MemNode::Control) != NULL 1367 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1368 && all_controls_dominate(base, phase->C->start())) { 1369 // A method-invariant, non-null address (constant or 'this' argument). 1370 set_req(MemNode::Control, NULL); 1371 } 1372 1373 if (EliminateAutoBox && can_reshape) { 1374 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); 1375 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1376 if (is_autobox_object(atp)) { 1377 Node* result = eliminate_autobox(phase); 1378 if (result != NULL) return result; 1379 } 1380 } 1381 } 1382 1383 Node* mem = in(MemNode::Memory); 1384 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1385 1386 if (addr_t != NULL) { 1387 // try to optimize our memory input 1388 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1389 if (opt_mem != mem) { 1390 set_req(MemNode::Memory, opt_mem); 1391 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1392 return this; 1393 } 1394 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1395 if (can_reshape && opt_mem->is_Phi() && 1396 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1397 // Split instance field load through Phi. 1398 Node* result = split_through_phi(phase); 1399 if (result != NULL) return result; 1400 } 1401 } 1402 1403 // Check for prior store with a different base or offset; make Load 1404 // independent. Skip through any number of them. Bail out if the stores 1405 // are in an endless dead cycle and report no progress. This is a key 1406 // transform for Reflection. However, if after skipping through the Stores 1407 // we can't then fold up against a prior store do NOT do the transform as 1408 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1409 // array memory alive twice: once for the hoisted Load and again after the 1410 // bypassed Store. This situation only works if EVERYBODY who does 1411 // anti-dependence work knows how to bypass. I.e. we need all 1412 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1413 // the alias index stuff. So instead, peek through Stores and IFF we can 1414 // fold up, do so. 1415 Node* prev_mem = find_previous_store(phase); 1416 // Steps (a), (b): Walk past independent stores to find an exact match. 1417 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1418 // (c) See if we can fold up on the spot, but don't fold up here. 1419 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1420 // just return a prior value, which is done by Identity calls. 1421 if (can_see_stored_value(prev_mem, phase)) { 1422 // Make ready for step (d): 1423 set_req(MemNode::Memory, prev_mem); 1424 return this; 1425 } 1426 } 1427 1428 return NULL; // No further progress 1429} 1430 1431// Helper to recognize certain Klass fields which are invariant across 1432// some group of array types (e.g., int[] or all T[] where T < Object). 1433const Type* 1434LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1435 ciKlass* klass) const { 1436 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1437 // The field is Klass::_modifier_flags. Return its (constant) value. 1438 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1439 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1440 return TypeInt::make(klass->modifier_flags()); 1441 } 1442 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1443 // The field is Klass::_access_flags. Return its (constant) value. 1444 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1445 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1446 return TypeInt::make(klass->access_flags()); 1447 } 1448 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1449 // The field is Klass::_layout_helper. Return its constant value if known. 1450 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1451 return TypeInt::make(klass->layout_helper()); 1452 } 1453 1454 // No match. 1455 return NULL; 1456} 1457 1458//------------------------------Value----------------------------------------- 1459const Type *LoadNode::Value( PhaseTransform *phase ) const { 1460 // Either input is TOP ==> the result is TOP 1461 Node* mem = in(MemNode::Memory); 1462 const Type *t1 = phase->type(mem); 1463 if (t1 == Type::TOP) return Type::TOP; 1464 Node* adr = in(MemNode::Address); 1465 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1466 if (tp == NULL || tp->empty()) return Type::TOP; 1467 int off = tp->offset(); 1468 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1469 1470 // Try to guess loaded type from pointer type 1471 if (tp->base() == Type::AryPtr) { 1472 const Type *t = tp->is_aryptr()->elem(); 1473 // Don't do this for integer types. There is only potential profit if 1474 // the element type t is lower than _type; that is, for int types, if _type is 1475 // more restrictive than t. This only happens here if one is short and the other 1476 // char (both 16 bits), and in those cases we've made an intentional decision 1477 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1478 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1479 // 1480 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1481 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1482 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1483 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1484 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1485 // In fact, that could have been the original type of p1, and p1 could have 1486 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1487 // expression (LShiftL quux 3) independently optimized to the constant 8. 1488 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1489 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1490 // t might actually be lower than _type, if _type is a unique 1491 // concrete subclass of abstract class t. 1492 // Make sure the reference is not into the header, by comparing 1493 // the offset against the offset of the start of the array's data. 1494 // Different array types begin at slightly different offsets (12 vs. 16). 1495 // We choose T_BYTE as an example base type that is least restrictive 1496 // as to alignment, which will therefore produce the smallest 1497 // possible base offset. 1498 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1499 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1500 const Type* jt = t->join(_type); 1501 // In any case, do not allow the join, per se, to empty out the type. 1502 if (jt->empty() && !t->empty()) { 1503 // This can happen if a interface-typed array narrows to a class type. 1504 jt = _type; 1505 } 1506 1507 if (EliminateAutoBox && adr->is_AddP()) { 1508 // The pointers in the autobox arrays are always non-null 1509 Node* base = adr->in(AddPNode::Base); 1510 if (base != NULL && 1511 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { 1512 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1513 if (is_autobox_cache(atp)) { 1514 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1515 } 1516 } 1517 } 1518 return jt; 1519 } 1520 } 1521 } else if (tp->base() == Type::InstPtr) { 1522 const TypeInstPtr* tinst = tp->is_instptr(); 1523 ciKlass* klass = tinst->klass(); 1524 assert( off != Type::OffsetBot || 1525 // arrays can be cast to Objects 1526 tp->is_oopptr()->klass()->is_java_lang_Object() || 1527 // unsafe field access may not have a constant offset 1528 phase->C->has_unsafe_access(), 1529 "Field accesses must be precise" ); 1530 // For oop loads, we expect the _type to be precise 1531 if (OptimizeStringConcat && klass == phase->C->env()->String_klass() && 1532 adr->is_AddP() && off != Type::OffsetBot) { 1533 // For constant Strings treat the fields as compile time constants. 1534 Node* base = adr->in(AddPNode::Base); 1535 if (base->Opcode() == Op_ConP) { 1536 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1537 ciObject* string = t->const_oop(); 1538 ciConstant constant = string->as_instance()->field_value_by_offset(off); 1539 if (constant.basic_type() == T_INT) { 1540 return TypeInt::make(constant.as_int()); 1541 } else if (constant.basic_type() == T_ARRAY) { 1542 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1543 return TypeNarrowOop::make_from_constant(constant.as_object()); 1544 } else { 1545 return TypeOopPtr::make_from_constant(constant.as_object()); 1546 } 1547 } 1548 } 1549 } 1550 } else if (tp->base() == Type::KlassPtr) { 1551 assert( off != Type::OffsetBot || 1552 // arrays can be cast to Objects 1553 tp->is_klassptr()->klass()->is_java_lang_Object() || 1554 // also allow array-loading from the primary supertype 1555 // array during subtype checks 1556 Opcode() == Op_LoadKlass, 1557 "Field accesses must be precise" ); 1558 // For klass/static loads, we expect the _type to be precise 1559 } 1560 1561 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1562 if (tkls != NULL && !StressReflectiveCode) { 1563 ciKlass* klass = tkls->klass(); 1564 if (klass->is_loaded() && tkls->klass_is_exact()) { 1565 // We are loading a field from a Klass metaobject whose identity 1566 // is known at compile time (the type is "exact" or "precise"). 1567 // Check for fields we know are maintained as constants by the VM. 1568 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1569 // The field is Klass::_super_check_offset. Return its (constant) value. 1570 // (Folds up type checking code.) 1571 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1572 return TypeInt::make(klass->super_check_offset()); 1573 } 1574 // Compute index into primary_supers array 1575 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1576 // Check for overflowing; use unsigned compare to handle the negative case. 1577 if( depth < ciKlass::primary_super_limit() ) { 1578 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1579 // (Folds up type checking code.) 1580 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1581 ciKlass *ss = klass->super_of_depth(depth); 1582 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1583 } 1584 const Type* aift = load_array_final_field(tkls, klass); 1585 if (aift != NULL) return aift; 1586 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1587 && klass->is_array_klass()) { 1588 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1589 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1590 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1591 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1592 } 1593 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1594 // The field is Klass::_java_mirror. Return its (constant) value. 1595 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1596 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1597 return TypeInstPtr::make(klass->java_mirror()); 1598 } 1599 } 1600 1601 // We can still check if we are loading from the primary_supers array at a 1602 // shallow enough depth. Even though the klass is not exact, entries less 1603 // than or equal to its super depth are correct. 1604 if (klass->is_loaded() ) { 1605 ciType *inner = klass->klass(); 1606 while( inner->is_obj_array_klass() ) 1607 inner = inner->as_obj_array_klass()->base_element_type(); 1608 if( inner->is_instance_klass() && 1609 !inner->as_instance_klass()->flags().is_interface() ) { 1610 // Compute index into primary_supers array 1611 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1612 // Check for overflowing; use unsigned compare to handle the negative case. 1613 if( depth < ciKlass::primary_super_limit() && 1614 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1615 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1616 // (Folds up type checking code.) 1617 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1618 ciKlass *ss = klass->super_of_depth(depth); 1619 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1620 } 1621 } 1622 } 1623 1624 // If the type is enough to determine that the thing is not an array, 1625 // we can give the layout_helper a positive interval type. 1626 // This will help short-circuit some reflective code. 1627 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1628 && !klass->is_array_klass() // not directly typed as an array 1629 && !klass->is_interface() // specifically not Serializable & Cloneable 1630 && !klass->is_java_lang_Object() // not the supertype of all T[] 1631 ) { 1632 // Note: When interfaces are reliable, we can narrow the interface 1633 // test to (klass != Serializable && klass != Cloneable). 1634 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1635 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1636 // The key property of this type is that it folds up tests 1637 // for array-ness, since it proves that the layout_helper is positive. 1638 // Thus, a generic value like the basic object layout helper works fine. 1639 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1640 } 1641 } 1642 1643 // If we are loading from a freshly-allocated object, produce a zero, 1644 // if the load is provably beyond the header of the object. 1645 // (Also allow a variable load from a fresh array to produce zero.) 1646 if (ReduceFieldZeroing) { 1647 Node* value = can_see_stored_value(mem,phase); 1648 if (value != NULL && value->is_Con()) 1649 return value->bottom_type(); 1650 } 1651 1652 const TypeOopPtr *tinst = tp->isa_oopptr(); 1653 if (tinst != NULL && tinst->is_known_instance_field()) { 1654 // If we have an instance type and our memory input is the 1655 // programs's initial memory state, there is no matching store, 1656 // so just return a zero of the appropriate type 1657 Node *mem = in(MemNode::Memory); 1658 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1659 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1660 return Type::get_zero_type(_type->basic_type()); 1661 } 1662 } 1663 return _type; 1664} 1665 1666//------------------------------match_edge------------------------------------- 1667// Do we Match on this edge index or not? Match only the address. 1668uint LoadNode::match_edge(uint idx) const { 1669 return idx == MemNode::Address; 1670} 1671 1672//--------------------------LoadBNode::Ideal-------------------------------------- 1673// 1674// If the previous store is to the same address as this load, 1675// and the value stored was larger than a byte, replace this load 1676// with the value stored truncated to a byte. If no truncation is 1677// needed, the replacement is done in LoadNode::Identity(). 