1 /* 2 * Copyright (c) 1997, 2019, 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 "opto/rootnode.hpp" 48 #include "utilities/align.hpp" 49 #include "utilities/copy.hpp" 50 #include "utilities/macros.hpp" 51 #include "utilities/vmError.hpp" 52 #if INCLUDE_ZGC 53 #include "gc/z/c2/zBarrierSetC2.hpp" 54 #endif 55 56 // Portions of code courtesy of Clifford Click 57 58 // Optimization - Graph Style 59 60 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 61 62 //============================================================================= 63 uint MemNode::size_of() const { return sizeof(*this); } 64 65 const TypePtr *MemNode::adr_type() const { 66 Node* adr = in(Address); 67 if (adr == NULL) return NULL; // node is dead 68 const TypePtr* cross_check = NULL; 69 DEBUG_ONLY(cross_check = _adr_type); 70 return calculate_adr_type(adr->bottom_type(), cross_check); 71 } 72 73 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 74 if (adr != NULL) { 75 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 76 return true; 77 } 78 } 79 return false; 80 } 81 82 #ifndef PRODUCT 83 void MemNode::dump_spec(outputStream *st) const { 84 if (in(Address) == NULL) return; // node is dead 85 #ifndef ASSERT 86 // fake the missing field 87 const TypePtr* _adr_type = NULL; 88 if (in(Address) != NULL) 89 _adr_type = in(Address)->bottom_type()->isa_ptr(); 90 #endif 91 dump_adr_type(this, _adr_type, st); 92 93 Compile* C = Compile::current(); 94 if (C->alias_type(_adr_type)->is_volatile()) { 95 st->print(" Volatile!"); 96 } 97 if (_unaligned_access) { 98 st->print(" unaligned"); 99 } 100 if (_mismatched_access) { 101 st->print(" mismatched"); 102 } 103 if (_unsafe_access) { 104 st->print(" unsafe"); 105 } 106 } 107 108 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 109 st->print(" @"); 110 if (adr_type == NULL) { 111 st->print("NULL"); 112 } else { 113 adr_type->dump_on(st); 114 Compile* C = Compile::current(); 115 Compile::AliasType* atp = NULL; 116 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 117 if (atp == NULL) 118 st->print(", idx=?\?;"); 119 else if (atp->index() == Compile::AliasIdxBot) 120 st->print(", idx=Bot;"); 121 else if (atp->index() == Compile::AliasIdxTop) 122 st->print(", idx=Top;"); 123 else if (atp->index() == Compile::AliasIdxRaw) 124 st->print(", idx=Raw;"); 125 else { 126 ciField* field = atp->field(); 127 if (field) { 128 st->print(", name="); 129 field->print_name_on(st); 130 } 131 st->print(", idx=%d;", atp->index()); 132 } 133 } 134 } 135 136 extern void print_alias_types(); 137 138 #endif 139 140 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 141 assert((t_oop != NULL), "sanity"); 142 bool is_instance = t_oop->is_known_instance_field(); 143 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 144 (load != NULL) && load->is_Load() && 145 (phase->is_IterGVN() != NULL); 146 if (!(is_instance || is_boxed_value_load)) 147 return mchain; // don't try to optimize non-instance types 148 uint instance_id = t_oop->instance_id(); 149 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory); 150 Node *prev = NULL; 151 Node *result = mchain; 152 while (prev != result) { 153 prev = result; 154 if (result == start_mem) 155 break; // hit one of our sentinels 156 // skip over a call which does not affect this memory slice 157 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 158 Node *proj_in = result->in(0); 159 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 160 break; // hit one of our sentinels 161 } else if (proj_in->is_Call()) { 162 // ArrayCopyNodes processed here as well 163 CallNode *call = proj_in->as_Call(); 164 if (!call->may_modify(t_oop, phase)) { // returns false for instances 165 result = call->in(TypeFunc::Memory); 166 } 167 } else if (proj_in->is_Initialize()) { 168 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 169 // Stop if this is the initialization for the object instance which 170 // contains this memory slice, otherwise skip over it. 171 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 172 break; 173 } 174 if (is_instance) { 175 result = proj_in->in(TypeFunc::Memory); 176 } else if (is_boxed_value_load) { 177 Node* klass = alloc->in(AllocateNode::KlassNode); 178 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 179 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 180 result = proj_in->in(TypeFunc::Memory); // not related allocation 181 } 182 } 183 } else if (proj_in->is_MemBar()) { 184 ArrayCopyNode* ac = NULL; 185 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 186 break; 187 } 188 result = proj_in->in(TypeFunc::Memory); 189 } else { 190 assert(false, "unexpected projection"); 191 } 192 } else if (result->is_ClearArray()) { 193 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 194 // Can not bypass initialization of the instance 195 // we are looking for. 196 break; 197 } 198 // Otherwise skip it (the call updated 'result' value). 199 } else if (result->is_MergeMem()) { 200 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 201 } 202 } 203 return result; 204 } 205 206 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 207 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 208 if (t_oop == NULL) 209 return mchain; // don't try to optimize non-oop types 210 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 211 bool is_instance = t_oop->is_known_instance_field(); 212 PhaseIterGVN *igvn = phase->is_IterGVN(); 213 if (is_instance && igvn != NULL && result->is_Phi()) { 214 PhiNode *mphi = result->as_Phi(); 215 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 216 const TypePtr *t = mphi->adr_type(); 217 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 218 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 219 t->is_oopptr()->cast_to_exactness(true) 220 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 221 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) { 222 // clone the Phi with our address type 223 result = mphi->split_out_instance(t_adr, igvn); 224 } else { 225 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 226 } 227 } 228 return result; 229 } 230 231 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 232 uint alias_idx = phase->C->get_alias_index(tp); 233 Node *mem = mmem; 234 #ifdef ASSERT 235 { 236 // Check that current type is consistent with the alias index used during graph construction 237 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 238 bool consistent = adr_check == NULL || adr_check->empty() || 239 phase->C->must_alias(adr_check, alias_idx ); 240 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 241 if( !consistent && adr_check != NULL && !adr_check->empty() && 242 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 243 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 244 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 245 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 246 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 247 // don't assert if it is dead code. 248 consistent = true; 249 } 250 if( !consistent ) { 251 st->print("alias_idx==%d, adr_check==", alias_idx); 252 if( adr_check == NULL ) { 253 st->print("NULL"); 254 } else { 255 adr_check->dump(); 256 } 257 st->cr(); 258 print_alias_types(); 259 assert(consistent, "adr_check must match alias idx"); 260 } 261 } 262 #endif 263 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 264 // means an array I have not precisely typed yet. Do not do any 265 // alias stuff with it any time soon. 266 const TypeOopPtr *toop = tp->isa_oopptr(); 267 if( tp->base() != Type::AnyPtr && 268 !(toop && 269 toop->klass() != NULL && 270 toop->klass()->is_java_lang_Object() && 271 toop->offset() == Type::OffsetBot) ) { 272 // compress paths and change unreachable cycles to TOP 273 // If not, we can update the input infinitely along a MergeMem cycle 274 // Equivalent code in PhiNode::Ideal 275 Node* m = phase->transform(mmem); 276 // If transformed to a MergeMem, get the desired slice 277 // Otherwise the returned node represents memory for every slice 278 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 279 // Update input if it is progress over what we have now 280 } 281 return mem; 282 } 283 284 //--------------------------Ideal_common--------------------------------------- 285 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 286 // Unhook non-raw memories from complete (macro-expanded) initializations. 287 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 288 // If our control input is a dead region, kill all below the region 289 Node *ctl = in(MemNode::Control); 290 if (ctl && remove_dead_region(phase, can_reshape)) 291 return this; 292 ctl = in(MemNode::Control); 293 // Don't bother trying to transform a dead node 294 if (ctl && ctl->is_top()) return NodeSentinel; 295 296 PhaseIterGVN *igvn = phase->is_IterGVN(); 297 // Wait if control on the worklist. 298 if (ctl && can_reshape && igvn != NULL) { 299 Node* bol = NULL; 300 Node* cmp = NULL; 301 if (ctl->in(0)->is_If()) { 302 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 303 bol = ctl->in(0)->in(1); 304 if (bol->is_Bool()) 305 cmp = ctl->in(0)->in(1)->in(1); 306 } 307 if (igvn->_worklist.member(ctl) || 308 (bol != NULL && igvn->_worklist.member(bol)) || 309 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 310 // This control path may be dead. 311 // Delay this memory node transformation until the control is processed. 312 phase->is_IterGVN()->_worklist.push(this); 313 return NodeSentinel; // caller will return NULL 314 } 315 } 316 // Ignore if memory is dead, or self-loop 317 Node *mem = in(MemNode::Memory); 318 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 319 assert(mem != this, "dead loop in MemNode::Ideal"); 320 321 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 322 // This memory slice may be dead. 323 // Delay this mem node transformation until the memory is processed. 324 phase->is_IterGVN()->_worklist.push(this); 325 return NodeSentinel; // caller will return NULL 326 } 327 328 Node *address = in(MemNode::Address); 329 const Type *t_adr = phase->type(address); 330 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 331 332 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) { 333 // Unsafe off-heap access with zero address. Remove access and other control users 334 // to not confuse optimizations and add a HaltNode to fail if this is ever executed. 335 assert(ctl != NULL, "unsafe accesses should be control dependent"); 336 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) { 337 Node* u = ctl->fast_out(i); 338 if (u != ctl) { 339 igvn->rehash_node_delayed(u); 340 int nb = u->replace_edge(ctl, phase->C->top()); 341 --i, imax -= nb; 342 } 343 } 344 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr)); 345 Node* halt = igvn->transform(new HaltNode(ctl, frame)); 346 phase->C->root()->add_req(halt); 347 return this; 348 } 349 350 if (can_reshape && igvn != NULL && 351 (igvn->_worklist.member(address) || 352 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 353 // The address's base and type may change when the address is processed. 354 // Delay this mem node transformation until the address is processed. 355 phase->is_IterGVN()->_worklist.push(this); 356 return NodeSentinel; // caller will return NULL 357 } 358 359 // Do NOT remove or optimize the next lines: ensure a new alias index 360 // is allocated for an oop pointer type before Escape Analysis. 361 // Note: C++ will not remove it since the call has side effect. 362 if (t_adr->isa_oopptr()) { 363 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 364 } 365 366 Node* base = NULL; 367 if (address->is_AddP()) { 368 base = address->in(AddPNode::Base); 369 } 370 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 371 !t_adr->isa_rawptr()) { 372 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 373 // Skip this node optimization if its address has TOP base. 374 return NodeSentinel; // caller will return NULL 375 } 376 377 // Avoid independent memory operations 378 Node* old_mem = mem; 379 380 // The code which unhooks non-raw memories from complete (macro-expanded) 381 // initializations was removed. After macro-expansion all stores catched 382 // by Initialize node became raw stores and there is no information 383 // which memory slices they modify. So it is unsafe to move any memory 384 // operation above these stores. Also in most cases hooked non-raw memories 385 // were already unhooked by using information from detect_ptr_independence() 386 // and find_previous_store(). 387 388 if (mem->is_MergeMem()) { 389 MergeMemNode* mmem = mem->as_MergeMem(); 390 const TypePtr *tp = t_adr->is_ptr(); 391 392 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 393 } 394 395 if (mem != old_mem) { 396 set_req(MemNode::Memory, mem); 397 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) { 398 igvn->_worklist.push(old_mem); 399 } 400 if (phase->type(mem) == Type::TOP) return NodeSentinel; 401 return this; 402 } 403 404 // let the subclass continue analyzing... 405 return NULL; 406 } 407 408 // Helper function for proving some simple control dominations. 409 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 410 // Already assumes that 'dom' is available at 'sub', and that 'sub' 411 // is not a constant (dominated by the method's StartNode). 412 // Used by MemNode::find_previous_store to prove that the 413 // control input of a memory operation predates (dominates) 414 // an allocation it wants to look past. 415 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 416 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 417 return false; // Conservative answer for dead code 418 419 // Check 'dom'. Skip Proj and CatchProj nodes. 420 dom = dom->find_exact_control(dom); 421 if (dom == NULL || dom->is_top()) 422 return false; // Conservative answer for dead code 423 424 if (dom == sub) { 425 // For the case when, for example, 'sub' is Initialize and the original 426 // 'dom' is Proj node of the 'sub'. 427 return false; 428 } 429 430 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 431 return true; 432 433 // 'dom' dominates 'sub' if its control edge and control edges 434 // of all its inputs dominate or equal to sub's control edge. 435 436 // Currently 'sub' is either Allocate, Initialize or Start nodes. 437 // Or Region for the check in LoadNode::Ideal(); 438 // 'sub' should have sub->in(0) != NULL. 439 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 440 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 441 442 // Get control edge of 'sub'. 443 Node* orig_sub = sub; 444 sub = sub->find_exact_control(sub->in(0)); 445 if (sub == NULL || sub->is_top()) 446 return false; // Conservative answer for dead code 447 448 assert(sub->is_CFG(), "expecting control"); 449 450 if (sub == dom) 451 return true; 452 453 if (sub->is_Start() || sub->is_Root()) 454 return false; 455 456 { 457 // Check all control edges of 'dom'. 458 459 ResourceMark rm; 460 Arena* arena = Thread::current()->resource_area(); 461 Node_List nlist(arena); 462 Unique_Node_List dom_list(arena); 463 464 dom_list.push(dom); 465 bool only_dominating_controls = false; 466 467 for (uint next = 0; next < dom_list.size(); next++) { 468 Node* n = dom_list.at(next); 469 if (n == orig_sub) 470 return false; // One of dom's inputs dominated by sub. 471 if (!n->is_CFG() && n->pinned()) { 472 // Check only own control edge for pinned non-control nodes. 473 n = n->find_exact_control(n->in(0)); 474 if (n == NULL || n->is_top()) 475 return false; // Conservative answer for dead code 476 assert(n->is_CFG(), "expecting control"); 477 dom_list.push(n); 478 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 479 only_dominating_controls = true; 480 } else if (n->is_CFG()) { 481 if (n->dominates(sub, nlist)) 482 only_dominating_controls = true; 483 else 484 return false; 485 } else { 486 // First, own control edge. 487 Node* m = n->find_exact_control(n->in(0)); 488 if (m != NULL) { 489 if (m->is_top()) 490 return false; // Conservative answer for dead code 491 dom_list.push(m); 492 } 493 // Now, the rest of edges. 494 uint cnt = n->req(); 495 for (uint i = 1; i < cnt; i++) { 496 m = n->find_exact_control(n->in(i)); 497 if (m == NULL || m->is_top()) 498 continue; 499 dom_list.push(m); 500 } 501 } 502 } 503 return only_dominating_controls; 504 } 505 } 506 507 //---------------------detect_ptr_independence--------------------------------- 508 // Used by MemNode::find_previous_store to prove that two base 509 // pointers are never equal. 510 // The pointers are accompanied by their associated allocations, 511 // if any, which have been previously discovered by the caller. 512 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 513 Node* p2, AllocateNode* a2, 514 PhaseTransform* phase) { 515 // Attempt to prove that these two pointers cannot be aliased. 516 // They may both manifestly be allocations, and they should differ. 517 // Or, if they are not both allocations, they can be distinct constants. 518 // Otherwise, one is an allocation and the other a pre-existing value. 519 if (a1 == NULL && a2 == NULL) { // neither an allocation 520 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 521 } else if (a1 != NULL && a2 != NULL) { // both allocations 522 return (a1 != a2); 523 } else if (a1 != NULL) { // one allocation a1 524 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 525 return all_controls_dominate(p2, a1); 526 } else { //(a2 != NULL) // one allocation a2 527 return all_controls_dominate(p1, a2); 528 } 529 return false; 530 } 531 532 533 // Find an arraycopy that must have set (can_see_stored_value=true) or 534 // could have set (can_see_stored_value=false) the value for this load 535 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 536 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 537 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 538 Node* mb = mem->in(0); 539 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 540 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 541 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 542 if (ac->is_clonebasic()) { 543 intptr_t offset; 544 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 545 if (alloc != NULL && alloc == ld_alloc) { 546 return ac; 547 } 548 } 549 } 550 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 551 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 552 553 if (ac->is_arraycopy_validated() || 554 ac->is_copyof_validated() || 555 ac->is_copyofrange_validated()) { 556 Node* ld_addp = in(MemNode::Address); 557 if (ld_addp->is_AddP()) { 558 Node* ld_base = ld_addp->in(AddPNode::Address); 559 Node* ld_offs = ld_addp->in(AddPNode::Offset); 560 561 Node* dest = ac->in(ArrayCopyNode::Dest); 562 563 if (dest == ld_base) { 564 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 565 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 566 return ac; 567 } 568 if (!can_see_stored_value) { 569 mem = ac->in(TypeFunc::Memory); 570 } 571 } 572 } 573 } 574 } 575 return NULL; 576 } 577 578 // The logic for reordering loads and stores uses four steps: 579 // (a) Walk carefully past stores and initializations which we 580 // can prove are independent of this load. 581 // (b) Observe that the next memory state makes an exact match 582 // with self (load or store), and locate the relevant store. 583 // (c) Ensure that, if we were to wire self directly to the store, 584 // the optimizer would fold it up somehow. 585 // (d) Do the rewiring, and return, depending on some other part of 586 // the optimizer to fold up the load. 587 // This routine handles steps (a) and (b). Steps (c) and (d) are 588 // specific to loads and stores, so they are handled by the callers. 589 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 590 // 591 Node* MemNode::find_previous_store(PhaseTransform* phase) { 592 Node* ctrl = in(MemNode::Control); 593 Node* adr = in(MemNode::Address); 594 intptr_t offset = 0; 595 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 596 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 597 598 if (offset == Type::OffsetBot) 599 return NULL; // cannot unalias unless there are precise offsets 600 601 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 602 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 603 604 intptr_t size_in_bytes = memory_size(); 605 606 Node* mem = in(MemNode::Memory); // start searching here... 607 608 int cnt = 50; // Cycle limiter 609 for (;;) { // While we can dance past unrelated stores... 610 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 611 612 Node* prev = mem; 613 if (mem->is_Store()) { 614 Node* st_adr = mem->in(MemNode::Address); 615 intptr_t st_offset = 0; 616 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 617 if (st_base == NULL) 618 break; // inscrutable pointer 619 620 // For raw accesses it's not enough to prove that constant offsets don't intersect. 621 // We need the bases to be the equal in order for the offset check to make sense. 622 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 623 break; 624 } 625 626 if (st_offset != offset && st_offset != Type::OffsetBot) { 627 const int MAX_STORE = BytesPerLong; 628 if (st_offset >= offset + size_in_bytes || 629 st_offset <= offset - MAX_STORE || 630 st_offset <= offset - mem->as_Store()->memory_size()) { 631 // Success: The offsets are provably independent. 632 // (You may ask, why not just test st_offset != offset and be done? 633 // The answer is that stores of different sizes can co-exist 634 // in the same sequence of RawMem effects. We sometimes initialize 635 // a whole 'tile' of array elements with a single jint or jlong.) 636 mem = mem->in(MemNode::Memory); 637 continue; // (a) advance through independent store memory 638 } 639 } 640 if (st_base != base && 641 detect_ptr_independence(base, alloc, 642 st_base, 643 AllocateNode::Ideal_allocation(st_base, phase), 644 phase)) { 645 // Success: The bases are provably independent. 646 mem = mem->in(MemNode::Memory); 647 continue; // (a) advance through independent store memory 648 } 649 650 // (b) At this point, if the bases or offsets do not agree, we lose, 651 // since we have not managed to prove 'this' and 'mem' independent. 652 if (st_base == base && st_offset == offset) { 653 return mem; // let caller handle steps (c), (d) 654 } 655 656 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 657 InitializeNode* st_init = mem->in(0)->as_Initialize(); 658 AllocateNode* st_alloc = st_init->allocation(); 659 if (st_alloc == NULL) 660 break; // something degenerated 661 bool known_identical = false; 662 bool known_independent = false; 663 if (alloc == st_alloc) 664 known_identical = true; 665 else if (alloc != NULL) 666 known_independent = true; 667 else if (all_controls_dominate(this, st_alloc)) 668 known_independent = true; 669 670 if (known_independent) { 671 // The bases are provably independent: Either they are 672 // manifestly distinct allocations, or else the control 673 // of this load dominates the store's allocation. 