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::may_modify(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::may_modify(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, "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 // While we can list what field types java.lang.String has, it is more 1681 // future-proof to handle all possible field types, anticipating future 1682 // changes and experiments in String code. 1683 Node* base = adr->in(AddPNode::Base); 1684 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1685 if (t != NULL && t->singleton()) { 1686 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1687 if (field != NULL && field->is_final()) { 1688 ciObject* string = t->const_oop(); 1689 ciConstant constant = string->as_instance()->field_value(field); 1690 // Type::make_from_constant does not handle narrow oops, so handle it here. 1691 // Everything else can use the factory method. 1692 if ((constant.basic_type() == T_ARRAY || constant.basic_type() == T_OBJECT) 1693 && adr->bottom_type()->is_ptr_to_narrowoop()) { 1694 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1695 } else { 1696 return Type::make_from_constant(constant, true); 1697 } 1698 } 1699 } 1700 } 1701 // Optimizations for constant objects 1702 ciObject* const_oop = tinst->const_oop(); 1703 if (const_oop != NULL) { 1704 // For constant Boxed value treat the target field as a compile time constant. 1705 if (tinst->is_ptr_to_boxed_value()) { 1706 return tinst->get_const_boxed_value(); 1707 } else 1708 // For constant CallSites treat the target field as a compile time constant. 1709 if (const_oop->is_call_site()) { 1710 ciCallSite* call_site = const_oop->as_call_site(); 1711 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1712 if (field != NULL && field->is_call_site_target()) { 1713 ciMethodHandle* target = call_site->get_target(); 1714 if (target != NULL) { // just in case 1715 ciConstant constant(T_OBJECT, target); 1716 const Type* t; 1717 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1718 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1719 } else { 1720 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1721 } 1722 // Add a dependence for invalidation of the optimization. 1723 if (!call_site->is_constant_call_site()) { 1724 C->dependencies()->assert_call_site_target_value(call_site, target); 1725 } 1726 return t; 1727 } 1728 } 1729 } 1730 } 1731 } else if (tp->base() == Type::KlassPtr) { 1732 assert( off != Type::OffsetBot || 1733 // arrays can be cast to Objects 1734 tp->is_klassptr()->klass()->is_java_lang_Object() || 1735 // also allow array-loading from the primary supertype 1736 // array during subtype checks 1737 Opcode() == Op_LoadKlass, 1738 "Field accesses must be precise" ); 1739 // For klass/static loads, we expect the _type to be precise 1740 } 1741 1742 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1743 if (tkls != NULL && !StressReflectiveCode) { 1744 ciKlass* klass = tkls->klass(); 1745 if (klass->is_loaded() && tkls->klass_is_exact()) { 1746 // We are loading a field from a Klass metaobject whose identity 1747 // is known at compile time (the type is "exact" or "precise"). 1748 // Check for fields we know are maintained as constants by the VM. 1749 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1750 // The field is Klass::_super_check_offset. Return its (constant) value. 1751 // (Folds up type checking code.) 1752 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1753 return TypeInt::make(klass->super_check_offset()); 1754 } 1755 // Compute index into primary_supers array 1756 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1757 // Check for overflowing; use unsigned compare to handle the negative case. 1758 if( depth < ciKlass::primary_super_limit() ) { 1759 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1760 // (Folds up type checking code.) 1761 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1762 ciKlass *ss = klass->super_of_depth(depth); 1763 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1764 } 1765 const Type* aift = load_array_final_field(tkls, klass); 1766 if (aift != NULL) return aift; 1767 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1768 // The field is Klass::_java_mirror. Return its (constant) value. 1769 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1770 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1771 return TypeInstPtr::make(klass->java_mirror()); 1772 } 1773 } 1774 1775 // We can still check if we are loading from the primary_supers array at a 1776 // shallow enough depth. Even though the klass is not exact, entries less 1777 // than or equal to its super depth are correct. 1778 if (klass->is_loaded() ) { 1779 ciType *inner = klass; 1780 while( inner->is_obj_array_klass() ) 1781 inner = inner->as_obj_array_klass()->base_element_type(); 1782 if( inner->is_instance_klass() && 1783 !inner->as_instance_klass()->flags().is_interface() ) { 1784 // Compute index into primary_supers array 1785 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1786 // Check for overflowing; use unsigned compare to handle the negative case. 1787 if( depth < ciKlass::primary_super_limit() && 1788 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1789 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1790 // (Folds up type checking code.) 1791 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1792 ciKlass *ss = klass->super_of_depth(depth); 1793 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1794 } 1795 } 1796 } 1797 1798 // If the type is enough to determine that the thing is not an array, 1799 // we can give the layout_helper a positive interval type. 1800 // This will help short-circuit some reflective code. 1801 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1802 && !klass->is_array_klass() // not directly typed as an array 1803 && !klass->is_interface() // specifically not Serializable & Cloneable 1804 && !klass->is_java_lang_Object() // not the supertype of all T[] 1805 ) { 1806 // Note: When interfaces are reliable, we can narrow the interface 1807 // test to (klass != Serializable && klass != Cloneable). 1808 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1809 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1810 // The key property of this type is that it folds up tests 1811 // for array-ness, since it proves that the layout_helper is positive. 1812 // Thus, a generic value like the basic object layout helper works fine. 1813 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1814 } 1815 } 1816 1817 // If we are loading from a freshly-allocated object, produce a zero, 1818 // if the load is provably beyond the header of the object. 1819 // (Also allow a variable load from a fresh array to produce zero.) 1820 const TypeOopPtr *tinst = tp->isa_oopptr(); 1821 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1822 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1823 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1824 Node* value = can_see_stored_value(mem,phase); 1825 if (value != NULL && value->is_Con()) { 1826 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1827 return value->bottom_type(); 1828 } 1829 } 1830 1831 if (is_instance) { 1832 // If we have an instance type and our memory input is the 1833 // programs's initial memory state, there is no matching store, 1834 // so just return a zero of the appropriate type 1835 Node *mem = in(MemNode::Memory); 1836 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1837 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1838 return Type::get_zero_type(_type->basic_type()); 1839 } 1840 } 1841 return _type; 1842 } 1843 1844 //------------------------------match_edge------------------------------------- 1845 // Do we Match on this edge index or not? Match only the address. 1846 uint LoadNode::match_edge(uint idx) const { 1847 return idx == MemNode::Address; 1848 } 1849 1850 //--------------------------LoadBNode::Ideal-------------------------------------- 1851 // 1852 // If the previous store is to the same address as this load, 1853 // and the value stored was larger than a byte, replace this load 1854 // with the value stored truncated to a byte. If no truncation is 1855 // needed, the replacement is done in LoadNode::Identity(). 1856 // 1857 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1858 Node* mem = in(MemNode::Memory); 1859 Node* value = can_see_stored_value(mem,phase); 1860 if( value && !phase->type(value)->higher_equal( _type ) ) { 1861 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1862 return new RShiftINode(result, phase->intcon(24)); 1863 } 1864 // Identity call will handle the case where truncation is not needed. 1865 return LoadNode::Ideal(phase, can_reshape); 1866 } 1867 1868 const Type* LoadBNode::Value(PhaseTransform *phase) const { 1869 Node* mem = in(MemNode::Memory); 1870 Node* value = can_see_stored_value(mem,phase); 1871 if (value != NULL && value->is_Con() && 1872 !value->bottom_type()->higher_equal(_type)) { 1873 // If the input to the store does not fit with the load's result type, 1874 // it must be truncated. We can't delay until Ideal call since 1875 // a singleton Value is needed for split_thru_phi optimization. 1876 int con = value->get_int(); 1877 return TypeInt::make((con << 24) >> 24); 1878 } 1879 return LoadNode::Value(phase); 1880 } 1881 1882 //--------------------------LoadUBNode::Ideal------------------------------------- 1883 // 1884 // If the previous store is to the same address as this load, 1885 // and the value stored was larger than a byte, replace this load 1886 // with the value stored truncated to a byte. If no truncation is 1887 // needed, the replacement is done in LoadNode::Identity(). 1888 // 1889 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1890 Node* mem = in(MemNode::Memory); 1891 Node* value = can_see_stored_value(mem, phase); 1892 if (value && !phase->type(value)->higher_equal(_type)) 1893 return new AndINode(value, phase->intcon(0xFF)); 1894 // Identity call will handle the case where truncation is not needed. 1895 return LoadNode::Ideal(phase, can_reshape); 1896 } 1897 1898 const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1899 Node* mem = in(MemNode::Memory); 1900 Node* value = can_see_stored_value(mem,phase); 1901 if (value != NULL && value->is_Con() && 1902 !value->bottom_type()->higher_equal(_type)) { 1903 // If the input to the store does not fit with the load's result type, 1904 // it must be truncated. We can't delay until Ideal call since 1905 // a singleton Value is needed for split_thru_phi optimization. 1906 int con = value->get_int(); 1907 return TypeInt::make(con & 0xFF); 1908 } 1909 return LoadNode::Value(phase); 1910 } 1911 1912 //--------------------------LoadUSNode::Ideal------------------------------------- 1913 // 1914 // If the previous store is to the same address as this load, 1915 // and the value stored was larger than a char, replace this load 1916 // with the value stored truncated to a char. If no truncation is 1917 // needed, the replacement is done in LoadNode::Identity(). 1918 // 1919 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1920 Node* mem = in(MemNode::Memory); 1921 Node* value = can_see_stored_value(mem,phase); 1922 if( value && !phase->type(value)->higher_equal( _type ) ) 1923 return new AndINode(value,phase->intcon(0xFFFF)); 1924 // Identity call will handle the case where truncation is not needed. 1925 return LoadNode::Ideal(phase, can_reshape); 1926 } 1927 1928 const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1929 Node* mem = in(MemNode::Memory); 1930 Node* value = can_see_stored_value(mem,phase); 1931 if (value != NULL && value->is_Con() && 1932 !value->bottom_type()->higher_equal(_type)) { 1933 // If the input to the store does not fit with the load's result type, 1934 // it must be truncated. We can't delay until Ideal call since 1935 // a singleton Value is needed for split_thru_phi optimization. 1936 int con = value->get_int(); 1937 return TypeInt::make(con & 0xFFFF); 1938 } 1939 return LoadNode::Value(phase); 1940 } 1941 1942 //--------------------------LoadSNode::Ideal-------------------------------------- 1943 // 1944 // If the previous store is to the same address as this load, 1945 // and the value stored was larger than a short, replace this load 1946 // with the value stored truncated to a short. If no truncation is 1947 // needed, the replacement is done in LoadNode::Identity(). 1948 // 1949 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1950 Node* mem = in(MemNode::Memory); 1951 Node* value = can_see_stored_value(mem,phase); 1952 if( value && !phase->type(value)->higher_equal( _type ) ) { 1953 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 1954 return new RShiftINode(result, phase->intcon(16)); 1955 } 1956 // Identity call will handle the case where truncation is not needed. 1957 return LoadNode::Ideal(phase, can_reshape); 1958 } 1959 1960 const Type* LoadSNode::Value(PhaseTransform *phase) const { 1961 Node* mem = in(MemNode::Memory); 1962 Node* value = can_see_stored_value(mem,phase); 1963 if (value != NULL && value->is_Con() && 1964 !value->bottom_type()->higher_equal(_type)) { 1965 // If the input to the store does not fit with the load's result type, 1966 // it must be truncated. We can't delay until Ideal call since 1967 // a singleton Value is needed for split_thru_phi optimization. 1968 int con = value->get_int(); 1969 return TypeInt::make((con << 16) >> 16); 1970 } 1971 return LoadNode::Value(phase); 1972 } 1973 1974 //============================================================================= 1975 //----------------------------LoadKlassNode::make------------------------------ 1976 // Polymorphic factory method: 1977 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 1978 // sanity check the alias category against the created node type 1979 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 1980 assert(adr_type != NULL, "expecting TypeKlassPtr"); 1981 #ifdef _LP64 1982 if (adr_type->is_ptr_to_narrowklass()) { 1983 assert(UseCompressedClassPointers, "no compressed klasses"); 1984 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 1985 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 1986 } 1987 #endif 1988 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1989 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 1990 } 1991 1992 //------------------------------Value------------------------------------------ 1993 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1994 return klass_value_common(phase); 1995 } 1996 1997 // In most cases, LoadKlassNode does not have the control input set. If the control 1998 // input is set, it must not be removed (by LoadNode::Ideal()). 