1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "oops/objArrayKlass.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/arraycopynode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/compile.hpp" 34 #include "opto/connode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/machnode.hpp" 38 #include "opto/matcher.hpp" 39 #include "opto/memnode.hpp" 40 #include "opto/mulnode.hpp" 41 #include "opto/narrowptrnode.hpp" 42 #include "opto/phaseX.hpp" 43 #include "opto/regmask.hpp" 44 #include "utilities/copy.hpp" 45 46 // Portions of code courtesy of Clifford Click 47 48 // Optimization - Graph Style 49 50 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 51 52 //============================================================================= 53 uint MemNode::size_of() const { return sizeof(*this); } 54 55 const TypePtr *MemNode::adr_type() const { 56 Node* adr = in(Address); 57 if (adr == NULL) return NULL; // node is dead 58 const TypePtr* cross_check = NULL; 59 DEBUG_ONLY(cross_check = _adr_type); 60 return calculate_adr_type(adr->bottom_type(), cross_check); 61 } 62 63 #ifndef PRODUCT 64 void MemNode::dump_spec(outputStream *st) const { 65 if (in(Address) == NULL) return; // node is dead 66 #ifndef ASSERT 67 // fake the missing field 68 const TypePtr* _adr_type = NULL; 69 if (in(Address) != NULL) 70 _adr_type = in(Address)->bottom_type()->isa_ptr(); 71 #endif 72 dump_adr_type(this, _adr_type, st); 73 74 Compile* C = Compile::current(); 75 if( C->alias_type(_adr_type)->is_volatile() ) 76 st->print(" Volatile!"); 77 } 78 79 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 80 st->print(" @"); 81 if (adr_type == NULL) { 82 st->print("NULL"); 83 } else { 84 adr_type->dump_on(st); 85 Compile* C = Compile::current(); 86 Compile::AliasType* atp = NULL; 87 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 88 if (atp == NULL) 89 st->print(", idx=?\?;"); 90 else if (atp->index() == Compile::AliasIdxBot) 91 st->print(", idx=Bot;"); 92 else if (atp->index() == Compile::AliasIdxTop) 93 st->print(", idx=Top;"); 94 else if (atp->index() == Compile::AliasIdxRaw) 95 st->print(", idx=Raw;"); 96 else { 97 ciField* field = atp->field(); 98 if (field) { 99 st->print(", name="); 100 field->print_name_on(st); 101 } 102 st->print(", idx=%d;", atp->index()); 103 } 104 } 105 } 106 107 extern void print_alias_types(); 108 109 #endif 110 111 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 112 assert((t_oop != NULL), "sanity"); 113 bool is_instance = t_oop->is_known_instance_field(); 114 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 115 (load != NULL) && load->is_Load() && 116 (phase->is_IterGVN() != NULL); 117 if (!(is_instance || is_boxed_value_load)) 118 return mchain; // don't try to optimize non-instance types 119 uint instance_id = t_oop->instance_id(); 120 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 121 Node *prev = NULL; 122 Node *result = mchain; 123 while (prev != result) { 124 prev = result; 125 if (result == start_mem) 126 break; // hit one of our sentinels 127 // skip over a call which does not affect this memory slice 128 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 129 Node *proj_in = result->in(0); 130 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 131 break; // hit one of our sentinels 132 } else if (proj_in->is_Call()) { 133 // ArrayCopyNodes processed here as well 134 CallNode *call = proj_in->as_Call(); 135 if (!call->may_modify(t_oop, phase)) { // returns false for instances 136 result = call->in(TypeFunc::Memory); 137 } 138 } else if (proj_in->is_Initialize()) { 139 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 140 // Stop if this is the initialization for the object instance which 141 // contains this memory slice, otherwise skip over it. 142 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 143 break; 144 } 145 if (is_instance) { 146 result = proj_in->in(TypeFunc::Memory); 147 } else if (is_boxed_value_load) { 148 Node* klass = alloc->in(AllocateNode::KlassNode); 149 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 150 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 151 result = proj_in->in(TypeFunc::Memory); // not related allocation 152 } 153 } 154 } else if (proj_in->is_MemBar()) { 155 if (ArrayCopyNode::membar_for_arraycopy(t_oop, proj_in->as_MemBar(), phase)) { 156 break; 157 } 158 result = proj_in->in(TypeFunc::Memory); 159 } else { 160 assert(false, "unexpected projection"); 161 } 162 } else if (result->is_ClearArray()) { 163 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 164 // Can not bypass initialization of the instance 165 // we are looking for. 166 break; 167 } 168 // Otherwise skip it (the call updated 'result' value). 169 } else if (result->is_MergeMem()) { 170 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 171 } 172 } 173 return result; 174 } 175 176 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 177 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 178 if (t_oop == NULL) 179 return mchain; // don't try to optimize non-oop types 180 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 181 bool is_instance = t_oop->is_known_instance_field(); 182 PhaseIterGVN *igvn = phase->is_IterGVN(); 183 if (is_instance && igvn != NULL && result->is_Phi()) { 184 PhiNode *mphi = result->as_Phi(); 185 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 186 const TypePtr *t = mphi->adr_type(); 187 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 188 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 189 t->is_oopptr()->cast_to_exactness(true) 190 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 191 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 192 // clone the Phi with our address type 193 result = mphi->split_out_instance(t_adr, igvn); 194 } else { 195 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 196 } 197 } 198 return result; 199 } 200 201 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 202 uint alias_idx = phase->C->get_alias_index(tp); 203 Node *mem = mmem; 204 #ifdef ASSERT 205 { 206 // Check that current type is consistent with the alias index used during graph construction 207 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 208 bool consistent = adr_check == NULL || adr_check->empty() || 209 phase->C->must_alias(adr_check, alias_idx ); 210 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 211 if( !consistent && adr_check != NULL && !adr_check->empty() && 212 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 213 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 214 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 215 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 216 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 217 // don't assert if it is dead code. 218 consistent = true; 219 } 220 if( !consistent ) { 221 st->print("alias_idx==%d, adr_check==", alias_idx); 222 if( adr_check == NULL ) { 223 st->print("NULL"); 224 } else { 225 adr_check->dump(); 226 } 227 st->cr(); 228 print_alias_types(); 229 assert(consistent, "adr_check must match alias idx"); 230 } 231 } 232 #endif 233 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 234 // means an array I have not precisely typed yet. Do not do any 235 // alias stuff with it any time soon. 236 const TypeOopPtr *toop = tp->isa_oopptr(); 237 if( tp->base() != Type::AnyPtr && 238 !(toop && 239 toop->klass() != NULL && 240 toop->klass()->is_java_lang_Object() && 241 toop->offset() == Type::OffsetBot) ) { 242 // compress paths and change unreachable cycles to TOP 243 // If not, we can update the input infinitely along a MergeMem cycle 244 // Equivalent code in PhiNode::Ideal 245 Node* m = phase->transform(mmem); 246 // If transformed to a MergeMem, get the desired slice 247 // Otherwise the returned node represents memory for every slice 248 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 249 // Update input if it is progress over what we have now 250 } 251 return mem; 252 } 253 254 //--------------------------Ideal_common--------------------------------------- 255 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 256 // Unhook non-raw memories from complete (macro-expanded) initializations. 257 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 258 // If our control input is a dead region, kill all below the region 259 Node *ctl = in(MemNode::Control); 260 if (ctl && remove_dead_region(phase, can_reshape)) 261 return this; 262 ctl = in(MemNode::Control); 263 // Don't bother trying to transform a dead node 264 if (ctl && ctl->is_top()) return NodeSentinel; 265 266 PhaseIterGVN *igvn = phase->is_IterGVN(); 267 // Wait if control on the worklist. 268 if (ctl && can_reshape && igvn != NULL) { 269 Node* bol = NULL; 270 Node* cmp = NULL; 271 if (ctl->in(0)->is_If()) { 272 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 273 bol = ctl->in(0)->in(1); 274 if (bol->is_Bool()) 275 cmp = ctl->in(0)->in(1)->in(1); 276 } 277 if (igvn->_worklist.member(ctl) || 278 (bol != NULL && igvn->_worklist.member(bol)) || 279 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 280 // This control path may be dead. 281 // Delay this memory node transformation until the control is processed. 282 phase->is_IterGVN()->_worklist.push(this); 283 return NodeSentinel; // caller will return NULL 284 } 285 } 286 // Ignore if memory is dead, or self-loop 287 Node *mem = in(MemNode::Memory); 288 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 289 assert(mem != this, "dead loop in MemNode::Ideal"); 290 291 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 292 // This memory slice may be dead. 293 // Delay this mem node transformation until the memory is processed. 294 phase->is_IterGVN()->_worklist.push(this); 295 return NodeSentinel; // caller will return NULL 296 } 297 298 Node *address = in(MemNode::Address); 299 const Type *t_adr = phase->type(address); 300 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 301 302 if (can_reshape && igvn != NULL && 303 (igvn->_worklist.member(address) || 304 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { 305 // The address's base and type may change when the address is processed. 306 // Delay this mem node transformation until the address is processed. 307 phase->is_IterGVN()->_worklist.push(this); 308 return NodeSentinel; // caller will return NULL 309 } 310 311 // Do NOT remove or optimize the next lines: ensure a new alias index 312 // is allocated for an oop pointer type before Escape Analysis. 313 // Note: C++ will not remove it since the call has side effect. 314 if (t_adr->isa_oopptr()) { 315 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 316 } 317 318 Node* base = NULL; 319 if (address->is_AddP()) { 320 base = address->in(AddPNode::Base); 321 } 322 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 323 !t_adr->isa_rawptr()) { 324 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 325 // Skip this node optimization if its address has TOP base. 326 return NodeSentinel; // caller will return NULL 327 } 328 329 // Avoid independent memory operations 330 Node* old_mem = mem; 331 332 // The code which unhooks non-raw memories from complete (macro-expanded) 333 // initializations was removed. After macro-expansion all stores catched 334 // by Initialize node became raw stores and there is no information 335 // which memory slices they modify. So it is unsafe to move any memory 336 // operation above these stores. Also in most cases hooked non-raw memories 337 // were already unhooked by using information from detect_ptr_independence() 338 // and find_previous_store(). 339 340 if (mem->is_MergeMem()) { 341 MergeMemNode* mmem = mem->as_MergeMem(); 342 const TypePtr *tp = t_adr->is_ptr(); 343 344 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 345 } 346 347 if (mem != old_mem) { 348 set_req(MemNode::Memory, mem); 349 if (can_reshape && old_mem->outcnt() == 0) { 350 igvn->_worklist.push(old_mem); 351 } 352 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 353 return this; 354 } 355 356 // let the subclass continue analyzing... 357 return NULL; 358 } 359 360 // Helper function for proving some simple control dominations. 361 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 362 // Already assumes that 'dom' is available at 'sub', and that 'sub' 363 // is not a constant (dominated by the method's StartNode). 364 // Used by MemNode::find_previous_store to prove that the 365 // control input of a memory operation predates (dominates) 366 // an allocation it wants to look past. 367 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 368 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 369 return false; // Conservative answer for dead code 370 371 // Check 'dom'. Skip Proj and CatchProj nodes. 372 dom = dom->find_exact_control(dom); 373 if (dom == NULL || dom->is_top()) 374 return false; // Conservative answer for dead code 375 376 if (dom == sub) { 377 // For the case when, for example, 'sub' is Initialize and the original 378 // 'dom' is Proj node of the 'sub'. 379 return false; 380 } 381 382 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 383 return true; 384 385 // 'dom' dominates 'sub' if its control edge and control edges 386 // of all its inputs dominate or equal to sub's control edge. 387 388 // Currently 'sub' is either Allocate, Initialize or Start nodes. 389 // Or Region for the check in LoadNode::Ideal(); 390 // 'sub' should have sub->in(0) != NULL. 391 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 392 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 393 394 // Get control edge of 'sub'. 395 Node* orig_sub = sub; 396 sub = sub->find_exact_control(sub->in(0)); 397 if (sub == NULL || sub->is_top()) 398 return false; // Conservative answer for dead code 399 400 assert(sub->is_CFG(), "expecting control"); 401 402 if (sub == dom) 403 return true; 404 405 if (sub->is_Start() || sub->is_Root()) 406 return false; 407 408 { 409 // Check all control edges of 'dom'. 410 411 ResourceMark rm; 412 Arena* arena = Thread::current()->resource_area(); 413 Node_List nlist(arena); 414 Unique_Node_List dom_list(arena); 415 416 dom_list.push(dom); 417 bool only_dominating_controls = false; 418 419 for (uint next = 0; next < dom_list.size(); next++) { 420 Node* n = dom_list.at(next); 421 if (n == orig_sub) 422 return false; // One of dom's inputs dominated by sub. 423 if (!n->is_CFG() && n->pinned()) { 424 // Check only own control edge for pinned non-control nodes. 425 n = n->find_exact_control(n->in(0)); 426 if (n == NULL || n->is_top()) 427 return false; // Conservative answer for dead code 428 assert(n->is_CFG(), "expecting control"); 429 dom_list.push(n); 430 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 431 only_dominating_controls = true; 432 } else if (n->is_CFG()) { 433 if (n->dominates(sub, nlist)) 434 only_dominating_controls = true; 435 else 436 return false; 437 } else { 438 // First, own control edge. 439 Node* m = n->find_exact_control(n->in(0)); 440 if (m != NULL) { 441 if (m->is_top()) 442 return false; // Conservative answer for dead code 443 dom_list.push(m); 444 } 445 // Now, the rest of edges. 446 uint cnt = n->req(); 447 for (uint i = 1; i < cnt; i++) { 448 m = n->find_exact_control(n->in(i)); 449 if (m == NULL || m->is_top()) 450 continue; 451 dom_list.push(m); 452 } 453 } 454 } 455 return only_dominating_controls; 456 } 457 } 458 459 //---------------------detect_ptr_independence--------------------------------- 460 // Used by MemNode::find_previous_store to prove that two base 461 // pointers are never equal. 462 // The pointers are accompanied by their associated allocations, 463 // if any, which have been previously discovered by the caller. 464 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 465 Node* p2, AllocateNode* a2, 466 PhaseTransform* phase) { 467 // Attempt to prove that these two pointers cannot be aliased. 468 // They may both manifestly be allocations, and they should differ. 469 // Or, if they are not both allocations, they can be distinct constants. 470 // Otherwise, one is an allocation and the other a pre-existing value. 471 if (a1 == NULL && a2 == NULL) { // neither an allocation 472 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 473 } else if (a1 != NULL && a2 != NULL) { // both allocations 474 return (a1 != a2); 475 } else if (a1 != NULL) { // one allocation a1 476 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 477 return all_controls_dominate(p2, a1); 478 } else { //(a2 != NULL) // one allocation a2 479 return all_controls_dominate(p1, a2); 480 } 481 return false; 482 } 483 484 485 // Find an arraycopy that must have set (can_see_stored_value=true) or 486 // could have set (can_see_stored_value=false) the value for this load 487 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 488 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 489 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 490 Node* mb = mem->in(0); 491 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 492 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 493 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 494 if (ac->is_clonebasic()) { 495 intptr_t offset; 496 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 497 assert(alloc != NULL && alloc->initialization()->is_complete_with_arraycopy(), "broken allocation"); 498 if (alloc == ld_alloc) { 499 return ac; 500 } 501 } 502 } 503 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 504 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 505 506 if (ac->is_arraycopy_validated() || 507 ac->is_copyof_validated() || 508 ac->is_copyofrange_validated()) { 509 Node* ld_addp = in(MemNode::Address); 510 if (ld_addp->is_AddP()) { 511 Node* ld_base = ld_addp->in(AddPNode::Address); 512 Node* ld_offs = ld_addp->in(AddPNode::Offset); 513 514 Node* dest = ac->in(ArrayCopyNode::Dest); 515 516 if (dest == ld_base) { 517 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 518 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 519 return ac; 520 } 521 if (!can_see_stored_value) { 522 mem = ac->in(TypeFunc::Memory); 523 } 524 } 525 } 526 } 527 } 528 return NULL; 529 } 530 531 // The logic for reordering loads and stores uses four steps: 532 // (a) Walk carefully past stores and initializations which we 533 // can prove are independent of this load. 534 // (b) Observe that the next memory state makes an exact match 535 // with self (load or store), and locate the relevant store. 536 // (c) Ensure that, if we were to wire self directly to the store, 537 // the optimizer would fold it up somehow. 538 // (d) Do the rewiring, and return, depending on some other part of 539 // the optimizer to fold up the load. 540 // This routine handles steps (a) and (b). Steps (c) and (d) are 541 // specific to loads and stores, so they are handled by the callers. 542 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 543 // 544 Node* MemNode::find_previous_store(PhaseTransform* phase) { 545 Node* ctrl = in(MemNode::Control); 546 Node* adr = in(MemNode::Address); 547 intptr_t offset = 0; 548 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 549 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 550 551 if (offset == Type::OffsetBot) 552 return NULL; // cannot unalias unless there are precise offsets 553 554 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 555 556 intptr_t size_in_bytes = memory_size(); 557 558 Node* mem = in(MemNode::Memory); // start searching here... 559 560 int cnt = 50; // Cycle limiter 561 for (;;) { // While we can dance past unrelated stores... 562 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 563 564 Node* prev = mem; 565 if (mem->is_Store()) { 566 Node* st_adr = mem->in(MemNode::Address); 567 intptr_t st_offset = 0; 568 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 569 if (st_base == NULL) 570 break; // inscrutable pointer 571 if (st_offset != offset && st_offset != Type::OffsetBot) { 572 const int MAX_STORE = BytesPerLong; 573 if (st_offset >= offset + size_in_bytes || 574 st_offset <= offset - MAX_STORE || 575 st_offset <= offset - mem->as_Store()->memory_size()) { 576 // Success: The offsets are provably independent. 577 // (You may ask, why not just test st_offset != offset and be done? 578 // The answer is that stores of different sizes can co-exist 579 // in the same sequence of RawMem effects. We sometimes initialize 580 // a whole 'tile' of array elements with a single jint or jlong.) 581 mem = mem->in(MemNode::Memory); 582 continue; // (a) advance through independent store memory 583 } 584 } 585 if (st_base != base && 586 detect_ptr_independence(base, alloc, 587 st_base, 588 AllocateNode::Ideal_allocation(st_base, phase), 589 phase)) { 590 // Success: The bases are provably independent. 591 mem = mem->in(MemNode::Memory); 592 continue; // (a) advance through independent store memory 593 } 594 595 // (b) At this point, if the bases or offsets do not agree, we lose, 596 // since we have not managed to prove 'this' and 'mem' independent. 597 if (st_base == base && st_offset == offset) { 598 return mem; // let caller handle steps (c), (d) 599 } 600 601 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 602 InitializeNode* st_init = mem->in(0)->as_Initialize(); 603 AllocateNode* st_alloc = st_init->allocation(); 604 if (st_alloc == NULL) 605 break; // something degenerated 606 bool known_identical = false; 607 bool known_independent = false; 608 if (alloc == st_alloc) 609 known_identical = true; 610 else if (alloc != NULL) 611 known_independent = true; 612 else if (all_controls_dominate(this, st_alloc)) 613 known_independent = true; 614 615 if (known_independent) { 616 // The bases are provably independent: Either they are 617 // manifestly distinct allocations, or else the control 618 // of this load dominates the store's allocation. 