1 /* 2 * Copyright (c) 1997, 2018, 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 "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "oops/objArrayKlass.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/compile.hpp" 37 #include "opto/connode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/loopnode.hpp" 40 #include "opto/machnode.hpp" 41 #include "opto/matcher.hpp" 42 #include "opto/memnode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/phaseX.hpp" 46 #include "opto/regmask.hpp" 47 #include "utilities/align.hpp" 48 #include "utilities/copy.hpp" 49 #include "utilities/macros.hpp" 50 #include "utilities/vmError.hpp" 51 #if INCLUDE_ZGC 52 #include "gc/z/c2/zBarrierSetC2.hpp" 53 #endif 54 55 // Portions of code courtesy of Clifford Click 56 57 // Optimization - Graph Style 58 59 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 60 61 //============================================================================= 62 uint MemNode::size_of() const { return sizeof(*this); } 63 64 const TypePtr *MemNode::adr_type() const { 65 Node* adr = in(Address); 66 if (adr == NULL) return NULL; // node is dead 67 const TypePtr* cross_check = NULL; 68 DEBUG_ONLY(cross_check = _adr_type); 69 return calculate_adr_type(adr->bottom_type(), cross_check); 70 } 71 72 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 73 if (adr != NULL) { 74 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 75 return true; 76 } 77 } 78 return false; 79 } 80 81 #ifndef PRODUCT 82 void MemNode::dump_spec(outputStream *st) const { 83 if (in(Address) == NULL) return; // node is dead 84 #ifndef ASSERT 85 // fake the missing field 86 const TypePtr* _adr_type = NULL; 87 if (in(Address) != NULL) 88 _adr_type = in(Address)->bottom_type()->isa_ptr(); 89 #endif 90 dump_adr_type(this, _adr_type, st); 91 92 Compile* C = Compile::current(); 93 if (C->alias_type(_adr_type)->is_volatile()) { 94 st->print(" Volatile!"); 95 } 96 if (_unaligned_access) { 97 st->print(" unaligned"); 98 } 99 if (_mismatched_access) { 100 st->print(" mismatched"); 101 } 102 } 103 104 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 105 st->print(" @"); 106 if (adr_type == NULL) { 107 st->print("NULL"); 108 } else { 109 adr_type->dump_on(st); 110 Compile* C = Compile::current(); 111 Compile::AliasType* atp = NULL; 112 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 113 if (atp == NULL) 114 st->print(", idx=?\?;"); 115 else if (atp->index() == Compile::AliasIdxBot) 116 st->print(", idx=Bot;"); 117 else if (atp->index() == Compile::AliasIdxTop) 118 st->print(", idx=Top;"); 119 else if (atp->index() == Compile::AliasIdxRaw) 120 st->print(", idx=Raw;"); 121 else { 122 ciField* field = atp->field(); 123 if (field) { 124 st->print(", name="); 125 field->print_name_on(st); 126 } 127 st->print(", idx=%d;", atp->index()); 128 } 129 } 130 } 131 132 extern void print_alias_types(); 133 134 #endif 135 136 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 137 assert((t_oop != NULL), "sanity"); 138 bool is_instance = t_oop->is_known_instance_field(); 139 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 140 (load != NULL) && load->is_Load() && 141 (phase->is_IterGVN() != NULL); 142 if (!(is_instance || is_boxed_value_load)) 143 return mchain; // don't try to optimize non-instance types 144 uint instance_id = t_oop->instance_id(); 145 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory); 146 Node *prev = NULL; 147 Node *result = mchain; 148 while (prev != result) { 149 prev = result; 150 if (result == start_mem) 151 break; // hit one of our sentinels 152 // skip over a call which does not affect this memory slice 153 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 154 Node *proj_in = result->in(0); 155 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 156 break; // hit one of our sentinels 157 } else if (proj_in->is_Call()) { 158 // ArrayCopyNodes processed here as well 159 CallNode *call = proj_in->as_Call(); 160 if (!call->may_modify(t_oop, phase)) { // returns false for instances 161 result = call->in(TypeFunc::Memory); 162 } 163 } else if (proj_in->is_Initialize()) { 164 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 165 // Stop if this is the initialization for the object instance which 166 // contains this memory slice, otherwise skip over it. 167 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 168 break; 169 } 170 if (is_instance) { 171 result = proj_in->in(TypeFunc::Memory); 172 } else if (is_boxed_value_load) { 173 Node* klass = alloc->in(AllocateNode::KlassNode); 174 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 175 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 176 result = proj_in->in(TypeFunc::Memory); // not related allocation 177 } 178 } 179 } else if (proj_in->is_MemBar()) { 180 ArrayCopyNode* ac = NULL; 181 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 182 break; 183 } 184 result = proj_in->in(TypeFunc::Memory); 185 } else { 186 assert(false, "unexpected projection"); 187 } 188 } else if (result->is_ClearArray()) { 189 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 190 // Can not bypass initialization of the instance 191 // we are looking for. 192 break; 193 } 194 // Otherwise skip it (the call updated 'result' value). 195 } else if (result->is_MergeMem()) { 196 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 197 } 198 } 199 return result; 200 } 201 202 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 203 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 204 if (t_oop == NULL) 205 return mchain; // don't try to optimize non-oop types 206 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 207 bool is_instance = t_oop->is_known_instance_field(); 208 PhaseIterGVN *igvn = phase->is_IterGVN(); 209 if (is_instance && igvn != NULL && result->is_Phi()) { 210 PhiNode *mphi = result->as_Phi(); 211 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 212 const TypePtr *t = mphi->adr_type(); 213 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 214 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 215 t->is_oopptr()->cast_to_exactness(true) 216 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 217 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) { 218 // clone the Phi with our address type 219 result = mphi->split_out_instance(t_adr, igvn); 220 } else { 221 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 222 } 223 } 224 return result; 225 } 226 227 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 228 uint alias_idx = phase->C->get_alias_index(tp); 229 Node *mem = mmem; 230 #ifdef ASSERT 231 { 232 // Check that current type is consistent with the alias index used during graph construction 233 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 234 bool consistent = adr_check == NULL || adr_check->empty() || 235 phase->C->must_alias(adr_check, alias_idx ); 236 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 237 if( !consistent && adr_check != NULL && !adr_check->empty() && 238 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 239 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 240 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 241 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 242 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 243 // don't assert if it is dead code. 244 consistent = true; 245 } 246 if( !consistent ) { 247 st->print("alias_idx==%d, adr_check==", alias_idx); 248 if( adr_check == NULL ) { 249 st->print("NULL"); 250 } else { 251 adr_check->dump(); 252 } 253 st->cr(); 254 print_alias_types(); 255 assert(consistent, "adr_check must match alias idx"); 256 } 257 } 258 #endif 259 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 260 // means an array I have not precisely typed yet. Do not do any 261 // alias stuff with it any time soon. 262 const TypeOopPtr *toop = tp->isa_oopptr(); 263 if( tp->base() != Type::AnyPtr && 264 !(toop && 265 toop->klass() != NULL && 266 toop->klass()->is_java_lang_Object() && 267 toop->offset() == Type::OffsetBot) ) { 268 // compress paths and change unreachable cycles to TOP 269 // If not, we can update the input infinitely along a MergeMem cycle 270 // Equivalent code in PhiNode::Ideal 271 Node* m = phase->transform(mmem); 272 // If transformed to a MergeMem, get the desired slice 273 // Otherwise the returned node represents memory for every slice 274 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 275 // Update input if it is progress over what we have now 276 } 277 return mem; 278 } 279 280 //--------------------------Ideal_common--------------------------------------- 281 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 282 // Unhook non-raw memories from complete (macro-expanded) initializations. 283 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 284 // If our control input is a dead region, kill all below the region 285 Node *ctl = in(MemNode::Control); 286 if (ctl && remove_dead_region(phase, can_reshape)) 287 return this; 288 ctl = in(MemNode::Control); 289 // Don't bother trying to transform a dead node 290 if (ctl && ctl->is_top()) return NodeSentinel; 291 292 PhaseIterGVN *igvn = phase->is_IterGVN(); 293 // Wait if control on the worklist. 294 if (ctl && can_reshape && igvn != NULL) { 295 Node* bol = NULL; 296 Node* cmp = NULL; 297 if (ctl->in(0)->is_If()) { 298 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 299 bol = ctl->in(0)->in(1); 300 if (bol->is_Bool()) 301 cmp = ctl->in(0)->in(1)->in(1); 302 } 303 if (igvn->_worklist.member(ctl) || 304 (bol != NULL && igvn->_worklist.member(bol)) || 305 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 306 // This control path may be dead. 307 // Delay this memory node transformation until the control is processed. 308 phase->is_IterGVN()->_worklist.push(this); 309 return NodeSentinel; // caller will return NULL 310 } 311 } 312 // Ignore if memory is dead, or self-loop 313 Node *mem = in(MemNode::Memory); 314 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 315 assert(mem != this, "dead loop in MemNode::Ideal"); 316 317 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 318 // This memory slice may be dead. 319 // Delay this mem node transformation until the memory is processed. 320 phase->is_IterGVN()->_worklist.push(this); 321 return NodeSentinel; // caller will return NULL 322 } 323 324 Node *address = in(MemNode::Address); 325 const Type *t_adr = phase->type(address); 326 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 327 328 if (can_reshape && igvn != NULL && 329 (igvn->_worklist.member(address) || 330 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 331 // The address's base and type may change when the address is processed. 332 // Delay this mem node transformation until the address is processed. 333 phase->is_IterGVN()->_worklist.push(this); 334 return NodeSentinel; // caller will return NULL 335 } 336 337 // Do NOT remove or optimize the next lines: ensure a new alias index 338 // is allocated for an oop pointer type before Escape Analysis. 339 // Note: C++ will not remove it since the call has side effect. 340 if (t_adr->isa_oopptr()) { 341 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 342 } 343 344 Node* base = NULL; 345 if (address->is_AddP()) { 346 base = address->in(AddPNode::Base); 347 } 348 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 349 !t_adr->isa_rawptr()) { 350 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 351 // Skip this node optimization if its address has TOP base. 352 return NodeSentinel; // caller will return NULL 353 } 354 355 // Avoid independent memory operations 356 Node* old_mem = mem; 357 358 // The code which unhooks non-raw memories from complete (macro-expanded) 359 // initializations was removed. After macro-expansion all stores catched 360 // by Initialize node became raw stores and there is no information 361 // which memory slices they modify. So it is unsafe to move any memory 362 // operation above these stores. Also in most cases hooked non-raw memories 363 // were already unhooked by using information from detect_ptr_independence() 364 // and find_previous_store(). 365 366 if (mem->is_MergeMem()) { 367 MergeMemNode* mmem = mem->as_MergeMem(); 368 const TypePtr *tp = t_adr->is_ptr(); 369 370 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 371 } 372 373 if (mem != old_mem) { 374 set_req(MemNode::Memory, mem); 375 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) { 376 igvn->_worklist.push(old_mem); 377 } 378 if (phase->type(mem) == Type::TOP) return NodeSentinel; 379 return this; 380 } 381 382 // let the subclass continue analyzing... 383 return NULL; 384 } 385 386 // Helper function for proving some simple control dominations. 387 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 388 // Already assumes that 'dom' is available at 'sub', and that 'sub' 389 // is not a constant (dominated by the method's StartNode). 390 // Used by MemNode::find_previous_store to prove that the 391 // control input of a memory operation predates (dominates) 392 // an allocation it wants to look past. 393 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 394 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 395 return false; // Conservative answer for dead code 396 397 // Check 'dom'. Skip Proj and CatchProj nodes. 398 dom = dom->find_exact_control(dom); 399 if (dom == NULL || dom->is_top()) 400 return false; // Conservative answer for dead code 401 402 if (dom == sub) { 403 // For the case when, for example, 'sub' is Initialize and the original 404 // 'dom' is Proj node of the 'sub'. 405 return false; 406 } 407 408 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 409 return true; 410 411 // 'dom' dominates 'sub' if its control edge and control edges 412 // of all its inputs dominate or equal to sub's control edge. 413 414 // Currently 'sub' is either Allocate, Initialize or Start nodes. 415 // Or Region for the check in LoadNode::Ideal(); 416 // 'sub' should have sub->in(0) != NULL. 417 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 418 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 419 420 // Get control edge of 'sub'. 421 Node* orig_sub = sub; 422 sub = sub->find_exact_control(sub->in(0)); 423 if (sub == NULL || sub->is_top()) 424 return false; // Conservative answer for dead code 425 426 assert(sub->is_CFG(), "expecting control"); 427 428 if (sub == dom) 429 return true; 430 431 if (sub->is_Start() || sub->is_Root()) 432 return false; 433 434 { 435 // Check all control edges of 'dom'. 436 437 ResourceMark rm; 438 Arena* arena = Thread::current()->resource_area(); 439 Node_List nlist(arena); 440 Unique_Node_List dom_list(arena); 441 442 dom_list.push(dom); 443 bool only_dominating_controls = false; 444 445 for (uint next = 0; next < dom_list.size(); next++) { 446 Node* n = dom_list.at(next); 447 if (n == orig_sub) 448 return false; // One of dom's inputs dominated by sub. 449 if (!n->is_CFG() && n->pinned()) { 450 // Check only own control edge for pinned non-control nodes. 451 n = n->find_exact_control(n->in(0)); 452 if (n == NULL || n->is_top()) 453 return false; // Conservative answer for dead code 454 assert(n->is_CFG(), "expecting control"); 455 dom_list.push(n); 456 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 457 only_dominating_controls = true; 458 } else if (n->is_CFG()) { 459 if (n->dominates(sub, nlist)) 460 only_dominating_controls = true; 461 else 462 return false; 463 } else { 464 // First, own control edge. 465 Node* m = n->find_exact_control(n->in(0)); 466 if (m != NULL) { 467 if (m->is_top()) 468 return false; // Conservative answer for dead code 469 dom_list.push(m); 470 } 471 // Now, the rest of edges. 472 uint cnt = n->req(); 473 for (uint i = 1; i < cnt; i++) { 474 m = n->find_exact_control(n->in(i)); 475 if (m == NULL || m->is_top()) 476 continue; 477 dom_list.push(m); 478 } 479 } 480 } 481 return only_dominating_controls; 482 } 483 } 484 485 //---------------------detect_ptr_independence--------------------------------- 486 // Used by MemNode::find_previous_store to prove that two base 487 // pointers are never equal. 488 // The pointers are accompanied by their associated allocations, 489 // if any, which have been previously discovered by the caller. 490 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 491 Node* p2, AllocateNode* a2, 492 PhaseTransform* phase) { 493 // Attempt to prove that these two pointers cannot be aliased. 494 // They may both manifestly be allocations, and they should differ. 495 // Or, if they are not both allocations, they can be distinct constants. 496 // Otherwise, one is an allocation and the other a pre-existing value. 497 if (a1 == NULL && a2 == NULL) { // neither an allocation 498 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 499 } else if (a1 != NULL && a2 != NULL) { // both allocations 500 return (a1 != a2); 501 } else if (a1 != NULL) { // one allocation a1 502 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 503 return all_controls_dominate(p2, a1); 504 } else { //(a2 != NULL) // one allocation a2 505 return all_controls_dominate(p1, a2); 506 } 507 return false; 508 } 509 510 511 // Find an arraycopy that must have set (can_see_stored_value=true) or 512 // could have set (can_see_stored_value=false) the value for this load 513 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 514 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 515 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 516 Node* mb = mem->in(0); 517 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 518 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 519 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 520 if (ac->is_clonebasic()) { 521 intptr_t offset; 522 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 523 if (alloc != NULL && alloc == ld_alloc) { 524 return ac; 525 } 526 } 527 } 528 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 529 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 530 531 if (ac->is_arraycopy_validated() || 532 ac->is_copyof_validated() || 533 ac->is_copyofrange_validated()) { 534 Node* ld_addp = in(MemNode::Address); 535 if (ld_addp->is_AddP()) { 536 Node* ld_base = ld_addp->in(AddPNode::Address); 537 Node* ld_offs = ld_addp->in(AddPNode::Offset); 538 539 Node* dest = ac->in(ArrayCopyNode::Dest); 540 541 if (dest == ld_base) { 542 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 543 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 544 return ac; 545 } 546 if (!can_see_stored_value) { 547 mem = ac->in(TypeFunc::Memory); 548 } 549 } 550 } 551 } 552 } 553 return NULL; 554 } 555 556 // The logic for reordering loads and stores uses four steps: 557 // (a) Walk carefully past stores and initializations which we 558 // can prove are independent of this load. 559 // (b) Observe that the next memory state makes an exact match 560 // with self (load or store), and locate the relevant store. 561 // (c) Ensure that, if we were to wire self directly to the store, 562 // the optimizer would fold it up somehow. 563 // (d) Do the rewiring, and return, depending on some other part of 564 // the optimizer to fold up the load. 565 // This routine handles steps (a) and (b). Steps (c) and (d) are 566 // specific to loads and stores, so they are handled by the callers. 567 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 568 // 569 Node* MemNode::find_previous_store(PhaseTransform* phase) { 570 Node* ctrl = in(MemNode::Control); 571 Node* adr = in(MemNode::Address); 572 intptr_t offset = 0; 573 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 574 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 575 576 if (offset == Type::OffsetBot) 577 return NULL; // cannot unalias unless there are precise offsets 578 579 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 580 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 581 582 intptr_t size_in_bytes = memory_size(); 583 584 Node* mem = in(MemNode::Memory); // start searching here... 585 586 int cnt = 50; // Cycle limiter 587 for (;;) { // While we can dance past unrelated stores... 588 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 589 590 Node* prev = mem; 591 if (mem->is_Store()) { 592 Node* st_adr = mem->in(MemNode::Address); 593 intptr_t st_offset = 0; 594 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 595 if (st_base == NULL) 596 break; // inscrutable pointer 597 598 // For raw accesses it's not enough to prove that constant offsets don't intersect. 599 // We need the bases to be the equal in order for the offset check to make sense. 600 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 601 break; 602 } 603 604 if (st_offset != offset && st_offset != Type::OffsetBot) { 605 const int MAX_STORE = BytesPerLong; 606 if (st_offset >= offset + size_in_bytes || 607 st_offset <= offset - MAX_STORE || 608 st_offset <= offset - mem->as_Store()->memory_size()) { 609 // Success: The offsets are provably independent. 610 // (You may ask, why not just test st_offset != offset and be done? 611 // The answer is that stores of different sizes can co-exist 612 // in the same sequence of RawMem effects. We sometimes initialize 613 // a whole 'tile' of array elements with a single jint or jlong.) 614 mem = mem->in(MemNode::Memory); 615 continue; // (a) advance through independent store memory 616 } 617 } 618 if (st_base != base && 619 detect_ptr_independence(base, alloc, 620 st_base, 621 AllocateNode::Ideal_allocation(st_base, phase), 622 phase)) { 623 // Success: The bases are provably independent. 624 mem = mem->in(MemNode::Memory); 625 continue; // (a) advance through independent store memory 626 } 627 628 // (b) At this point, if the bases or offsets do not agree, we lose, 629 // since we have not managed to prove 'this' and 'mem' independent. 630 if (st_base == base && st_offset == offset) { 631 return mem; // let caller handle steps (c), (d) 632 } 633 634 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 635 InitializeNode* st_init = mem->in(0)->as_Initialize(); 636 AllocateNode* st_alloc = st_init->allocation(); 637 if (st_alloc == NULL) 638 break; // something degenerated 639 bool known_identical = false; 640 bool known_independent = false; 641 if (alloc == st_alloc) 642 known_identical = true; 643 else if (alloc != NULL) 644 known_independent = true; 645 else if (all_controls_dominate(this, st_alloc)) 646 known_independent = true; 647 648 if (known_independent) { 649 // The bases are provably independent: Either they are 650 // manifestly distinct allocations, or else the control 651 // of this load dominates the store's allocation. 652 int alias_idx = phase->C->get_alias_index(adr_type()); 653 if (alias_idx == Compile::AliasIdxRaw) { 654 mem = st_alloc->in(TypeFunc::Memory); 655 } else { 656 mem = st_init->memory(alias_idx); 657 } 658 continue; // (a) advance through independent store memory 659 } 660 661 // (b) at this point, if we are not looking at a store initializing 662 // the same allocation we are loading from, we lose. 663 if (known_identical) { 664 // From caller, can_see_stored_value will consult find_captured_store. 