674 int alias_idx = phase->C->get_alias_index(adr_type()); 675 if (alias_idx == Compile::AliasIdxRaw) { 676 mem = st_alloc->in(TypeFunc::Memory); 677 } else { 678 mem = st_init->memory(alias_idx); 679 } 680 continue; // (a) advance through independent store memory 681 } 682 683 // (b) at this point, if we are not looking at a store initializing 684 // the same allocation we are loading from, we lose. 685 if (known_identical) { 686 // From caller, can_see_stored_value will consult find_captured_store. 687 return mem; // let caller handle steps (c), (d) 688 } 689 690 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 691 if (prev != mem) { 692 // Found an arraycopy but it doesn't affect that load 693 continue; 694 } 695 // Found an arraycopy that may affect that load 696 return mem; 697 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 698 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 699 if (mem->is_Proj() && mem->in(0)->is_Call()) { 700 // ArrayCopyNodes processed here as well. 701 CallNode *call = mem->in(0)->as_Call(); 702 if (!call->may_modify(addr_t, phase)) { 703 mem = call->in(TypeFunc::Memory); 704 continue; // (a) advance through independent call memory 705 } 706 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 707 ArrayCopyNode* ac = NULL; 708 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 709 break; 710 } 711 mem = mem->in(0)->in(TypeFunc::Memory); 712 continue; // (a) advance through independent MemBar memory 713 } else if (mem->is_ClearArray()) { 714 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 715 // (the call updated 'mem' value) 716 continue; // (a) advance through independent allocation memory 717 } else { 718 // Can not bypass initialization of the instance 719 // we are looking for. 720 return mem; 721 } 722 } else if (mem->is_MergeMem()) { 723 int alias_idx = phase->C->get_alias_index(adr_type()); 724 mem = mem->as_MergeMem()->memory_at(alias_idx); 725 continue; // (a) advance through independent MergeMem memory 726 } 727 } 728 729 // Unless there is an explicit 'continue', we must bail out here, 730 // because 'mem' is an inscrutable memory state (e.g., a call). 731 break; 732 } 733 734 return NULL; // bail out 735 } 736 737 //----------------------calculate_adr_type------------------------------------- 738 // Helper function. Notices when the given type of address hits top or bottom. 739 // Also, asserts a cross-check of the type against the expected address type. 740 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 741 if (t == Type::TOP) return NULL; // does not touch memory any more? 742 #ifdef PRODUCT 743 cross_check = NULL; 744 #else 745 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL; 746 #endif 747 const TypePtr* tp = t->isa_ptr(); 748 if (tp == NULL) { 749 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 750 return TypePtr::BOTTOM; // touches lots of memory 751 } else { 752 #ifdef ASSERT 753 // %%%% [phh] We don't check the alias index if cross_check is 754 // TypeRawPtr::BOTTOM. Needs to be investigated. 755 if (cross_check != NULL && 756 cross_check != TypePtr::BOTTOM && 757 cross_check != TypeRawPtr::BOTTOM) { 758 // Recheck the alias index, to see if it has changed (due to a bug). 759 Compile* C = Compile::current(); 760 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 761 "must stay in the original alias category"); 762 // The type of the address must be contained in the adr_type, 763 // disregarding "null"-ness. 764 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 765 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 766 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 767 "real address must not escape from expected memory type"); 768 } 769 #endif 770 return tp; 771 } 772 } 773 774 //============================================================================= 775 // Should LoadNode::Ideal() attempt to remove control edges? 776 bool LoadNode::can_remove_control() const { 777 return true; 778 } 779 uint LoadNode::size_of() const { return sizeof(*this); } 780 bool LoadNode::cmp( const Node &n ) const 781 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 782 const Type *LoadNode::bottom_type() const { return _type; } 783 uint LoadNode::ideal_reg() const { 784 return _type->ideal_reg(); 785 } 786 787 #ifndef PRODUCT 788 void LoadNode::dump_spec(outputStream *st) const { 789 MemNode::dump_spec(st); 790 if( !Verbose && !WizardMode ) { 791 // standard dump does this in Verbose and WizardMode 792 st->print(" #"); _type->dump_on(st); 793 } 794 if (!depends_only_on_test()) { 795 st->print(" (does not depend only on test)"); 796 } 797 } 798 #endif 799 800 #ifdef ASSERT 801 //----------------------------is_immutable_value------------------------------- 802 // Helper function to allow a raw load without control edge for some cases 803 bool LoadNode::is_immutable_value(Node* adr) { 804 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 805 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 806 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 807 in_bytes(JavaThread::osthread_offset()))); 808 } 809 #endif 810 811 //----------------------------LoadNode::make----------------------------------- 812 // Polymorphic factory method: 813 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 814 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 815 Compile* C = gvn.C; 816 817 // sanity check the alias category against the created node type 818 assert(!(adr_type->isa_oopptr() && 819 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 820 "use LoadKlassNode instead"); 821 assert(!(adr_type->isa_aryptr() && 822 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 823 "use LoadRangeNode instead"); 824 // Check control edge of raw loads 825 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 826 // oop will be recorded in oop map if load crosses safepoint 827 rt->isa_oopptr() || is_immutable_value(adr), 828 "raw memory operations should have control edge"); 829 LoadNode* load = NULL; 830 switch (bt) { 831 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 832 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 833 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 834 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 835 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 836 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 837 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 838 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 839 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 840 case T_OBJECT: 841 #ifdef _LP64 842 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 843 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 844 } else 845 #endif 846 { 847 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 848 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 849 } 850 break; 851 default: 852 ShouldNotReachHere(); 853 break; 854 } 855 assert(load != NULL, "LoadNode should have been created"); 856 if (unaligned) { 857 load->set_unaligned_access(); 858 } 859 if (mismatched) { 860 load->set_mismatched_access(); 861 } 862 if (unsafe) { 863 load->set_unsafe_access(); 864 } 865 if (load->Opcode() == Op_LoadN) { 866 Node* ld = gvn.transform(load); 867 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 868 } 869 870 return load; 871 } 872 873 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 874 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 875 bool require_atomic = true; 876 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 877 if (unaligned) { 878 load->set_unaligned_access(); 879 } 880 if (mismatched) { 881 load->set_mismatched_access(); 882 } 883 if (unsafe) { 884 load->set_unsafe_access(); 885 } 886 return load; 887 } 888 889 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 890 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) { 891 bool require_atomic = true; 892 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 893 if (unaligned) { 894 load->set_unaligned_access(); 895 } 896 if (mismatched) { 897 load->set_mismatched_access(); 898 } 899 if (unsafe) { 900 load->set_unsafe_access(); 901 } 902 return load; 903 } 904 905 906 907 //------------------------------hash------------------------------------------- 908 uint LoadNode::hash() const { 909 // unroll addition of interesting fields 910 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 911 } 912 913 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 914 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 915 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 916 bool is_stable_ary = FoldStableValues && 917 (tp != NULL) && (tp->isa_aryptr() != NULL) && 918 tp->isa_aryptr()->is_stable(); 919 920 return (eliminate_boxing && non_volatile) || is_stable_ary; 921 } 922 923 return false; 924 } 925 926 // Is the value loaded previously stored by an arraycopy? If so return 927 // a load node that reads from the source array so we may be able to 928 // optimize out the ArrayCopy node later. 929 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 930 Node* ld_adr = in(MemNode::Address); 931 intptr_t ld_off = 0; 932 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 933 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 934 if (ac != NULL) { 935 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 936 937 Node* mem = ac->in(TypeFunc::Memory); 938 Node* ctl = ac->in(0); 939 Node* src = ac->in(ArrayCopyNode::Src); 940 941 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 942 return NULL; 943 } 944 945 LoadNode* ld = clone()->as_Load(); 946 Node* addp = in(MemNode::Address)->clone(); 947 if (ac->as_ArrayCopy()->is_clonebasic()) { 948 assert(ld_alloc != NULL, "need an alloc"); 949 assert(addp->is_AddP(), "address must be addp"); 950 assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address"); 951 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 952 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern"); 953 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern"); 954 addp->set_req(AddPNode::Base, src->in(AddPNode::Base)); 955 addp->set_req(AddPNode::Address, src->in(AddPNode::Address)); 956 } else { 957 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 958 ac->as_ArrayCopy()->is_copyof_validated() || 959 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 960 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 961 addp->set_req(AddPNode::Base, src); 962 addp->set_req(AddPNode::Address, src); 963 964 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 965 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 966 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 967 uint shift = exact_log2(type2aelembytes(ary_elem)); 968 969 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 970 #ifdef _LP64 971 diff = phase->transform(new ConvI2LNode(diff)); 972 #endif 973 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 974 975 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 976 addp->set_req(AddPNode::Offset, offset); 977 } 978 addp = phase->transform(addp); 979 #ifdef ASSERT 980 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 981 ld->_adr_type = adr_type; 982 #endif 983 ld->set_req(MemNode::Address, addp); 984 ld->set_req(0, ctl); 985 ld->set_req(MemNode::Memory, mem); 986 // load depends on the tests that validate the arraycopy 987 ld->_control_dependency = UnknownControl; 988 return ld; 989 } 990 return NULL; 991 } 992 993 994 //---------------------------can_see_stored_value------------------------------ 995 // This routine exists to make sure this set of tests is done the same 996 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 997 // will change the graph shape in a way which makes memory alive twice at the 998 // same time (uses the Oracle model of aliasing), then some 999 // LoadXNode::Identity will fold things back to the equivalence-class model 1000 // of aliasing. 1001 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 1002 Node* ld_adr = in(MemNode::Address); 1003 intptr_t ld_off = 0; 1004 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off); 1005 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase); 1006 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 1007 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 1008 // This is more general than load from boxing objects. 1009 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1010 uint alias_idx = atp->index(); 1011 bool final = !atp->is_rewritable(); 1012 Node* result = NULL; 1013 Node* current = st; 1014 // Skip through chains of MemBarNodes checking the MergeMems for 1015 // new states for the slice of this load. Stop once any other 1016 // kind of node is encountered. Loads from final memory can skip 1017 // through any kind of MemBar but normal loads shouldn't skip 1018 // through MemBarAcquire since the could allow them to move out of 1019 // a synchronized region. 1020 while (current->is_Proj()) { 1021 int opc = current->in(0)->Opcode(); 1022 if ((final && (opc == Op_MemBarAcquire || 1023 opc == Op_MemBarAcquireLock || 1024 opc == Op_LoadFence)) || 1025 opc == Op_MemBarRelease || 1026 opc == Op_StoreFence || 1027 opc == Op_MemBarReleaseLock || 1028 opc == Op_MemBarStoreStore || 1029 opc == Op_MemBarCPUOrder) { 1030 Node* mem = current->in(0)->in(TypeFunc::Memory); 1031 if (mem->is_MergeMem()) { 1032 MergeMemNode* merge = mem->as_MergeMem(); 1033 Node* new_st = merge->memory_at(alias_idx); 1034 if (new_st == merge->base_memory()) { 1035 // Keep searching 1036 current = new_st; 1037 continue; 1038 } 1039 // Save the new memory state for the slice and fall through 1040 // to exit. 1041 result = new_st; 1042 } 1043 } 1044 break; 1045 } 1046 if (result != NULL) { 1047 st = result; 1048 } 1049 } 1050 1051 // Loop around twice in the case Load -> Initialize -> Store. 1052 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1053 for (int trip = 0; trip <= 1; trip++) { 1054 1055 if (st->is_Store()) { 1056 Node* st_adr = st->in(MemNode::Address); 1057 if (!phase->eqv(st_adr, ld_adr)) { 1058 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers). 1059 intptr_t st_off = 0; 1060 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off); 1061 if (ld_base == NULL) return NULL; 1062 if (st_base == NULL) return NULL; 1063 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL; 1064 if (ld_off != st_off) return NULL; 1065 if (ld_off == Type::OffsetBot) return NULL; 1066 // Same base, same offset. 1067 // Possible improvement for arrays: check index value instead of absolute offset. 1068 1069 // At this point we have proven something like this setup: 1070 // B = << base >> 1071 // L = LoadQ(AddP(Check/CastPP(B), #Off)) 1072 // S = StoreQ(AddP( B , #Off), V) 1073 // (Actually, we haven't yet proven the Q's are the same.) 1074 // In other words, we are loading from a casted version of 1075 // the same pointer-and-offset that we stored to. 1076 // Casted version may carry a dependency and it is respected. 1077 // Thus, we are able to replace L by V. 1078 } 1079 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1080 if (store_Opcode() != st->Opcode()) 1081 return NULL; 1082 return st->in(MemNode::ValueIn); 1083 } 1084 1085 // A load from a freshly-created object always returns zero. 1086 // (This can happen after LoadNode::Ideal resets the load's memory input 1087 // to find_captured_store, which returned InitializeNode::zero_memory.) 1088 if (st->is_Proj() && st->in(0)->is_Allocate() && 1089 (st->in(0) == ld_alloc) && 1090 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1091 // return a zero value for the load's basic type 1092 // (This is one of the few places where a generic PhaseTransform 1093 // can create new nodes. Think of it as lazily manifesting 1094 // virtually pre-existing constants.) 1095 return phase->zerocon(memory_type()); 1096 } 1097 1098 // A load from an initialization barrier can match a captured store. 1099 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1100 InitializeNode* init = st->in(0)->as_Initialize(); 1101 AllocateNode* alloc = init->allocation(); 1102 if ((alloc != NULL) && (alloc == ld_alloc)) { 1103 // examine a captured store value 1104 st = init->find_captured_store(ld_off, memory_size(), phase); 1105 if (st != NULL) { 1106 continue; // take one more trip around 1107 } 1108 } 1109 } 1110 1111 // Load boxed value from result of valueOf() call is input parameter. 1112 if (this->is_Load() && ld_adr->is_AddP() && 1113 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1114 intptr_t ignore = 0; 1115 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1116 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1117 base = bs->step_over_gc_barrier(base); 1118 if (base != NULL && base->is_Proj() && 1119 base->as_Proj()->_con == TypeFunc::Parms && 1120 base->in(0)->is_CallStaticJava() && 1121 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1122 return base->in(0)->in(TypeFunc::Parms); 1123 } 1124 } 1125 1126 break; 1127 } 1128 1129 return NULL; 1130 } 1131 1132 //----------------------is_instance_field_load_with_local_phi------------------ 1133 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1134 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1135 in(Address)->is_AddP() ) { 1136 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1137 // Only instances and boxed values. 1138 if( t_oop != NULL && 1139 (t_oop->is_ptr_to_boxed_value() || 1140 t_oop->is_known_instance_field()) && 1141 t_oop->offset() != Type::OffsetBot && 1142 t_oop->offset() != Type::OffsetTop) { 1143 return true; 1144 } 1145 } 1146 return false; 1147 } 1148 1149 //------------------------------Identity--------------------------------------- 1150 // Loads are identity if previous store is to same address 1151 Node* LoadNode::Identity(PhaseGVN* phase) { 1152 // If the previous store-maker is the right kind of Store, and the store is 1153 // to the same address, then we are equal to the value stored. 1154 Node* mem = in(Memory); 1155 Node* value = can_see_stored_value(mem, phase); 1156 if( value ) { 1157 // byte, short & char stores truncate naturally. 1158 // A load has to load the truncated value which requires 1159 // some sort of masking operation and that requires an 1160 // Ideal call instead of an Identity call. 1161 if (memory_size() < BytesPerInt) { 1162 // If the input to the store does not fit with the load's result type, 1163 // it must be truncated via an Ideal call. 1164 if (!phase->type(value)->higher_equal(phase->type(this))) 1165 return this; 1166 } 1167 // (This works even when value is a Con, but LoadNode::Value 1168 // usually runs first, producing the singleton type of the Con.) 1169 return value; 1170 } 1171 1172 // Search for an existing data phi which was generated before for the same 1173 // instance's field to avoid infinite generation of phis in a loop. 1174 Node *region = mem->in(0); 1175 if (is_instance_field_load_with_local_phi(region)) { 1176 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1177 int this_index = phase->C->get_alias_index(addr_t); 1178 int this_offset = addr_t->offset(); 1179 int this_iid = addr_t->instance_id(); 1180 if (!addr_t->is_known_instance() && 1181 addr_t->is_ptr_to_boxed_value()) { 1182 // Use _idx of address base (could be Phi node) for boxed values. 1183 intptr_t ignore = 0; 1184 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1185 if (base == NULL) { 1186 return this; 1187 } 1188 this_iid = base->_idx; 1189 } 1190 const Type* this_type = bottom_type(); 1191 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1192 Node* phi = region->fast_out(i); 1193 if (phi->is_Phi() && phi != mem && 1194 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1195 return phi; 1196 } 1197 } 1198 } 1199 1200 return this; 1201 } 1202 1203 // Construct an equivalent unsigned load. 1204 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1205 BasicType bt = T_ILLEGAL; 1206 const Type* rt = NULL; 1207 switch (Opcode()) { 1208 case Op_LoadUB: return this; 1209 case Op_LoadUS: return this; 1210 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1211 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1212 default: 1213 assert(false, "no unsigned variant: %s", Name()); 1214 return NULL; 1215 } 1216 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1217 raw_adr_type(), rt, bt, _mo, _control_dependency, 1218 is_unaligned_access(), is_mismatched_access()); 1219 } 1220 1221 // Construct an equivalent signed load. 1222 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1223 BasicType bt = T_ILLEGAL; 1224 const Type* rt = NULL; 1225 switch (Opcode()) { 1226 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1227 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1228 case Op_LoadB: // fall through 1229 case Op_LoadS: // fall through 1230 case Op_LoadI: // fall through 1231 case Op_LoadL: return this; 1232 default: 1233 assert(false, "no signed variant: %s", Name()); 1234 return NULL; 1235 } 1236 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1237 raw_adr_type(), rt, bt, _mo, _control_dependency, 1238 is_unaligned_access(), is_mismatched_access()); 1239 } 1240 1241 // We're loading from an object which has autobox behaviour. 1242 // If this object is result of a valueOf call we'll have a phi 1243 // merging a newly allocated object and a load from the cache. 1244 // We want to replace this load with the original incoming 1245 // argument to the valueOf call. 1246 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1247 assert(phase->C->eliminate_boxing(), "sanity"); 1248 intptr_t ignore = 0; 1249 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1250 if ((base == NULL) || base->is_Phi()) { 1251 // Push the loads from the phi that comes from valueOf up 1252 // through it to allow elimination of the loads and the recovery 1253 // of the original value. It is done in split_through_phi(). 1254 return NULL; 1255 } else if (base->is_Load() || 1256 (base->is_DecodeN() && base->in(1)->is_Load())) { 1257 // Eliminate the load of boxed value for integer types from the cache 1258 // array by deriving the value from the index into the array. 1259 // Capture the offset of the load and then reverse the computation. 1260 1261 // Get LoadN node which loads a boxing object from 'cache' array. 1262 if (base->is_DecodeN()) { 1263 base = base->in(1); 1264 } 1265 if (!base->in(Address)->is_AddP()) { 1266 return NULL; // Complex address 1267 } 1268 AddPNode* address = base->in(Address)->as_AddP(); 1269 Node* cache_base = address->in(AddPNode::Base); 1270 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1271 // Get ConP node which is static 'cache' field. 1272 cache_base = cache_base->in(1); 1273 } 1274 if ((cache_base != NULL) && cache_base->is_Con()) { 1275 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1276 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1277 Node* elements[4]; 1278 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1279 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1280 if (count > 0 && elements[0]->is_Con() && 1281 (count == 1 || 1282 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1283 elements[1]->in(2) == phase->intcon(shift)))) { 1284 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1285 // Fetch the box object cache[0] at the base of the array and get its value 1286 ciInstance* box = array->obj_at(0)->as_instance(); 1287 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1288 assert(ik->is_box_klass(), "sanity"); 1289 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1290 if (ik->nof_nonstatic_fields() == 1) { 1291 // This should be true nonstatic_field_at requires calling 1292 // nof_nonstatic_fields so check it anyway 1293 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1294 BasicType bt = c.basic_type(); 1295 // Only integer types have boxing cache. 