1999 bool LoadKlassNode::can_remove_control() const { 2000 return false; 2001 } 2002 2003 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 2004 // Either input is TOP ==> the result is TOP 2005 const Type *t1 = phase->type( in(MemNode::Memory) ); 2006 if (t1 == Type::TOP) return Type::TOP; 2007 Node *adr = in(MemNode::Address); 2008 const Type *t2 = phase->type( adr ); 2009 if (t2 == Type::TOP) return Type::TOP; 2010 const TypePtr *tp = t2->is_ptr(); 2011 if (TypePtr::above_centerline(tp->ptr()) || 2012 tp->ptr() == TypePtr::Null) return Type::TOP; 2013 2014 // Return a more precise klass, if possible 2015 const TypeInstPtr *tinst = tp->isa_instptr(); 2016 if (tinst != NULL) { 2017 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2018 int offset = tinst->offset(); 2019 if (ik == phase->C->env()->Class_klass() 2020 && (offset == java_lang_Class::klass_offset_in_bytes() || 2021 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2022 // We are loading a special hidden field from a Class mirror object, 2023 // the field which points to the VM's Klass metaobject. 2024 ciType* t = tinst->java_mirror_type(); 2025 // java_mirror_type returns non-null for compile-time Class constants. 2026 if (t != NULL) { 2027 // constant oop => constant klass 2028 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2029 if (t->is_void()) { 2030 // We cannot create a void array. Since void is a primitive type return null 2031 // klass. Users of this result need to do a null check on the returned klass. 2032 return TypePtr::NULL_PTR; 2033 } 2034 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2035 } 2036 if (!t->is_klass()) { 2037 // a primitive Class (e.g., int.class) has NULL for a klass field 2038 return TypePtr::NULL_PTR; 2039 } 2040 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2041 return TypeKlassPtr::make(t->as_klass()); 2042 } 2043 // non-constant mirror, so we can't tell what's going on 2044 } 2045 if( !ik->is_loaded() ) 2046 return _type; // Bail out if not loaded 2047 if (offset == oopDesc::klass_offset_in_bytes()) { 2048 if (tinst->klass_is_exact()) { 2049 return TypeKlassPtr::make(ik); 2050 } 2051 // See if we can become precise: no subklasses and no interface 2052 // (Note: We need to support verified interfaces.) 2053 if (!ik->is_interface() && !ik->has_subklass()) { 2054 //assert(!UseExactTypes, "this code should be useless with exact types"); 2055 // Add a dependence; if any subclass added we need to recompile 2056 if (!ik->is_final()) { 2057 // %%% should use stronger assert_unique_concrete_subtype instead 2058 phase->C->dependencies()->assert_leaf_type(ik); 2059 } 2060 // Return precise klass 2061 return TypeKlassPtr::make(ik); 2062 } 2063 2064 // Return root of possible klass 2065 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2066 } 2067 } 2068 2069 // Check for loading klass from an array 2070 const TypeAryPtr *tary = tp->isa_aryptr(); 2071 if( tary != NULL ) { 2072 ciKlass *tary_klass = tary->klass(); 2073 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2074 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2075 if (tary->klass_is_exact()) { 2076 return TypeKlassPtr::make(tary_klass); 2077 } 2078 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2079 // If the klass is an object array, we defer the question to the 2080 // array component klass. 2081 if( ak->is_obj_array_klass() ) { 2082 assert( ak->is_loaded(), "" ); 2083 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2084 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2085 ciInstanceKlass* ik = base_k->as_instance_klass(); 2086 // See if we can become precise: no subklasses and no interface 2087 if (!ik->is_interface() && !ik->has_subklass()) { 2088 //assert(!UseExactTypes, "this code should be useless with exact types"); 2089 // Add a dependence; if any subclass added we need to recompile 2090 if (!ik->is_final()) { 2091 phase->C->dependencies()->assert_leaf_type(ik); 2092 } 2093 // Return precise array klass 2094 return TypeKlassPtr::make(ak); 2095 } 2096 } 2097 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2098 } else { // Found a type-array? 2099 //assert(!UseExactTypes, "this code should be useless with exact types"); 2100 assert( ak->is_type_array_klass(), "" ); 2101 return TypeKlassPtr::make(ak); // These are always precise 2102 } 2103 } 2104 } 2105 2106 // Check for loading klass from an array klass 2107 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2108 if (tkls != NULL && !StressReflectiveCode) { 2109 ciKlass* klass = tkls->klass(); 2110 if( !klass->is_loaded() ) 2111 return _type; // Bail out if not loaded 2112 if( klass->is_obj_array_klass() && 2113 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2114 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2115 // // Always returning precise element type is incorrect, 2116 // // e.g., element type could be object and array may contain strings 2117 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2118 2119 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2120 // according to the element type's subclassing. 2121 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2122 } 2123 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2124 tkls->offset() == in_bytes(Klass::super_offset())) { 2125 ciKlass* sup = klass->as_instance_klass()->super(); 2126 // The field is Klass::_super. Return its (constant) value. 2127 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2128 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2129 } 2130 } 2131 2132 // Bailout case 2133 return LoadNode::Value(phase); 2134 } 2135 2136 //------------------------------Identity--------------------------------------- 2137 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2138 // Also feed through the klass in Allocate(...klass...)._klass. 2139 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2140 return klass_identity_common(phase); 2141 } 2142 2143 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2144 Node* x = LoadNode::Identity(phase); 2145 if (x != this) return x; 2146 2147 // Take apart the address into an oop and and offset. 2148 // Return 'this' if we cannot. 2149 Node* adr = in(MemNode::Address); 2150 intptr_t offset = 0; 2151 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2152 if (base == NULL) return this; 2153 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2154 if (toop == NULL) return this; 2155 2156 // We can fetch the klass directly through an AllocateNode. 2157 // This works even if the klass is not constant (clone or newArray). 2158 if (offset == oopDesc::klass_offset_in_bytes()) { 2159 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2160 if (allocated_klass != NULL) { 2161 return allocated_klass; 2162 } 2163 } 2164 2165 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2166 // See inline_native_Class_query for occurrences of these patterns. 2167 // Java Example: x.getClass().isAssignableFrom(y) 2168 // 2169 // This improves reflective code, often making the Class 2170 // mirror go completely dead. (Current exception: Class 2171 // mirrors may appear in debug info, but we could clean them out by 2172 // introducing a new debug info operator for Klass*.java_mirror). 2173 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2174 && offset == java_lang_Class::klass_offset_in_bytes()) { 2175 // We are loading a special hidden field from a Class mirror, 2176 // the field which points to its Klass or ArrayKlass metaobject. 2177 if (base->is_Load()) { 2178 Node* adr2 = base->in(MemNode::Address); 2179 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2180 if (tkls != NULL && !tkls->empty() 2181 && (tkls->klass()->is_instance_klass() || 2182 tkls->klass()->is_array_klass()) 2183 && adr2->is_AddP() 2184 ) { 2185 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2186 if (tkls->offset() == mirror_field) { 2187 return adr2->in(AddPNode::Base); 2188 } 2189 } 2190 } 2191 } 2192 2193 return this; 2194 } 2195 2196 2197 //------------------------------Value------------------------------------------ 2198 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2199 const Type *t = klass_value_common(phase); 2200 if (t == Type::TOP) 2201 return t; 2202 2203 return t->make_narrowklass(); 2204 } 2205 2206 //------------------------------Identity--------------------------------------- 2207 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2208 // Also feed through the klass in Allocate(...klass...)._klass. 2209 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2210 Node *x = klass_identity_common(phase); 2211 2212 const Type *t = phase->type( x ); 2213 if( t == Type::TOP ) return x; 2214 if( t->isa_narrowklass()) return x; 2215 assert (!t->isa_narrowoop(), "no narrow oop here"); 2216 2217 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2218 } 2219 2220 //------------------------------Value----------------------------------------- 2221 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2222 // Either input is TOP ==> the result is TOP 2223 const Type *t1 = phase->type( in(MemNode::Memory) ); 2224 if( t1 == Type::TOP ) return Type::TOP; 2225 Node *adr = in(MemNode::Address); 2226 const Type *t2 = phase->type( adr ); 2227 if( t2 == Type::TOP ) return Type::TOP; 2228 const TypePtr *tp = t2->is_ptr(); 2229 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2230 const TypeAryPtr *tap = tp->isa_aryptr(); 2231 if( !tap ) return _type; 2232 return tap->size(); 2233 } 2234 2235 //-------------------------------Ideal--------------------------------------- 2236 // Feed through the length in AllocateArray(...length...)._length. 2237 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2238 Node* p = MemNode::Ideal_common(phase, can_reshape); 2239 if (p) return (p == NodeSentinel) ? NULL : p; 2240 2241 // Take apart the address into an oop and and offset. 2242 // Return 'this' if we cannot. 2243 Node* adr = in(MemNode::Address); 2244 intptr_t offset = 0; 2245 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2246 if (base == NULL) return NULL; 2247 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2248 if (tary == NULL) return NULL; 2249 2250 // We can fetch the length directly through an AllocateArrayNode. 2251 // This works even if the length is not constant (clone or newArray). 2252 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2253 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2254 if (alloc != NULL) { 2255 Node* allocated_length = alloc->Ideal_length(); 2256 Node* len = alloc->make_ideal_length(tary, phase); 2257 if (allocated_length != len) { 2258 // New CastII improves on this. 2259 return len; 2260 } 2261 } 2262 } 2263 2264 return NULL; 2265 } 2266 2267 //------------------------------Identity--------------------------------------- 2268 // Feed through the length in AllocateArray(...length...)._length. 2269 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2270 Node* x = LoadINode::Identity(phase); 2271 if (x != this) return x; 2272 2273 // Take apart the address into an oop and and offset. 2274 // Return 'this' if we cannot. 2275 Node* adr = in(MemNode::Address); 2276 intptr_t offset = 0; 2277 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2278 if (base == NULL) return this; 2279 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2280 if (tary == NULL) return this; 2281 2282 // We can fetch the length directly through an AllocateArrayNode. 2283 // This works even if the length is not constant (clone or newArray). 2284 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2285 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2286 if (alloc != NULL) { 2287 Node* allocated_length = alloc->Ideal_length(); 2288 // Do not allow make_ideal_length to allocate a CastII node. 2289 Node* len = alloc->make_ideal_length(tary, phase, false); 2290 if (allocated_length == len) { 2291 // Return allocated_length only if it would not be improved by a CastII. 2292 return allocated_length; 2293 } 2294 } 2295 } 2296 2297 return this; 2298 2299 } 2300 2301 //============================================================================= 2302 //---------------------------StoreNode::make----------------------------------- 2303 // Polymorphic factory method: 2304 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2305 assert((mo == unordered || mo == release), "unexpected"); 2306 Compile* C = gvn.C; 2307 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2308 ctl != NULL, "raw memory operations should have control edge"); 2309 2310 switch (bt) { 2311 case T_BOOLEAN: 2312 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2313 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2314 case T_CHAR: 2315 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2316 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2317 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2318 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2319 case T_METADATA: 2320 case T_ADDRESS: 2321 case T_OBJECT: 2322 #ifdef _LP64 2323 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2324 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2325 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2326 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2327 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2328 adr->bottom_type()->isa_rawptr())) { 2329 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2330 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2331 } 2332 #endif 2333 { 2334 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2335 } 2336 } 2337 ShouldNotReachHere(); 2338 return (StoreNode*)NULL; 2339 } 2340 2341 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2342 bool require_atomic = true; 2343 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2344 } 2345 2346 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2347 bool require_atomic = true; 2348 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2349 } 2350 2351 2352 //--------------------------bottom_type---------------------------------------- 2353 const Type *StoreNode::bottom_type() const { 2354 return Type::MEMORY; 2355 } 2356 2357 //------------------------------hash------------------------------------------- 2358 uint StoreNode::hash() const { 2359 // unroll addition of interesting fields 2360 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2361 2362 // Since they are not commoned, do not hash them: 2363 return NO_HASH; 2364 } 2365 2366 //------------------------------Ideal------------------------------------------ 2367 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2368 // When a store immediately follows a relevant allocation/initialization, 2369 // try to capture it into the initialization, or hoist it above. 