665 return mem; // let caller handle steps (c), (d) 666 } 667 668 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 669 if (prev != mem) { 670 // Found an arraycopy but it doesn't affect that load 671 continue; 672 } 673 // Found an arraycopy that may affect that load 674 return mem; 675 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 676 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 677 if (mem->is_Proj() && mem->in(0)->is_Call()) { 678 // ArrayCopyNodes processed here as well. 679 CallNode *call = mem->in(0)->as_Call(); 680 if (!call->may_modify(addr_t, phase)) { 681 mem = call->in(TypeFunc::Memory); 682 continue; // (a) advance through independent call memory 683 } 684 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 685 ArrayCopyNode* ac = NULL; 686 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 687 break; 688 } 689 mem = mem->in(0)->in(TypeFunc::Memory); 690 continue; // (a) advance through independent MemBar memory 691 } else if (mem->is_ClearArray()) { 692 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 693 // (the call updated 'mem' value) 694 continue; // (a) advance through independent allocation memory 695 } else { 696 // Can not bypass initialization of the instance 697 // we are looking for. 698 return mem; 699 } 700 } else if (mem->is_MergeMem()) { 701 int alias_idx = phase->C->get_alias_index(adr_type()); 702 mem = mem->as_MergeMem()->memory_at(alias_idx); 703 continue; // (a) advance through independent MergeMem memory 704 } 705 } 706 707 // Unless there is an explicit 'continue', we must bail out here, 708 // because 'mem' is an inscrutable memory state (e.g., a call). 709 break; 710 } 711 712 return NULL; // bail out 713 } 714 715 //----------------------calculate_adr_type------------------------------------- 716 // Helper function. Notices when the given type of address hits top or bottom. 717 // Also, asserts a cross-check of the type against the expected address type. 718 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 719 if (t == Type::TOP) return NULL; // does not touch memory any more? 720 #ifdef PRODUCT 721 cross_check = NULL; 722 #else 723 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL; 724 #endif 725 const TypePtr* tp = t->isa_ptr(); 726 if (tp == NULL) { 727 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 728 return TypePtr::BOTTOM; // touches lots of memory 729 } else { 730 #ifdef ASSERT 731 // %%%% [phh] We don't check the alias index if cross_check is 732 // TypeRawPtr::BOTTOM. Needs to be investigated. 733 if (cross_check != NULL && 734 cross_check != TypePtr::BOTTOM && 735 cross_check != TypeRawPtr::BOTTOM) { 736 // Recheck the alias index, to see if it has changed (due to a bug). 737 Compile* C = Compile::current(); 738 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 739 "must stay in the original alias category"); 740 // The type of the address must be contained in the adr_type, 741 // disregarding "null"-ness. 742 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 743 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 744 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 745 "real address must not escape from expected memory type"); 746 } 747 #endif 748 return tp; 749 } 750 } 751 752 //============================================================================= 753 // Should LoadNode::Ideal() attempt to remove control edges? 754 bool LoadNode::can_remove_control() const { 755 return true; 756 } 757 uint LoadNode::size_of() const { return sizeof(*this); } 758 uint LoadNode::cmp( const Node &n ) const 759 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 760 const Type *LoadNode::bottom_type() const { return _type; } 761 uint LoadNode::ideal_reg() const { 762 return _type->ideal_reg(); 763 } 764 765 #ifndef PRODUCT 766 void LoadNode::dump_spec(outputStream *st) const { 767 MemNode::dump_spec(st); 768 if( !Verbose && !WizardMode ) { 769 // standard dump does this in Verbose and WizardMode 770 st->print(" #"); _type->dump_on(st); 771 } 772 if (!depends_only_on_test()) { 773 st->print(" (does not depend only on test)"); 774 } 775 } 776 #endif 777 778 #ifdef ASSERT 779 //----------------------------is_immutable_value------------------------------- 780 // Helper function to allow a raw load without control edge for some cases 781 bool LoadNode::is_immutable_value(Node* adr) { 782 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 783 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 784 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 785 in_bytes(JavaThread::osthread_offset()))); 786 } 787 #endif 788 789 //----------------------------LoadNode::make----------------------------------- 790 // Polymorphic factory method: 791 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 792 ControlDependency control_dependency, bool unaligned, bool mismatched) { 793 Compile* C = gvn.C; 794 795 // sanity check the alias category against the created node type 796 assert(!(adr_type->isa_oopptr() && 797 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 798 "use LoadKlassNode instead"); 799 assert(!(adr_type->isa_aryptr() && 800 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 801 "use LoadRangeNode instead"); 802 // Check control edge of raw loads 803 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 804 // oop will be recorded in oop map if load crosses safepoint 805 rt->isa_oopptr() || is_immutable_value(adr), 806 "raw memory operations should have control edge"); 807 LoadNode* load = NULL; 808 switch (bt) { 809 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 810 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 811 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 812 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 813 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 814 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 815 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 816 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 817 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 818 case T_OBJECT: 819 #ifdef _LP64 820 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 821 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 822 } else 823 #endif 824 { 825 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 826 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 827 } 828 break; 829 default: 830 ShouldNotReachHere(); 831 break; 832 } 833 assert(load != NULL, "LoadNode should have been created"); 834 if (unaligned) { 835 load->set_unaligned_access(); 836 } 837 if (mismatched) { 838 load->set_mismatched_access(); 839 } 840 if (load->Opcode() == Op_LoadN) { 841 Node* ld = gvn.transform(load); 842 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 843 } 844 845 return load; 846 } 847 848 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 849 ControlDependency control_dependency, bool unaligned, bool mismatched) { 850 bool require_atomic = true; 851 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 852 if (unaligned) { 853 load->set_unaligned_access(); 854 } 855 if (mismatched) { 856 load->set_mismatched_access(); 857 } 858 return load; 859 } 860 861 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 862 ControlDependency control_dependency, bool unaligned, bool mismatched) { 863 bool require_atomic = true; 864 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 865 if (unaligned) { 866 load->set_unaligned_access(); 867 } 868 if (mismatched) { 869 load->set_mismatched_access(); 870 } 871 return load; 872 } 873 874 875 876 //------------------------------hash------------------------------------------- 877 uint LoadNode::hash() const { 878 // unroll addition of interesting fields 879 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 880 } 881 882 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 883 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 884 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 885 bool is_stable_ary = FoldStableValues && 886 (tp != NULL) && (tp->isa_aryptr() != NULL) && 887 tp->isa_aryptr()->is_stable(); 888 889 return (eliminate_boxing && non_volatile) || is_stable_ary; 890 } 891 892 return false; 893 } 894 895 // Is the value loaded previously stored by an arraycopy? If so return 896 // a load node that reads from the source array so we may be able to 897 // optimize out the ArrayCopy node later. 898 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 899 #if INCLUDE_ZGC 900 if (UseZGC) { 901 if (bottom_type()->make_oopptr() != NULL) { 902 return NULL; 903 } 904 } 905 #endif 906 907 Node* ld_adr = in(MemNode::Address); 908 intptr_t ld_off = 0; 909 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 910 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 911 if (ac != NULL) { 912 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 913 914 Node* mem = ac->in(TypeFunc::Memory); 915 Node* ctl = ac->in(0); 916 Node* src = ac->in(ArrayCopyNode::Src); 917 918 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 919 return NULL; 920 } 921 922 LoadNode* ld = clone()->as_Load(); 923 Node* addp = in(MemNode::Address)->clone(); 924 if (ac->as_ArrayCopy()->is_clonebasic()) { 925 assert(ld_alloc != NULL, "need an alloc"); 926 assert(addp->is_AddP(), "address must be addp"); 927 assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address"); 928 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 929 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern"); 930 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern"); 931 addp->set_req(AddPNode::Base, src->in(AddPNode::Base)); 932 addp->set_req(AddPNode::Address, src->in(AddPNode::Address)); 933 } else { 934 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 935 ac->as_ArrayCopy()->is_copyof_validated() || 936 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 937 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 938 addp->set_req(AddPNode::Base, src); 939 addp->set_req(AddPNode::Address, src); 940 941 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 942 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 943 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 944 uint shift = exact_log2(type2aelembytes(ary_elem)); 945 946 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 947 #ifdef _LP64 948 diff = phase->transform(new ConvI2LNode(diff)); 949 #endif 950 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 951 952 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 953 addp->set_req(AddPNode::Offset, offset); 954 } 955 addp = phase->transform(addp); 956 #ifdef ASSERT 957 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 958 ld->_adr_type = adr_type; 959 #endif 960 ld->set_req(MemNode::Address, addp); 961 ld->set_req(0, ctl); 962 ld->set_req(MemNode::Memory, mem); 963 // load depends on the tests that validate the arraycopy 964 ld->_control_dependency = Pinned; 965 return ld; 966 } 967 return NULL; 968 } 969 970 971 //---------------------------can_see_stored_value------------------------------ 972 // This routine exists to make sure this set of tests is done the same 973 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 974 // will change the graph shape in a way which makes memory alive twice at the 975 // same time (uses the Oracle model of aliasing), then some 976 // LoadXNode::Identity will fold things back to the equivalence-class model 977 // of aliasing. 978 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 979 Node* ld_adr = in(MemNode::Address); 980 intptr_t ld_off = 0; 981 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 982 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 983 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 984 // This is more general than load from boxing objects. 985 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 986 uint alias_idx = atp->index(); 987 bool final = !atp->is_rewritable(); 988 Node* result = NULL; 989 Node* current = st; 990 // Skip through chains of MemBarNodes checking the MergeMems for 991 // new states for the slice of this load. Stop once any other 992 // kind of node is encountered. Loads from final memory can skip 993 // through any kind of MemBar but normal loads shouldn't skip 994 // through MemBarAcquire since the could allow them to move out of 995 // a synchronized region. 996 while (current->is_Proj()) { 997 int opc = current->in(0)->Opcode(); 998 if ((final && (opc == Op_MemBarAcquire || 999 opc == Op_MemBarAcquireLock || 1000 opc == Op_LoadFence)) || 1001 opc == Op_MemBarRelease || 1002 opc == Op_StoreFence || 1003 opc == Op_MemBarReleaseLock || 1004 opc == Op_MemBarStoreStore || 1005 opc == Op_MemBarCPUOrder) { 1006 Node* mem = current->in(0)->in(TypeFunc::Memory); 1007 if (mem->is_MergeMem()) { 1008 MergeMemNode* merge = mem->as_MergeMem(); 1009 Node* new_st = merge->memory_at(alias_idx); 1010 if (new_st == merge->base_memory()) { 1011 // Keep searching 1012 current = new_st; 1013 continue; 1014 } 1015 // Save the new memory state for the slice and fall through 1016 // to exit. 1017 result = new_st; 1018 } 1019 } 1020 break; 1021 } 1022 if (result != NULL) { 1023 st = result; 1024 } 1025 } 1026 1027 // Loop around twice in the case Load -> Initialize -> Store. 1028 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1029 for (int trip = 0; trip <= 1; trip++) { 1030 1031 if (st->is_Store()) { 1032 Node* st_adr = st->in(MemNode::Address); 1033 if (!phase->eqv(st_adr, ld_adr)) { 1034 // Try harder before giving up... Match raw and non-raw pointers. 1035 intptr_t st_off = 0; 1036 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 1037 if (alloc == NULL) return NULL; 1038 if (alloc != ld_alloc) return NULL; 1039 if (ld_off != st_off) return NULL; 1040 // At this point we have proven something like this setup: 1041 // A = Allocate(...) 1042 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 1043 // S = StoreQ(, AddP(, A.Parm , #Off), V) 1044 // (Actually, we haven't yet proven the Q's are the same.) 1045 // In other words, we are loading from a casted version of 1046 // the same pointer-and-offset that we stored to. 1047 // Thus, we are able to replace L by V. 1048 } 1049 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1050 if (store_Opcode() != st->Opcode()) 1051 return NULL; 1052 return st->in(MemNode::ValueIn); 1053 } 1054 1055 // A load from a freshly-created object always returns zero. 1056 // (This can happen after LoadNode::Ideal resets the load's memory input 1057 // to find_captured_store, which returned InitializeNode::zero_memory.) 1058 if (st->is_Proj() && st->in(0)->is_Allocate() && 1059 (st->in(0) == ld_alloc) && 1060 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1061 // return a zero value for the load's basic type 1062 // (This is one of the few places where a generic PhaseTransform 1063 // can create new nodes. Think of it as lazily manifesting 1064 // virtually pre-existing constants.) 1065 return phase->zerocon(memory_type()); 1066 } 1067 1068 // A load from an initialization barrier can match a captured store. 1069 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1070 InitializeNode* init = st->in(0)->as_Initialize(); 1071 AllocateNode* alloc = init->allocation(); 1072 if ((alloc != NULL) && (alloc == ld_alloc)) { 1073 // examine a captured store value 1074 st = init->find_captured_store(ld_off, memory_size(), phase); 1075 if (st != NULL) { 1076 continue; // take one more trip around 1077 } 1078 } 1079 } 1080 1081 // Load boxed value from result of valueOf() call is input parameter. 1082 if (this->is_Load() && ld_adr->is_AddP() && 1083 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1084 intptr_t ignore = 0; 1085 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1086 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1087 base = bs->step_over_gc_barrier(base); 1088 if (base != NULL && base->is_Proj() && 1089 base->as_Proj()->_con == TypeFunc::Parms && 1090 base->in(0)->is_CallStaticJava() && 1091 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1092 return base->in(0)->in(TypeFunc::Parms); 1093 } 1094 } 1095 1096 break; 1097 } 1098 1099 return NULL; 1100 } 1101 1102 //----------------------is_instance_field_load_with_local_phi------------------ 1103 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1104 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1105 in(Address)->is_AddP() ) { 1106 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1107 // Only instances and boxed values. 1108 if( t_oop != NULL && 1109 (t_oop->is_ptr_to_boxed_value() || 1110 t_oop->is_known_instance_field()) && 1111 t_oop->offset() != Type::OffsetBot && 1112 t_oop->offset() != Type::OffsetTop) { 1113 return true; 1114 } 1115 } 1116 return false; 1117 } 1118 1119 //------------------------------Identity--------------------------------------- 1120 // Loads are identity if previous store is to same address 1121 Node* LoadNode::Identity(PhaseGVN* phase) { 1122 // If the previous store-maker is the right kind of Store, and the store is 1123 // to the same address, then we are equal to the value stored. 1124 Node* mem = in(Memory); 1125 Node* value = can_see_stored_value(mem, phase); 1126 if( value ) { 1127 // byte, short & char stores truncate naturally. 1128 // A load has to load the truncated value which requires 1129 // some sort of masking operation and that requires an 1130 // Ideal call instead of an Identity call. 1131 if (memory_size() < BytesPerInt) { 1132 // If the input to the store does not fit with the load's result type, 1133 // it must be truncated via an Ideal call. 1134 if (!phase->type(value)->higher_equal(phase->type(this))) 1135 return this; 1136 } 1137 // (This works even when value is a Con, but LoadNode::Value 1138 // usually runs first, producing the singleton type of the Con.) 1139 return value; 1140 } 1141 1142 // Search for an existing data phi which was generated before for the same 1143 // instance's field to avoid infinite generation of phis in a loop. 1144 Node *region = mem->in(0); 1145 if (is_instance_field_load_with_local_phi(region)) { 1146 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1147 int this_index = phase->C->get_alias_index(addr_t); 1148 int this_offset = addr_t->offset(); 1149 int this_iid = addr_t->instance_id(); 1150 if (!addr_t->is_known_instance() && 1151 addr_t->is_ptr_to_boxed_value()) { 1152 // Use _idx of address base (could be Phi node) for boxed values. 1153 intptr_t ignore = 0; 1154 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1155 if (base == NULL) { 1156 return this; 1157 } 1158 this_iid = base->_idx; 1159 } 1160 const Type* this_type = bottom_type(); 1161 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1162 Node* phi = region->fast_out(i); 1163 if (phi->is_Phi() && phi != mem && 1164 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1165 return phi; 1166 } 1167 } 1168 } 1169 1170 return this; 1171 } 1172 1173 // Construct an equivalent unsigned load. 1174 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1175 BasicType bt = T_ILLEGAL; 1176 const Type* rt = NULL; 1177 switch (Opcode()) { 1178 case Op_LoadUB: return this; 1179 case Op_LoadUS: return this; 1180 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1181 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1182 default: 1183 assert(false, "no unsigned variant: %s", Name()); 1184 return NULL; 1185 } 1186 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1187 raw_adr_type(), rt, bt, _mo, _control_dependency, 1188 is_unaligned_access(), is_mismatched_access()); 1189 } 1190 1191 // Construct an equivalent signed load. 1192 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1193 BasicType bt = T_ILLEGAL; 1194 const Type* rt = NULL; 1195 switch (Opcode()) { 1196 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1197 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1198 case Op_LoadB: // fall through 1199 case Op_LoadS: // fall through 1200 case Op_LoadI: // fall through 1201 case Op_LoadL: return this; 1202 default: 1203 assert(false, "no signed variant: %s", Name()); 1204 return NULL; 1205 } 1206 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1207 raw_adr_type(), rt, bt, _mo, _control_dependency, 1208 is_unaligned_access(), is_mismatched_access()); 1209 } 1210 1211 // We're loading from an object which has autobox behaviour. 1212 // If this object is result of a valueOf call we'll have a phi 1213 // merging a newly allocated object and a load from the cache. 1214 // We want to replace this load with the original incoming 1215 // argument to the valueOf call. 1216 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1217 assert(phase->C->eliminate_boxing(), "sanity"); 1218 intptr_t ignore = 0; 1219 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1220 if ((base == NULL) || base->is_Phi()) { 1221 // Push the loads from the phi that comes from valueOf up 1222 // through it to allow elimination of the loads and the recovery 1223 // of the original value. It is done in split_through_phi(). 1224 return NULL; 1225 } else if (base->is_Load() || 1226 (base->is_DecodeN() && base->in(1)->is_Load())) { 1227 // Eliminate the load of boxed value for integer types from the cache 1228 // array by deriving the value from the index into the array. 1229 // Capture the offset of the load and then reverse the computation. 1230 1231 // Get LoadN node which loads a boxing object from 'cache' array. 1232 if (base->is_DecodeN()) { 1233 base = base->in(1); 1234 } 1235 if (!base->in(Address)->is_AddP()) { 1236 return NULL; // Complex address 1237 } 1238 AddPNode* address = base->in(Address)->as_AddP(); 1239 Node* cache_base = address->in(AddPNode::Base); 1240 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1241 // Get ConP node which is static 'cache' field. 1242 cache_base = cache_base->in(1); 1243 } 1244 if ((cache_base != NULL) && cache_base->is_Con()) { 1245 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1246 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1247 Node* elements[4]; 1248 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1249 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1250 if (count > 0 && elements[0]->is_Con() && 1251 (count == 1 || 1252 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1253 elements[1]->in(2) == phase->intcon(shift)))) { 1254 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1255 // Fetch the box object cache[0] at the base of the array and get its value 1256 ciInstance* box = array->obj_at(0)->as_instance(); 1257 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1258 assert(ik->is_box_klass(), "sanity"); 1259 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1260 if (ik->nof_nonstatic_fields() == 1) { 1261 // This should be true nonstatic_field_at requires calling 1262 // nof_nonstatic_fields so check it anyway 1263 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1264 BasicType bt = c.basic_type(); 1265 // Only integer types have boxing cache. 