619 int alias_idx = phase->C->get_alias_index(adr_type()); 620 if (alias_idx == Compile::AliasIdxRaw) { 621 mem = st_alloc->in(TypeFunc::Memory); 622 } else { 623 mem = st_init->memory(alias_idx); 624 } 625 continue; // (a) advance through independent store memory 626 } 627 628 // (b) at this point, if we are not looking at a store initializing 629 // the same allocation we are loading from, we lose. 630 if (known_identical) { 631 // From caller, can_see_stored_value will consult find_captured_store. 632 return mem; // let caller handle steps (c), (d) 633 } 634 635 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 636 if (prev != mem) { 637 // Found an arraycopy but it doesn't affect that load 638 continue; 639 } 640 // Found an arraycopy that may affect that load 641 return mem; 642 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 643 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 644 if (mem->is_Proj() && mem->in(0)->is_Call()) { 645 // ArrayCopyNodes processed here as well. 646 CallNode *call = mem->in(0)->as_Call(); 647 if (!call->may_modify(addr_t, phase)) { 648 mem = call->in(TypeFunc::Memory); 649 continue; // (a) advance through independent call memory 650 } 651 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 652 if (ArrayCopyNode::membar_for_arraycopy(addr_t, mem->in(0)->as_MemBar(), phase)) { 653 break; 654 } 655 mem = mem->in(0)->in(TypeFunc::Memory); 656 continue; // (a) advance through independent MemBar memory 657 } else if (mem->is_ClearArray()) { 658 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 659 // (the call updated 'mem' value) 660 continue; // (a) advance through independent allocation memory 661 } else { 662 // Can not bypass initialization of the instance 663 // we are looking for. 664 return mem; 665 } 666 } else if (mem->is_MergeMem()) { 667 int alias_idx = phase->C->get_alias_index(adr_type()); 668 mem = mem->as_MergeMem()->memory_at(alias_idx); 669 continue; // (a) advance through independent MergeMem memory 670 } 671 } 672 673 // Unless there is an explicit 'continue', we must bail out here, 674 // because 'mem' is an inscrutable memory state (e.g., a call). 675 break; 676 } 677 678 return NULL; // bail out 679 } 680 681 //----------------------calculate_adr_type------------------------------------- 682 // Helper function. Notices when the given type of address hits top or bottom. 683 // Also, asserts a cross-check of the type against the expected address type. 684 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 685 if (t == Type::TOP) return NULL; // does not touch memory any more? 686 #ifdef PRODUCT 687 cross_check = NULL; 688 #else 689 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 690 #endif 691 const TypePtr* tp = t->isa_ptr(); 692 if (tp == NULL) { 693 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 694 return TypePtr::BOTTOM; // touches lots of memory 695 } else { 696 #ifdef ASSERT 697 // %%%% [phh] We don't check the alias index if cross_check is 698 // TypeRawPtr::BOTTOM. Needs to be investigated. 699 if (cross_check != NULL && 700 cross_check != TypePtr::BOTTOM && 701 cross_check != TypeRawPtr::BOTTOM) { 702 // Recheck the alias index, to see if it has changed (due to a bug). 703 Compile* C = Compile::current(); 704 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 705 "must stay in the original alias category"); 706 // The type of the address must be contained in the adr_type, 707 // disregarding "null"-ness. 708 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 709 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 710 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 711 "real address must not escape from expected memory type"); 712 } 713 #endif 714 return tp; 715 } 716 } 717 718 //============================================================================= 719 // Should LoadNode::Ideal() attempt to remove control edges? 720 bool LoadNode::can_remove_control() const { 721 return true; 722 } 723 uint LoadNode::size_of() const { return sizeof(*this); } 724 uint LoadNode::cmp( const Node &n ) const 725 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 726 const Type *LoadNode::bottom_type() const { return _type; } 727 uint LoadNode::ideal_reg() const { 728 return _type->ideal_reg(); 729 } 730 731 #ifndef PRODUCT 732 void LoadNode::dump_spec(outputStream *st) const { 733 MemNode::dump_spec(st); 734 if( !Verbose && !WizardMode ) { 735 // standard dump does this in Verbose and WizardMode 736 st->print(" #"); _type->dump_on(st); 737 } 738 if (!_depends_only_on_test) { 739 st->print(" (does not depend only on test)"); 740 } 741 } 742 #endif 743 744 #ifdef ASSERT 745 //----------------------------is_immutable_value------------------------------- 746 // Helper function to allow a raw load without control edge for some cases 747 bool LoadNode::is_immutable_value(Node* adr) { 748 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 749 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 750 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 751 in_bytes(JavaThread::osthread_offset()))); 752 } 753 #endif 754 755 //----------------------------LoadNode::make----------------------------------- 756 // Polymorphic factory method: 757 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, ControlDependency control_dependency) { 758 Compile* C = gvn.C; 759 760 // sanity check the alias category against the created node type 761 assert(!(adr_type->isa_oopptr() && 762 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 763 "use LoadKlassNode instead"); 764 assert(!(adr_type->isa_aryptr() && 765 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 766 "use LoadRangeNode instead"); 767 // Check control edge of raw loads 768 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 769 // oop will be recorded in oop map if load crosses safepoint 770 rt->isa_oopptr() || is_immutable_value(adr), 771 "raw memory operations should have control edge"); 772 switch (bt) { 773 case T_BOOLEAN: return new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 774 case T_BYTE: return new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 775 case T_INT: return new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 776 case T_CHAR: return new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 777 case T_SHORT: return new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 778 case T_LONG: return new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); 779 case T_FLOAT: return new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); 780 case T_DOUBLE: return new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); 781 case T_ADDRESS: return new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 782 case T_OBJECT: 783 #ifdef _LP64 784 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 785 Node* load = gvn.transform(new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency)); 786 return new DecodeNNode(load, load->bottom_type()->make_ptr()); 787 } else 788 #endif 789 { 790 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 791 return new LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo, control_dependency); 792 } 793 } 794 ShouldNotReachHere(); 795 return (LoadNode*)NULL; 796 } 797 798 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) { 799 bool require_atomic = true; 800 return new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 801 } 802 803 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) { 804 bool require_atomic = true; 805 return new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 806 } 807 808 809 810 //------------------------------hash------------------------------------------- 811 uint LoadNode::hash() const { 812 // unroll addition of interesting fields 813 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 814 } 815 816 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 817 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 818 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 819 bool is_stable_ary = FoldStableValues && 820 (tp != NULL) && (tp->isa_aryptr() != NULL) && 821 tp->isa_aryptr()->is_stable(); 822 823 return (eliminate_boxing && non_volatile) || is_stable_ary; 824 } 825 826 return false; 827 } 828 829 // Is the value loaded previously stored by an arraycopy? If so return 830 // a load node that reads from the source array so we may be able to 831 // optimize out the ArrayCopy node later. 832 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const { 833 Node* ld_adr = in(MemNode::Address); 834 intptr_t ld_off = 0; 835 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 836 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 837 if (ac != NULL) { 838 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 839 840 Node* ld = clone(); 841 if (ac->as_ArrayCopy()->is_clonebasic()) { 842 assert(ld_alloc != NULL, "need an alloc"); 843 Node* addp = in(MemNode::Address)->clone(); 844 assert(addp->is_AddP(), "address must be addp"); 845 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern"); 846 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern"); 847 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base)); 848 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address)); 849 ld->set_req(MemNode::Address, phase->transform(addp)); 850 if (in(0) != NULL) { 851 assert(ld_alloc->in(0) != NULL, "alloc must have control"); 852 ld->set_req(0, ld_alloc->in(0)); 853 } 854 } else { 855 Node* addp = in(MemNode::Address)->clone(); 856 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 857 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)); 858 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)); 859 860 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 861 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 862 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 863 uint shift = exact_log2(type2aelembytes(ary_elem)); 864 865 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 866 #ifdef _LP64 867 diff = phase->transform(new ConvI2LNode(diff)); 868 #endif 869 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 870 871 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 872 addp->set_req(AddPNode::Offset, offset); 873 ld->set_req(MemNode::Address, phase->transform(addp)); 874 875 if (in(0) != NULL) { 876 assert(ac->in(0) != NULL, "alloc must have control"); 877 ld->set_req(0, ac->in(0)); 878 } 879 } 880 // load depends on the tests that validate the arraycopy 881 ld->as_Load()->_depends_only_on_test = Pinned; 882 return ld; 883 } 884 return NULL; 885 } 886 887 888 //---------------------------can_see_stored_value------------------------------ 889 // This routine exists to make sure this set of tests is done the same 890 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 891 // will change the graph shape in a way which makes memory alive twice at the 892 // same time (uses the Oracle model of aliasing), then some 893 // LoadXNode::Identity will fold things back to the equivalence-class model 894 // of aliasing. 895 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 896 Node* ld_adr = in(MemNode::Address); 897 intptr_t ld_off = 0; 898 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 899 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 900 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 901 // This is more general than load from boxing objects. 902 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 903 uint alias_idx = atp->index(); 904 bool final = !atp->is_rewritable(); 905 Node* result = NULL; 906 Node* current = st; 907 // Skip through chains of MemBarNodes checking the MergeMems for 908 // new states for the slice of this load. Stop once any other 909 // kind of node is encountered. Loads from final memory can skip 910 // through any kind of MemBar but normal loads shouldn't skip 911 // through MemBarAcquire since the could allow them to move out of 912 // a synchronized region. 913 while (current->is_Proj()) { 914 int opc = current->in(0)->Opcode(); 915 if ((final && (opc == Op_MemBarAcquire || 916 opc == Op_MemBarAcquireLock || 917 opc == Op_LoadFence)) || 918 opc == Op_MemBarRelease || 919 opc == Op_StoreFence || 920 opc == Op_MemBarReleaseLock || 921 opc == Op_MemBarStoreStore || 922 opc == Op_MemBarCPUOrder) { 923 Node* mem = current->in(0)->in(TypeFunc::Memory); 924 if (mem->is_MergeMem()) { 925 MergeMemNode* merge = mem->as_MergeMem(); 926 Node* new_st = merge->memory_at(alias_idx); 927 if (new_st == merge->base_memory()) { 928 // Keep searching 929 current = new_st; 930 continue; 931 } 932 // Save the new memory state for the slice and fall through 933 // to exit. 934 result = new_st; 935 } 936 } 937 break; 938 } 939 if (result != NULL) { 940 st = result; 941 } 942 } 943 944 // Loop around twice in the case Load -> Initialize -> Store. 945 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 946 for (int trip = 0; trip <= 1; trip++) { 947 948 if (st->is_Store()) { 949 Node* st_adr = st->in(MemNode::Address); 950 if (!phase->eqv(st_adr, ld_adr)) { 951 // Try harder before giving up... Match raw and non-raw pointers. 952 intptr_t st_off = 0; 953 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 954 if (alloc == NULL) return NULL; 955 if (alloc != ld_alloc) return NULL; 956 if (ld_off != st_off) return NULL; 957 // At this point we have proven something like this setup: 958 // A = Allocate(...) 959 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 960 // S = StoreQ(, AddP(, A.Parm , #Off), V) 961 // (Actually, we haven't yet proven the Q's are the same.) 962 // In other words, we are loading from a casted version of 963 // the same pointer-and-offset that we stored to. 964 // Thus, we are able to replace L by V. 965 } 966 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 967 if (store_Opcode() != st->Opcode()) 968 return NULL; 969 return st->in(MemNode::ValueIn); 970 } 971 972 // A load from a freshly-created object always returns zero. 973 // (This can happen after LoadNode::Ideal resets the load's memory input 974 // to find_captured_store, which returned InitializeNode::zero_memory.) 975 if (st->is_Proj() && st->in(0)->is_Allocate() && 976 (st->in(0) == ld_alloc) && 977 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 978 // return a zero value for the load's basic type 979 // (This is one of the few places where a generic PhaseTransform 980 // can create new nodes. Think of it as lazily manifesting 981 // virtually pre-existing constants.) 982 return phase->zerocon(memory_type()); 983 } 984 985 // A load from an initialization barrier can match a captured store. 986 if (st->is_Proj() && st->in(0)->is_Initialize()) { 987 InitializeNode* init = st->in(0)->as_Initialize(); 988 AllocateNode* alloc = init->allocation(); 989 if ((alloc != NULL) && (alloc == ld_alloc)) { 990 // examine a captured store value 991 st = init->find_captured_store(ld_off, memory_size(), phase); 992 if (st != NULL) { 993 continue; // take one more trip around 994 } 995 } 996 } 997 998 // Load boxed value from result of valueOf() call is input parameter. 999 if (this->is_Load() && ld_adr->is_AddP() && 1000 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1001 intptr_t ignore = 0; 1002 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1003 if (base != NULL && base->is_Proj() && 1004 base->as_Proj()->_con == TypeFunc::Parms && 1005 base->in(0)->is_CallStaticJava() && 1006 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1007 return base->in(0)->in(TypeFunc::Parms); 1008 } 1009 } 1010 1011 break; 1012 } 1013 1014 return NULL; 1015 } 1016 1017 //----------------------is_instance_field_load_with_local_phi------------------ 1018 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1019 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1020 in(Address)->is_AddP() ) { 1021 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1022 // Only instances and boxed values. 1023 if( t_oop != NULL && 1024 (t_oop->is_ptr_to_boxed_value() || 1025 t_oop->is_known_instance_field()) && 1026 t_oop->offset() != Type::OffsetBot && 1027 t_oop->offset() != Type::OffsetTop) { 1028 return true; 1029 } 1030 } 1031 return false; 1032 } 1033 1034 //------------------------------Identity--------------------------------------- 1035 // Loads are identity if previous store is to same address 1036 Node *LoadNode::Identity( PhaseTransform *phase ) { 1037 // If the previous store-maker is the right kind of Store, and the store is 1038 // to the same address, then we are equal to the value stored. 1039 Node* mem = in(Memory); 1040 Node* value = can_see_stored_value(mem, phase); 1041 if( value ) { 1042 // byte, short & char stores truncate naturally. 1043 // A load has to load the truncated value which requires 1044 // some sort of masking operation and that requires an 1045 // Ideal call instead of an Identity call. 1046 if (memory_size() < BytesPerInt) { 1047 // If the input to the store does not fit with the load's result type, 1048 // it must be truncated via an Ideal call. 1049 if (!phase->type(value)->higher_equal(phase->type(this))) 1050 return this; 1051 } 1052 // (This works even when value is a Con, but LoadNode::Value 1053 // usually runs first, producing the singleton type of the Con.) 1054 return value; 1055 } 1056 1057 // Search for an existing data phi which was generated before for the same 1058 // instance's field to avoid infinite generation of phis in a loop. 1059 Node *region = mem->in(0); 1060 if (is_instance_field_load_with_local_phi(region)) { 1061 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1062 int this_index = phase->C->get_alias_index(addr_t); 1063 int this_offset = addr_t->offset(); 1064 int this_iid = addr_t->instance_id(); 1065 if (!addr_t->is_known_instance() && 1066 addr_t->is_ptr_to_boxed_value()) { 1067 // Use _idx of address base (could be Phi node) for boxed values. 1068 intptr_t ignore = 0; 1069 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1070 this_iid = base->_idx; 1071 } 1072 const Type* this_type = bottom_type(); 1073 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1074 Node* phi = region->fast_out(i); 1075 if (phi->is_Phi() && phi != mem && 1076 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) { 1077 return phi; 1078 } 1079 } 1080 } 1081 1082 return this; 1083 } 1084 1085 // We're loading from an object which has autobox behaviour. 1086 // If this object is result of a valueOf call we'll have a phi 1087 // merging a newly allocated object and a load from the cache. 1088 // We want to replace this load with the original incoming 1089 // argument to the valueOf call. 1090 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1091 assert(phase->C->eliminate_boxing(), "sanity"); 1092 intptr_t ignore = 0; 1093 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1094 if ((base == NULL) || base->is_Phi()) { 1095 // Push the loads from the phi that comes from valueOf up 1096 // through it to allow elimination of the loads and the recovery 1097 // of the original value. It is done in split_through_phi(). 1098 return NULL; 1099 } else if (base->is_Load() || 1100 base->is_DecodeN() && base->in(1)->is_Load()) { 1101 // Eliminate the load of boxed value for integer types from the cache 1102 // array by deriving the value from the index into the array. 1103 // Capture the offset of the load and then reverse the computation. 1104 1105 // Get LoadN node which loads a boxing object from 'cache' array. 1106 if (base->is_DecodeN()) { 1107 base = base->in(1); 1108 } 1109 if (!base->in(Address)->is_AddP()) { 1110 return NULL; // Complex address 1111 } 1112 AddPNode* address = base->in(Address)->as_AddP(); 1113 Node* cache_base = address->in(AddPNode::Base); 1114 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1115 // Get ConP node which is static 'cache' field. 1116 cache_base = cache_base->in(1); 1117 } 1118 if ((cache_base != NULL) && cache_base->is_Con()) { 1119 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1120 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1121 Node* elements[4]; 1122 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1123 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1124 if ((count > 0) && elements[0]->is_Con() && 1125 ((count == 1) || 1126 (count == 2) && elements[1]->Opcode() == Op_LShiftX && 1127 elements[1]->in(2) == phase->intcon(shift))) { 1128 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1129 // Fetch the box object cache[0] at the base of the array and get its value 1130 ciInstance* box = array->obj_at(0)->as_instance(); 1131 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1132 assert(ik->is_box_klass(), "sanity"); 1133 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1134 if (ik->nof_nonstatic_fields() == 1) { 1135 // This should be true nonstatic_field_at requires calling 1136 // nof_nonstatic_fields so check it anyway 1137 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1138 BasicType bt = c.basic_type(); 1139 // Only integer types have boxing cache. 1140 assert(bt == T_BOOLEAN || bt == T_CHAR || 1141 bt == T_BYTE || bt == T_SHORT || 1142 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt))); 1143 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1144 if (cache_low != (int)cache_low) { 1145 return NULL; // should not happen since cache is array indexed by value 1146 } 1147 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1148 if (offset != (int)offset) { 1149 return NULL; // should not happen since cache is array indexed by value 1150 } 1151 // Add up all the offsets making of the address of the load 1152 Node* result = elements[0]; 1153 for (int i = 1; i < count; i++) { 1154 result = phase->transform(new AddXNode(result, elements[i])); 1155 } 1156 // Remove the constant offset from the address and then 1157 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1158 // remove the scaling of the offset to recover the original index. 