2370 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2371 Node* p = MemNode::Ideal_common(phase, can_reshape); 2372 if (p) return (p == NodeSentinel) ? NULL : p; 2373 2374 Node* mem = in(MemNode::Memory); 2375 Node* address = in(MemNode::Address); 2376 // Back-to-back stores to same address? Fold em up. Generally 2377 // unsafe if I have intervening uses... Also disallowed for StoreCM 2378 // since they must follow each StoreP operation. Redundant StoreCMs 2379 // are eliminated just before matching in final_graph_reshape. 2380 { 2381 Node* st = mem; 2382 // If Store 'st' has more than one use, we cannot fold 'st' away. 2383 // For example, 'st' might be the final state at a conditional 2384 // return. Or, 'st' might be used by some node which is live at 2385 // the same time 'st' is live, which might be unschedulable. So, 2386 // require exactly ONE user until such time as we clone 'mem' for 2387 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2388 // true). 2389 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2390 // Looking at a dead closed cycle of memory? 2391 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2392 assert(Opcode() == st->Opcode() || 2393 st->Opcode() == Op_StoreVector || 2394 Opcode() == Op_StoreVector || 2395 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2396 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI), // expanded ClearArrayNode 2397 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2398 2399 if (st->in(MemNode::Address)->eqv_uncast(address) && 2400 st->as_Store()->memory_size() <= this->memory_size()) { 2401 Node* use = st->raw_out(0); 2402 phase->igvn_rehash_node_delayed(use); 2403 if (can_reshape) { 2404 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2405 } else { 2406 // It's OK to do this in the parser, since DU info is always accurate, 2407 // and the parser always refers to nodes via SafePointNode maps. 2408 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2409 } 2410 return this; 2411 } 2412 st = st->in(MemNode::Memory); 2413 } 2414 } 2415 2416 2417 // Capture an unaliased, unconditional, simple store into an initializer. 2418 // Or, if it is independent of the allocation, hoist it above the allocation. 2419 if (ReduceFieldZeroing && /*can_reshape &&*/ 2420 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2421 InitializeNode* init = mem->in(0)->as_Initialize(); 2422 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2423 if (offset > 0) { 2424 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2425 // If the InitializeNode captured me, it made a raw copy of me, 2426 // and I need to disappear. 2427 if (moved != NULL) { 2428 // %%% hack to ensure that Ideal returns a new node: 2429 mem = MergeMemNode::make(mem); 2430 return mem; // fold me away 2431 } 2432 } 2433 } 2434 2435 return NULL; // No further progress 2436 } 2437 2438 //------------------------------Value----------------------------------------- 2439 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2440 // Either input is TOP ==> the result is TOP 2441 const Type *t1 = phase->type( in(MemNode::Memory) ); 2442 if( t1 == Type::TOP ) return Type::TOP; 2443 const Type *t2 = phase->type( in(MemNode::Address) ); 2444 if( t2 == Type::TOP ) return Type::TOP; 2445 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2446 if( t3 == Type::TOP ) return Type::TOP; 2447 return Type::MEMORY; 2448 } 2449 2450 //------------------------------Identity--------------------------------------- 2451 // Remove redundant stores: 2452 // Store(m, p, Load(m, p)) changes to m. 2453 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2454 Node *StoreNode::Identity( PhaseTransform *phase ) { 2455 Node* mem = in(MemNode::Memory); 2456 Node* adr = in(MemNode::Address); 2457 Node* val = in(MemNode::ValueIn); 2458 2459 // Load then Store? Then the Store is useless 2460 if (val->is_Load() && 2461 val->in(MemNode::Address)->eqv_uncast(adr) && 2462 val->in(MemNode::Memory )->eqv_uncast(mem) && 2463 val->as_Load()->store_Opcode() == Opcode()) { 2464 return mem; 2465 } 2466 2467 // Two stores in a row of the same value? 2468 if (mem->is_Store() && 2469 mem->in(MemNode::Address)->eqv_uncast(adr) && 2470 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2471 mem->Opcode() == Opcode()) { 2472 return mem; 2473 } 2474 2475 // Store of zero anywhere into a freshly-allocated object? 2476 // Then the store is useless. 2477 // (It must already have been captured by the InitializeNode.) 2478 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2479 // a newly allocated object is already all-zeroes everywhere 2480 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2481 return mem; 2482 } 2483 2484 // the store may also apply to zero-bits in an earlier object 2485 Node* prev_mem = find_previous_store(phase); 2486 // Steps (a), (b): Walk past independent stores to find an exact match. 2487 if (prev_mem != NULL) { 2488 Node* prev_val = can_see_stored_value(prev_mem, phase); 2489 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2490 // prev_val and val might differ by a cast; it would be good 2491 // to keep the more informative of the two. 2492 return mem; 2493 } 2494 } 2495 } 2496 2497 return this; 2498 } 2499 2500 //------------------------------match_edge------------------------------------- 2501 // Do we Match on this edge index or not? Match only memory & value 2502 uint StoreNode::match_edge(uint idx) const { 2503 return idx == MemNode::Address || idx == MemNode::ValueIn; 2504 } 2505 2506 //------------------------------cmp-------------------------------------------- 2507 // Do not common stores up together. They generally have to be split 2508 // back up anyways, so do not bother. 2509 uint StoreNode::cmp( const Node &n ) const { 2510 return (&n == this); // Always fail except on self 2511 } 2512 2513 //------------------------------Ideal_masked_input----------------------------- 2514 // Check for a useless mask before a partial-word store 2515 // (StoreB ... (AndI valIn conIa) ) 2516 // If (conIa & mask == mask) this simplifies to 2517 // (StoreB ... (valIn) ) 2518 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2519 Node *val = in(MemNode::ValueIn); 2520 if( val->Opcode() == Op_AndI ) { 2521 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2522 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2523 set_req(MemNode::ValueIn, val->in(1)); 2524 return this; 2525 } 2526 } 2527 return NULL; 2528 } 2529 2530 2531 //------------------------------Ideal_sign_extended_input---------------------- 2532 // Check for useless sign-extension before a partial-word store 2533 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2534 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2535 // (StoreB ... (valIn) ) 2536 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2537 Node *val = in(MemNode::ValueIn); 2538 if( val->Opcode() == Op_RShiftI ) { 2539 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2540 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2541 Node *shl = val->in(1); 2542 if( shl->Opcode() == Op_LShiftI ) { 2543 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2544 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2545 set_req(MemNode::ValueIn, shl->in(1)); 2546 return this; 2547 } 2548 } 2549 } 2550 } 2551 return NULL; 2552 } 2553 2554 //------------------------------value_never_loaded----------------------------------- 2555 // Determine whether there are any possible loads of the value stored. 2556 // For simplicity, we actually check if there are any loads from the 2557 // address stored to, not just for loads of the value stored by this node. 2558 // 2559 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2560 Node *adr = in(Address); 2561 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2562 if (adr_oop == NULL) 2563 return false; 2564 if (!adr_oop->is_known_instance_field()) 2565 return false; // if not a distinct instance, there may be aliases of the address 2566 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2567 Node *use = adr->fast_out(i); 2568 if (use->is_Load() || use->is_LoadStore()) { 2569 return false; 2570 } 2571 } 2572 return true; 2573 } 2574 2575 //============================================================================= 2576 //------------------------------Ideal------------------------------------------ 2577 // If the store is from an AND mask that leaves the low bits untouched, then 2578 // we can skip the AND operation. If the store is from a sign-extension 2579 // (a left shift, then right shift) we can skip both. 2580 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2581 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2582 if( progress != NULL ) return progress; 2583 2584 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2585 if( progress != NULL ) return progress; 2586 2587 // Finally check the default case 2588 return StoreNode::Ideal(phase, can_reshape); 2589 } 2590 2591 //============================================================================= 2592 //------------------------------Ideal------------------------------------------ 2593 // If the store is from an AND mask that leaves the low bits untouched, then 2594 // we can skip the AND operation 2595 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2596 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2597 if( progress != NULL ) return progress; 2598 2599 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2600 if( progress != NULL ) return progress; 2601 2602 // Finally check the default case 2603 return StoreNode::Ideal(phase, can_reshape); 2604 } 2605 2606 //============================================================================= 2607 //------------------------------Identity--------------------------------------- 2608 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2609 // No need to card mark when storing a null ptr 2610 Node* my_store = in(MemNode::OopStore); 2611 if (my_store->is_Store()) { 2612 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2613 if( t1 == TypePtr::NULL_PTR ) { 2614 return in(MemNode::Memory); 2615 } 2616 } 2617 return this; 2618 } 2619 2620 //============================================================================= 2621 //------------------------------Ideal--------------------------------------- 2622 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2623 Node* progress = StoreNode::Ideal(phase, can_reshape); 2624 if (progress != NULL) return progress; 2625 2626 Node* my_store = in(MemNode::OopStore); 2627 if (my_store->is_MergeMem()) { 2628 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2629 set_req(MemNode::OopStore, mem); 2630 return this; 2631 } 2632 2633 return NULL; 2634 } 2635 2636 //------------------------------Value----------------------------------------- 2637 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2638 // Either input is TOP ==> the result is TOP 2639 const Type *t = phase->type( in(MemNode::Memory) ); 2640 if( t == Type::TOP ) return Type::TOP; 2641 t = phase->type( in(MemNode::Address) ); 2642 if( t == Type::TOP ) return Type::TOP; 2643 t = phase->type( in(MemNode::ValueIn) ); 2644 if( t == Type::TOP ) return Type::TOP; 2645 // If extra input is TOP ==> the result is TOP 2646 t = phase->type( in(MemNode::OopStore) ); 2647 if( t == Type::TOP ) return Type::TOP; 2648 2649 return StoreNode::Value( phase ); 2650 } 2651 2652 2653 //============================================================================= 2654 //----------------------------------SCMemProjNode------------------------------ 2655 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2656 { 2657 return bottom_type(); 2658 } 2659 2660 //============================================================================= 2661 //----------------------------------LoadStoreNode------------------------------ 2662 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2663 : Node(required), 2664 _type(rt), 2665 _adr_type(at) 2666 { 2667 init_req(MemNode::Control, c ); 2668 init_req(MemNode::Memory , mem); 2669 init_req(MemNode::Address, adr); 2670 init_req(MemNode::ValueIn, val); 2671 init_class_id(Class_LoadStore); 2672 } 2673 2674 uint LoadStoreNode::ideal_reg() const { 2675 return _type->ideal_reg(); 2676 } 2677 2678 bool LoadStoreNode::result_not_used() const { 2679 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2680 Node *x = fast_out(i); 2681 if (x->Opcode() == Op_SCMemProj) continue; 2682 return false; 2683 } 2684 return true; 2685 } 2686 2687 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2688 2689 //============================================================================= 2690 //----------------------------------LoadStoreConditionalNode-------------------- 2691 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2692 init_req(ExpectedIn, ex ); 2693 } 2694 2695 //============================================================================= 2696 //-------------------------------adr_type-------------------------------------- 2697 const TypePtr* ClearArrayNode::adr_type() const { 2698 Node *adr = in(3); 2699 if (adr == NULL) return NULL; // node is dead 2700 return MemNode::calculate_adr_type(adr->bottom_type()); 2701 } 2702 2703 //------------------------------match_edge------------------------------------- 2704 // Do we Match on this edge index or not? Do not match memory 2705 uint ClearArrayNode::match_edge(uint idx) const { 2706 return idx > 1; 2707 } 2708 2709 //------------------------------Identity--------------------------------------- 2710 // Clearing a zero length array does nothing 2711 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2712 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2713 } 2714 2715 //------------------------------Idealize--------------------------------------- 2716 // Clearing a short array is faster with stores 2717 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2718 const int unit = BytesPerLong; 2719 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2720 if (!t) return NULL; 2721 if (!t->is_con()) return NULL; 2722 intptr_t raw_count = t->get_con(); 2723 intptr_t size = raw_count; 2724 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2725 // Clearing nothing uses the Identity call. 2726 // Negative clears are possible on dead ClearArrays 2727 // (see jck test stmt114.stmt11402.val). 