1678// 1679Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1680 Node* mem = in(MemNode::Memory); 1681 Node* value = can_see_stored_value(mem,phase); 1682 if( value && !phase->type(value)->higher_equal( _type ) ) { 1683 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1684 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1685 } 1686 // Identity call will handle the case where truncation is not needed. 1687 return LoadNode::Ideal(phase, can_reshape); 1688} 1689 1690//--------------------------LoadUBNode::Ideal------------------------------------- 1691// 1692// If the previous store is to the same address as this load, 1693// and the value stored was larger than a byte, replace this load 1694// with the value stored truncated to a byte. If no truncation is 1695// needed, the replacement is done in LoadNode::Identity(). 1696// 1697Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1698 Node* mem = in(MemNode::Memory); 1699 Node* value = can_see_stored_value(mem, phase); 1700 if (value && !phase->type(value)->higher_equal(_type)) 1701 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); 1702 // Identity call will handle the case where truncation is not needed. 1703 return LoadNode::Ideal(phase, can_reshape); 1704} 1705 1706//--------------------------LoadUSNode::Ideal------------------------------------- 1707// 1708// If the previous store is to the same address as this load, 1709// and the value stored was larger than a char, replace this load 1710// with the value stored truncated to a char. If no truncation is 1711// needed, the replacement is done in LoadNode::Identity(). 1712// 1713Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1714 Node* mem = in(MemNode::Memory); 1715 Node* value = can_see_stored_value(mem,phase); 1716 if( value && !phase->type(value)->higher_equal( _type ) ) 1717 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1718 // Identity call will handle the case where truncation is not needed. 1719 return LoadNode::Ideal(phase, can_reshape); 1720} 1721 1722//--------------------------LoadSNode::Ideal-------------------------------------- 1723// 1724// If the previous store is to the same address as this load, 1725// and the value stored was larger than a short, replace this load 1726// with the value stored truncated to a short. If no truncation is 1727// needed, the replacement is done in LoadNode::Identity(). 1728// 1729Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1730 Node* mem = in(MemNode::Memory); 1731 Node* value = can_see_stored_value(mem,phase); 1732 if( value && !phase->type(value)->higher_equal( _type ) ) { 1733 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1734 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1735 } 1736 // Identity call will handle the case where truncation is not needed. 1737 return LoadNode::Ideal(phase, can_reshape); 1738} 1739 1740//============================================================================= 1741//----------------------------LoadKlassNode::make------------------------------ 1742// Polymorphic factory method: 1743Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1744 Compile* C = gvn.C; 1745 Node *ctl = NULL; 1746 // sanity check the alias category against the created node type 1747 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); 1748 assert(adr_type != NULL, "expecting TypeOopPtr"); 1749#ifdef _LP64 1750 if (adr_type->is_ptr_to_narrowoop()) { 1751 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1752 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1753 } 1754#endif 1755 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1756 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1757} 1758 1759//------------------------------Value------------------------------------------ 1760const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1761 return klass_value_common(phase); 1762} 1763 1764const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1765 // Either input is TOP ==> the result is TOP 1766 const Type *t1 = phase->type( in(MemNode::Memory) ); 1767 if (t1 == Type::TOP) return Type::TOP; 1768 Node *adr = in(MemNode::Address); 1769 const Type *t2 = phase->type( adr ); 1770 if (t2 == Type::TOP) return Type::TOP; 1771 const TypePtr *tp = t2->is_ptr(); 1772 if (TypePtr::above_centerline(tp->ptr()) || 1773 tp->ptr() == TypePtr::Null) return Type::TOP; 1774 1775 // Return a more precise klass, if possible 1776 const TypeInstPtr *tinst = tp->isa_instptr(); 1777 if (tinst != NULL) { 1778 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1779 int offset = tinst->offset(); 1780 if (ik == phase->C->env()->Class_klass() 1781 && (offset == java_lang_Class::klass_offset_in_bytes() || 1782 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1783 // We are loading a special hidden field from a Class mirror object, 1784 // the field which points to the VM's Klass metaobject. 1785 ciType* t = tinst->java_mirror_type(); 1786 // java_mirror_type returns non-null for compile-time Class constants. 1787 if (t != NULL) { 1788 // constant oop => constant klass 1789 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1790 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1791 } 1792 if (!t->is_klass()) { 1793 // a primitive Class (e.g., int.class) has NULL for a klass field 1794 return TypePtr::NULL_PTR; 1795 } 1796 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1797 return TypeKlassPtr::make(t->as_klass()); 1798 } 1799 // non-constant mirror, so we can't tell what's going on 1800 } 1801 if( !ik->is_loaded() ) 1802 return _type; // Bail out if not loaded 1803 if (offset == oopDesc::klass_offset_in_bytes()) { 1804 if (tinst->klass_is_exact()) { 1805 return TypeKlassPtr::make(ik); 1806 } 1807 // See if we can become precise: no subklasses and no interface 1808 // (Note: We need to support verified interfaces.) 1809 if (!ik->is_interface() && !ik->has_subklass()) { 1810 //assert(!UseExactTypes, "this code should be useless with exact types"); 1811 // Add a dependence; if any subclass added we need to recompile 1812 if (!ik->is_final()) { 1813 // %%% should use stronger assert_unique_concrete_subtype instead 1814 phase->C->dependencies()->assert_leaf_type(ik); 1815 } 1816 // Return precise klass 1817 return TypeKlassPtr::make(ik); 1818 } 1819 1820 // Return root of possible klass 1821 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1822 } 1823 } 1824 1825 // Check for loading klass from an array 1826 const TypeAryPtr *tary = tp->isa_aryptr(); 1827 if( tary != NULL ) { 1828 ciKlass *tary_klass = tary->klass(); 1829 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1830 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1831 if (tary->klass_is_exact()) { 1832 return TypeKlassPtr::make(tary_klass); 1833 } 1834 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1835 // If the klass is an object array, we defer the question to the 1836 // array component klass. 1837 if( ak->is_obj_array_klass() ) { 1838 assert( ak->is_loaded(), "" ); 1839 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1840 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1841 ciInstanceKlass* ik = base_k->as_instance_klass(); 1842 // See if we can become precise: no subklasses and no interface 1843 if (!ik->is_interface() && !ik->has_subklass()) { 1844 //assert(!UseExactTypes, "this code should be useless with exact types"); 1845 // Add a dependence; if any subclass added we need to recompile 1846 if (!ik->is_final()) { 1847 phase->C->dependencies()->assert_leaf_type(ik); 1848 } 1849 // Return precise array klass 1850 return TypeKlassPtr::make(ak); 1851 } 1852 } 1853 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1854 } else { // Found a type-array? 1855 //assert(!UseExactTypes, "this code should be useless with exact types"); 1856 assert( ak->is_type_array_klass(), "" ); 1857 return TypeKlassPtr::make(ak); // These are always precise 1858 } 1859 } 1860 } 1861 1862 // Check for loading klass from an array klass 1863 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1864 if (tkls != NULL && !StressReflectiveCode) { 1865 ciKlass* klass = tkls->klass(); 1866 if( !klass->is_loaded() ) 1867 return _type; // Bail out if not loaded 1868 if( klass->is_obj_array_klass() && 1869 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1870 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1871 // // Always returning precise element type is incorrect, 1872 // // e.g., element type could be object and array may contain strings 1873 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1874 1875 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1876 // according to the element type's subclassing. 1877 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1878 } 1879 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1880 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1881 ciKlass* sup = klass->as_instance_klass()->super(); 1882 // The field is Klass::_super. Return its (constant) value. 1883 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1884 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1885 } 1886 } 1887 1888 // Bailout case 1889 return LoadNode::Value(phase); 1890} 1891 1892//------------------------------Identity--------------------------------------- 1893// To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1894// Also feed through the klass in Allocate(...klass...)._klass. 1895Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1896 return klass_identity_common(phase); 1897} 1898 1899Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 1900 Node* x = LoadNode::Identity(phase); 1901 if (x != this) return x; 1902 1903 // Take apart the address into an oop and and offset. 1904 // Return 'this' if we cannot. 1905 Node* adr = in(MemNode::Address); 1906 intptr_t offset = 0; 1907 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1908 if (base == NULL) return this; 1909 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1910 if (toop == NULL) return this; 1911 1912 // We can fetch the klass directly through an AllocateNode. 1913 // This works even if the klass is not constant (clone or newArray). 1914 if (offset == oopDesc::klass_offset_in_bytes()) { 1915 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1916 if (allocated_klass != NULL) { 1917 return allocated_klass; 1918 } 1919 } 1920 1921 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1922 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1923 // See inline_native_Class_query for occurrences of these patterns. 1924 // Java Example: x.getClass().isAssignableFrom(y) 1925 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1926 // 1927 // This improves reflective code, often making the Class 1928 // mirror go completely dead. (Current exception: Class 1929 // mirrors may appear in debug info, but we could clean them out by 1930 // introducing a new debug info operator for klassOop.java_mirror). 1931 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1932 && (offset == java_lang_Class::klass_offset_in_bytes() || 1933 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1934 // We are loading a special hidden field from a Class mirror, 1935 // the field which points to its Klass or arrayKlass metaobject. 1936 if (base->is_Load()) { 1937 Node* adr2 = base->in(MemNode::Address); 1938 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1939 if (tkls != NULL && !tkls->empty() 1940 && (tkls->klass()->is_instance_klass() || 1941 tkls->klass()->is_array_klass()) 1942 && adr2->is_AddP() 1943 ) { 1944 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1945 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1946 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1947 } 1948 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1949 return adr2->in(AddPNode::Base); 1950 } 1951 } 1952 } 1953 } 1954 1955 return this; 1956} 1957 1958 1959//------------------------------Value------------------------------------------ 1960const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 1961 const Type *t = klass_value_common(phase); 1962 if (t == Type::TOP) 1963 return t; 1964 1965 return t->make_narrowoop(); 1966} 1967 1968//------------------------------Identity--------------------------------------- 1969// To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 1970// Also feed through the klass in Allocate(...klass...)._klass. 1971Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 1972 Node *x = klass_identity_common(phase); 1973 1974 const Type *t = phase->type( x ); 1975 if( t == Type::TOP ) return x; 1976 if( t->isa_narrowoop()) return x; 1977 1978 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 1979} 1980 1981//------------------------------Value----------------------------------------- 1982const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1983 // Either input is TOP ==> the result is TOP 1984 const Type *t1 = phase->type( in(MemNode::Memory) ); 1985 if( t1 == Type::TOP ) return Type::TOP; 1986 Node *adr = in(MemNode::Address); 1987 const Type *t2 = phase->type( adr ); 1988 if( t2 == Type::TOP ) return Type::TOP; 1989 const TypePtr *tp = t2->is_ptr(); 1990 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 1991 const TypeAryPtr *tap = tp->isa_aryptr(); 1992 if( !tap ) return _type; 1993 return tap->size(); 1994} 1995 1996//-------------------------------Ideal--------------------------------------- 1997// Feed through the length in AllocateArray(...length...)._length. 1998Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1999 Node* p = MemNode::Ideal_common(phase, can_reshape); 2000 if (p) return (p == NodeSentinel) ? NULL : p; 2001 2002 // Take apart the address into an oop and and offset. 2003 // Return 'this' if we cannot. 2004 Node* adr = in(MemNode::Address); 2005 intptr_t offset = 0; 2006 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2007 if (base == NULL) return NULL; 2008 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2009 if (tary == NULL) return NULL; 2010 2011 // We can fetch the length directly through an AllocateArrayNode. 