1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "oops/objArrayKlass.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/compile.hpp" 37 #include "opto/connode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/loopnode.hpp" 40 #include "opto/machnode.hpp" 41 #include "opto/matcher.hpp" 42 #include "opto/memnode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/phaseX.hpp" 46 #include "opto/regmask.hpp" 47 #include "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 uint 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 #if INCLUDE_ZGC 931 if (UseZGC) { 932 if (bottom_type()->make_oopptr() != NULL) { 933 return NULL; 934 } 935 } 936 #endif 937 938 Node* ld_adr = in(MemNode::Address); 939 intptr_t ld_off = 0; 940 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 941 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 942 if (ac != NULL) { 943 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 944 945 Node* mem = ac->in(TypeFunc::Memory); 946 Node* ctl = ac->in(0); 947 Node* src = ac->in(ArrayCopyNode::Src); 948 949 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 950 return NULL; 951 } 952 953 LoadNode* ld = clone()->as_Load(); 954 Node* addp = in(MemNode::Address)->clone(); 955 if (ac->as_ArrayCopy()->is_clonebasic()) { 956 assert(ld_alloc != NULL, "need an alloc"); 957 assert(addp->is_AddP(), "address must be addp"); 958 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern"); 959 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern"); 960 addp->set_req(AddPNode::Base, src->in(AddPNode::Base)); 961 addp->set_req(AddPNode::Address, src->in(AddPNode::Address)); 962 } else { 963 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 964 ac->as_ArrayCopy()->is_copyof_validated() || 965 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 966 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 967 addp->set_req(AddPNode::Base, src); 968 addp->set_req(AddPNode::Address, src); 969 970 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 971 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 972 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 973 uint shift = exact_log2(type2aelembytes(ary_elem)); 974 975 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 976 #ifdef _LP64 977 diff = phase->transform(new ConvI2LNode(diff)); 978 #endif 979 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 980 981 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 982 addp->set_req(AddPNode::Offset, offset); 983 } 984 addp = phase->transform(addp); 985 #ifdef ASSERT 986 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 987 ld->_adr_type = adr_type; 988 #endif 989 ld->set_req(MemNode::Address, addp); 990 ld->set_req(0, ctl); 991 ld->set_req(MemNode::Memory, mem); 992 // load depends on the tests that validate the arraycopy 993 ld->_control_dependency = Pinned; 994 return ld; 995 } 996 return NULL; 997 } 998 999 1000 //---------------------------can_see_stored_value------------------------------ 1001 // This routine exists to make sure this set of tests is done the same 1002 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 1003 // will change the graph shape in a way which makes memory alive twice at the 1004 // same time (uses the Oracle model of aliasing), then some 1005 // LoadXNode::Identity will fold things back to the equivalence-class model 1006 // of aliasing. 1007 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 1008 Node* ld_adr = in(MemNode::Address); 1009 intptr_t ld_off = 0; 1010 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 1011 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 1012 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 1013 // This is more general than load from boxing objects. 1014 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1015 uint alias_idx = atp->index(); 1016 bool final = !atp->is_rewritable(); 1017 Node* result = NULL; 1018 Node* current = st; 1019 // Skip through chains of MemBarNodes checking the MergeMems for 1020 // new states for the slice of this load. Stop once any other 1021 // kind of node is encountered. Loads from final memory can skip 1022 // through any kind of MemBar but normal loads shouldn't skip 1023 // through MemBarAcquire since the could allow them to move out of 1024 // a synchronized region. 