2728 if (size <= 0 || size % unit != 0) return NULL; 2729 intptr_t count = size / unit; 2730 // Length too long; use fast hardware clear 2731 if (size > Matcher::init_array_short_size) return NULL; 2732 Node *mem = in(1); 2733 if( phase->type(mem)==Type::TOP ) return NULL; 2734 Node *adr = in(3); 2735 const Type* at = phase->type(adr); 2736 if( at==Type::TOP ) return NULL; 2737 const TypePtr* atp = at->isa_ptr(); 2738 // adjust atp to be the correct array element address type 2739 if (atp == NULL) atp = TypePtr::BOTTOM; 2740 else atp = atp->add_offset(Type::OffsetBot); 2741 // Get base for derived pointer purposes 2742 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2743 Node *base = adr->in(1); 2744 2745 Node *zero = phase->makecon(TypeLong::ZERO); 2746 Node *off = phase->MakeConX(BytesPerLong); 2747 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2748 count--; 2749 while( count-- ) { 2750 mem = phase->transform(mem); 2751 adr = phase->transform(new AddPNode(base,adr,off)); 2752 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2753 } 2754 return mem; 2755 } 2756 2757 //----------------------------step_through---------------------------------- 2758 // Return allocation input memory edge if it is different instance 2759 // or itself if it is the one we are looking for. 2760 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2761 Node* n = *np; 2762 assert(n->is_ClearArray(), "sanity"); 2763 intptr_t offset; 2764 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2765 // This method is called only before Allocate nodes are expanded 2766 // during macro nodes expansion. Before that ClearArray nodes are 2767 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2768 // Allocate nodes are expanded) which follows allocations. 2769 assert(alloc != NULL, "should have allocation"); 2770 if (alloc->_idx == instance_id) { 2771 // Can not bypass initialization of the instance we are looking for. 2772 return false; 2773 } 2774 // Otherwise skip it. 2775 InitializeNode* init = alloc->initialization(); 2776 if (init != NULL) 2777 *np = init->in(TypeFunc::Memory); 2778 else 2779 *np = alloc->in(TypeFunc::Memory); 2780 return true; 2781 } 2782 2783 //----------------------------clear_memory------------------------------------- 2784 // Generate code to initialize object storage to zero. 2785 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2786 intptr_t start_offset, 2787 Node* end_offset, 2788 PhaseGVN* phase) { 2789 intptr_t offset = start_offset; 2790 2791 int unit = BytesPerLong; 2792 if ((offset % unit) != 0) { 2793 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2794 adr = phase->transform(adr); 2795 const TypePtr* atp = TypeRawPtr::BOTTOM; 2796 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2797 mem = phase->transform(mem); 2798 offset += BytesPerInt; 2799 } 2800 assert((offset % unit) == 0, ""); 2801 2802 // Initialize the remaining stuff, if any, with a ClearArray. 2803 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2804 } 2805 2806 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2807 Node* start_offset, 2808 Node* end_offset, 2809 PhaseGVN* phase) { 2810 if (start_offset == end_offset) { 2811 // nothing to do 2812 return mem; 2813 } 2814 2815 int unit = BytesPerLong; 2816 Node* zbase = start_offset; 2817 Node* zend = end_offset; 2818 2819 // Scale to the unit required by the CPU: 2820 if (!Matcher::init_array_count_is_in_bytes) { 2821 Node* shift = phase->intcon(exact_log2(unit)); 2822 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2823 zend = phase->transform(new URShiftXNode(zend, shift) ); 2824 } 2825 2826 // Bulk clear double-words 2827 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2828 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2829 mem = new ClearArrayNode(ctl, mem, zsize, adr); 2830 return phase->transform(mem); 2831 } 2832 2833 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2834 intptr_t start_offset, 2835 intptr_t end_offset, 2836 PhaseGVN* phase) { 2837 if (start_offset == end_offset) { 2838 // nothing to do 2839 return mem; 2840 } 2841 2842 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2843 intptr_t done_offset = end_offset; 2844 if ((done_offset % BytesPerLong) != 0) { 2845 done_offset -= BytesPerInt; 2846 } 2847 if (done_offset > start_offset) { 2848 mem = clear_memory(ctl, mem, dest, 2849 start_offset, phase->MakeConX(done_offset), phase); 2850 } 2851 if (done_offset < end_offset) { // emit the final 32-bit store 2852 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 2853 adr = phase->transform(adr); 2854 const TypePtr* atp = TypeRawPtr::BOTTOM; 2855 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2856 mem = phase->transform(mem); 2857 done_offset += BytesPerInt; 2858 } 2859 assert(done_offset == end_offset, ""); 2860 return mem; 2861 } 2862 2863 //============================================================================= 2864 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2865 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2866 _adr_type(C->get_adr_type(alias_idx)) 2867 { 2868 init_class_id(Class_MemBar); 2869 Node* top = C->top(); 2870 init_req(TypeFunc::I_O,top); 2871 init_req(TypeFunc::FramePtr,top); 2872 init_req(TypeFunc::ReturnAdr,top); 2873 if (precedent != NULL) 2874 init_req(TypeFunc::Parms, precedent); 2875 } 2876 2877 //------------------------------cmp-------------------------------------------- 2878 uint MemBarNode::hash() const { return NO_HASH; } 2879 uint MemBarNode::cmp( const Node &n ) const { 2880 return (&n == this); // Always fail except on self 2881 } 2882 2883 //------------------------------make------------------------------------------- 2884 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2885 switch (opcode) { 2886 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 2887 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 2888 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 2889 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 2890 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 2891 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 2892 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 2893 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 2894 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 2895 case Op_Initialize: return new InitializeNode(C, atp, pn); 2896 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 2897 default: ShouldNotReachHere(); return NULL; 2898 } 2899 } 2900 2901 //------------------------------Ideal------------------------------------------ 2902 // Return a node which is more "ideal" than the current node. Strip out 2903 // control copies 2904 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2905 if (remove_dead_region(phase, can_reshape)) return this; 2906 // Don't bother trying to transform a dead node 2907 if (in(0) && in(0)->is_top()) { 2908 return NULL; 2909 } 2910 2911 bool progress = false; 2912 // Eliminate volatile MemBars for scalar replaced objects. 2913 if (can_reshape && req() == (Precedent+1)) { 2914 bool eliminate = false; 2915 int opc = Opcode(); 2916 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2917 // Volatile field loads and stores. 2918 Node* my_mem = in(MemBarNode::Precedent); 2919 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2920 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2921 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2922 // replace this Precedent (decodeN) with the Load instead. 2923 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2924 Node* load_node = my_mem->in(1); 2925 set_req(MemBarNode::Precedent, load_node); 2926 phase->is_IterGVN()->_worklist.push(my_mem); 2927 my_mem = load_node; 2928 } else { 2929 assert(my_mem->unique_out() == this, "sanity"); 2930 del_req(Precedent); 2931 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2932 my_mem = NULL; 2933 } 2934 progress = true; 2935 } 2936 if (my_mem != NULL && my_mem->is_Mem()) { 2937 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2938 // Check for scalar replaced object reference. 2939 if( t_oop != NULL && t_oop->is_known_instance_field() && 2940 t_oop->offset() != Type::OffsetBot && 2941 t_oop->offset() != Type::OffsetTop) { 2942 eliminate = true; 2943 } 2944 } 2945 } else if (opc == Op_MemBarRelease) { 2946 // Final field stores. 2947 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 2948 if ((alloc != NULL) && alloc->is_Allocate() && 2949 alloc->as_Allocate()->does_not_escape_thread()) { 2950 // The allocated object does not escape. 2951 eliminate = true; 2952 } 2953 } 2954 if (eliminate) { 2955 // Replace MemBar projections by its inputs. 2956 PhaseIterGVN* igvn = phase->is_IterGVN(); 2957 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2958 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2959 // Must return either the original node (now dead) or a new node 2960 // (Do not return a top here, since that would break the uniqueness of top.) 2961 return new ConINode(TypeInt::ZERO); 2962 } 2963 } 2964 return progress ? this : NULL; 2965 } 2966 2967 //------------------------------Value------------------------------------------ 2968 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2969 if( !in(0) ) return Type::TOP; 2970 if( phase->type(in(0)) == Type::TOP ) 2971 return Type::TOP; 2972 return TypeTuple::MEMBAR; 2973 } 2974 2975 //------------------------------match------------------------------------------ 2976 // Construct projections for memory. 2977 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2978 switch (proj->_con) { 2979 case TypeFunc::Control: 2980 case TypeFunc::Memory: 2981 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2982 } 2983 ShouldNotReachHere(); 2984 return NULL; 2985 } 2986 2987 //===========================InitializeNode==================================== 2988 // SUMMARY: 2989 // This node acts as a memory barrier on raw memory, after some raw stores. 2990 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 2991 // The Initialize can 'capture' suitably constrained stores as raw inits. 2992 // It can coalesce related raw stores into larger units (called 'tiles'). 2993 // It can avoid zeroing new storage for memory units which have raw inits. 2994 // At macro-expansion, it is marked 'complete', and does not optimize further. 2995 // 2996 // EXAMPLE: 2997 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2998 // ctl = incoming control; mem* = incoming memory 2999 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3000 // First allocate uninitialized memory and fill in the header: 3001 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3002 // ctl := alloc.Control; mem* := alloc.Memory* 3003 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3004 // Then initialize to zero the non-header parts of the raw memory block: 3005 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3006 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3007 // After the initialize node executes, the object is ready for service: 3008 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3009 // Suppose its body is immediately initialized as {1,2}: 3010 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3011 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3012 // mem.SLICE(#short[*]) := store2 3013 // 3014 // DETAILS: 3015 // An InitializeNode collects and isolates object initialization after 3016 // an AllocateNode and before the next possible safepoint. As a 3017 // memory barrier (MemBarNode), it keeps critical stores from drifting 3018 // down past any safepoint or any publication of the allocation. 3019 // Before this barrier, a newly-allocated object may have uninitialized bits. 3020 // After this barrier, it may be treated as a real oop, and GC is allowed. 3021 // 3022 // The semantics of the InitializeNode include an implicit zeroing of 3023 // the new object from object header to the end of the object. 3024 // (The object header and end are determined by the AllocateNode.) 3025 // 3026 // Certain stores may be added as direct inputs to the InitializeNode. 3027 // These stores must update raw memory, and they must be to addresses 3028 // derived from the raw address produced by AllocateNode, and with 3029 // a constant offset. They must be ordered by increasing offset. 3030 // The first one is at in(RawStores), the last at in(req()-1). 3031 // Unlike most memory operations, they are not linked in a chain, 3032 // but are displayed in parallel as users of the rawmem output of 3033 // the allocation. 3034 // 3035 // (See comments in InitializeNode::capture_store, which continue 3036 // the example given above.) 3037 // 3038 // When the associated Allocate is macro-expanded, the InitializeNode 3039 // may be rewritten to optimize collected stores. A ClearArrayNode 3040 // may also be created at that point to represent any required zeroing. 3041 // The InitializeNode is then marked 'complete', prohibiting further 3042 // capturing of nearby memory operations. 