1266 assert(bt == T_BOOLEAN || bt == T_CHAR || 1267 bt == T_BYTE || bt == T_SHORT || 1268 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1269 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1270 if (cache_low != (int)cache_low) { 1271 return NULL; // should not happen since cache is array indexed by value 1272 } 1273 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1274 if (offset != (int)offset) { 1275 return NULL; // should not happen since cache is array indexed by value 1276 } 1277 // Add up all the offsets making of the address of the load 1278 Node* result = elements[0]; 1279 for (int i = 1; i < count; i++) { 1280 result = phase->transform(new AddXNode(result, elements[i])); 1281 } 1282 // Remove the constant offset from the address and then 1283 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1284 // remove the scaling of the offset to recover the original index. 1285 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1286 // Peel the shift off directly but wrap it in a dummy node 1287 // since Ideal can't return existing nodes 1288 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1289 } else if (result->is_Add() && result->in(2)->is_Con() && 1290 result->in(1)->Opcode() == Op_LShiftX && 1291 result->in(1)->in(2) == phase->intcon(shift)) { 1292 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1293 // but for boxing cache access we know that X<<Z will not overflow 1294 // (there is range check) so we do this optimizatrion by hand here. 1295 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1296 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1297 } else { 1298 result = new RShiftXNode(result, phase->intcon(shift)); 1299 } 1300 #ifdef _LP64 1301 if (bt != T_LONG) { 1302 result = new ConvL2INode(phase->transform(result)); 1303 } 1304 #else 1305 if (bt == T_LONG) { 1306 result = new ConvI2LNode(phase->transform(result)); 1307 } 1308 #endif 1309 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1310 // Need to preserve unboxing load type if it is unsigned. 1311 switch(this->Opcode()) { 1312 case Op_LoadUB: 1313 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1314 break; 1315 case Op_LoadUS: 1316 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1317 break; 1318 } 1319 return result; 1320 } 1321 } 1322 } 1323 } 1324 } 1325 return NULL; 1326 } 1327 1328 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1329 Node* region = phi->in(0); 1330 if (region == NULL) { 1331 return false; // Wait stable graph 1332 } 1333 uint cnt = phi->req(); 1334 for (uint i = 1; i < cnt; i++) { 1335 Node* rc = region->in(i); 1336 if (rc == NULL || phase->type(rc) == Type::TOP) 1337 return false; // Wait stable graph 1338 Node* in = phi->in(i); 1339 if (in == NULL || phase->type(in) == Type::TOP) 1340 return false; // Wait stable graph 1341 } 1342 return true; 1343 } 1344 //------------------------------split_through_phi------------------------------ 1345 // Split instance or boxed field load through Phi. 1346 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1347 Node* mem = in(Memory); 1348 Node* address = in(Address); 1349 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1350 1351 assert((t_oop != NULL) && 1352 (t_oop->is_known_instance_field() || 1353 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1354 1355 Compile* C = phase->C; 1356 intptr_t ignore = 0; 1357 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1358 bool base_is_phi = (base != NULL) && base->is_Phi(); 1359 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1360 (base != NULL) && (base == address->in(AddPNode::Base)) && 1361 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1362 1363 if (!((mem->is_Phi() || base_is_phi) && 1364 (load_boxed_values || t_oop->is_known_instance_field()))) { 1365 return NULL; // memory is not Phi 1366 } 1367 1368 if (mem->is_Phi()) { 1369 if (!stable_phi(mem->as_Phi(), phase)) { 1370 return NULL; // Wait stable graph 1371 } 1372 uint cnt = mem->req(); 1373 // Check for loop invariant memory. 1374 if (cnt == 3) { 1375 for (uint i = 1; i < cnt; i++) { 1376 Node* in = mem->in(i); 1377 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1378 if (m == mem) { 1379 set_req(Memory, mem->in(cnt - i)); 1380 return this; // made change 1381 } 1382 } 1383 } 1384 } 1385 if (base_is_phi) { 1386 if (!stable_phi(base->as_Phi(), phase)) { 1387 return NULL; // Wait stable graph 1388 } 1389 uint cnt = base->req(); 1390 // Check for loop invariant memory. 1391 if (cnt == 3) { 1392 for (uint i = 1; i < cnt; i++) { 1393 if (base->in(i) == base) { 1394 return NULL; // Wait stable graph 1395 } 1396 } 1397 } 1398 } 1399 1400 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1401 1402 // Split through Phi (see original code in loopopts.cpp). 1403 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1404 1405 // Do nothing here if Identity will find a value 1406 // (to avoid infinite chain of value phis generation). 1407 if (!phase->eqv(this, this->Identity(phase))) 1408 return NULL; 1409 1410 // Select Region to split through. 1411 Node* region; 1412 if (!base_is_phi) { 1413 assert(mem->is_Phi(), "sanity"); 1414 region = mem->in(0); 1415 // Skip if the region dominates some control edge of the address. 1416 if (!MemNode::all_controls_dominate(address, region)) 1417 return NULL; 1418 } else if (!mem->is_Phi()) { 1419 assert(base_is_phi, "sanity"); 1420 region = base->in(0); 1421 // Skip if the region dominates some control edge of the memory. 1422 if (!MemNode::all_controls_dominate(mem, region)) 1423 return NULL; 1424 } else if (base->in(0) != mem->in(0)) { 1425 assert(base_is_phi && mem->is_Phi(), "sanity"); 1426 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1427 region = base->in(0); 1428 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1429 region = mem->in(0); 1430 } else { 1431 return NULL; // complex graph 1432 } 1433 } else { 1434 assert(base->in(0) == mem->in(0), "sanity"); 1435 region = mem->in(0); 1436 } 1437 1438 const Type* this_type = this->bottom_type(); 1439 int this_index = C->get_alias_index(t_oop); 1440 int this_offset = t_oop->offset(); 1441 int this_iid = t_oop->instance_id(); 1442 if (!t_oop->is_known_instance() && load_boxed_values) { 1443 // Use _idx of address base for boxed values. 1444 this_iid = base->_idx; 1445 } 1446 PhaseIterGVN* igvn = phase->is_IterGVN(); 1447 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1448 for (uint i = 1; i < region->req(); i++) { 1449 Node* x; 1450 Node* the_clone = NULL; 1451 if (region->in(i) == C->top()) { 1452 x = C->top(); // Dead path? Use a dead data op 1453 } else { 1454 x = this->clone(); // Else clone up the data op 1455 the_clone = x; // Remember for possible deletion. 1456 // Alter data node to use pre-phi inputs 1457 if (this->in(0) == region) { 1458 x->set_req(0, region->in(i)); 1459 } else { 1460 x->set_req(0, NULL); 1461 } 1462 if (mem->is_Phi() && (mem->in(0) == region)) { 1463 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1464 } 1465 if (address->is_Phi() && address->in(0) == region) { 1466 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1467 } 1468 if (base_is_phi && (base->in(0) == region)) { 1469 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1470 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1471 x->set_req(Address, adr_x); 1472 } 1473 } 1474 // Check for a 'win' on some paths 1475 const Type *t = x->Value(igvn); 1476 1477 bool singleton = t->singleton(); 1478 1479 // See comments in PhaseIdealLoop::split_thru_phi(). 1480 if (singleton && t == Type::TOP) { 1481 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1482 } 1483 1484 if (singleton) { 1485 x = igvn->makecon(t); 1486 } else { 1487 // We now call Identity to try to simplify the cloned node. 1488 // Note that some Identity methods call phase->type(this). 1489 // Make sure that the type array is big enough for 1490 // our new node, even though we may throw the node away. 1491 // (This tweaking with igvn only works because x is a new node.) 1492 igvn->set_type(x, t); 1493 // If x is a TypeNode, capture any more-precise type permanently into Node 1494 // otherwise it will be not updated during igvn->transform since 1495 // igvn->type(x) is set to x->Value() already. 1496 x->raise_bottom_type(t); 1497 Node *y = x->Identity(igvn); 1498 if (y != x) { 1499 x = y; 1500 } else { 1501 y = igvn->hash_find_insert(x); 1502 if (y) { 1503 x = y; 1504 } else { 1505 // Else x is a new node we are keeping 1506 // We do not need register_new_node_with_optimizer 1507 // because set_type has already been called. 1508 igvn->_worklist.push(x); 1509 } 1510 } 1511 } 1512 if (x != the_clone && the_clone != NULL) { 1513 igvn->remove_dead_node(the_clone); 1514 } 1515 phi->set_req(i, x); 1516 } 1517 // Record Phi 1518 igvn->register_new_node_with_optimizer(phi); 1519 return phi; 1520 } 1521 1522 //------------------------------Ideal------------------------------------------ 1523 // If the load is from Field memory and the pointer is non-null, it might be possible to 1524 // zero out the control input. 1525 // If the offset is constant and the base is an object allocation, 1526 // try to hook me up to the exact initializing store. 1527 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1528 Node* p = MemNode::Ideal_common(phase, can_reshape); 1529 if (p) return (p == NodeSentinel) ? NULL : p; 1530 1531 Node* ctrl = in(MemNode::Control); 1532 Node* address = in(MemNode::Address); 1533 bool progress = false; 1534 1535 // Skip up past a SafePoint control. Cannot do this for Stores because 1536 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1537 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1538 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1539 ctrl = ctrl->in(0); 1540 set_req(MemNode::Control,ctrl); 1541 progress = true; 1542 } 1543 1544 intptr_t ignore = 0; 1545 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1546 if (base != NULL 1547 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1548 // Check for useless control edge in some common special cases 1549 if (in(MemNode::Control) != NULL 1550 && can_remove_control() 1551 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1552 && all_controls_dominate(base, phase->C->start())) { 1553 // A method-invariant, non-null address (constant or 'this' argument). 1554 set_req(MemNode::Control, NULL); 1555 progress = true; 1556 } 1557 } 1558 1559 Node* mem = in(MemNode::Memory); 1560 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1561 1562 if (can_reshape && (addr_t != NULL)) { 1563 // try to optimize our memory input 1564 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1565 if (opt_mem != mem) { 1566 set_req(MemNode::Memory, opt_mem); 1567 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1568 return this; 1569 } 1570 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1571 if ((t_oop != NULL) && 1572 (t_oop->is_known_instance_field() || 1573 t_oop->is_ptr_to_boxed_value())) { 1574 PhaseIterGVN *igvn = phase->is_IterGVN(); 1575 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1576 // Delay this transformation until memory Phi is processed. 1577 phase->is_IterGVN()->_worklist.push(this); 1578 return NULL; 1579 } 1580 // Split instance field load through Phi. 1581 Node* result = split_through_phi(phase); 1582 if (result != NULL) return result; 1583 1584 if (t_oop->is_ptr_to_boxed_value()) { 1585 Node* result = eliminate_autobox(phase); 1586 if (result != NULL) return result; 1587 } 1588 } 1589 } 1590 1591 // Is there a dominating load that loads the same value? Leave 1592 // anything that is not a load of a field/array element (like 1593 // barriers etc.) alone 1594 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) { 1595 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1596 Node *use = mem->fast_out(i); 1597 if (use != this && 1598 use->Opcode() == Opcode() && 1599 use->in(0) != NULL && 1600 use->in(0) != in(0) && 1601 use->in(Address) == in(Address)) { 1602 Node* ctl = in(0); 1603 for (int i = 0; i < 10 && ctl != NULL; i++) { 1604 ctl = IfNode::up_one_dom(ctl); 1605 if (ctl == use->in(0)) { 1606 set_req(0, use->in(0)); 1607 return this; 1608 } 1609 } 1610 } 1611 } 1612 } 1613 1614 // Check for prior store with a different base or offset; make Load 1615 // independent. Skip through any number of them. Bail out if the stores 1616 // are in an endless dead cycle and report no progress. This is a key 1617 // transform for Reflection. However, if after skipping through the Stores 1618 // we can't then fold up against a prior store do NOT do the transform as 1619 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1620 // array memory alive twice: once for the hoisted Load and again after the 1621 // bypassed Store. This situation only works if EVERYBODY who does 1622 // anti-dependence work knows how to bypass. I.e. we need all 1623 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1624 // the alias index stuff. So instead, peek through Stores and IFF we can 1625 // fold up, do so. 1626 Node* prev_mem = find_previous_store(phase); 1627 if (prev_mem != NULL) { 1628 Node* value = can_see_arraycopy_value(prev_mem, phase); 1629 if (value != NULL) { 1630 return value; 1631 } 1632 } 1633 // Steps (a), (b): Walk past independent stores to find an exact match. 1634 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1635 // (c) See if we can fold up on the spot, but don't fold up here. 1636 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1637 // just return a prior value, which is done by Identity calls. 1638 if (can_see_stored_value(prev_mem, phase)) { 1639 // Make ready for step (d): 1640 set_req(MemNode::Memory, prev_mem); 1641 return this; 1642 } 1643 } 1644 1645 return progress ? this : NULL; 1646 } 1647 1648 // Helper to recognize certain Klass fields which are invariant across 1649 // some group of array types (e.g., int[] or all T[] where T < Object). 1650 const Type* 1651 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1652 ciKlass* klass) const { 1653 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1654 // The field is Klass::_modifier_flags. Return its (constant) value. 1655 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1656 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1657 return TypeInt::make(klass->modifier_flags()); 1658 } 1659 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1660 // The field is Klass::_access_flags. Return its (constant) value. 1661 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1662 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1663 return TypeInt::make(klass->access_flags()); 1664 } 1665 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1666 // The field is Klass::_layout_helper. Return its constant value if known. 1667 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1668 return TypeInt::make(klass->layout_helper()); 1669 } 1670 1671 // No match. 1672 return NULL; 1673 } 1674 1675 //------------------------------Value----------------------------------------- 1676 const Type* LoadNode::Value(PhaseGVN* phase) const { 1677 // Either input is TOP ==> the result is TOP 1678 Node* mem = in(MemNode::Memory); 1679 const Type *t1 = phase->type(mem); 1680 if (t1 == Type::TOP) return Type::TOP; 1681 Node* adr = in(MemNode::Address); 1682 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1683 if (tp == NULL || tp->empty()) return Type::TOP; 1684 int off = tp->offset(); 1685 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1686 Compile* C = phase->C; 1687 1688 // Try to guess loaded type from pointer type 1689 if (tp->isa_aryptr()) { 1690 const TypeAryPtr* ary = tp->is_aryptr(); 1691 const Type* t = ary->elem(); 1692 1693 // Determine whether the reference is beyond the header or not, by comparing 1694 // the offset against the offset of the start of the array's data. 1695 // Different array types begin at slightly different offsets (12 vs. 16). 1696 // We choose T_BYTE as an example base type that is least restrictive 1697 // as to alignment, which will therefore produce the smallest 1698 // possible base offset. 1699 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1700 const bool off_beyond_header = (off >= min_base_off); 1701 1702 // Try to constant-fold a stable array element. 1703 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1704 // Make sure the reference is not into the header and the offset is constant 1705 ciObject* aobj = ary->const_oop(); 1706 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1707 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1708 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1709 stable_dimension, 1710 memory_type(), is_unsigned()); 1711 if (con_type != NULL) { 1712 return con_type; 1713 } 1714 } 1715 } 1716 1717 // Don't do this for integer types. There is only potential profit if 1718 // the element type t is lower than _type; that is, for int types, if _type is 1719 // more restrictive than t. This only happens here if one is short and the other 1720 // char (both 16 bits), and in those cases we've made an intentional decision 1721 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1722 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1723 // 1724 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1725 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1726 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1727 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1728 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1729 // In fact, that could have been the original type of p1, and p1 could have 1730 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1731 // expression (LShiftL quux 3) independently optimized to the constant 8. 1732 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1733 && (_type->isa_vect() == NULL) 1734 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1735 // t might actually be lower than _type, if _type is a unique 1736 // concrete subclass of abstract class t. 1737 if (off_beyond_header) { // is the offset beyond the header? 1738 const Type* jt = t->join_speculative(_type); 1739 // In any case, do not allow the join, per se, to empty out the type. 1740 if (jt->empty() && !t->empty()) { 1741 // This can happen if a interface-typed array narrows to a class type. 1742 jt = _type; 1743 } 1744 #ifdef ASSERT 1745 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1746 // The pointers in the autobox arrays are always non-null 1747 Node* base = adr->in(AddPNode::Base); 1748 if ((base != NULL) && base->is_DecodeN()) { 1749 // Get LoadN node which loads IntegerCache.cache field 1750 base = base->in(1); 1751 } 1752 if ((base != NULL) && base->is_Con()) { 1753 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1754 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1755 // It could be narrow oop 1756 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1757 } 1758 } 1759 } 1760 #endif 1761 return jt; 1762 } 1763 } 1764 } else if (tp->base() == Type::InstPtr) { 1765 assert( off != Type::OffsetBot || 1766 // arrays can be cast to Objects 1767 tp->is_oopptr()->klass()->is_java_lang_Object() || 1768 // unsafe field access may not have a constant offset 1769 C->has_unsafe_access(), 1770 "Field accesses must be precise" ); 1771 // For oop loads, we expect the _type to be precise. 1772 1773 // Optimize loads from constant fields. 1774 const TypeInstPtr* tinst = tp->is_instptr(); 1775 ciObject* const_oop = tinst->const_oop(); 1776 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1777 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type()); 1778 if (con_type != NULL) { 1779 return con_type; 1780 } 1781 } 1782 } else if (tp->base() == Type::KlassPtr) { 1783 assert( off != Type::OffsetBot || 1784 // arrays can be cast to Objects 1785 tp->is_klassptr()->klass()->is_java_lang_Object() || 1786 // also allow array-loading from the primary supertype 1787 // array during subtype checks 1788 Opcode() == Op_LoadKlass, 1789 "Field accesses must be precise" ); 1790 // For klass/static loads, we expect the _type to be precise 1791 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) { 1792 /* With mirrors being an indirect in the Klass* 1793 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 1794 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 1795 * 1796 * So check the type and klass of the node before the LoadP. 1797 */ 1798 Node* adr2 = adr->in(MemNode::Address); 1799 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1800 if (tkls != NULL && !StressReflectiveCode) { 1801 ciKlass* klass = tkls->klass(); 1802 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1803 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1804 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1805 return TypeInstPtr::make(klass->java_mirror()); 1806 } 1807 } 1808 } 1809 1810 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1811 if (tkls != NULL && !StressReflectiveCode) { 1812 ciKlass* klass = tkls->klass(); 1813 if (klass->is_loaded() && tkls->klass_is_exact()) { 1814 // We are loading a field from a Klass metaobject whose identity 1815 // is known at compile time (the type is "exact" or "precise"). 1816 // Check for fields we know are maintained as constants by the VM. 1817 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1818 // The field is Klass::_super_check_offset. Return its (constant) value. 1819 // (Folds up type checking code.) 1820 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1821 return TypeInt::make(klass->super_check_offset()); 1822 } 1823 // Compute index into primary_supers array 1824 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1825 // Check for overflowing; use unsigned compare to handle the negative case. 1826 if( depth < ciKlass::primary_super_limit() ) { 1827 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1828 // (Folds up type checking code.) 1829 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1830 ciKlass *ss = klass->super_of_depth(depth); 1831 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1832 } 1833 const Type* aift = load_array_final_field(tkls, klass); 1834 if (aift != NULL) return aift; 1835 } 1836 1837 // We can still check if we are loading from the primary_supers array at a 1838 // shallow enough depth. Even though the klass is not exact, entries less 1839 // than or equal to its super depth are correct. 