1296 assert(bt == T_BOOLEAN || bt == T_CHAR || 1297 bt == T_BYTE || bt == T_SHORT || 1298 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1299 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1300 if (cache_low != (int)cache_low) { 1301 return NULL; // should not happen since cache is array indexed by value 1302 } 1303 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1304 if (offset != (int)offset) { 1305 return NULL; // should not happen since cache is array indexed by value 1306 } 1307 // Add up all the offsets making of the address of the load 1308 Node* result = elements[0]; 1309 for (int i = 1; i < count; i++) { 1310 result = phase->transform(new AddXNode(result, elements[i])); 1311 } 1312 // Remove the constant offset from the address and then 1313 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1314 // remove the scaling of the offset to recover the original index. 1315 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1316 // Peel the shift off directly but wrap it in a dummy node 1317 // since Ideal can't return existing nodes 1318 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1319 } else if (result->is_Add() && result->in(2)->is_Con() && 1320 result->in(1)->Opcode() == Op_LShiftX && 1321 result->in(1)->in(2) == phase->intcon(shift)) { 1322 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1323 // but for boxing cache access we know that X<<Z will not overflow 1324 // (there is range check) so we do this optimizatrion by hand here. 1325 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1326 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1327 } else { 1328 result = new RShiftXNode(result, phase->intcon(shift)); 1329 } 1330 #ifdef _LP64 1331 if (bt != T_LONG) { 1332 result = new ConvL2INode(phase->transform(result)); 1333 } 1334 #else 1335 if (bt == T_LONG) { 1336 result = new ConvI2LNode(phase->transform(result)); 1337 } 1338 #endif 1339 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1340 // Need to preserve unboxing load type if it is unsigned. 1341 switch(this->Opcode()) { 1342 case Op_LoadUB: 1343 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1344 break; 1345 case Op_LoadUS: 1346 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1347 break; 1348 } 1349 return result; 1350 } 1351 } 1352 } 1353 } 1354 } 1355 return NULL; 1356 } 1357 1358 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1359 Node* region = phi->in(0); 1360 if (region == NULL) { 1361 return false; // Wait stable graph 1362 } 1363 uint cnt = phi->req(); 1364 for (uint i = 1; i < cnt; i++) { 1365 Node* rc = region->in(i); 1366 if (rc == NULL || phase->type(rc) == Type::TOP) 1367 return false; // Wait stable graph 1368 Node* in = phi->in(i); 1369 if (in == NULL || phase->type(in) == Type::TOP) 1370 return false; // Wait stable graph 1371 } 1372 return true; 1373 } 1374 //------------------------------split_through_phi------------------------------ 1375 // Split instance or boxed field load through Phi. 1376 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1377 Node* mem = in(Memory); 1378 Node* address = in(Address); 1379 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1380 1381 assert((t_oop != NULL) && 1382 (t_oop->is_known_instance_field() || 1383 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1384 1385 Compile* C = phase->C; 1386 intptr_t ignore = 0; 1387 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1388 bool base_is_phi = (base != NULL) && base->is_Phi(); 1389 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1390 (base != NULL) && (base == address->in(AddPNode::Base)) && 1391 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1392 1393 if (!((mem->is_Phi() || base_is_phi) && 1394 (load_boxed_values || t_oop->is_known_instance_field()))) { 1395 return NULL; // memory is not Phi 1396 } 1397 1398 if (mem->is_Phi()) { 1399 if (!stable_phi(mem->as_Phi(), phase)) { 1400 return NULL; // Wait stable graph 1401 } 1402 uint cnt = mem->req(); 1403 // Check for loop invariant memory. 1404 if (cnt == 3) { 1405 for (uint i = 1; i < cnt; i++) { 1406 Node* in = mem->in(i); 1407 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1408 if (m == mem) { 1409 if (i == 1) { 1410 // if the first edge was a loop, check second edge too. 1411 // If both are replaceable - we are in an infinite loop 1412 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1413 if (n == mem) { 1414 break; 1415 } 1416 } 1417 set_req(Memory, mem->in(cnt - i)); 1418 return this; // made change 1419 } 1420 } 1421 } 1422 } 1423 if (base_is_phi) { 1424 if (!stable_phi(base->as_Phi(), phase)) { 1425 return NULL; // Wait stable graph 1426 } 1427 uint cnt = base->req(); 1428 // Check for loop invariant memory. 1429 if (cnt == 3) { 1430 for (uint i = 1; i < cnt; i++) { 1431 if (base->in(i) == base) { 1432 return NULL; // Wait stable graph 1433 } 1434 } 1435 } 1436 } 1437 1438 // Split through Phi (see original code in loopopts.cpp). 1439 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1440 1441 // Do nothing here if Identity will find a value 1442 // (to avoid infinite chain of value phis generation). 1443 if (!phase->eqv(this, phase->apply_identity(this))) 1444 return NULL; 1445 1446 // Select Region to split through. 1447 Node* region; 1448 if (!base_is_phi) { 1449 assert(mem->is_Phi(), "sanity"); 1450 region = mem->in(0); 1451 // Skip if the region dominates some control edge of the address. 1452 if (!MemNode::all_controls_dominate(address, region)) 1453 return NULL; 1454 } else if (!mem->is_Phi()) { 1455 assert(base_is_phi, "sanity"); 1456 region = base->in(0); 1457 // Skip if the region dominates some control edge of the memory. 1458 if (!MemNode::all_controls_dominate(mem, region)) 1459 return NULL; 1460 } else if (base->in(0) != mem->in(0)) { 1461 assert(base_is_phi && mem->is_Phi(), "sanity"); 1462 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1463 region = base->in(0); 1464 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1465 region = mem->in(0); 1466 } else { 1467 return NULL; // complex graph 1468 } 1469 } else { 1470 assert(base->in(0) == mem->in(0), "sanity"); 1471 region = mem->in(0); 1472 } 1473 1474 const Type* this_type = this->bottom_type(); 1475 int this_index = C->get_alias_index(t_oop); 1476 int this_offset = t_oop->offset(); 1477 int this_iid = t_oop->instance_id(); 1478 if (!t_oop->is_known_instance() && load_boxed_values) { 1479 // Use _idx of address base for boxed values. 1480 this_iid = base->_idx; 1481 } 1482 PhaseIterGVN* igvn = phase->is_IterGVN(); 1483 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1484 for (uint i = 1; i < region->req(); i++) { 1485 Node* x; 1486 Node* the_clone = NULL; 1487 if (region->in(i) == C->top()) { 1488 x = C->top(); // Dead path? Use a dead data op 1489 } else { 1490 x = this->clone(); // Else clone up the data op 1491 the_clone = x; // Remember for possible deletion. 1492 // Alter data node to use pre-phi inputs 1493 if (this->in(0) == region) { 1494 x->set_req(0, region->in(i)); 1495 } else { 1496 x->set_req(0, NULL); 1497 } 1498 if (mem->is_Phi() && (mem->in(0) == region)) { 1499 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1500 } 1501 if (address->is_Phi() && address->in(0) == region) { 1502 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1503 } 1504 if (base_is_phi && (base->in(0) == region)) { 1505 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1506 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1507 x->set_req(Address, adr_x); 1508 } 1509 } 1510 // Check for a 'win' on some paths 1511 const Type *t = x->Value(igvn); 1512 1513 bool singleton = t->singleton(); 1514 1515 // See comments in PhaseIdealLoop::split_thru_phi(). 1516 if (singleton && t == Type::TOP) { 1517 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1518 } 1519 1520 if (singleton) { 1521 x = igvn->makecon(t); 1522 } else { 1523 // We now call Identity to try to simplify the cloned node. 1524 // Note that some Identity methods call phase->type(this). 1525 // Make sure that the type array is big enough for 1526 // our new node, even though we may throw the node away. 1527 // (This tweaking with igvn only works because x is a new node.) 1528 igvn->set_type(x, t); 1529 // If x is a TypeNode, capture any more-precise type permanently into Node 1530 // otherwise it will be not updated during igvn->transform since 1531 // igvn->type(x) is set to x->Value() already. 1532 x->raise_bottom_type(t); 1533 Node *y = igvn->apply_identity(x); 1534 if (y != x) { 1535 x = y; 1536 } else { 1537 y = igvn->hash_find_insert(x); 1538 if (y) { 1539 x = y; 1540 } else { 1541 // Else x is a new node we are keeping 1542 // We do not need register_new_node_with_optimizer 1543 // because set_type has already been called. 1544 igvn->_worklist.push(x); 1545 } 1546 } 1547 } 1548 if (x != the_clone && the_clone != NULL) { 1549 igvn->remove_dead_node(the_clone); 1550 } 1551 phi->set_req(i, x); 1552 } 1553 // Record Phi 1554 igvn->register_new_node_with_optimizer(phi); 1555 return phi; 1556 } 1557 1558 //------------------------------Ideal------------------------------------------ 1559 // If the load is from Field memory and the pointer is non-null, it might be possible to 1560 // zero out the control input. 1561 // If the offset is constant and the base is an object allocation, 1562 // try to hook me up to the exact initializing store. 1563 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1564 Node* p = MemNode::Ideal_common(phase, can_reshape); 1565 if (p) return (p == NodeSentinel) ? NULL : p; 1566 1567 Node* ctrl = in(MemNode::Control); 1568 Node* address = in(MemNode::Address); 1569 bool progress = false; 1570 1571 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1572 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1573 1574 // Skip up past a SafePoint control. Cannot do this for Stores because 1575 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1576 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1577 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1578 !addr_mark && 1579 (depends_only_on_test() || has_unknown_control_dependency())) { 1580 ctrl = ctrl->in(0); 1581 set_req(MemNode::Control,ctrl); 1582 progress = true; 1583 } 1584 1585 intptr_t ignore = 0; 1586 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1587 if (base != NULL 1588 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1589 // Check for useless control edge in some common special cases 1590 if (in(MemNode::Control) != NULL 1591 && can_remove_control() 1592 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1593 && all_controls_dominate(base, phase->C->start())) { 1594 // A method-invariant, non-null address (constant or 'this' argument). 1595 set_req(MemNode::Control, NULL); 1596 progress = true; 1597 } 1598 } 1599 1600 Node* mem = in(MemNode::Memory); 1601 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1602 1603 if (can_reshape && (addr_t != NULL)) { 1604 // try to optimize our memory input 1605 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1606 if (opt_mem != mem) { 1607 set_req(MemNode::Memory, opt_mem); 1608 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1609 return this; 1610 } 1611 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1612 if ((t_oop != NULL) && 1613 (t_oop->is_known_instance_field() || 1614 t_oop->is_ptr_to_boxed_value())) { 1615 PhaseIterGVN *igvn = phase->is_IterGVN(); 1616 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1617 // Delay this transformation until memory Phi is processed. 1618 phase->is_IterGVN()->_worklist.push(this); 1619 return NULL; 1620 } 1621 // Split instance field load through Phi. 1622 Node* result = split_through_phi(phase); 1623 if (result != NULL) return result; 1624 1625 if (t_oop->is_ptr_to_boxed_value()) { 1626 Node* result = eliminate_autobox(phase); 1627 if (result != NULL) return result; 1628 } 1629 } 1630 } 1631 1632 // Is there a dominating load that loads the same value? Leave 1633 // anything that is not a load of a field/array element (like 1634 // barriers etc.) alone 1635 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) { 1636 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1637 Node *use = mem->fast_out(i); 1638 if (use != this && 1639 use->Opcode() == Opcode() && 1640 use->in(0) != NULL && 1641 use->in(0) != in(0) && 1642 use->in(Address) == in(Address)) { 1643 Node* ctl = in(0); 1644 for (int i = 0; i < 10 && ctl != NULL; i++) { 1645 ctl = IfNode::up_one_dom(ctl); 1646 if (ctl == use->in(0)) { 1647 set_req(0, use->in(0)); 1648 return this; 1649 } 1650 } 1651 } 1652 } 1653 } 1654 1655 // Check for prior store with a different base or offset; make Load 1656 // independent. Skip through any number of them. Bail out if the stores 1657 // are in an endless dead cycle and report no progress. This is a key 1658 // transform for Reflection. However, if after skipping through the Stores 1659 // we can't then fold up against a prior store do NOT do the transform as 1660 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1661 // array memory alive twice: once for the hoisted Load and again after the 1662 // bypassed Store. This situation only works if EVERYBODY who does 1663 // anti-dependence work knows how to bypass. I.e. we need all 1664 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1665 // the alias index stuff. So instead, peek through Stores and IFF we can 1666 // fold up, do so. 1667 Node* prev_mem = find_previous_store(phase); 1668 if (prev_mem != NULL) { 1669 Node* value = can_see_arraycopy_value(prev_mem, phase); 1670 if (value != NULL) { 1671 return value; 1672 } 1673 } 1674 // Steps (a), (b): Walk past independent stores to find an exact match. 1675 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1676 // (c) See if we can fold up on the spot, but don't fold up here. 1677 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1678 // just return a prior value, which is done by Identity calls. 1679 if (can_see_stored_value(prev_mem, phase)) { 1680 // Make ready for step (d): 1681 set_req(MemNode::Memory, prev_mem); 1682 return this; 1683 } 1684 } 1685 1686 return progress ? this : NULL; 1687 } 1688 1689 // Helper to recognize certain Klass fields which are invariant across 1690 // some group of array types (e.g., int[] or all T[] where T < Object). 1691 const Type* 1692 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1693 ciKlass* klass) const { 1694 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1695 // The field is Klass::_modifier_flags. Return its (constant) value. 1696 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1697 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1698 return TypeInt::make(klass->modifier_flags()); 1699 } 1700 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1701 // The field is Klass::_access_flags. Return its (constant) value. 1702 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1703 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1704 return TypeInt::make(klass->access_flags()); 1705 } 1706 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1707 // The field is Klass::_layout_helper. Return its constant value if known. 1708 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1709 return TypeInt::make(klass->layout_helper()); 1710 } 1711 1712 // No match. 1713 return NULL; 1714 } 1715 1716 //------------------------------Value----------------------------------------- 1717 const Type* LoadNode::Value(PhaseGVN* phase) const { 1718 // Either input is TOP ==> the result is TOP 1719 Node* mem = in(MemNode::Memory); 1720 const Type *t1 = phase->type(mem); 1721 if (t1 == Type::TOP) return Type::TOP; 1722 Node* adr = in(MemNode::Address); 1723 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1724 if (tp == NULL || tp->empty()) return Type::TOP; 1725 int off = tp->offset(); 1726 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1727 Compile* C = phase->C; 1728 1729 // Try to guess loaded type from pointer type 1730 if (tp->isa_aryptr()) { 1731 const TypeAryPtr* ary = tp->is_aryptr(); 1732 const Type* t = ary->elem(); 1733 1734 // Determine whether the reference is beyond the header or not, by comparing 1735 // the offset against the offset of the start of the array's data. 1736 // Different array types begin at slightly different offsets (12 vs. 16). 1737 // We choose T_BYTE as an example base type that is least restrictive 1738 // as to alignment, which will therefore produce the smallest 1739 // possible base offset. 1740 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1741 const bool off_beyond_header = (off >= min_base_off); 1742 1743 // Try to constant-fold a stable array element. 1744 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1745 // Make sure the reference is not into the header and the offset is constant 1746 ciObject* aobj = ary->const_oop(); 1747 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1748 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1749 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1750 stable_dimension, 1751 memory_type(), is_unsigned()); 1752 if (con_type != NULL) { 1753 return con_type; 1754 } 1755 } 1756 } 1757 1758 // Don't do this for integer types. There is only potential profit if 1759 // the element type t is lower than _type; that is, for int types, if _type is 1760 // more restrictive than t. This only happens here if one is short and the other 1761 // char (both 16 bits), and in those cases we've made an intentional decision 1762 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1763 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1764 // 1765 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1766 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1767 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1768 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1769 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1770 // In fact, that could have been the original type of p1, and p1 could have 1771 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1772 // expression (LShiftL quux 3) independently optimized to the constant 8. 1773 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1774 && (_type->isa_vect() == NULL) 1775 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1776 // t might actually be lower than _type, if _type is a unique 1777 // concrete subclass of abstract class t. 1778 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header? 1779 const Type* jt = t->join_speculative(_type); 1780 // In any case, do not allow the join, per se, to empty out the type. 1781 if (jt->empty() && !t->empty()) { 1782 // This can happen if a interface-typed array narrows to a class type. 1783 jt = _type; 1784 } 1785 #ifdef ASSERT 1786 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1787 // The pointers in the autobox arrays are always non-null 1788 Node* base = adr->in(AddPNode::Base); 1789 if ((base != NULL) && base->is_DecodeN()) { 1790 // Get LoadN node which loads IntegerCache.cache field 1791 base = base->in(1); 1792 } 1793 if ((base != NULL) && base->is_Con()) { 1794 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1795 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1796 // It could be narrow oop 1797 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1798 } 1799 } 1800 } 1801 #endif 1802 return jt; 1803 } 1804 } 1805 } else if (tp->base() == Type::InstPtr) { 1806 assert( off != Type::OffsetBot || 1807 // arrays can be cast to Objects 1808 tp->is_oopptr()->klass()->is_java_lang_Object() || 1809 // unsafe field access may not have a constant offset 1810 C->has_unsafe_access(), 1811 "Field accesses must be precise" ); 1812 // For oop loads, we expect the _type to be precise. 1813 1814 // Optimize loads from constant fields. 1815 const TypeInstPtr* tinst = tp->is_instptr(); 1816 ciObject* const_oop = tinst->const_oop(); 1817 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1818 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type()); 1819 if (con_type != NULL) { 1820 return con_type; 1821 } 1822 } 1823 } else if (tp->base() == Type::KlassPtr) { 1824 assert( off != Type::OffsetBot || 1825 // arrays can be cast to Objects 1826 tp->is_klassptr()->klass()->is_java_lang_Object() || 1827 // also allow array-loading from the primary supertype 1828 // array during subtype checks 1829 Opcode() == Op_LoadKlass, 1830 "Field accesses must be precise" ); 1831 // For klass/static loads, we expect the _type to be precise 1832 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) { 1833 /* With mirrors being an indirect in the Klass* 1834 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 1835 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 1836 * 1837 * So check the type and klass of the node before the LoadP. 1838 */ 1839 Node* adr2 = adr->in(MemNode::Address); 1840 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1841 if (tkls != NULL && !StressReflectiveCode) { 1842 ciKlass* klass = tkls->klass(); 1843 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1844 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1845 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1846 return TypeInstPtr::make(klass->java_mirror()); 1847 } 1848 } 1849 } 1850 1851 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1852 if (tkls != NULL && !StressReflectiveCode) { 1853 ciKlass* klass = tkls->klass(); 1854 if (klass->is_loaded() && tkls->klass_is_exact()) { 1855 // We are loading a field from a Klass metaobject whose identity 1856 // is known at compile time (the type is "exact" or "precise"). 1857 // Check for fields we know are maintained as constants by the VM. 1858 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1859 // The field is Klass::_super_check_offset. Return its (constant) value. 1860 // (Folds up type checking code.) 1861 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1862 return TypeInt::make(klass->super_check_offset()); 1863 } 1864 // Compute index into primary_supers array 1865 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1866 // Check for overflowing; use unsigned compare to handle the negative case. 1867 if( depth < ciKlass::primary_super_limit() ) { 1868 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1869 // (Folds up type checking code.) 1870 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1871 ciKlass *ss = klass->super_of_depth(depth); 1872 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1873 } 1874 const Type* aift = load_array_final_field(tkls, klass); 1875 if (aift != NULL) return aift; 1876 } 1877 1878 // We can still check if we are loading from the primary_supers array at a 1879 // shallow enough depth. Even though the klass is not exact, entries less 1880 // than or equal to its super depth are correct. 