3043 // 3044 // During macro-expansion, all captured initializations which store 3045 // constant values of 32 bits or smaller are coalesced (if advantageous) 3046 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3047 // initialized in fewer memory operations. Memory words which are 3048 // covered by neither tiles nor non-constant stores are pre-zeroed 3049 // by explicit stores of zero. (The code shape happens to do all 3050 // zeroing first, then all other stores, with both sequences occurring 3051 // in order of ascending offsets.) 3052 // 3053 // Alternatively, code may be inserted between an AllocateNode and its 3054 // InitializeNode, to perform arbitrary initialization of the new object. 3055 // E.g., the object copying intrinsics insert complex data transfers here. 3056 // The initialization must then be marked as 'complete' disable the 3057 // built-in zeroing semantics and the collection of initializing stores. 3058 // 3059 // While an InitializeNode is incomplete, reads from the memory state 3060 // produced by it are optimizable if they match the control edge and 3061 // new oop address associated with the allocation/initialization. 3062 // They return a stored value (if the offset matches) or else zero. 3063 // A write to the memory state, if it matches control and address, 3064 // and if it is to a constant offset, may be 'captured' by the 3065 // InitializeNode. It is cloned as a raw memory operation and rewired 3066 // inside the initialization, to the raw oop produced by the allocation. 3067 // Operations on addresses which are provably distinct (e.g., to 3068 // other AllocateNodes) are allowed to bypass the initialization. 3069 // 3070 // The effect of all this is to consolidate object initialization 3071 // (both arrays and non-arrays, both piecewise and bulk) into a 3072 // single location, where it can be optimized as a unit. 3073 // 3074 // Only stores with an offset less than TrackedInitializationLimit words 3075 // will be considered for capture by an InitializeNode. This puts a 3076 // reasonable limit on the complexity of optimized initializations. 3077 3078 //---------------------------InitializeNode------------------------------------ 3079 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3080 : _is_complete(Incomplete), _does_not_escape(false), 3081 MemBarNode(C, adr_type, rawoop) 3082 { 3083 init_class_id(Class_Initialize); 3084 3085 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3086 assert(in(RawAddress) == rawoop, "proper init"); 3087 // Note: allocation() can be NULL, for secondary initialization barriers 3088 } 3089 3090 // Since this node is not matched, it will be processed by the 3091 // register allocator. Declare that there are no constraints 3092 // on the allocation of the RawAddress edge. 3093 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3094 // This edge should be set to top, by the set_complete. But be conservative. 3095 if (idx == InitializeNode::RawAddress) 3096 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3097 return RegMask::Empty; 3098 } 3099 3100 Node* InitializeNode::memory(uint alias_idx) { 3101 Node* mem = in(Memory); 3102 if (mem->is_MergeMem()) { 3103 return mem->as_MergeMem()->memory_at(alias_idx); 3104 } else { 3105 // incoming raw memory is not split 3106 return mem; 3107 } 3108 } 3109 3110 bool InitializeNode::is_non_zero() { 3111 if (is_complete()) return false; 3112 remove_extra_zeroes(); 3113 return (req() > RawStores); 3114 } 3115 3116 void InitializeNode::set_complete(PhaseGVN* phase) { 3117 assert(!is_complete(), "caller responsibility"); 3118 _is_complete = Complete; 3119 3120 // After this node is complete, it contains a bunch of 3121 // raw-memory initializations. There is no need for 3122 // it to have anything to do with non-raw memory effects. 3123 // Therefore, tell all non-raw users to re-optimize themselves, 3124 // after skipping the memory effects of this initialization. 3125 PhaseIterGVN* igvn = phase->is_IterGVN(); 3126 if (igvn) igvn->add_users_to_worklist(this); 3127 } 3128 3129 // convenience function 3130 // return false if the init contains any stores already 3131 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3132 InitializeNode* init = initialization(); 3133 if (init == NULL || init->is_complete()) return false; 3134 init->remove_extra_zeroes(); 3135 // for now, if this allocation has already collected any inits, bail: 3136 if (init->is_non_zero()) return false; 3137 init->set_complete(phase); 3138 return true; 3139 } 3140 3141 void InitializeNode::remove_extra_zeroes() { 3142 if (req() == RawStores) return; 3143 Node* zmem = zero_memory(); 3144 uint fill = RawStores; 3145 for (uint i = fill; i < req(); i++) { 3146 Node* n = in(i); 3147 if (n->is_top() || n == zmem) continue; // skip 3148 if (fill < i) set_req(fill, n); // compact 3149 ++fill; 3150 } 3151 // delete any empty spaces created: 3152 while (fill < req()) { 3153 del_req(fill); 3154 } 3155 } 3156 3157 // Helper for remembering which stores go with which offsets. 3158 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3159 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3160 intptr_t offset = -1; 3161 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3162 phase, offset); 3163 if (base == NULL) return -1; // something is dead, 3164 if (offset < 0) return -1; // dead, dead 3165 return offset; 3166 } 3167 3168 // Helper for proving that an initialization expression is 3169 // "simple enough" to be folded into an object initialization. 3170 // Attempts to prove that a store's initial value 'n' can be captured 3171 // within the initialization without creating a vicious cycle, such as: 3172 // { Foo p = new Foo(); p.next = p; } 3173 // True for constants and parameters and small combinations thereof. 3174 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3175 if (n == NULL) return true; // (can this really happen?) 3176 if (n->is_Proj()) n = n->in(0); 3177 if (n == this) return false; // found a cycle 3178 if (n->is_Con()) return true; 3179 if (n->is_Start()) return true; // params, etc., are OK 3180 if (n->is_Root()) return true; // even better 3181 3182 Node* ctl = n->in(0); 3183 if (ctl != NULL && !ctl->is_top()) { 3184 if (ctl->is_Proj()) ctl = ctl->in(0); 3185 if (ctl == this) return false; 3186 3187 // If we already know that the enclosing memory op is pinned right after 3188 // the init, then any control flow that the store has picked up 3189 // must have preceded the init, or else be equal to the init. 3190 // Even after loop optimizations (which might change control edges) 3191 // a store is never pinned *before* the availability of its inputs. 3192 if (!MemNode::all_controls_dominate(n, this)) 3193 return false; // failed to prove a good control 3194 } 3195 3196 // Check data edges for possible dependencies on 'this'. 3197 if ((count += 1) > 20) return false; // complexity limit 3198 for (uint i = 1; i < n->req(); i++) { 3199 Node* m = n->in(i); 3200 if (m == NULL || m == n || m->is_top()) continue; 3201 uint first_i = n->find_edge(m); 3202 if (i != first_i) continue; // process duplicate edge just once 3203 if (!detect_init_independence(m, count)) { 3204 return false; 3205 } 3206 } 3207 3208 return true; 3209 } 3210 3211 // Here are all the checks a Store must pass before it can be moved into 3212 // an initialization. Returns zero if a check fails. 3213 // On success, returns the (constant) offset to which the store applies, 3214 // within the initialized memory. 3215 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3216 const int FAIL = 0; 3217 if (st->req() != MemNode::ValueIn + 1) 3218 return FAIL; // an inscrutable StoreNode (card mark?) 3219 Node* ctl = st->in(MemNode::Control); 3220 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3221 return FAIL; // must be unconditional after the initialization 3222 Node* mem = st->in(MemNode::Memory); 3223 if (!(mem->is_Proj() && mem->in(0) == this)) 3224 return FAIL; // must not be preceded by other stores 3225 Node* adr = st->in(MemNode::Address); 3226 intptr_t offset; 3227 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3228 if (alloc == NULL) 3229 return FAIL; // inscrutable address 3230 if (alloc != allocation()) 3231 return FAIL; // wrong allocation! (store needs to float up) 3232 Node* val = st->in(MemNode::ValueIn); 3233 int complexity_count = 0; 3234 if (!detect_init_independence(val, complexity_count)) 3235 return FAIL; // stored value must be 'simple enough' 3236 3237 // The Store can be captured only if nothing after the allocation 3238 // and before the Store is using the memory location that the store 3239 // overwrites. 3240 bool failed = false; 3241 // If is_complete_with_arraycopy() is true the shape of the graph is 3242 // well defined and is safe so no need for extra checks. 3243 if (!is_complete_with_arraycopy()) { 3244 // We are going to look at each use of the memory state following 3245 // the allocation to make sure nothing reads the memory that the 3246 // Store writes. 3247 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3248 int alias_idx = phase->C->get_alias_index(t_adr); 3249 ResourceMark rm; 3250 Unique_Node_List mems; 3251 mems.push(mem); 3252 Node* unique_merge = NULL; 3253 for (uint next = 0; next < mems.size(); ++next) { 3254 Node *m = mems.at(next); 3255 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3256 Node *n = m->fast_out(j); 3257 if (n->outcnt() == 0) { 3258 continue; 3259 } 3260 if (n == st) { 3261 continue; 3262 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3263 // If the control of this use is different from the control 3264 // of the Store which is right after the InitializeNode then 3265 // this node cannot be between the InitializeNode and the 3266 // Store. 3267 continue; 3268 } else if (n->is_MergeMem()) { 3269 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3270 // We can hit a MergeMemNode (that will likely go away 3271 // later) that is a direct use of the memory state 3272 // following the InitializeNode on the same slice as the 3273 // store node that we'd like to capture. We need to check 3274 // the uses of the MergeMemNode. 3275 mems.push(n); 3276 } 3277 } else if (n->is_Mem()) { 3278 Node* other_adr = n->in(MemNode::Address); 3279 if (other_adr == adr) { 3280 failed = true; 3281 break; 3282 } else { 3283 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3284 if (other_t_adr != NULL) { 3285 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3286 if (other_alias_idx == alias_idx) { 3287 // A load from the same memory slice as the store right 3288 // after the InitializeNode. We check the control of the 3289 // object/array that is loaded from. If it's the same as 3290 // the store control then we cannot capture the store. 3291 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3292 Node* base = other_adr; 3293 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3294 base = base->in(AddPNode::Base); 3295 if (base != NULL) { 3296 base = base->uncast(); 3297 if (base->is_Proj() && base->in(0) == alloc) { 3298 failed = true; 3299 break; 3300 } 3301 } 3302 } 3303 } 3304 } 3305 } else { 3306 failed = true; 3307 break; 3308 } 3309 } 3310 } 3311 } 3312 if (failed) { 3313 if (!can_reshape) { 3314 // We decided we couldn't capture the store during parsing. We 3315 // should try again during the next IGVN once the graph is 3316 // cleaner. 3317 phase->C->record_for_igvn(st); 3318 } 3319 return FAIL; 3320 } 3321 3322 return offset; // success 3323 } 3324 3325 // Find the captured store in(i) which corresponds to the range 3326 // [start..start+size) in the initialized object. 3327 // If there is one, return its index i. If there isn't, return the 3328 // negative of the index where it should be inserted. 3329 // Return 0 if the queried range overlaps an initialization boundary 3330 // or if dead code is encountered. 3331 // If size_in_bytes is zero, do not bother with overlap checks. 3332 int InitializeNode::captured_store_insertion_point(intptr_t start, 3333 int size_in_bytes, 3334 PhaseTransform* phase) { 3335 const int FAIL = 0, MAX_STORE = BytesPerLong; 3336 3337 if (is_complete()) 3338 return FAIL; // arraycopy got here first; punt 3339 3340 assert(allocation() != NULL, "must be present"); 3341 3342 // no negatives, no header fields: 3343 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3344 3345 // after a certain size, we bail out on tracking all the stores: 3346 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3347 if (start >= ti_limit) return FAIL; 3348 3349 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3350 if (i >= limit) return -(int)i; // not found; here is where to put it 3351 3352 Node* st = in(i); 3353 intptr_t st_off = get_store_offset(st, phase); 3354 if (st_off < 0) { 3355 if (st != zero_memory()) { 3356 return FAIL; // bail out if there is dead garbage 3357 } 3358 } else if (st_off > start) { 3359 // ...we are done, since stores are ordered 3360 if (st_off < start + size_in_bytes) { 3361 return FAIL; // the next store overlaps 3362 } 3363 return -(int)i; // not found; here is where to put it 3364 } else if (st_off < start) { 3365 if (size_in_bytes != 0 && 3366 start < st_off + MAX_STORE && 3367 start < st_off + st->as_Store()->memory_size()) { 3368 return FAIL; // the previous store overlaps 3369 } 3370 } else { 3371 if (size_in_bytes != 0 && 3372 st->as_Store()->memory_size() != size_in_bytes) { 3373 return FAIL; // mismatched store size 3374 } 3375 return i; 3376 } 3377 3378 ++i; 3379 } 3380 } 3381 3382 // Look for a captured store which initializes at the offset 'start' 3383 // with the given size. If there is no such store, and no other 3384 // initialization interferes, then return zero_memory (the memory 3385 // projection of the AllocateNode). 3386 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3387 PhaseTransform* phase) { 3388 assert(stores_are_sane(phase), ""); 3389 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3390 if (i == 0) { 3391 return NULL; // something is dead 3392 } else if (i < 0) { 3393 return zero_memory(); // just primordial zero bits here 3394 } else { 3395 Node* st = in(i); // here is the store at this position 3396 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3397 return st; 3398 } 3399 } 3400 3401 // Create, as a raw pointer, an address within my new object at 'offset'. 