1159 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1160 // Peel the shift off directly but wrap it in a dummy node 1161 // since Ideal can't return existing nodes 1162 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1163 } else if (result->is_Add() && result->in(2)->is_Con() && 1164 result->in(1)->Opcode() == Op_LShiftX && 1165 result->in(1)->in(2) == phase->intcon(shift)) { 1166 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1167 // but for boxing cache access we know that X<<Z will not overflow 1168 // (there is range check) so we do this optimizatrion by hand here. 1169 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1170 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1171 } else { 1172 result = new RShiftXNode(result, phase->intcon(shift)); 1173 } 1174 #ifdef _LP64 1175 if (bt != T_LONG) { 1176 result = new ConvL2INode(phase->transform(result)); 1177 } 1178 #else 1179 if (bt == T_LONG) { 1180 result = new ConvI2LNode(phase->transform(result)); 1181 } 1182 #endif 1183 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1184 // Need to preserve unboxing load type if it is unsigned. 1185 switch(this->Opcode()) { 1186 case Op_LoadUB: 1187 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1188 break; 1189 case Op_LoadUS: 1190 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1191 break; 1192 } 1193 return result; 1194 } 1195 } 1196 } 1197 } 1198 } 1199 return NULL; 1200 } 1201 1202 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1203 Node* region = phi->in(0); 1204 if (region == NULL) { 1205 return false; // Wait stable graph 1206 } 1207 uint cnt = phi->req(); 1208 for (uint i = 1; i < cnt; i++) { 1209 Node* rc = region->in(i); 1210 if (rc == NULL || phase->type(rc) == Type::TOP) 1211 return false; // Wait stable graph 1212 Node* in = phi->in(i); 1213 if (in == NULL || phase->type(in) == Type::TOP) 1214 return false; // Wait stable graph 1215 } 1216 return true; 1217 } 1218 //------------------------------split_through_phi------------------------------ 1219 // Split instance or boxed field load through Phi. 1220 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1221 Node* mem = in(Memory); 1222 Node* address = in(Address); 1223 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1224 1225 assert((t_oop != NULL) && 1226 (t_oop->is_known_instance_field() || 1227 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1228 1229 Compile* C = phase->C; 1230 intptr_t ignore = 0; 1231 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1232 bool base_is_phi = (base != NULL) && base->is_Phi(); 1233 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1234 (base != NULL) && (base == address->in(AddPNode::Base)) && 1235 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1236 1237 if (!((mem->is_Phi() || base_is_phi) && 1238 (load_boxed_values || t_oop->is_known_instance_field()))) { 1239 return NULL; // memory is not Phi 1240 } 1241 1242 if (mem->is_Phi()) { 1243 if (!stable_phi(mem->as_Phi(), phase)) { 1244 return NULL; // Wait stable graph 1245 } 1246 uint cnt = mem->req(); 1247 // Check for loop invariant memory. 1248 if (cnt == 3) { 1249 for (uint i = 1; i < cnt; i++) { 1250 Node* in = mem->in(i); 1251 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1252 if (m == mem) { 1253 set_req(Memory, mem->in(cnt - i)); 1254 return this; // made change 1255 } 1256 } 1257 } 1258 } 1259 if (base_is_phi) { 1260 if (!stable_phi(base->as_Phi(), phase)) { 1261 return NULL; // Wait stable graph 1262 } 1263 uint cnt = base->req(); 1264 // Check for loop invariant memory. 1265 if (cnt == 3) { 1266 for (uint i = 1; i < cnt; i++) { 1267 if (base->in(i) == base) { 1268 return NULL; // Wait stable graph 1269 } 1270 } 1271 } 1272 } 1273 1274 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1275 1276 // Split through Phi (see original code in loopopts.cpp). 1277 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1278 1279 // Do nothing here if Identity will find a value 1280 // (to avoid infinite chain of value phis generation). 1281 if (!phase->eqv(this, this->Identity(phase))) 1282 return NULL; 1283 1284 // Select Region to split through. 1285 Node* region; 1286 if (!base_is_phi) { 1287 assert(mem->is_Phi(), "sanity"); 1288 region = mem->in(0); 1289 // Skip if the region dominates some control edge of the address. 1290 if (!MemNode::all_controls_dominate(address, region)) 1291 return NULL; 1292 } else if (!mem->is_Phi()) { 1293 assert(base_is_phi, "sanity"); 1294 region = base->in(0); 1295 // Skip if the region dominates some control edge of the memory. 1296 if (!MemNode::all_controls_dominate(mem, region)) 1297 return NULL; 1298 } else if (base->in(0) != mem->in(0)) { 1299 assert(base_is_phi && mem->is_Phi(), "sanity"); 1300 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1301 region = base->in(0); 1302 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1303 region = mem->in(0); 1304 } else { 1305 return NULL; // complex graph 1306 } 1307 } else { 1308 assert(base->in(0) == mem->in(0), "sanity"); 1309 region = mem->in(0); 1310 } 1311 1312 const Type* this_type = this->bottom_type(); 1313 int this_index = C->get_alias_index(t_oop); 1314 int this_offset = t_oop->offset(); 1315 int this_iid = t_oop->instance_id(); 1316 if (!t_oop->is_known_instance() && load_boxed_values) { 1317 // Use _idx of address base for boxed values. 1318 this_iid = base->_idx; 1319 } 1320 PhaseIterGVN* igvn = phase->is_IterGVN(); 1321 Node* phi = new PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1322 for (uint i = 1; i < region->req(); i++) { 1323 Node* x; 1324 Node* the_clone = NULL; 1325 if (region->in(i) == C->top()) { 1326 x = C->top(); // Dead path? Use a dead data op 1327 } else { 1328 x = this->clone(); // Else clone up the data op 1329 the_clone = x; // Remember for possible deletion. 1330 // Alter data node to use pre-phi inputs 1331 if (this->in(0) == region) { 1332 x->set_req(0, region->in(i)); 1333 } else { 1334 x->set_req(0, NULL); 1335 } 1336 if (mem->is_Phi() && (mem->in(0) == region)) { 1337 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1338 } 1339 if (address->is_Phi() && address->in(0) == region) { 1340 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1341 } 1342 if (base_is_phi && (base->in(0) == region)) { 1343 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1344 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1345 x->set_req(Address, adr_x); 1346 } 1347 } 1348 // Check for a 'win' on some paths 1349 const Type *t = x->Value(igvn); 1350 1351 bool singleton = t->singleton(); 1352 1353 // See comments in PhaseIdealLoop::split_thru_phi(). 1354 if (singleton && t == Type::TOP) { 1355 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1356 } 1357 1358 if (singleton) { 1359 x = igvn->makecon(t); 1360 } else { 1361 // We now call Identity to try to simplify the cloned node. 1362 // Note that some Identity methods call phase->type(this). 1363 // Make sure that the type array is big enough for 1364 // our new node, even though we may throw the node away. 1365 // (This tweaking with igvn only works because x is a new node.) 1366 igvn->set_type(x, t); 1367 // If x is a TypeNode, capture any more-precise type permanently into Node 1368 // otherwise it will be not updated during igvn->transform since 1369 // igvn->type(x) is set to x->Value() already. 1370 x->raise_bottom_type(t); 1371 Node *y = x->Identity(igvn); 1372 if (y != x) { 1373 x = y; 1374 } else { 1375 y = igvn->hash_find_insert(x); 1376 if (y) { 1377 x = y; 1378 } else { 1379 // Else x is a new node we are keeping 1380 // We do not need register_new_node_with_optimizer 1381 // because set_type has already been called. 1382 igvn->_worklist.push(x); 1383 } 1384 } 1385 } 1386 if (x != the_clone && the_clone != NULL) { 1387 igvn->remove_dead_node(the_clone); 1388 } 1389 phi->set_req(i, x); 1390 } 1391 // Record Phi 1392 igvn->register_new_node_with_optimizer(phi); 1393 return phi; 1394 } 1395 1396 //------------------------------Ideal------------------------------------------ 1397 // If the load is from Field memory and the pointer is non-null, it might be possible to 1398 // zero out the control input. 1399 // If the offset is constant and the base is an object allocation, 1400 // try to hook me up to the exact initializing store. 1401 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1402 Node* p = MemNode::Ideal_common(phase, can_reshape); 1403 if (p) return (p == NodeSentinel) ? NULL : p; 1404 1405 Node* ctrl = in(MemNode::Control); 1406 Node* address = in(MemNode::Address); 1407 bool progress = false; 1408 1409 // Skip up past a SafePoint control. Cannot do this for Stores because 1410 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1411 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1412 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1413 ctrl = ctrl->in(0); 1414 set_req(MemNode::Control,ctrl); 1415 progress = true; 1416 } 1417 1418 intptr_t ignore = 0; 1419 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1420 if (base != NULL 1421 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1422 // Check for useless control edge in some common special cases 1423 if (in(MemNode::Control) != NULL 1424 && can_remove_control() 1425 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1426 && all_controls_dominate(base, phase->C->start())) { 1427 // A method-invariant, non-null address (constant or 'this' argument). 1428 set_req(MemNode::Control, NULL); 1429 progress = true; 1430 } 1431 } 1432 1433 Node* mem = in(MemNode::Memory); 1434 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1435 1436 if (can_reshape && (addr_t != NULL)) { 1437 // try to optimize our memory input 1438 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1439 if (opt_mem != mem) { 1440 set_req(MemNode::Memory, opt_mem); 1441 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1442 return this; 1443 } 1444 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1445 if ((t_oop != NULL) && 1446 (t_oop->is_known_instance_field() || 1447 t_oop->is_ptr_to_boxed_value())) { 1448 PhaseIterGVN *igvn = phase->is_IterGVN(); 1449 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1450 // Delay this transformation until memory Phi is processed. 1451 phase->is_IterGVN()->_worklist.push(this); 1452 return NULL; 1453 } 1454 // Split instance field load through Phi. 1455 Node* result = split_through_phi(phase); 1456 if (result != NULL) return result; 1457 1458 if (t_oop->is_ptr_to_boxed_value()) { 1459 Node* result = eliminate_autobox(phase); 1460 if (result != NULL) return result; 1461 } 1462 } 1463 } 1464 1465 // Is there a dominating load that loads the same value? Leave 1466 // anything that is not a load of a field/array element (like 1467 // barriers etc.) alone 1468 if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) { 1469 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1470 Node *use = mem->fast_out(i); 1471 if (use != this && 1472 use->Opcode() == Opcode() && 1473 use->in(0) != NULL && 1474 use->in(0) != in(0) && 1475 use->in(Address) == in(Address)) { 1476 Node* ctl = in(0); 1477 for (int i = 0; i < 10 && ctl != NULL; i++) { 1478 ctl = IfNode::up_one_dom(ctl); 1479 if (ctl == use->in(0)) { 1480 set_req(0, use->in(0)); 1481 return this; 1482 } 1483 } 1484 } 1485 } 1486 } 1487 1488 // Check for prior store with a different base or offset; make Load 1489 // independent. Skip through any number of them. Bail out if the stores 1490 // are in an endless dead cycle and report no progress. This is a key 1491 // transform for Reflection. However, if after skipping through the Stores 1492 // we can't then fold up against a prior store do NOT do the transform as 1493 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1494 // array memory alive twice: once for the hoisted Load and again after the 1495 // bypassed Store. This situation only works if EVERYBODY who does 1496 // anti-dependence work knows how to bypass. I.e. we need all 1497 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1498 // the alias index stuff. So instead, peek through Stores and IFF we can 1499 // fold up, do so. 1500 Node* prev_mem = find_previous_store(phase); 1501 if (prev_mem != NULL) { 1502 Node* value = can_see_arraycopy_value(prev_mem, phase); 1503 if (value != NULL) { 1504 return value; 1505 } 1506 } 1507 // Steps (a), (b): Walk past independent stores to find an exact match. 1508 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1509 // (c) See if we can fold up on the spot, but don't fold up here. 1510 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1511 // just return a prior value, which is done by Identity calls. 1512 if (can_see_stored_value(prev_mem, phase)) { 1513 // Make ready for step (d): 1514 set_req(MemNode::Memory, prev_mem); 1515 return this; 1516 } 1517 } 1518 1519 return progress ? this : NULL; 1520 } 1521 1522 // Helper to recognize certain Klass fields which are invariant across 1523 // some group of array types (e.g., int[] or all T[] where T < Object). 1524 const Type* 1525 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1526 ciKlass* klass) const { 1527 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1528 // The field is Klass::_modifier_flags. Return its (constant) value. 1529 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1530 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1531 return TypeInt::make(klass->modifier_flags()); 1532 } 1533 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1534 // The field is Klass::_access_flags. Return its (constant) value. 1535 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1536 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1537 return TypeInt::make(klass->access_flags()); 1538 } 1539 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1540 // The field is Klass::_layout_helper. Return its constant value if known. 1541 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1542 return TypeInt::make(klass->layout_helper()); 1543 } 1544 1545 // No match. 1546 return NULL; 1547 } 1548 1549 // Try to constant-fold a stable array element. 1550 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) { 1551 assert(ary->const_oop(), "array should be constant"); 1552 assert(ary->is_stable(), "array should be stable"); 1553 1554 // Decode the results of GraphKit::array_element_address. 1555 ciArray* aobj = ary->const_oop()->as_array(); 1556 ciConstant con = aobj->element_value_by_offset(off); 1557 1558 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) { 1559 const Type* con_type = Type::make_from_constant(con); 1560 if (con_type != NULL) { 1561 if (con_type->isa_aryptr()) { 1562 // Join with the array element type, in case it is also stable. 1563 int dim = ary->stable_dimension(); 1564 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1); 1565 } 1566 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) { 1567 con_type = con_type->make_narrowoop(); 1568 } 1569 #ifndef PRODUCT 1570 if (TraceIterativeGVN) { 1571 tty->print("FoldStableValues: array element [off=%d]: con_type=", off); 1572 con_type->dump(); tty->cr(); 1573 } 1574 #endif //PRODUCT 1575 return con_type; 1576 } 1577 } 1578 return NULL; 1579 } 1580 1581 //------------------------------Value----------------------------------------- 1582 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1583 // Either input is TOP ==> the result is TOP 1584 Node* mem = in(MemNode::Memory); 1585 const Type *t1 = phase->type(mem); 1586 if (t1 == Type::TOP) return Type::TOP; 1587 Node* adr = in(MemNode::Address); 1588 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1589 if (tp == NULL || tp->empty()) return Type::TOP; 1590 int off = tp->offset(); 1591 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1592 Compile* C = phase->C; 1593 1594 // Try to guess loaded type from pointer type 1595 if (tp->isa_aryptr()) { 1596 const TypeAryPtr* ary = tp->is_aryptr(); 1597 const Type* t = ary->elem(); 1598 1599 // Determine whether the reference is beyond the header or not, by comparing 1600 // the offset against the offset of the start of the array's data. 1601 // Different array types begin at slightly different offsets (12 vs. 16). 1602 // We choose T_BYTE as an example base type that is least restrictive 1603 // as to alignment, which will therefore produce the smallest 1604 // possible base offset. 1605 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1606 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1607 1608 // Try to constant-fold a stable array element. 1609 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) { 1610 // Make sure the reference is not into the header and the offset is constant 1611 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1612 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type()); 1613 if (con_type != NULL) { 1614 return con_type; 1615 } 1616 } 1617 } 1618 1619 // Don't do this for integer types. There is only potential profit if 1620 // the element type t is lower than _type; that is, for int types, if _type is 1621 // more restrictive than t. This only happens here if one is short and the other 1622 // char (both 16 bits), and in those cases we've made an intentional decision 1623 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1624 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1625 // 1626 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1627 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1628 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1629 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1630 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1631 // In fact, that could have been the original type of p1, and p1 could have 1632 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1633 // expression (LShiftL quux 3) independently optimized to the constant 8. 1634 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1635 && (_type->isa_vect() == NULL) 1636 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1637 // t might actually be lower than _type, if _type is a unique 1638 // concrete subclass of abstract class t. 1639 if (off_beyond_header) { // is the offset beyond the header? 1640 const Type* jt = t->join_speculative(_type); 1641 // In any case, do not allow the join, per se, to empty out the type. 1642 if (jt->empty() && !t->empty()) { 1643 // This can happen if a interface-typed array narrows to a class type. 1644 jt = _type; 1645 } 1646 #ifdef ASSERT 1647 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1648 // The pointers in the autobox arrays are always non-null 1649 Node* base = adr->in(AddPNode::Base); 1650 if ((base != NULL) && base->is_DecodeN()) { 1651 // Get LoadN node which loads IntegerCache.cache field 1652 base = base->in(1); 1653 } 1654 if ((base != NULL) && base->is_Con()) { 1655 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1656 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1657 // It could be narrow oop 1658 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1659 } 1660 } 1661 } 1662 #endif 1663 return jt; 1664 } 1665 } 1666 } else if (tp->base() == Type::InstPtr) { 1667 ciEnv* env = C->env(); 1668 const TypeInstPtr* tinst = tp->is_instptr(); 1669 ciKlass* klass = tinst->klass(); 1670 assert( off != Type::OffsetBot || 1671 // arrays can be cast to Objects 1672 tp->is_oopptr()->klass()->is_java_lang_Object() || 1673 // unsafe field access may not have a constant offset 1674 C->has_unsafe_access(), 1675 "Field accesses must be precise" ); 1676 // For oop loads, we expect the _type to be precise 1677 if (klass == env->String_klass() && 1678 adr->is_AddP() && off != Type::OffsetBot) { 1679 // For constant Strings treat the final fields as compile time constants. 1680 Node* base = adr->in(AddPNode::Base); 1681 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1682 if (t != NULL && t->singleton()) { 1683 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1684 if (field != NULL && field->is_final()) { 1685 ciObject* string = t->const_oop(); 1686 ciConstant constant = string->as_instance()->field_value(field); 1687 if (constant.basic_type() == T_INT) { 1688 return TypeInt::make(constant.as_int()); 1689 } else if (constant.basic_type() == T_ARRAY) { 1690 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1691 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1692 } else { 1693 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1694 } 1695 } 1696 } 1697 } 1698 } 1699 // Optimizations for constant objects 1700 ciObject* const_oop = tinst->const_oop(); 1701 if (const_oop != NULL) { 1702 // For constant Boxed value treat the target field as a compile time constant. 1703 if (tinst->is_ptr_to_boxed_value()) { 1704 return tinst->get_const_boxed_value(); 1705 } else 1706 // For constant CallSites treat the target field as a compile time constant. 1707 if (const_oop->is_call_site()) { 1708 ciCallSite* call_site = const_oop->as_call_site(); 1709 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1710 if (field != NULL && field->is_call_site_target()) { 1711 ciMethodHandle* target = call_site->get_target(); 1712 if (target != NULL) { // just in case 1713 ciConstant constant(T_OBJECT, target); 1714 const Type* t; 1715 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1716 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1717 } else { 1718 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1719 } 1720 // Add a dependence for invalidation of the optimization. 1721 if (!call_site->is_constant_call_site()) { 1722 C->dependencies()->assert_call_site_target_value(call_site, target); 1723 } 1724 return t; 1725 } 1726 } 1727 } 1728 } 1729 } else if (tp->base() == Type::KlassPtr) { 1730 assert( off != Type::OffsetBot || 1731 // arrays can be cast to Objects 1732 tp->is_klassptr()->klass()->is_java_lang_Object() || 1733 // also allow array-loading from the primary supertype 1734 // array during subtype checks 1735 Opcode() == Op_LoadKlass, 1736 "Field accesses must be precise" ); 1737 // For klass/static loads, we expect the _type to be precise 1738 } 1739 1740 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1741 if (tkls != NULL && !StressReflectiveCode) { 1742 ciKlass* klass = tkls->klass(); 1743 if (klass->is_loaded() && tkls->klass_is_exact()) { 1744 // We are loading a field from a Klass metaobject whose identity 1745 // is known at compile time (the type is "exact" or "precise"). 