1881 if (klass->is_loaded() ) { 1882 ciType *inner = klass; 1883 while( inner->is_obj_array_klass() ) 1884 inner = inner->as_obj_array_klass()->base_element_type(); 1885 if( inner->is_instance_klass() && 1886 !inner->as_instance_klass()->flags().is_interface() ) { 1887 // Compute index into primary_supers array 1888 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1889 // Check for overflowing; use unsigned compare to handle the negative case. 1890 if( depth < ciKlass::primary_super_limit() && 1891 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1892 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1893 // (Folds up type checking code.) 1894 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1895 ciKlass *ss = klass->super_of_depth(depth); 1896 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1897 } 1898 } 1899 } 1900 1901 // If the type is enough to determine that the thing is not an array, 1902 // we can give the layout_helper a positive interval type. 1903 // This will help short-circuit some reflective code. 1904 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1905 && !klass->is_array_klass() // not directly typed as an array 1906 && !klass->is_interface() // specifically not Serializable & Cloneable 1907 && !klass->is_java_lang_Object() // not the supertype of all T[] 1908 ) { 1909 // Note: When interfaces are reliable, we can narrow the interface 1910 // test to (klass != Serializable && klass != Cloneable). 1911 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1912 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1913 // The key property of this type is that it folds up tests 1914 // for array-ness, since it proves that the layout_helper is positive. 1915 // Thus, a generic value like the basic object layout helper works fine. 1916 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1917 } 1918 } 1919 1920 // If we are loading from a freshly-allocated object, produce a zero, 1921 // if the load is provably beyond the header of the object. 1922 // (Also allow a variable load from a fresh array to produce zero.) 1923 const TypeOopPtr *tinst = tp->isa_oopptr(); 1924 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1925 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1926 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1927 Node* value = can_see_stored_value(mem,phase); 1928 if (value != NULL && value->is_Con()) { 1929 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1930 return value->bottom_type(); 1931 } 1932 } 1933 1934 if (is_instance) { 1935 // If we have an instance type and our memory input is the 1936 // programs's initial memory state, there is no matching store, 1937 // so just return a zero of the appropriate type 1938 Node *mem = in(MemNode::Memory); 1939 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1940 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1941 return Type::get_zero_type(_type->basic_type()); 1942 } 1943 } 1944 return _type; 1945 } 1946 1947 //------------------------------match_edge------------------------------------- 1948 // Do we Match on this edge index or not? Match only the address. 1949 uint LoadNode::match_edge(uint idx) const { 1950 return idx == MemNode::Address; 1951 } 1952 1953 //--------------------------LoadBNode::Ideal-------------------------------------- 1954 // 1955 // If the previous store is to the same address as this load, 1956 // and the value stored was larger than a byte, replace this load 1957 // with the value stored truncated to a byte. If no truncation is 1958 // needed, the replacement is done in LoadNode::Identity(). 1959 // 1960 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1961 Node* mem = in(MemNode::Memory); 1962 Node* value = can_see_stored_value(mem,phase); 1963 if( value && !phase->type(value)->higher_equal( _type ) ) { 1964 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1965 return new RShiftINode(result, phase->intcon(24)); 1966 } 1967 // Identity call will handle the case where truncation is not needed. 1968 return LoadNode::Ideal(phase, can_reshape); 1969 } 1970 1971 const Type* LoadBNode::Value(PhaseGVN* phase) const { 1972 Node* mem = in(MemNode::Memory); 1973 Node* value = can_see_stored_value(mem,phase); 1974 if (value != NULL && value->is_Con() && 1975 !value->bottom_type()->higher_equal(_type)) { 1976 // If the input to the store does not fit with the load's result type, 1977 // it must be truncated. We can't delay until Ideal call since 1978 // a singleton Value is needed for split_thru_phi optimization. 1979 int con = value->get_int(); 1980 return TypeInt::make((con << 24) >> 24); 1981 } 1982 return LoadNode::Value(phase); 1983 } 1984 1985 //--------------------------LoadUBNode::Ideal------------------------------------- 1986 // 1987 // If the previous store is to the same address as this load, 1988 // and the value stored was larger than a byte, replace this load 1989 // with the value stored truncated to a byte. If no truncation is 1990 // needed, the replacement is done in LoadNode::Identity(). 1991 // 1992 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1993 Node* mem = in(MemNode::Memory); 1994 Node* value = can_see_stored_value(mem, phase); 1995 if (value && !phase->type(value)->higher_equal(_type)) 1996 return new AndINode(value, phase->intcon(0xFF)); 1997 // Identity call will handle the case where truncation is not needed. 1998 return LoadNode::Ideal(phase, can_reshape); 1999 } 2000 2001 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2002 Node* mem = in(MemNode::Memory); 2003 Node* value = can_see_stored_value(mem,phase); 2004 if (value != NULL && value->is_Con() && 2005 !value->bottom_type()->higher_equal(_type)) { 2006 // If the input to the store does not fit with the load's result type, 2007 // it must be truncated. We can't delay until Ideal call since 2008 // a singleton Value is needed for split_thru_phi optimization. 2009 int con = value->get_int(); 2010 return TypeInt::make(con & 0xFF); 2011 } 2012 return LoadNode::Value(phase); 2013 } 2014 2015 //--------------------------LoadUSNode::Ideal------------------------------------- 2016 // 2017 // If the previous store is to the same address as this load, 2018 // and the value stored was larger than a char, replace this load 2019 // with the value stored truncated to a char. If no truncation is 2020 // needed, the replacement is done in LoadNode::Identity(). 2021 // 2022 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2023 Node* mem = in(MemNode::Memory); 2024 Node* value = can_see_stored_value(mem,phase); 2025 if( value && !phase->type(value)->higher_equal( _type ) ) 2026 return new AndINode(value,phase->intcon(0xFFFF)); 2027 // Identity call will handle the case where truncation is not needed. 2028 return LoadNode::Ideal(phase, can_reshape); 2029 } 2030 2031 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2032 Node* mem = in(MemNode::Memory); 2033 Node* value = can_see_stored_value(mem,phase); 2034 if (value != NULL && value->is_Con() && 2035 !value->bottom_type()->higher_equal(_type)) { 2036 // If the input to the store does not fit with the load's result type, 2037 // it must be truncated. We can't delay until Ideal call since 2038 // a singleton Value is needed for split_thru_phi optimization. 2039 int con = value->get_int(); 2040 return TypeInt::make(con & 0xFFFF); 2041 } 2042 return LoadNode::Value(phase); 2043 } 2044 2045 //--------------------------LoadSNode::Ideal-------------------------------------- 2046 // 2047 // If the previous store is to the same address as this load, 2048 // and the value stored was larger than a short, replace this load 2049 // with the value stored truncated to a short. If no truncation is 2050 // needed, the replacement is done in LoadNode::Identity(). 2051 // 2052 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2053 Node* mem = in(MemNode::Memory); 2054 Node* value = can_see_stored_value(mem,phase); 2055 if( value && !phase->type(value)->higher_equal( _type ) ) { 2056 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 2057 return new RShiftINode(result, phase->intcon(16)); 2058 } 2059 // Identity call will handle the case where truncation is not needed. 2060 return LoadNode::Ideal(phase, can_reshape); 2061 } 2062 2063 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2064 Node* mem = in(MemNode::Memory); 2065 Node* value = can_see_stored_value(mem,phase); 2066 if (value != NULL && value->is_Con() && 2067 !value->bottom_type()->higher_equal(_type)) { 2068 // If the input to the store does not fit with the load's result type, 2069 // it must be truncated. We can't delay until Ideal call since 2070 // a singleton Value is needed for split_thru_phi optimization. 2071 int con = value->get_int(); 2072 return TypeInt::make((con << 16) >> 16); 2073 } 2074 return LoadNode::Value(phase); 2075 } 2076 2077 //============================================================================= 2078 //----------------------------LoadKlassNode::make------------------------------ 2079 // Polymorphic factory method: 2080 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2081 // sanity check the alias category against the created node type 2082 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2083 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2084 #ifdef _LP64 2085 if (adr_type->is_ptr_to_narrowklass()) { 2086 assert(UseCompressedClassPointers, "no compressed klasses"); 2087 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2088 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2089 } 2090 #endif 2091 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2092 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2093 } 2094 2095 //------------------------------Value------------------------------------------ 2096 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2097 return klass_value_common(phase); 2098 } 2099 2100 // In most cases, LoadKlassNode does not have the control input set. If the control 2101 // input is set, it must not be removed (by LoadNode::Ideal()). 2102 bool LoadKlassNode::can_remove_control() const { 2103 return false; 2104 } 2105 2106 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2107 // Either input is TOP ==> the result is TOP 2108 const Type *t1 = phase->type( in(MemNode::Memory) ); 2109 if (t1 == Type::TOP) return Type::TOP; 2110 Node *adr = in(MemNode::Address); 2111 const Type *t2 = phase->type( adr ); 2112 if (t2 == Type::TOP) return Type::TOP; 2113 const TypePtr *tp = t2->is_ptr(); 2114 if (TypePtr::above_centerline(tp->ptr()) || 2115 tp->ptr() == TypePtr::Null) return Type::TOP; 2116 2117 // Return a more precise klass, if possible 2118 const TypeInstPtr *tinst = tp->isa_instptr(); 2119 if (tinst != NULL) { 2120 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2121 int offset = tinst->offset(); 2122 if (ik == phase->C->env()->Class_klass() 2123 && (offset == java_lang_Class::klass_offset_in_bytes() || 2124 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2125 // We are loading a special hidden field from a Class mirror object, 2126 // the field which points to the VM's Klass metaobject. 2127 ciType* t = tinst->java_mirror_type(); 2128 // java_mirror_type returns non-null for compile-time Class constants. 2129 if (t != NULL) { 2130 // constant oop => constant klass 2131 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2132 if (t->is_void()) { 2133 // We cannot create a void array. Since void is a primitive type return null 2134 // klass. Users of this result need to do a null check on the returned klass. 2135 return TypePtr::NULL_PTR; 2136 } 2137 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2138 } 2139 if (!t->is_klass()) { 2140 // a primitive Class (e.g., int.class) has NULL for a klass field 2141 return TypePtr::NULL_PTR; 2142 } 2143 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2144 return TypeKlassPtr::make(t->as_klass()); 2145 } 2146 // non-constant mirror, so we can't tell what's going on 2147 } 2148 if( !ik->is_loaded() ) 2149 return _type; // Bail out if not loaded 2150 if (offset == oopDesc::klass_offset_in_bytes()) { 2151 if (tinst->klass_is_exact()) { 2152 return TypeKlassPtr::make(ik); 2153 } 2154 // See if we can become precise: no subklasses and no interface 2155 // (Note: We need to support verified interfaces.) 2156 if (!ik->is_interface() && !ik->has_subklass()) { 2157 //assert(!UseExactTypes, "this code should be useless with exact types"); 2158 // Add a dependence; if any subclass added we need to recompile 2159 if (!ik->is_final()) { 2160 // %%% should use stronger assert_unique_concrete_subtype instead 2161 phase->C->dependencies()->assert_leaf_type(ik); 2162 } 2163 // Return precise klass 2164 return TypeKlassPtr::make(ik); 2165 } 2166 2167 // Return root of possible klass 2168 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2169 } 2170 } 2171 2172 // Check for loading klass from an array 2173 const TypeAryPtr *tary = tp->isa_aryptr(); 2174 if( tary != NULL ) { 2175 ciKlass *tary_klass = tary->klass(); 2176 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2177 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2178 if (tary->klass_is_exact()) { 2179 return TypeKlassPtr::make(tary_klass); 2180 } 2181 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2182 // If the klass is an object array, we defer the question to the 2183 // array component klass. 2184 if( ak->is_obj_array_klass() ) { 2185 assert( ak->is_loaded(), "" ); 2186 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2187 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2188 ciInstanceKlass* ik = base_k->as_instance_klass(); 2189 // See if we can become precise: no subklasses and no interface 2190 if (!ik->is_interface() && !ik->has_subklass()) { 2191 //assert(!UseExactTypes, "this code should be useless with exact types"); 2192 // Add a dependence; if any subclass added we need to recompile 2193 if (!ik->is_final()) { 2194 phase->C->dependencies()->assert_leaf_type(ik); 2195 } 2196 // Return precise array klass 2197 return TypeKlassPtr::make(ak); 2198 } 2199 } 2200 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2201 } else { // Found a type-array? 2202 //assert(!UseExactTypes, "this code should be useless with exact types"); 2203 assert( ak->is_type_array_klass(), "" ); 2204 return TypeKlassPtr::make(ak); // These are always precise 2205 } 2206 } 2207 } 2208 2209 // Check for loading klass from an array klass 2210 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2211 if (tkls != NULL && !StressReflectiveCode) { 2212 ciKlass* klass = tkls->klass(); 2213 if( !klass->is_loaded() ) 2214 return _type; // Bail out if not loaded 2215 if( klass->is_obj_array_klass() && 2216 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2217 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2218 // // Always returning precise element type is incorrect, 2219 // // e.g., element type could be object and array may contain strings 2220 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2221 2222 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2223 // according to the element type's subclassing. 2224 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2225 } 2226 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2227 tkls->offset() == in_bytes(Klass::super_offset())) { 2228 ciKlass* sup = klass->as_instance_klass()->super(); 2229 // The field is Klass::_super. Return its (constant) value. 2230 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2231 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2232 } 2233 } 2234 2235 // Bailout case 2236 return LoadNode::Value(phase); 2237 } 2238 2239 //------------------------------Identity--------------------------------------- 2240 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2241 // Also feed through the klass in Allocate(...klass...)._klass. 2242 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2243 return klass_identity_common(phase); 2244 } 2245 2246 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2247 Node* x = LoadNode::Identity(phase); 2248 if (x != this) return x; 2249 2250 // Take apart the address into an oop and and offset. 2251 // Return 'this' if we cannot. 2252 Node* adr = in(MemNode::Address); 2253 intptr_t offset = 0; 2254 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2255 if (base == NULL) return this; 2256 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2257 if (toop == NULL) return this; 2258 2259 // Step over potential GC barrier for OopHandle resolve 2260 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2261 if (bs->is_gc_barrier_node(base)) { 2262 base = bs->step_over_gc_barrier(base); 2263 } 2264 2265 // We can fetch the klass directly through an AllocateNode. 2266 // This works even if the klass is not constant (clone or newArray). 2267 if (offset == oopDesc::klass_offset_in_bytes()) { 2268 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2269 if (allocated_klass != NULL) { 2270 return allocated_klass; 2271 } 2272 } 2273 2274 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2275 // See inline_native_Class_query for occurrences of these patterns. 2276 // Java Example: x.getClass().isAssignableFrom(y) 2277 // 2278 // This improves reflective code, often making the Class 2279 // mirror go completely dead. (Current exception: Class 2280 // mirrors may appear in debug info, but we could clean them out by 2281 // introducing a new debug info operator for Klass.java_mirror). 2282 2283 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2284 && offset == java_lang_Class::klass_offset_in_bytes()) { 2285 if (base->is_Load()) { 2286 Node* base2 = base->in(MemNode::Address); 2287 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2288 Node* adr2 = base2->in(MemNode::Address); 2289 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2290 if (tkls != NULL && !tkls->empty() 2291 && (tkls->klass()->is_instance_klass() || 2292 tkls->klass()->is_array_klass()) 2293 && adr2->is_AddP() 2294 ) { 2295 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2296 if (tkls->offset() == mirror_field) { 2297 return adr2->in(AddPNode::Base); 2298 } 2299 } 2300 } 2301 } 2302 } 2303 2304 return this; 2305 } 2306 2307 2308 //------------------------------Value------------------------------------------ 2309 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2310 const Type *t = klass_value_common(phase); 2311 if (t == Type::TOP) 2312 return t; 2313 2314 return t->make_narrowklass(); 2315 } 2316 2317 //------------------------------Identity--------------------------------------- 2318 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2319 // Also feed through the klass in Allocate(...klass...)._klass. 2320 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2321 Node *x = klass_identity_common(phase); 2322 2323 const Type *t = phase->type( x ); 2324 if( t == Type::TOP ) return x; 2325 if( t->isa_narrowklass()) return x; 2326 assert (!t->isa_narrowoop(), "no narrow oop here"); 2327 2328 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2329 } 2330 2331 //------------------------------Value----------------------------------------- 2332 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2333 // Either input is TOP ==> the result is TOP 2334 const Type *t1 = phase->type( in(MemNode::Memory) ); 2335 if( t1 == Type::TOP ) return Type::TOP; 2336 Node *adr = in(MemNode::Address); 2337 const Type *t2 = phase->type( adr ); 2338 if( t2 == Type::TOP ) return Type::TOP; 2339 const TypePtr *tp = t2->is_ptr(); 2340 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2341 const TypeAryPtr *tap = tp->isa_aryptr(); 2342 if( !tap ) return _type; 2343 return tap->size(); 2344 } 2345 2346 //-------------------------------Ideal--------------------------------------- 2347 // Feed through the length in AllocateArray(...length...)._length. 2348 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2349 Node* p = MemNode::Ideal_common(phase, can_reshape); 2350 if (p) return (p == NodeSentinel) ? NULL : p; 2351 2352 // Take apart the address into an oop and and offset. 2353 // Return 'this' if we cannot. 2354 Node* adr = in(MemNode::Address); 2355 intptr_t offset = 0; 2356 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2357 if (base == NULL) return NULL; 2358 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2359 if (tary == NULL) return NULL; 2360 2361 // We can fetch the length directly through an AllocateArrayNode. 2362 // This works even if the length is not constant (clone or newArray). 2363 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2364 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2365 if (alloc != NULL) { 2366 Node* allocated_length = alloc->Ideal_length(); 2367 Node* len = alloc->make_ideal_length(tary, phase); 2368 if (allocated_length != len) { 2369 // New CastII improves on this. 2370 return len; 2371 } 2372 } 2373 } 2374 2375 return NULL; 2376 } 2377 2378 //------------------------------Identity--------------------------------------- 2379 // Feed through the length in AllocateArray(...length...)._length. 2380 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2381 Node* x = LoadINode::Identity(phase); 2382 if (x != this) return x; 2383 2384 // Take apart the address into an oop and and offset. 2385 // Return 'this' if we cannot. 2386 Node* adr = in(MemNode::Address); 2387 intptr_t offset = 0; 2388 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2389 if (base == NULL) return this; 2390 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2391 if (tary == NULL) return this; 2392 2393 // We can fetch the length directly through an AllocateArrayNode. 2394 // This works even if the length is not constant (clone or newArray). 2395 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2396 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2397 if (alloc != NULL) { 2398 Node* allocated_length = alloc->Ideal_length(); 2399 // Do not allow make_ideal_length to allocate a CastII node. 2400 Node* len = alloc->make_ideal_length(tary, phase, false); 2401 if (allocated_length == len) { 2402 // Return allocated_length only if it would not be improved by a CastII. 