1840 if (klass->is_loaded() ) { 1841 ciType *inner = klass; 1842 while( inner->is_obj_array_klass() ) 1843 inner = inner->as_obj_array_klass()->base_element_type(); 1844 if( inner->is_instance_klass() && 1845 !inner->as_instance_klass()->flags().is_interface() ) { 1846 // Compute index into primary_supers array 1847 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1848 // Check for overflowing; use unsigned compare to handle the negative case. 1849 if( depth < ciKlass::primary_super_limit() && 1850 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1851 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1852 // (Folds up type checking code.) 1853 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1854 ciKlass *ss = klass->super_of_depth(depth); 1855 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1856 } 1857 } 1858 } 1859 1860 // If the type is enough to determine that the thing is not an array, 1861 // we can give the layout_helper a positive interval type. 1862 // This will help short-circuit some reflective code. 1863 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1864 && !klass->is_array_klass() // not directly typed as an array 1865 && !klass->is_interface() // specifically not Serializable & Cloneable 1866 && !klass->is_java_lang_Object() // not the supertype of all T[] 1867 ) { 1868 // Note: When interfaces are reliable, we can narrow the interface 1869 // test to (klass != Serializable && klass != Cloneable). 1870 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1871 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1872 // The key property of this type is that it folds up tests 1873 // for array-ness, since it proves that the layout_helper is positive. 1874 // Thus, a generic value like the basic object layout helper works fine. 1875 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1876 } 1877 } 1878 1879 // If we are loading from a freshly-allocated object, produce a zero, 1880 // if the load is provably beyond the header of the object. 1881 // (Also allow a variable load from a fresh array to produce zero.) 1882 const TypeOopPtr *tinst = tp->isa_oopptr(); 1883 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1884 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1885 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1886 Node* value = can_see_stored_value(mem,phase); 1887 if (value != NULL && value->is_Con()) { 1888 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1889 return value->bottom_type(); 1890 } 1891 } 1892 1893 if (is_instance) { 1894 // If we have an instance type and our memory input is the 1895 // programs's initial memory state, there is no matching store, 1896 // so just return a zero of the appropriate type 1897 Node *mem = in(MemNode::Memory); 1898 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1899 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1900 return Type::get_zero_type(_type->basic_type()); 1901 } 1902 } 1903 return _type; 1904 } 1905 1906 //------------------------------match_edge------------------------------------- 1907 // Do we Match on this edge index or not? Match only the address. 1908 uint LoadNode::match_edge(uint idx) const { 1909 return idx == MemNode::Address; 1910 } 1911 1912 //--------------------------LoadBNode::Ideal-------------------------------------- 1913 // 1914 // If the previous store is to the same address as this load, 1915 // and the value stored was larger than a byte, replace this load 1916 // with the value stored truncated to a byte. If no truncation is 1917 // needed, the replacement is done in LoadNode::Identity(). 1918 // 1919 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1920 Node* mem = in(MemNode::Memory); 1921 Node* value = can_see_stored_value(mem,phase); 1922 if( value && !phase->type(value)->higher_equal( _type ) ) { 1923 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1924 return new RShiftINode(result, phase->intcon(24)); 1925 } 1926 // Identity call will handle the case where truncation is not needed. 1927 return LoadNode::Ideal(phase, can_reshape); 1928 } 1929 1930 const Type* LoadBNode::Value(PhaseGVN* 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 << 24) >> 24); 1940 } 1941 return LoadNode::Value(phase); 1942 } 1943 1944 //--------------------------LoadUBNode::Ideal------------------------------------- 1945 // 1946 // If the previous store is to the same address as this load, 1947 // and the value stored was larger than a byte, replace this load 1948 // with the value stored truncated to a byte. If no truncation is 1949 // needed, the replacement is done in LoadNode::Identity(). 1950 // 1951 Node* LoadUBNode::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 return new AndINode(value, phase->intcon(0xFF)); 1956 // Identity call will handle the case where truncation is not needed. 1957 return LoadNode::Ideal(phase, can_reshape); 1958 } 1959 1960 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 1961 Node* mem = in(MemNode::Memory); 1962 Node* value = can_see_stored_value(mem,phase); 1963 if (value != NULL && value->is_Con() && 1964 !value->bottom_type()->higher_equal(_type)) { 1965 // If the input to the store does not fit with the load's result type, 1966 // it must be truncated. We can't delay until Ideal call since 1967 // a singleton Value is needed for split_thru_phi optimization. 1968 int con = value->get_int(); 1969 return TypeInt::make(con & 0xFF); 1970 } 1971 return LoadNode::Value(phase); 1972 } 1973 1974 //--------------------------LoadUSNode::Ideal------------------------------------- 1975 // 1976 // If the previous store is to the same address as this load, 1977 // and the value stored was larger than a char, replace this load 1978 // with the value stored truncated to a char. If no truncation is 1979 // needed, the replacement is done in LoadNode::Identity(). 1980 // 1981 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1982 Node* mem = in(MemNode::Memory); 1983 Node* value = can_see_stored_value(mem,phase); 1984 if( value && !phase->type(value)->higher_equal( _type ) ) 1985 return new AndINode(value,phase->intcon(0xFFFF)); 1986 // Identity call will handle the case where truncation is not needed. 1987 return LoadNode::Ideal(phase, can_reshape); 1988 } 1989 1990 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 1991 Node* mem = in(MemNode::Memory); 1992 Node* value = can_see_stored_value(mem,phase); 1993 if (value != NULL && value->is_Con() && 1994 !value->bottom_type()->higher_equal(_type)) { 1995 // If the input to the store does not fit with the load's result type, 1996 // it must be truncated. We can't delay until Ideal call since 1997 // a singleton Value is needed for split_thru_phi optimization. 1998 int con = value->get_int(); 1999 return TypeInt::make(con & 0xFFFF); 2000 } 2001 return LoadNode::Value(phase); 2002 } 2003 2004 //--------------------------LoadSNode::Ideal-------------------------------------- 2005 // 2006 // If the previous store is to the same address as this load, 2007 // and the value stored was larger than a short, replace this load 2008 // with the value stored truncated to a short. If no truncation is 2009 // needed, the replacement is done in LoadNode::Identity(). 2010 // 2011 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2012 Node* mem = in(MemNode::Memory); 2013 Node* value = can_see_stored_value(mem,phase); 2014 if( value && !phase->type(value)->higher_equal( _type ) ) { 2015 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 2016 return new RShiftINode(result, phase->intcon(16)); 2017 } 2018 // Identity call will handle the case where truncation is not needed. 2019 return LoadNode::Ideal(phase, can_reshape); 2020 } 2021 2022 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2023 Node* mem = in(MemNode::Memory); 2024 Node* value = can_see_stored_value(mem,phase); 2025 if (value != NULL && value->is_Con() && 2026 !value->bottom_type()->higher_equal(_type)) { 2027 // If the input to the store does not fit with the load's result type, 2028 // it must be truncated. We can't delay until Ideal call since 2029 // a singleton Value is needed for split_thru_phi optimization. 2030 int con = value->get_int(); 2031 return TypeInt::make((con << 16) >> 16); 2032 } 2033 return LoadNode::Value(phase); 2034 } 2035 2036 //============================================================================= 2037 //----------------------------LoadKlassNode::make------------------------------ 2038 // Polymorphic factory method: 2039 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2040 // sanity check the alias category against the created node type 2041 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2042 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2043 #ifdef _LP64 2044 if (adr_type->is_ptr_to_narrowklass()) { 2045 assert(UseCompressedClassPointers, "no compressed klasses"); 2046 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2047 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2048 } 2049 #endif 2050 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2051 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2052 } 2053 2054 //------------------------------Value------------------------------------------ 2055 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2056 return klass_value_common(phase); 2057 } 2058 2059 // In most cases, LoadKlassNode does not have the control input set. If the control 2060 // input is set, it must not be removed (by LoadNode::Ideal()). 2061 bool LoadKlassNode::can_remove_control() const { 2062 return false; 2063 } 2064 2065 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2066 // Either input is TOP ==> the result is TOP 2067 const Type *t1 = phase->type( in(MemNode::Memory) ); 2068 if (t1 == Type::TOP) return Type::TOP; 2069 Node *adr = in(MemNode::Address); 2070 const Type *t2 = phase->type( adr ); 2071 if (t2 == Type::TOP) return Type::TOP; 2072 const TypePtr *tp = t2->is_ptr(); 2073 if (TypePtr::above_centerline(tp->ptr()) || 2074 tp->ptr() == TypePtr::Null) return Type::TOP; 2075 2076 // Return a more precise klass, if possible 2077 const TypeInstPtr *tinst = tp->isa_instptr(); 2078 if (tinst != NULL) { 2079 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2080 int offset = tinst->offset(); 2081 if (ik == phase->C->env()->Class_klass() 2082 && (offset == java_lang_Class::klass_offset_in_bytes() || 2083 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2084 // We are loading a special hidden field from a Class mirror object, 2085 // the field which points to the VM's Klass metaobject. 2086 ciType* t = tinst->java_mirror_type(); 2087 // java_mirror_type returns non-null for compile-time Class constants. 2088 if (t != NULL) { 2089 // constant oop => constant klass 2090 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2091 if (t->is_void()) { 2092 // We cannot create a void array. Since void is a primitive type return null 2093 // klass. Users of this result need to do a null check on the returned klass. 2094 return TypePtr::NULL_PTR; 2095 } 2096 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2097 } 2098 if (!t->is_klass()) { 2099 // a primitive Class (e.g., int.class) has NULL for a klass field 2100 return TypePtr::NULL_PTR; 2101 } 2102 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2103 return TypeKlassPtr::make(t->as_klass()); 2104 } 2105 // non-constant mirror, so we can't tell what's going on 2106 } 2107 if( !ik->is_loaded() ) 2108 return _type; // Bail out if not loaded 2109 if (offset == oopDesc::klass_offset_in_bytes()) { 2110 if (tinst->klass_is_exact()) { 2111 return TypeKlassPtr::make(ik); 2112 } 2113 // See if we can become precise: no subklasses and no interface 2114 // (Note: We need to support verified interfaces.) 2115 if (!ik->is_interface() && !ik->has_subklass()) { 2116 //assert(!UseExactTypes, "this code should be useless with exact types"); 2117 // Add a dependence; if any subclass added we need to recompile 2118 if (!ik->is_final()) { 2119 // %%% should use stronger assert_unique_concrete_subtype instead 2120 phase->C->dependencies()->assert_leaf_type(ik); 2121 } 2122 // Return precise klass 2123 return TypeKlassPtr::make(ik); 2124 } 2125 2126 // Return root of possible klass 2127 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2128 } 2129 } 2130 2131 // Check for loading klass from an array 2132 const TypeAryPtr *tary = tp->isa_aryptr(); 2133 if( tary != NULL ) { 2134 ciKlass *tary_klass = tary->klass(); 2135 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2136 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2137 if (tary->klass_is_exact()) { 2138 return TypeKlassPtr::make(tary_klass); 2139 } 2140 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2141 // If the klass is an object array, we defer the question to the 2142 // array component klass. 2143 if( ak->is_obj_array_klass() ) { 2144 assert( ak->is_loaded(), "" ); 2145 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2146 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2147 ciInstanceKlass* ik = base_k->as_instance_klass(); 2148 // See if we can become precise: no subklasses and no interface 2149 if (!ik->is_interface() && !ik->has_subklass()) { 2150 //assert(!UseExactTypes, "this code should be useless with exact types"); 2151 // Add a dependence; if any subclass added we need to recompile 2152 if (!ik->is_final()) { 2153 phase->C->dependencies()->assert_leaf_type(ik); 2154 } 2155 // Return precise array klass 2156 return TypeKlassPtr::make(ak); 2157 } 2158 } 2159 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2160 } else { // Found a type-array? 2161 //assert(!UseExactTypes, "this code should be useless with exact types"); 2162 assert( ak->is_type_array_klass(), "" ); 2163 return TypeKlassPtr::make(ak); // These are always precise 2164 } 2165 } 2166 } 2167 2168 // Check for loading klass from an array klass 2169 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2170 if (tkls != NULL && !StressReflectiveCode) { 2171 ciKlass* klass = tkls->klass(); 2172 if( !klass->is_loaded() ) 2173 return _type; // Bail out if not loaded 2174 if( klass->is_obj_array_klass() && 2175 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2176 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2177 // // Always returning precise element type is incorrect, 2178 // // e.g., element type could be object and array may contain strings 2179 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2180 2181 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2182 // according to the element type's subclassing. 2183 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2184 } 2185 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2186 tkls->offset() == in_bytes(Klass::super_offset())) { 2187 ciKlass* sup = klass->as_instance_klass()->super(); 2188 // The field is Klass::_super. Return its (constant) value. 2189 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2190 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2191 } 2192 } 2193 2194 // Bailout case 2195 return LoadNode::Value(phase); 2196 } 2197 2198 //------------------------------Identity--------------------------------------- 2199 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2200 // Also feed through the klass in Allocate(...klass...)._klass. 2201 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2202 return klass_identity_common(phase); 2203 } 2204 2205 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2206 Node* x = LoadNode::Identity(phase); 2207 if (x != this) return x; 2208 2209 // Take apart the address into an oop and and offset. 2210 // Return 'this' if we cannot. 2211 Node* adr = in(MemNode::Address); 2212 intptr_t offset = 0; 2213 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2214 if (base == NULL) return this; 2215 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2216 if (toop == NULL) return this; 2217 2218 // Step over potential GC barrier for OopHandle resolve 2219 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2220 if (bs->is_gc_barrier_node(base)) { 2221 base = bs->step_over_gc_barrier(base); 2222 } 2223 2224 // We can fetch the klass directly through an AllocateNode. 2225 // This works even if the klass is not constant (clone or newArray). 2226 if (offset == oopDesc::klass_offset_in_bytes()) { 2227 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2228 if (allocated_klass != NULL) { 2229 return allocated_klass; 2230 } 2231 } 2232 2233 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2234 // See inline_native_Class_query for occurrences of these patterns. 2235 // Java Example: x.getClass().isAssignableFrom(y) 2236 // 2237 // This improves reflective code, often making the Class 2238 // mirror go completely dead. (Current exception: Class 2239 // mirrors may appear in debug info, but we could clean them out by 2240 // introducing a new debug info operator for Klass.java_mirror). 2241 2242 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2243 && offset == java_lang_Class::klass_offset_in_bytes()) { 2244 if (base->is_Load()) { 2245 Node* base2 = base->in(MemNode::Address); 2246 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2247 Node* adr2 = base2->in(MemNode::Address); 2248 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2249 if (tkls != NULL && !tkls->empty() 2250 && (tkls->klass()->is_instance_klass() || 2251 tkls->klass()->is_array_klass()) 2252 && adr2->is_AddP() 2253 ) { 2254 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2255 if (tkls->offset() == mirror_field) { 2256 return adr2->in(AddPNode::Base); 2257 } 2258 } 2259 } 2260 } 2261 } 2262 2263 return this; 2264 } 2265 2266 2267 //------------------------------Value------------------------------------------ 2268 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2269 const Type *t = klass_value_common(phase); 2270 if (t == Type::TOP) 2271 return t; 2272 2273 return t->make_narrowklass(); 2274 } 2275 2276 //------------------------------Identity--------------------------------------- 2277 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2278 // Also feed through the klass in Allocate(...klass...)._klass. 2279 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2280 Node *x = klass_identity_common(phase); 2281 2282 const Type *t = phase->type( x ); 2283 if( t == Type::TOP ) return x; 2284 if( t->isa_narrowklass()) return x; 2285 assert (!t->isa_narrowoop(), "no narrow oop here"); 2286 2287 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2288 } 2289 2290 //------------------------------Value----------------------------------------- 2291 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2292 // Either input is TOP ==> the result is TOP 2293 const Type *t1 = phase->type( in(MemNode::Memory) ); 2294 if( t1 == Type::TOP ) return Type::TOP; 2295 Node *adr = in(MemNode::Address); 2296 const Type *t2 = phase->type( adr ); 2297 if( t2 == Type::TOP ) return Type::TOP; 2298 const TypePtr *tp = t2->is_ptr(); 2299 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2300 const TypeAryPtr *tap = tp->isa_aryptr(); 2301 if( !tap ) return _type; 2302 return tap->size(); 2303 } 2304 2305 //-------------------------------Ideal--------------------------------------- 2306 // Feed through the length in AllocateArray(...length...)._length. 2307 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2308 Node* p = MemNode::Ideal_common(phase, can_reshape); 2309 if (p) return (p == NodeSentinel) ? NULL : p; 2310 2311 // Take apart the address into an oop and and offset. 2312 // Return 'this' if we cannot. 2313 Node* adr = in(MemNode::Address); 2314 intptr_t offset = 0; 2315 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2316 if (base == NULL) return NULL; 2317 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2318 if (tary == NULL) return NULL; 2319 2320 // We can fetch the length directly through an AllocateArrayNode. 2321 // This works even if the length is not constant (clone or newArray). 2322 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2323 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2324 if (alloc != NULL) { 2325 Node* allocated_length = alloc->Ideal_length(); 2326 Node* len = alloc->make_ideal_length(tary, phase); 2327 if (allocated_length != len) { 2328 // New CastII improves on this. 2329 return len; 2330 } 2331 } 2332 } 2333 2334 return NULL; 2335 } 2336 2337 //------------------------------Identity--------------------------------------- 2338 // Feed through the length in AllocateArray(...length...)._length. 2339 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2340 Node* x = LoadINode::Identity(phase); 2341 if (x != this) return x; 2342 2343 // Take apart the address into an oop and and offset. 2344 // Return 'this' if we cannot. 2345 Node* adr = in(MemNode::Address); 2346 intptr_t offset = 0; 2347 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2348 if (base == NULL) return this; 2349 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2350 if (tary == NULL) return this; 2351 2352 // We can fetch the length directly through an AllocateArrayNode. 2353 // This works even if the length is not constant (clone or newArray). 2354 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2355 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2356 if (alloc != NULL) { 2357 Node* allocated_length = alloc->Ideal_length(); 2358 // Do not allow make_ideal_length to allocate a CastII node. 2359 Node* len = alloc->make_ideal_length(tary, phase, false); 2360 if (allocated_length == len) { 2361 // Return allocated_length only if it would not be improved by a CastII. 2362 return allocated_length; 2363 } 2364 } 2365 } 2366 2367 return this; 2368 2369 } 2370 2371 //============================================================================= 2372 //---------------------------StoreNode::make----------------------------------- 2373 // Polymorphic factory method: 2374 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2375 assert((mo == unordered || mo == release), "unexpected"); 2376 Compile* C = gvn.C; 2377 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2378 ctl != NULL, "raw memory operations should have control edge"); 2379 2380 switch (bt) { 2381 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2382 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2383 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2384 case T_CHAR: 2385 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2386 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2387 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2388 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2389 case T_METADATA: 2390 case T_ADDRESS: 2391 case T_OBJECT: 2392 #ifdef _LP64 2393 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2394 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2395 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2396 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2397 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2398 adr->bottom_type()->isa_rawptr())) { 2399 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2400 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2401 } 2402 #endif 2403 { 2404 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2405 } 2406 default: 2407 ShouldNotReachHere(); 2408 return (StoreNode*)NULL; 2409 } 2410 } 2411 2412 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2413 bool require_atomic = true; 2414 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2415 } 2416 2417 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2418 bool require_atomic = true; 2419 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2420 } 2421 2422 2423 //--------------------------bottom_type---------------------------------------- 2424 const Type *StoreNode::bottom_type() const { 2425 return Type::MEMORY; 2426 } 2427 2428 //------------------------------hash------------------------------------------- 2429 uint StoreNode::hash() const { 2430 // unroll addition of interesting fields 2431 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2432 2433 // Since they are not commoned, do not hash them: 2434 return NO_HASH; 2435 } 2436 2437 //------------------------------Ideal------------------------------------------ 2438 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2439 // When a store immediately follows a relevant allocation/initialization, 2440 // try to capture it into the initialization, or hoist it above. 