2012 // This works even if the length is not constant (clone or newArray). 2013 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2014 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2015 if (alloc != NULL) { 2016 Node* allocated_length = alloc->Ideal_length(); 2017 Node* len = alloc->make_ideal_length(tary, phase); 2018 if (allocated_length != len) { 2019 // New CastII improves on this. 2020 return len; 2021 } 2022 } 2023 } 2024 2025 return NULL; 2026} 2027 2028//------------------------------Identity--------------------------------------- 2029// Feed through the length in AllocateArray(...length...)._length. 2030Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2031 Node* x = LoadINode::Identity(phase); 2032 if (x != this) return x; 2033 2034 // Take apart the address into an oop and and offset. 2035 // Return 'this' if we cannot. 2036 Node* adr = in(MemNode::Address); 2037 intptr_t offset = 0; 2038 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2039 if (base == NULL) return this; 2040 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2041 if (tary == NULL) return this; 2042 2043 // We can fetch the length directly through an AllocateArrayNode. 2044 // This works even if the length is not constant (clone or newArray). 2045 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2046 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2047 if (alloc != NULL) { 2048 Node* allocated_length = alloc->Ideal_length(); 2049 // Do not allow make_ideal_length to allocate a CastII node. 2050 Node* len = alloc->make_ideal_length(tary, phase, false); 2051 if (allocated_length == len) { 2052 // Return allocated_length only if it would not be improved by a CastII. 2053 return allocated_length; 2054 } 2055 } 2056 } 2057 2058 return this; 2059 2060} 2061 2062//============================================================================= 2063//---------------------------StoreNode::make----------------------------------- 2064// Polymorphic factory method: 2065StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 2066 Compile* C = gvn.C; 2067 2068 switch (bt) { 2069 case T_BOOLEAN: 2070 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 2071 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 2072 case T_CHAR: 2073 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 2074 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 2075 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 2076 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 2077 case T_ADDRESS: 2078 case T_OBJECT: 2079#ifdef _LP64 2080 if (adr->bottom_type()->is_ptr_to_narrowoop() || 2081 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 2082 adr->bottom_type()->isa_rawptr())) { 2083 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2084 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 2085 } else 2086#endif 2087 { 2088 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 2089 } 2090 } 2091 ShouldNotReachHere(); 2092 return (StoreNode*)NULL; 2093} 2094 2095StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2096 bool require_atomic = true; 2097 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2098} 2099 2100 2101//--------------------------bottom_type---------------------------------------- 2102const Type *StoreNode::bottom_type() const { 2103 return Type::MEMORY; 2104} 2105 2106//------------------------------hash------------------------------------------- 2107uint StoreNode::hash() const { 2108 // unroll addition of interesting fields 2109 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2110 2111 // Since they are not commoned, do not hash them: 2112 return NO_HASH; 2113} 2114 2115//------------------------------Ideal------------------------------------------ 2116// Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2117// When a store immediately follows a relevant allocation/initialization, 2118// try to capture it into the initialization, or hoist it above. 2119Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2120 Node* p = MemNode::Ideal_common(phase, can_reshape); 2121 if (p) return (p == NodeSentinel) ? NULL : p; 2122 2123 Node* mem = in(MemNode::Memory); 2124 Node* address = in(MemNode::Address); 2125 2126 // Back-to-back stores to same address? Fold em up. 2127 // Generally unsafe if I have intervening uses... 2128 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 2129 // Looking at a dead closed cycle of memory? 2130 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2131 2132 assert(Opcode() == mem->Opcode() || 2133 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2134 "no mismatched stores, except on raw memory"); 2135 2136 if (mem->outcnt() == 1 && // check for intervening uses 2137 mem->as_Store()->memory_size() <= this->memory_size()) { 2138 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2139 // For example, 'mem' might be the final state at a conditional return. 2140 // Or, 'mem' might be used by some node which is live at the same time 2141 // 'this' is live, which might be unschedulable. So, require exactly 2142 // ONE user, the 'this' store, until such time as we clone 'mem' for 2143 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2144 if (can_reshape) { // (%%% is this an anachronism?) 2145 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2146 phase->is_IterGVN()); 2147 } else { 2148 // It's OK to do this in the parser, since DU info is always accurate, 2149 // and the parser always refers to nodes via SafePointNode maps. 2150 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2151 } 2152 return this; 2153 } 2154 } 2155 2156 // Capture an unaliased, unconditional, simple store into an initializer. 2157 // Or, if it is independent of the allocation, hoist it above the allocation. 2158 if (ReduceFieldZeroing && /*can_reshape &&*/ 2159 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2160 InitializeNode* init = mem->in(0)->as_Initialize(); 2161 intptr_t offset = init->can_capture_store(this, phase); 2162 if (offset > 0) { 2163 Node* moved = init->capture_store(this, offset, phase); 2164 // If the InitializeNode captured me, it made a raw copy of me, 2165 // and I need to disappear. 2166 if (moved != NULL) { 2167 // %%% hack to ensure that Ideal returns a new node: 2168 mem = MergeMemNode::make(phase->C, mem); 2169 return mem; // fold me away 2170 } 2171 } 2172 } 2173 2174 return NULL; // No further progress 2175} 2176 2177//------------------------------Value----------------------------------------- 2178const Type *StoreNode::Value( PhaseTransform *phase ) const { 2179 // Either input is TOP ==> the result is TOP 2180 const Type *t1 = phase->type( in(MemNode::Memory) ); 2181 if( t1 == Type::TOP ) return Type::TOP; 2182 const Type *t2 = phase->type( in(MemNode::Address) ); 2183 if( t2 == Type::TOP ) return Type::TOP; 2184 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2185 if( t3 == Type::TOP ) return Type::TOP; 2186 return Type::MEMORY; 2187} 2188 2189//------------------------------Identity--------------------------------------- 2190// Remove redundant stores: 2191// Store(m, p, Load(m, p)) changes to m. 2192// Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2193Node *StoreNode::Identity( PhaseTransform *phase ) { 2194 Node* mem = in(MemNode::Memory); 2195 Node* adr = in(MemNode::Address); 2196 Node* val = in(MemNode::ValueIn); 2197 2198 // Load then Store? Then the Store is useless 2199 if (val->is_Load() && 2200 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 2201 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 2202 val->as_Load()->store_Opcode() == Opcode()) { 2203 return mem; 2204 } 2205 2206 // Two stores in a row of the same value? 2207 if (mem->is_Store() && 2208 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 2209 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 2210 mem->Opcode() == Opcode()) { 2211 return mem; 2212 } 2213 2214 // Store of zero anywhere into a freshly-allocated object? 2215 // Then the store is useless. 2216 // (It must already have been captured by the InitializeNode.) 2217 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2218 // a newly allocated object is already all-zeroes everywhere 2219 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2220 return mem; 2221 } 2222 2223 // the store may also apply to zero-bits in an earlier object 2224 Node* prev_mem = find_previous_store(phase); 2225 // Steps (a), (b): Walk past independent stores to find an exact match. 2226 if (prev_mem != NULL) { 2227 Node* prev_val = can_see_stored_value(prev_mem, phase); 2228 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2229 // prev_val and val might differ by a cast; it would be good 2230 // to keep the more informative of the two. 2231 return mem; 2232 } 2233 } 2234 } 2235 2236 return this; 2237} 2238 2239//------------------------------match_edge------------------------------------- 2240// Do we Match on this edge index or not? Match only memory & value 2241uint StoreNode::match_edge(uint idx) const { 2242 return idx == MemNode::Address || idx == MemNode::ValueIn; 2243} 2244 2245//------------------------------cmp-------------------------------------------- 2246// Do not common stores up together. They generally have to be split 2247// back up anyways, so do not bother. 2248uint StoreNode::cmp( const Node &n ) const { 2249 return (&n == this); // Always fail except on self 2250} 2251 2252//------------------------------Ideal_masked_input----------------------------- 2253// Check for a useless mask before a partial-word store 2254// (StoreB ... (AndI valIn conIa) ) 2255// If (conIa & mask == mask) this simplifies to 2256// (StoreB ... (valIn) ) 2257Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2258 Node *val = in(MemNode::ValueIn); 2259 if( val->Opcode() == Op_AndI ) { 2260 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2261 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2262 set_req(MemNode::ValueIn, val->in(1)); 2263 return this; 2264 } 2265 } 2266 return NULL; 2267} 2268 2269 2270//------------------------------Ideal_sign_extended_input---------------------- 2271// Check for useless sign-extension before a partial-word store 2272// (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2273// If (conIL == conIR && conIR <= num_bits) this simplifies to 2274// (StoreB ... (valIn) ) 2275Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2276 Node *val = in(MemNode::ValueIn); 2277 if( val->Opcode() == Op_RShiftI ) { 2278 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2279 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2280 Node *shl = val->in(1); 2281 if( shl->Opcode() == Op_LShiftI ) { 2282 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2283 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2284 set_req(MemNode::ValueIn, shl->in(1)); 2285 return this; 2286 } 2287 } 2288 } 2289 } 2290 return NULL; 2291} 2292 2293//------------------------------value_never_loaded----------------------------------- 2294// Determine whether there are any possible loads of the value stored. 2295// For simplicity, we actually check if there are any loads from the 2296// address stored to, not just for loads of the value stored by this node. 2297// 2298bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2299 Node *adr = in(Address); 2300 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2301 if (adr_oop == NULL) 2302 return false; 2303 if (!adr_oop->is_known_instance_field()) 2304 return false; // if not a distinct instance, there may be aliases of the address 2305 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2306 Node *use = adr->fast_out(i); 2307 int opc = use->Opcode(); 2308 if (use->is_Load() || use->is_LoadStore()) { 2309 return false; 2310 } 2311 } 2312 return true; 2313} 2314 2315//============================================================================= 2316//------------------------------Ideal------------------------------------------ 2317// If the store is from an AND mask that leaves the low bits untouched, then 2318// we can skip the AND operation. If the store is from a sign-extension 2319// (a left shift, then right shift) we can skip both. 2320Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2321 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2322 if( progress != NULL ) return progress; 2323 2324 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2325 if( progress != NULL ) return progress; 2326 2327 // Finally check the default case 2328 return StoreNode::Ideal(phase, can_reshape); 2329} 2330 2331//============================================================================= 2332//------------------------------Ideal------------------------------------------ 2333// If the store is from an AND mask that leaves the low bits untouched, then 2334// we can skip the AND operation 2335Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2336 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2337 if( progress != NULL ) return progress; 2338 2339 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2340 if( progress != NULL ) return progress; 2341 2342 // Finally check the default case 2343 return StoreNode::Ideal(phase, can_reshape); 2344} 2345 2346//============================================================================= 2347//------------------------------Identity--------------------------------------- 2348Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2349 // No need to card mark when storing a null ptr 2350 Node* my_store = in(MemNode::OopStore); 2351 if (my_store->is_Store()) { 2352 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2353 if( t1 == TypePtr::NULL_PTR ) { 2354 return in(MemNode::Memory); 2355 } 2356 } 2357 return this; 2358} 2359 2360//============================================================================= 2361//------------------------------Ideal--------------------------------------- 2362Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2363 Node* progress = StoreNode::Ideal(phase, can_reshape); 2364 if (progress != NULL) return progress; 2365 2366 Node* my_store = in(MemNode::OopStore); 2367 if (my_store->is_MergeMem()) { 2368 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2369 