2403 return allocated_length; 2404 } 2405 } 2406 } 2407 2408 return this; 2409 2410 } 2411 2412 //============================================================================= 2413 //---------------------------StoreNode::make----------------------------------- 2414 // Polymorphic factory method: 2415 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2416 assert((mo == unordered || mo == release), "unexpected"); 2417 Compile* C = gvn.C; 2418 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2419 ctl != NULL, "raw memory operations should have control edge"); 2420 2421 switch (bt) { 2422 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2423 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2424 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2425 case T_CHAR: 2426 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2427 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2428 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2429 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2430 case T_METADATA: 2431 case T_ADDRESS: 2432 case T_OBJECT: 2433 #ifdef _LP64 2434 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2435 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2436 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2437 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2438 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2439 adr->bottom_type()->isa_rawptr())) { 2440 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2441 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2442 } 2443 #endif 2444 { 2445 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2446 } 2447 default: 2448 ShouldNotReachHere(); 2449 return (StoreNode*)NULL; 2450 } 2451 } 2452 2453 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2454 bool require_atomic = true; 2455 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2456 } 2457 2458 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2459 bool require_atomic = true; 2460 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2461 } 2462 2463 2464 //--------------------------bottom_type---------------------------------------- 2465 const Type *StoreNode::bottom_type() const { 2466 return Type::MEMORY; 2467 } 2468 2469 //------------------------------hash------------------------------------------- 2470 uint StoreNode::hash() const { 2471 // unroll addition of interesting fields 2472 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2473 2474 // Since they are not commoned, do not hash them: 2475 return NO_HASH; 2476 } 2477 2478 //------------------------------Ideal------------------------------------------ 2479 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2480 // When a store immediately follows a relevant allocation/initialization, 2481 // try to capture it into the initialization, or hoist it above. 2482 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2483 Node* p = MemNode::Ideal_common(phase, can_reshape); 2484 if (p) return (p == NodeSentinel) ? NULL : p; 2485 2486 Node* mem = in(MemNode::Memory); 2487 Node* address = in(MemNode::Address); 2488 // Back-to-back stores to same address? Fold em up. Generally 2489 // unsafe if I have intervening uses... Also disallowed for StoreCM 2490 // since they must follow each StoreP operation. Redundant StoreCMs 2491 // are eliminated just before matching in final_graph_reshape. 2492 { 2493 Node* st = mem; 2494 // If Store 'st' has more than one use, we cannot fold 'st' away. 2495 // For example, 'st' might be the final state at a conditional 2496 // return. Or, 'st' might be used by some node which is live at 2497 // the same time 'st' is live, which might be unschedulable. So, 2498 // require exactly ONE user until such time as we clone 'mem' for 2499 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2500 // true). 2501 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2502 // Looking at a dead closed cycle of memory? 2503 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2504 assert(Opcode() == st->Opcode() || 2505 st->Opcode() == Op_StoreVector || 2506 Opcode() == Op_StoreVector || 2507 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2508 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2509 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2510 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2511 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2512 2513 if (st->in(MemNode::Address)->eqv_uncast(address) && 2514 st->as_Store()->memory_size() <= this->memory_size()) { 2515 Node* use = st->raw_out(0); 2516 phase->igvn_rehash_node_delayed(use); 2517 if (can_reshape) { 2518 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2519 } else { 2520 // It's OK to do this in the parser, since DU info is always accurate, 2521 // and the parser always refers to nodes via SafePointNode maps. 2522 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2523 } 2524 return this; 2525 } 2526 st = st->in(MemNode::Memory); 2527 } 2528 } 2529 2530 2531 // Capture an unaliased, unconditional, simple store into an initializer. 2532 // Or, if it is independent of the allocation, hoist it above the allocation. 2533 if (ReduceFieldZeroing && /*can_reshape &&*/ 2534 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2535 InitializeNode* init = mem->in(0)->as_Initialize(); 2536 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2537 if (offset > 0) { 2538 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2539 // If the InitializeNode captured me, it made a raw copy of me, 2540 // and I need to disappear. 2541 if (moved != NULL) { 2542 // %%% hack to ensure that Ideal returns a new node: 2543 mem = MergeMemNode::make(mem); 2544 return mem; // fold me away 2545 } 2546 } 2547 } 2548 2549 return NULL; // No further progress 2550 } 2551 2552 //------------------------------Value----------------------------------------- 2553 const Type* StoreNode::Value(PhaseGVN* phase) const { 2554 // Either input is TOP ==> the result is TOP 2555 const Type *t1 = phase->type( in(MemNode::Memory) ); 2556 if( t1 == Type::TOP ) return Type::TOP; 2557 const Type *t2 = phase->type( in(MemNode::Address) ); 2558 if( t2 == Type::TOP ) return Type::TOP; 2559 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2560 if( t3 == Type::TOP ) return Type::TOP; 2561 return Type::MEMORY; 2562 } 2563 2564 //------------------------------Identity--------------------------------------- 2565 // Remove redundant stores: 2566 // Store(m, p, Load(m, p)) changes to m. 2567 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2568 Node* StoreNode::Identity(PhaseGVN* phase) { 2569 Node* mem = in(MemNode::Memory); 2570 Node* adr = in(MemNode::Address); 2571 Node* val = in(MemNode::ValueIn); 2572 2573 Node* result = this; 2574 2575 // Load then Store? Then the Store is useless 2576 if (val->is_Load() && 2577 val->in(MemNode::Address)->eqv_uncast(adr) && 2578 val->in(MemNode::Memory )->eqv_uncast(mem) && 2579 val->as_Load()->store_Opcode() == Opcode()) { 2580 result = mem; 2581 } 2582 2583 // Two stores in a row of the same value? 2584 if (result == this && 2585 mem->is_Store() && 2586 mem->in(MemNode::Address)->eqv_uncast(adr) && 2587 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2588 mem->Opcode() == Opcode()) { 2589 result = mem; 2590 } 2591 2592 // Store of zero anywhere into a freshly-allocated object? 2593 // Then the store is useless. 2594 // (It must already have been captured by the InitializeNode.) 2595 if (result == this && 2596 ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2597 // a newly allocated object is already all-zeroes everywhere 2598 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2599 result = mem; 2600 } 2601 2602 if (result == this) { 2603 // the store may also apply to zero-bits in an earlier object 2604 Node* prev_mem = find_previous_store(phase); 2605 // Steps (a), (b): Walk past independent stores to find an exact match. 2606 if (prev_mem != NULL) { 2607 Node* prev_val = can_see_stored_value(prev_mem, phase); 2608 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2609 // prev_val and val might differ by a cast; it would be good 2610 // to keep the more informative of the two. 2611 result = mem; 2612 } 2613 } 2614 } 2615 } 2616 2617 if (result != this && phase->is_IterGVN() != NULL) { 2618 MemBarNode* trailing = trailing_membar(); 2619 if (trailing != NULL) { 2620 #ifdef ASSERT 2621 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2622 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects"); 2623 #endif 2624 PhaseIterGVN* igvn = phase->is_IterGVN(); 2625 trailing->remove(igvn); 2626 } 2627 } 2628 2629 return result; 2630 } 2631 2632 //------------------------------match_edge------------------------------------- 2633 // Do we Match on this edge index or not? Match only memory & value 2634 uint StoreNode::match_edge(uint idx) const { 2635 return idx == MemNode::Address || idx == MemNode::ValueIn; 2636 } 2637 2638 //------------------------------cmp-------------------------------------------- 2639 // Do not common stores up together. They generally have to be split 2640 // back up anyways, so do not bother. 2641 bool StoreNode::cmp( const Node &n ) const { 2642 return (&n == this); // Always fail except on self 2643 } 2644 2645 //------------------------------Ideal_masked_input----------------------------- 2646 // Check for a useless mask before a partial-word store 2647 // (StoreB ... (AndI valIn conIa) ) 2648 // If (conIa & mask == mask) this simplifies to 2649 // (StoreB ... (valIn) ) 2650 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2651 Node *val = in(MemNode::ValueIn); 2652 if( val->Opcode() == Op_AndI ) { 2653 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2654 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2655 set_req(MemNode::ValueIn, val->in(1)); 2656 return this; 2657 } 2658 } 2659 return NULL; 2660 } 2661 2662 2663 //------------------------------Ideal_sign_extended_input---------------------- 2664 // Check for useless sign-extension before a partial-word store 2665 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2666 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2667 // (StoreB ... (valIn) ) 2668 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2669 Node *val = in(MemNode::ValueIn); 2670 if( val->Opcode() == Op_RShiftI ) { 2671 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2672 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2673 Node *shl = val->in(1); 2674 if( shl->Opcode() == Op_LShiftI ) { 2675 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2676 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2677 set_req(MemNode::ValueIn, shl->in(1)); 2678 return this; 2679 } 2680 } 2681 } 2682 } 2683 return NULL; 2684 } 2685 2686 //------------------------------value_never_loaded----------------------------------- 2687 // Determine whether there are any possible loads of the value stored. 2688 // For simplicity, we actually check if there are any loads from the 2689 // address stored to, not just for loads of the value stored by this node. 2690 // 2691 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2692 Node *adr = in(Address); 2693 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2694 if (adr_oop == NULL) 2695 return false; 2696 if (!adr_oop->is_known_instance_field()) 2697 return false; // if not a distinct instance, there may be aliases of the address 2698 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2699 Node *use = adr->fast_out(i); 2700 if (use->is_Load() || use->is_LoadStore()) { 2701 return false; 2702 } 2703 } 2704 return true; 2705 } 2706 2707 MemBarNode* StoreNode::trailing_membar() const { 2708 if (is_release()) { 2709 MemBarNode* trailing_mb = NULL; 2710 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2711 Node* u = fast_out(i); 2712 if (u->is_MemBar()) { 2713 if (u->as_MemBar()->trailing_store()) { 2714 assert(u->Opcode() == Op_MemBarVolatile, ""); 2715 assert(trailing_mb == NULL, "only one"); 2716 trailing_mb = u->as_MemBar(); 2717 #ifdef ASSERT 2718 Node* leading = u->as_MemBar()->leading_membar(); 2719 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2720 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2721 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2722 #endif 2723 } else { 2724 assert(u->as_MemBar()->standalone(), ""); 2725 } 2726 } 2727 } 2728 return trailing_mb; 2729 } 2730 return NULL; 2731 } 2732 2733 2734 //============================================================================= 2735 //------------------------------Ideal------------------------------------------ 2736 // If the store is from an AND mask that leaves the low bits untouched, then 2737 // we can skip the AND operation. If the store is from a sign-extension 2738 // (a left shift, then right shift) we can skip both. 2739 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2740 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2741 if( progress != NULL ) return progress; 2742 2743 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2744 if( progress != NULL ) return progress; 2745 2746 // Finally check the default case 2747 return StoreNode::Ideal(phase, can_reshape); 2748 } 2749 2750 //============================================================================= 2751 //------------------------------Ideal------------------------------------------ 2752 // If the store is from an AND mask that leaves the low bits untouched, then 2753 // we can skip the AND operation 2754 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2755 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2756 if( progress != NULL ) return progress; 2757 2758 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2759 if( progress != NULL ) return progress; 2760 2761 // Finally check the default case 2762 return StoreNode::Ideal(phase, can_reshape); 2763 } 2764 2765 //============================================================================= 2766 //------------------------------Identity--------------------------------------- 2767 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2768 // No need to card mark when storing a null ptr 2769 Node* my_store = in(MemNode::OopStore); 2770 if (my_store->is_Store()) { 2771 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2772 if( t1 == TypePtr::NULL_PTR ) { 2773 return in(MemNode::Memory); 2774 } 2775 } 2776 return this; 2777 } 2778 2779 //============================================================================= 2780 //------------------------------Ideal--------------------------------------- 2781 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2782 Node* progress = StoreNode::Ideal(phase, can_reshape); 2783 if (progress != NULL) return progress; 2784 2785 Node* my_store = in(MemNode::OopStore); 2786 if (my_store->is_MergeMem()) { 2787 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2788 set_req(MemNode::OopStore, mem); 2789 return this; 2790 } 2791 2792 return NULL; 2793 } 2794 2795 //------------------------------Value----------------------------------------- 2796 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2797 // Either input is TOP ==> the result is TOP 2798 const Type *t = phase->type( in(MemNode::Memory) ); 2799 if( t == Type::TOP ) return Type::TOP; 2800 t = phase->type( in(MemNode::Address) ); 2801 if( t == Type::TOP ) return Type::TOP; 2802 t = phase->type( in(MemNode::ValueIn) ); 2803 if( t == Type::TOP ) return Type::TOP; 2804 // If extra input is TOP ==> the result is TOP 2805 t = phase->type( in(MemNode::OopStore) ); 2806 if( t == Type::TOP ) return Type::TOP; 2807 2808 return StoreNode::Value( phase ); 2809 } 2810 2811 2812 //============================================================================= 2813 //----------------------------------SCMemProjNode------------------------------ 2814 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2815 { 2816 return bottom_type(); 2817 } 2818 2819 //============================================================================= 2820 //----------------------------------LoadStoreNode------------------------------ 2821 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2822 : Node(required), 2823 _type(rt), 2824 _adr_type(at), 2825 _has_barrier(false) 2826 { 2827 init_req(MemNode::Control, c ); 2828 init_req(MemNode::Memory , mem); 2829 init_req(MemNode::Address, adr); 2830 init_req(MemNode::ValueIn, val); 2831 init_class_id(Class_LoadStore); 2832 } 2833 2834 uint LoadStoreNode::ideal_reg() const { 2835 return _type->ideal_reg(); 2836 } 2837 2838 bool LoadStoreNode::result_not_used() const { 2839 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2840 Node *x = fast_out(i); 2841 if (x->Opcode() == Op_SCMemProj) continue; 2842 return false; 2843 } 2844 return true; 2845 } 2846 2847 MemBarNode* LoadStoreNode::trailing_membar() const { 2848 MemBarNode* trailing = NULL; 2849 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2850 Node* u = fast_out(i); 2851 if (u->is_MemBar()) { 2852 if (u->as_MemBar()->trailing_load_store()) { 2853 assert(u->Opcode() == Op_MemBarAcquire, ""); 2854 assert(trailing == NULL, "only one"); 2855 trailing = u->as_MemBar(); 2856 #ifdef ASSERT 2857 Node* leading = trailing->leading_membar(); 2858 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2859 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 2860 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 2861 #endif 2862 } else { 2863 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 2864 } 2865 } 2866 } 2867 2868 return trailing; 2869 } 2870 2871 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2872 2873 //============================================================================= 2874 //----------------------------------LoadStoreConditionalNode-------------------- 2875 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2876 init_req(ExpectedIn, ex ); 2877 } 2878 2879 //============================================================================= 2880 //-------------------------------adr_type-------------------------------------- 2881 const TypePtr* ClearArrayNode::adr_type() const { 2882 Node *adr = in(3); 2883 if (adr == NULL) return NULL; // node is dead 2884 return MemNode::calculate_adr_type(adr->bottom_type()); 2885 } 2886 2887 //------------------------------match_edge------------------------------------- 2888 // Do we Match on this edge index or not? Do not match memory 2889 uint ClearArrayNode::match_edge(uint idx) const { 2890 return idx > 1; 2891 } 2892 2893 //------------------------------Identity--------------------------------------- 2894 // Clearing a zero length array does nothing 2895 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2896 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2897 } 2898 2899 //------------------------------Idealize--------------------------------------- 2900 // Clearing a short array is faster with stores 2901 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2902 // Already know this is a large node, do not try to ideal it 2903 if (!IdealizeClearArrayNode || _is_large) return NULL; 2904 2905 const int unit = BytesPerLong; 2906 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2907 if (!t) return NULL; 2908 if (!t->is_con()) return NULL; 2909 intptr_t raw_count = t->get_con(); 2910 intptr_t size = raw_count; 2911 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2912 // Clearing nothing uses the Identity call. 2913 // Negative clears are possible on dead ClearArrays 2914 // (see jck test stmt114.stmt11402.val). 2915 if (size <= 0 || size % unit != 0) return NULL; 2916 intptr_t count = size / unit; 2917 // Length too long; communicate this to matchers and assemblers. 2918 // Assemblers are responsible to produce fast hardware clears for it. 2919 if (size > InitArrayShortSize) { 2920 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 2921 } 2922 Node *mem = in(1); 2923 if( phase->type(mem)==Type::TOP ) return NULL; 2924 Node *adr = in(3); 2925 const Type* at = phase->type(adr); 2926 if( at==Type::TOP ) return NULL; 2927 const TypePtr* atp = at->isa_ptr(); 2928 // adjust atp to be the correct array element address type 2929 if (atp == NULL) atp = TypePtr::BOTTOM; 2930 else atp = atp->add_offset(Type::OffsetBot); 2931 // Get base for derived pointer purposes 2932 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2933 Node *base = adr->in(1); 2934 2935 Node *zero = phase->makecon(TypeLong::ZERO); 2936 Node *off = phase->MakeConX(BytesPerLong); 2937 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2938 count--; 2939 while( count-- ) { 2940 mem = phase->transform(mem); 2941 adr = phase->transform(new AddPNode(base,adr,off)); 2942 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2943 } 2944 return mem; 2945 } 2946 2947 //----------------------------step_through---------------------------------- 2948 // Return allocation input memory edge if it is different instance 2949 // or itself if it is the one we are looking for. 2950 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2951 Node* n = *np; 2952 assert(n->is_ClearArray(), "sanity"); 2953 intptr_t offset; 2954 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2955 // This method is called only before Allocate nodes are expanded 2956 // during macro nodes expansion. Before that ClearArray nodes are 2957 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2958 // Allocate nodes are expanded) which follows allocations. 2959 assert(alloc != NULL, "should have allocation"); 2960 if (alloc->_idx == instance_id) { 2961 // Can not bypass initialization of the instance we are looking for. 2962 return false; 2963 } 2964 // Otherwise skip it. 2965 InitializeNode* init = alloc->initialization(); 2966 if (init != NULL) 2967 *np = init->in(TypeFunc::Memory); 2968 else 2969 *np = alloc->in(TypeFunc::Memory); 2970 return true; 2971 } 2972 2973 //----------------------------clear_memory------------------------------------- 2974 // Generate code to initialize object storage to zero. 2975 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2976 intptr_t start_offset, 2977 Node* end_offset, 2978 PhaseGVN* phase) { 2979 intptr_t offset = start_offset; 2980 2981 int unit = BytesPerLong; 2982 if ((offset % unit) != 0) { 2983 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2984 adr = phase->transform(adr); 2985 const TypePtr* atp = TypeRawPtr::BOTTOM; 2986 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2987 mem = phase->transform(mem); 2988 offset += BytesPerInt; 2989 } 2990 assert((offset % unit) == 0, ""); 2991 2992 // Initialize the remaining stuff, if any, with a ClearArray. 