3402 Node* InitializeNode::make_raw_address(intptr_t offset, 3403 PhaseTransform* phase) { 3404 Node* addr = in(RawAddress); 3405 if (offset != 0) { 3406 Compile* C = phase->C; 3407 addr = phase->transform( new AddPNode(C->top(), addr, 3408 phase->MakeConX(offset)) ); 3409 } 3410 return addr; 3411 } 3412 3413 // Clone the given store, converting it into a raw store 3414 // initializing a field or element of my new object. 3415 // Caller is responsible for retiring the original store, 3416 // with subsume_node or the like. 3417 // 3418 // From the example above InitializeNode::InitializeNode, 3419 // here are the old stores to be captured: 3420 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3421 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3422 // 3423 // Here is the changed code; note the extra edges on init: 3424 // alloc = (Allocate ...) 3425 // rawoop = alloc.RawAddress 3426 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3427 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3428 // init = (Initialize alloc.Control alloc.Memory rawoop 3429 // rawstore1 rawstore2) 3430 // 3431 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3432 PhaseTransform* phase, bool can_reshape) { 3433 assert(stores_are_sane(phase), ""); 3434 3435 if (start < 0) return NULL; 3436 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3437 3438 Compile* C = phase->C; 3439 int size_in_bytes = st->memory_size(); 3440 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3441 if (i == 0) return NULL; // bail out 3442 Node* prev_mem = NULL; // raw memory for the captured store 3443 if (i > 0) { 3444 prev_mem = in(i); // there is a pre-existing store under this one 3445 set_req(i, C->top()); // temporarily disconnect it 3446 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3447 } else { 3448 i = -i; // no pre-existing store 3449 prev_mem = zero_memory(); // a slice of the newly allocated object 3450 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3451 set_req(--i, C->top()); // reuse this edge; it has been folded away 3452 else 3453 ins_req(i, C->top()); // build a new edge 3454 } 3455 Node* new_st = st->clone(); 3456 new_st->set_req(MemNode::Control, in(Control)); 3457 new_st->set_req(MemNode::Memory, prev_mem); 3458 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3459 new_st = phase->transform(new_st); 3460 3461 // At this point, new_st might have swallowed a pre-existing store 3462 // at the same offset, or perhaps new_st might have disappeared, 3463 // if it redundantly stored the same value (or zero to fresh memory). 3464 3465 // In any case, wire it in: 3466 phase->igvn_rehash_node_delayed(this); 3467 set_req(i, new_st); 3468 3469 // The caller may now kill the old guy. 3470 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3471 assert(check_st == new_st || check_st == NULL, "must be findable"); 3472 assert(!is_complete(), ""); 3473 return new_st; 3474 } 3475 3476 static bool store_constant(jlong* tiles, int num_tiles, 3477 intptr_t st_off, int st_size, 3478 jlong con) { 3479 if ((st_off & (st_size-1)) != 0) 3480 return false; // strange store offset (assume size==2**N) 3481 address addr = (address)tiles + st_off; 3482 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3483 switch (st_size) { 3484 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3485 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3486 case sizeof(jint): *(jint*) addr = (jint) con; break; 3487 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3488 default: return false; // strange store size (detect size!=2**N here) 3489 } 3490 return true; // return success to caller 3491 } 3492 3493 // Coalesce subword constants into int constants and possibly 3494 // into long constants. The goal, if the CPU permits, 3495 // is to initialize the object with a small number of 64-bit tiles. 3496 // Also, convert floating-point constants to bit patterns. 3497 // Non-constants are not relevant to this pass. 3498 // 3499 // In terms of the running example on InitializeNode::InitializeNode 3500 // and InitializeNode::capture_store, here is the transformation 3501 // of rawstore1 and rawstore2 into rawstore12: 3502 // alloc = (Allocate ...) 3503 // rawoop = alloc.RawAddress 3504 // tile12 = 0x00010002 3505 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3506 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3507 // 3508 void 3509 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3510 Node* size_in_bytes, 3511 PhaseGVN* phase) { 3512 Compile* C = phase->C; 3513 3514 assert(stores_are_sane(phase), ""); 3515 // Note: After this pass, they are not completely sane, 3516 // since there may be some overlaps. 3517 3518 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3519 3520 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3521 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3522 size_limit = MIN2(size_limit, ti_limit); 3523 size_limit = align_size_up(size_limit, BytesPerLong); 3524 int num_tiles = size_limit / BytesPerLong; 3525 3526 // allocate space for the tile map: 3527 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3528 jlong tiles_buf[small_len]; 3529 Node* nodes_buf[small_len]; 3530 jlong inits_buf[small_len]; 3531 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3532 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3533 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3534 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3535 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3536 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3537 // tiles: exact bitwise model of all primitive constants 3538 // nodes: last constant-storing node subsumed into the tiles model 3539 // inits: which bytes (in each tile) are touched by any initializations 3540 3541 //// Pass A: Fill in the tile model with any relevant stores. 3542 3543 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3544 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3545 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3546 Node* zmem = zero_memory(); // initially zero memory state 3547 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3548 Node* st = in(i); 3549 intptr_t st_off = get_store_offset(st, phase); 3550 3551 // Figure out the store's offset and constant value: 3552 if (st_off < header_size) continue; //skip (ignore header) 3553 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3554 int st_size = st->as_Store()->memory_size(); 3555 if (st_off + st_size > size_limit) break; 3556 3557 // Record which bytes are touched, whether by constant or not. 3558 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3559 continue; // skip (strange store size) 3560 3561 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3562 if (!val->singleton()) continue; //skip (non-con store) 3563 BasicType type = val->basic_type(); 3564 3565 jlong con = 0; 3566 switch (type) { 3567 case T_INT: con = val->is_int()->get_con(); break; 3568 case T_LONG: con = val->is_long()->get_con(); break; 3569 case T_FLOAT: con = jint_cast(val->getf()); break; 3570 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3571 default: continue; //skip (odd store type) 3572 } 3573 3574 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3575 st->Opcode() == Op_StoreL) { 3576 continue; // This StoreL is already optimal. 3577 } 3578 3579 // Store down the constant. 3580 store_constant(tiles, num_tiles, st_off, st_size, con); 3581 3582 intptr_t j = st_off >> LogBytesPerLong; 3583 3584 if (type == T_INT && st_size == BytesPerInt 3585 && (st_off & BytesPerInt) == BytesPerInt) { 3586 jlong lcon = tiles[j]; 3587 if (!Matcher::isSimpleConstant64(lcon) && 3588 st->Opcode() == Op_StoreI) { 3589 // This StoreI is already optimal by itself. 3590 jint* intcon = (jint*) &tiles[j]; 3591 intcon[1] = 0; // undo the store_constant() 3592 3593 // If the previous store is also optimal by itself, back up and 3594 // undo the action of the previous loop iteration... if we can. 3595 // But if we can't, just let the previous half take care of itself. 3596 st = nodes[j]; 3597 st_off -= BytesPerInt; 3598 con = intcon[0]; 3599 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3600 assert(st_off >= header_size, "still ignoring header"); 3601 assert(get_store_offset(st, phase) == st_off, "must be"); 3602 assert(in(i-1) == zmem, "must be"); 3603 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3604 assert(con == tcon->is_int()->get_con(), "must be"); 3605 // Undo the effects of the previous loop trip, which swallowed st: 3606 intcon[0] = 0; // undo store_constant() 3607 set_req(i-1, st); // undo set_req(i, zmem) 3608 nodes[j] = NULL; // undo nodes[j] = st 3609 --old_subword; // undo ++old_subword 3610 } 3611 continue; // This StoreI is already optimal. 3612 } 3613 } 3614 3615 // This store is not needed. 3616 set_req(i, zmem); 3617 nodes[j] = st; // record for the moment 3618 if (st_size < BytesPerLong) // something has changed 3619 ++old_subword; // includes int/float, but who's counting... 3620 else ++old_long; 3621 } 3622 3623 if ((old_subword + old_long) == 0) 3624 return; // nothing more to do 3625 3626 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3627 // Be sure to insert them before overlapping non-constant stores. 3628 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3629 for (int j = 0; j < num_tiles; j++) { 3630 jlong con = tiles[j]; 3631 jlong init = inits[j]; 3632 if (con == 0) continue; 3633 jint con0, con1; // split the constant, address-wise 3634 jint init0, init1; // split the init map, address-wise 3635 { union { jlong con; jint intcon[2]; } u; 3636 u.con = con; 3637 con0 = u.intcon[0]; 3638 con1 = u.intcon[1]; 3639 u.con = init; 3640 init0 = u.intcon[0]; 3641 init1 = u.intcon[1]; 3642 } 3643 3644 Node* old = nodes[j]; 3645 assert(old != NULL, "need the prior store"); 3646 intptr_t offset = (j * BytesPerLong); 3647 3648 bool split = !Matcher::isSimpleConstant64(con); 3649 3650 if (offset < header_size) { 3651 assert(offset + BytesPerInt >= header_size, "second int counts"); 3652 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3653 split = true; // only the second word counts 3654 // Example: int a[] = { 42 ... } 3655 } else if (con0 == 0 && init0 == -1) { 3656 split = true; // first word is covered by full inits 3657 // Example: int a[] = { ... foo(), 42 ... } 3658 } else if (con1 == 0 && init1 == -1) { 3659 split = true; // second word is covered by full inits 3660 // Example: int a[] = { ... 42, foo() ... } 3661 } 3662 3663 // Here's a case where init0 is neither 0 nor -1: 3664 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3665 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3666 // In this case the tile is not split; it is (jlong)42. 3667 // The big tile is stored down, and then the foo() value is inserted. 3668 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3669 3670 Node* ctl = old->in(MemNode::Control); 3671 Node* adr = make_raw_address(offset, phase); 3672 const TypePtr* atp = TypeRawPtr::BOTTOM; 3673 3674 // One or two coalesced stores to plop down. 3675 Node* st[2]; 3676 intptr_t off[2]; 3677 int nst = 0; 3678 if (!split) { 3679 ++new_long; 3680 off[nst] = offset; 3681 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3682 phase->longcon(con), T_LONG, MemNode::unordered); 3683 } else { 3684 // Omit either if it is a zero. 3685 if (con0 != 0) { 3686 ++new_int; 3687 off[nst] = offset; 3688 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3689 phase->intcon(con0), T_INT, MemNode::unordered); 3690 } 3691 if (con1 != 0) { 3692 ++new_int; 3693 offset += BytesPerInt; 3694 adr = make_raw_address(offset, phase); 3695 off[nst] = offset; 3696 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3697 phase->intcon(con1), T_INT, MemNode::unordered); 3698 } 3699 } 3700 3701 // Insert second store first, then the first before the second. 3702 // Insert each one just before any overlapping non-constant stores. 3703 while (nst > 0) { 3704 Node* st1 = st[--nst]; 3705 C->copy_node_notes_to(st1, old); 3706 st1 = phase->transform(st1); 3707 offset = off[nst]; 3708 assert(offset >= header_size, "do not smash header"); 3709 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3710 guarantee(ins_idx != 0, "must re-insert constant store"); 3711 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3712 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3713 set_req(--ins_idx, st1); 3714 else 3715 ins_req(ins_idx, st1); 3716 } 3717 } 3718 3719 if (PrintCompilation && WizardMode) 3720 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3721 old_subword, old_long, new_int, new_long); 3722 if (C->log() != NULL) 3723 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3724 old_subword, old_long, new_int, new_long); 3725 3726 // Clean up any remaining occurrences of zmem: 3727 remove_extra_zeroes(); 3728 } 3729 3730 // Explore forward from in(start) to find the first fully initialized 3731 // word, and return its offset. Skip groups of subword stores which 3732 // together initialize full words. If in(start) is itself part of a 3733 // fully initialized word, return the offset of in(start). If there 3734 // are no following full-word stores, or if something is fishy, return 3735 // a negative value. 3736 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3737 int int_map = 0; 3738 intptr_t int_map_off = 0; 3739 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3740 3741 for (uint i = start, limit = req(); i < limit; i++) { 3742 Node* st = in(i); 3743 3744 intptr_t st_off = get_store_offset(st, phase); 3745 if (st_off < 0) break; // return conservative answer 3746 3747 int st_size = st->as_Store()->memory_size(); 3748 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3749 return st_off; // we found a complete word init 3750 } 3751 3752 // update the map: 3753 3754 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3755 if (this_int_off != int_map_off) { 3756 // reset the map: 3757 int_map = 0; 3758 int_map_off = this_int_off; 3759 } 3760 3761 int subword_off = st_off - this_int_off; 3762 int_map |= right_n_bits(st_size) << subword_off; 3763 if ((int_map & FULL_MAP) == FULL_MAP) { 3764 return this_int_off; // we found a complete word init 3765 } 3766 3767 // Did this store hit or cross the word boundary? 