set_req(MemNode::OopStore, mem); 2370 return this; 2371 } 2372 2373 return NULL; 2374} 2375 2376//------------------------------Value----------------------------------------- 2377const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2378 // Either input is TOP ==> the result is TOP 2379 const Type *t = phase->type( in(MemNode::Memory) ); 2380 if( t == Type::TOP ) return Type::TOP; 2381 t = phase->type( in(MemNode::Address) ); 2382 if( t == Type::TOP ) return Type::TOP; 2383 t = phase->type( in(MemNode::ValueIn) ); 2384 if( t == Type::TOP ) return Type::TOP; 2385 // If extra input is TOP ==> the result is TOP 2386 t = phase->type( in(MemNode::OopStore) ); 2387 if( t == Type::TOP ) return Type::TOP; 2388 2389 return StoreNode::Value( phase ); 2390} 2391 2392 2393//============================================================================= 2394//----------------------------------SCMemProjNode------------------------------ 2395const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2396{ 2397 return bottom_type(); 2398} 2399 2400//============================================================================= 2401LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2402 init_req(MemNode::Control, c ); 2403 init_req(MemNode::Memory , mem); 2404 init_req(MemNode::Address, adr); 2405 init_req(MemNode::ValueIn, val); 2406 init_req( ExpectedIn, ex ); 2407 init_class_id(Class_LoadStore); 2408 2409} 2410 2411//============================================================================= 2412//-------------------------------adr_type-------------------------------------- 2413// Do we Match on this edge index or not? Do not match memory 2414const TypePtr* ClearArrayNode::adr_type() const { 2415 Node *adr = in(3); 2416 return MemNode::calculate_adr_type(adr->bottom_type()); 2417} 2418 2419//------------------------------match_edge------------------------------------- 2420// Do we Match on this edge index or not? Do not match memory 2421uint ClearArrayNode::match_edge(uint idx) const { 2422 return idx > 1; 2423} 2424 2425//------------------------------Identity--------------------------------------- 2426// Clearing a zero length array does nothing 2427Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2428 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2429} 2430 2431//------------------------------Idealize--------------------------------------- 2432// Clearing a short array is faster with stores 2433Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2434 const int unit = BytesPerLong; 2435 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2436 if (!t) return NULL; 2437 if (!t->is_con()) return NULL; 2438 intptr_t raw_count = t->get_con(); 2439 intptr_t size = raw_count; 2440 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2441 // Clearing nothing uses the Identity call. 2442 // Negative clears are possible on dead ClearArrays 2443 // (see jck test stmt114.stmt11402.val). 2444 if (size <= 0 || size % unit != 0) return NULL; 2445 intptr_t count = size / unit; 2446 // Length too long; use fast hardware clear 2447 if (size > Matcher::init_array_short_size) return NULL; 2448 Node *mem = in(1); 2449 if( phase->type(mem)==Type::TOP ) return NULL; 2450 Node *adr = in(3); 2451 const Type* at = phase->type(adr); 2452 if( at==Type::TOP ) return NULL; 2453 const TypePtr* atp = at->isa_ptr(); 2454 // adjust atp to be the correct array element address type 2455 if (atp == NULL) atp = TypePtr::BOTTOM; 2456 else atp = atp->add_offset(Type::OffsetBot); 2457 // Get base for derived pointer purposes 2458 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2459 Node *base = adr->in(1); 2460 2461 Node *zero = phase->makecon(TypeLong::ZERO); 2462 Node *off = phase->MakeConX(BytesPerLong); 2463 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2464 count--; 2465 while( count-- ) { 2466 mem = phase->transform(mem); 2467 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2468 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2469 } 2470 return mem; 2471} 2472 2473//----------------------------step_through---------------------------------- 2474// Return allocation input memory edge if it is different instance 2475// or itself if it is the one we are looking for. 2476bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2477 Node* n = *np; 2478 assert(n->is_ClearArray(), "sanity"); 2479 intptr_t offset; 2480 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2481 // This method is called only before Allocate nodes are expanded during 2482 // macro nodes expansion. Before that ClearArray nodes are only generated 2483 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2484 assert(alloc != NULL, "should have allocation"); 2485 if (alloc->_idx == instance_id) { 2486 // Can not bypass initialization of the instance we are looking for. 2487 return false; 2488 } 2489 // Otherwise skip it. 2490 InitializeNode* init = alloc->initialization(); 2491 if (init != NULL) 2492 *np = init->in(TypeFunc::Memory); 2493 else 2494 *np = alloc->in(TypeFunc::Memory); 2495 return true; 2496} 2497 2498//----------------------------clear_memory------------------------------------- 2499// Generate code to initialize object storage to zero. 2500Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2501 intptr_t start_offset, 2502 Node* end_offset, 2503 PhaseGVN* phase) { 2504 Compile* C = phase->C; 2505 intptr_t offset = start_offset; 2506 2507 int unit = BytesPerLong; 2508 if ((offset % unit) != 0) { 2509 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2510 adr = phase->transform(adr); 2511 const TypePtr* atp = TypeRawPtr::BOTTOM; 2512 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2513 mem = phase->transform(mem); 2514 offset += BytesPerInt; 2515 } 2516 assert((offset % unit) == 0, ""); 2517 2518 // Initialize the remaining stuff, if any, with a ClearArray. 2519 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2520} 2521 2522Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2523 Node* start_offset, 2524 Node* end_offset, 2525 PhaseGVN* phase) { 2526 if (start_offset == end_offset) { 2527 // nothing to do 2528 return mem; 2529 } 2530 2531 Compile* C = phase->C; 2532 int unit = BytesPerLong; 2533 Node* zbase = start_offset; 2534 Node* zend = end_offset; 2535 2536 // Scale to the unit required by the CPU: 2537 if (!Matcher::init_array_count_is_in_bytes) { 2538 Node* shift = phase->intcon(exact_log2(unit)); 2539 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2540 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2541 } 2542 2543 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2544 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2545 2546 // Bulk clear double-words 2547 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2548 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2549 return phase->transform(mem); 2550} 2551 2552Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2553 intptr_t start_offset, 2554 intptr_t end_offset, 2555 PhaseGVN* phase) { 2556 if (start_offset == end_offset) { 2557 // nothing to do 2558 return mem; 2559 } 2560 2561 Compile* C = phase->C; 2562 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2563 intptr_t done_offset = end_offset; 2564 if ((done_offset % BytesPerLong) != 0) { 2565 done_offset -= BytesPerInt; 2566 } 2567 if (done_offset > start_offset) { 2568 mem = clear_memory(ctl, mem, dest, 2569 start_offset, phase->MakeConX(done_offset), phase); 2570 } 2571 if (done_offset < end_offset) { // emit the final 32-bit store 2572 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2573 adr = phase->transform(adr); 2574 const TypePtr* atp = TypeRawPtr::BOTTOM; 2575 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2576 mem = phase->transform(mem); 2577 done_offset += BytesPerInt; 2578 } 2579 assert(done_offset == end_offset, ""); 2580 return mem; 2581} 2582 2583//============================================================================= 2584// Do we match on this edge? No memory edges 2585uint StrCompNode::match_edge(uint idx) const { 2586 return idx == 2 || idx == 3; // StrComp (Binary str1 cnt1) (Binary str2 cnt2) 2587} 2588 2589//------------------------------Ideal------------------------------------------ 2590// Return a node which is more "ideal" than the current node. Strip out 2591// control copies 2592Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2593 return remove_dead_region(phase, can_reshape) ? this : NULL; 2594} 2595 2596//============================================================================= 2597// Do we match on this edge? No memory edges 2598uint StrEqualsNode::match_edge(uint idx) const { 2599 return idx == 2 || idx == 3; // StrEquals (Binary str1 str2) cnt 2600} 2601 2602//------------------------------Ideal------------------------------------------ 2603// Return a node which is more "ideal" than the current node. Strip out 2604// control copies 2605Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2606 return remove_dead_region(phase, can_reshape) ? this : NULL; 2607} 2608 2609//============================================================================= 2610// Do we match on this edge? No memory edges 2611uint StrIndexOfNode::match_edge(uint idx) const { 2612 return idx == 2 || idx == 3; // StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2) 2613} 2614 2615//------------------------------Ideal------------------------------------------ 2616// Return a node which is more "ideal" than the current node. Strip out 2617// control copies 2618Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2619 return remove_dead_region(phase, can_reshape) ? this : NULL; 2620} 2621 2622//============================================================================= 2623// Do we match on this edge? No memory edges 2624uint AryEqNode::match_edge(uint idx) const { 2625 return idx == 2 || idx == 3; // StrEquals ary1 ary2 2626} 2627//------------------------------Ideal------------------------------------------ 2628// Return a node which is more "ideal" than the current node. Strip out 2629// control copies 2630Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2631 return remove_dead_region(phase, can_reshape) ? this : NULL; 2632} 2633 2634//============================================================================= 2635MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2636 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2637 _adr_type(C->get_adr_type(alias_idx)) 2638{ 2639 init_class_id(Class_MemBar); 2640 Node* top = C->top(); 2641 init_req(TypeFunc::I_O,top); 2642 init_req(TypeFunc::FramePtr,top); 2643 init_req(TypeFunc::ReturnAdr,top); 2644 if (precedent != NULL) 2645 init_req(TypeFunc::Parms, precedent); 2646} 2647 2648//------------------------------cmp-------------------------------------------- 2649uint MemBarNode::hash() const { return NO_HASH; } 2650uint MemBarNode::cmp( const Node &n ) const { 2651 return (&n == this); // Always fail except on self 2652} 2653 2654//------------------------------make------------------------------------------- 2655MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2656 int len = Precedent + (pn == NULL? 0: 1); 2657 switch (opcode) { 2658 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2659 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2660 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2661 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2662 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2663 default: ShouldNotReachHere(); return NULL; 2664 } 2665} 2666 2667//------------------------------Ideal------------------------------------------ 2668// Return a node which is more "ideal" than the current node. Strip out 2669// control copies 2670Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2671 if (remove_dead_region(phase, can_reshape)) return this; 2672 2673 // Eliminate volatile MemBars for scalar replaced objects. 2674 if (can_reshape && req() == (Precedent+1) && 2675 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { 2676 // Volatile field loads and stores. 2677 Node* my_mem = in(MemBarNode::Precedent); 2678 if (my_mem != NULL && my_mem->is_Mem()) { 2679 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2680 // Check for scalar replaced object reference. 2681 if( t_oop != NULL && t_oop->is_known_instance_field() && 2682 t_oop->offset() != Type::OffsetBot && 2683 t_oop->offset() != Type::OffsetTop) { 2684 // Replace MemBar projections by its inputs. 2685 PhaseIterGVN* igvn = phase->is_IterGVN(); 2686 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2687 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2688 // Must return either the original node (now dead) or a new node 2689 // (Do not return a top here, since that would break the uniqueness of top.) 2690 return new (phase->C, 1) ConINode(TypeInt::ZERO); 2691 } 2692 } 2693 } 2694 return NULL; 2695} 2696 2697//------------------------------Value------------------------------------------ 2698const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2699 if( !in(0) ) return Type::TOP; 2700 if( phase->type(in(0)) == Type::TOP ) 2701 return Type::TOP; 2702 return TypeTuple::MEMBAR; 2703} 2704 2705//------------------------------match------------------------------------------ 2706// Construct projections for memory. 2707Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2708 switch (proj->_con) { 2709 case TypeFunc::Control: 2710 case TypeFunc::Memory: 2711 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2712 } 2713 ShouldNotReachHere(); 2714 return NULL; 2715} 2716 2717//===========================InitializeNode==================================== 2718// SUMMARY: 2719// This node acts as a memory barrier on raw memory, after some raw stores. 2720// The 'cooked' oop value feeds from the Initialize, not the Allocation. 2721// The Initialize can 'capture' suitably constrained stores as raw inits. 2722// It can coalesce related raw stores into larger units (called 'tiles'). 2723// It can avoid zeroing new storage for memory units which have raw inits. 2724// At macro-expansion, it is marked 'complete', and does not optimize further. 