2993 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2994 } 2995 2996 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2997 Node* start_offset, 2998 Node* end_offset, 2999 PhaseGVN* phase) { 3000 if (start_offset == end_offset) { 3001 // nothing to do 3002 return mem; 3003 } 3004 3005 int unit = BytesPerLong; 3006 Node* zbase = start_offset; 3007 Node* zend = end_offset; 3008 3009 // Scale to the unit required by the CPU: 3010 if (!Matcher::init_array_count_is_in_bytes) { 3011 Node* shift = phase->intcon(exact_log2(unit)); 3012 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3013 zend = phase->transform(new URShiftXNode(zend, shift) ); 3014 } 3015 3016 // Bulk clear double-words 3017 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3018 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3019 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 3020 return phase->transform(mem); 3021 } 3022 3023 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3024 intptr_t start_offset, 3025 intptr_t end_offset, 3026 PhaseGVN* phase) { 3027 if (start_offset == end_offset) { 3028 // nothing to do 3029 return mem; 3030 } 3031 3032 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3033 intptr_t done_offset = end_offset; 3034 if ((done_offset % BytesPerLong) != 0) { 3035 done_offset -= BytesPerInt; 3036 } 3037 if (done_offset > start_offset) { 3038 mem = clear_memory(ctl, mem, dest, 3039 start_offset, phase->MakeConX(done_offset), phase); 3040 } 3041 if (done_offset < end_offset) { // emit the final 32-bit store 3042 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3043 adr = phase->transform(adr); 3044 const TypePtr* atp = TypeRawPtr::BOTTOM; 3045 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3046 mem = phase->transform(mem); 3047 done_offset += BytesPerInt; 3048 } 3049 assert(done_offset == end_offset, ""); 3050 return mem; 3051 } 3052 3053 //============================================================================= 3054 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3055 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 3056 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3057 #ifdef ASSERT 3058 , _pair_idx(0) 3059 #endif 3060 { 3061 init_class_id(Class_MemBar); 3062 Node* top = C->top(); 3063 init_req(TypeFunc::I_O,top); 3064 init_req(TypeFunc::FramePtr,top); 3065 init_req(TypeFunc::ReturnAdr,top); 3066 if (precedent != NULL) 3067 init_req(TypeFunc::Parms, precedent); 3068 } 3069 3070 //------------------------------cmp-------------------------------------------- 3071 uint MemBarNode::hash() const { return NO_HASH; } 3072 bool MemBarNode::cmp( const Node &n ) const { 3073 return (&n == this); // Always fail except on self 3074 } 3075 3076 //------------------------------make------------------------------------------- 3077 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3078 switch (opcode) { 3079 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3080 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3081 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3082 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3083 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3084 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3085 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3086 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3087 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3088 case Op_Initialize: return new InitializeNode(C, atp, pn); 3089 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3090 default: ShouldNotReachHere(); return NULL; 3091 } 3092 } 3093 3094 void MemBarNode::remove(PhaseIterGVN *igvn) { 3095 if (outcnt() != 2) { 3096 return; 3097 } 3098 if (trailing_store() || trailing_load_store()) { 3099 MemBarNode* leading = leading_membar(); 3100 if (leading != NULL) { 3101 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3102 leading->remove(igvn); 3103 } 3104 } 3105 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3106 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3107 } 3108 3109 //------------------------------Ideal------------------------------------------ 3110 // Return a node which is more "ideal" than the current node. Strip out 3111 // control copies 3112 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3113 if (remove_dead_region(phase, can_reshape)) return this; 3114 // Don't bother trying to transform a dead node 3115 if (in(0) && in(0)->is_top()) { 3116 return NULL; 3117 } 3118 3119 bool progress = false; 3120 // Eliminate volatile MemBars for scalar replaced objects. 3121 if (can_reshape && req() == (Precedent+1)) { 3122 bool eliminate = false; 3123 int opc = Opcode(); 3124 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3125 // Volatile field loads and stores. 3126 Node* my_mem = in(MemBarNode::Precedent); 3127 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3128 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3129 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3130 // replace this Precedent (decodeN) with the Load instead. 3131 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3132 Node* load_node = my_mem->in(1); 3133 set_req(MemBarNode::Precedent, load_node); 3134 phase->is_IterGVN()->_worklist.push(my_mem); 3135 my_mem = load_node; 3136 } else { 3137 assert(my_mem->unique_out() == this, "sanity"); 3138 del_req(Precedent); 3139 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3140 my_mem = NULL; 3141 } 3142 progress = true; 3143 } 3144 if (my_mem != NULL && my_mem->is_Mem()) { 3145 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3146 // Check for scalar replaced object reference. 3147 if( t_oop != NULL && t_oop->is_known_instance_field() && 3148 t_oop->offset() != Type::OffsetBot && 3149 t_oop->offset() != Type::OffsetTop) { 3150 eliminate = true; 3151 } 3152 } 3153 } else if (opc == Op_MemBarRelease) { 3154 // Final field stores. 3155 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3156 if ((alloc != NULL) && alloc->is_Allocate() && 3157 alloc->as_Allocate()->does_not_escape_thread()) { 3158 // The allocated object does not escape. 3159 eliminate = true; 3160 } 3161 } 3162 if (eliminate) { 3163 // Replace MemBar projections by its inputs. 3164 PhaseIterGVN* igvn = phase->is_IterGVN(); 3165 remove(igvn); 3166 // Must return either the original node (now dead) or a new node 3167 // (Do not return a top here, since that would break the uniqueness of top.) 3168 return new ConINode(TypeInt::ZERO); 3169 } 3170 } 3171 return progress ? this : NULL; 3172 } 3173 3174 //------------------------------Value------------------------------------------ 3175 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3176 if( !in(0) ) return Type::TOP; 3177 if( phase->type(in(0)) == Type::TOP ) 3178 return Type::TOP; 3179 return TypeTuple::MEMBAR; 3180 } 3181 3182 //------------------------------match------------------------------------------ 3183 // Construct projections for memory. 3184 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3185 switch (proj->_con) { 3186 case TypeFunc::Control: 3187 case TypeFunc::Memory: 3188 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3189 } 3190 ShouldNotReachHere(); 3191 return NULL; 3192 } 3193 3194 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3195 trailing->_kind = TrailingStore; 3196 leading->_kind = LeadingStore; 3197 #ifdef ASSERT 3198 trailing->_pair_idx = leading->_idx; 3199 leading->_pair_idx = leading->_idx; 3200 #endif 3201 } 3202 3203 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3204 trailing->_kind = TrailingLoadStore; 3205 leading->_kind = LeadingLoadStore; 3206 #ifdef ASSERT 3207 trailing->_pair_idx = leading->_idx; 3208 leading->_pair_idx = leading->_idx; 3209 #endif 3210 } 3211 3212 MemBarNode* MemBarNode::trailing_membar() const { 3213 ResourceMark rm; 3214 Node* trailing = (Node*)this; 3215 VectorSet seen(Thread::current()->resource_area()); 3216 Node_Stack multis(0); 3217 do { 3218 Node* c = trailing; 3219 uint i = 0; 3220 do { 3221 trailing = NULL; 3222 for (; i < c->outcnt(); i++) { 3223 Node* next = c->raw_out(i); 3224 if (next != c && next->is_CFG()) { 3225 if (c->is_MultiBranch()) { 3226 if (multis.node() == c) { 3227 multis.set_index(i+1); 3228 } else { 3229 multis.push(c, i+1); 3230 } 3231 } 3232 trailing = next; 3233 break; 3234 } 3235 } 3236 if (trailing != NULL && !seen.test_set(trailing->_idx)) { 3237 break; 3238 } 3239 while (multis.size() > 0) { 3240 c = multis.node(); 3241 i = multis.index(); 3242 if (i < c->req()) { 3243 break; 3244 } 3245 multis.pop(); 3246 } 3247 } while (multis.size() > 0); 3248 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3249 3250 MemBarNode* mb = trailing->as_MemBar(); 3251 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3252 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3253 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3254 return mb; 3255 } 3256 3257 MemBarNode* MemBarNode::leading_membar() const { 3258 ResourceMark rm; 3259 VectorSet seen(Thread::current()->resource_area()); 3260 Node_Stack regions(0); 3261 Node* leading = in(0); 3262 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3263 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) { 3264 leading = NULL; 3265 while (regions.size() > 0 && leading == NULL) { 3266 Node* r = regions.node(); 3267 uint i = regions.index(); 3268 if (i < r->req()) { 3269 leading = r->in(i); 3270 regions.set_index(i+1); 3271 } else { 3272 regions.pop(); 3273 } 3274 } 3275 if (leading == NULL) { 3276 assert(regions.size() == 0, "all paths should have been tried"); 3277 return NULL; 3278 } 3279 } 3280 if (leading->is_Region()) { 3281 regions.push(leading, 2); 3282 leading = leading->in(1); 3283 } else { 3284 leading = leading->in(0); 3285 } 3286 } 3287 #ifdef ASSERT 3288 Unique_Node_List wq; 3289 wq.push((Node*)this); 3290 uint found = 0; 3291 for (uint i = 0; i < wq.size(); i++) { 3292 Node* n = wq.at(i); 3293 if (n->is_Region()) { 3294 for (uint j = 1; j < n->req(); j++) { 3295 Node* in = n->in(j); 3296 if (in != NULL && !in->is_top()) { 3297 wq.push(in); 3298 } 3299 } 3300 } else { 3301 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3302 assert(n == leading, "consistency check failed"); 3303 found++; 3304 } else { 3305 Node* in = n->in(0); 3306 if (in != NULL && !in->is_top()) { 3307 wq.push(in); 3308 } 3309 } 3310 } 3311 } 3312 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed"); 3313 #endif 3314 if (leading == NULL) { 3315 return NULL; 3316 } 3317 MemBarNode* mb = leading->as_MemBar(); 3318 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3319 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3320 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3321 return mb; 3322 } 3323 3324 //===========================InitializeNode==================================== 3325 // SUMMARY: 3326 // This node acts as a memory barrier on raw memory, after some raw stores. 3327 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3328 // The Initialize can 'capture' suitably constrained stores as raw inits. 3329 // It can coalesce related raw stores into larger units (called 'tiles'). 3330 // It can avoid zeroing new storage for memory units which have raw inits. 3331 // At macro-expansion, it is marked 'complete', and does not optimize further. 3332 // 3333 // EXAMPLE: 3334 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3335 // ctl = incoming control; mem* = incoming memory 3336 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3337 // First allocate uninitialized memory and fill in the header: 3338 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3339 // ctl := alloc.Control; mem* := alloc.Memory* 3340 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3341 // Then initialize to zero the non-header parts of the raw memory block: 3342 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3343 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3344 // After the initialize node executes, the object is ready for service: 3345 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3346 // Suppose its body is immediately initialized as {1,2}: 3347 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3348 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3349 // mem.SLICE(#short[*]) := store2 3350 // 3351 // DETAILS: 3352 // An InitializeNode collects and isolates object initialization after 3353 // an AllocateNode and before the next possible safepoint. As a 3354 // memory barrier (MemBarNode), it keeps critical stores from drifting 3355 // down past any safepoint or any publication of the allocation. 3356 // Before this barrier, a newly-allocated object may have uninitialized bits. 3357 // After this barrier, it may be treated as a real oop, and GC is allowed. 3358 // 3359 // The semantics of the InitializeNode include an implicit zeroing of 3360 // the new object from object header to the end of the object. 3361 // (The object header and end are determined by the AllocateNode.) 3362 // 3363 // Certain stores may be added as direct inputs to the InitializeNode. 3364 // These stores must update raw memory, and they must be to addresses 3365 // derived from the raw address produced by AllocateNode, and with 3366 // a constant offset. They must be ordered by increasing offset. 3367 // The first one is at in(RawStores), the last at in(req()-1). 3368 // Unlike most memory operations, they are not linked in a chain, 3369 // but are displayed in parallel as users of the rawmem output of 3370 // the allocation. 3371 // 3372 // (See comments in InitializeNode::capture_store, which continue 3373 // the example given above.) 3374 // 3375 // When the associated Allocate is macro-expanded, the InitializeNode 3376 // may be rewritten to optimize collected stores. A ClearArrayNode 3377 // may also be created at that point to represent any required zeroing. 3378 // The InitializeNode is then marked 'complete', prohibiting further 3379 // capturing of nearby memory operations. 3380 // 3381 // During macro-expansion, all captured initializations which store 3382 // constant values of 32 bits or smaller are coalesced (if advantageous) 3383 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3384 // initialized in fewer memory operations. Memory words which are 3385 // covered by neither tiles nor non-constant stores are pre-zeroed 3386 // by explicit stores of zero. (The code shape happens to do all 3387 // zeroing first, then all other stores, with both sequences occurring 3388 // in order of ascending offsets.) 3389 // 3390 // Alternatively, code may be inserted between an AllocateNode and its 3391 // InitializeNode, to perform arbitrary initialization of the new object. 3392 // E.g., the object copying intrinsics insert complex data transfers here. 3393 // The initialization must then be marked as 'complete' disable the 3394 // built-in zeroing semantics and the collection of initializing stores. 3395 // 3396 // While an InitializeNode is incomplete, reads from the memory state 3397 // produced by it are optimizable if they match the control edge and 3398 // new oop address associated with the allocation/initialization. 3399 // They return a stored value (if the offset matches) or else zero. 3400 // A write to the memory state, if it matches control and address, 3401 // and if it is to a constant offset, may be 'captured' by the 3402 // InitializeNode. It is cloned as a raw memory operation and rewired 3403 // inside the initialization, to the raw oop produced by the allocation. 3404 // Operations on addresses which are provably distinct (e.g., to 3405 // other AllocateNodes) are allowed to bypass the initialization. 3406 // 3407 // The effect of all this is to consolidate object initialization 3408 // (both arrays and non-arrays, both piecewise and bulk) into a 3409 // single location, where it can be optimized as a unit. 3410 // 3411 // Only stores with an offset less than TrackedInitializationLimit words 3412 // will be considered for capture by an InitializeNode. This puts a 3413 // reasonable limit on the complexity of optimized initializations. 3414 3415 //---------------------------InitializeNode------------------------------------ 3416 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3417 : MemBarNode(C, adr_type, rawoop), 3418 _is_complete(Incomplete), _does_not_escape(false) 3419 { 3420 init_class_id(Class_Initialize); 3421 3422 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3423 assert(in(RawAddress) == rawoop, "proper init"); 3424 // Note: allocation() can be NULL, for secondary initialization barriers 3425 } 3426 3427 // Since this node is not matched, it will be processed by the 3428 // register allocator. Declare that there are no constraints 3429 // on the allocation of the RawAddress edge. 3430 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3431 // This edge should be set to top, by the set_complete. But be conservative. 3432 if (idx == InitializeNode::RawAddress) 3433 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3434 return RegMask::Empty; 3435 } 3436 3437 Node* InitializeNode::memory(uint alias_idx) { 3438 Node* mem = in(Memory); 3439 if (mem->is_MergeMem()) { 3440 return mem->as_MergeMem()->memory_at(alias_idx); 3441 } else { 3442 // incoming raw memory is not split 3443 return mem; 3444 } 3445 } 3446 3447 bool InitializeNode::is_non_zero() { 3448 if (is_complete()) return false; 3449 remove_extra_zeroes(); 3450 return (req() > RawStores); 3451 } 3452 3453 void InitializeNode::set_complete(PhaseGVN* phase) { 3454 assert(!is_complete(), "caller responsibility"); 3455 _is_complete = Complete; 3456 3457 // After this node is complete, it contains a bunch of 3458 // raw-memory initializations. There is no need for 3459 // it to have anything to do with non-raw memory effects. 3460 // Therefore, tell all non-raw users to re-optimize themselves, 3461 // after skipping the memory effects of this initialization. 3462 PhaseIterGVN* igvn = phase->is_IterGVN(); 3463 if (igvn) igvn->add_users_to_worklist(this); 3464 } 3465 3466 // convenience function 3467 // return false if the init contains any stores already 3468 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3469 InitializeNode* init = initialization(); 3470 if (init == NULL || init->is_complete()) return false; 3471 init->remove_extra_zeroes(); 3472 // for now, if this allocation has already collected any inits, bail: 3473 if (init->is_non_zero()) return false; 3474 init->set_complete(phase); 3475 return true; 3476 } 3477 3478 void InitializeNode::remove_extra_zeroes() { 3479 if (req() == RawStores) return; 3480 Node* zmem = zero_memory(); 3481 uint fill = RawStores; 3482 for (uint i = fill; i < req(); i++) { 3483 Node* n = in(i); 3484 if (n->is_top() || n == zmem) continue; // skip 3485 if (fill < i) set_req(fill, n); // compact 3486 ++fill; 3487 } 3488 // delete any empty spaces created: 3489 while (fill < req()) { 3490 del_req(fill); 3491 } 3492 } 3493 3494 // Helper for remembering which stores go with which offsets. 3495 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3496 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3497 intptr_t offset = -1; 3498 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3499 phase, offset); 3500 if (base == NULL) return -1; // something is dead, 3501 if (offset < 0) return -1; // dead, dead 3502 return offset; 3503 } 3504 3505 // Helper for proving that an initialization expression is 3506 // "simple enough" to be folded into an object initialization. 3507 // Attempts to prove that a store's initial value 'n' can be captured 3508 // within the initialization without creating a vicious cycle, such as: 3509 // { Foo p = new Foo(); p.next = p; } 3510 // True for constants and parameters and small combinations thereof. 3511 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3512 if (n == NULL) return true; // (can this really happen?) 3513 if (n->is_Proj()) n = n->in(0); 3514 if (n == this) return false; // found a cycle 3515 if (n->is_Con()) return true; 3516 if (n->is_Start()) return true; // params, etc., are OK 3517 if (n->is_Root()) return true; // even better 3518 3519 Node* ctl = n->in(0); 3520 if (ctl != NULL && !ctl->is_top()) { 3521 if (ctl->is_Proj()) ctl = ctl->in(0); 3522 if (ctl == this) return false; 3523 3524 // If we already know that the enclosing memory op is pinned right after 3525 // the init, then any control flow that the store has picked up 3526 // must have preceded the init, or else be equal to the init. 3527 // Even after loop optimizations (which might change control edges) 3528 // a store is never pinned *before* the availability of its inputs. 3529 if (!MemNode::all_controls_dominate(n, this)) 3530 return false; // failed to prove a good control 3531 } 3532 3533 // Check data edges for possible dependencies on 'this'. 3534 if ((count += 1) > 20) return false; // complexity limit 3535 for (uint i = 1; i < n->req(); i++) { 3536 Node* m = n->in(i); 3537 if (m == NULL || m == n || m->is_top()) continue; 3538 uint first_i = n->find_edge(m); 3539 if (i != first_i) continue; // process duplicate edge just once 3540 if (!detect_init_independence(m, count)) { 3541 return false; 3542 } 3543 } 3544 3545 return true; 3546 } 3547 3548 // Here are all the checks a Store must pass before it can be moved into 3549 // an initialization. Returns zero if a check fails. 3550 // On success, returns the (constant) offset to which the store applies, 3551 // within the initialized memory. 3552 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3553 const int FAIL = 0; 3554 if (st->req() != MemNode::ValueIn + 1) 3555 return FAIL; // an inscrutable StoreNode (card mark?) 3556 Node* ctl = st->in(MemNode::Control); 3557 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3558 return FAIL; // must be unconditional after the initialization 3559 Node* mem = st->in(MemNode::Memory); 3560 if (!(mem->is_Proj() && mem->in(0) == this)) 3561 return FAIL; // must not be preceded by other stores 3562 Node* adr = st->in(MemNode::Address); 3563 intptr_t offset; 3564 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3565 if (alloc == NULL) 3566 return FAIL; // inscrutable address 3567 if (alloc != allocation()) 3568 return FAIL; // wrong allocation! (store needs to float up) 3569 int size_in_bytes = st->memory_size(); 3570 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) { 3571 return FAIL; // mismatched access 3572 } 3573 Node* val = st->in(MemNode::ValueIn); 3574 int complexity_count = 0; 3575 if (!detect_init_independence(val, complexity_count)) 3576 return FAIL; // stored value must be 'simple enough' 3577 3578 // The Store can be captured only if nothing after the allocation 3579 // and before the Store is using the memory location that the store 3580 // overwrites. 3581 bool failed = false; 3582 // If is_complete_with_arraycopy() is true the shape of the graph is 3583 // well defined and is safe so no need for extra checks. 3584 if (!is_complete_with_arraycopy()) { 3585 // We are going to look at each use of the memory state following 3586 // the allocation to make sure nothing reads the memory that the 3587 // Store writes. 3588 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3589 int alias_idx = phase->C->get_alias_index(t_adr); 3590 ResourceMark rm; 3591 Unique_Node_List mems; 3592 mems.push(mem); 3593 Node* unique_merge = NULL; 3594 for (uint next = 0; next < mems.size(); ++next) { 3595 Node *m = mems.at(next); 3596 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3597 Node *n = m->fast_out(j); 3598 if (n->outcnt() == 0) { 3599 continue; 3600 } 3601 if (n == st) { 3602 continue; 3603 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3604 // If the control of this use is different from the control 3605 // of the Store which is right after the InitializeNode then 3606 // this node cannot be between the InitializeNode and the 3607 // Store. 3608 continue; 3609 } else if (n->is_MergeMem()) { 3610 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3611 // We can hit a MergeMemNode (that will likely go away 3612 // later) that is a direct use of the memory state 3613 // following the InitializeNode on the same slice as the 3614 // store node that we'd like to capture. We need to check 3615 // the uses of the MergeMemNode. 3616 mems.push(n); 3617 } 3618 } else if (n->is_Mem()) { 3619 Node* other_adr = n->in(MemNode::Address); 3620 if (other_adr == adr) { 3621 failed = true; 3622 break; 3623 } else { 3624 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3625 if (other_t_adr != NULL) { 3626 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3627 if (other_alias_idx == alias_idx) { 3628 // A load from the same memory slice as the store right 3629 // after the InitializeNode. We check the control of the 3630 // object/array that is loaded from. If it's the same as 3631 // the store control then we cannot capture the store. 