2441 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2442 Node* p = MemNode::Ideal_common(phase, can_reshape); 2443 if (p) return (p == NodeSentinel) ? NULL : p; 2444 2445 Node* mem = in(MemNode::Memory); 2446 Node* address = in(MemNode::Address); 2447 // Back-to-back stores to same address? Fold em up. Generally 2448 // unsafe if I have intervening uses... Also disallowed for StoreCM 2449 // since they must follow each StoreP operation. Redundant StoreCMs 2450 // are eliminated just before matching in final_graph_reshape. 2451 { 2452 Node* st = mem; 2453 // If Store 'st' has more than one use, we cannot fold 'st' away. 2454 // For example, 'st' might be the final state at a conditional 2455 // return. Or, 'st' might be used by some node which is live at 2456 // the same time 'st' is live, which might be unschedulable. So, 2457 // require exactly ONE user until such time as we clone 'mem' for 2458 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2459 // true). 2460 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2461 // Looking at a dead closed cycle of memory? 2462 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2463 assert(Opcode() == st->Opcode() || 2464 st->Opcode() == Op_StoreVector || 2465 Opcode() == Op_StoreVector || 2466 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2467 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2468 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2469 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2470 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2471 2472 if (st->in(MemNode::Address)->eqv_uncast(address) && 2473 st->as_Store()->memory_size() <= this->memory_size()) { 2474 Node* use = st->raw_out(0); 2475 phase->igvn_rehash_node_delayed(use); 2476 if (can_reshape) { 2477 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2478 } else { 2479 // It's OK to do this in the parser, since DU info is always accurate, 2480 // and the parser always refers to nodes via SafePointNode maps. 2481 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2482 } 2483 return this; 2484 } 2485 st = st->in(MemNode::Memory); 2486 } 2487 } 2488 2489 2490 // Capture an unaliased, unconditional, simple store into an initializer. 2491 // Or, if it is independent of the allocation, hoist it above the allocation. 2492 if (ReduceFieldZeroing && /*can_reshape &&*/ 2493 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2494 InitializeNode* init = mem->in(0)->as_Initialize(); 2495 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2496 if (offset > 0) { 2497 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2498 // If the InitializeNode captured me, it made a raw copy of me, 2499 // and I need to disappear. 2500 if (moved != NULL) { 2501 // %%% hack to ensure that Ideal returns a new node: 2502 mem = MergeMemNode::make(mem); 2503 return mem; // fold me away 2504 } 2505 } 2506 } 2507 2508 return NULL; // No further progress 2509 } 2510 2511 //------------------------------Value----------------------------------------- 2512 const Type* StoreNode::Value(PhaseGVN* phase) const { 2513 // Either input is TOP ==> the result is TOP 2514 const Type *t1 = phase->type( in(MemNode::Memory) ); 2515 if( t1 == Type::TOP ) return Type::TOP; 2516 const Type *t2 = phase->type( in(MemNode::Address) ); 2517 if( t2 == Type::TOP ) return Type::TOP; 2518 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2519 if( t3 == Type::TOP ) return Type::TOP; 2520 return Type::MEMORY; 2521 } 2522 2523 //------------------------------Identity--------------------------------------- 2524 // Remove redundant stores: 2525 // Store(m, p, Load(m, p)) changes to m. 2526 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2527 Node* StoreNode::Identity(PhaseGVN* phase) { 2528 Node* mem = in(MemNode::Memory); 2529 Node* adr = in(MemNode::Address); 2530 Node* val = in(MemNode::ValueIn); 2531 2532 Node* result = this; 2533 2534 // Load then Store? Then the Store is useless 2535 if (val->is_Load() && 2536 val->in(MemNode::Address)->eqv_uncast(adr) && 2537 val->in(MemNode::Memory )->eqv_uncast(mem) && 2538 val->as_Load()->store_Opcode() == Opcode()) { 2539 result = mem; 2540 } 2541 2542 // Two stores in a row of the same value? 2543 if (result == this && 2544 mem->is_Store() && 2545 mem->in(MemNode::Address)->eqv_uncast(adr) && 2546 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2547 mem->Opcode() == Opcode()) { 2548 result = mem; 2549 } 2550 2551 // Store of zero anywhere into a freshly-allocated object? 2552 // Then the store is useless. 2553 // (It must already have been captured by the InitializeNode.) 2554 if (result == this && 2555 ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2556 // a newly allocated object is already all-zeroes everywhere 2557 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2558 result = mem; 2559 } 2560 2561 if (result == this) { 2562 // the store may also apply to zero-bits in an earlier object 2563 Node* prev_mem = find_previous_store(phase); 2564 // Steps (a), (b): Walk past independent stores to find an exact match. 2565 if (prev_mem != NULL) { 2566 Node* prev_val = can_see_stored_value(prev_mem, phase); 2567 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2568 // prev_val and val might differ by a cast; it would be good 2569 // to keep the more informative of the two. 2570 result = mem; 2571 } 2572 } 2573 } 2574 } 2575 2576 if (result != this && phase->is_IterGVN() != NULL) { 2577 MemBarNode* trailing = trailing_membar(); 2578 if (trailing != NULL) { 2579 #ifdef ASSERT 2580 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2581 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects"); 2582 #endif 2583 PhaseIterGVN* igvn = phase->is_IterGVN(); 2584 trailing->remove(igvn); 2585 } 2586 } 2587 2588 return result; 2589 } 2590 2591 //------------------------------match_edge------------------------------------- 2592 // Do we Match on this edge index or not? Match only memory & value 2593 uint StoreNode::match_edge(uint idx) const { 2594 return idx == MemNode::Address || idx == MemNode::ValueIn; 2595 } 2596 2597 //------------------------------cmp-------------------------------------------- 2598 // Do not common stores up together. They generally have to be split 2599 // back up anyways, so do not bother. 2600 uint StoreNode::cmp( const Node &n ) const { 2601 return (&n == this); // Always fail except on self 2602 } 2603 2604 //------------------------------Ideal_masked_input----------------------------- 2605 // Check for a useless mask before a partial-word store 2606 // (StoreB ... (AndI valIn conIa) ) 2607 // If (conIa & mask == mask) this simplifies to 2608 // (StoreB ... (valIn) ) 2609 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2610 Node *val = in(MemNode::ValueIn); 2611 if( val->Opcode() == Op_AndI ) { 2612 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2613 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2614 set_req(MemNode::ValueIn, val->in(1)); 2615 return this; 2616 } 2617 } 2618 return NULL; 2619 } 2620 2621 2622 //------------------------------Ideal_sign_extended_input---------------------- 2623 // Check for useless sign-extension before a partial-word store 2624 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2625 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2626 // (StoreB ... (valIn) ) 2627 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2628 Node *val = in(MemNode::ValueIn); 2629 if( val->Opcode() == Op_RShiftI ) { 2630 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2631 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2632 Node *shl = val->in(1); 2633 if( shl->Opcode() == Op_LShiftI ) { 2634 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2635 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2636 set_req(MemNode::ValueIn, shl->in(1)); 2637 return this; 2638 } 2639 } 2640 } 2641 } 2642 return NULL; 2643 } 2644 2645 //------------------------------value_never_loaded----------------------------------- 2646 // Determine whether there are any possible loads of the value stored. 2647 // For simplicity, we actually check if there are any loads from the 2648 // address stored to, not just for loads of the value stored by this node. 2649 // 2650 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2651 Node *adr = in(Address); 2652 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2653 if (adr_oop == NULL) 2654 return false; 2655 if (!adr_oop->is_known_instance_field()) 2656 return false; // if not a distinct instance, there may be aliases of the address 2657 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2658 Node *use = adr->fast_out(i); 2659 if (use->is_Load() || use->is_LoadStore()) { 2660 return false; 2661 } 2662 } 2663 return true; 2664 } 2665 2666 MemBarNode* StoreNode::trailing_membar() const { 2667 if (is_release()) { 2668 MemBarNode* trailing_mb = NULL; 2669 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2670 Node* u = fast_out(i); 2671 if (u->is_MemBar()) { 2672 if (u->as_MemBar()->trailing_store()) { 2673 assert(u->Opcode() == Op_MemBarVolatile, ""); 2674 assert(trailing_mb == NULL, "only one"); 2675 trailing_mb = u->as_MemBar(); 2676 #ifdef ASSERT 2677 Node* leading = u->as_MemBar()->leading_membar(); 2678 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2679 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2680 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2681 #endif 2682 } else { 2683 assert(u->as_MemBar()->standalone(), ""); 2684 } 2685 } 2686 } 2687 return trailing_mb; 2688 } 2689 return NULL; 2690 } 2691 2692 2693 //============================================================================= 2694 //------------------------------Ideal------------------------------------------ 2695 // If the store is from an AND mask that leaves the low bits untouched, then 2696 // we can skip the AND operation. If the store is from a sign-extension 2697 // (a left shift, then right shift) we can skip both. 2698 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2699 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2700 if( progress != NULL ) return progress; 2701 2702 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2703 if( progress != NULL ) return progress; 2704 2705 // Finally check the default case 2706 return StoreNode::Ideal(phase, can_reshape); 2707 } 2708 2709 //============================================================================= 2710 //------------------------------Ideal------------------------------------------ 2711 // If the store is from an AND mask that leaves the low bits untouched, then 2712 // we can skip the AND operation 2713 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2714 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2715 if( progress != NULL ) return progress; 2716 2717 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2718 if( progress != NULL ) return progress; 2719 2720 // Finally check the default case 2721 return StoreNode::Ideal(phase, can_reshape); 2722 } 2723 2724 //============================================================================= 2725 //------------------------------Identity--------------------------------------- 2726 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2727 // No need to card mark when storing a null ptr 2728 Node* my_store = in(MemNode::OopStore); 2729 if (my_store->is_Store()) { 2730 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2731 if( t1 == TypePtr::NULL_PTR ) { 2732 return in(MemNode::Memory); 2733 } 2734 } 2735 return this; 2736 } 2737 2738 //============================================================================= 2739 //------------------------------Ideal--------------------------------------- 2740 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2741 Node* progress = StoreNode::Ideal(phase, can_reshape); 2742 if (progress != NULL) return progress; 2743 2744 Node* my_store = in(MemNode::OopStore); 2745 if (my_store->is_MergeMem()) { 2746 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2747 set_req(MemNode::OopStore, mem); 2748 return this; 2749 } 2750 2751 return NULL; 2752 } 2753 2754 //------------------------------Value----------------------------------------- 2755 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2756 // Either input is TOP ==> the result is TOP 2757 const Type *t = phase->type( in(MemNode::Memory) ); 2758 if( t == Type::TOP ) return Type::TOP; 2759 t = phase->type( in(MemNode::Address) ); 2760 if( t == Type::TOP ) return Type::TOP; 2761 t = phase->type( in(MemNode::ValueIn) ); 2762 if( t == Type::TOP ) return Type::TOP; 2763 // If extra input is TOP ==> the result is TOP 2764 t = phase->type( in(MemNode::OopStore) ); 2765 if( t == Type::TOP ) return Type::TOP; 2766 2767 return StoreNode::Value( phase ); 2768 } 2769 2770 2771 //============================================================================= 2772 //----------------------------------SCMemProjNode------------------------------ 2773 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2774 { 2775 return bottom_type(); 2776 } 2777 2778 //============================================================================= 2779 //----------------------------------LoadStoreNode------------------------------ 2780 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2781 : Node(required), 2782 _type(rt), 2783 _adr_type(at) 2784 { 2785 init_req(MemNode::Control, c ); 2786 init_req(MemNode::Memory , mem); 2787 init_req(MemNode::Address, adr); 2788 init_req(MemNode::ValueIn, val); 2789 init_class_id(Class_LoadStore); 2790 } 2791 2792 uint LoadStoreNode::ideal_reg() const { 2793 return _type->ideal_reg(); 2794 } 2795 2796 bool LoadStoreNode::result_not_used() const { 2797 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2798 Node *x = fast_out(i); 2799 if (x->Opcode() == Op_SCMemProj) continue; 2800 return false; 2801 } 2802 return true; 2803 } 2804 2805 MemBarNode* LoadStoreNode::trailing_membar() const { 2806 MemBarNode* trailing = NULL; 2807 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2808 Node* u = fast_out(i); 2809 if (u->is_MemBar()) { 2810 if (u->as_MemBar()->trailing_load_store()) { 2811 assert(u->Opcode() == Op_MemBarAcquire, ""); 2812 assert(trailing == NULL, "only one"); 2813 trailing = u->as_MemBar(); 2814 #ifdef ASSERT 2815 Node* leading = trailing->leading_membar(); 2816 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2817 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 2818 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 2819 #endif 2820 } else { 2821 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 2822 } 2823 } 2824 } 2825 2826 return trailing; 2827 } 2828 2829 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2830 2831 //============================================================================= 2832 //----------------------------------LoadStoreConditionalNode-------------------- 2833 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2834 init_req(ExpectedIn, ex ); 2835 } 2836 2837 //============================================================================= 2838 //-------------------------------adr_type-------------------------------------- 2839 const TypePtr* ClearArrayNode::adr_type() const { 2840 Node *adr = in(3); 2841 if (adr == NULL) return NULL; // node is dead 2842 return MemNode::calculate_adr_type(adr->bottom_type()); 2843 } 2844 2845 //------------------------------match_edge------------------------------------- 2846 // Do we Match on this edge index or not? Do not match memory 2847 uint ClearArrayNode::match_edge(uint idx) const { 2848 return idx > 1; 2849 } 2850 2851 //------------------------------Identity--------------------------------------- 2852 // Clearing a zero length array does nothing 2853 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2854 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2855 } 2856 2857 //------------------------------Idealize--------------------------------------- 2858 // Clearing a short array is faster with stores 2859 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2860 // Already know this is a large node, do not try to ideal it 2861 if (!IdealizeClearArrayNode || _is_large) return NULL; 2862 2863 const int unit = BytesPerLong; 2864 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2865 if (!t) return NULL; 2866 if (!t->is_con()) return NULL; 2867 intptr_t raw_count = t->get_con(); 2868 intptr_t size = raw_count; 2869 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2870 // Clearing nothing uses the Identity call. 2871 // Negative clears are possible on dead ClearArrays 2872 // (see jck test stmt114.stmt11402.val). 2873 if (size <= 0 || size % unit != 0) return NULL; 2874 intptr_t count = size / unit; 2875 // Length too long; communicate this to matchers and assemblers. 2876 // Assemblers are responsible to produce fast hardware clears for it. 2877 if (size > InitArrayShortSize) { 2878 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 2879 } 2880 Node *mem = in(1); 2881 if( phase->type(mem)==Type::TOP ) return NULL; 2882 Node *adr = in(3); 2883 const Type* at = phase->type(adr); 2884 if( at==Type::TOP ) return NULL; 2885 const TypePtr* atp = at->isa_ptr(); 2886 // adjust atp to be the correct array element address type 2887 if (atp == NULL) atp = TypePtr::BOTTOM; 2888 else atp = atp->add_offset(Type::OffsetBot); 2889 // Get base for derived pointer purposes 2890 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2891 Node *base = adr->in(1); 2892 2893 Node *zero = phase->makecon(TypeLong::ZERO); 2894 Node *off = phase->MakeConX(BytesPerLong); 2895 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2896 count--; 2897 while( count-- ) { 2898 mem = phase->transform(mem); 2899 adr = phase->transform(new AddPNode(base,adr,off)); 2900 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2901 } 2902 return mem; 2903 } 2904 2905 //----------------------------step_through---------------------------------- 2906 // Return allocation input memory edge if it is different instance 2907 // or itself if it is the one we are looking for. 2908 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2909 Node* n = *np; 2910 assert(n->is_ClearArray(), "sanity"); 2911 intptr_t offset; 2912 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2913 // This method is called only before Allocate nodes are expanded 2914 // during macro nodes expansion. Before that ClearArray nodes are 2915 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2916 // Allocate nodes are expanded) which follows allocations. 2917 assert(alloc != NULL, "should have allocation"); 2918 if (alloc->_idx == instance_id) { 2919 // Can not bypass initialization of the instance we are looking for. 2920 return false; 2921 } 2922 // Otherwise skip it. 2923 InitializeNode* init = alloc->initialization(); 2924 if (init != NULL) 2925 *np = init->in(TypeFunc::Memory); 2926 else 2927 *np = alloc->in(TypeFunc::Memory); 2928 return true; 2929 } 2930 2931 //----------------------------clear_memory------------------------------------- 2932 // Generate code to initialize object storage to zero. 2933 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2934 intptr_t start_offset, 2935 Node* end_offset, 2936 PhaseGVN* phase) { 2937 intptr_t offset = start_offset; 2938 2939 int unit = BytesPerLong; 2940 if ((offset % unit) != 0) { 2941 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2942 adr = phase->transform(adr); 2943 const TypePtr* atp = TypeRawPtr::BOTTOM; 2944 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2945 mem = phase->transform(mem); 2946 offset += BytesPerInt; 2947 } 2948 assert((offset % unit) == 0, ""); 2949 2950 // Initialize the remaining stuff, if any, with a ClearArray. 