2725// 2726// EXAMPLE: 2727// The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2728// ctl = incoming control; mem* = incoming memory 2729// (Note: A star * on a memory edge denotes I/O and other standard edges.) 2730// First allocate uninitialized memory and fill in the header: 2731// alloc = (Allocate ctl mem* 16 #short[].klass ...) 2732// ctl := alloc.Control; mem* := alloc.Memory* 2733// rawmem = alloc.Memory; rawoop = alloc.RawAddress 2734// Then initialize to zero the non-header parts of the raw memory block: 2735// init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2736// ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2737// After the initialize node executes, the object is ready for service: 2738// oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2739// Suppose its body is immediately initialized as {1,2}: 2740// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2741// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2742// mem.SLICE(#short[*]) := store2 2743// 2744// DETAILS: 2745// An InitializeNode collects and isolates object initialization after 2746// an AllocateNode and before the next possible safepoint. As a 2747// memory barrier (MemBarNode), it keeps critical stores from drifting 2748// down past any safepoint or any publication of the allocation. 2749// Before this barrier, a newly-allocated object may have uninitialized bits. 2750// After this barrier, it may be treated as a real oop, and GC is allowed. 2751// 2752// The semantics of the InitializeNode include an implicit zeroing of 2753// the new object from object header to the end of the object. 2754// (The object header and end are determined by the AllocateNode.) 2755// 2756// Certain stores may be added as direct inputs to the InitializeNode. 2757// These stores must update raw memory, and they must be to addresses 2758// derived from the raw address produced by AllocateNode, and with 2759// a constant offset. They must be ordered by increasing offset. 2760// The first one is at in(RawStores), the last at in(req()-1). 2761// Unlike most memory operations, they are not linked in a chain, 2762// but are displayed in parallel as users of the rawmem output of 2763// the allocation. 2764// 2765// (See comments in InitializeNode::capture_store, which continue 2766// the example given above.) 2767// 2768// When the associated Allocate is macro-expanded, the InitializeNode 2769// may be rewritten to optimize collected stores. A ClearArrayNode 2770// may also be created at that point to represent any required zeroing. 2771// The InitializeNode is then marked 'complete', prohibiting further 2772// capturing of nearby memory operations. 2773// 2774// During macro-expansion, all captured initializations which store 2775// constant values of 32 bits or smaller are coalesced (if advantageous) 2776// into larger 'tiles' 32 or 64 bits. This allows an object to be 2777// initialized in fewer memory operations. Memory words which are 2778// covered by neither tiles nor non-constant stores are pre-zeroed 2779// by explicit stores of zero. (The code shape happens to do all 2780// zeroing first, then all other stores, with both sequences occurring 2781// in order of ascending offsets.) 2782// 2783// Alternatively, code may be inserted between an AllocateNode and its 2784// InitializeNode, to perform arbitrary initialization of the new object. 2785// E.g., the object copying intrinsics insert complex data transfers here. 2786// The initialization must then be marked as 'complete' disable the 2787// built-in zeroing semantics and the collection of initializing stores. 2788// 2789// While an InitializeNode is incomplete, reads from the memory state 2790// produced by it are optimizable if they match the control edge and 2791// new oop address associated with the allocation/initialization. 2792// They return a stored value (if the offset matches) or else zero. 2793// A write to the memory state, if it matches control and address, 2794// and if it is to a constant offset, may be 'captured' by the 2795// InitializeNode. It is cloned as a raw memory operation and rewired 2796// inside the initialization, to the raw oop produced by the allocation. 2797// Operations on addresses which are provably distinct (e.g., to 2798// other AllocateNodes) are allowed to bypass the initialization. 2799// 2800// The effect of all this is to consolidate object initialization 2801// (both arrays and non-arrays, both piecewise and bulk) into a 2802// single location, where it can be optimized as a unit. 2803// 2804// Only stores with an offset less than TrackedInitializationLimit words 2805// will be considered for capture by an InitializeNode. This puts a 2806// reasonable limit on the complexity of optimized initializations. 2807 2808//---------------------------InitializeNode------------------------------------ 2809InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2810 : _is_complete(false), 2811 MemBarNode(C, adr_type, rawoop) 2812{ 2813 init_class_id(Class_Initialize); 2814 2815 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2816 assert(in(RawAddress) == rawoop, "proper init"); 2817 // Note: allocation() can be NULL, for secondary initialization barriers 2818} 2819 2820// Since this node is not matched, it will be processed by the 2821// register allocator. Declare that there are no constraints 2822// on the allocation of the RawAddress edge. 2823const RegMask &InitializeNode::in_RegMask(uint idx) const { 2824 // This edge should be set to top, by the set_complete. But be conservative. 2825 if (idx == InitializeNode::RawAddress) 2826 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2827 return RegMask::Empty; 2828} 2829 2830Node* InitializeNode::memory(uint alias_idx) { 2831 Node* mem = in(Memory); 2832 if (mem->is_MergeMem()) { 2833 return mem->as_MergeMem()->memory_at(alias_idx); 2834 } else { 2835 // incoming raw memory is not split 2836 return mem; 2837 } 2838} 2839 2840bool InitializeNode::is_non_zero() { 2841 if (is_complete()) return false; 2842 remove_extra_zeroes(); 2843 return (req() > RawStores); 2844} 2845 2846void InitializeNode::set_complete(PhaseGVN* phase) { 2847 assert(!is_complete(), "caller responsibility"); 2848 _is_complete = true; 2849 2850 // After this node is complete, it contains a bunch of 2851 // raw-memory initializations. There is no need for 2852 // it to have anything to do with non-raw memory effects. 2853 // Therefore, tell all non-raw users to re-optimize themselves, 2854 // after skipping the memory effects of this initialization. 2855 PhaseIterGVN* igvn = phase->is_IterGVN(); 2856 if (igvn) igvn->add_users_to_worklist(this); 2857} 2858 2859// convenience function 2860// return false if the init contains any stores already 2861bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2862 InitializeNode* init = initialization(); 2863 if (init == NULL || init->is_complete()) return false; 2864 init->remove_extra_zeroes(); 2865 // for now, if this allocation has already collected any inits, bail: 2866 if (init->is_non_zero()) return false; 2867 init->set_complete(phase); 2868 return true; 2869} 2870 2871void InitializeNode::remove_extra_zeroes() { 2872 if (req() == RawStores) return; 2873 Node* zmem = zero_memory(); 2874 uint fill = RawStores; 2875 for (uint i = fill; i < req(); i++) { 2876 Node* n = in(i); 2877 if (n->is_top() || n == zmem) continue; // skip 2878 if (fill < i) set_req(fill, n); // compact 2879 ++fill; 2880 } 2881 // delete any empty spaces created: 2882 while (fill < req()) { 2883 del_req(fill); 2884 } 2885} 2886 2887// Helper for remembering which stores go with which offsets. 2888intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2889 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2890 intptr_t offset = -1; 2891 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2892 phase, offset); 2893 if (base == NULL) return -1; // something is dead, 2894 if (offset < 0) return -1; // dead, dead 2895 return offset; 2896} 2897 2898// Helper for proving that an initialization expression is 2899// "simple enough" to be folded into an object initialization. 2900// Attempts to prove that a store's initial value 'n' can be captured 2901// within the initialization without creating a vicious cycle, such as: 2902// { Foo p = new Foo(); p.next = p; } 2903// True for constants and parameters and small combinations thereof. 2904bool InitializeNode::detect_init_independence(Node* n, 2905 bool st_is_pinned, 2906 int& count) { 2907 if (n == NULL) return true; // (can this really happen?) 2908 if (n->is_Proj()) n = n->in(0); 2909 if (n == this) return false; // found a cycle 2910 if (n->is_Con()) return true; 2911 if (n->is_Start()) return true; // params, etc., are OK 2912 if (n->is_Root()) return true; // even better 2913 2914 Node* ctl = n->in(0); 2915 if (ctl != NULL && !ctl->is_top()) { 2916 if (ctl->is_Proj()) ctl = ctl->in(0); 2917 if (ctl == this) return false; 2918 2919 // If we already know that the enclosing memory op is pinned right after 2920 // the init, then any control flow that the store has picked up 2921 // must have preceded the init, or else be equal to the init. 2922 // Even after loop optimizations (which might change control edges) 2923 // a store is never pinned *before* the availability of its inputs. 2924 if (!MemNode::all_controls_dominate(n, this)) 2925 return false; // failed to prove a good control 2926 2927 } 2928 2929 // Check data edges for possible dependencies on 'this'. 2930 if ((count += 1) > 20) return false; // complexity limit 2931 for (uint i = 1; i < n->req(); i++) { 2932 Node* m = n->in(i); 2933 if (m == NULL || m == n || m->is_top()) continue; 2934 uint first_i = n->find_edge(m); 2935 if (i != first_i) continue; // process duplicate edge just once 2936 if (!detect_init_independence(m, st_is_pinned, count)) { 2937 return false; 2938 } 2939 } 2940 2941 return true; 2942} 2943 2944// Here are all the checks a Store must pass before it can be moved into 2945// an initialization. Returns zero if a check fails. 2946// On success, returns the (constant) offset to which the store applies, 2947// within the initialized memory. 2948intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2949 const int FAIL = 0; 2950 if (st->req() != MemNode::ValueIn + 1) 2951 return FAIL; // an inscrutable StoreNode (card mark?) 2952 Node* ctl = st->in(MemNode::Control); 2953 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2954 return FAIL; // must be unconditional after the initialization 2955 Node* mem = st->in(MemNode::Memory); 2956 if (!(mem->is_Proj() && mem->in(0) == this)) 2957 return FAIL; // must not be preceded by other stores 2958 Node* adr = st->in(MemNode::Address); 2959 intptr_t offset; 2960 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2961 if (alloc == NULL) 2962 return FAIL; // inscrutable address 2963 if (alloc != allocation()) 2964 return FAIL; // wrong allocation! (store needs to float up) 2965 Node* val = st->in(MemNode::ValueIn); 2966 int complexity_count = 0; 2967 if (!detect_init_independence(val, true, complexity_count)) 2968 return FAIL; // stored value must be 'simple enough' 2969 2970 return offset; // success 2971} 2972 2973// Find the captured store in(i) which corresponds to the range 2974// [start..start+size) in the initialized object. 2975// If there is one, return its index i. If there isn't, return the 2976// negative of the index where it should be inserted. 2977// Return 0 if the queried range overlaps an initialization boundary 2978// or if dead code is encountered. 2979// If size_in_bytes is zero, do not bother with overlap checks. 2980int InitializeNode::captured_store_insertion_point(intptr_t start, 2981 int size_in_bytes, 2982 PhaseTransform* phase) { 2983 const int FAIL = 0, MAX_STORE = BytesPerLong; 2984 2985 if (is_complete()) 2986 return FAIL; // arraycopy got here first; punt 2987 2988 assert(allocation() != NULL, "must be present"); 2989 2990 // no negatives, no header fields: 2991 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 2992 2993 // after a certain size, we bail out on tracking all the stores: 2994 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2995 if (start >= ti_limit) return FAIL; 2996 2997 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 2998 if (i >= limit) return -(int)i; // not found; here is where to put it 2999 3000 Node* st = in(i); 3001 intptr_t st_off = get_store_offset(st, phase); 3002 if (st_off < 0) { 3003 if (st != zero_memory()) { 3004 return FAIL; // bail out if there is dead garbage 3005 } 3006 } else if (st_off > start) { 3007 // ...we are done, since stores are ordered 3008 if (st_off < start + size_in_bytes) { 3009 return FAIL; // the next store overlaps 3010 } 3011 return -(int)i; // not found; here is where to put it 3012 } else if (st_off < start) { 3013 if (size_in_bytes != 0 && 3014 start < st_off + MAX_STORE && 3015 start < st_off + st->as_Store()->memory_size()) { 3016 return FAIL; // the previous store overlaps 3017 } 3018 } else { 3019 if (size_in_bytes != 0 && 3020 st->as_Store()->memory_size() != size_in_bytes) { 3021 return FAIL; // mismatched store size 3022 } 3023 return i; 3024 } 3025 3026 ++i; 3027 } 3028} 3029 3030// Look for a captured store which initializes at the offset 'start' 3031// with the given size. If there is no such store, and no other 3032// initialization interferes, then return zero_memory (the memory 3033// projection of the AllocateNode). 3034Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3035 PhaseTransform* phase) { 3036 assert(stores_are_sane(phase), ""); 3037 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3038 if (i == 0) { 3039 return NULL; // something is dead 3040 } else if (i < 0) { 3041 return zero_memory(); // just primordial zero bits here 3042 } else { 3043 Node* st = in(i); // here is the store at this position 3044 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3045 return st; 3046 } 3047} 3048 3049// Create, as a raw pointer, an address within my new object at 'offset'. 