3632 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3633 Node* base = other_adr; 3634 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3635 base = base->in(AddPNode::Base); 3636 if (base != NULL) { 3637 base = base->uncast(); 3638 if (base->is_Proj() && base->in(0) == alloc) { 3639 failed = true; 3640 break; 3641 } 3642 } 3643 } 3644 } 3645 } 3646 } else { 3647 failed = true; 3648 break; 3649 } 3650 } 3651 } 3652 } 3653 if (failed) { 3654 if (!can_reshape) { 3655 // We decided we couldn't capture the store during parsing. We 3656 // should try again during the next IGVN once the graph is 3657 // cleaner. 3658 phase->C->record_for_igvn(st); 3659 } 3660 return FAIL; 3661 } 3662 3663 return offset; // success 3664 } 3665 3666 // Find the captured store in(i) which corresponds to the range 3667 // [start..start+size) in the initialized object. 3668 // If there is one, return its index i. If there isn't, return the 3669 // negative of the index where it should be inserted. 3670 // Return 0 if the queried range overlaps an initialization boundary 3671 // or if dead code is encountered. 3672 // If size_in_bytes is zero, do not bother with overlap checks. 3673 int InitializeNode::captured_store_insertion_point(intptr_t start, 3674 int size_in_bytes, 3675 PhaseTransform* phase) { 3676 const int FAIL = 0, MAX_STORE = BytesPerLong; 3677 3678 if (is_complete()) 3679 return FAIL; // arraycopy got here first; punt 3680 3681 assert(allocation() != NULL, "must be present"); 3682 3683 // no negatives, no header fields: 3684 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3685 3686 // after a certain size, we bail out on tracking all the stores: 3687 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3688 if (start >= ti_limit) return FAIL; 3689 3690 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3691 if (i >= limit) return -(int)i; // not found; here is where to put it 3692 3693 Node* st = in(i); 3694 intptr_t st_off = get_store_offset(st, phase); 3695 if (st_off < 0) { 3696 if (st != zero_memory()) { 3697 return FAIL; // bail out if there is dead garbage 3698 } 3699 } else if (st_off > start) { 3700 // ...we are done, since stores are ordered 3701 if (st_off < start + size_in_bytes) { 3702 return FAIL; // the next store overlaps 3703 } 3704 return -(int)i; // not found; here is where to put it 3705 } else if (st_off < start) { 3706 if (size_in_bytes != 0 && 3707 start < st_off + MAX_STORE && 3708 start < st_off + st->as_Store()->memory_size()) { 3709 return FAIL; // the previous store overlaps 3710 } 3711 } else { 3712 if (size_in_bytes != 0 && 3713 st->as_Store()->memory_size() != size_in_bytes) { 3714 return FAIL; // mismatched store size 3715 } 3716 return i; 3717 } 3718 3719 ++i; 3720 } 3721 } 3722 3723 // Look for a captured store which initializes at the offset 'start' 3724 // with the given size. If there is no such store, and no other 3725 // initialization interferes, then return zero_memory (the memory 3726 // projection of the AllocateNode). 3727 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3728 PhaseTransform* phase) { 3729 assert(stores_are_sane(phase), ""); 3730 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3731 if (i == 0) { 3732 return NULL; // something is dead 3733 } else if (i < 0) { 3734 return zero_memory(); // just primordial zero bits here 3735 } else { 3736 Node* st = in(i); // here is the store at this position 3737 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3738 return st; 3739 } 3740 } 3741 3742 // Create, as a raw pointer, an address within my new object at 'offset'. 3743 Node* InitializeNode::make_raw_address(intptr_t offset, 3744 PhaseTransform* phase) { 3745 Node* addr = in(RawAddress); 3746 if (offset != 0) { 3747 Compile* C = phase->C; 3748 addr = phase->transform( new AddPNode(C->top(), addr, 3749 phase->MakeConX(offset)) ); 3750 } 3751 return addr; 3752 } 3753 3754 // Clone the given store, converting it into a raw store 3755 // initializing a field or element of my new object. 3756 // Caller is responsible for retiring the original store, 3757 // with subsume_node or the like. 3758 // 3759 // From the example above InitializeNode::InitializeNode, 3760 // here are the old stores to be captured: 3761 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3762 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3763 // 3764 // Here is the changed code; note the extra edges on init: 3765 // alloc = (Allocate ...) 3766 // rawoop = alloc.RawAddress 3767 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3768 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3769 // init = (Initialize alloc.Control alloc.Memory rawoop 3770 // rawstore1 rawstore2) 3771 // 3772 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3773 PhaseTransform* phase, bool can_reshape) { 3774 assert(stores_are_sane(phase), ""); 3775 3776 if (start < 0) return NULL; 3777 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3778 3779 Compile* C = phase->C; 3780 int size_in_bytes = st->memory_size(); 3781 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3782 if (i == 0) return NULL; // bail out 3783 Node* prev_mem = NULL; // raw memory for the captured store 3784 if (i > 0) { 3785 prev_mem = in(i); // there is a pre-existing store under this one 3786 set_req(i, C->top()); // temporarily disconnect it 3787 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3788 } else { 3789 i = -i; // no pre-existing store 3790 prev_mem = zero_memory(); // a slice of the newly allocated object 3791 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3792 set_req(--i, C->top()); // reuse this edge; it has been folded away 3793 else 3794 ins_req(i, C->top()); // build a new edge 3795 } 3796 Node* new_st = st->clone(); 3797 new_st->set_req(MemNode::Control, in(Control)); 3798 new_st->set_req(MemNode::Memory, prev_mem); 3799 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3800 new_st = phase->transform(new_st); 3801 3802 // At this point, new_st might have swallowed a pre-existing store 3803 // at the same offset, or perhaps new_st might have disappeared, 3804 // if it redundantly stored the same value (or zero to fresh memory). 3805 3806 // In any case, wire it in: 3807 phase->igvn_rehash_node_delayed(this); 3808 set_req(i, new_st); 3809 3810 // The caller may now kill the old guy. 3811 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3812 assert(check_st == new_st || check_st == NULL, "must be findable"); 3813 assert(!is_complete(), ""); 3814 return new_st; 3815 } 3816 3817 static bool store_constant(jlong* tiles, int num_tiles, 3818 intptr_t st_off, int st_size, 3819 jlong con) { 3820 if ((st_off & (st_size-1)) != 0) 3821 return false; // strange store offset (assume size==2**N) 3822 address addr = (address)tiles + st_off; 3823 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3824 switch (st_size) { 3825 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3826 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3827 case sizeof(jint): *(jint*) addr = (jint) con; break; 3828 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3829 default: return false; // strange store size (detect size!=2**N here) 3830 } 3831 return true; // return success to caller 3832 } 3833 3834 // Coalesce subword constants into int constants and possibly 3835 // into long constants. The goal, if the CPU permits, 3836 // is to initialize the object with a small number of 64-bit tiles. 3837 // Also, convert floating-point constants to bit patterns. 3838 // Non-constants are not relevant to this pass. 3839 // 3840 // In terms of the running example on InitializeNode::InitializeNode 3841 // and InitializeNode::capture_store, here is the transformation 3842 // of rawstore1 and rawstore2 into rawstore12: 3843 // alloc = (Allocate ...) 3844 // rawoop = alloc.RawAddress 3845 // tile12 = 0x00010002 3846 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3847 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3848 // 3849 void 3850 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3851 Node* size_in_bytes, 3852 PhaseGVN* phase) { 3853 Compile* C = phase->C; 3854 3855 assert(stores_are_sane(phase), ""); 3856 // Note: After this pass, they are not completely sane, 3857 // since there may be some overlaps. 3858 3859 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3860 3861 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3862 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3863 size_limit = MIN2(size_limit, ti_limit); 3864 size_limit = align_up(size_limit, BytesPerLong); 3865 int num_tiles = size_limit / BytesPerLong; 3866 3867 // allocate space for the tile map: 3868 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3869 jlong tiles_buf[small_len]; 3870 Node* nodes_buf[small_len]; 3871 jlong inits_buf[small_len]; 3872 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3873 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3874 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3875 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3876 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3877 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3878 // tiles: exact bitwise model of all primitive constants 3879 // nodes: last constant-storing node subsumed into the tiles model 3880 // inits: which bytes (in each tile) are touched by any initializations 3881 3882 //// Pass A: Fill in the tile model with any relevant stores. 3883 3884 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3885 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3886 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3887 Node* zmem = zero_memory(); // initially zero memory state 3888 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3889 Node* st = in(i); 3890 intptr_t st_off = get_store_offset(st, phase); 3891 3892 // Figure out the store's offset and constant value: 3893 if (st_off < header_size) continue; //skip (ignore header) 3894 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3895 int st_size = st->as_Store()->memory_size(); 3896 if (st_off + st_size > size_limit) break; 3897 3898 // Record which bytes are touched, whether by constant or not. 3899 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3900 continue; // skip (strange store size) 3901 3902 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3903 if (!val->singleton()) continue; //skip (non-con store) 3904 BasicType type = val->basic_type(); 3905 3906 jlong con = 0; 3907 switch (type) { 3908 case T_INT: con = val->is_int()->get_con(); break; 3909 case T_LONG: con = val->is_long()->get_con(); break; 3910 case T_FLOAT: con = jint_cast(val->getf()); break; 3911 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3912 default: continue; //skip (odd store type) 3913 } 3914 3915 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3916 st->Opcode() == Op_StoreL) { 3917 continue; // This StoreL is already optimal. 3918 } 3919 3920 // Store down the constant. 3921 store_constant(tiles, num_tiles, st_off, st_size, con); 3922 3923 intptr_t j = st_off >> LogBytesPerLong; 3924 3925 if (type == T_INT && st_size == BytesPerInt 3926 && (st_off & BytesPerInt) == BytesPerInt) { 3927 jlong lcon = tiles[j]; 3928 if (!Matcher::isSimpleConstant64(lcon) && 3929 st->Opcode() == Op_StoreI) { 3930 // This StoreI is already optimal by itself. 3931 jint* intcon = (jint*) &tiles[j]; 3932 intcon[1] = 0; // undo the store_constant() 3933 3934 // If the previous store is also optimal by itself, back up and 3935 // undo the action of the previous loop iteration... if we can. 3936 // But if we can't, just let the previous half take care of itself. 3937 st = nodes[j]; 3938 st_off -= BytesPerInt; 3939 con = intcon[0]; 3940 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3941 assert(st_off >= header_size, "still ignoring header"); 3942 assert(get_store_offset(st, phase) == st_off, "must be"); 3943 assert(in(i-1) == zmem, "must be"); 3944 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3945 assert(con == tcon->is_int()->get_con(), "must be"); 3946 // Undo the effects of the previous loop trip, which swallowed st: 3947 intcon[0] = 0; // undo store_constant() 3948 set_req(i-1, st); // undo set_req(i, zmem) 3949 nodes[j] = NULL; // undo nodes[j] = st 3950 --old_subword; // undo ++old_subword 3951 } 3952 continue; // This StoreI is already optimal. 3953 } 3954 } 3955 3956 // This store is not needed. 3957 set_req(i, zmem); 3958 nodes[j] = st; // record for the moment 3959 if (st_size < BytesPerLong) // something has changed 3960 ++old_subword; // includes int/float, but who's counting... 3961 else ++old_long; 3962 } 3963 3964 if ((old_subword + old_long) == 0) 3965 return; // nothing more to do 3966 3967 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3968 // Be sure to insert them before overlapping non-constant stores. 3969 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3970 for (int j = 0; j < num_tiles; j++) { 3971 jlong con = tiles[j]; 3972 jlong init = inits[j]; 3973 if (con == 0) continue; 3974 jint con0, con1; // split the constant, address-wise 3975 jint init0, init1; // split the init map, address-wise 3976 { union { jlong con; jint intcon[2]; } u; 3977 u.con = con; 3978 con0 = u.intcon[0]; 3979 con1 = u.intcon[1]; 3980 u.con = init; 3981 init0 = u.intcon[0]; 3982 init1 = u.intcon[1]; 3983 } 3984 3985 Node* old = nodes[j]; 3986 assert(old != NULL, "need the prior store"); 3987 intptr_t offset = (j * BytesPerLong); 3988 3989 bool split = !Matcher::isSimpleConstant64(con); 3990 3991 if (offset < header_size) { 3992 assert(offset + BytesPerInt >= header_size, "second int counts"); 3993 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3994 split = true; // only the second word counts 3995 // Example: int a[] = { 42 ... } 3996 } else if (con0 == 0 && init0 == -1) { 3997 split = true; // first word is covered by full inits 3998 // Example: int a[] = { ... foo(), 42 ... } 3999 } else if (con1 == 0 && init1 == -1) { 4000 split = true; // second word is covered by full inits 4001 // Example: int a[] = { ... 42, foo() ... } 4002 } 4003 4004 // Here's a case where init0 is neither 0 nor -1: 4005 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4006 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4007 // In this case the tile is not split; it is (jlong)42. 4008 // The big tile is stored down, and then the foo() value is inserted. 4009 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4010 4011 Node* ctl = old->in(MemNode::Control); 4012 Node* adr = make_raw_address(offset, phase); 4013 const TypePtr* atp = TypeRawPtr::BOTTOM; 4014 4015 // One or two coalesced stores to plop down. 4016 Node* st[2]; 4017 intptr_t off[2]; 4018 int nst = 0; 4019 if (!split) { 4020 ++new_long; 4021 off[nst] = offset; 4022 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4023 phase->longcon(con), T_LONG, MemNode::unordered); 4024 } else { 4025 // Omit either if it is a zero. 4026 if (con0 != 0) { 4027 ++new_int; 4028 off[nst] = offset; 4029 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4030 phase->intcon(con0), T_INT, MemNode::unordered); 4031 } 4032 if (con1 != 0) { 4033 ++new_int; 4034 offset += BytesPerInt; 4035 adr = make_raw_address(offset, phase); 4036 off[nst] = offset; 4037 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4038 phase->intcon(con1), T_INT, MemNode::unordered); 4039 } 4040 } 4041 4042 // Insert second store first, then the first before the second. 4043 // Insert each one just before any overlapping non-constant stores. 4044 while (nst > 0) { 4045 Node* st1 = st[--nst]; 4046 C->copy_node_notes_to(st1, old); 4047 st1 = phase->transform(st1); 4048 offset = off[nst]; 4049 assert(offset >= header_size, "do not smash header"); 4050 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4051 guarantee(ins_idx != 0, "must re-insert constant store"); 4052 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4053 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4054 set_req(--ins_idx, st1); 4055 else 4056 ins_req(ins_idx, st1); 4057 } 4058 } 4059 4060 if (PrintCompilation && WizardMode) 4061 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4062 old_subword, old_long, new_int, new_long); 4063 if (C->log() != NULL) 4064 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4065 old_subword, old_long, new_int, new_long); 4066 4067 // Clean up any remaining occurrences of zmem: 4068 remove_extra_zeroes(); 4069 } 4070 4071 // Explore forward from in(start) to find the first fully initialized 4072 // word, and return its offset. Skip groups of subword stores which 4073 // together initialize full words. If in(start) is itself part of a 4074 // fully initialized word, return the offset of in(start). If there 4075 // are no following full-word stores, or if something is fishy, return 4076 // a negative value. 4077 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4078 int int_map = 0; 4079 intptr_t int_map_off = 0; 4080 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4081 4082 for (uint i = start, limit = req(); i < limit; i++) { 4083 Node* st = in(i); 4084 4085 intptr_t st_off = get_store_offset(st, phase); 4086 if (st_off < 0) break; // return conservative answer 4087 4088 int st_size = st->as_Store()->memory_size(); 4089 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4090 return st_off; // we found a complete word init 4091 } 4092 4093 // update the map: 4094 4095 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4096 if (this_int_off != int_map_off) { 4097 // reset the map: 4098 int_map = 0; 4099 int_map_off = this_int_off; 4100 } 4101 4102 int subword_off = st_off - this_int_off; 4103 int_map |= right_n_bits(st_size) << subword_off; 4104 if ((int_map & FULL_MAP) == FULL_MAP) { 4105 return this_int_off; // we found a complete word init 4106 } 4107 4108 // Did this store hit or cross the word boundary? 4109 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4110 if (next_int_off == this_int_off + BytesPerInt) { 4111 // We passed the current int, without fully initializing it. 4112 int_map_off = next_int_off; 4113 int_map >>= BytesPerInt; 4114 } else if (next_int_off > this_int_off + BytesPerInt) { 4115 // We passed the current and next int. 4116 return this_int_off + BytesPerInt; 4117 } 4118 } 4119 4120 return -1; 4121 } 4122 4123 4124 // Called when the associated AllocateNode is expanded into CFG. 4125 // At this point, we may perform additional optimizations. 4126 // Linearize the stores by ascending offset, to make memory 4127 // activity as coherent as possible. 4128 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4129 intptr_t header_size, 4130 Node* size_in_bytes, 4131 PhaseGVN* phase) { 4132 assert(!is_complete(), "not already complete"); 4133 assert(stores_are_sane(phase), ""); 4134 assert(allocation() != NULL, "must be present"); 4135 4136 remove_extra_zeroes(); 4137 4138 if (ReduceFieldZeroing || ReduceBulkZeroing) 4139 // reduce instruction count for common initialization patterns 4140 coalesce_subword_stores(header_size, size_in_bytes, phase); 4141 4142 Node* zmem = zero_memory(); // initially zero memory state 4143 Node* inits = zmem; // accumulating a linearized chain of inits 4144 #ifdef ASSERT 4145 intptr_t first_offset = allocation()->minimum_header_size(); 4146 intptr_t last_init_off = first_offset; // previous init offset 4147 intptr_t last_init_end = first_offset; // previous init offset+size 4148 intptr_t last_tile_end = first_offset; // previous tile offset+size 4149 #endif 4150 intptr_t zeroes_done = header_size; 4151 4152 bool do_zeroing = true; // we might give up if inits are very sparse 4153 int big_init_gaps = 0; // how many large gaps have we seen? 4154 4155 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4156 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4157 4158 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4159 Node* st = in(i); 4160 intptr_t st_off = get_store_offset(st, phase); 4161 if (st_off < 0) 4162 break; // unknown junk in the inits 4163 if (st->in(MemNode::Memory) != zmem) 4164 break; // complicated store chains somehow in list 4165 4166 int st_size = st->as_Store()->memory_size(); 4167 intptr_t next_init_off = st_off + st_size; 4168 4169 if (do_zeroing && zeroes_done < next_init_off) { 4170 // See if this store needs a zero before it or under it. 4171 intptr_t zeroes_needed = st_off; 4172 4173 if (st_size < BytesPerInt) { 4174 // Look for subword stores which only partially initialize words. 4175 // If we find some, we must lay down some word-level zeroes first, 4176 // underneath the subword stores. 4177 // 4178 // Examples: 4179 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4180 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4181 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4182 // 4183 // Note: coalesce_subword_stores may have already done this, 4184 // if it was prompted by constant non-zero subword initializers. 4185 // But this case can still arise with non-constant stores. 4186 4187 intptr_t next_full_store = find_next_fullword_store(i, phase); 4188 4189 // In the examples above: 4190 // in(i) p q r s x y z 4191 // st_off 12 13 14 15 12 13 14 4192 // st_size 1 1 1 1 1 1 1 4193 // next_full_s. 12 16 16 16 16 16 16 4194 // z's_done 12 16 16 16 12 16 12 4195 // z's_needed 12 16 16 16 16 16 16 4196 // zsize 0 0 0 0 4 0 4 4197 if (next_full_store < 0) { 4198 // Conservative tack: Zero to end of current word. 4199 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4200 } else { 4201 // Zero to beginning of next fully initialized word. 4202 // Or, don't zero at all, if we are already in that word. 4203 assert(next_full_store >= zeroes_needed, "must go forward"); 4204 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4205 zeroes_needed = next_full_store; 4206 } 4207 } 4208 4209 if (zeroes_needed > zeroes_done) { 4210 intptr_t zsize = zeroes_needed - zeroes_done; 4211 // Do some incremental zeroing on rawmem, in parallel with inits. 4212 zeroes_done = align_down(zeroes_done, BytesPerInt); 4213 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4214 zeroes_done, zeroes_needed, 4215 phase); 4216 zeroes_done = zeroes_needed; 4217 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4218 do_zeroing = false; // leave the hole, next time 4219 } 4220 } 4221 4222 // Collect the store and move on: 4223 st->set_req(MemNode::Memory, inits); 4224 inits = st; // put it on the linearized chain 4225 set_req(i, zmem); // unhook from previous position 4226 4227 if (zeroes_done == st_off) 4228 zeroes_done = next_init_off; 4229 4230 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4231 4232 #ifdef ASSERT 4233 // Various order invariants. Weaker than stores_are_sane because 4234 // a large constant tile can be filled in by smaller non-constant stores. 