2951 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2952 } 2953 2954 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2955 Node* start_offset, 2956 Node* end_offset, 2957 PhaseGVN* phase) { 2958 if (start_offset == end_offset) { 2959 // nothing to do 2960 return mem; 2961 } 2962 2963 int unit = BytesPerLong; 2964 Node* zbase = start_offset; 2965 Node* zend = end_offset; 2966 2967 // Scale to the unit required by the CPU: 2968 if (!Matcher::init_array_count_is_in_bytes) { 2969 Node* shift = phase->intcon(exact_log2(unit)); 2970 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2971 zend = phase->transform(new URShiftXNode(zend, shift) ); 2972 } 2973 2974 // Bulk clear double-words 2975 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2976 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2977 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 2978 return phase->transform(mem); 2979 } 2980 2981 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2982 intptr_t start_offset, 2983 intptr_t end_offset, 2984 PhaseGVN* phase) { 2985 if (start_offset == end_offset) { 2986 // nothing to do 2987 return mem; 2988 } 2989 2990 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2991 intptr_t done_offset = end_offset; 2992 if ((done_offset % BytesPerLong) != 0) { 2993 done_offset -= BytesPerInt; 2994 } 2995 if (done_offset > start_offset) { 2996 mem = clear_memory(ctl, mem, dest, 2997 start_offset, phase->MakeConX(done_offset), phase); 2998 } 2999 if (done_offset < end_offset) { // emit the final 32-bit store 3000 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3001 adr = phase->transform(adr); 3002 const TypePtr* atp = TypeRawPtr::BOTTOM; 3003 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3004 mem = phase->transform(mem); 3005 done_offset += BytesPerInt; 3006 } 3007 assert(done_offset == end_offset, ""); 3008 return mem; 3009 } 3010 3011 //============================================================================= 3012 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3013 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 3014 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3015 #ifdef ASSERT 3016 , _pair_idx(0) 3017 #endif 3018 { 3019 init_class_id(Class_MemBar); 3020 Node* top = C->top(); 3021 init_req(TypeFunc::I_O,top); 3022 init_req(TypeFunc::FramePtr,top); 3023 init_req(TypeFunc::ReturnAdr,top); 3024 if (precedent != NULL) 3025 init_req(TypeFunc::Parms, precedent); 3026 } 3027 3028 //------------------------------cmp-------------------------------------------- 3029 uint MemBarNode::hash() const { return NO_HASH; } 3030 uint MemBarNode::cmp( const Node &n ) const { 3031 return (&n == this); // Always fail except on self 3032 } 3033 3034 //------------------------------make------------------------------------------- 3035 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3036 switch (opcode) { 3037 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3038 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3039 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3040 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3041 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3042 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3043 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3044 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3045 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3046 case Op_Initialize: return new InitializeNode(C, atp, pn); 3047 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3048 default: ShouldNotReachHere(); return NULL; 3049 } 3050 } 3051 3052 void MemBarNode::remove(PhaseIterGVN *igvn) { 3053 if (outcnt() != 2) { 3054 return; 3055 } 3056 if (trailing_store() || trailing_load_store()) { 3057 MemBarNode* leading = leading_membar(); 3058 if (leading != NULL) { 3059 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3060 leading->remove(igvn); 3061 } 3062 } 3063 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3064 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3065 } 3066 3067 //------------------------------Ideal------------------------------------------ 3068 // Return a node which is more "ideal" than the current node. Strip out 3069 // control copies 3070 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3071 if (remove_dead_region(phase, can_reshape)) return this; 3072 // Don't bother trying to transform a dead node 3073 if (in(0) && in(0)->is_top()) { 3074 return NULL; 3075 } 3076 3077 #if INCLUDE_ZGC 3078 if (UseZGC) { 3079 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) { 3080 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop); 3081 set_req(MemBarNode::Precedent, load_node); 3082 return this; 3083 } 3084 } 3085 #endif 3086 3087 bool progress = false; 3088 // Eliminate volatile MemBars for scalar replaced objects. 3089 if (can_reshape && req() == (Precedent+1)) { 3090 bool eliminate = false; 3091 int opc = Opcode(); 3092 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3093 // Volatile field loads and stores. 3094 Node* my_mem = in(MemBarNode::Precedent); 3095 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3096 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3097 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3098 // replace this Precedent (decodeN) with the Load instead. 3099 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3100 Node* load_node = my_mem->in(1); 3101 set_req(MemBarNode::Precedent, load_node); 3102 phase->is_IterGVN()->_worklist.push(my_mem); 3103 my_mem = load_node; 3104 } else { 3105 assert(my_mem->unique_out() == this, "sanity"); 3106 del_req(Precedent); 3107 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3108 my_mem = NULL; 3109 } 3110 progress = true; 3111 } 3112 if (my_mem != NULL && my_mem->is_Mem()) { 3113 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3114 // Check for scalar replaced object reference. 3115 if( t_oop != NULL && t_oop->is_known_instance_field() && 3116 t_oop->offset() != Type::OffsetBot && 3117 t_oop->offset() != Type::OffsetTop) { 3118 eliminate = true; 3119 } 3120 } 3121 } else if (opc == Op_MemBarRelease) { 3122 // Final field stores. 3123 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3124 if ((alloc != NULL) && alloc->is_Allocate() && 3125 alloc->as_Allocate()->does_not_escape_thread()) { 3126 // The allocated object does not escape. 3127 eliminate = true; 3128 } 3129 } 3130 if (eliminate) { 3131 // Replace MemBar projections by its inputs. 3132 PhaseIterGVN* igvn = phase->is_IterGVN(); 3133 remove(igvn); 3134 // Must return either the original node (now dead) or a new node 3135 // (Do not return a top here, since that would break the uniqueness of top.) 3136 return new ConINode(TypeInt::ZERO); 3137 } 3138 } 3139 return progress ? this : NULL; 3140 } 3141 3142 //------------------------------Value------------------------------------------ 3143 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3144 if( !in(0) ) return Type::TOP; 3145 if( phase->type(in(0)) == Type::TOP ) 3146 return Type::TOP; 3147 return TypeTuple::MEMBAR; 3148 } 3149 3150 //------------------------------match------------------------------------------ 3151 // Construct projections for memory. 3152 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3153 switch (proj->_con) { 3154 case TypeFunc::Control: 3155 case TypeFunc::Memory: 3156 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3157 } 3158 ShouldNotReachHere(); 3159 return NULL; 3160 } 3161 3162 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3163 trailing->_kind = TrailingStore; 3164 leading->_kind = LeadingStore; 3165 #ifdef ASSERT 3166 trailing->_pair_idx = leading->_idx; 3167 leading->_pair_idx = leading->_idx; 3168 #endif 3169 } 3170 3171 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3172 trailing->_kind = TrailingLoadStore; 3173 leading->_kind = LeadingLoadStore; 3174 #ifdef ASSERT 3175 trailing->_pair_idx = leading->_idx; 3176 leading->_pair_idx = leading->_idx; 3177 #endif 3178 } 3179 3180 MemBarNode* MemBarNode::trailing_membar() const { 3181 ResourceMark rm; 3182 Node* trailing = (Node*)this; 3183 VectorSet seen(Thread::current()->resource_area()); 3184 Node_Stack multis(0); 3185 do { 3186 Node* c = trailing; 3187 uint i = 0; 3188 do { 3189 trailing = NULL; 3190 for (; i < c->outcnt(); i++) { 3191 Node* next = c->raw_out(i); 3192 if (next != c && next->is_CFG()) { 3193 if (c->is_MultiBranch()) { 3194 if (multis.node() == c) { 3195 multis.set_index(i+1); 3196 } else { 3197 multis.push(c, i+1); 3198 } 3199 } 3200 trailing = next; 3201 break; 3202 } 3203 } 3204 if (trailing != NULL && !seen.test_set(trailing->_idx)) { 3205 break; 3206 } 3207 while (multis.size() > 0) { 3208 c = multis.node(); 3209 i = multis.index(); 3210 if (i < c->req()) { 3211 break; 3212 } 3213 multis.pop(); 3214 } 3215 } while (multis.size() > 0); 3216 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3217 3218 MemBarNode* mb = trailing->as_MemBar(); 3219 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3220 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3221 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3222 return mb; 3223 } 3224 3225 MemBarNode* MemBarNode::leading_membar() const { 3226 ResourceMark rm; 3227 VectorSet seen(Thread::current()->resource_area()); 3228 Node_Stack regions(0); 3229 Node* leading = in(0); 3230 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3231 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) { 3232 leading = NULL; 3233 while (regions.size() > 0) { 3234 Node* r = regions.node(); 3235 uint i = regions.index(); 3236 if (i < r->req()) { 3237 leading = r->in(i); 3238 regions.set_index(i+1); 3239 } else { 3240 regions.pop(); 3241 } 3242 } 3243 if (leading == NULL) { 3244 assert(regions.size() == 0, "all paths should have been tried"); 3245 return NULL; 3246 } 3247 } 3248 if (leading->is_Region()) { 3249 regions.push(leading, 2); 3250 leading = leading->in(1); 3251 } else { 3252 leading = leading->in(0); 3253 } 3254 } 3255 #ifdef ASSERT 3256 Unique_Node_List wq; 3257 wq.push((Node*)this); 3258 uint found = 0; 3259 for (uint i = 0; i < wq.size(); i++) { 3260 Node* n = wq.at(i); 3261 if (n->is_Region()) { 3262 for (uint j = 1; j < n->req(); j++) { 3263 Node* in = n->in(j); 3264 if (in != NULL && !in->is_top()) { 3265 wq.push(in); 3266 } 3267 } 3268 } else { 3269 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3270 assert(n == leading, "consistency check failed"); 3271 found++; 3272 } else { 3273 Node* in = n->in(0); 3274 if (in != NULL && !in->is_top()) { 3275 wq.push(in); 3276 } 3277 } 3278 } 3279 } 3280 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed"); 3281 #endif 3282 if (leading == NULL) { 3283 return NULL; 3284 } 3285 MemBarNode* mb = leading->as_MemBar(); 3286 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3287 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3288 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3289 return mb; 3290 } 3291 3292 //===========================InitializeNode==================================== 3293 // SUMMARY: 3294 // This node acts as a memory barrier on raw memory, after some raw stores. 3295 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3296 // The Initialize can 'capture' suitably constrained stores as raw inits. 3297 // It can coalesce related raw stores into larger units (called 'tiles'). 3298 // It can avoid zeroing new storage for memory units which have raw inits. 3299 // At macro-expansion, it is marked 'complete', and does not optimize further. 3300 // 3301 // EXAMPLE: 3302 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3303 // ctl = incoming control; mem* = incoming memory 3304 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3305 // First allocate uninitialized memory and fill in the header: 3306 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3307 // ctl := alloc.Control; mem* := alloc.Memory* 3308 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3309 // Then initialize to zero the non-header parts of the raw memory block: 3310 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3311 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3312 // After the initialize node executes, the object is ready for service: 3313 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3314 // Suppose its body is immediately initialized as {1,2}: 3315 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3316 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3317 // mem.SLICE(#short[*]) := store2 3318 // 3319 // DETAILS: 3320 // An InitializeNode collects and isolates object initialization after 3321 // an AllocateNode and before the next possible safepoint. As a 3322 // memory barrier (MemBarNode), it keeps critical stores from drifting 3323 // down past any safepoint or any publication of the allocation. 3324 // Before this barrier, a newly-allocated object may have uninitialized bits. 3325 // After this barrier, it may be treated as a real oop, and GC is allowed. 3326 // 3327 // The semantics of the InitializeNode include an implicit zeroing of 3328 // the new object from object header to the end of the object. 3329 // (The object header and end are determined by the AllocateNode.) 3330 // 3331 // Certain stores may be added as direct inputs to the InitializeNode. 3332 // These stores must update raw memory, and they must be to addresses 3333 // derived from the raw address produced by AllocateNode, and with 3334 // a constant offset. They must be ordered by increasing offset. 3335 // The first one is at in(RawStores), the last at in(req()-1). 3336 // Unlike most memory operations, they are not linked in a chain, 3337 // but are displayed in parallel as users of the rawmem output of 3338 // the allocation. 3339 // 3340 // (See comments in InitializeNode::capture_store, which continue 3341 // the example given above.) 3342 // 3343 // When the associated Allocate is macro-expanded, the InitializeNode 3344 // may be rewritten to optimize collected stores. A ClearArrayNode 3345 // may also be created at that point to represent any required zeroing. 3346 // The InitializeNode is then marked 'complete', prohibiting further 3347 // capturing of nearby memory operations. 3348 // 3349 // During macro-expansion, all captured initializations which store 3350 // constant values of 32 bits or smaller are coalesced (if advantageous) 3351 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3352 // initialized in fewer memory operations. Memory words which are 3353 // covered by neither tiles nor non-constant stores are pre-zeroed 3354 // by explicit stores of zero. (The code shape happens to do all 3355 // zeroing first, then all other stores, with both sequences occurring 3356 // in order of ascending offsets.) 3357 // 3358 // Alternatively, code may be inserted between an AllocateNode and its 3359 // InitializeNode, to perform arbitrary initialization of the new object. 3360 // E.g., the object copying intrinsics insert complex data transfers here. 3361 // The initialization must then be marked as 'complete' disable the 3362 // built-in zeroing semantics and the collection of initializing stores. 3363 // 3364 // While an InitializeNode is incomplete, reads from the memory state 3365 // produced by it are optimizable if they match the control edge and 3366 // new oop address associated with the allocation/initialization. 3367 // They return a stored value (if the offset matches) or else zero. 3368 // A write to the memory state, if it matches control and address, 3369 // and if it is to a constant offset, may be 'captured' by the 3370 // InitializeNode. It is cloned as a raw memory operation and rewired 3371 // inside the initialization, to the raw oop produced by the allocation. 3372 // Operations on addresses which are provably distinct (e.g., to 3373 // other AllocateNodes) are allowed to bypass the initialization. 3374 // 3375 // The effect of all this is to consolidate object initialization 3376 // (both arrays and non-arrays, both piecewise and bulk) into a 3377 // single location, where it can be optimized as a unit. 3378 // 3379 // Only stores with an offset less than TrackedInitializationLimit words 3380 // will be considered for capture by an InitializeNode. This puts a 3381 // reasonable limit on the complexity of optimized initializations. 3382 3383 //---------------------------InitializeNode------------------------------------ 3384 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3385 : MemBarNode(C, adr_type, rawoop), 3386 _is_complete(Incomplete), _does_not_escape(false) 3387 { 3388 init_class_id(Class_Initialize); 3389 3390 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3391 assert(in(RawAddress) == rawoop, "proper init"); 3392 // Note: allocation() can be NULL, for secondary initialization barriers 3393 } 3394 3395 // Since this node is not matched, it will be processed by the 3396 // register allocator. Declare that there are no constraints 3397 // on the allocation of the RawAddress edge. 3398 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3399 // This edge should be set to top, by the set_complete. But be conservative. 3400 if (idx == InitializeNode::RawAddress) 3401 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3402 return RegMask::Empty; 3403 } 3404 3405 Node* InitializeNode::memory(uint alias_idx) { 3406 Node* mem = in(Memory); 3407 if (mem->is_MergeMem()) { 3408 return mem->as_MergeMem()->memory_at(alias_idx); 3409 } else { 3410 // incoming raw memory is not split 3411 return mem; 3412 } 3413 } 3414 3415 bool InitializeNode::is_non_zero() { 3416 if (is_complete()) return false; 3417 remove_extra_zeroes(); 3418 return (req() > RawStores); 3419 } 3420 3421 void InitializeNode::set_complete(PhaseGVN* phase) { 3422 assert(!is_complete(), "caller responsibility"); 3423 _is_complete = Complete; 3424 3425 // After this node is complete, it contains a bunch of 3426 // raw-memory initializations. There is no need for 3427 // it to have anything to do with non-raw memory effects. 3428 // Therefore, tell all non-raw users to re-optimize themselves, 3429 // after skipping the memory effects of this initialization. 3430 PhaseIterGVN* igvn = phase->is_IterGVN(); 3431 if (igvn) igvn->add_users_to_worklist(this); 3432 } 3433 3434 // convenience function 3435 // return false if the init contains any stores already 3436 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3437 InitializeNode* init = initialization(); 3438 if (init == NULL || init->is_complete()) return false; 3439 init->remove_extra_zeroes(); 3440 // for now, if this allocation has already collected any inits, bail: 3441 if (init->is_non_zero()) return false; 3442 init->set_complete(phase); 3443 return true; 3444 } 3445 3446 void InitializeNode::remove_extra_zeroes() { 3447 if (req() == RawStores) return; 3448 Node* zmem = zero_memory(); 3449 uint fill = RawStores; 3450 for (uint i = fill; i < req(); i++) { 3451 Node* n = in(i); 3452 if (n->is_top() || n == zmem) continue; // skip 3453 if (fill < i) set_req(fill, n); // compact 3454 ++fill; 3455 } 3456 // delete any empty spaces created: 3457 while (fill < req()) { 3458 del_req(fill); 3459 } 3460 } 3461 3462 // Helper for remembering which stores go with which offsets. 3463 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3464 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3465 intptr_t offset = -1; 3466 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3467 phase, offset); 3468 if (base == NULL) return -1; // something is dead, 3469 if (offset < 0) return -1; // dead, dead 3470 return offset; 3471 } 3472 3473 // Helper for proving that an initialization expression is 3474 // "simple enough" to be folded into an object initialization. 3475 // Attempts to prove that a store's initial value 'n' can be captured 3476 // within the initialization without creating a vicious cycle, such as: 3477 // { Foo p = new Foo(); p.next = p; } 3478 // True for constants and parameters and small combinations thereof. 3479 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3480 if (n == NULL) return true; // (can this really happen?) 3481 if (n->is_Proj()) n = n->in(0); 3482 if (n == this) return false; // found a cycle 3483 if (n->is_Con()) return true; 3484 if (n->is_Start()) return true; // params, etc., are OK 3485 if (n->is_Root()) return true; // even better 3486 3487 Node* ctl = n->in(0); 3488 if (ctl != NULL && !ctl->is_top()) { 3489 if (ctl->is_Proj()) ctl = ctl->in(0); 3490 if (ctl == this) return false; 3491 3492 // If we already know that the enclosing memory op is pinned right after 3493 // the init, then any control flow that the store has picked up 3494 // must have preceded the init, or else be equal to the init. 3495 // Even after loop optimizations (which might change control edges) 3496 // a store is never pinned *before* the availability of its inputs. 3497 if (!MemNode::all_controls_dominate(n, this)) 3498 return false; // failed to prove a good control 3499 } 3500 3501 // Check data edges for possible dependencies on 'this'. 3502 if ((count += 1) > 20) return false; // complexity limit 3503 for (uint i = 1; i < n->req(); i++) { 3504 Node* m = n->in(i); 3505 if (m == NULL || m == n || m->is_top()) continue; 3506 uint first_i = n->find_edge(m); 3507 if (i != first_i) continue; // process duplicate edge just once 3508 if (!detect_init_independence(m, count)) { 3509 return false; 3510 } 3511 } 3512 3513 return true; 3514 } 3515 3516 // Here are all the checks a Store must pass before it can be moved into 3517 // an initialization. Returns zero if a check fails. 3518 // On success, returns the (constant) offset to which the store applies, 3519 // within the initialized memory. 3520 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3521 const int FAIL = 0; 3522 if (st->is_unaligned_access()) { 3523 return FAIL; 3524 } 3525 if (st->req() != MemNode::ValueIn + 1) 3526 return FAIL; // an inscrutable StoreNode (card mark?) 3527 Node* ctl = st->in(MemNode::Control); 3528 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3529 return FAIL; // must be unconditional after the initialization 3530 Node* mem = st->in(MemNode::Memory); 3531 if (!(mem->is_Proj() && mem->in(0) == this)) 3532 return FAIL; // must not be preceded by other stores 3533 Node* adr = st->in(MemNode::Address); 3534 intptr_t offset; 3535 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3536 if (alloc == NULL) 3537 return FAIL; // inscrutable address 3538 if (alloc != allocation()) 3539 return FAIL; // wrong allocation! (store needs to float up) 3540 Node* val = st->in(MemNode::ValueIn); 3541 int complexity_count = 0; 3542 if (!detect_init_independence(val, complexity_count)) 3543 return FAIL; // stored value must be 'simple enough' 3544 3545 // The Store can be captured only if nothing after the allocation 3546 // and before the Store is using the memory location that the store 3547 // overwrites. 3548 bool failed = false; 3549 // If is_complete_with_arraycopy() is true the shape of the graph is 3550 // well defined and is safe so no need for extra checks. 3551 if (!is_complete_with_arraycopy()) { 3552 // We are going to look at each use of the memory state following 3553 // the allocation to make sure nothing reads the memory that the 3554 // Store writes. 3555 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3556 int alias_idx = phase->C->get_alias_index(t_adr); 3557 ResourceMark rm; 3558 Unique_Node_List mems; 3559 mems.push(mem); 3560 Node* unique_merge = NULL; 3561 for (uint next = 0; next < mems.size(); ++next) { 3562 Node *m = mems.at(next); 3563 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3564 Node *n = m->fast_out(j); 3565 if (n->outcnt() == 0) { 3566 continue; 3567 } 3568 if (n == st) { 3569 continue; 3570 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3571 // If the control of this use is different from the control 3572 // of the Store which is right after the InitializeNode then 3573 // this node cannot be between the InitializeNode and the 3574 // Store. 3575 continue; 3576 } else if (n->is_MergeMem()) { 3577 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3578 // We can hit a MergeMemNode (that will likely go away 3579 // later) that is a direct use of the memory state 3580 // following the InitializeNode on the same slice as the 3581 // store node that we'd like to capture. We need to check 3582 // the uses of the MergeMemNode. 3583 mems.push(n); 3584 } 3585 } else if (n->is_Mem()) { 3586 Node* other_adr = n->in(MemNode::Address); 3587 if (other_adr == adr) { 3588 failed = true; 3589 break; 3590 } else { 3591 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3592 if (other_t_adr != NULL) { 3593 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3594 if (other_alias_idx == alias_idx) { 3595 // A load from the same memory slice as the store right 3596 // after the InitializeNode. We check the control of the 3597 // object/array that is loaded from. If it's the same as 3598 // the store control then we cannot capture the store. 3599 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3600 Node* base = other_adr; 3601 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3602 base = base->in(AddPNode::Base); 3603 if (base != NULL) { 3604 base = base->uncast(); 3605 if (base->is_Proj() && base->in(0) == alloc) { 3606 failed = true; 3607 break; 3608 } 3609 } 3610 } 3611 } 3612 } 3613 } else { 3614 failed = true; 3615 break; 3616 } 3617 } 3618 } 3619 } 3620 if (failed) { 3621 if (!can_reshape) { 3622 // We decided we couldn't capture the store during parsing. We 3623 // should try again during the next IGVN once the graph is 3624 // cleaner. 