1746 // Check for fields we know are maintained as constants by the VM. 1747 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1748 // The field is Klass::_super_check_offset. Return its (constant) value. 1749 // (Folds up type checking code.) 1750 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1751 return TypeInt::make(klass->super_check_offset()); 1752 } 1753 // Compute index into primary_supers array 1754 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1755 // Check for overflowing; use unsigned compare to handle the negative case. 1756 if( depth < ciKlass::primary_super_limit() ) { 1757 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1758 // (Folds up type checking code.) 1759 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1760 ciKlass *ss = klass->super_of_depth(depth); 1761 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1762 } 1763 const Type* aift = load_array_final_field(tkls, klass); 1764 if (aift != NULL) return aift; 1765 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1766 // The field is Klass::_java_mirror. Return its (constant) value. 1767 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1768 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1769 return TypeInstPtr::make(klass->java_mirror()); 1770 } 1771 } 1772 1773 // We can still check if we are loading from the primary_supers array at a 1774 // shallow enough depth. Even though the klass is not exact, entries less 1775 // than or equal to its super depth are correct. 1776 if (klass->is_loaded() ) { 1777 ciType *inner = klass; 1778 while( inner->is_obj_array_klass() ) 1779 inner = inner->as_obj_array_klass()->base_element_type(); 1780 if( inner->is_instance_klass() && 1781 !inner->as_instance_klass()->flags().is_interface() ) { 1782 // Compute index into primary_supers array 1783 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1784 // Check for overflowing; use unsigned compare to handle the negative case. 1785 if( depth < ciKlass::primary_super_limit() && 1786 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1787 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1788 // (Folds up type checking code.) 1789 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1790 ciKlass *ss = klass->super_of_depth(depth); 1791 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1792 } 1793 } 1794 } 1795 1796 // If the type is enough to determine that the thing is not an array, 1797 // we can give the layout_helper a positive interval type. 1798 // This will help short-circuit some reflective code. 1799 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1800 && !klass->is_array_klass() // not directly typed as an array 1801 && !klass->is_interface() // specifically not Serializable & Cloneable 1802 && !klass->is_java_lang_Object() // not the supertype of all T[] 1803 ) { 1804 // Note: When interfaces are reliable, we can narrow the interface 1805 // test to (klass != Serializable && klass != Cloneable). 1806 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1807 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1808 // The key property of this type is that it folds up tests 1809 // for array-ness, since it proves that the layout_helper is positive. 1810 // Thus, a generic value like the basic object layout helper works fine. 1811 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1812 } 1813 } 1814 1815 // If we are loading from a freshly-allocated object, produce a zero, 1816 // if the load is provably beyond the header of the object. 1817 // (Also allow a variable load from a fresh array to produce zero.) 1818 const TypeOopPtr *tinst = tp->isa_oopptr(); 1819 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1820 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1821 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1822 Node* value = can_see_stored_value(mem,phase); 1823 if (value != NULL && value->is_Con()) { 1824 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1825 return value->bottom_type(); 1826 } 1827 } 1828 1829 if (is_instance) { 1830 // If we have an instance type and our memory input is the 1831 // programs's initial memory state, there is no matching store, 1832 // so just return a zero of the appropriate type 1833 Node *mem = in(MemNode::Memory); 1834 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1835 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1836 return Type::get_zero_type(_type->basic_type()); 1837 } 1838 } 1839 return _type; 1840 } 1841 1842 //------------------------------match_edge------------------------------------- 1843 // Do we Match on this edge index or not? Match only the address. 1844 uint LoadNode::match_edge(uint idx) const { 1845 return idx == MemNode::Address; 1846 } 1847 1848 //--------------------------LoadBNode::Ideal-------------------------------------- 1849 // 1850 // If the previous store is to the same address as this load, 1851 // and the value stored was larger than a byte, replace this load 1852 // with the value stored truncated to a byte. If no truncation is 1853 // needed, the replacement is done in LoadNode::Identity(). 1854 // 1855 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1856 Node* mem = in(MemNode::Memory); 1857 Node* value = can_see_stored_value(mem,phase); 1858 if( value && !phase->type(value)->higher_equal( _type ) ) { 1859 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1860 return new RShiftINode(result, phase->intcon(24)); 1861 } 1862 // Identity call will handle the case where truncation is not needed. 1863 return LoadNode::Ideal(phase, can_reshape); 1864 } 1865 1866 const Type* LoadBNode::Value(PhaseTransform *phase) const { 1867 Node* mem = in(MemNode::Memory); 1868 Node* value = can_see_stored_value(mem,phase); 1869 if (value != NULL && value->is_Con() && 1870 !value->bottom_type()->higher_equal(_type)) { 1871 // If the input to the store does not fit with the load's result type, 1872 // it must be truncated. We can't delay until Ideal call since 1873 // a singleton Value is needed for split_thru_phi optimization. 1874 int con = value->get_int(); 1875 return TypeInt::make((con << 24) >> 24); 1876 } 1877 return LoadNode::Value(phase); 1878 } 1879 1880 //--------------------------LoadUBNode::Ideal------------------------------------- 1881 // 1882 // If the previous store is to the same address as this load, 1883 // and the value stored was larger than a byte, replace this load 1884 // with the value stored truncated to a byte. If no truncation is 1885 // needed, the replacement is done in LoadNode::Identity(). 1886 // 1887 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1888 Node* mem = in(MemNode::Memory); 1889 Node* value = can_see_stored_value(mem, phase); 1890 if (value && !phase->type(value)->higher_equal(_type)) 1891 return new AndINode(value, phase->intcon(0xFF)); 1892 // Identity call will handle the case where truncation is not needed. 1893 return LoadNode::Ideal(phase, can_reshape); 1894 } 1895 1896 const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1897 Node* mem = in(MemNode::Memory); 1898 Node* value = can_see_stored_value(mem,phase); 1899 if (value != NULL && value->is_Con() && 1900 !value->bottom_type()->higher_equal(_type)) { 1901 // If the input to the store does not fit with the load's result type, 1902 // it must be truncated. We can't delay until Ideal call since 1903 // a singleton Value is needed for split_thru_phi optimization. 1904 int con = value->get_int(); 1905 return TypeInt::make(con & 0xFF); 1906 } 1907 return LoadNode::Value(phase); 1908 } 1909 1910 //--------------------------LoadUSNode::Ideal------------------------------------- 1911 // 1912 // If the previous store is to the same address as this load, 1913 // and the value stored was larger than a char, replace this load 1914 // with the value stored truncated to a char. If no truncation is 1915 // needed, the replacement is done in LoadNode::Identity(). 1916 // 1917 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1918 Node* mem = in(MemNode::Memory); 1919 Node* value = can_see_stored_value(mem,phase); 1920 if( value && !phase->type(value)->higher_equal( _type ) ) 1921 return new AndINode(value,phase->intcon(0xFFFF)); 1922 // Identity call will handle the case where truncation is not needed. 1923 return LoadNode::Ideal(phase, can_reshape); 1924 } 1925 1926 const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1927 Node* mem = in(MemNode::Memory); 1928 Node* value = can_see_stored_value(mem,phase); 1929 if (value != NULL && value->is_Con() && 1930 !value->bottom_type()->higher_equal(_type)) { 1931 // If the input to the store does not fit with the load's result type, 1932 // it must be truncated. We can't delay until Ideal call since 1933 // a singleton Value is needed for split_thru_phi optimization. 1934 int con = value->get_int(); 1935 return TypeInt::make(con & 0xFFFF); 1936 } 1937 return LoadNode::Value(phase); 1938 } 1939 1940 //--------------------------LoadSNode::Ideal-------------------------------------- 1941 // 1942 // If the previous store is to the same address as this load, 1943 // and the value stored was larger than a short, replace this load 1944 // with the value stored truncated to a short. If no truncation is 1945 // needed, the replacement is done in LoadNode::Identity(). 1946 // 1947 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1948 Node* mem = in(MemNode::Memory); 1949 Node* value = can_see_stored_value(mem,phase); 1950 if( value && !phase->type(value)->higher_equal( _type ) ) { 1951 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 1952 return new RShiftINode(result, phase->intcon(16)); 1953 } 1954 // Identity call will handle the case where truncation is not needed. 1955 return LoadNode::Ideal(phase, can_reshape); 1956 } 1957 1958 const Type* LoadSNode::Value(PhaseTransform *phase) const { 1959 Node* mem = in(MemNode::Memory); 1960 Node* value = can_see_stored_value(mem,phase); 1961 if (value != NULL && value->is_Con() && 1962 !value->bottom_type()->higher_equal(_type)) { 1963 // If the input to the store does not fit with the load's result type, 1964 // it must be truncated. We can't delay until Ideal call since 1965 // a singleton Value is needed for split_thru_phi optimization. 1966 int con = value->get_int(); 1967 return TypeInt::make((con << 16) >> 16); 1968 } 1969 return LoadNode::Value(phase); 1970 } 1971 1972 //============================================================================= 1973 //----------------------------LoadKlassNode::make------------------------------ 1974 // Polymorphic factory method: 1975 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 1976 // sanity check the alias category against the created node type 1977 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 1978 assert(adr_type != NULL, "expecting TypeKlassPtr"); 1979 #ifdef _LP64 1980 if (adr_type->is_ptr_to_narrowklass()) { 1981 assert(UseCompressedClassPointers, "no compressed klasses"); 1982 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 1983 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 1984 } 1985 #endif 1986 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1987 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 1988 } 1989 1990 //------------------------------Value------------------------------------------ 1991 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1992 return klass_value_common(phase); 1993 } 1994 1995 // In most cases, LoadKlassNode does not have the control input set. If the control 1996 // input is set, it must not be removed (by LoadNode::Ideal()). 1997 bool LoadKlassNode::can_remove_control() const { 1998 return false; 1999 } 2000 2001 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 2002 // Either input is TOP ==> the result is TOP 2003 const Type *t1 = phase->type( in(MemNode::Memory) ); 2004 if (t1 == Type::TOP) return Type::TOP; 2005 Node *adr = in(MemNode::Address); 2006 const Type *t2 = phase->type( adr ); 2007 if (t2 == Type::TOP) return Type::TOP; 2008 const TypePtr *tp = t2->is_ptr(); 2009 if (TypePtr::above_centerline(tp->ptr()) || 2010 tp->ptr() == TypePtr::Null) return Type::TOP; 2011 2012 // Return a more precise klass, if possible 2013 const TypeInstPtr *tinst = tp->isa_instptr(); 2014 if (tinst != NULL) { 2015 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2016 int offset = tinst->offset(); 2017 if (ik == phase->C->env()->Class_klass() 2018 && (offset == java_lang_Class::klass_offset_in_bytes() || 2019 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2020 // We are loading a special hidden field from a Class mirror object, 2021 // the field which points to the VM's Klass metaobject. 2022 ciType* t = tinst->java_mirror_type(); 2023 // java_mirror_type returns non-null for compile-time Class constants. 2024 if (t != NULL) { 2025 // constant oop => constant klass 2026 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2027 if (t->is_void()) { 2028 // We cannot create a void array. Since void is a primitive type return null 2029 // klass. Users of this result need to do a null check on the returned klass. 2030 return TypePtr::NULL_PTR; 2031 } 2032 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2033 } 2034 if (!t->is_klass()) { 2035 // a primitive Class (e.g., int.class) has NULL for a klass field 2036 return TypePtr::NULL_PTR; 2037 } 2038 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2039 return TypeKlassPtr::make(t->as_klass()); 2040 } 2041 // non-constant mirror, so we can't tell what's going on 2042 } 2043 if( !ik->is_loaded() ) 2044 return _type; // Bail out if not loaded 2045 if (offset == oopDesc::klass_offset_in_bytes()) { 2046 if (tinst->klass_is_exact()) { 2047 return TypeKlassPtr::make(ik); 2048 } 2049 // See if we can become precise: no subklasses and no interface 2050 // (Note: We need to support verified interfaces.) 2051 if (!ik->is_interface() && !ik->has_subklass()) { 2052 //assert(!UseExactTypes, "this code should be useless with exact types"); 2053 // Add a dependence; if any subclass added we need to recompile 2054 if (!ik->is_final()) { 2055 // %%% should use stronger assert_unique_concrete_subtype instead 2056 phase->C->dependencies()->assert_leaf_type(ik); 2057 } 2058 // Return precise klass 2059 return TypeKlassPtr::make(ik); 2060 } 2061 2062 // Return root of possible klass 2063 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2064 } 2065 } 2066 2067 // Check for loading klass from an array 2068 const TypeAryPtr *tary = tp->isa_aryptr(); 2069 if( tary != NULL ) { 2070 ciKlass *tary_klass = tary->klass(); 2071 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2072 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2073 if (tary->klass_is_exact()) { 2074 return TypeKlassPtr::make(tary_klass); 2075 } 2076 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2077 // If the klass is an object array, we defer the question to the 2078 // array component klass. 2079 if( ak->is_obj_array_klass() ) { 2080 assert( ak->is_loaded(), "" ); 2081 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2082 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2083 ciInstanceKlass* ik = base_k->as_instance_klass(); 2084 // See if we can become precise: no subklasses and no interface 2085 if (!ik->is_interface() && !ik->has_subklass()) { 2086 //assert(!UseExactTypes, "this code should be useless with exact types"); 2087 // Add a dependence; if any subclass added we need to recompile 2088 if (!ik->is_final()) { 2089 phase->C->dependencies()->assert_leaf_type(ik); 2090 } 2091 // Return precise array klass 2092 return TypeKlassPtr::make(ak); 2093 } 2094 } 2095 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2096 } else { // Found a type-array? 2097 //assert(!UseExactTypes, "this code should be useless with exact types"); 2098 assert( ak->is_type_array_klass(), "" ); 2099 return TypeKlassPtr::make(ak); // These are always precise 2100 } 2101 } 2102 } 2103 2104 // Check for loading klass from an array klass 2105 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2106 if (tkls != NULL && !StressReflectiveCode) { 2107 ciKlass* klass = tkls->klass(); 2108 if( !klass->is_loaded() ) 2109 return _type; // Bail out if not loaded 2110 if( klass->is_obj_array_klass() && 2111 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2112 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2113 // // Always returning precise element type is incorrect, 2114 // // e.g., element type could be object and array may contain strings 2115 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2116 2117 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2118 // according to the element type's subclassing. 2119 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2120 } 2121 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2122 tkls->offset() == in_bytes(Klass::super_offset())) { 2123 ciKlass* sup = klass->as_instance_klass()->super(); 2124 // The field is Klass::_super. Return its (constant) value. 2125 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2126 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2127 } 2128 } 2129 2130 // Bailout case 2131 return LoadNode::Value(phase); 2132 } 2133 2134 //------------------------------Identity--------------------------------------- 2135 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2136 // Also feed through the klass in Allocate(...klass...)._klass. 2137 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2138 return klass_identity_common(phase); 2139 } 2140 2141 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2142 Node* x = LoadNode::Identity(phase); 2143 if (x != this) return x; 2144 2145 // Take apart the address into an oop and and offset. 2146 // Return 'this' if we cannot. 2147 Node* adr = in(MemNode::Address); 2148 intptr_t offset = 0; 2149 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2150 if (base == NULL) return this; 2151 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2152 if (toop == NULL) return this; 2153 2154 // We can fetch the klass directly through an AllocateNode. 2155 // This works even if the klass is not constant (clone or newArray). 2156 if (offset == oopDesc::klass_offset_in_bytes()) { 2157 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2158 if (allocated_klass != NULL) { 2159 return allocated_klass; 2160 } 2161 } 2162 2163 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2164 // See inline_native_Class_query for occurrences of these patterns. 2165 // Java Example: x.getClass().isAssignableFrom(y) 2166 // 2167 // This improves reflective code, often making the Class 2168 // mirror go completely dead. (Current exception: Class 2169 // mirrors may appear in debug info, but we could clean them out by 2170 // introducing a new debug info operator for Klass*.java_mirror). 2171 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2172 && offset == java_lang_Class::klass_offset_in_bytes()) { 2173 // We are loading a special hidden field from a Class mirror, 2174 // the field which points to its Klass or ArrayKlass metaobject. 2175 if (base->is_Load()) { 2176 Node* adr2 = base->in(MemNode::Address); 2177 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2178 if (tkls != NULL && !tkls->empty() 2179 && (tkls->klass()->is_instance_klass() || 2180 tkls->klass()->is_array_klass()) 2181 && adr2->is_AddP() 2182 ) { 2183 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2184 if (tkls->offset() == mirror_field) { 2185 return adr2->in(AddPNode::Base); 2186 } 2187 } 2188 } 2189 } 2190 2191 return this; 2192 } 2193 2194 2195 //------------------------------Value------------------------------------------ 2196 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2197 const Type *t = klass_value_common(phase); 2198 if (t == Type::TOP) 2199 return t; 2200 2201 return t->make_narrowklass(); 2202 } 2203 2204 //------------------------------Identity--------------------------------------- 2205 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2206 // Also feed through the klass in Allocate(...klass...)._klass. 2207 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2208 Node *x = klass_identity_common(phase); 2209 2210 const Type *t = phase->type( x ); 2211 if( t == Type::TOP ) return x; 2212 if( t->isa_narrowklass()) return x; 2213 assert (!t->isa_narrowoop(), "no narrow oop here"); 2214 2215 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2216 } 2217 2218 //------------------------------Value----------------------------------------- 2219 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2220 // Either input is TOP ==> the result is TOP 2221 const Type *t1 = phase->type( in(MemNode::Memory) ); 2222 if( t1 == Type::TOP ) return Type::TOP; 2223 Node *adr = in(MemNode::Address); 2224 const Type *t2 = phase->type( adr ); 2225 if( t2 == Type::TOP ) return Type::TOP; 2226 const TypePtr *tp = t2->is_ptr(); 2227 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2228 const TypeAryPtr *tap = tp->isa_aryptr(); 2229 if( !tap ) return _type; 2230 return tap->size(); 2231 } 2232 2233 //-------------------------------Ideal--------------------------------------- 2234 // Feed through the length in AllocateArray(...length...)._length. 2235 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2236 Node* p = MemNode::Ideal_common(phase, can_reshape); 2237 if (p) return (p == NodeSentinel) ? NULL : p; 2238 2239 // Take apart the address into an oop and and offset. 2240 // Return 'this' if we cannot. 2241 Node* adr = in(MemNode::Address); 2242 intptr_t offset = 0; 2243 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2244 if (base == NULL) return NULL; 2245 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2246 if (tary == NULL) return NULL; 2247 2248 // We can fetch the length directly through an AllocateArrayNode. 2249 // This works even if the length is not constant (clone or newArray). 2250 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2251 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2252 if (alloc != NULL) { 2253 Node* allocated_length = alloc->Ideal_length(); 2254 Node* len = alloc->make_ideal_length(tary, phase); 2255 if (allocated_length != len) { 2256 // New CastII improves on this. 2257 return len; 2258 } 2259 } 2260 } 2261 2262 return NULL; 2263 } 2264 2265 //------------------------------Identity--------------------------------------- 2266 // Feed through the length in AllocateArray(...length...)