3768 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3769 if (next_int_off == this_int_off + BytesPerInt) { 3770 // We passed the current int, without fully initializing it. 3771 int_map_off = next_int_off; 3772 int_map >>= BytesPerInt; 3773 } else if (next_int_off > this_int_off + BytesPerInt) { 3774 // We passed the current and next int. 3775 return this_int_off + BytesPerInt; 3776 } 3777 } 3778 3779 return -1; 3780 } 3781 3782 3783 // Called when the associated AllocateNode is expanded into CFG. 3784 // At this point, we may perform additional optimizations. 3785 // Linearize the stores by ascending offset, to make memory 3786 // activity as coherent as possible. 3787 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3788 intptr_t header_size, 3789 Node* size_in_bytes, 3790 PhaseGVN* phase) { 3791 assert(!is_complete(), "not already complete"); 3792 assert(stores_are_sane(phase), ""); 3793 assert(allocation() != NULL, "must be present"); 3794 3795 remove_extra_zeroes(); 3796 3797 if (ReduceFieldZeroing || ReduceBulkZeroing) 3798 // reduce instruction count for common initialization patterns 3799 coalesce_subword_stores(header_size, size_in_bytes, phase); 3800 3801 Node* zmem = zero_memory(); // initially zero memory state 3802 Node* inits = zmem; // accumulating a linearized chain of inits 3803 #ifdef ASSERT 3804 intptr_t first_offset = allocation()->minimum_header_size(); 3805 intptr_t last_init_off = first_offset; // previous init offset 3806 intptr_t last_init_end = first_offset; // previous init offset+size 3807 intptr_t last_tile_end = first_offset; // previous tile offset+size 3808 #endif 3809 intptr_t zeroes_done = header_size; 3810 3811 bool do_zeroing = true; // we might give up if inits are very sparse 3812 int big_init_gaps = 0; // how many large gaps have we seen? 3813 3814 if (ZeroTLAB) do_zeroing = false; 3815 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3816 3817 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3818 Node* st = in(i); 3819 intptr_t st_off = get_store_offset(st, phase); 3820 if (st_off < 0) 3821 break; // unknown junk in the inits 3822 if (st->in(MemNode::Memory) != zmem) 3823 break; // complicated store chains somehow in list 3824 3825 int st_size = st->as_Store()->memory_size(); 3826 intptr_t next_init_off = st_off + st_size; 3827 3828 if (do_zeroing && zeroes_done < next_init_off) { 3829 // See if this store needs a zero before it or under it. 3830 intptr_t zeroes_needed = st_off; 3831 3832 if (st_size < BytesPerInt) { 3833 // Look for subword stores which only partially initialize words. 3834 // If we find some, we must lay down some word-level zeroes first, 3835 // underneath the subword stores. 3836 // 3837 // Examples: 3838 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3839 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3840 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3841 // 3842 // Note: coalesce_subword_stores may have already done this, 3843 // if it was prompted by constant non-zero subword initializers. 3844 // But this case can still arise with non-constant stores. 3845 3846 intptr_t next_full_store = find_next_fullword_store(i, phase); 3847 3848 // In the examples above: 3849 // in(i) p q r s x y z 3850 // st_off 12 13 14 15 12 13 14 3851 // st_size 1 1 1 1 1 1 1 3852 // next_full_s. 12 16 16 16 16 16 16 3853 // z's_done 12 16 16 16 12 16 12 3854 // z's_needed 12 16 16 16 16 16 16 3855 // zsize 0 0 0 0 4 0 4 3856 if (next_full_store < 0) { 3857 // Conservative tack: Zero to end of current word. 3858 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3859 } else { 3860 // Zero to beginning of next fully initialized word. 3861 // Or, don't zero at all, if we are already in that word. 3862 assert(next_full_store >= zeroes_needed, "must go forward"); 3863 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3864 zeroes_needed = next_full_store; 3865 } 3866 } 3867 3868 if (zeroes_needed > zeroes_done) { 3869 intptr_t zsize = zeroes_needed - zeroes_done; 3870 // Do some incremental zeroing on rawmem, in parallel with inits. 3871 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3872 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3873 zeroes_done, zeroes_needed, 3874 phase); 3875 zeroes_done = zeroes_needed; 3876 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3877 do_zeroing = false; // leave the hole, next time 3878 } 3879 } 3880 3881 // Collect the store and move on: 3882 st->set_req(MemNode::Memory, inits); 3883 inits = st; // put it on the linearized chain 3884 set_req(i, zmem); // unhook from previous position 3885 3886 if (zeroes_done == st_off) 3887 zeroes_done = next_init_off; 3888 3889 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3890 3891 #ifdef ASSERT 3892 // Various order invariants. Weaker than stores_are_sane because 3893 // a large constant tile can be filled in by smaller non-constant stores. 3894 assert(st_off >= last_init_off, "inits do not reverse"); 3895 last_init_off = st_off; 3896 const Type* val = NULL; 3897 if (st_size >= BytesPerInt && 3898 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3899 (int)val->basic_type() < (int)T_OBJECT) { 3900 assert(st_off >= last_tile_end, "tiles do not overlap"); 3901 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3902 last_tile_end = MAX2(last_tile_end, next_init_off); 3903 } else { 3904 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3905 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3906 assert(st_off >= last_init_end, "inits do not overlap"); 3907 last_init_end = next_init_off; // it's a non-tile 3908 } 3909 #endif //ASSERT 3910 } 3911 3912 remove_extra_zeroes(); // clear out all the zmems left over 3913 add_req(inits); 3914 3915 if (!ZeroTLAB) { 3916 // If anything remains to be zeroed, zero it all now. 3917 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3918 // if it is the last unused 4 bytes of an instance, forget about it 3919 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3920 if (zeroes_done + BytesPerLong >= size_limit) { 3921 assert(allocation() != NULL, ""); 3922 if (allocation()->Opcode() == Op_Allocate) { 3923 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3924 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3925 if (zeroes_done == k->layout_helper()) 3926 zeroes_done = size_limit; 3927 } 3928 } 3929 if (zeroes_done < size_limit) { 3930 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3931 zeroes_done, size_in_bytes, phase); 3932 } 3933 } 3934 3935 set_complete(phase); 3936 return rawmem; 3937 } 3938 3939 3940 #ifdef ASSERT 3941 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3942 if (is_complete()) 3943 return true; // stores could be anything at this point 3944 assert(allocation() != NULL, "must be present"); 3945 intptr_t last_off = allocation()->minimum_header_size(); 3946 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3947 Node* st = in(i); 3948 intptr_t st_off = get_store_offset(st, phase); 3949 if (st_off < 0) continue; // ignore dead garbage 3950 if (last_off > st_off) { 3951 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 3952 this->dump(2); 3953 assert(false, "ascending store offsets"); 3954 return false; 3955 } 3956 last_off = st_off + st->as_Store()->memory_size(); 3957 } 3958 return true; 3959 } 3960 #endif //ASSERT 3961 3962 3963 3964 3965 //============================MergeMemNode===================================== 3966 // 3967 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3968 // contributing store or call operations. Each contributor provides the memory 3969 // state for a particular "alias type" (see Compile::alias_type). For example, 3970 // if a MergeMem has an input X for alias category #6, then any memory reference 3971 // to alias category #6 may use X as its memory state input, as an exact equivalent 3972 // to using the MergeMem as a whole. 3973 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3974 // 3975 // (Here, the <N> notation gives the index of the relevant adr_type.) 3976 // 3977 // In one special case (and more cases in the future), alias categories overlap. 3978 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3979 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3980 // it is exactly equivalent to that state W: 3981 // MergeMem(<Bot>: W) <==> W 3982 // 3983 // Usually, the merge has more than one input. In that case, where inputs 3984 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3985 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3986 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3987 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3988 // 3989 // A merge can take a "wide" memory state as one of its narrow inputs. 3990 // This simply means that the merge observes out only the relevant parts of 3991 // the wide input. That is, wide memory states arriving at narrow merge inputs 3992 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3993 // 3994 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3995 // and that memory slices "leak through": 3996 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3997 // 3998 // But, in such a cascade, repeated memory slices can "block the leak": 3999 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4000 // 4001 // In the last example, Y is not part of the combined memory state of the 4002 // outermost MergeMem. The system must, of course, prevent unschedulable 4003 // memory states from arising, so you can be sure that the state Y is somehow 4004 // a precursor to state Y'. 4005 // 4006 // 4007 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4008 // of each MergeMemNode array are exactly the numerical alias indexes, including 4009 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4010 // Compile::alias_type (and kin) produce and manage these indexes. 4011 // 4012 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4013 // (Note that this provides quick access to the top node inside MergeMem methods, 4014 // without the need to reach out via TLS to Compile::current.) 4015 // 4016 // As a consequence of what was just described, a MergeMem that represents a full 4017 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4018 // containing all alias categories. 4019 // 4020 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4021 // 4022 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4023 // a memory state for the alias type <N>, or else the top node, meaning that 4024 // there is no particular input for that alias type. Note that the length of 4025 // a MergeMem is variable, and may be extended at any time to accommodate new 4026 // memory states at larger alias indexes. When merges grow, they are of course 4027 // filled with "top" in the unused in() positions. 4028 // 4029 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4030 // (Top was chosen because it works smoothly with passes like GCM.) 4031 // 4032 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4033 // the type of random VM bits like TLS references.) Since it is always the 4034 // first non-Bot memory slice, some low-level loops use it to initialize an 4035 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4036 // 4037 // 4038 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4039 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4040 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4041 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4042 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4043 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4044 // 4045 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4046 // really that different from the other memory inputs. An abbreviation called 4047 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4048 // 4049 // 4050 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4051 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4052 // that "emerges though" the base memory will be marked as excluding the alias types 4053 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4054 // 4055 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4056 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4057 // 4058 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4059 // (It is currently unimplemented.) As you can see, the resulting merge is 4060 // actually a disjoint union of memory states, rather than an overlay. 4061 // 4062 4063 //------------------------------MergeMemNode----------------------------------- 4064 Node* MergeMemNode::make_empty_memory() { 4065 Node* empty_memory = (Node*) Compile::current()->top(); 4066 assert(empty_memory->is_top(), "correct sentinel identity"); 4067 return empty_memory; 4068 } 4069 4070 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4071 init_class_id(Class_MergeMem); 4072 // all inputs are nullified in Node::Node(int) 4073 // set_input(0, NULL); // no control input 4074 4075 // Initialize the edges uniformly to top, for starters. 4076 Node* empty_mem = make_empty_memory(); 4077 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4078 init_req(i,empty_mem); 4079 } 4080 assert(empty_memory() == empty_mem, ""); 4081 4082 if( new_base != NULL && new_base->is_MergeMem() ) { 4083 MergeMemNode* mdef = new_base->as_MergeMem(); 4084 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4085 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4086 mms.set_memory(mms.memory2()); 4087 } 4088 assert(base_memory() == mdef->base_memory(), ""); 4089 } else { 4090 set_base_memory(new_base); 4091 } 4092 } 4093 4094 // Make a new, untransformed MergeMem with the same base as 'mem'. 4095 // If mem is itself a MergeMem, populate the result with the same edges. 4096 MergeMemNode* MergeMemNode::make(Node* mem) { 4097 return new MergeMemNode(mem); 4098 } 4099 4100 //------------------------------cmp-------------------------------------------- 4101 uint MergeMemNode::hash() const { return NO_HASH; } 4102 uint MergeMemNode::cmp( const Node &n ) const { 4103 return (&n == this); // Always fail except on self 4104 } 4105 4106 //------------------------------Identity--------------------------------------- 4107 Node* MergeMemNode::Identity(PhaseTransform *phase) { 4108 // Identity if this merge point does not record any interesting memory 4109 // disambiguations. 