3625 phase->C->record_for_igvn(st); 3626 } 3627 return FAIL; 3628 } 3629 3630 return offset; // success 3631 } 3632 3633 // Find the captured store in(i) which corresponds to the range 3634 // [start..start+size) in the initialized object. 3635 // If there is one, return its index i. If there isn't, return the 3636 // negative of the index where it should be inserted. 3637 // Return 0 if the queried range overlaps an initialization boundary 3638 // or if dead code is encountered. 3639 // If size_in_bytes is zero, do not bother with overlap checks. 3640 int InitializeNode::captured_store_insertion_point(intptr_t start, 3641 int size_in_bytes, 3642 PhaseTransform* phase) { 3643 const int FAIL = 0, MAX_STORE = BytesPerLong; 3644 3645 if (is_complete()) 3646 return FAIL; // arraycopy got here first; punt 3647 3648 assert(allocation() != NULL, "must be present"); 3649 3650 // no negatives, no header fields: 3651 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3652 3653 // after a certain size, we bail out on tracking all the stores: 3654 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3655 if (start >= ti_limit) return FAIL; 3656 3657 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3658 if (i >= limit) return -(int)i; // not found; here is where to put it 3659 3660 Node* st = in(i); 3661 intptr_t st_off = get_store_offset(st, phase); 3662 if (st_off < 0) { 3663 if (st != zero_memory()) { 3664 return FAIL; // bail out if there is dead garbage 3665 } 3666 } else if (st_off > start) { 3667 // ...we are done, since stores are ordered 3668 if (st_off < start + size_in_bytes) { 3669 return FAIL; // the next store overlaps 3670 } 3671 return -(int)i; // not found; here is where to put it 3672 } else if (st_off < start) { 3673 if (size_in_bytes != 0 && 3674 start < st_off + MAX_STORE && 3675 start < st_off + st->as_Store()->memory_size()) { 3676 return FAIL; // the previous store overlaps 3677 } 3678 } else { 3679 if (size_in_bytes != 0 && 3680 st->as_Store()->memory_size() != size_in_bytes) { 3681 return FAIL; // mismatched store size 3682 } 3683 return i; 3684 } 3685 3686 ++i; 3687 } 3688 } 3689 3690 // Look for a captured store which initializes at the offset 'start' 3691 // with the given size. If there is no such store, and no other 3692 // initialization interferes, then return zero_memory (the memory 3693 // projection of the AllocateNode). 3694 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3695 PhaseTransform* phase) { 3696 assert(stores_are_sane(phase), ""); 3697 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3698 if (i == 0) { 3699 return NULL; // something is dead 3700 } else if (i < 0) { 3701 return zero_memory(); // just primordial zero bits here 3702 } else { 3703 Node* st = in(i); // here is the store at this position 3704 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3705 return st; 3706 } 3707 } 3708 3709 // Create, as a raw pointer, an address within my new object at 'offset'. 3710 Node* InitializeNode::make_raw_address(intptr_t offset, 3711 PhaseTransform* phase) { 3712 Node* addr = in(RawAddress); 3713 if (offset != 0) { 3714 Compile* C = phase->C; 3715 addr = phase->transform( new AddPNode(C->top(), addr, 3716 phase->MakeConX(offset)) ); 3717 } 3718 return addr; 3719 } 3720 3721 // Clone the given store, converting it into a raw store 3722 // initializing a field or element of my new object. 3723 // Caller is responsible for retiring the original store, 3724 // with subsume_node or the like. 3725 // 3726 // From the example above InitializeNode::InitializeNode, 3727 // here are the old stores to be captured: 3728 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3729 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3730 // 3731 // Here is the changed code; note the extra edges on init: 3732 // alloc = (Allocate ...) 3733 // rawoop = alloc.RawAddress 3734 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3735 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3736 // init = (Initialize alloc.Control alloc.Memory rawoop 3737 // rawstore1 rawstore2) 3738 // 3739 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3740 PhaseTransform* phase, bool can_reshape) { 3741 assert(stores_are_sane(phase), ""); 3742 3743 if (start < 0) return NULL; 3744 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3745 3746 Compile* C = phase->C; 3747 int size_in_bytes = st->memory_size(); 3748 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3749 if (i == 0) return NULL; // bail out 3750 Node* prev_mem = NULL; // raw memory for the captured store 3751 if (i > 0) { 3752 prev_mem = in(i); // there is a pre-existing store under this one 3753 set_req(i, C->top()); // temporarily disconnect it 3754 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3755 } else { 3756 i = -i; // no pre-existing store 3757 prev_mem = zero_memory(); // a slice of the newly allocated object 3758 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3759 set_req(--i, C->top()); // reuse this edge; it has been folded away 3760 else 3761 ins_req(i, C->top()); // build a new edge 3762 } 3763 Node* new_st = st->clone(); 3764 new_st->set_req(MemNode::Control, in(Control)); 3765 new_st->set_req(MemNode::Memory, prev_mem); 3766 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3767 new_st = phase->transform(new_st); 3768 3769 // At this point, new_st might have swallowed a pre-existing store 3770 // at the same offset, or perhaps new_st might have disappeared, 3771 // if it redundantly stored the same value (or zero to fresh memory). 3772 3773 // In any case, wire it in: 3774 phase->igvn_rehash_node_delayed(this); 3775 set_req(i, new_st); 3776 3777 // The caller may now kill the old guy. 3778 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3779 assert(check_st == new_st || check_st == NULL, "must be findable"); 3780 assert(!is_complete(), ""); 3781 return new_st; 3782 } 3783 3784 static bool store_constant(jlong* tiles, int num_tiles, 3785 intptr_t st_off, int st_size, 3786 jlong con) { 3787 if ((st_off & (st_size-1)) != 0) 3788 return false; // strange store offset (assume size==2**N) 3789 address addr = (address)tiles + st_off; 3790 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3791 switch (st_size) { 3792 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3793 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3794 case sizeof(jint): *(jint*) addr = (jint) con; break; 3795 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3796 default: return false; // strange store size (detect size!=2**N here) 3797 } 3798 return true; // return success to caller 3799 } 3800 3801 // Coalesce subword constants into int constants and possibly 3802 // into long constants. The goal, if the CPU permits, 3803 // is to initialize the object with a small number of 64-bit tiles. 3804 // Also, convert floating-point constants to bit patterns. 3805 // Non-constants are not relevant to this pass. 3806 // 3807 // In terms of the running example on InitializeNode::InitializeNode 3808 // and InitializeNode::capture_store, here is the transformation 3809 // of rawstore1 and rawstore2 into rawstore12: 3810 // alloc = (Allocate ...) 3811 // rawoop = alloc.RawAddress 3812 // tile12 = 0x00010002 3813 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3814 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3815 // 3816 void 3817 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3818 Node* size_in_bytes, 3819 PhaseGVN* phase) { 3820 Compile* C = phase->C; 3821 3822 assert(stores_are_sane(phase), ""); 3823 // Note: After this pass, they are not completely sane, 3824 // since there may be some overlaps. 3825 3826 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3827 3828 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3829 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3830 size_limit = MIN2(size_limit, ti_limit); 3831 size_limit = align_up(size_limit, BytesPerLong); 3832 int num_tiles = size_limit / BytesPerLong; 3833 3834 // allocate space for the tile map: 3835 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3836 jlong tiles_buf[small_len]; 3837 Node* nodes_buf[small_len]; 3838 jlong inits_buf[small_len]; 3839 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3840 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3841 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3842 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3843 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3844 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3845 // tiles: exact bitwise model of all primitive constants 3846 // nodes: last constant-storing node subsumed into the tiles model 3847 // inits: which bytes (in each tile) are touched by any initializations 3848 3849 //// Pass A: Fill in the tile model with any relevant stores. 3850 3851 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3852 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3853 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3854 Node* zmem = zero_memory(); // initially zero memory state 3855 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3856 Node* st = in(i); 3857 intptr_t st_off = get_store_offset(st, phase); 3858 3859 // Figure out the store's offset and constant value: 3860 if (st_off < header_size) continue; //skip (ignore header) 3861 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3862 int st_size = st->as_Store()->memory_size(); 3863 if (st_off + st_size > size_limit) break; 3864 3865 // Record which bytes are touched, whether by constant or not. 3866 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3867 continue; // skip (strange store size) 3868 3869 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3870 if (!val->singleton()) continue; //skip (non-con store) 3871 BasicType type = val->basic_type(); 3872 3873 jlong con = 0; 3874 switch (type) { 3875 case T_INT: con = val->is_int()->get_con(); break; 3876 case T_LONG: con = val->is_long()->get_con(); break; 3877 case T_FLOAT: con = jint_cast(val->getf()); break; 3878 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3879 default: continue; //skip (odd store type) 3880 } 3881 3882 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3883 st->Opcode() == Op_StoreL) { 3884 continue; // This StoreL is already optimal. 3885 } 3886 3887 // Store down the constant. 3888 store_constant(tiles, num_tiles, st_off, st_size, con); 3889 3890 intptr_t j = st_off >> LogBytesPerLong; 3891 3892 if (type == T_INT && st_size == BytesPerInt 3893 && (st_off & BytesPerInt) == BytesPerInt) { 3894 jlong lcon = tiles[j]; 3895 if (!Matcher::isSimpleConstant64(lcon) && 3896 st->Opcode() == Op_StoreI) { 3897 // This StoreI is already optimal by itself. 3898 jint* intcon = (jint*) &tiles[j]; 3899 intcon[1] = 0; // undo the store_constant() 3900 3901 // If the previous store is also optimal by itself, back up and 3902 // undo the action of the previous loop iteration... if we can. 3903 // But if we can't, just let the previous half take care of itself. 3904 st = nodes[j]; 3905 st_off -= BytesPerInt; 3906 con = intcon[0]; 3907 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3908 assert(st_off >= header_size, "still ignoring header"); 3909 assert(get_store_offset(st, phase) == st_off, "must be"); 3910 assert(in(i-1) == zmem, "must be"); 3911 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3912 assert(con == tcon->is_int()->get_con(), "must be"); 3913 // Undo the effects of the previous loop trip, which swallowed st: 3914 intcon[0] = 0; // undo store_constant() 3915 set_req(i-1, st); // undo set_req(i, zmem) 3916 nodes[j] = NULL; // undo nodes[j] = st 3917 --old_subword; // undo ++old_subword 3918 } 3919 continue; // This StoreI is already optimal. 3920 } 3921 } 3922 3923 // This store is not needed. 3924 set_req(i, zmem); 3925 nodes[j] = st; // record for the moment 3926 if (st_size < BytesPerLong) // something has changed 3927 ++old_subword; // includes int/float, but who's counting... 3928 else ++old_long; 3929 } 3930 3931 if ((old_subword + old_long) == 0) 3932 return; // nothing more to do 3933 3934 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3935 // Be sure to insert them before overlapping non-constant stores. 3936 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3937 for (int j = 0; j < num_tiles; j++) { 3938 jlong con = tiles[j]; 3939 jlong init = inits[j]; 3940 if (con == 0) continue; 3941 jint con0, con1; // split the constant, address-wise 3942 jint init0, init1; // split the init map, address-wise 3943 { union { jlong con; jint intcon[2]; } u; 3944 u.con = con; 3945 con0 = u.intcon[0]; 3946 con1 = u.intcon[1]; 3947 u.con = init; 3948 init0 = u.intcon[0]; 3949 init1 = u.intcon[1]; 3950 } 3951 3952 Node* old = nodes[j]; 3953 assert(old != NULL, "need the prior store"); 3954 intptr_t offset = (j * BytesPerLong); 3955 3956 bool split = !Matcher::isSimpleConstant64(con); 3957 3958 if (offset < header_size) { 3959 assert(offset + BytesPerInt >= header_size, "second int counts"); 3960 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3961 split = true; // only the second word counts 3962 // Example: int a[] = { 42 ... } 3963 } else if (con0 == 0 && init0 == -1) { 3964 split = true; // first word is covered by full inits 3965 // Example: int a[] = { ... foo(), 42 ... } 3966 } else if (con1 == 0 && init1 == -1) { 3967 split = true; // second word is covered by full inits 3968 // Example: int a[] = { ... 42, foo() ... } 3969 } 3970 3971 // Here's a case where init0 is neither 0 nor -1: 3972 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3973 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3974 // In this case the tile is not split; it is (jlong)42. 3975 // The big tile is stored down, and then the foo() value is inserted. 3976 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3977 3978 Node* ctl = old->in(MemNode::Control); 3979 Node* adr = make_raw_address(offset, phase); 3980 const TypePtr* atp = TypeRawPtr::BOTTOM; 3981 3982 // One or two coalesced stores to plop down. 3983 Node* st[2]; 3984 intptr_t off[2]; 3985 int nst = 0; 3986 if (!split) { 3987 ++new_long; 3988 off[nst] = offset; 3989 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3990 phase->longcon(con), T_LONG, MemNode::unordered); 3991 } else { 3992 // Omit either if it is a zero. 3993 if (con0 != 0) { 3994 ++new_int; 3995 off[nst] = offset; 3996 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3997 phase->intcon(con0), T_INT, MemNode::unordered); 3998 } 3999 if (con1 != 0) { 4000 ++new_int; 4001 offset += BytesPerInt; 4002 adr = make_raw_address(offset, phase); 4003 off[nst] = offset; 4004 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4005 phase->intcon(con1), T_INT, MemNode::unordered); 4006 } 4007 } 4008 4009 // Insert second store first, then the first before the second. 4010 // Insert each one just before any overlapping non-constant stores. 4011 while (nst > 0) { 4012 Node* st1 = st[--nst]; 4013 C->copy_node_notes_to(st1, old); 4014 st1 = phase->transform(st1); 4015 offset = off[nst]; 4016 assert(offset >= header_size, "do not smash header"); 4017 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4018 guarantee(ins_idx != 0, "must re-insert constant store"); 4019 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4020 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4021 set_req(--ins_idx, st1); 4022 else 4023 ins_req(ins_idx, st1); 4024 } 4025 } 4026 4027 if (PrintCompilation && WizardMode) 4028 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4029 old_subword, old_long, new_int, new_long); 4030 if (C->log() != NULL) 4031 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4032 old_subword, old_long, new_int, new_long); 4033 4034 // Clean up any remaining occurrences of zmem: 4035 remove_extra_zeroes(); 4036 } 4037 4038 // Explore forward from in(start) to find the first fully initialized 4039 // word, and return its offset. Skip groups of subword stores which 4040 // together initialize full words. If in(start) is itself part of a 4041 // fully initialized word, return the offset of in(start). If there 4042 // are no following full-word stores, or if something is fishy, return 4043 // a negative value. 4044 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4045 int int_map = 0; 4046 intptr_t int_map_off = 0; 4047 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4048 4049 for (uint i = start, limit = req(); i < limit; i++) { 4050 Node* st = in(i); 4051 4052 intptr_t st_off = get_store_offset(st, phase); 4053 if (st_off < 0) break; // return conservative answer 4054 4055 int st_size = st->as_Store()->memory_size(); 4056 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4057 return st_off; // we found a complete word init 4058 } 4059 4060 // update the map: 4061 4062 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4063 if (this_int_off != int_map_off) { 4064 // reset the map: 4065 int_map = 0; 4066 int_map_off = this_int_off; 4067 } 4068 4069 int subword_off = st_off - this_int_off; 4070 int_map |= right_n_bits(st_size) << subword_off; 4071 if ((int_map & FULL_MAP) == FULL_MAP) { 4072 return this_int_off; // we found a complete word init 4073 } 4074 4075 // Did this store hit or cross the word boundary? 4076 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4077 if (next_int_off == this_int_off + BytesPerInt) { 4078 // We passed the current int, without fully initializing it. 4079 int_map_off = next_int_off; 4080 int_map >>= BytesPerInt; 4081 } else if (next_int_off > this_int_off + BytesPerInt) { 4082 // We passed the current and next int. 4083 return this_int_off + BytesPerInt; 4084 } 4085 } 4086 4087 return -1; 4088 } 4089 4090 4091 // Called when the associated AllocateNode is expanded into CFG. 4092 // At this point, we may perform additional optimizations. 4093 // Linearize the stores by ascending offset, to make memory 4094 // activity as coherent as possible. 4095 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4096 intptr_t header_size, 4097 Node* size_in_bytes, 4098 PhaseGVN* phase) { 4099 assert(!is_complete(), "not already complete"); 4100 assert(stores_are_sane(phase), ""); 4101 assert(allocation() != NULL, "must be present"); 4102 4103 remove_extra_zeroes(); 4104 4105 if (ReduceFieldZeroing || ReduceBulkZeroing) 4106 // reduce instruction count for common initialization patterns 4107 coalesce_subword_stores(header_size, size_in_bytes, phase); 4108 4109 Node* zmem = zero_memory(); // initially zero memory state 4110 Node* inits = zmem; // accumulating a linearized chain of inits 4111 #ifdef ASSERT 4112 intptr_t first_offset = allocation()->minimum_header_size(); 4113 intptr_t last_init_off = first_offset; // previous init offset 4114 intptr_t last_init_end = first_offset; // previous init offset+size 4115 intptr_t last_tile_end = first_offset; // previous tile offset+size 4116 #endif 4117 intptr_t zeroes_done = header_size; 4118 4119 bool do_zeroing = true; // we might give up if inits are very sparse 4120 int big_init_gaps = 0; // how many large gaps have we seen? 4121 4122 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4123 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4124 4125 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4126 Node* st = in(i); 4127 intptr_t st_off = get_store_offset(st, phase); 4128 if (st_off < 0) 4129 break; // unknown junk in the inits 4130 if (st->in(MemNode::Memory) != zmem) 4131 break; // complicated store chains somehow in list 4132 4133 int st_size = st->as_Store()->memory_size(); 4134 intptr_t next_init_off = st_off + st_size; 4135 4136 if (do_zeroing && zeroes_done < next_init_off) { 4137 // See if this store needs a zero before it or under it. 4138 intptr_t zeroes_needed = st_off; 4139 4140 if (st_size < BytesPerInt) { 4141 // Look for subword stores which only partially initialize words. 4142 // If we find some, we must lay down some word-level zeroes first, 4143 // underneath the subword stores. 4144 // 4145 // Examples: 4146 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4147 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4148 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4149 // 4150 // Note: coalesce_subword_stores may have already done this, 4151 // if it was prompted by constant non-zero subword initializers. 4152 // But this case can still arise with non-constant stores. 4153 4154 intptr_t next_full_store = find_next_fullword_store(i, phase); 4155 4156 // In the examples above: 4157 // in(i) p q r s x y z 4158 // st_off 12 13 14 15 12 13 14 4159 // st_size 1 1 1 1 1 1 1 4160 // next_full_s. 12 16 16 16 16 16 16 4161 // z's_done 12 16 16 16 12 16 12 4162 // z's_needed 12 16 16 16 16 16 16 4163 // zsize 0 0 0 0 4 0 4 4164 if (next_full_store < 0) { 4165 // Conservative tack: Zero to end of current word. 4166 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4167 } else { 4168 // Zero to beginning of next fully initialized word. 4169 // Or, don't zero at all, if we are already in that word. 4170 assert(next_full_store >= zeroes_needed, "must go forward"); 4171 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4172 zeroes_needed = next_full_store; 4173 } 4174 } 4175 4176 if (zeroes_needed > zeroes_done) { 4177 intptr_t zsize = zeroes_needed - zeroes_done; 4178 // Do some incremental zeroing on rawmem, in parallel with inits. 4179 zeroes_done = align_down(zeroes_done, BytesPerInt); 4180 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4181 zeroes_done, zeroes_needed, 4182 phase); 4183 zeroes_done = zeroes_needed; 4184 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4185 do_zeroing = false; // leave the hole, next time 4186 } 4187 } 4188 4189 // Collect the store and move on: 4190 st->set_req(MemNode::Memory, inits); 4191 inits = st; // put it on the linearized chain 4192 set_req(i, zmem); // unhook from previous position 4193 4194 if (zeroes_done == st_off) 4195 zeroes_done = next_init_off; 4196 4197 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4198 4199 #ifdef ASSERT 4200 // Various order invariants. Weaker than stores_are_sane because 4201 // a large constant tile can be filled in by smaller non-constant stores. 4202 assert(st_off >= last_init_off, "inits do not reverse"); 4203 last_init_off = st_off; 4204 const Type* val = NULL; 4205 if (st_size >= BytesPerInt && 4206 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4207 (int)val->basic_type() < (int)T_OBJECT) { 4208 assert(st_off >= last_tile_end, "tiles do not overlap"); 4209 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4210 last_tile_end = MAX2(last_tile_end, next_init_off); 4211 } else { 4212 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4213 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4214 assert(st_off >= last_init_end, "inits do not overlap"); 4215 last_init_end = next_init_off; // it's a non-tile 4216 } 4217 #endif //ASSERT 4218 } 4219 4220 remove_extra_zeroes(); // clear out all the zmems left over 4221 add_req(inits); 4222 4223 if (!