._length. 2267 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2268 Node* x = LoadINode::Identity(phase); 2269 if (x != this) return x; 2270 2271 // Take apart the address into an oop and and offset. 2272 // Return 'this' if we cannot. 2273 Node* adr = in(MemNode::Address); 2274 intptr_t offset = 0; 2275 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2276 if (base == NULL) return this; 2277 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2278 if (tary == NULL) return this; 2279 2280 // We can fetch the length directly through an AllocateArrayNode. 2281 // This works even if the length is not constant (clone or newArray). 2282 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2283 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2284 if (alloc != NULL) { 2285 Node* allocated_length = alloc->Ideal_length(); 2286 // Do not allow make_ideal_length to allocate a CastII node. 2287 Node* len = alloc->make_ideal_length(tary, phase, false); 2288 if (allocated_length == len) { 2289 // Return allocated_length only if it would not be improved by a CastII. 2290 return allocated_length; 2291 } 2292 } 2293 } 2294 2295 return this; 2296 2297 } 2298 2299 //============================================================================= 2300 //---------------------------StoreNode::make----------------------------------- 2301 // Polymorphic factory method: 2302 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2303 assert((mo == unordered || mo == release), "unexpected"); 2304 Compile* C = gvn.C; 2305 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2306 ctl != NULL, "raw memory operations should have control edge"); 2307 2308 switch (bt) { 2309 case T_BOOLEAN: 2310 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2311 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2312 case T_CHAR: 2313 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2314 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2315 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2316 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2317 case T_METADATA: 2318 case T_ADDRESS: 2319 case T_OBJECT: 2320 #ifdef _LP64 2321 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2322 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2323 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2324 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2325 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2326 adr->bottom_type()->isa_rawptr())) { 2327 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2328 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2329 } 2330 #endif 2331 { 2332 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2333 } 2334 } 2335 ShouldNotReachHere(); 2336 return (StoreNode*)NULL; 2337 } 2338 2339 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2340 bool require_atomic = true; 2341 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2342 } 2343 2344 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2345 bool require_atomic = true; 2346 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2347 } 2348 2349 2350 //--------------------------bottom_type---------------------------------------- 2351 const Type *StoreNode::bottom_type() const { 2352 return Type::MEMORY; 2353 } 2354 2355 //------------------------------hash------------------------------------------- 2356 uint StoreNode::hash() const { 2357 // unroll addition of interesting fields 2358 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2359 2360 // Since they are not commoned, do not hash them: 2361 return NO_HASH; 2362 } 2363 2364 //------------------------------Ideal------------------------------------------ 2365 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2366 // When a store immediately follows a relevant allocation/initialization, 2367 // try to capture it into the initialization, or hoist it above. 2368 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2369 Node* p = MemNode::Ideal_common(phase, can_reshape); 2370 if (p) return (p == NodeSentinel) ? NULL : p; 2371 2372 Node* mem = in(MemNode::Memory); 2373 Node* address = in(MemNode::Address); 2374 2375 // Back-to-back stores to same address? Fold em up. Generally 2376 // unsafe if I have intervening uses... Also disallowed for StoreCM 2377 // since they must follow each StoreP operation. Redundant StoreCMs 2378 // are eliminated just before matching in final_graph_reshape. 2379 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && 2380 mem->Opcode() != Op_StoreCM) { 2381 // Looking at a dead closed cycle of memory? 2382 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2383 2384 assert(Opcode() == mem->Opcode() || 2385 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2386 "no mismatched stores, except on raw memory"); 2387 2388 if (mem->outcnt() == 1 && // check for intervening uses 2389 mem->as_Store()->memory_size() <= this->memory_size()) { 2390 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2391 // For example, 'mem' might be the final state at a conditional return. 2392 // Or, 'mem' might be used by some node which is live at the same time 2393 // 'this' is live, which might be unschedulable. So, require exactly 2394 // ONE user, the 'this' store, until such time as we clone 'mem' for 2395 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2396 if (can_reshape) { // (%%% is this an anachronism?) 2397 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2398 phase->is_IterGVN()); 2399 } else { 2400 // It's OK to do this in the parser, since DU info is always accurate, 2401 // and the parser always refers to nodes via SafePointNode maps. 2402 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2403 } 2404 return this; 2405 } 2406 } 2407 2408 // Capture an unaliased, unconditional, simple store into an initializer. 2409 // Or, if it is independent of the allocation, hoist it above the allocation. 2410 if (ReduceFieldZeroing && /*can_reshape &&*/ 2411 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2412 InitializeNode* init = mem->in(0)->as_Initialize(); 2413 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2414 if (offset > 0) { 2415 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2416 // If the InitializeNode captured me, it made a raw copy of me, 2417 // and I need to disappear. 2418 if (moved != NULL) { 2419 // %%% hack to ensure that Ideal returns a new node: 2420 mem = MergeMemNode::make(mem); 2421 return mem; // fold me away 2422 } 2423 } 2424 } 2425 2426 return NULL; // No further progress 2427 } 2428 2429 //------------------------------Value----------------------------------------- 2430 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2431 // Either input is TOP ==> the result is TOP 2432 const Type *t1 = phase->type( in(MemNode::Memory) ); 2433 if( t1 == Type::TOP ) return Type::TOP; 2434 const Type *t2 = phase->type( in(MemNode::Address) ); 2435 if( t2 == Type::TOP ) return Type::TOP; 2436 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2437 if( t3 == Type::TOP ) return Type::TOP; 2438 return Type::MEMORY; 2439 } 2440 2441 //------------------------------Identity--------------------------------------- 2442 // Remove redundant stores: 2443 // Store(m, p, Load(m, p)) changes to m. 2444 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2445 Node *StoreNode::Identity( PhaseTransform *phase ) { 2446 Node* mem = in(MemNode::Memory); 2447 Node* adr = in(MemNode::Address); 2448 Node* val = in(MemNode::ValueIn); 2449 2450 // Load then Store? Then the Store is useless 2451 if (val->is_Load() && 2452 val->in(MemNode::Address)->eqv_uncast(adr) && 2453 val->in(MemNode::Memory )->eqv_uncast(mem) && 2454 val->as_Load()->store_Opcode() == Opcode()) { 2455 return mem; 2456 } 2457 2458 // Two stores in a row of the same value? 2459 if (mem->is_Store() && 2460 mem->in(MemNode::Address)->eqv_uncast(adr) && 2461 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2462 mem->Opcode() == Opcode()) { 2463 return mem; 2464 } 2465 2466 // Store of zero anywhere into a freshly-allocated object? 2467 // Then the store is useless. 2468 // (It must already have been captured by the InitializeNode.) 2469 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2470 // a newly allocated object is already all-zeroes everywhere 2471 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2472 return mem; 2473 } 2474 2475 // the store may also apply to zero-bits in an earlier object 2476 Node* prev_mem = find_previous_store(phase); 2477 // Steps (a), (b): Walk past independent stores to find an exact match. 2478 if (prev_mem != NULL) { 2479 Node* prev_val = can_see_stored_value(prev_mem, phase); 2480 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2481 // prev_val and val might differ by a cast; it would be good 2482 // to keep the more informative of the two. 2483 return mem; 2484 } 2485 } 2486 } 2487 2488 return this; 2489 } 2490 2491 //------------------------------match_edge------------------------------------- 2492 // Do we Match on this edge index or not? Match only memory & value 2493 uint StoreNode::match_edge(uint idx) const { 2494 return idx == MemNode::Address || idx == MemNode::ValueIn; 2495 } 2496 2497 //------------------------------cmp-------------------------------------------- 2498 // Do not common stores up together. They generally have to be split 2499 // back up anyways, so do not bother. 2500 uint StoreNode::cmp( const Node &n ) const { 2501 return (&n == this); // Always fail except on self 2502 } 2503 2504 //------------------------------Ideal_masked_input----------------------------- 2505 // Check for a useless mask before a partial-word store 2506 // (StoreB ... (AndI valIn conIa) ) 2507 // If (conIa & mask == mask) this simplifies to 2508 // (StoreB ... (valIn) ) 2509 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2510 Node *val = in(MemNode::ValueIn); 2511 if( val->Opcode() == Op_AndI ) { 2512 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2513 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2514 set_req(MemNode::ValueIn, val->in(1)); 2515 return this; 2516 } 2517 } 2518 return NULL; 2519 } 2520 2521 2522 //------------------------------Ideal_sign_extended_input---------------------- 2523 // Check for useless sign-extension before a partial-word store 2524 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2525 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2526 // (StoreB ... (valIn) ) 2527 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2528 Node *val = in(MemNode::ValueIn); 2529 if( val->Opcode() == Op_RShiftI ) { 2530 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2531 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2532 Node *shl = val->in(1); 2533 if( shl->Opcode() == Op_LShiftI ) { 2534 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2535 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2536 set_req(MemNode::ValueIn, shl->in(1)); 2537 return this; 2538 } 2539 } 2540 } 2541 } 2542 return NULL; 2543 } 2544 2545 //------------------------------value_never_loaded----------------------------------- 2546 // Determine whether there are any possible loads of the value stored. 2547 // For simplicity, we actually check if there are any loads from the 2548 // address stored to, not just for loads of the value stored by this node. 2549 // 2550 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2551 Node *adr = in(Address); 2552 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2553 if (adr_oop == NULL) 2554 return false; 2555 if (!adr_oop->is_known_instance_field()) 2556 return false; // if not a distinct instance, there may be aliases of the address 2557 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2558 Node *use = adr->fast_out(i); 2559 if (use->is_Load() || use->is_LoadStore()) { 2560 return false; 2561 } 2562 } 2563 return true; 2564 } 2565 2566 //============================================================================= 2567 //------------------------------Ideal------------------------------------------ 2568 // If the store is from an AND mask that leaves the low bits untouched, then 2569 // we can skip the AND operation. If the store is from a sign-extension 2570 // (a left shift, then right shift) we can skip both. 2571 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2572 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2573 if( progress != NULL ) return progress; 2574 2575 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2576 if( progress != NULL ) return progress; 2577 2578 // Finally check the default case 2579 return StoreNode::Ideal(phase, can_reshape); 2580 } 2581 2582 //============================================================================= 2583 //------------------------------Ideal------------------------------------------ 2584 // If the store is from an AND mask that leaves the low bits untouched, then 2585 // we can skip the AND operation 2586 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2587 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2588 if( progress != NULL ) return progress; 2589 2590 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2591 if( progress != NULL ) return progress; 2592 2593 // Finally check the default case 2594 return StoreNode::Ideal(phase, can_reshape); 2595 } 2596 2597 //============================================================================= 2598 //------------------------------Identity--------------------------------------- 2599 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2600 // No need to card mark when storing a null ptr 2601 Node* my_store = in(MemNode::OopStore); 2602 if (my_store->is_Store()) { 2603 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2604 if( t1 == TypePtr::NULL_PTR ) { 2605 return in(MemNode::Memory); 2606 } 2607 } 2608 return this; 2609 } 2610 2611 //============================================================================= 2612 //------------------------------Ideal--------------------------------------- 2613 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2614 Node* progress = StoreNode::Ideal(phase, can_reshape); 2615 if (progress != NULL) return progress; 2616 2617 Node* my_store = in(MemNode::OopStore); 2618 if (my_store->is_MergeMem()) { 2619 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2620 set_req(MemNode::OopStore, mem); 2621 return this; 2622 } 2623 2624 return NULL; 2625 } 2626 2627 //------------------------------Value----------------------------------------- 2628 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2629 // Either input is TOP ==> the result is TOP 2630 const Type *t = phase->type( in(MemNode::Memory) ); 2631 if( t == Type::TOP ) return Type::TOP; 2632 t = phase->type( in(MemNode::Address) ); 2633 if( t == Type::TOP ) return Type::TOP; 2634 t = phase->type( in(MemNode::ValueIn) ); 2635 if( t == Type::TOP ) return Type::TOP; 2636 // If extra input is TOP ==> the result is TOP 2637 t = phase->type( in(MemNode::OopStore) ); 2638 if( t == Type::TOP ) return Type::TOP; 2639 2640 return StoreNode::Value( phase ); 2641 } 2642 2643 2644 //============================================================================= 2645 //----------------------------------SCMemProjNode------------------------------ 2646 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2647 { 2648 return bottom_type(); 2649 } 2650 2651 //============================================================================= 2652 //----------------------------------LoadStoreNode------------------------------ 2653 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2654 : Node(required), 2655 _type(rt), 2656 _adr_type(at) 2657 { 2658 init_req(MemNode::Control, c ); 2659 init_req(MemNode::Memory , mem); 2660 init_req(MemNode::Address, adr); 2661 init_req(MemNode::ValueIn, val); 2662 init_class_id(Class_LoadStore); 2663 } 2664 2665 uint LoadStoreNode::ideal_reg() const { 2666 return _type->ideal_reg(); 2667 } 2668 2669 bool LoadStoreNode::result_not_used() const { 2670 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2671 Node *x = fast_out(i); 2672 if (x->Opcode() == Op_SCMemProj) continue; 2673 return false; 2674 } 2675 return true; 2676 } 2677 2678 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2679 2680 //============================================================================= 2681 //----------------------------------LoadStoreConditionalNode-------------------- 2682 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2683 init_req(ExpectedIn, ex ); 2684 } 2685 2686 //============================================================================= 2687 //-------------------------------adr_type-------------------------------------- 2688 const TypePtr* ClearArrayNode::adr_type() const { 2689 Node *adr = in(3); 2690 if (adr == NULL) return NULL; // node is dead 2691 return MemNode::calculate_adr_type(adr->bottom_type()); 2692 } 2693 2694 //------------------------------match_edge------------------------------------- 2695 // Do we Match on this edge index or not? Do not match memory 2696 uint ClearArrayNode::match_edge(uint idx) const { 2697 return idx > 1; 2698 } 2699 2700 //------------------------------Identity--------------------------------------- 2701 // Clearing a zero length array does nothing 2702 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2703 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2704 } 2705 2706 //------------------------------Idealize--------------------------------------- 2707 // Clearing a short array is faster with stores 2708 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2709 const int unit = BytesPerLong; 2710 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2711 if (!t) return NULL; 2712 if (!t->is_con()) return NULL; 2713 intptr_t raw_count = t->get_con(); 2714 intptr_t size = raw_count; 2715 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2716 // Clearing nothing uses the Identity call. 2717 // Negative clears are possible on dead ClearArrays 2718 // (see jck test stmt114.stmt11402.val). 2719 if (size <= 0 || size % unit != 0) return NULL; 2720 intptr_t count = size / unit; 2721 // Length too long; use fast hardware clear 2722 if (size > Matcher::init_array_short_size) return NULL; 2723 Node *mem = in(1); 2724 if( phase->type(mem)==Type::TOP ) return NULL; 2725 Node *adr = in(3); 2726 const Type* at = phase->type(adr); 2727 if( at==Type::TOP ) return NULL; 2728 const TypePtr* atp = at->isa_ptr(); 2729 // adjust atp to be the correct array element address type 2730 if (atp == NULL) atp = TypePtr::BOTTOM; 2731 else atp = atp->add_offset(Type::OffsetBot); 2732 // Get base for derived pointer purposes 2733 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2734 Node *base = adr->in(1); 2735 2736 Node *zero = phase->makecon(TypeLong::ZERO); 2737 Node *off = phase->MakeConX(BytesPerLong); 2738 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2739 count--; 2740 while( count-- ) { 2741 mem = phase->transform(mem); 2742 adr = phase->transform(new AddPNode(base,adr,off)); 2743 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2744 } 2745 return mem; 2746 } 2747 2748 //----------------------------step_through---------------------------------- 2749 // Return allocation input memory edge if it is different instance 2750 // or itself if it is the one we are looking for. 2751 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2752 Node* n = *np; 2753 assert(n->is_ClearArray(), "sanity"); 2754 intptr_t offset; 2755 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2756 // This method is called only before Allocate nodes are expanded 2757 // during macro nodes expansion. Before that ClearArray nodes are 2758 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2759 // Allocate nodes are expanded) which follows allocations. 2760 assert(alloc != NULL, "should have allocation"); 2761 if (alloc->_idx == instance_id) { 2762 // Can not bypass initialization of the instance we are looking for. 2763 return false; 2764 } 2765 // Otherwise skip it. 2766 InitializeNode* init = alloc->initialization(); 2767 if (init != NULL) 2768 *np = init->in(TypeFunc::Memory); 2769 else 2770 *np = alloc->in(TypeFunc::Memory); 2771 return true; 2772 } 2773 2774 //----------------------------clear_memory------------------------------------- 2775 // Generate code to initialize object storage to zero. 2776 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2777 intptr_t start_offset, 2778 Node* end_offset, 2779 PhaseGVN* phase) { 2780 intptr_t offset = start_offset; 2781 2782 int unit = BytesPerLong; 2783 if ((offset % unit) != 0) { 2784 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2785 adr = phase->transform(adr); 2786 const TypePtr* atp = TypeRawPtr::BOTTOM; 2787 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2788 mem = phase->transform(mem); 2789 offset += BytesPerInt; 2790 } 2791 assert((offset % unit) == 0, ""); 2792 2793 // Initialize the remaining stuff, if any, with a ClearArray. 