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