(UseTLAB && ZeroTLAB)) { 4224 // If anything remains to be zeroed, zero it all now. 4225 zeroes_done = align_down(zeroes_done, BytesPerInt); 4226 // if it is the last unused 4 bytes of an instance, forget about it 4227 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4228 if (zeroes_done + BytesPerLong >= size_limit) { 4229 AllocateNode* alloc = allocation(); 4230 assert(alloc != NULL, "must be present"); 4231 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 4232 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4233 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4234 if (zeroes_done == k->layout_helper()) 4235 zeroes_done = size_limit; 4236 } 4237 } 4238 if (zeroes_done < size_limit) { 4239 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4240 zeroes_done, size_in_bytes, phase); 4241 } 4242 } 4243 4244 set_complete(phase); 4245 return rawmem; 4246 } 4247 4248 4249 #ifdef ASSERT 4250 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4251 if (is_complete()) 4252 return true; // stores could be anything at this point 4253 assert(allocation() != NULL, "must be present"); 4254 intptr_t last_off = allocation()->minimum_header_size(); 4255 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4256 Node* st = in(i); 4257 intptr_t st_off = get_store_offset(st, phase); 4258 if (st_off < 0) continue; // ignore dead garbage 4259 if (last_off > st_off) { 4260 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4261 this->dump(2); 4262 assert(false, "ascending store offsets"); 4263 return false; 4264 } 4265 last_off = st_off + st->as_Store()->memory_size(); 4266 } 4267 return true; 4268 } 4269 #endif //ASSERT 4270 4271 4272 4273 4274 //============================MergeMemNode===================================== 4275 // 4276 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4277 // contributing store or call operations. Each contributor provides the memory 4278 // state for a particular "alias type" (see Compile::alias_type). For example, 4279 // if a MergeMem has an input X for alias category #6, then any memory reference 4280 // to alias category #6 may use X as its memory state input, as an exact equivalent 4281 // to using the MergeMem as a whole. 4282 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4283 // 4284 // (Here, the <N> notation gives the index of the relevant adr_type.) 4285 // 4286 // In one special case (and more cases in the future), alias categories overlap. 4287 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4288 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4289 // it is exactly equivalent to that state W: 4290 // MergeMem(<Bot>: W) <==> W 4291 // 4292 // Usually, the merge has more than one input. In that case, where inputs 4293 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4294 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4295 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4296 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4297 // 4298 // A merge can take a "wide" memory state as one of its narrow inputs. 4299 // This simply means that the merge observes out only the relevant parts of 4300 // the wide input. That is, wide memory states arriving at narrow merge inputs 4301 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4302 // 4303 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4304 // and that memory slices "leak through": 4305 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4306 // 4307 // But, in such a cascade, repeated memory slices can "block the leak": 4308 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4309 // 4310 // In the last example, Y is not part of the combined memory state of the 4311 // outermost MergeMem. The system must, of course, prevent unschedulable 4312 // memory states from arising, so you can be sure that the state Y is somehow 4313 // a precursor to state Y'. 4314 // 4315 // 4316 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4317 // of each MergeMemNode array are exactly the numerical alias indexes, including 4318 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4319 // Compile::alias_type (and kin) produce and manage these indexes. 4320 // 4321 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4322 // (Note that this provides quick access to the top node inside MergeMem methods, 4323 // without the need to reach out via TLS to Compile::current.) 4324 // 4325 // As a consequence of what was just described, a MergeMem that represents a full 4326 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4327 // containing all alias categories. 4328 // 4329 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4330 // 4331 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4332 // a memory state for the alias type <N>, or else the top node, meaning that 4333 // there is no particular input for that alias type. Note that the length of 4334 // a MergeMem is variable, and may be extended at any time to accommodate new 4335 // memory states at larger alias indexes. When merges grow, they are of course 4336 // filled with "top" in the unused in() positions. 4337 // 4338 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4339 // (Top was chosen because it works smoothly with passes like GCM.) 4340 // 4341 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4342 // the type of random VM bits like TLS references.) Since it is always the 4343 // first non-Bot memory slice, some low-level loops use it to initialize an 4344 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4345 // 4346 // 4347 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4348 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4349 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4350 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4351 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4352 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4353 // 4354 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4355 // really that different from the other memory inputs. An abbreviation called 4356 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4357 // 4358 // 4359 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4360 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4361 // that "emerges though" the base memory will be marked as excluding the alias types 4362 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4363 // 4364 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4365 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4366 // 4367 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4368 // (It is currently unimplemented.) As you can see, the resulting merge is 4369 // actually a disjoint union of memory states, rather than an overlay. 4370 // 4371 4372 //------------------------------MergeMemNode----------------------------------- 4373 Node* MergeMemNode::make_empty_memory() { 4374 Node* empty_memory = (Node*) Compile::current()->top(); 4375 assert(empty_memory->is_top(), "correct sentinel identity"); 4376 return empty_memory; 4377 } 4378 4379 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4380 init_class_id(Class_MergeMem); 4381 // all inputs are nullified in Node::Node(int) 4382 // set_input(0, NULL); // no control input 4383 4384 // Initialize the edges uniformly to top, for starters. 4385 Node* empty_mem = make_empty_memory(); 4386 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4387 init_req(i,empty_mem); 4388 } 4389 assert(empty_memory() == empty_mem, ""); 4390 4391 if( new_base != NULL && new_base->is_MergeMem() ) { 4392 MergeMemNode* mdef = new_base->as_MergeMem(); 4393 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4394 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4395 mms.set_memory(mms.memory2()); 4396 } 4397 assert(base_memory() == mdef->base_memory(), ""); 4398 } else { 4399 set_base_memory(new_base); 4400 } 4401 } 4402 4403 // Make a new, untransformed MergeMem with the same base as 'mem'. 4404 // If mem is itself a MergeMem, populate the result with the same edges. 4405 MergeMemNode* MergeMemNode::make(Node* mem) { 4406 return new MergeMemNode(mem); 4407 } 4408 4409 //------------------------------cmp-------------------------------------------- 4410 uint MergeMemNode::hash() const { return NO_HASH; } 4411 uint MergeMemNode::cmp( const Node &n ) const { 4412 return (&n == this); // Always fail except on self 4413 } 4414 4415 //------------------------------Identity--------------------------------------- 4416 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4417 // Identity if this merge point does not record any interesting memory 4418 // disambiguations. 4419 Node* base_mem = base_memory(); 4420 Node* empty_mem = empty_memory(); 4421 if (base_mem != empty_mem) { // Memory path is not dead? 4422 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4423 Node* mem = in(i); 4424 if (mem != empty_mem && mem != base_mem) { 4425 return this; // Many memory splits; no change 4426 } 4427 } 4428 } 4429 return base_mem; // No memory splits; ID on the one true input 4430 } 4431 4432 //------------------------------Ideal------------------------------------------ 4433 // This method is invoked recursively on chains of MergeMem nodes 4434 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4435 // Remove chain'd MergeMems 4436 // 4437 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4438 // relative to the "in(Bot)". Since we are patching both at the same time, 4439 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4440 // but rewrite each "in(i)" relative to the new "in(Bot)". 4441 Node *progress = NULL; 4442 4443 4444 Node* old_base = base_memory(); 4445 Node* empty_mem = empty_memory(); 4446 if (old_base == empty_mem) 4447 return NULL; // Dead memory path. 4448 4449 MergeMemNode* old_mbase; 4450 if (old_base != NULL && old_base->is_MergeMem()) 4451 old_mbase = old_base->as_MergeMem(); 4452 else 4453 old_mbase = NULL; 4454 Node* new_base = old_base; 4455 4456 // simplify stacked MergeMems in base memory 4457 if (old_mbase) new_base = old_mbase->base_memory(); 4458 4459 // the base memory might contribute new slices beyond my req() 4460 if (old_mbase) grow_to_match(old_mbase); 4461 4462 // Look carefully at the base node if it is a phi. 4463 PhiNode* phi_base; 4464 if (new_base != NULL && new_base->is_Phi()) 4465 phi_base = new_base->as_Phi(); 4466 else 4467 phi_base = NULL; 4468 4469 Node* phi_reg = NULL; 4470 uint phi_len = (uint)-1; 4471 if (phi_base != NULL && !phi_base->is_copy()) { 4472 // do not examine phi if degraded to a copy 4473 phi_reg = phi_base->region(); 4474 phi_len = phi_base->req(); 4475 // see if the phi is unfinished 4476 for (uint i = 1; i < phi_len; i++) { 4477 if (phi_base->in(i) == NULL) { 4478 // incomplete phi; do not look at it yet! 4479 phi_reg = NULL; 4480 phi_len = (uint)-1; 4481 break; 4482 } 4483 } 4484 } 4485 4486 // Note: We do not call verify_sparse on entry, because inputs 4487 // can normalize to the base_memory via subsume_node or similar 4488 // mechanisms. This method repairs that damage. 4489 4490 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4491 4492 // Look at each slice. 4493 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4494 Node* old_in = in(i); 4495 // calculate the old memory value 4496 Node* old_mem = old_in; 4497 if (old_mem == empty_mem) old_mem = old_base; 4498 assert(old_mem == memory_at(i), ""); 4499 4500 // maybe update (reslice) the old memory value 4501 4502 // simplify stacked MergeMems 4503 Node* new_mem = old_mem; 4504 MergeMemNode* old_mmem; 4505 if (old_mem != NULL && old_mem->is_MergeMem()) 4506 old_mmem = old_mem->as_MergeMem(); 4507 else 4508 old_mmem = NULL; 4509 if (old_mmem == this) { 4510 // This can happen if loops break up and safepoints disappear. 4511 // A merge of BotPtr (default) with a RawPtr memory derived from a 4512 // safepoint can be rewritten to a merge of the same BotPtr with 4513 // the BotPtr phi coming into the loop. If that phi disappears 4514 // also, we can end up with a self-loop of the mergemem. 4515 // In general, if loops degenerate and memory effects disappear, 4516 // a mergemem can be left looking at itself. This simply means 4517 // that the mergemem's default should be used, since there is 4518 // no longer any apparent effect on this slice. 4519 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4520 // from start. Update the input to TOP. 4521 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4522 } 4523 else if (old_mmem != NULL) { 4524 new_mem = old_mmem->memory_at(i); 4525 } 4526 // else preceding memory was not a MergeMem 4527 4528 // replace equivalent phis (unfortunately, they do not GVN together) 4529 if (new_mem != NULL && new_mem != new_base && 4530 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4531 if (new_mem->is_Phi()) { 4532 PhiNode* phi_mem = new_mem->as_Phi(); 4533 for (uint i = 1; i < phi_len; i++) { 4534 if (phi_base->in(i) != phi_mem->in(i)) { 4535 phi_mem = NULL; 4536 break; 4537 } 4538 } 4539 if (phi_mem != NULL) { 4540 // equivalent phi nodes; revert to the def 4541 new_mem = new_base; 4542 } 4543 } 4544 } 4545 4546 // maybe store down a new value 4547 Node* new_in = new_mem; 4548 if (new_in == new_base) new_in = empty_mem; 4549 4550 if (new_in != old_in) { 4551 // Warning: Do not combine this "if" with the previous "if" 4552 // A memory slice might have be be rewritten even if it is semantically 4553 // unchanged, if the base_memory value has changed. 4554 set_req(i, new_in); 4555 progress = this; // Report progress 4556 } 4557 } 4558 4559 if (new_base != old_base) { 4560 set_req(Compile::AliasIdxBot, new_base); 4561 // Don't use set_base_memory(new_base), because we need to update du. 4562 assert(base_memory() == new_base, ""); 4563 progress = this; 4564 } 4565 4566 if( base_memory() == this ) { 4567 // a self cycle indicates this memory path is dead 4568 set_req(Compile::AliasIdxBot, empty_mem); 4569 } 4570 4571 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4572 // Recursion must occur after the self cycle check above 4573 if( base_memory()->is_MergeMem() ) { 4574 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4575 Node *m = phase->transform(new_mbase); // Rollup any cycles 4576 if( m != NULL && 4577 (m->is_top() || 4578 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4579 // propagate rollup of dead cycle to self 4580 set_req(Compile::AliasIdxBot, empty_mem); 4581 } 4582 } 4583 4584 if( base_memory() == empty_mem ) { 4585 progress = this; 4586 // Cut inputs during Parse phase only. 4587 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4588 if( !can_reshape ) { 4589 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4590 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4591 } 4592 } 4593 } 4594 4595 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4596 // Check if PhiNode::Ideal's "Split phis through memory merges" 4597 // transform should be attempted. Look for this->phi->this cycle. 4598 uint merge_width = req(); 4599 if (merge_width > Compile::AliasIdxRaw) { 4600 PhiNode* phi = base_memory()->as_Phi(); 4601 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4602 if (phi->in(i) == this) { 4603 phase->is_IterGVN()->_worklist.push(phi); 4604 break; 4605 } 4606 } 4607 } 4608 } 4609 4610 assert(progress || verify_sparse(), "please, no dups of base"); 4611 return progress; 4612 } 4613 4614 //-------------------------set_base_memory------------------------------------- 4615 void MergeMemNode::set_base_memory(Node *new_base) { 4616 Node* empty_mem = empty_memory(); 4617 set_req(Compile::AliasIdxBot, new_base); 4618 assert(memory_at(req()) == new_base, "must set default memory"); 4619 // Clear out other occurrences of new_base: 4620 if (new_base != empty_mem) { 4621 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4622 if (in(i) == new_base) set_req(i, empty_mem); 4623 } 4624 } 4625 } 4626 4627 //------------------------------out_RegMask------------------------------------ 4628 const RegMask &MergeMemNode::out_RegMask() const { 4629 return RegMask::Empty; 4630 } 4631 4632 //------------------------------dump_spec-------------------------------------- 4633 #ifndef PRODUCT 4634 void MergeMemNode::dump_spec(outputStream *st) const { 4635 st->print(" {"); 4636 Node* base_mem = base_memory(); 4637 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4638 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4639 if (mem == base_mem) { st->print(" -"); continue; } 4640 st->print( " N%d:", mem->_idx ); 4641 Compile::current()->get_adr_type(i)->dump_on(st); 4642 } 4643 st->print(" }"); 4644 } 4645 #endif // !PRODUCT 4646 4647 4648 #ifdef ASSERT 4649 static bool might_be_same(Node* a, Node* b) { 4650 if (a == b) return true; 4651 if (!(a->is_Phi() || b->is_Phi())) return false; 4652 // phis shift around during optimization 4653 return true; // pretty stupid... 4654 } 4655 4656 // verify a narrow slice (either incoming or outgoing) 4657 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4658 if (!VerifyAliases) return; // don't bother to verify unless requested 4659 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4660 if (Node::in_dump()) return; // muzzle asserts when printing 4661 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4662 assert(n != NULL, ""); 4663 // Elide intervening MergeMem's 4664 while (n->is_MergeMem()) { 4665 n = n->as_MergeMem()->memory_at(alias_idx); 4666 } 4667 Compile* C = Compile::current(); 4668 const TypePtr* n_adr_type = n->adr_type(); 4669 if (n == m->empty_memory()) { 4670 // Implicit copy of base_memory() 4671 } else if (n_adr_type != TypePtr::BOTTOM) { 4672 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4673 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4674 } else { 4675 // A few places like make_runtime_call "know" that VM calls are narrow, 4676 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4677 bool expected_wide_mem = false; 4678 if (n == m->base_memory()) { 4679 expected_wide_mem = true; 4680 } else if (alias_idx == Compile::AliasIdxRaw || 4681 n == m->memory_at(Compile::AliasIdxRaw)) { 4682 expected_wide_mem = true; 4683 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4684 // memory can "leak through" calls on channels that 4685 // are write-once. Allow this also. 4686 expected_wide_mem = true; 4687 } 4688 assert(expected_wide_mem, "expected narrow slice replacement"); 4689 } 4690 } 4691 #else // !ASSERT 4692 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4693 #endif 4694 4695 4696 //-----------------------------memory_at--------------------------------------- 4697 Node* MergeMemNode::memory_at(uint alias_idx) const { 4698 assert(alias_idx >= Compile::AliasIdxRaw || 4699 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4700 "must avoid base_memory and AliasIdxTop"); 4701 4702 // Otherwise, it is a narrow slice. 4703 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4704 Compile *C = Compile::current(); 4705 if (is_empty_memory(n)) { 4706 // the array is sparse; empty slots are the "top" node 4707 n = base_memory(); 4708 assert(Node::in_dump() 4709 || n == NULL || n->bottom_type() == Type::TOP 4710 || n->adr_type() == NULL // address is TOP 4711 || n->adr_type() == TypePtr::BOTTOM 4712 || n->adr_type() == TypeRawPtr::BOTTOM 4713 || Compile::current()->AliasLevel() == 0, 4714 "must be a wide memory"); 4715 // AliasLevel == 0 if we are organizing the memory states manually. 4716 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4717 } else { 4718 // make sure the stored slice is sane 4719 #ifdef ASSERT 4720 if (VMError::is_error_reported() || Node::in_dump()) { 4721 } else if (might_be_same(n, base_memory())) { 4722 // Give it a pass: It is a mostly harmless repetition of the base. 4723 // This can arise normally from node subsumption during optimization. 4724 } else { 4725 verify_memory_slice(this, alias_idx, n); 4726 } 4727 #endif 4728 } 4729 return n; 4730 } 4731 4732 //---------------------------set_memory_at------------------------------------- 4733 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4734 verify_memory_slice(this, alias_idx, n); 4735 Node* empty_mem = empty_memory(); 4736 if (n == base_memory()) n = empty_mem; // collapse default 4737 uint need_req = alias_idx+1; 4738 if (req() < need_req) { 4739 if (n == empty_mem) return; // already the default, so do not grow me 4740 // grow the sparse array 4741 do { 4742 add_req(empty_mem); 4743 } while (req() < need_req); 4744 } 4745 set_req( alias_idx, n ); 4746 } 4747 4748 4749 4750 //--------------------------iteration_setup------------------------------------ 4751 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4752 if (other != NULL) { 4753 grow_to_match(other); 4754 // invariant: the finite support of mm2 is within mm->req() 4755 #ifdef ASSERT 4756 for (uint i = req(); i < other->req(); i++) { 4757 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4758 } 4759 #endif 4760 } 4761 // Replace spurious copies of base_memory by top. 4762 Node* base_mem = base_memory(); 4763 if (base_mem != NULL && !base_mem->is_top()) { 4764 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4765 if (in(i) == base_mem) 4766 set_req(i, empty_memory()); 4767 } 4768 } 4769 } 4770 4771 //---------------------------grow_to_match------------------------------------- 4772 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4773 Node* empty_mem = empty_memory(); 4774 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4775 // look for the finite support of the other memory 4776 for (uint i = other->req(); --i >= req(); ) { 4777 if (other->in(i) != empty_mem) { 4778 uint new_len = i+1; 4779 while (req() < new_len) add_req(empty_mem); 4780 break; 4781 } 4782 } 4783 } 4784 4785 //---------------------------verify_sparse------------------------------------- 4786 #ifndef PRODUCT 4787 bool MergeMemNode::verify_sparse() const { 4788 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4789 Node* base_mem = base_memory(); 4790 // The following can happen in degenerate cases, since empty==top. 4791 if (is_empty_memory(base_mem)) return true; 4792 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4793 assert(in(i) != NULL, "sane slice"); 4794 if (in(i) == base_mem) return false; // should have been the sentinel value! 4795 } 4796 return true; 4797 } 4798 4799 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4800 Node* n; 4801 n = mm->in(idx); 4802 if (mem == n) return true; // might be empty_memory() 4803 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4804 if (mem == n) return true; 4805 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4806 if (mem == n) return true; 4807 if (n == NULL) break; 4808 } 4809 return false; 4810 } 4811 #endif // !PRODUCT