2794 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2795 } 2796 2797 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2798 Node* start_offset, 2799 Node* end_offset, 2800 PhaseGVN* phase) { 2801 if (start_offset == end_offset) { 2802 // nothing to do 2803 return mem; 2804 } 2805 2806 int unit = BytesPerLong; 2807 Node* zbase = start_offset; 2808 Node* zend = end_offset; 2809 2810 // Scale to the unit required by the CPU: 2811 if (!Matcher::init_array_count_is_in_bytes) { 2812 Node* shift = phase->intcon(exact_log2(unit)); 2813 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2814 zend = phase->transform(new URShiftXNode(zend, shift) ); 2815 } 2816 2817 // Bulk clear double-words 2818 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2819 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2820 mem = new ClearArrayNode(ctl, mem, zsize, adr); 2821 return phase->transform(mem); 2822 } 2823 2824 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2825 intptr_t start_offset, 2826 intptr_t end_offset, 2827 PhaseGVN* phase) { 2828 if (start_offset == end_offset) { 2829 // nothing to do 2830 return mem; 2831 } 2832 2833 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2834 intptr_t done_offset = end_offset; 2835 if ((done_offset % BytesPerLong) != 0) { 2836 done_offset -= BytesPerInt; 2837 } 2838 if (done_offset > start_offset) { 2839 mem = clear_memory(ctl, mem, dest, 2840 start_offset, phase->MakeConX(done_offset), phase); 2841 } 2842 if (done_offset < end_offset) { // emit the final 32-bit store 2843 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 2844 adr = phase->transform(adr); 2845 const TypePtr* atp = TypeRawPtr::BOTTOM; 2846 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2847 mem = phase->transform(mem); 2848 done_offset += BytesPerInt; 2849 } 2850 assert(done_offset == end_offset, ""); 2851 return mem; 2852 } 2853 2854 //============================================================================= 2855 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2856 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2857 _adr_type(C->get_adr_type(alias_idx)) 2858 { 2859 init_class_id(Class_MemBar); 2860 Node* top = C->top(); 2861 init_req(TypeFunc::I_O,top); 2862 init_req(TypeFunc::FramePtr,top); 2863 init_req(TypeFunc::ReturnAdr,top); 2864 if (precedent != NULL) 2865 init_req(TypeFunc::Parms, precedent); 2866 } 2867 2868 //------------------------------cmp-------------------------------------------- 2869 uint MemBarNode::hash() const { return NO_HASH; } 2870 uint MemBarNode::cmp( const Node &n ) const { 2871 return (&n == this); // Always fail except on self 2872 } 2873 2874 //------------------------------make------------------------------------------- 2875 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2876 switch (opcode) { 2877 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 2878 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 2879 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 2880 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 2881 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 2882 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 2883 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 2884 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 2885 case Op_Initialize: return new InitializeNode(C, atp, pn); 2886 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 2887 default: ShouldNotReachHere(); return NULL; 2888 } 2889 } 2890 2891 //------------------------------Ideal------------------------------------------ 2892 // Return a node which is more "ideal" than the current node. Strip out 2893 // control copies 2894 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2895 if (remove_dead_region(phase, can_reshape)) return this; 2896 // Don't bother trying to transform a dead node 2897 if (in(0) && in(0)->is_top()) { 2898 return NULL; 2899 } 2900 2901 bool progress = false; 2902 // Eliminate volatile MemBars for scalar replaced objects. 2903 if (can_reshape && req() == (Precedent+1)) { 2904 bool eliminate = false; 2905 int opc = Opcode(); 2906 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2907 // Volatile field loads and stores. 2908 Node* my_mem = in(MemBarNode::Precedent); 2909 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2910 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2911 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2912 // replace this Precedent (decodeN) with the Load instead. 2913 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2914 Node* load_node = my_mem->in(1); 2915 set_req(MemBarNode::Precedent, load_node); 2916 phase->is_IterGVN()->_worklist.push(my_mem); 2917 my_mem = load_node; 2918 } else { 2919 assert(my_mem->unique_out() == this, "sanity"); 2920 del_req(Precedent); 2921 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2922 my_mem = NULL; 2923 } 2924 progress = true; 2925 } 2926 if (my_mem != NULL && my_mem->is_Mem()) { 2927 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2928 // Check for scalar replaced object reference. 2929 if( t_oop != NULL && t_oop->is_known_instance_field() && 2930 t_oop->offset() != Type::OffsetBot && 2931 t_oop->offset() != Type::OffsetTop) { 2932 eliminate = true; 2933 } 2934 } 2935 } else if (opc == Op_MemBarRelease) { 2936 // Final field stores. 2937 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 2938 if ((alloc != NULL) && alloc->is_Allocate() && 2939 alloc->as_Allocate()->_is_non_escaping) { 2940 // The allocated object does not escape. 2941 eliminate = true; 2942 } 2943 } 2944 if (eliminate) { 2945 // Replace MemBar projections by its inputs. 2946 PhaseIterGVN* igvn = phase->is_IterGVN(); 2947 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2948 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2949 // Must return either the original node (now dead) or a new node 2950 // (Do not return a top here, since that would break the uniqueness of top.) 2951 return new ConINode(TypeInt::ZERO); 2952 } 2953 } 2954 return progress ? this : NULL; 2955 } 2956 2957 //------------------------------Value------------------------------------------ 2958 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2959 if( !in(0) ) return Type::TOP; 2960 if( phase->type(in(0)) == Type::TOP ) 2961 return Type::TOP; 2962 return TypeTuple::MEMBAR; 2963 } 2964 2965 //------------------------------match------------------------------------------ 2966 // Construct projections for memory. 2967 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2968 switch (proj->_con) { 2969 case TypeFunc::Control: 2970 case TypeFunc::Memory: 2971 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2972 } 2973 ShouldNotReachHere(); 2974 return NULL; 2975 } 2976 2977 //===========================InitializeNode==================================== 2978 // SUMMARY: 2979 // This node acts as a memory barrier on raw memory, after some raw stores. 2980 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2981 // The Initialize can 'capture' suitably constrained stores as raw inits. 2982 // It can coalesce related raw stores into larger units (called 'tiles'). 2983 // It can avoid zeroing new storage for memory units which have raw inits. 2984 // At macro-expansion, it is marked 'complete', and does not optimize further. 2985 // 2986 // EXAMPLE: 2987 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2988 // ctl = incoming control; mem* = incoming memory 2989 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 2990 // First allocate uninitialized memory and fill in the header: 2991 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 2992 // ctl := alloc.Control; mem* := alloc.Memory* 2993 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 2994 // Then initialize to zero the non-header parts of the raw memory block: 2995 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2996 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2997 // After the initialize node executes, the object is ready for service: 2998 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2999 // Suppose its body is immediately initialized as {1,2}: 3000 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3001 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3002 // mem.SLICE(#short[*]) := store2 3003 // 3004 // DETAILS: 3005 // An InitializeNode collects and isolates object initialization after 3006 // an AllocateNode and before the next possible safepoint. As a 3007 // memory barrier (MemBarNode), it keeps critical stores from drifting 3008 // down past any safepoint or any publication of the allocation. 3009 // Before this barrier, a newly-allocated object may have uninitialized bits. 3010 // After this barrier, it may be treated as a real oop, and GC is allowed. 3011 // 3012 // The semantics of the InitializeNode include an implicit zeroing of 3013 // the new object from object header to the end of the object. 3014 // (The object header and end are determined by the AllocateNode.) 3015 // 3016 // Certain stores may be added as direct inputs to the InitializeNode. 3017 // These stores must update raw memory, and they must be to addresses 3018 // derived from the raw address produced by AllocateNode, and with 3019 // a constant offset. They must be ordered by increasing offset. 3020 // The first one is at in(RawStores), the last at in(req()-1). 3021 // Unlike most memory operations, they are not linked in a chain, 3022 // but are displayed in parallel as users of the rawmem output of 3023 // the allocation. 3024 // 3025 // (See comments in InitializeNode::capture_store, which continue 3026 // the example given above.) 3027 // 3028 // When the associated Allocate is macro-expanded, the InitializeNode 3029 // may be rewritten to optimize collected stores. A ClearArrayNode 3030 // may also be created at that point to represent any required zeroing. 3031 // The InitializeNode is then marked 'complete', prohibiting further 3032 // capturing of nearby memory operations. 3033 // 3034 // During macro-expansion, all captured initializations which store 3035 // constant values of 32 bits or smaller are coalesced (if advantageous) 3036 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3037 // initialized in fewer memory operations. Memory words which are 3038 // covered by neither tiles nor non-constant stores are pre-zeroed 3039 // by explicit stores of zero. (The code shape happens to do all 3040 // zeroing first, then all other stores, with both sequences occurring 3041 // in order of ascending offsets.) 3042 // 3043 // Alternatively, code may be inserted between an AllocateNode and its 3044 // InitializeNode, to perform arbitrary initialization of the new object. 3045 // E.g., the object copying intrinsics insert complex data transfers here. 3046 // The initialization must then be marked as 'complete' disable the 3047 // built-in zeroing semantics and the collection of initializing stores. 3048 // 3049 // While an InitializeNode is incomplete, reads from the memory state 3050 // produced by it are optimizable if they match the control edge and 3051 // new oop address associated with the allocation/initialization. 3052 // They return a stored value (if the offset matches) or else zero. 3053 // A write to the memory state, if it matches control and address, 3054 // and if it is to a constant offset, may be 'captured' by the 3055 // InitializeNode. It is cloned as a raw memory operation and rewired 3056 // inside the initialization, to the raw oop produced by the allocation. 3057 // Operations on addresses which are provably distinct (e.g., to 3058 // other AllocateNodes) are allowed to bypass the initialization. 3059 // 3060 // The effect of all this is to consolidate object initialization 3061 // (both arrays and non-arrays, both piecewise and bulk) into a 3062 // single location, where it can be optimized as a unit. 3063 // 3064 // Only stores with an offset less than TrackedInitializationLimit words 3065 // will be considered for capture by an InitializeNode. This puts a 3066 // reasonable limit on the complexity of optimized initializations. 3067 3068 //---------------------------InitializeNode------------------------------------ 3069 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3070 : _is_complete(Incomplete), _does_not_escape(false), 3071 MemBarNode(C, adr_type, rawoop) 3072 { 3073 init_class_id(Class_Initialize); 3074 3075 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3076 assert(in(RawAddress) == rawoop, "proper init"); 3077 // Note: allocation() can be NULL, for secondary initialization barriers 3078 } 3079 3080 // Since this node is not matched, it will be processed by the 3081 // register allocator. Declare that there are no constraints 3082 // on the allocation of the RawAddress edge. 3083 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3084 // This edge should be set to top, by the set_complete. But be conservative. 3085 if (idx == InitializeNode::RawAddress) 3086 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3087 return RegMask::Empty; 3088 } 3089 3090 Node* InitializeNode::memory(uint alias_idx) { 3091 Node* mem = in(Memory); 3092 if (mem->is_MergeMem()) { 3093 return mem->as_MergeMem()->memory_at(alias_idx); 3094 } else { 3095 // incoming raw memory is not split 3096 return mem; 3097 } 3098 } 3099 3100 bool InitializeNode::is_non_zero() { 3101 if (is_complete()) return false; 3102 remove_extra_zeroes(); 3103 return (req() > RawStores); 3104 } 3105 3106 void InitializeNode::set_complete(PhaseGVN* phase) { 3107 assert(!is_complete(), "caller responsibility"); 3108 _is_complete = Complete; 3109 3110 // After this node is complete, it contains a bunch of 3111 // raw-memory initializations. There is no need for 3112 // it to have anything to do with non-raw memory effects. 3113 // Therefore, tell all non-raw users to re-optimize themselves, 3114 // after skipping the memory effects of this initialization. 3115 PhaseIterGVN* igvn = phase->is_IterGVN(); 3116 if (igvn) igvn->add_users_to_worklist(this); 3117 } 3118 3119 // convenience function 3120 // return false if the init contains any stores already 3121 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3122 InitializeNode* init = initialization(); 3123 if (init == NULL || init->is_complete()) return false; 3124 init->remove_extra_zeroes(); 3125 // for now, if this allocation has already collected any inits, bail: 3126 if (init->is_non_zero()) return false; 3127 init->set_complete(phase); 3128 return true; 3129 } 3130 3131 void InitializeNode::remove_extra_zeroes() { 3132 if (req() == RawStores) return; 3133 Node* zmem = zero_memory(); 3134 uint fill = RawStores; 3135 for (uint i = fill; i < req(); i++) { 3136 Node* n = in(i); 3137 if (n->is_top() || n == zmem) continue; // skip 3138 if (fill < i) set_req(fill, n); // compact 3139 ++fill; 3140 } 3141 // delete any empty spaces created: 3142 while (fill < req()) { 3143 del_req(fill); 3144 } 3145 } 3146 3147 // Helper for remembering which stores go with which offsets. 3148 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3149 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3150 intptr_t offset = -1; 3151 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3152 phase, offset); 3153 if (base == NULL) return -1; // something is dead, 3154 if (offset < 0) return -1; // dead, dead 3155 return offset; 3156 } 3157 3158 // Helper for proving that an initialization expression is 3159 // "simple enough" to be folded into an object initialization. 3160 // Attempts to prove that a store's initial value 'n' can be captured 3161 // within the initialization without creating a vicious cycle, such as: 3162 // { Foo p = new Foo(); p.next = p; } 3163 // True for constants and parameters and small combinations thereof. 3164 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3165 if (n == NULL) return true; // (can this really happen?) 3166 if (n->is_Proj()) n = n->in(0); 3167 if (n == this) return false; // found a cycle 3168 if (n->is_Con()) return true; 3169 if (n->is_Start()) return true; // params, etc., are OK 3170 if (n->is_Root()) return true; // even better 3171 3172 Node* ctl = n->in(0); 3173 if (ctl != NULL && !ctl->is_top()) { 3174 if (ctl->is_Proj()) ctl = ctl->in(0); 3175 if (ctl == this) return false; 3176 3177 // If we already know that the enclosing memory op is pinned right after 3178 // the init, then any control flow that the store has picked up 3179 // must have preceded the init, or else be equal to the init. 3180 // Even after loop optimizations (which might change control edges) 3181 // a store is never pinned *before* the availability of its inputs. 3182 if (!MemNode::all_controls_dominate(n, this)) 3183 return false; // failed to prove a good control 3184 } 3185 3186 // Check data edges for possible dependencies on 'this'. 3187 if ((count += 1) > 20) return false; // complexity limit 3188 for (uint i = 1; i < n->req(); i++) { 3189 Node* m = n->in(i); 3190 if (m == NULL || m == n || m->is_top()) continue; 3191 uint first_i = n->find_edge(m); 3192 if (i != first_i) continue; // process duplicate edge just once 3193 if (!detect_init_independence(m, count)) { 3194 return false; 3195 } 3196 } 3197 3198 return true; 3199 } 3200 3201 // Here are all the checks a Store must pass before it can be moved into 3202 // an initialization. Returns zero if a check fails. 3203 // On success, returns the (constant) offset to which the store applies, 3204 // within the initialized memory. 3205 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3206 const int FAIL = 0; 3207 if (st->req() != MemNode::ValueIn + 1) 3208 return FAIL; // an inscrutable StoreNode (card mark?) 3209 Node* ctl = st->in(MemNode::Control); 3210 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3211 return FAIL; // must be unconditional after the initialization 3212 Node* mem = st->in(MemNode::Memory); 3213 if (!(mem->is_Proj() && mem->in(0) == this)) 3214 return FAIL; // must not be preceded by other stores 3215 Node* adr = st->in(MemNode::Address); 3216 intptr_t offset; 3217 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3218 if (alloc == NULL) 3219 return FAIL; // inscrutable address 3220 if (alloc != allocation()) 3221 return FAIL; // wrong allocation! (store needs to float up) 3222 Node* val = st->in(MemNode::ValueIn); 3223 int complexity_count = 0; 3224 if (!detect_init_independence(val, complexity_count)) 3225 return FAIL; // stored value must be 'simple enough' 3226 3227 // The Store can be captured only if nothing after the allocation 3228 // and before the Store is using the memory location that the store 3229 // overwrites. 3230 bool failed = false; 3231 // If is_complete_with_arraycopy() is true the shape of the graph is 3232 // well defined and is safe so no need for extra checks. 3233 if (!is_complete_with_arraycopy()) { 3234 // We are going to look at each use of the memory state following 3235 // the allocation to make sure nothing reads the memory that the 3236 // Store writes. 3237 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3238 int alias_idx = phase->C->get_alias_index(t_adr); 3239 ResourceMark rm; 3240 Unique_Node_List mems; 3241 mems.push(mem); 3242 Node* unique_merge = NULL; 3243 for (uint next = 0; next < mems.size(); ++next) { 3244 Node *m = mems.at(next); 3245 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3246 Node *n = m->fast_out(j); 3247 if (n->outcnt() == 0) { 3248 continue; 3249 } 3250 if (n == st) { 3251 continue; 3252 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3253 // If the control of this use is different from the control 3254 // of the Store which is right after the InitializeNode then 3255 // this node cannot be between the InitializeNode and the 3256 // Store. 3257 continue; 3258 } else if (n->is_MergeMem()) { 3259 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3260 // We can hit a MergeMemNode (that will likely go away 3261 // later) that is a direct use of the memory state 3262 // following the InitializeNode on the same slice as the 3263 // store node that we'd like to capture. We need to check 3264 // the uses of the MergeMemNode. 3265 mems.push(n); 3266 } 3267 } else if (n->is_Mem()) { 3268 Node* other_adr = n->in(MemNode::Address); 3269 if (other_adr == adr) { 3270 failed = true; 3271 break; 3272 } else { 3273 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3274 if (other_t_adr != NULL) { 3275 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3276 if (other_alias_idx == alias_idx) { 3277 // A load from the same memory slice as the store right 3278 // after the InitializeNode. We check the control of the 3279 // object/array that is loaded from. If it's the same as 3280 // the store control then we cannot capture the store. 3281 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3282 Node* base = other_adr; 3283 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); 3284 base = base->in(AddPNode::Base); 3285 if (base != NULL) { 3286 base = base->uncast(); 3287 if (base->is_Proj() && base->in(0) == alloc) { 3288 failed = true; 3289 break; 3290 } 3291 } 3292 } 3293 } 3294 } 3295 } else { 3296 failed = true; 3297 break; 3298 } 3299 } 3300 } 3301 } 3302 if (failed) { 3303 if (!can_reshape) { 3304 // We decided we couldn't capture the store during parsing. We 3305 // should try again during the next IGVN once the graph is 3306 // cleaner. 3307 phase->C->record_for_igvn(st); 3308 } 3309 return FAIL; 3310 } 3311 3312 return offset; // success 3313 } 3314 3315 // Find the captured store in(i) which corresponds to the range 3316 // [start..start+size) in the initialized object. 3317 // If there is one, return its index i. If there isn't, return the 3318 // negative of the index where it should be inserted. 3319 // Return 0 if the queried range overlaps an initialization boundary 3320 // or if dead code is encountered. 3321 // If size_in_bytes is zero, do not bother with overlap checks. 3322 int InitializeNode::captured_store_insertion_point(intptr_t start, 3323 int size_in_bytes, 3324 PhaseTransform* phase) { 3325 const int FAIL = 0, MAX_STORE = BytesPerLong; 3326 3327 if (is_complete()) 3328 return FAIL; // arraycopy got here first; punt 3329 3330 assert(allocation() != NULL, "must be present"); 3331 3332 // no negatives, no header fields: 3333 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3334 3335 // after a certain size, we bail out on tracking all the stores: 3336 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3337 if (start >= ti_limit) return FAIL; 3338 3339 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3340 if (i >= limit) return -(int)i; // not found; here is where to put it 3341 3342 Node* st = in(i); 3343 intptr_t st_off = get_store_offset(st, phase); 3344 if (st_off < 0) { 3345 if (st != zero_memory()) { 3346 return FAIL; // bail out if there is dead garbage 3347 } 3348 } else if (st_off > start) { 3349 // ...we are done, since stores are ordered 3350 if (st_off < start + size_in_bytes) { 3351 return FAIL; // the next store overlaps 3352 } 3353 return -(int)i; // not found; here is where to put it 3354 } else if (st_off < start) { 3355 if (size_in_bytes != 0 && 3356 start < st_off + MAX_STORE && 3357 start < st_off + st->as_Store()->memory_size()) { 3358 return FAIL; // the previous store overlaps 3359 } 3360 } else { 3361 if (size_in_bytes != 0 && 3362 st->as_Store()->memory_size() != size_in_bytes) { 3363 return FAIL; // mismatched store size 3364 } 3365 return i; 3366 } 3367 3368 ++i; 3369 } 3370 } 3371 3372 // Look for a captured store which initializes at the offset 'start' 3373 // with the given size. If there is no such store, and no other 3374 // initialization interferes, then return zero_memory (the memory 3375 // projection of the AllocateNode). 