4235 assert(st_off >= last_init_off, "inits do not reverse"); 4236 last_init_off = st_off; 4237 const Type* val = NULL; 4238 if (st_size >= BytesPerInt && 4239 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4240 (int)val->basic_type() < (int)T_OBJECT) { 4241 assert(st_off >= last_tile_end, "tiles do not overlap"); 4242 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4243 last_tile_end = MAX2(last_tile_end, next_init_off); 4244 } else { 4245 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4246 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4247 assert(st_off >= last_init_end, "inits do not overlap"); 4248 last_init_end = next_init_off; // it's a non-tile 4249 } 4250 #endif //ASSERT 4251 } 4252 4253 remove_extra_zeroes(); // clear out all the zmems left over 4254 add_req(inits); 4255 4256 if (!(UseTLAB && ZeroTLAB)) { 4257 // If anything remains to be zeroed, zero it all now. 4258 zeroes_done = align_down(zeroes_done, BytesPerInt); 4259 // if it is the last unused 4 bytes of an instance, forget about it 4260 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4261 if (zeroes_done + BytesPerLong >= size_limit) { 4262 AllocateNode* alloc = allocation(); 4263 assert(alloc != NULL, "must be present"); 4264 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 4265 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4266 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4267 if (zeroes_done == k->layout_helper()) 4268 zeroes_done = size_limit; 4269 } 4270 } 4271 if (zeroes_done < size_limit) { 4272 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4273 zeroes_done, size_in_bytes, phase); 4274 } 4275 } 4276 4277 set_complete(phase); 4278 return rawmem; 4279 } 4280 4281 4282 #ifdef ASSERT 4283 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4284 if (is_complete()) 4285 return true; // stores could be anything at this point 4286 assert(allocation() != NULL, "must be present"); 4287 intptr_t last_off = allocation()->minimum_header_size(); 4288 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4289 Node* st = in(i); 4290 intptr_t st_off = get_store_offset(st, phase); 4291 if (st_off < 0) continue; // ignore dead garbage 4292 if (last_off > st_off) { 4293 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4294 this->dump(2); 4295 assert(false, "ascending store offsets"); 4296 return false; 4297 } 4298 last_off = st_off + st->as_Store()->memory_size(); 4299 } 4300 return true; 4301 } 4302 #endif //ASSERT 4303 4304 4305 4306 4307 //============================MergeMemNode===================================== 4308 // 4309 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4310 // contributing store or call operations. Each contributor provides the memory 4311 // state for a particular "alias type" (see Compile::alias_type). For example, 4312 // if a MergeMem has an input X for alias category #6, then any memory reference 4313 // to alias category #6 may use X as its memory state input, as an exact equivalent 4314 // to using the MergeMem as a whole. 4315 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4316 // 4317 // (Here, the <N> notation gives the index of the relevant adr_type.) 4318 // 4319 // In one special case (and more cases in the future), alias categories overlap. 4320 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4321 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4322 // it is exactly equivalent to that state W: 4323 // MergeMem(<Bot>: W) <==> W 4324 // 4325 // Usually, the merge has more than one input. In that case, where inputs 4326 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4327 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4328 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4329 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4330 // 4331 // A merge can take a "wide" memory state as one of its narrow inputs. 4332 // This simply means that the merge observes out only the relevant parts of 4333 // the wide input. That is, wide memory states arriving at narrow merge inputs 4334 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4335 // 4336 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4337 // and that memory slices "leak through": 4338 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4339 // 4340 // But, in such a cascade, repeated memory slices can "block the leak": 4341 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4342 // 4343 // In the last example, Y is not part of the combined memory state of the 4344 // outermost MergeMem. The system must, of course, prevent unschedulable 4345 // memory states from arising, so you can be sure that the state Y is somehow 4346 // a precursor to state Y'. 4347 // 4348 // 4349 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4350 // of each MergeMemNode array are exactly the numerical alias indexes, including 4351 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4352 // Compile::alias_type (and kin) produce and manage these indexes. 4353 // 4354 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4355 // (Note that this provides quick access to the top node inside MergeMem methods, 4356 // without the need to reach out via TLS to Compile::current.) 4357 // 4358 // As a consequence of what was just described, a MergeMem that represents a full 4359 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4360 // containing all alias categories. 4361 // 4362 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4363 // 4364 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4365 // a memory state for the alias type <N>, or else the top node, meaning that 4366 // there is no particular input for that alias type. Note that the length of 4367 // a MergeMem is variable, and may be extended at any time to accommodate new 4368 // memory states at larger alias indexes. When merges grow, they are of course 4369 // filled with "top" in the unused in() positions. 4370 // 4371 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4372 // (Top was chosen because it works smoothly with passes like GCM.) 4373 // 4374 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4375 // the type of random VM bits like TLS references.) Since it is always the 4376 // first non-Bot memory slice, some low-level loops use it to initialize an 4377 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4378 // 4379 // 4380 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4381 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4382 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4383 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4384 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4385 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4386 // 4387 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4388 // really that different from the other memory inputs. An abbreviation called 4389 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4390 // 4391 // 4392 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4393 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4394 // that "emerges though" the base memory will be marked as excluding the alias types 4395 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4396 // 4397 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4398 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4399 // 4400 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4401 // (It is currently unimplemented.) As you can see, the resulting merge is 4402 // actually a disjoint union of memory states, rather than an overlay. 4403 // 4404 4405 //------------------------------MergeMemNode----------------------------------- 4406 Node* MergeMemNode::make_empty_memory() { 4407 Node* empty_memory = (Node*) Compile::current()->top(); 4408 assert(empty_memory->is_top(), "correct sentinel identity"); 4409 return empty_memory; 4410 } 4411 4412 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4413 init_class_id(Class_MergeMem); 4414 // all inputs are nullified in Node::Node(int) 4415 // set_input(0, NULL); // no control input 4416 4417 // Initialize the edges uniformly to top, for starters. 4418 Node* empty_mem = make_empty_memory(); 4419 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4420 init_req(i,empty_mem); 4421 } 4422 assert(empty_memory() == empty_mem, ""); 4423 4424 if( new_base != NULL && new_base->is_MergeMem() ) { 4425 MergeMemNode* mdef = new_base->as_MergeMem(); 4426 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4427 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4428 mms.set_memory(mms.memory2()); 4429 } 4430 assert(base_memory() == mdef->base_memory(), ""); 4431 } else { 4432 set_base_memory(new_base); 4433 } 4434 } 4435 4436 // Make a new, untransformed MergeMem with the same base as 'mem'. 4437 // If mem is itself a MergeMem, populate the result with the same edges. 4438 MergeMemNode* MergeMemNode::make(Node* mem) { 4439 return new MergeMemNode(mem); 4440 } 4441 4442 //------------------------------cmp-------------------------------------------- 4443 uint MergeMemNode::hash() const { return NO_HASH; } 4444 bool MergeMemNode::cmp( const Node &n ) const { 4445 return (&n == this); // Always fail except on self 4446 } 4447 4448 //------------------------------Identity--------------------------------------- 4449 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4450 // Identity if this merge point does not record any interesting memory 4451 // disambiguations. 4452 Node* base_mem = base_memory(); 4453 Node* empty_mem = empty_memory(); 4454 if (base_mem != empty_mem) { // Memory path is not dead? 4455 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4456 Node* mem = in(i); 4457 if (mem != empty_mem && mem != base_mem) { 4458 return this; // Many memory splits; no change 4459 } 4460 } 4461 } 4462 return base_mem; // No memory splits; ID on the one true input 4463 } 4464 4465 //------------------------------Ideal------------------------------------------ 4466 // This method is invoked recursively on chains of MergeMem nodes 4467 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4468 // Remove chain'd MergeMems 4469 // 4470 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4471 // relative to the "in(Bot)". Since we are patching both at the same time, 4472 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4473 // but rewrite each "in(i)" relative to the new "in(Bot)". 4474 Node *progress = NULL; 4475 4476 4477 Node* old_base = base_memory(); 4478 Node* empty_mem = empty_memory(); 4479 if (old_base == empty_mem) 4480 return NULL; // Dead memory path. 4481 4482 MergeMemNode* old_mbase; 4483 if (old_base != NULL && old_base->is_MergeMem()) 4484 old_mbase = old_base->as_MergeMem(); 4485 else 4486 old_mbase = NULL; 4487 Node* new_base = old_base; 4488 4489 // simplify stacked MergeMems in base memory 4490 if (old_mbase) new_base = old_mbase->base_memory(); 4491 4492 // the base memory might contribute new slices beyond my req() 4493 if (old_mbase) grow_to_match(old_mbase); 4494 4495 // Look carefully at the base node if it is a phi. 4496 PhiNode* phi_base; 4497 if (new_base != NULL && new_base->is_Phi()) 4498 phi_base = new_base->as_Phi(); 4499 else 4500 phi_base = NULL; 4501 4502 Node* phi_reg = NULL; 4503 uint phi_len = (uint)-1; 4504 if (phi_base != NULL && !phi_base->is_copy()) { 4505 // do not examine phi if degraded to a copy 4506 phi_reg = phi_base->region(); 4507 phi_len = phi_base->req(); 4508 // see if the phi is unfinished 4509 for (uint i = 1; i < phi_len; i++) { 4510 if (phi_base->in(i) == NULL) { 4511 // incomplete phi; do not look at it yet! 4512 phi_reg = NULL; 4513 phi_len = (uint)-1; 4514 break; 4515 } 4516 } 4517 } 4518 4519 // Note: We do not call verify_sparse on entry, because inputs 4520 // can normalize to the base_memory via subsume_node or similar 4521 // mechanisms. This method repairs that damage. 4522 4523 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4524 4525 // Look at each slice. 4526 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4527 Node* old_in = in(i); 4528 // calculate the old memory value 4529 Node* old_mem = old_in; 4530 if (old_mem == empty_mem) old_mem = old_base; 4531 assert(old_mem == memory_at(i), ""); 4532 4533 // maybe update (reslice) the old memory value 4534 4535 // simplify stacked MergeMems 4536 Node* new_mem = old_mem; 4537 MergeMemNode* old_mmem; 4538 if (old_mem != NULL && old_mem->is_MergeMem()) 4539 old_mmem = old_mem->as_MergeMem(); 4540 else 4541 old_mmem = NULL; 4542 if (old_mmem == this) { 4543 // This can happen if loops break up and safepoints disappear. 4544 // A merge of BotPtr (default) with a RawPtr memory derived from a 4545 // safepoint can be rewritten to a merge of the same BotPtr with 4546 // the BotPtr phi coming into the loop. If that phi disappears 4547 // also, we can end up with a self-loop of the mergemem. 4548 // In general, if loops degenerate and memory effects disappear, 4549 // a mergemem can be left looking at itself. This simply means 4550 // that the mergemem's default should be used, since there is 4551 // no longer any apparent effect on this slice. 4552 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4553 // from start. Update the input to TOP. 4554 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4555 } 4556 else if (old_mmem != NULL) { 4557 new_mem = old_mmem->memory_at(i); 4558 } 4559 // else preceding memory was not a MergeMem 4560 4561 // replace equivalent phis (unfortunately, they do not GVN together) 4562 if (new_mem != NULL && new_mem != new_base && 4563 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4564 if (new_mem->is_Phi()) { 4565 PhiNode* phi_mem = new_mem->as_Phi(); 4566 for (uint i = 1; i < phi_len; i++) { 4567 if (phi_base->in(i) != phi_mem->in(i)) { 4568 phi_mem = NULL; 4569 break; 4570 } 4571 } 4572 if (phi_mem != NULL) { 4573 // equivalent phi nodes; revert to the def 4574 new_mem = new_base; 4575 } 4576 } 4577 } 4578 4579 // maybe store down a new value 4580 Node* new_in = new_mem; 4581 if (new_in == new_base) new_in = empty_mem; 4582 4583 if (new_in != old_in) { 4584 // Warning: Do not combine this "if" with the previous "if" 4585 // A memory slice might have be be rewritten even if it is semantically 4586 // unchanged, if the base_memory value has changed. 4587 set_req(i, new_in); 4588 progress = this; // Report progress 4589 } 4590 } 4591 4592 if (new_base != old_base) { 4593 set_req(Compile::AliasIdxBot, new_base); 4594 // Don't use set_base_memory(new_base), because we need to update du. 4595 assert(base_memory() == new_base, ""); 4596 progress = this; 4597 } 4598 4599 if( base_memory() == this ) { 4600 // a self cycle indicates this memory path is dead 4601 set_req(Compile::AliasIdxBot, empty_mem); 4602 } 4603 4604 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4605 // Recursion must occur after the self cycle check above 4606 if( base_memory()->is_MergeMem() ) { 4607 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4608 Node *m = phase->transform(new_mbase); // Rollup any cycles 4609 if( m != NULL && 4610 (m->is_top() || 4611 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4612 // propagate rollup of dead cycle to self 4613 set_req(Compile::AliasIdxBot, empty_mem); 4614 } 4615 } 4616 4617 if( base_memory() == empty_mem ) { 4618 progress = this; 4619 // Cut inputs during Parse phase only. 4620 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4621 if( !can_reshape ) { 4622 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4623 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4624 } 4625 } 4626 } 4627 4628 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4629 // Check if PhiNode::Ideal's "Split phis through memory merges" 4630 // transform should be attempted. Look for this->phi->this cycle. 4631 uint merge_width = req(); 4632 if (merge_width > Compile::AliasIdxRaw) { 4633 PhiNode* phi = base_memory()->as_Phi(); 4634 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4635 if (phi->in(i) == this) { 4636 phase->is_IterGVN()->_worklist.push(phi); 4637 break; 4638 } 4639 } 4640 } 4641 } 4642 4643 assert(progress || verify_sparse(), "please, no dups of base"); 4644 return progress; 4645 } 4646 4647 //-------------------------set_base_memory------------------------------------- 4648 void MergeMemNode::set_base_memory(Node *new_base) { 4649 Node* empty_mem = empty_memory(); 4650 set_req(Compile::AliasIdxBot, new_base); 4651 assert(memory_at(req()) == new_base, "must set default memory"); 4652 // Clear out other occurrences of new_base: 4653 if (new_base != empty_mem) { 4654 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4655 if (in(i) == new_base) set_req(i, empty_mem); 4656 } 4657 } 4658 } 4659 4660 //------------------------------out_RegMask------------------------------------ 4661 const RegMask &MergeMemNode::out_RegMask() const { 4662 return RegMask::Empty; 4663 } 4664 4665 //------------------------------dump_spec-------------------------------------- 4666 #ifndef PRODUCT 4667 void MergeMemNode::dump_spec(outputStream *st) const { 4668 st->print(" {"); 4669 Node* base_mem = base_memory(); 4670 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4671 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4672 if (mem == base_mem) { st->print(" -"); continue; } 4673 st->print( " N%d:", mem->_idx ); 4674 Compile::current()->get_adr_type(i)->dump_on(st); 4675 } 4676 st->print(" }"); 4677 } 4678 #endif // !PRODUCT 4679 4680 4681 #ifdef ASSERT 4682 static bool might_be_same(Node* a, Node* b) { 4683 if (a == b) return true; 4684 if (!(a->is_Phi() || b->is_Phi())) return false; 4685 // phis shift around during optimization 4686 return true; // pretty stupid... 4687 } 4688 4689 // verify a narrow slice (either incoming or outgoing) 4690 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4691 if (!VerifyAliases) return; // don't bother to verify unless requested 4692 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4693 if (Node::in_dump()) return; // muzzle asserts when printing 4694 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4695 assert(n != NULL, ""); 4696 // Elide intervening MergeMem's 4697 while (n->is_MergeMem()) { 4698 n = n->as_MergeMem()->memory_at(alias_idx); 4699 } 4700 Compile* C = Compile::current(); 4701 const TypePtr* n_adr_type = n->adr_type(); 4702 if (n == m->empty_memory()) { 4703 // Implicit copy of base_memory() 4704 } else if (n_adr_type != TypePtr::BOTTOM) { 4705 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4706 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4707 } else { 4708 // A few places like make_runtime_call "know" that VM calls are narrow, 4709 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4710 bool expected_wide_mem = false; 4711 if (n == m->base_memory()) { 4712 expected_wide_mem = true; 4713 } else if (alias_idx == Compile::AliasIdxRaw || 4714 n == m->memory_at(Compile::AliasIdxRaw)) { 4715 expected_wide_mem = true; 4716 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4717 // memory can "leak through" calls on channels that 4718 // are write-once. Allow this also. 4719 expected_wide_mem = true; 4720 } 4721 assert(expected_wide_mem, "expected narrow slice replacement"); 4722 } 4723 } 4724 #else // !ASSERT 4725 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4726 #endif 4727 4728 4729 //-----------------------------memory_at--------------------------------------- 4730 Node* MergeMemNode::memory_at(uint alias_idx) const { 4731 assert(alias_idx >= Compile::AliasIdxRaw || 4732 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4733 "must avoid base_memory and AliasIdxTop"); 4734 4735 // Otherwise, it is a narrow slice. 4736 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4737 Compile *C = Compile::current(); 4738 if (is_empty_memory(n)) { 4739 // the array is sparse; empty slots are the "top" node 4740 n = base_memory(); 4741 assert(Node::in_dump() 4742 || n == NULL || n->bottom_type() == Type::TOP 4743 || n->adr_type() == NULL // address is TOP 4744 || n->adr_type() == TypePtr::BOTTOM 4745 || n->adr_type() == TypeRawPtr::BOTTOM 4746 || Compile::current()->AliasLevel() == 0, 4747 "must be a wide memory"); 4748 // AliasLevel == 0 if we are organizing the memory states manually. 4749 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4750 } else { 4751 // make sure the stored slice is sane 4752 #ifdef ASSERT 4753 if (VMError::is_error_reported() || Node::in_dump()) { 4754 } else if (might_be_same(n, base_memory())) { 4755 // Give it a pass: It is a mostly harmless repetition of the base. 4756 // This can arise normally from node subsumption during optimization. 4757 } else { 4758 verify_memory_slice(this, alias_idx, n); 4759 } 4760 #endif 4761 } 4762 return n; 4763 } 4764 4765 //---------------------------set_memory_at------------------------------------- 4766 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4767 verify_memory_slice(this, alias_idx, n); 4768 Node* empty_mem = empty_memory(); 4769 if (n == base_memory()) n = empty_mem; // collapse default 4770 uint need_req = alias_idx+1; 4771 if (req() < need_req) { 4772 if (n == empty_mem) return; // already the default, so do not grow me 4773 // grow the sparse array 4774 do { 4775 add_req(empty_mem); 4776 } while (req() < need_req); 4777 } 4778 set_req( alias_idx, n ); 4779 } 4780 4781 4782 4783 //--------------------------iteration_setup------------------------------------ 4784 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4785 if (other != NULL) { 4786 grow_to_match(other); 4787 // invariant: the finite support of mm2 is within mm->req() 4788 #ifdef ASSERT 4789 for (uint i = req(); i < other->req(); i++) { 4790 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4791 } 4792 #endif 4793 } 4794 // Replace spurious copies of base_memory by top. 4795 Node* base_mem = base_memory(); 4796 if (base_mem != NULL && !base_mem->is_top()) { 4797 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4798 if (in(i) == base_mem) 4799 set_req(i, empty_memory()); 4800 } 4801 } 4802 } 4803 4804 //---------------------------grow_to_match------------------------------------- 4805 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4806 Node* empty_mem = empty_memory(); 4807 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4808 // look for the finite support of the other memory 4809 for (uint i = other->req(); --i >= req(); ) { 4810 if (other->in(i) != empty_mem) { 4811 uint new_len = i+1; 4812 while (req() < new_len) add_req(empty_mem); 4813 break; 4814 } 4815 } 4816 } 4817 4818 //---------------------------verify_sparse------------------------------------- 4819 #ifndef PRODUCT 4820 bool MergeMemNode::verify_sparse() const { 4821 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4822 Node* base_mem = base_memory(); 4823 // The following can happen in degenerate cases, since empty==top. 4824 if (is_empty_memory(base_mem)) return true; 4825 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4826 assert(in(i) != NULL, "sane slice"); 4827 if (in(i) == base_mem) return false; // should have been the sentinel value! 4828 } 4829 return true; 4830 } 4831 4832 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4833 Node* n; 4834 n = mm->in(idx); 4835 if (mem == n) return true; // might be empty_memory() 4836 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4837 if (mem == n) return true; 4838 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4839 if (mem == n) return true; 4840 if (n == NULL) break; 4841 } 4842 return false; 4843 } 4844 #endif // !PRODUCT