3050Node* InitializeNode::make_raw_address(intptr_t offset, 3051 PhaseTransform* phase) { 3052 Node* addr = in(RawAddress); 3053 if (offset != 0) { 3054 Compile* C = phase->C; 3055 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 3056 phase->MakeConX(offset)) ); 3057 } 3058 return addr; 3059} 3060 3061// Clone the given store, converting it into a raw store 3062// initializing a field or element of my new object. 3063// Caller is responsible for retiring the original store, 3064// with subsume_node or the like. 3065// 3066// From the example above InitializeNode::InitializeNode, 3067// here are the old stores to be captured: 3068// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3069// store2 = (StoreC init.Control store1 (+ oop 14) 2) 3070// 3071// Here is the changed code; note the extra edges on init: 3072// alloc = (Allocate ...) 3073// rawoop = alloc.RawAddress 3074// rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3075// rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3076// init = (Initialize alloc.Control alloc.Memory rawoop 3077// rawstore1 rawstore2) 3078// 3079Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3080 PhaseTransform* phase) { 3081 assert(stores_are_sane(phase), ""); 3082 3083 if (start < 0) return NULL; 3084 assert(can_capture_store(st, phase) == start, "sanity"); 3085 3086 Compile* C = phase->C; 3087 int size_in_bytes = st->memory_size(); 3088 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3089 if (i == 0) return NULL; // bail out 3090 Node* prev_mem = NULL; // raw memory for the captured store 3091 if (i > 0) { 3092 prev_mem = in(i); // there is a pre-existing store under this one 3093 set_req(i, C->top()); // temporarily disconnect it 3094 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3095 } else { 3096 i = -i; // no pre-existing store 3097 prev_mem = zero_memory(); // a slice of the newly allocated object 3098 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3099 set_req(--i, C->top()); // reuse this edge; it has been folded away 3100 else 3101 ins_req(i, C->top()); // build a new edge 3102 } 3103 Node* new_st = st->clone(); 3104 new_st->set_req(MemNode::Control, in(Control)); 3105 new_st->set_req(MemNode::Memory, prev_mem); 3106 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3107 new_st = phase->transform(new_st); 3108 3109 // At this point, new_st might have swallowed a pre-existing store 3110 // at the same offset, or perhaps new_st might have disappeared, 3111 // if it redundantly stored the same value (or zero to fresh memory). 3112 3113 // In any case, wire it in: 3114 set_req(i, new_st); 3115 3116 // The caller may now kill the old guy. 3117 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3118 assert(check_st == new_st || check_st == NULL, "must be findable"); 3119 assert(!is_complete(), ""); 3120 return new_st; 3121} 3122 3123static bool store_constant(jlong* tiles, int num_tiles, 3124 intptr_t st_off, int st_size, 3125 jlong con) { 3126 if ((st_off & (st_size-1)) != 0) 3127 return false; // strange store offset (assume size==2**N) 3128 address addr = (address)tiles + st_off; 3129 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3130 switch (st_size) { 3131 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3132 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3133 case sizeof(jint): *(jint*) addr = (jint) con; break; 3134 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3135 default: return false; // strange store size (detect size!=2**N here) 3136 } 3137 return true; // return success to caller 3138} 3139 3140// Coalesce subword constants into int constants and possibly 3141// into long constants. The goal, if the CPU permits, 3142// is to initialize the object with a small number of 64-bit tiles. 3143// Also, convert floating-point constants to bit patterns. 3144// Non-constants are not relevant to this pass. 3145// 3146// In terms of the running example on InitializeNode::InitializeNode 3147// and InitializeNode::capture_store, here is the transformation 3148// of rawstore1 and rawstore2 into rawstore12: 3149// alloc = (Allocate ...) 3150// rawoop = alloc.RawAddress 3151// tile12 = 0x00010002 3152// rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3153// init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3154// 3155void 3156InitializeNode::coalesce_subword_stores(intptr_t header_size, 3157 Node* size_in_bytes, 3158 PhaseGVN* phase) { 3159 Compile* C = phase->C; 3160 3161 assert(stores_are_sane(phase), ""); 3162 // Note: After this pass, they are not completely sane, 3163 // since there may be some overlaps. 3164 3165 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3166 3167 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3168 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3169 size_limit = MIN2(size_limit, ti_limit); 3170 size_limit = align_size_up(size_limit, BytesPerLong); 3171 int num_tiles = size_limit / BytesPerLong; 3172 3173 // allocate space for the tile map: 3174 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3175 jlong tiles_buf[small_len]; 3176 Node* nodes_buf[small_len]; 3177 jlong inits_buf[small_len]; 3178 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3179 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3180 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3181 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3182 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3183 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3184 // tiles: exact bitwise model of all primitive constants 3185 // nodes: last constant-storing node subsumed into the tiles model 3186 // inits: which bytes (in each tile) are touched by any initializations 3187 3188 //// Pass A: Fill in the tile model with any relevant stores. 3189 3190 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3191 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3192 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3193 Node* zmem = zero_memory(); // initially zero memory state 3194 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3195 Node* st = in(i); 3196 intptr_t st_off = get_store_offset(st, phase); 3197 3198 // Figure out the store's offset and constant value: 3199 if (st_off < header_size) continue; //skip (ignore header) 3200 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3201 int st_size = st->as_Store()->memory_size(); 3202 if (st_off + st_size > size_limit) break; 3203 3204 // Record which bytes are touched, whether by constant or not. 3205 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3206 continue; // skip (strange store size) 3207 3208 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3209 if (!val->singleton()) continue; //skip (non-con store) 3210 BasicType type = val->basic_type(); 3211 3212 jlong con = 0; 3213 switch (type) { 3214 case T_INT: con = val->is_int()->get_con(); break; 3215 case T_LONG: con = val->is_long()->get_con(); break; 3216 case T_FLOAT: con = jint_cast(val->getf()); break; 3217 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3218 default: continue; //skip (odd store type) 3219 } 3220 3221 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3222 st->Opcode() == Op_StoreL) { 3223 continue; // This StoreL is already optimal. 3224 } 3225 3226 // Store down the constant. 3227 store_constant(tiles, num_tiles, st_off, st_size, con); 3228 3229 intptr_t j = st_off >> LogBytesPerLong; 3230 3231 if (type == T_INT && st_size == BytesPerInt 3232 && (st_off & BytesPerInt) == BytesPerInt) { 3233 jlong lcon = tiles[j]; 3234 if (!Matcher::isSimpleConstant64(lcon) && 3235 st->Opcode() == Op_StoreI) { 3236 // This StoreI is already optimal by itself. 3237 jint* intcon = (jint*) &tiles[j]; 3238 intcon[1] = 0; // undo the store_constant() 3239 3240 // If the previous store is also optimal by itself, back up and 3241 // undo the action of the previous loop iteration... if we can. 3242 // But if we can't, just let the previous half take care of itself. 3243 st = nodes[j]; 3244 st_off -= BytesPerInt; 3245 con = intcon[0]; 3246 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3247 assert(st_off >= header_size, "still ignoring header"); 3248 assert(get_store_offset(st, phase) == st_off, "must be"); 3249 assert(in(i-1) == zmem, "must be"); 3250 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3251 assert(con == tcon->is_int()->get_con(), "must be"); 3252 // Undo the effects of the previous loop trip, which swallowed st: 3253 intcon[0] = 0; // undo store_constant() 3254 set_req(i-1, st); // undo set_req(i, zmem) 3255 nodes[j] = NULL; // undo nodes[j] = st 3256 --old_subword; // undo ++old_subword 3257 } 3258 continue; // This StoreI is already optimal. 3259 } 3260 } 3261 3262 // This store is not needed. 3263 set_req(i, zmem); 3264 nodes[j] = st; // record for the moment 3265 if (st_size < BytesPerLong) // something has changed 3266 ++old_subword; // includes int/float, but who's counting... 3267 else ++old_long; 3268 } 3269 3270 if ((old_subword + old_long) == 0) 3271 return; // nothing more to do 3272 3273 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3274 // Be sure to insert them before overlapping non-constant stores. 3275 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3276 for (int j = 0; j < num_tiles; j++) { 3277 jlong con = tiles[j]; 3278 jlong init = inits[j]; 3279 if (con == 0) continue; 3280 jint con0, con1; // split the constant, address-wise 3281 jint init0, init1; // split the init map, address-wise 3282 { union { jlong con; jint intcon[2]; } u; 3283 u.con = con; 3284 con0 = u.intcon[0]; 3285 con1 = u.intcon[1]; 3286 u.con = init; 3287 init0 = u.intcon[0]; 3288 init1 = u.intcon[1]; 3289 } 3290 3291 Node* old = nodes[j]; 3292 assert(old != NULL, "need the prior store"); 3293 intptr_t offset = (j * BytesPerLong); 3294 3295 bool split = !Matcher::isSimpleConstant64(con); 3296 3297 if (offset < header_size) { 3298 assert(offset + BytesPerInt >= header_size, "second int counts"); 3299 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3300 split = true; // only the second word counts 3301 // Example: int a[] = { 42 ... } 3302 } else if (con0 == 0 && init0 == -1) { 3303 split = true; // first word is covered by full inits 3304 // Example: int a[] = { ... foo(), 42 ... } 3305 } else if (con1 == 0 && init1 == -1) { 3306 split = true; // second word is covered by full inits 3307 // Example: int a[] = { ... 42, foo() ... } 3308 } 3309 3310 // Here's a case where init0 is neither 0 nor -1: 3311 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3312 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3313 // In this case the tile is not split; it is (jlong)42. 3314 // The big tile is stored down, and then the foo() value is inserted. 3315 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3316 3317 Node* ctl = old->in(MemNode::Control); 3318 Node* adr = make_raw_address(offset, phase); 3319 const TypePtr* atp = TypeRawPtr::BOTTOM; 3320 3321 // One or two coalesced stores to plop down. 3322 Node* st[2]; 3323 intptr_t off[2]; 3324 int nst = 0; 3325 if (!split) { 3326 ++new_long; 3327 off[nst] = offset; 3328 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3329 phase->longcon(con), T_LONG); 3330 } else { 3331 // Omit either if it is a zero. 3332 if (con0 != 0) { 3333 ++new_int; 3334 off[nst] = offset; 3335 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3336 phase->intcon(con0), T_INT); 3337 } 3338 if (con1 != 0) { 3339 ++new_int; 3340 offset += BytesPerInt; 3341 adr = make_raw_address(offset, phase); 3342 off[nst] = offset; 3343 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3344 phase->intcon(con1), T_INT); 3345 } 3346 } 3347 3348 // Insert second store first, then the first before the second. 3349 // Insert each one just before any overlapping non-constant stores. 3350 while (nst > 0) { 3351 Node* st1 = st[--nst]; 3352 C->copy_node_notes_to(st1, old); 3353 st1 = phase->transform(st1); 3354 offset = off[nst]; 3355 assert(offset >= header_size, "do not smash header"); 3356 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3357 guarantee(ins_idx != 0, "must re-insert constant store"); 3358 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3359 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3360 set_req(--ins_idx, st1); 3361 else 3362 ins_req(ins_idx, st1); 3363 } 3364 } 3365 3366 if (PrintCompilation && WizardMode) 3367 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3368 old_subword, old_long, new_int, new_long); 3369 if (C->log() != NULL) 3370 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3371 old_subword, old_long, new_int, new_long); 3372 3373 // Clean up any remaining occurrences of zmem: 3374 remove_extra_zeroes(); 3375} 3376 3377// Explore forward from in(start) to find the first fully initialized 3378// word, and return its offset. Skip groups of subword stores which 3379// together initialize full words. If in(start) is itself part of a 3380// fully initialized word, return the offset of in(start). If there 3381// are no following full-word stores, or if something is fishy, return 3382// a negative value. 3383intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3384 int int_map = 0; 3385 intptr_t int_map_off = 0; 3386 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3387 3388 for (uint i = start, limit = req(); i < limit; i++) { 3389 Node* st = in(i); 3390 3391 intptr_t st_off = get_store_offset(st, phase); 3392 if (st_off < 0) break; // return conservative answer 3393 3394 int st_size = st->as_Store()->memory_size(); 3395 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3396 return st_off; // we found a complete word init 3397 } 3398 3399 // update the map: 3400 3401 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3402 if (this_int_off != int_map_off) { 3403 // reset the map: 3404 int_map = 0; 3405 int_map_off = this_int_off; 3406 } 3407 3408 int subword_off = st_off - this_int_off; 3409 int_map |= right_n_bits(st_size) << subword_off; 3410 if ((int_map & FULL_MAP) == FULL_MAP) { 3411 return this_int_off; // we found a complete word init 3412 } 3413 3414 // Did this store hit or cross the word boundary? 