3376 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3377 PhaseTransform* phase) { 3378 assert(stores_are_sane(phase), ""); 3379 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3380 if (i == 0) { 3381 return NULL; // something is dead 3382 } else if (i < 0) { 3383 return zero_memory(); // just primordial zero bits here 3384 } else { 3385 Node* st = in(i); // here is the store at this position 3386 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3387 return st; 3388 } 3389 } 3390 3391 // Create, as a raw pointer, an address within my new object at 'offset'. 3392 Node* InitializeNode::make_raw_address(intptr_t offset, 3393 PhaseTransform* phase) { 3394 Node* addr = in(RawAddress); 3395 if (offset != 0) { 3396 Compile* C = phase->C; 3397 addr = phase->transform( new AddPNode(C->top(), addr, 3398 phase->MakeConX(offset)) ); 3399 } 3400 return addr; 3401 } 3402 3403 // Clone the given store, converting it into a raw store 3404 // initializing a field or element of my new object. 3405 // Caller is responsible for retiring the original store, 3406 // with subsume_node or the like. 3407 // 3408 // From the example above InitializeNode::InitializeNode, 3409 // here are the old stores to be captured: 3410 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3411 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3412 // 3413 // Here is the changed code; note the extra edges on init: 3414 // alloc = (Allocate ...) 3415 // rawoop = alloc.RawAddress 3416 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3417 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3418 // init = (Initialize alloc.Control alloc.Memory rawoop 3419 // rawstore1 rawstore2) 3420 // 3421 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3422 PhaseTransform* phase, bool can_reshape) { 3423 assert(stores_are_sane(phase), ""); 3424 3425 if (start < 0) return NULL; 3426 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3427 3428 Compile* C = phase->C; 3429 int size_in_bytes = st->memory_size(); 3430 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3431 if (i == 0) return NULL; // bail out 3432 Node* prev_mem = NULL; // raw memory for the captured store 3433 if (i > 0) { 3434 prev_mem = in(i); // there is a pre-existing store under this one 3435 set_req(i, C->top()); // temporarily disconnect it 3436 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3437 } else { 3438 i = -i; // no pre-existing store 3439 prev_mem = zero_memory(); // a slice of the newly allocated object 3440 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3441 set_req(--i, C->top()); // reuse this edge; it has been folded away 3442 else 3443 ins_req(i, C->top()); // build a new edge 3444 } 3445 Node* new_st = st->clone(); 3446 new_st->set_req(MemNode::Control, in(Control)); 3447 new_st->set_req(MemNode::Memory, prev_mem); 3448 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3449 new_st = phase->transform(new_st); 3450 3451 // At this point, new_st might have swallowed a pre-existing store 3452 // at the same offset, or perhaps new_st might have disappeared, 3453 // if it redundantly stored the same value (or zero to fresh memory). 3454 3455 // In any case, wire it in: 3456 phase->igvn_rehash_node_delayed(this); 3457 set_req(i, new_st); 3458 3459 // The caller may now kill the old guy. 3460 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3461 assert(check_st == new_st || check_st == NULL, "must be findable"); 3462 assert(!is_complete(), ""); 3463 return new_st; 3464 } 3465 3466 static bool store_constant(jlong* tiles, int num_tiles, 3467 intptr_t st_off, int st_size, 3468 jlong con) { 3469 if ((st_off & (st_size-1)) != 0) 3470 return false; // strange store offset (assume size==2**N) 3471 address addr = (address)tiles + st_off; 3472 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3473 switch (st_size) { 3474 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3475 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3476 case sizeof(jint): *(jint*) addr = (jint) con; break; 3477 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3478 default: return false; // strange store size (detect size!=2**N here) 3479 } 3480 return true; // return success to caller 3481 } 3482 3483 // Coalesce subword constants into int constants and possibly 3484 // into long constants. The goal, if the CPU permits, 3485 // is to initialize the object with a small number of 64-bit tiles. 3486 // Also, convert floating-point constants to bit patterns. 3487 // Non-constants are not relevant to this pass. 3488 // 3489 // In terms of the running example on InitializeNode::InitializeNode 3490 // and InitializeNode::capture_store, here is the transformation 3491 // of rawstore1 and rawstore2 into rawstore12: 3492 // alloc = (Allocate ...) 3493 // rawoop = alloc.RawAddress 3494 // tile12 = 0x00010002 3495 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3496 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3497 // 3498 void 3499 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3500 Node* size_in_bytes, 3501 PhaseGVN* phase) { 3502 Compile* C = phase->C; 3503 3504 assert(stores_are_sane(phase), ""); 3505 // Note: After this pass, they are not completely sane, 3506 // since there may be some overlaps. 3507 3508 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3509 3510 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3511 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3512 size_limit = MIN2(size_limit, ti_limit); 3513 size_limit = align_size_up(size_limit, BytesPerLong); 3514 int num_tiles = size_limit / BytesPerLong; 3515 3516 // allocate space for the tile map: 3517 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3518 jlong tiles_buf[small_len]; 3519 Node* nodes_buf[small_len]; 3520 jlong inits_buf[small_len]; 3521 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3522 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3523 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3524 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3525 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3526 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3527 // tiles: exact bitwise model of all primitive constants 3528 // nodes: last constant-storing node subsumed into the tiles model 3529 // inits: which bytes (in each tile) are touched by any initializations 3530 3531 //// Pass A: Fill in the tile model with any relevant stores. 3532 3533 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3534 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3535 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3536 Node* zmem = zero_memory(); // initially zero memory state 3537 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3538 Node* st = in(i); 3539 intptr_t st_off = get_store_offset(st, phase); 3540 3541 // Figure out the store's offset and constant value: 3542 if (st_off < header_size) continue; //skip (ignore header) 3543 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3544 int st_size = st->as_Store()->memory_size(); 3545 if (st_off + st_size > size_limit) break; 3546 3547 // Record which bytes are touched, whether by constant or not. 3548 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3549 continue; // skip (strange store size) 3550 3551 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3552 if (!val->singleton()) continue; //skip (non-con store) 3553 BasicType type = val->basic_type(); 3554 3555 jlong con = 0; 3556 switch (type) { 3557 case T_INT: con = val->is_int()->get_con(); break; 3558 case T_LONG: con = val->is_long()->get_con(); break; 3559 case T_FLOAT: con = jint_cast(val->getf()); break; 3560 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3561 default: continue; //skip (odd store type) 3562 } 3563 3564 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3565 st->Opcode() == Op_StoreL) { 3566 continue; // This StoreL is already optimal. 3567 } 3568 3569 // Store down the constant. 3570 store_constant(tiles, num_tiles, st_off, st_size, con); 3571 3572 intptr_t j = st_off >> LogBytesPerLong; 3573 3574 if (type == T_INT && st_size == BytesPerInt 3575 && (st_off & BytesPerInt) == BytesPerInt) { 3576 jlong lcon = tiles[j]; 3577 if (!Matcher::isSimpleConstant64(lcon) && 3578 st->Opcode() == Op_StoreI) { 3579 // This StoreI is already optimal by itself. 3580 jint* intcon = (jint*) &tiles[j]; 3581 intcon[1] = 0; // undo the store_constant() 3582 3583 // If the previous store is also optimal by itself, back up and 3584 // undo the action of the previous loop iteration... if we can. 3585 // But if we can't, just let the previous half take care of itself. 3586 st = nodes[j]; 3587 st_off -= BytesPerInt; 3588 con = intcon[0]; 3589 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3590 assert(st_off >= header_size, "still ignoring header"); 3591 assert(get_store_offset(st, phase) == st_off, "must be"); 3592 assert(in(i-1) == zmem, "must be"); 3593 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3594 assert(con == tcon->is_int()->get_con(), "must be"); 3595 // Undo the effects of the previous loop trip, which swallowed st: 3596 intcon[0] = 0; // undo store_constant() 3597 set_req(i-1, st); // undo set_req(i, zmem) 3598 nodes[j] = NULL; // undo nodes[j] = st 3599 --old_subword; // undo ++old_subword 3600 } 3601 continue; // This StoreI is already optimal. 3602 } 3603 } 3604 3605 // This store is not needed. 3606 set_req(i, zmem); 3607 nodes[j] = st; // record for the moment 3608 if (st_size < BytesPerLong) // something has changed 3609 ++old_subword; // includes int/float, but who's counting... 3610 else ++old_long; 3611 } 3612 3613 if ((old_subword + old_long) == 0) 3614 return; // nothing more to do 3615 3616 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3617 // Be sure to insert them before overlapping non-constant stores. 3618 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3619 for (int j = 0; j < num_tiles; j++) { 3620 jlong con = tiles[j]; 3621 jlong init = inits[j]; 3622 if (con == 0) continue; 3623 jint con0, con1; // split the constant, address-wise 3624 jint init0, init1; // split the init map, address-wise 3625 { union { jlong con; jint intcon[2]; } u; 3626 u.con = con; 3627 con0 = u.intcon[0]; 3628 con1 = u.intcon[1]; 3629 u.con = init; 3630 init0 = u.intcon[0]; 3631 init1 = u.intcon[1]; 3632 } 3633 3634 Node* old = nodes[j]; 3635 assert(old != NULL, "need the prior store"); 3636 intptr_t offset = (j * BytesPerLong); 3637 3638 bool split = !Matcher::isSimpleConstant64(con); 3639 3640 if (offset < header_size) { 3641 assert(offset + BytesPerInt >= header_size, "second int counts"); 3642 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3643 split = true; // only the second word counts 3644 // Example: int a[] = { 42 ... } 3645 } else if (con0 == 0 && init0 == -1) { 3646 split = true; // first word is covered by full inits 3647 // Example: int a[] = { ... foo(), 42 ... } 3648 } else if (con1 == 0 && init1 == -1) { 3649 split = true; // second word is covered by full inits 3650 // Example: int a[] = { ... 42, foo() ... } 3651 } 3652 3653 // Here's a case where init0 is neither 0 nor -1: 3654 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3655 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3656 // In this case the tile is not split; it is (jlong)42. 3657 // The big tile is stored down, and then the foo() value is inserted. 3658 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3659 3660 Node* ctl = old->in(MemNode::Control); 3661 Node* adr = make_raw_address(offset, phase); 3662 const TypePtr* atp = TypeRawPtr::BOTTOM; 3663 3664 // One or two coalesced stores to plop down. 3665 Node* st[2]; 3666 intptr_t off[2]; 3667 int nst = 0; 3668 if (!split) { 3669 ++new_long; 3670 off[nst] = offset; 3671 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3672 phase->longcon(con), T_LONG, MemNode::unordered); 3673 } else { 3674 // Omit either if it is a zero. 3675 if (con0 != 0) { 3676 ++new_int; 3677 off[nst] = offset; 3678 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3679 phase->intcon(con0), T_INT, MemNode::unordered); 3680 } 3681 if (con1 != 0) { 3682 ++new_int; 3683 offset += BytesPerInt; 3684 adr = make_raw_address(offset, phase); 3685 off[nst] = offset; 3686 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3687 phase->intcon(con1), T_INT, MemNode::unordered); 3688 } 3689 } 3690 3691 // Insert second store first, then the first before the second. 3692 // Insert each one just before any overlapping non-constant stores. 3693 while (nst > 0) { 3694 Node* st1 = st[--nst]; 3695 C->copy_node_notes_to(st1, old); 3696 st1 = phase->transform(st1); 3697 offset = off[nst]; 3698 assert(offset >= header_size, "do not smash header"); 3699 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3700 guarantee(ins_idx != 0, "must re-insert constant store"); 3701 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3702 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3703 set_req(--ins_idx, st1); 3704 else 3705 ins_req(ins_idx, st1); 3706 } 3707 } 3708 3709 if (PrintCompilation && WizardMode) 3710 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3711 old_subword, old_long, new_int, new_long); 3712 if (C->log() != NULL) 3713 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3714 old_subword, old_long, new_int, new_long); 3715 3716 // Clean up any remaining occurrences of zmem: 3717 remove_extra_zeroes(); 3718 } 3719 3720 // Explore forward from in(start) to find the first fully initialized 3721 // word, and return its offset. Skip groups of subword stores which 3722 // together initialize full words. If in(start) is itself part of a 3723 // fully initialized word, return the offset of in(start). If there 3724 // are no following full-word stores, or if something is fishy, return 3725 // a negative value. 3726 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3727 int int_map = 0; 3728 intptr_t int_map_off = 0; 3729 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3730 3731 for (uint i = start, limit = req(); i < limit; i++) { 3732 Node* st = in(i); 3733 3734 intptr_t st_off = get_store_offset(st, phase); 3735 if (st_off < 0) break; // return conservative answer 3736 3737 int st_size = st->as_Store()->memory_size(); 3738 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3739 return st_off; // we found a complete word init 3740 } 3741 3742 // update the map: 3743 3744 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3745 if (this_int_off != int_map_off) { 3746 // reset the map: 3747 int_map = 0; 3748 int_map_off = this_int_off; 3749 } 3750 3751 int subword_off = st_off - this_int_off; 3752 int_map |= right_n_bits(st_size) << subword_off; 3753 if ((int_map & FULL_MAP) == FULL_MAP) { 3754 return this_int_off; // we found a complete word init 3755 } 3756 3757 // Did this store hit or cross the word boundary? 3758 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3759 if (next_int_off == this_int_off + BytesPerInt) { 3760 // We passed the current int, without fully initializing it. 3761 int_map_off = next_int_off; 3762 int_map >>= BytesPerInt; 3763 } else if (next_int_off > this_int_off + BytesPerInt) { 3764 // We passed the current and next int. 3765 return this_int_off + BytesPerInt; 3766 } 3767 } 3768 3769 return -1; 3770 } 3771 3772 3773 // Called when the associated AllocateNode is expanded into CFG. 3774 // At this point, we may perform additional optimizations. 3775 // Linearize the stores by ascending offset, to make memory 3776 // activity as coherent as possible. 3777 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3778 intptr_t header_size, 3779 Node* size_in_bytes, 3780 PhaseGVN* phase) { 3781 assert(!is_complete(), "not already complete"); 3782 assert(stores_are_sane(phase), ""); 3783 assert(allocation() != NULL, "must be present"); 3784 3785 remove_extra_zeroes(); 3786 3787 if (ReduceFieldZeroing || ReduceBulkZeroing) 3788 // reduce instruction count for common initialization patterns 3789 coalesce_subword_stores(header_size, size_in_bytes, phase); 3790 3791 Node* zmem = zero_memory(); // initially zero memory state 3792 Node* inits = zmem; // accumulating a linearized chain of inits 3793 #ifdef ASSERT 3794 intptr_t first_offset = allocation()->minimum_header_size(); 3795 intptr_t last_init_off = first_offset; // previous init offset 3796 intptr_t last_init_end = first_offset; // previous init offset+size 3797 intptr_t last_tile_end = first_offset; // previous tile offset+size 3798 #endif 3799 intptr_t zeroes_done = header_size; 3800 3801 bool do_zeroing = true; // we might give up if inits are very sparse 3802 int big_init_gaps = 0; // how many large gaps have we seen? 3803 3804 if (ZeroTLAB) do_zeroing = false; 3805 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3806 3807 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3808 Node* st = in(i); 3809 intptr_t st_off = get_store_offset(st, phase); 3810 if (st_off < 0) 3811 break; // unknown junk in the inits 3812 if (st->in(MemNode::Memory) != zmem) 3813 break; // complicated store chains somehow in list 3814 3815 int st_size = st->as_Store()->memory_size(); 3816 intptr_t next_init_off = st_off + st_size; 3817 3818 if (do_zeroing && zeroes_done < next_init_off) { 3819 // See if this store needs a zero before it or under it. 3820 intptr_t zeroes_needed = st_off; 3821 3822 if (st_size < BytesPerInt) { 3823 // Look for subword stores which only partially initialize words. 3824 // If we find some, we must lay down some word-level zeroes first, 3825 // underneath the subword stores. 3826 // 3827 // Examples: 3828 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3829 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3830 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3831 // 3832 // Note: coalesce_subword_stores may have already done this, 3833 // if it was prompted by constant non-zero subword initializers. 3834 // But this case can still arise with non-constant stores. 3835 3836 intptr_t next_full_store = find_next_fullword_store(i, phase); 3837 3838 // In the examples above: 3839 // in(i) p q r s x y z 3840 // st_off 12 13 14 15 12 13 14 3841 // st_size 1 1 1 1 1 1 1 3842 // next_full_s. 12 16 16 16 16 16 16 3843 // z's_done 12 16 16 16 12 16 12 3844 // z's_needed 12 16 16 16 16 16 16 3845 // zsize 0 0 0 0 4 0 4 3846 if (next_full_store < 0) { 3847 // Conservative tack: Zero to end of current word. 3848 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3849 } else { 3850 // Zero to beginning of next fully initialized word. 3851 // Or, don't zero at all, if we are already in that word. 3852 assert(next_full_store >= zeroes_needed, "must go forward"); 3853 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3854 zeroes_needed = next_full_store; 3855 } 3856 } 3857 3858 if (zeroes_needed > zeroes_done) { 3859 intptr_t zsize = zeroes_needed - zeroes_done; 3860 // Do some incremental zeroing on rawmem, in parallel with inits. 3861 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3862 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3863 zeroes_done, zeroes_needed, 3864 phase); 3865 zeroes_done = zeroes_needed; 3866 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3867 do_zeroing = false; // leave the hole, next time 3868 } 3869 } 3870 3871 // Collect the store and move on: 3872 st->set_req(MemNode::Memory, inits); 3873 inits = st; // put it on the linearized chain 3874 set_req(i, zmem); // unhook from previous position 3875 3876 if (zeroes_done == st_off) 3877 zeroes_done = next_init_off; 3878 3879 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3880 3881 #ifdef ASSERT 3882 // Various order invariants. Weaker than stores_are_sane because 3883 // a large constant tile can be filled in by smaller non-constant stores. 