1025 while (current->is_Proj()) { 1026 int opc = current->in(0)->Opcode(); 1027 if ((final && (opc == Op_MemBarAcquire || 1028 opc == Op_MemBarAcquireLock || 1029 opc == Op_LoadFence)) || 1030 opc == Op_MemBarRelease || 1031 opc == Op_StoreFence || 1032 opc == Op_MemBarReleaseLock || 1033 opc == Op_MemBarStoreStore || 1034 opc == Op_MemBarCPUOrder) { 1035 Node* mem = current->in(0)->in(TypeFunc::Memory); 1036 if (mem->is_MergeMem()) { 1037 MergeMemNode* merge = mem->as_MergeMem(); 1038 Node* new_st = merge->memory_at(alias_idx); 1039 if (new_st == merge->base_memory()) { 1040 // Keep searching 1041 current = new_st; 1042 continue; 1043 } 1044 // Save the new memory state for the slice and fall through 1045 // to exit. 1046 result = new_st; 1047 } 1048 } 1049 break; 1050 } 1051 if (result != NULL) { 1052 st = result; 1053 } 1054 } 1055 1056 // Loop around twice in the case Load -> Initialize -> Store. 1057 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1058 for (int trip = 0; trip <= 1; trip++) { 1059 1060 if (st->is_Store()) { 1061 Node* st_adr = st->in(MemNode::Address); 1062 if (!phase->eqv(st_adr, ld_adr)) { 1063 // Try harder before giving up... Match raw and non-raw pointers. 1064 intptr_t st_off = 0; 1065 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 1066 if (alloc == NULL) return NULL; 1067 if (alloc != ld_alloc) return NULL; 1068 if (ld_off != st_off) return NULL; 1069 // At this point we have proven something like this setup: 1070 // A = Allocate(...) 1071 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 1072 // S = StoreQ(, AddP(, A.Parm , #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 // Thus, we are able to replace L by V. 1077 } 1078 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1079 if (store_Opcode() != st->Opcode()) 1080 return NULL; 1081 return st->in(MemNode::ValueIn); 1082 } 1083 1084 // A load from a freshly-created object always returns zero. 1085 // (This can happen after LoadNode::Ideal resets the load's memory input 1086 // to find_captured_store, which returned InitializeNode::zero_memory.) 1087 if (st->is_Proj() && st->in(0)->is_Allocate() && 1088 (st->in(0) == ld_alloc) && 1089 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1090 // return a zero value for the load's basic type 1091 // (This is one of the few places where a generic PhaseTransform 1092 // can create new nodes. Think of it as lazily manifesting 1093 // virtually pre-existing constants.) 1094 return phase->zerocon(memory_type()); 1095 } 1096 1097 // A load from an initialization barrier can match a captured store. 1098 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1099 InitializeNode* init = st->in(0)->as_Initialize(); 1100 AllocateNode* alloc = init->allocation(); 1101 if ((alloc != NULL) && (alloc == ld_alloc)) { 1102 // examine a captured store value 1103 st = init->find_captured_store(ld_off, memory_size(), phase); 1104 if (st != NULL) { 1105 continue; // take one more trip around 1106 } 1107 } 1108 } 1109 1110 // Load boxed value from result of valueOf() call is input parameter. 1111 if (this->is_Load() && ld_adr->is_AddP() && 1112 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1113 intptr_t ignore = 0; 1114 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1115 if (base != NULL && base->is_Proj() && 1116 base->as_Proj()->_con == TypeFunc::Parms && 1117 base->in(0)->is_CallStaticJava() && 1118 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1119 return base->in(0)->in(TypeFunc::Parms); 1120 } 1121 } 1122 1123 break; 1124 } 1125 1126 return NULL; 1127 } 1128 1129 //----------------------is_instance_field_load_with_local_phi------------------ 1130 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1131 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1132 in(Address)->is_AddP() ) { 1133 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1134 // Only instances and boxed values. 1135 if( t_oop != NULL && 1136 (t_oop->is_ptr_to_boxed_value() || 1137 t_oop->is_known_instance_field()) && 1138 t_oop->offset() != Type::OffsetBot && 1139 t_oop->offset() != Type::OffsetTop) { 1140 return true; 1141 } 1142 } 1143 return false; 1144 } 1145 1146 //------------------------------Identity--------------------------------------- 1147 // Loads are identity if previous store is to same address 1148 Node* LoadNode::Identity(PhaseGVN* phase) { 1149 // If the previous store-maker is the right kind of Store, and the store is 1150 // to the same address, then we are equal to the value stored. 