3415 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3416 if (next_int_off == this_int_off + BytesPerInt) { 3417 // We passed the current int, without fully initializing it. 3418 int_map_off = next_int_off; 3419 int_map >>= BytesPerInt; 3420 } else if (next_int_off > this_int_off + BytesPerInt) { 3421 // We passed the current and next int. 3422 return this_int_off + BytesPerInt; 3423 } 3424 } 3425 3426 return -1; 3427} 3428 3429 3430// Called when the associated AllocateNode is expanded into CFG. 3431// At this point, we may perform additional optimizations. 3432// Linearize the stores by ascending offset, to make memory 3433// activity as coherent as possible. 3434Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3435 intptr_t header_size, 3436 Node* size_in_bytes, 3437 PhaseGVN* phase) { 3438 assert(!is_complete(), "not already complete"); 3439 assert(stores_are_sane(phase), ""); 3440 assert(allocation() != NULL, "must be present"); 3441 3442 remove_extra_zeroes(); 3443 3444 if (ReduceFieldZeroing || ReduceBulkZeroing) 3445 // reduce instruction count for common initialization patterns 3446 coalesce_subword_stores(header_size, size_in_bytes, phase); 3447 3448 Node* zmem = zero_memory(); // initially zero memory state 3449 Node* inits = zmem; // accumulating a linearized chain of inits 3450 #ifdef ASSERT 3451 intptr_t first_offset = allocation()->minimum_header_size(); 3452 intptr_t last_init_off = first_offset; // previous init offset 3453 intptr_t last_init_end = first_offset; // previous init offset+size 3454 intptr_t last_tile_end = first_offset; // previous tile offset+size 3455 #endif 3456 intptr_t zeroes_done = header_size; 3457 3458 bool do_zeroing = true; // we might give up if inits are very sparse 3459 int big_init_gaps = 0; // how many large gaps have we seen? 3460 3461 if (ZeroTLAB) do_zeroing = false; 3462 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3463 3464 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3465 Node* st = in(i); 3466 intptr_t st_off = get_store_offset(st, phase); 3467 if (st_off < 0) 3468 break; // unknown junk in the inits 3469 if (st->in(MemNode::Memory) != zmem) 3470 break; // complicated store chains somehow in list 3471 3472 int st_size = st->as_Store()->memory_size(); 3473 intptr_t next_init_off = st_off + st_size; 3474 3475 if (do_zeroing && zeroes_done < next_init_off) { 3476 // See if this store needs a zero before it or under it. 3477 intptr_t zeroes_needed = st_off; 3478 3479 if (st_size < BytesPerInt) { 3480 // Look for subword stores which only partially initialize words. 3481 // If we find some, we must lay down some word-level zeroes first, 3482 // underneath the subword stores. 3483 // 3484 // Examples: 3485 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3486 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3487 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3488 // 3489 // Note: coalesce_subword_stores may have already done this, 3490 // if it was prompted by constant non-zero subword initializers. 3491 // But this case can still arise with non-constant stores. 3492 3493 intptr_t next_full_store = find_next_fullword_store(i, phase); 3494 3495 // In the examples above: 3496 // in(i) p q r s x y z 3497 // st_off 12 13 14 15 12 13 14 3498 // st_size 1 1 1 1 1 1 1 3499 // next_full_s. 12 16 16 16 16 16 16 3500 // z's_done 12 16 16 16 12 16 12 3501 // z's_needed 12 16 16 16 16 16 16 3502 // zsize 0 0 0 0 4 0 4 3503 if (next_full_store < 0) { 3504 // Conservative tack: Zero to end of current word. 3505 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3506 } else { 3507 // Zero to beginning of next fully initialized word. 3508 // Or, don't zero at all, if we are already in that word. 3509 assert(next_full_store >= zeroes_needed, "must go forward"); 3510 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3511 zeroes_needed = next_full_store; 3512 } 3513 } 3514 3515 if (zeroes_needed > zeroes_done) { 3516 intptr_t zsize = zeroes_needed - zeroes_done; 3517 // Do some incremental zeroing on rawmem, in parallel with inits. 3518 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3519 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3520 zeroes_done, zeroes_needed, 3521 phase); 3522 zeroes_done = zeroes_needed; 3523 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3524 do_zeroing = false; // leave the hole, next time 3525 } 3526 } 3527 3528 // Collect the store and move on: 3529 st->set_req(MemNode::Memory, inits); 3530 inits = st; // put it on the linearized chain 3531 set_req(i, zmem); // unhook from previous position 3532 3533 if (zeroes_done == st_off) 3534 zeroes_done = next_init_off; 3535 3536 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3537 3538 #ifdef ASSERT 3539 // Various order invariants. Weaker than stores_are_sane because 3540 // a large constant tile can be filled in by smaller non-constant stores. 3541 assert(st_off >= last_init_off, "inits do not reverse"); 3542 last_init_off = st_off; 3543 const Type* val = NULL; 3544 if (st_size >= BytesPerInt && 3545 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3546 (int)val->basic_type() < (int)T_OBJECT) { 3547 assert(st_off >= last_tile_end, "tiles do not overlap"); 3548 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3549 last_tile_end = MAX2(last_tile_end, next_init_off); 3550 } else { 3551 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3552 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3553 assert(st_off >= last_init_end, "inits do not overlap"); 3554 last_init_end = next_init_off; // it's a non-tile 3555 } 3556 #endif //ASSERT 3557 } 3558 3559 remove_extra_zeroes(); // clear out all the zmems left over 3560 add_req(inits); 3561 3562 if (!ZeroTLAB) { 3563 // If anything remains to be zeroed, zero it all now. 3564 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3565 // if it is the last unused 4 bytes of an instance, forget about it 3566 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3567 if (zeroes_done + BytesPerLong >= size_limit) { 3568 assert(allocation() != NULL, ""); 3569 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3570 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3571 if (zeroes_done == k->layout_helper()) 3572 zeroes_done = size_limit; 3573 } 3574 if (zeroes_done < size_limit) { 3575 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3576 zeroes_done, size_in_bytes, phase); 3577 } 3578 } 3579 3580 set_complete(phase); 3581 return rawmem; 3582} 3583 3584 3585#ifdef ASSERT 3586bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3587 if (is_complete()) 3588 return true; // stores could be anything at this point 3589 assert(allocation() != NULL, "must be present"); 3590 intptr_t last_off = allocation()->minimum_header_size(); 3591 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3592 Node* st = in(i); 3593 intptr_t st_off = get_store_offset(st, phase); 3594 if (st_off < 0) continue; // ignore dead garbage 3595 if (last_off > st_off) { 3596 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3597 this->dump(2); 3598 assert(false, "ascending store offsets"); 3599 return false; 3600 } 3601 last_off = st_off + st->as_Store()->memory_size(); 3602 } 3603 return true; 3604} 3605#endif //ASSERT 3606 3607 3608 3609 3610//============================MergeMemNode===================================== 3611// 3612// SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3613// contributing store or call operations. Each contributor provides the memory 3614// state for a particular "alias type" (see Compile::alias_type). For example, 3615// if a MergeMem has an input X for alias category #6, then any memory reference 3616// to alias category #6 may use X as its memory state input, as an exact equivalent 3617// to using the MergeMem as a whole. 3618// Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3619// 3620// (Here, the <N> notation gives the index of the relevant adr_type.) 3621// 3622// In one special case (and more cases in the future), alias categories overlap. 3623// The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3624// states. Therefore, if a MergeMem has only one contributing input W for Bot, 3625// it is exactly equivalent to that state W: 3626// MergeMem(<Bot>: W) <==> W 3627// 3628// Usually, the merge has more than one input. In that case, where inputs 3629// overlap (i.e., one is Bot), the narrower alias type determines the memory 3630// state for that type, and the wider alias type (Bot) fills in everywhere else: 3631// Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3632// Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3633// 3634// A merge can take a "wide" memory state as one of its narrow inputs. 3635// This simply means that the merge observes out only the relevant parts of 3636// the wide input. That is, wide memory states arriving at narrow merge inputs 3637// are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3638// 3639// These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3640// and that memory slices "leak through": 3641// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3642// 3643// But, in such a cascade, repeated memory slices can "block the leak": 3644// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3645// 3646// In the last example, Y is not part of the combined memory state of the 3647// outermost MergeMem. The system must, of course, prevent unschedulable 3648// memory states from arising, so you can be sure that the state Y is somehow 3649// a precursor to state Y'. 3650// 3651// 3652// REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3653// of each MergeMemNode array are exactly the numerical alias indexes, including 3654// but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3655// Compile::alias_type (and kin) produce and manage these indexes. 3656// 3657// By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3658// (Note that this provides quick access to the top node inside MergeMem methods, 3659// without the need to reach out via TLS to Compile::current.) 3660// 3661// As a consequence of what was just described, a MergeMem that represents a full 3662// memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3663// containing all alias categories. 3664// 3665// MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3666// 3667// All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3668// a memory state for the alias type <N>, or else the top node, meaning that 3669// there is no particular input for that alias type. Note that the length of 3670// a MergeMem is variable, and may be extended at any time to accommodate new 3671// memory states at larger alias indexes. When merges grow, they are of course 3672// filled with "top" in the unused in() positions. 3673// 3674// This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3675// (Top was chosen because it works smoothly with passes like GCM.) 3676// 3677// For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3678// the type of random VM bits like TLS references.) Since it is always the 3679// first non-Bot memory slice, some low-level loops use it to initialize an 3680// index variable: for (i = AliasIdxRaw; i < req(); i++). 3681// 3682// 3683// ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3684// the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3685// the memory state for alias type <N>, or (if there is no particular slice at <N>, 3686// it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3687// or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3688// MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3689// 3690// %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3691// really that different from the other memory inputs. An abbreviation called 3692// "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3693// 3694// 3695// PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3696// partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3697// that "emerges though" the base memory will be marked as excluding the alias types 3698// of the other (narrow-memory) copies which "emerged through" the narrow edges: 3699// 3700// Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3701// ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3702// 3703// This strange "subtraction" effect is necessary to ensure IGVN convergence. 3704// (It is currently unimplemented.) As you can see, the resulting merge is 3705// actually a disjoint union of memory states, rather than an overlay. 3706// 3707 3708//------------------------------MergeMemNode----------------------------------- 3709Node* MergeMemNode::make_empty_memory() { 3710 Node* empty_memory = (Node*) Compile::current()->top(); 3711 assert(empty_memory->is_top(), "correct sentinel identity"); 3712 return empty_memory; 3713} 3714 3715MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3716 init_class_id(Class_MergeMem); 3717 // all inputs are nullified in Node::Node(int) 3718 // set_input(0, NULL); // no control input 3719 3720 // Initialize the edges uniformly to top, for starters. 3721 Node* empty_mem = make_empty_memory(); 3722 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3723 init_req(i,empty_mem); 3724 } 3725 assert(empty_memory() == empty_mem, ""); 3726 3727 if( new_base != NULL && new_base->is_MergeMem() ) { 3728 MergeMemNode* mdef = new_base->as_MergeMem(); 3729 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3730 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3731 mms.set_memory(mms.memory2()); 3732 } 3733 assert(base_memory() == mdef->base_memory(), ""); 3734 } else { 3735 set_base_memory(new_base); 3736 } 3737} 3738 3739// Make a new, untransformed MergeMem with the same base as 'mem'. 3740// If mem is itself a MergeMem, populate the result with the same edges. 3741MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3742 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3743} 3744 3745//------------------------------cmp-------------------------------------------- 3746uint MergeMemNode::hash() const { return NO_HASH; } 3747uint MergeMemNode::cmp( const Node &n ) const { 3748 return (&n == this); // Always fail except on self 3749} 3750 3751//------------------------------Identity--------------------------------------- 3752Node* MergeMemNode::Identity(PhaseTransform *phase) { 3753 // Identity if this merge point does not record any interesting memory 3754 // disambiguations. 