3884 assert(st_off >= last_init_off, "inits do not reverse"); 3885 last_init_off = st_off; 3886 const Type* val = NULL; 3887 if (st_size >= BytesPerInt && 3888 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3889 (int)val->basic_type() < (int)T_OBJECT) { 3890 assert(st_off >= last_tile_end, "tiles do not overlap"); 3891 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3892 last_tile_end = MAX2(last_tile_end, next_init_off); 3893 } else { 3894 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3895 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3896 assert(st_off >= last_init_end, "inits do not overlap"); 3897 last_init_end = next_init_off; // it's a non-tile 3898 } 3899 #endif //ASSERT 3900 } 3901 3902 remove_extra_zeroes(); // clear out all the zmems left over 3903 add_req(inits); 3904 3905 if (!ZeroTLAB) { 3906 // If anything remains to be zeroed, zero it all now. 3907 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3908 // if it is the last unused 4 bytes of an instance, forget about it 3909 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3910 if (zeroes_done + BytesPerLong >= size_limit) { 3911 assert(allocation() != NULL, ""); 3912 if (allocation()->Opcode() == Op_Allocate) { 3913 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3914 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3915 if (zeroes_done == k->layout_helper()) 3916 zeroes_done = size_limit; 3917 } 3918 } 3919 if (zeroes_done < size_limit) { 3920 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3921 zeroes_done, size_in_bytes, phase); 3922 } 3923 } 3924 3925 set_complete(phase); 3926 return rawmem; 3927 } 3928 3929 3930 #ifdef ASSERT 3931 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3932 if (is_complete()) 3933 return true; // stores could be anything at this point 3934 assert(allocation() != NULL, "must be present"); 3935 intptr_t last_off = allocation()->minimum_header_size(); 3936 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3937 Node* st = in(i); 3938 intptr_t st_off = get_store_offset(st, phase); 3939 if (st_off < 0) continue; // ignore dead garbage 3940 if (last_off > st_off) { 3941 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 3942 this->dump(2); 3943 assert(false, "ascending store offsets"); 3944 return false; 3945 } 3946 last_off = st_off + st->as_Store()->memory_size(); 3947 } 3948 return true; 3949 } 3950 #endif //ASSERT 3951 3952 3953 3954 3955 //============================MergeMemNode===================================== 3956 // 3957 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3958 // contributing store or call operations. Each contributor provides the memory 3959 // state for a particular "alias type" (see Compile::alias_type). For example, 3960 // if a MergeMem has an input X for alias category #6, then any memory reference 3961 // to alias category #6 may use X as its memory state input, as an exact equivalent 3962 // to using the MergeMem as a whole. 3963 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3964 // 3965 // (Here, the <N> notation gives the index of the relevant adr_type.) 3966 // 3967 // In one special case (and more cases in the future), alias categories overlap. 3968 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3969 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3970 // it is exactly equivalent to that state W: 3971 // MergeMem(<Bot>: W) <==> W 3972 // 3973 // Usually, the merge has more than one input. In that case, where inputs 3974 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3975 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3976 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3977 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3978 // 3979 // A merge can take a "wide" memory state as one of its narrow inputs. 3980 // This simply means that the merge observes out only the relevant parts of 3981 // the wide input. That is, wide memory states arriving at narrow merge inputs 3982 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3983 // 3984 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3985 // and that memory slices "leak through": 3986 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3987 // 3988 // But, in such a cascade, repeated memory slices can "block the leak": 3989 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3990 // 3991 // In the last example, Y is not part of the combined memory state of the 3992 // outermost MergeMem. The system must, of course, prevent unschedulable 3993 // memory states from arising, so you can be sure that the state Y is somehow 3994 // a precursor to state Y'. 3995 // 3996 // 3997 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3998 // of each MergeMemNode array are exactly the numerical alias indexes, including 3999 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4000 // Compile::alias_type (and kin) produce and manage these indexes. 4001 // 4002 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4003 // (Note that this provides quick access to the top node inside MergeMem methods, 4004 // without the need to reach out via TLS to Compile::current.) 4005 // 4006 // As a consequence of what was just described, a MergeMem that represents a full 4007 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4008 // containing all alias categories. 4009 // 4010 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4011 // 4012 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4013 // a memory state for the alias type <N>, or else the top node, meaning that 4014 // there is no particular input for that alias type. Note that the length of 4015 // a MergeMem is variable, and may be extended at any time to accommodate new 4016 // memory states at larger alias indexes. When merges grow, they are of course 4017 // filled with "top" in the unused in() positions. 4018 // 4019 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4020 // (Top was chosen because it works smoothly with passes like GCM.) 4021 // 4022 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4023 // the type of random VM bits like TLS references.) Since it is always the 4024 // first non-Bot memory slice, some low-level loops use it to initialize an 4025 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4026 // 4027 // 4028 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4029 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4030 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4031 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4032 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4033 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4034 // 4035 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4036 // really that different from the other memory inputs. An abbreviation called 4037 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4038 // 4039 // 4040 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4041 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4042 // that "emerges though" the base memory will be marked as excluding the alias types 4043 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4044 // 4045 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4046 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4047 // 4048 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4049 // (It is currently unimplemented.) As you can see, the resulting merge is 4050 // actually a disjoint union of memory states, rather than an overlay. 4051 // 4052 4053 //------------------------------MergeMemNode----------------------------------- 4054 Node* MergeMemNode::make_empty_memory() { 4055 Node* empty_memory = (Node*) Compile::current()->top(); 4056 assert(empty_memory->is_top(), "correct sentinel identity"); 4057 return empty_memory; 4058 } 4059 4060 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4061 init_class_id(Class_MergeMem); 4062 // all inputs are nullified in Node::Node(int) 4063 // set_input(0, NULL); // no control input 4064 4065 // Initialize the edges uniformly to top, for starters. 4066 Node* empty_mem = make_empty_memory(); 4067 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4068 init_req(i,empty_mem); 4069 } 4070 assert(empty_memory() == empty_mem, ""); 4071 4072 if( new_base != NULL && new_base->is_MergeMem() ) { 4073 MergeMemNode* mdef = new_base->as_MergeMem(); 4074 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4075 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4076 mms.set_memory(mms.memory2()); 4077 } 4078 assert(base_memory() == mdef->base_memory(), ""); 4079 } else { 4080 set_base_memory(new_base); 4081 } 4082 } 4083 4084 // Make a new, untransformed MergeMem with the same base as 'mem'. 4085 // If mem is itself a MergeMem, populate the result with the same edges. 4086 MergeMemNode* MergeMemNode::make(Node* mem) { 4087 return new MergeMemNode(mem); 4088 } 4089 4090 //------------------------------cmp-------------------------------------------- 4091 uint MergeMemNode::hash() const { return NO_HASH; } 4092 uint MergeMemNode::cmp( const Node &n ) const { 4093 return (&n == this); // Always fail except on self 4094 } 4095 4096 //------------------------------Identity--------------------------------------- 4097 Node* MergeMemNode::Identity(PhaseTransform *phase) { 4098 // Identity if this merge point does not record any interesting memory 4099 // disambiguations. 4100 Node* base_mem = base_memory(); 4101 Node* empty_mem = empty_memory(); 4102 if (base_mem != empty_mem) { // Memory path is not dead? 4103 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4104 Node* mem = in(i); 4105 if (mem != empty_mem && mem != base_mem) { 4106 return this; // Many memory splits; no change 4107 } 4108 } 4109 } 4110 return base_mem; // No memory splits; ID on the one true input 4111 } 4112 4113 //------------------------------Ideal------------------------------------------ 4114 // This method is invoked recursively on chains of MergeMem nodes 4115 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4116 // Remove chain'd MergeMems 4117 // 4118 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4119 // relative to the "in(Bot)". Since we are patching both at the same time, 4120 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4121 // but rewrite each "in(i)" relative to the new "in(Bot)". 4122 Node *progress = NULL; 4123 4124 4125 Node* old_base = base_memory(); 4126 Node* empty_mem = empty_memory(); 4127 if (old_base == empty_mem) 4128 return NULL; // Dead memory path. 4129 4130 MergeMemNode* old_mbase; 4131 if (old_base != NULL && old_base->is_MergeMem()) 4132 old_mbase = old_base->as_MergeMem(); 4133 else 4134 old_mbase = NULL; 4135 Node* new_base = old_base; 4136 4137 // simplify stacked MergeMems in base memory 4138 if (old_mbase) new_base = old_mbase->base_memory(); 4139 4140 // the base memory might contribute new slices beyond my req() 4141 if (old_mbase) grow_to_match(old_mbase); 4142 4143 // Look carefully at the base node if it is a phi. 4144 PhiNode* phi_base; 4145 if (new_base != NULL && new_base->is_Phi()) 4146 phi_base = new_base->as_Phi(); 4147 else 4148 phi_base = NULL; 4149 4150 Node* phi_reg = NULL; 4151 uint phi_len = (uint)-1; 4152 if (phi_base != NULL && !phi_base->is_copy()) { 4153 // do not examine phi if degraded to a copy 4154 phi_reg = phi_base->region(); 4155 phi_len = phi_base->req(); 4156 // see if the phi is unfinished 4157 for (uint i = 1; i < phi_len; i++) { 4158 if (phi_base->in(i) == NULL) { 4159 // incomplete phi; do not look at it yet! 4160 phi_reg = NULL; 4161 phi_len = (uint)-1; 4162 break; 4163 } 4164 } 4165 } 4166 4167 // Note: We do not call verify_sparse on entry, because inputs 4168 // can normalize to the base_memory via subsume_node or similar 4169 // mechanisms. This method repairs that damage. 4170 4171 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4172 4173 // Look at each slice. 4174 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4175 Node* old_in = in(i); 4176 // calculate the old memory value 4177 Node* old_mem = old_in; 4178 if (old_mem == empty_mem) old_mem = old_base; 4179 assert(old_mem == memory_at(i), ""); 4180 4181 // maybe update (reslice) the old memory value 4182 4183 // simplify stacked MergeMems 4184 Node* new_mem = old_mem; 4185 MergeMemNode* old_mmem; 4186 if (old_mem != NULL && old_mem->is_MergeMem()) 4187 old_mmem = old_mem->as_MergeMem(); 4188 else 4189 old_mmem = NULL; 4190 if (old_mmem == this) { 4191 // This can happen if loops break up and safepoints disappear. 4192 // A merge of BotPtr (default) with a RawPtr memory derived from a 4193 // safepoint can be rewritten to a merge of the same BotPtr with 4194 // the BotPtr phi coming into the loop. If that phi disappears 4195 // also, we can end up with a self-loop of the mergemem. 4196 // In general, if loops degenerate and memory effects disappear, 4197 // a mergemem can be left looking at itself. This simply means 4198 // that the mergemem's default should be used, since there is 4199 // no longer any apparent effect on this slice. 4200 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4201 // from start. Update the input to TOP. 4202 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4203 } 4204 else if (old_mmem != NULL) { 4205 new_mem = old_mmem->memory_at(i); 4206 } 4207 // else preceding memory was not a MergeMem 4208 4209 // replace equivalent phis (unfortunately, they do not GVN together) 4210 if (new_mem != NULL && new_mem != new_base && 4211 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4212 if (new_mem->is_Phi()) { 4213 PhiNode* phi_mem = new_mem->as_Phi(); 4214 for (uint i = 1; i < phi_len; i++) { 4215 if (phi_base->in(i) != phi_mem->in(i)) { 4216 phi_mem = NULL; 4217 break; 4218 } 4219 } 4220 if (phi_mem != NULL) { 4221 // equivalent phi nodes; revert to the def 4222 new_mem = new_base; 4223 } 4224 } 4225 } 4226 4227 // maybe store down a new value 4228 Node* new_in = new_mem; 4229 if (new_in == new_base) new_in = empty_mem; 4230 4231 if (new_in != old_in) { 4232 // Warning: Do not combine this "if" with the previous "if" 4233 // A memory slice might have be be rewritten even if it is semantically 4234 // unchanged, if the base_memory value has changed. 4235 set_req(i, new_in); 4236 progress = this; // Report progress 4237 } 4238 } 4239 4240 if (new_base != old_base) { 4241 set_req(Compile::AliasIdxBot, new_base); 4242 // Don't use set_base_memory(new_base), because we need to update du. 4243 assert(base_memory() == new_base, ""); 4244 progress = this; 4245 } 4246 4247 if( base_memory() == this ) { 4248 // a self cycle indicates this memory path is dead 4249 set_req(Compile::AliasIdxBot, empty_mem); 4250 } 4251 4252 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4253 // Recursion must occur after the self cycle check above 4254 if( base_memory()->is_MergeMem() ) { 4255 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4256 Node *m = phase->transform(new_mbase); // Rollup any cycles 4257 if( m != NULL && (m->is_top() || 4258 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4259 // propagate rollup of dead cycle to self 4260 set_req(Compile::AliasIdxBot, empty_mem); 4261 } 4262 } 4263 4264 if( base_memory() == empty_mem ) { 4265 progress = this; 4266 // Cut inputs during Parse phase only. 4267 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4268 if( !can_reshape ) { 4269 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4270 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4271 } 4272 } 4273 } 4274 4275 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4276 // Check if PhiNode::Ideal's "Split phis through memory merges" 4277 // transform should be attempted. Look for this->phi->this cycle. 4278 uint merge_width = req(); 4279 if (merge_width > Compile::AliasIdxRaw) { 4280 PhiNode* phi = base_memory()->as_Phi(); 4281 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4282 if (phi->in(i) == this) { 4283 phase->is_IterGVN()->_worklist.push(phi); 4284 break; 4285 } 4286 } 4287 } 4288 } 4289 4290 assert(progress || verify_sparse(), "please, no dups of base"); 4291 return progress; 4292 } 4293 4294 //-------------------------set_base_memory------------------------------------- 4295 void MergeMemNode::set_base_memory(Node *new_base) { 4296 Node* empty_mem = empty_memory(); 4297 set_req(Compile::AliasIdxBot, new_base); 4298 assert(memory_at(req()) == new_base, "must set default memory"); 4299 // Clear out other occurrences of new_base: 4300 if (new_base != empty_mem) { 4301 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4302 if (in(i) == new_base) set_req(i, empty_mem); 4303 } 4304 } 4305 } 4306 4307 //------------------------------out_RegMask------------------------------------ 4308 const RegMask &MergeMemNode::out_RegMask() const { 4309 return RegMask::Empty; 4310 } 4311 4312 //------------------------------dump_spec-------------------------------------- 4313 #ifndef PRODUCT 4314 void MergeMemNode::dump_spec(outputStream *st) const { 4315 st->print(" {"); 4316 Node* base_mem = base_memory(); 4317 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4318 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4319 if (mem == base_mem) { st->print(" -"); continue; } 4320 st->print( " N%d:", mem->_idx ); 4321 Compile::current()->get_adr_type(i)->dump_on(st); 4322 } 4323 st->print(" }"); 4324 } 4325 #endif // !PRODUCT 4326 4327 4328 #ifdef ASSERT 4329 static bool might_be_same(Node* a, Node* b) { 4330 if (a == b) return true; 4331 if (!(a->is_Phi() || b->is_Phi())) return false; 4332 // phis shift around during optimization 4333 return true; // pretty stupid... 4334 } 4335 4336 // verify a narrow slice (either incoming or outgoing) 4337 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4338 if (!VerifyAliases) return; // don't bother to verify unless requested 4339 if (is_error_reported()) return; // muzzle asserts when debugging an error 4340 if (Node::in_dump()) return; // muzzle asserts when printing 4341 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4342 assert(n != NULL, ""); 4343 // Elide intervening MergeMem's 4344 while (n->is_MergeMem()) { 4345 n = n->as_MergeMem()->memory_at(alias_idx); 4346 } 4347 Compile* C = Compile::current(); 4348 const TypePtr* n_adr_type = n->adr_type(); 4349 if (n == m->empty_memory()) { 4350 // Implicit copy of base_memory() 4351 } else if (n_adr_type != TypePtr::BOTTOM) { 4352 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4353 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4354 } else { 4355 // A few places like make_runtime_call "know" that VM calls are narrow, 4356 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4357 bool expected_wide_mem = false; 4358 if (n == m->base_memory()) { 4359 expected_wide_mem = true; 4360 } else if (alias_idx == Compile::AliasIdxRaw || 4361 n == m->memory_at(Compile::AliasIdxRaw)) { 4362 expected_wide_mem = true; 4363 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4364 // memory can "leak through" calls on channels that 4365 // are write-once. Allow this also. 4366 expected_wide_mem = true; 4367 } 4368 assert(expected_wide_mem, "expected narrow slice replacement"); 4369 } 4370 } 4371 #else // !ASSERT 4372 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4373 #endif 4374 4375 4376 //-----------------------------memory_at--------------------------------------- 4377 Node* MergeMemNode::memory_at(uint alias_idx) const { 4378 assert(alias_idx >= Compile::AliasIdxRaw || 4379 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4380 "must avoid base_memory and AliasIdxTop"); 4381 4382 // Otherwise, it is a narrow slice. 4383 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4384 Compile *C = Compile::current(); 4385 if (is_empty_memory(n)) { 4386 // the array is sparse; empty slots are the "top" node 4387 n = base_memory(); 4388 assert(Node::in_dump() 4389 || n == NULL || n->bottom_type() == Type::TOP 4390 || n->adr_type() == NULL // address is TOP 4391 || n->adr_type() == TypePtr::BOTTOM 4392 || n->adr_type() == TypeRawPtr::BOTTOM 4393 || Compile::current()->AliasLevel() == 0, 4394 "must be a wide memory"); 4395 // AliasLevel == 0 if we are organizing the memory states manually. 4396 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4397 } else { 4398 // make sure the stored slice is sane 4399 #ifdef ASSERT 4400 if (is_error_reported() || Node::in_dump()) { 4401 } else if (might_be_same(n, base_memory())) { 4402 // Give it a pass: It is a mostly harmless repetition of the base. 4403 // This can arise normally from node subsumption during optimization. 4404 } else { 4405 verify_memory_slice(this, alias_idx, n); 4406 } 4407 #endif 4408 } 4409 return n; 4410 } 4411 4412 //---------------------------set_memory_at------------------------------------- 4413 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4414 verify_memory_slice(this, alias_idx, n); 4415 Node* empty_mem = empty_memory(); 4416 if (n == base_memory()) n = empty_mem; // collapse default 4417 uint need_req = alias_idx+1; 4418 if (req() < need_req) { 4419 if (n == empty_mem) return; // already the default, so do not grow me 4420 // grow the sparse array 4421 do { 4422 add_req(empty_mem); 4423 } while (req() < need_req); 4424 } 4425 set_req( alias_idx, n ); 4426 } 4427 4428 4429 4430 //--------------------------iteration_setup------------------------------------ 4431 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4432 if (other != NULL) { 4433 grow_to_match(other); 4434 // invariant: the finite support of mm2 is within mm->req() 4435 #ifdef ASSERT 4436 for (uint i = req(); i < other->req(); i++) { 4437 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4438 } 4439 #endif 4440 } 4441 // Replace spurious copies of base_memory by top. 4442 Node* base_mem = base_memory(); 4443 if (base_mem != NULL && !base_mem->is_top()) { 4444 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4445 if (in(i) == base_mem) 4446 set_req(i, empty_memory()); 4447 } 4448 } 4449 } 4450 4451 //---------------------------grow_to_match------------------------------------- 4452 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4453 Node* empty_mem = empty_memory(); 4454 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4455 // look for the finite support of the other memory 4456 for (uint i = other->req(); --i >= req(); ) { 4457 if (other->in(i) != empty_mem) { 4458 uint new_len = i+1; 4459 while (req() < new_len) add_req(empty_mem); 4460 break; 4461 } 4462 } 4463 } 4464 4465 //---------------------------verify_sparse------------------------------------- 4466 #ifndef PRODUCT 4467 bool MergeMemNode::verify_sparse() const { 4468 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4469 Node* base_mem = base_memory(); 4470 // The following can happen in degenerate cases, since empty==top. 4471 if (is_empty_memory(base_mem)) return true; 4472 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4473 assert(in(i) != NULL, "sane slice"); 4474 if (in(i) == base_mem) return false; // should have been the sentinel value! 4475 } 4476 return true; 4477 } 4478 4479 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4480 Node* n; 4481 n = mm->in(idx); 4482 if (mem == n) return true; // might be empty_memory() 4483 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4484 if (mem == n) return true; 4485 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4486 if (mem == n) return true; 4487 if (n == NULL) break; 4488 } 4489 return false; 4490 } 4491 #endif // !PRODUCT