4110 Node* base_mem = base_memory(); 4111 Node* empty_mem = empty_memory(); 4112 if (base_mem != empty_mem) { // Memory path is not dead? 4113 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4114 Node* mem = in(i); 4115 if (mem != empty_mem && mem != base_mem) { 4116 return this; // Many memory splits; no change 4117 } 4118 } 4119 } 4120 return base_mem; // No memory splits; ID on the one true input 4121 } 4122 4123 //------------------------------Ideal------------------------------------------ 4124 // This method is invoked recursively on chains of MergeMem nodes 4125 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4126 // Remove chain'd MergeMems 4127 // 4128 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4129 // relative to the "in(Bot)". Since we are patching both at the same time, 4130 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4131 // but rewrite each "in(i)" relative to the new "in(Bot)". 4132 Node *progress = NULL; 4133 4134 4135 Node* old_base = base_memory(); 4136 Node* empty_mem = empty_memory(); 4137 if (old_base == empty_mem) 4138 return NULL; // Dead memory path. 4139 4140 MergeMemNode* old_mbase; 4141 if (old_base != NULL && old_base->is_MergeMem()) 4142 old_mbase = old_base->as_MergeMem(); 4143 else 4144 old_mbase = NULL; 4145 Node* new_base = old_base; 4146 4147 // simplify stacked MergeMems in base memory 4148 if (old_mbase) new_base = old_mbase->base_memory(); 4149 4150 // the base memory might contribute new slices beyond my req() 4151 if (old_mbase) grow_to_match(old_mbase); 4152 4153 // Look carefully at the base node if it is a phi. 4154 PhiNode* phi_base; 4155 if (new_base != NULL && new_base->is_Phi()) 4156 phi_base = new_base->as_Phi(); 4157 else 4158 phi_base = NULL; 4159 4160 Node* phi_reg = NULL; 4161 uint phi_len = (uint)-1; 4162 if (phi_base != NULL && !phi_base->is_copy()) { 4163 // do not examine phi if degraded to a copy 4164 phi_reg = phi_base->region(); 4165 phi_len = phi_base->req(); 4166 // see if the phi is unfinished 4167 for (uint i = 1; i < phi_len; i++) { 4168 if (phi_base->in(i) == NULL) { 4169 // incomplete phi; do not look at it yet! 4170 phi_reg = NULL; 4171 phi_len = (uint)-1; 4172 break; 4173 } 4174 } 4175 } 4176 4177 // Note: We do not call verify_sparse on entry, because inputs 4178 // can normalize to the base_memory via subsume_node or similar 4179 // mechanisms. This method repairs that damage. 4180 4181 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4182 4183 // Look at each slice. 4184 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4185 Node* old_in = in(i); 4186 // calculate the old memory value 4187 Node* old_mem = old_in; 4188 if (old_mem == empty_mem) old_mem = old_base; 4189 assert(old_mem == memory_at(i), ""); 4190 4191 // maybe update (reslice) the old memory value 4192 4193 // simplify stacked MergeMems 4194 Node* new_mem = old_mem; 4195 MergeMemNode* old_mmem; 4196 if (old_mem != NULL && old_mem->is_MergeMem()) 4197 old_mmem = old_mem->as_MergeMem(); 4198 else 4199 old_mmem = NULL; 4200 if (old_mmem == this) { 4201 // This can happen if loops break up and safepoints disappear. 4202 // A merge of BotPtr (default) with a RawPtr memory derived from a 4203 // safepoint can be rewritten to a merge of the same BotPtr with 4204 // the BotPtr phi coming into the loop. If that phi disappears 4205 // also, we can end up with a self-loop of the mergemem. 4206 // In general, if loops degenerate and memory effects disappear, 4207 // a mergemem can be left looking at itself. This simply means 4208 // that the mergemem's default should be used, since there is 4209 // no longer any apparent effect on this slice. 4210 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4211 // from start. Update the input to TOP. 4212 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4213 } 4214 else if (old_mmem != NULL) { 4215 new_mem = old_mmem->memory_at(i); 4216 } 4217 // else preceding memory was not a MergeMem 4218 4219 // replace equivalent phis (unfortunately, they do not GVN together) 4220 if (new_mem != NULL && new_mem != new_base && 4221 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4222 if (new_mem->is_Phi()) { 4223 PhiNode* phi_mem = new_mem->as_Phi(); 4224 for (uint i = 1; i < phi_len; i++) { 4225 if (phi_base->in(i) != phi_mem->in(i)) { 4226 phi_mem = NULL; 4227 break; 4228 } 4229 } 4230 if (phi_mem != NULL) { 4231 // equivalent phi nodes; revert to the def 4232 new_mem = new_base; 4233 } 4234 } 4235 } 4236 4237 // maybe store down a new value 4238 Node* new_in = new_mem; 4239 if (new_in == new_base) new_in = empty_mem; 4240 4241 if (new_in != old_in) { 4242 // Warning: Do not combine this "if" with the previous "if" 4243 // A memory slice might have be be rewritten even if it is semantically 4244 // unchanged, if the base_memory value has changed. 4245 set_req(i, new_in); 4246 progress = this; // Report progress 4247 } 4248 } 4249 4250 if (new_base != old_base) { 4251 set_req(Compile::AliasIdxBot, new_base); 4252 // Don't use set_base_memory(new_base), because we need to update du. 4253 assert(base_memory() == new_base, ""); 4254 progress = this; 4255 } 4256 4257 if( base_memory() == this ) { 4258 // a self cycle indicates this memory path is dead 4259 set_req(Compile::AliasIdxBot, empty_mem); 4260 } 4261 4262 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4263 // Recursion must occur after the self cycle check above 4264 if( base_memory()->is_MergeMem() ) { 4265 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4266 Node *m = phase->transform(new_mbase); // Rollup any cycles 4267 if( m != NULL && (m->is_top() || 4268 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4269 // propagate rollup of dead cycle to self 4270 set_req(Compile::AliasIdxBot, empty_mem); 4271 } 4272 } 4273 4274 if( base_memory() == empty_mem ) { 4275 progress = this; 4276 // Cut inputs during Parse phase only. 4277 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4278 if( !can_reshape ) { 4279 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4280 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4281 } 4282 } 4283 } 4284 4285 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4286 // Check if PhiNode::Ideal's "Split phis through memory merges" 4287 // transform should be attempted. Look for this->phi->this cycle. 4288 uint merge_width = req(); 4289 if (merge_width > Compile::AliasIdxRaw) { 4290 PhiNode* phi = base_memory()->as_Phi(); 4291 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4292 if (phi->in(i) == this) { 4293 phase->is_IterGVN()->_worklist.push(phi); 4294 break; 4295 } 4296 } 4297 } 4298 } 4299 4300 assert(progress || verify_sparse(), "please, no dups of base"); 4301 return progress; 4302 } 4303 4304 //-------------------------set_base_memory------------------------------------- 4305 void MergeMemNode::set_base_memory(Node *new_base) { 4306 Node* empty_mem = empty_memory(); 4307 set_req(Compile::AliasIdxBot, new_base); 4308 assert(memory_at(req()) == new_base, "must set default memory"); 4309 // Clear out other occurrences of new_base: 4310 if (new_base != empty_mem) { 4311 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4312 if (in(i) == new_base) set_req(i, empty_mem); 4313 } 4314 } 4315 } 4316 4317 //------------------------------out_RegMask------------------------------------ 4318 const RegMask &MergeMemNode::out_RegMask() const { 4319 return RegMask::Empty; 4320 } 4321 4322 //------------------------------dump_spec-------------------------------------- 4323 #ifndef PRODUCT 4324 void MergeMemNode::dump_spec(outputStream *st) const { 4325 st->print(" {"); 4326 Node* base_mem = base_memory(); 4327 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4328 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4329 if (mem == base_mem) { st->print(" -"); continue; } 4330 st->print( " N%d:", mem->_idx ); 4331 Compile::current()->get_adr_type(i)->dump_on(st); 4332 } 4333 st->print(" }"); 4334 } 4335 #endif // !PRODUCT 4336 4337 4338 #ifdef ASSERT 4339 static bool might_be_same(Node* a, Node* b) { 4340 if (a == b) return true; 4341 if (!(a->is_Phi() || b->is_Phi())) return false; 4342 // phis shift around during optimization 4343 return true; // pretty stupid... 4344 } 4345 4346 // verify a narrow slice (either incoming or outgoing) 4347 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4348 if (!VerifyAliases) return; // don't bother to verify unless requested 4349 if (is_error_reported()) return; // muzzle asserts when debugging an error 4350 if (Node::in_dump()) return; // muzzle asserts when printing 4351 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4352 assert(n != NULL, ""); 4353 // Elide intervening MergeMem's 4354 while (n->is_MergeMem()) { 4355 n = n->as_MergeMem()->memory_at(alias_idx); 4356 } 4357 Compile* C = Compile::current(); 4358 const TypePtr* n_adr_type = n->adr_type(); 4359 if (n == m->empty_memory()) { 4360 // Implicit copy of base_memory() 4361 } else if (n_adr_type != TypePtr::BOTTOM) { 4362 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4363 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4364 } else { 4365 // A few places like make_runtime_call "know" that VM calls are narrow, 4366 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4367 bool expected_wide_mem = false; 4368 if (n == m->base_memory()) { 4369 expected_wide_mem = true; 4370 } else if (alias_idx == Compile::AliasIdxRaw || 4371 n == m->memory_at(Compile::AliasIdxRaw)) { 4372 expected_wide_mem = true; 4373 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4374 // memory can "leak through" calls on channels that 4375 // are write-once. Allow this also. 4376 expected_wide_mem = true; 4377 } 4378 assert(expected_wide_mem, "expected narrow slice replacement"); 4379 } 4380 } 4381 #else // !ASSERT 4382 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4383 #endif 4384 4385 4386 //-----------------------------memory_at--------------------------------------- 4387 Node* MergeMemNode::memory_at(uint alias_idx) const { 4388 assert(alias_idx >= Compile::AliasIdxRaw || 4389 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4390 "must avoid base_memory and AliasIdxTop"); 4391 4392 // Otherwise, it is a narrow slice. 4393 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4394 Compile *C = Compile::current(); 4395 if (is_empty_memory(n)) { 4396 // the array is sparse; empty slots are the "top" node 4397 n = base_memory(); 4398 assert(Node::in_dump() 4399 || n == NULL || n->bottom_type() == Type::TOP 4400 || n->adr_type() == NULL // address is TOP 4401 || n->adr_type() == TypePtr::BOTTOM 4402 || n->adr_type() == TypeRawPtr::BOTTOM 4403 || Compile::current()->AliasLevel() == 0, 4404 "must be a wide memory"); 4405 // AliasLevel == 0 if we are organizing the memory states manually. 4406 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4407 } else { 4408 // make sure the stored slice is sane 4409 #ifdef ASSERT 4410 if (is_error_reported() || Node::in_dump()) { 4411 } else if (might_be_same(n, base_memory())) { 4412 // Give it a pass: It is a mostly harmless repetition of the base. 4413 // This can arise normally from node subsumption during optimization. 4414 } else { 4415 verify_memory_slice(this, alias_idx, n); 4416 } 4417 #endif 4418 } 4419 return n; 4420 } 4421 4422 //---------------------------set_memory_at------------------------------------- 4423 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4424 verify_memory_slice(this, alias_idx, n); 4425 Node* empty_mem = empty_memory(); 4426 if (n == base_memory()) n = empty_mem; // collapse default 4427 uint need_req = alias_idx+1; 4428 if (req() < need_req) { 4429 if (n == empty_mem) return; // already the default, so do not grow me 4430 // grow the sparse array 4431 do { 4432 add_req(empty_mem); 4433 } while (req() < need_req); 4434 } 4435 set_req( alias_idx, n ); 4436 } 4437 4438 4439 4440 //--------------------------iteration_setup------------------------------------ 4441 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4442 if (other != NULL) { 4443 grow_to_match(other); 4444 // invariant: the finite support of mm2 is within mm->req() 4445 #ifdef ASSERT 4446 for (uint i = req(); i < other->req(); i++) { 4447 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4448 } 4449 #endif 4450 } 4451 // Replace spurious copies of base_memory by top. 4452 Node* base_mem = base_memory(); 4453 if (base_mem != NULL && !base_mem->is_top()) { 4454 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4455 if (in(i) == base_mem) 4456 set_req(i, empty_memory()); 4457 } 4458 } 4459 } 4460 4461 //---------------------------grow_to_match------------------------------------- 4462 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4463 Node* empty_mem = empty_memory(); 4464 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4465 // look for the finite support of the other memory 4466 for (uint i = other->req(); --i >= req(); ) { 4467 if (other->in(i) != empty_mem) { 4468 uint new_len = i+1; 4469 while (req() < new_len) add_req(empty_mem); 4470 break; 4471 } 4472 } 4473 } 4474 4475 //---------------------------verify_sparse------------------------------------- 4476 #ifndef PRODUCT 4477 bool MergeMemNode::verify_sparse() const { 4478 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4479 Node* base_mem = base_memory(); 4480 // The following can happen in degenerate cases, since empty==top. 4481 if (is_empty_memory(base_mem)) return true; 4482 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4483 assert(in(i) != NULL, "sane slice"); 4484 if (in(i) == base_mem) return false; // should have been the sentinel value! 4485 } 4486 return true; 4487 } 4488 4489 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4490 Node* n; 4491 n = mm->in(idx); 4492 if (mem == n) return true; // might be empty_memory() 4493 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4494 if (mem == n) return true; 4495 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4496 if (mem == n) return true; 4497 if (n == NULL) break; 4498 } 4499 return false; 4500 } 4501 #endif // !PRODUCT