3755 Node* base_mem = base_memory(); 3756 Node* empty_mem = empty_memory(); 3757 if (base_mem != empty_mem) { // Memory path is not dead? 3758 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3759 Node* mem = in(i); 3760 if (mem != empty_mem && mem != base_mem) { 3761 return this; // Many memory splits; no change 3762 } 3763 } 3764 } 3765 return base_mem; // No memory splits; ID on the one true input 3766} 3767 3768//------------------------------Ideal------------------------------------------ 3769// This method is invoked recursively on chains of MergeMem nodes 3770Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3771 // Remove chain'd MergeMems 3772 // 3773 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3774 // relative to the "in(Bot)". Since we are patching both at the same time, 3775 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3776 // but rewrite each "in(i)" relative to the new "in(Bot)". 3777 Node *progress = NULL; 3778 3779 3780 Node* old_base = base_memory(); 3781 Node* empty_mem = empty_memory(); 3782 if (old_base == empty_mem) 3783 return NULL; // Dead memory path. 3784 3785 MergeMemNode* old_mbase; 3786 if (old_base != NULL && old_base->is_MergeMem()) 3787 old_mbase = old_base->as_MergeMem(); 3788 else 3789 old_mbase = NULL; 3790 Node* new_base = old_base; 3791 3792 // simplify stacked MergeMems in base memory 3793 if (old_mbase) new_base = old_mbase->base_memory(); 3794 3795 // the base memory might contribute new slices beyond my req() 3796 if (old_mbase) grow_to_match(old_mbase); 3797 3798 // Look carefully at the base node if it is a phi. 3799 PhiNode* phi_base; 3800 if (new_base != NULL && new_base->is_Phi()) 3801 phi_base = new_base->as_Phi(); 3802 else 3803 phi_base = NULL; 3804 3805 Node* phi_reg = NULL; 3806 uint phi_len = (uint)-1; 3807 if (phi_base != NULL && !phi_base->is_copy()) { 3808 // do not examine phi if degraded to a copy 3809 phi_reg = phi_base->region(); 3810 phi_len = phi_base->req(); 3811 // see if the phi is unfinished 3812 for (uint i = 1; i < phi_len; i++) { 3813 if (phi_base->in(i) == NULL) { 3814 // incomplete phi; do not look at it yet! 3815 phi_reg = NULL; 3816 phi_len = (uint)-1; 3817 break; 3818 } 3819 } 3820 } 3821 3822 // Note: We do not call verify_sparse on entry, because inputs 3823 // can normalize to the base_memory via subsume_node or similar 3824 // mechanisms. This method repairs that damage. 3825 3826 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3827 3828 // Look at each slice. 3829 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3830 Node* old_in = in(i); 3831 // calculate the old memory value 3832 Node* old_mem = old_in; 3833 if (old_mem == empty_mem) old_mem = old_base; 3834 assert(old_mem == memory_at(i), ""); 3835 3836 // maybe update (reslice) the old memory value 3837 3838 // simplify stacked MergeMems 3839 Node* new_mem = old_mem; 3840 MergeMemNode* old_mmem; 3841 if (old_mem != NULL && old_mem->is_MergeMem()) 3842 old_mmem = old_mem->as_MergeMem(); 3843 else 3844 old_mmem = NULL; 3845 if (old_mmem == this) { 3846 // This can happen if loops break up and safepoints disappear. 3847 // A merge of BotPtr (default) with a RawPtr memory derived from a 3848 // safepoint can be rewritten to a merge of the same BotPtr with 3849 // the BotPtr phi coming into the loop. If that phi disappears 3850 // also, we can end up with a self-loop of the mergemem. 3851 // In general, if loops degenerate and memory effects disappear, 3852 // a mergemem can be left looking at itself. This simply means 3853 // that the mergemem's default should be used, since there is 3854 // no longer any apparent effect on this slice. 3855 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3856 // from start. Update the input to TOP. 3857 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3858 } 3859 else if (old_mmem != NULL) { 3860 new_mem = old_mmem->memory_at(i); 3861 } 3862 // else preceding memory was not a MergeMem 3863 3864 // replace equivalent phis (unfortunately, they do not GVN together) 3865 if (new_mem != NULL && new_mem != new_base && 3866 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3867 if (new_mem->is_Phi()) { 3868 PhiNode* phi_mem = new_mem->as_Phi(); 3869 for (uint i = 1; i < phi_len; i++) { 3870 if (phi_base->in(i) != phi_mem->in(i)) { 3871 phi_mem = NULL; 3872 break; 3873 } 3874 } 3875 if (phi_mem != NULL) { 3876 // equivalent phi nodes; revert to the def 3877 new_mem = new_base; 3878 } 3879 } 3880 } 3881 3882 // maybe store down a new value 3883 Node* new_in = new_mem; 3884 if (new_in == new_base) new_in = empty_mem; 3885 3886 if (new_in != old_in) { 3887 // Warning: Do not combine this "if" with the previous "if" 3888 // A memory slice might have be be rewritten even if it is semantically 3889 // unchanged, if the base_memory value has changed. 3890 set_req(i, new_in); 3891 progress = this; // Report progress 3892 } 3893 } 3894 3895 if (new_base != old_base) { 3896 set_req(Compile::AliasIdxBot, new_base); 3897 // Don't use set_base_memory(new_base), because we need to update du. 3898 assert(base_memory() == new_base, ""); 3899 progress = this; 3900 } 3901 3902 if( base_memory() == this ) { 3903 // a self cycle indicates this memory path is dead 3904 set_req(Compile::AliasIdxBot, empty_mem); 3905 } 3906 3907 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3908 // Recursion must occur after the self cycle check above 3909 if( base_memory()->is_MergeMem() ) { 3910 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3911 Node *m = phase->transform(new_mbase); // Rollup any cycles 3912 if( m != NULL && (m->is_top() || 3913 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3914 // propagate rollup of dead cycle to self 3915 set_req(Compile::AliasIdxBot, empty_mem); 3916 } 3917 } 3918 3919 if( base_memory() == empty_mem ) { 3920 progress = this; 3921 // Cut inputs during Parse phase only. 3922 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3923 if( !can_reshape ) { 3924 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3925 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3926 } 3927 } 3928 } 3929 3930 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3931 // Check if PhiNode::Ideal's "Split phis through memory merges" 3932 // transform should be attempted. Look for this->phi->this cycle. 3933 uint merge_width = req(); 3934 if (merge_width > Compile::AliasIdxRaw) { 3935 PhiNode* phi = base_memory()->as_Phi(); 3936 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3937 if (phi->in(i) == this) { 3938 phase->is_IterGVN()->_worklist.push(phi); 3939 break; 3940 } 3941 } 3942 } 3943 } 3944 3945 assert(progress || verify_sparse(), "please, no dups of base"); 3946 return progress; 3947} 3948 3949//-------------------------set_base_memory------------------------------------- 3950void MergeMemNode::set_base_memory(Node *new_base) { 3951 Node* empty_mem = empty_memory(); 3952 set_req(Compile::AliasIdxBot, new_base); 3953 assert(memory_at(req()) == new_base, "must set default memory"); 3954 // Clear out other occurrences of new_base: 3955 if (new_base != empty_mem) { 3956 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3957 if (in(i) == new_base) set_req(i, empty_mem); 3958 } 3959 } 3960} 3961 3962//------------------------------out_RegMask------------------------------------ 3963const RegMask &MergeMemNode::out_RegMask() const { 3964 return RegMask::Empty; 3965} 3966 3967//------------------------------dump_spec-------------------------------------- 3968#ifndef PRODUCT 3969void MergeMemNode::dump_spec(outputStream *st) const { 3970 st->print(" {"); 3971 Node* base_mem = base_memory(); 3972 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3973 Node* mem = memory_at(i); 3974 if (mem == base_mem) { st->print(" -"); continue; } 3975 st->print( " N%d:", mem->_idx ); 3976 Compile::current()->get_adr_type(i)->dump_on(st); 3977 } 3978 st->print(" }"); 3979} 3980#endif // !PRODUCT 3981 3982 3983#ifdef ASSERT 3984static bool might_be_same(Node* a, Node* b) { 3985 if (a == b) return true; 3986 if (!(a->is_Phi() || b->is_Phi())) return false; 3987 // phis shift around during optimization 3988 return true; // pretty stupid... 3989} 3990 3991// verify a narrow slice (either incoming or outgoing) 3992static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 3993 if (!VerifyAliases) return; // don't bother to verify unless requested 3994 if (is_error_reported()) return; // muzzle asserts when debugging an error 3995 if (Node::in_dump()) return; // muzzle asserts when printing 3996 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 3997 assert(n != NULL, ""); 3998 // Elide intervening MergeMem's 3999 while (n->is_MergeMem()) { 4000 n = n->as_MergeMem()->memory_at(alias_idx); 4001 } 4002 Compile* C = Compile::current(); 4003 const TypePtr* n_adr_type = n->adr_type(); 4004 if (n == m->empty_memory()) { 4005 // Implicit copy of base_memory() 4006 } else if (n_adr_type != TypePtr::BOTTOM) { 4007 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4008 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4009 } else { 4010 // A few places like make_runtime_call "know" that VM calls are narrow, 4011 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4012 bool expected_wide_mem = false; 4013 if (n == m->base_memory()) { 4014 expected_wide_mem = true; 4015 } else if (alias_idx == Compile::AliasIdxRaw || 4016 n == m->memory_at(Compile::AliasIdxRaw)) { 4017 expected_wide_mem = true; 4018 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4019 // memory can "leak through" calls on channels that 4020 // are write-once. Allow this also. 4021 expected_wide_mem = true; 4022 } 4023 assert(expected_wide_mem, "expected narrow slice replacement"); 4024 } 4025} 4026#else // !ASSERT 4027#define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 4028#endif 4029 4030 4031//-----------------------------memory_at--------------------------------------- 4032Node* MergeMemNode::memory_at(uint alias_idx) const { 4033 assert(alias_idx >= Compile::AliasIdxRaw || 4034 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4035 "must avoid base_memory and AliasIdxTop"); 4036 4037 // Otherwise, it is a narrow slice. 4038 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4039 Compile *C = Compile::current(); 4040 if (is_empty_memory(n)) { 4041 // the array is sparse; empty slots are the "top" node 4042 n = base_memory(); 4043 assert(Node::in_dump() 4044 || n == NULL || n->bottom_type() == Type::TOP 4045 || n->adr_type() == TypePtr::BOTTOM 4046 || n->adr_type() == TypeRawPtr::BOTTOM 4047 || Compile::current()->AliasLevel() == 0, 4048 "must be a wide memory"); 4049 // AliasLevel == 0 if we are organizing the memory states manually. 4050 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4051 } else { 4052 // make sure the stored slice is sane 4053 #ifdef ASSERT 4054 if (is_error_reported() || Node::in_dump()) { 4055 } else if (might_be_same(n, base_memory())) { 4056 // Give it a pass: It is a mostly harmless repetition of the base. 4057 // This can arise normally from node subsumption during optimization. 4058 } else { 4059 verify_memory_slice(this, alias_idx, n); 4060 } 4061 #endif 4062 } 4063 return n; 4064} 4065 4066//---------------------------set_memory_at------------------------------------- 4067void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4068 verify_memory_slice(this, alias_idx, n); 4069 Node* empty_mem = empty_memory(); 4070 if (n == base_memory()) n = empty_mem; // collapse default 4071 uint need_req = alias_idx+1; 4072 if (req() < need_req) { 4073 if (n == empty_mem) return; // already the default, so do not grow me 4074 // grow the sparse array 4075 do { 4076 add_req(empty_mem); 4077 } while (req() < need_req); 4078 } 4079 set_req( alias_idx, n ); 4080} 4081 4082 4083 4084//--------------------------iteration_setup------------------------------------ 4085void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4086 if (other != NULL) { 4087 grow_to_match(other); 4088 // invariant: the finite support of mm2 is within mm->req() 4089 #ifdef ASSERT 4090 for (uint i = req(); i < other->req(); i++) { 4091 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4092 } 4093 #endif 4094 } 4095 // Replace spurious copies of base_memory by top. 4096 Node* base_mem = base_memory(); 4097 if (base_mem != NULL && !base_mem->is_top()) { 4098 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4099 if (in(i) == base_mem) 4100 set_req(i, empty_memory()); 4101 } 4102 } 4103} 4104 4105//---------------------------grow_to_match------------------------------------- 4106void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4107 Node* empty_mem = empty_memory(); 4108 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4109 // look for the finite support of the other memory 4110 for (uint i = other->req(); --i >= req(); ) { 4111 if (other->in(i) != empty_mem) { 4112 uint new_len = i+1; 4113 while (req() < new_len) add_req(empty_mem); 4114 break; 4115 } 4116 } 4117} 4118 4119//---------------------------verify_sparse------------------------------------- 4120#ifndef PRODUCT 4121bool MergeMemNode::verify_sparse() const { 4122 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4123 Node* base_mem = base_memory(); 4124 // The following can happen in degenerate cases, since empty==top. 4125 if (is_empty_memory(base_mem)) return true; 4126 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4127 assert(in(i) != NULL, "sane slice"); 4128 if (in(i) == base_mem) return false; // should have been the sentinel value! 4129 } 4130 return true; 4131} 4132 4133bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4134 Node* n; 4135 n = mm->in(idx); 4136 if (mem == n) return true; // might be empty_memory() 4137 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4138 if (mem == n) return true; 4139 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4140 if (mem == n) return true; 4141 if (n == NULL) break; 4142 } 4143 return false; 4144} 4145#endif // !PRODUCT 4146