1151 Node* mem = in(Memory); 1152 Node* value = can_see_stored_value(mem, phase); 1153 if( value ) { 1154 // byte, short & char stores truncate naturally. 1155 // A load has to load the truncated value which requires 1156 // some sort of masking operation and that requires an 1157 // Ideal call instead of an Identity call. 1158 if (memory_size() < BytesPerInt) { 1159 // If the input to the store does not fit with the load's result type, 1160 // it must be truncated via an Ideal call. 1161 if (!phase->type(value)->higher_equal(phase->type(this))) 1162 return this; 1163 } 1164 // (This works even when value is a Con, but LoadNode::Value 1165 // usually runs first, producing the singleton type of the Con.) 1166 return value; 1167 } 1168 1169 // Search for an existing data phi which was generated before for the same 1170 // instance's field to avoid infinite generation of phis in a loop. 1171 Node *region = mem->in(0); 1172 if (is_instance_field_load_with_local_phi(region)) { 1173 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1174 int this_index = phase->C->get_alias_index(addr_t); 1175 int this_offset = addr_t->offset(); 1176 int this_iid = addr_t->instance_id(); 1177 if (!addr_t->is_known_instance() && 1178 addr_t->is_ptr_to_boxed_value()) { 1179 // Use _idx of address base (could be Phi node) for boxed values. 1180 intptr_t ignore = 0; 1181 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1182 if (base == NULL) { 1183 return this; 1184 } 1185 this_iid = base->_idx; 1186 } 1187 const Type* this_type = bottom_type(); 1188 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1189 Node* phi = region->fast_out(i); 1190 if (phi->is_Phi() && phi != mem && 1191 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1192 return phi; 1193 } 1194 } 1195 } 1196 1197 return this; 1198 } 1199 1200 // Construct an equivalent unsigned load. 1201 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1202 BasicType bt = T_ILLEGAL; 1203 const Type* rt = NULL; 1204 switch (Opcode()) { 1205 case Op_LoadUB: return this; 1206 case Op_LoadUS: return this; 1207 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1208 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1209 default: 1210 assert(false, "no unsigned variant: %s", Name()); 1211 return NULL; 1212 } 1213 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1214 raw_adr_type(), rt, bt, _mo, _control_dependency, 1215 is_unaligned_access(), is_mismatched_access()); 1216 } 1217 1218 // Construct an equivalent signed load. 1219 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1220 BasicType bt = T_ILLEGAL; 1221 const Type* rt = NULL; 1222 switch (Opcode()) { 1223 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1224 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1225 case Op_LoadB: // fall through 1226 case Op_LoadS: // fall through 1227 case Op_LoadI: // fall through 1228 case Op_LoadL: return this; 1229 default: 1230 assert(false, "no signed variant: %s", Name()); 1231 return NULL; 1232 } 1233 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1234 raw_adr_type(), rt, bt, _mo, _control_dependency, 1235 is_unaligned_access(), is_mismatched_access()); 1236 } 1237 1238 // We're loading from an object which has autobox behaviour. 1239 // If this object is result of a valueOf call we'll have a phi 1240 // merging a newly allocated object and a load from the cache. 1241 // We want to replace this load with the original incoming 1242 // argument to the valueOf call. 1243 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1244 assert(phase->C->eliminate_boxing(), "sanity"); 1245 intptr_t ignore = 0; 1246 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1247 if ((base == NULL) || base->is_Phi()) { 1248 // Push the loads from the phi that comes from valueOf up 1249 // through it to allow elimination of the loads and the recovery 1250 // of the original value. It is done in split_through_phi(). 1251 return NULL; 1252 } else if (base->is_Load() || 1253 (base->is_DecodeN() && base->in(1)->is_Load())) { 1254 // Eliminate the load of boxed value for integer types from the cache 1255 // array by deriving the value from the index into the array. 1256 // Capture the offset of the load and then reverse the computation. 1257 1258 // Get LoadN node which loads a boxing object from 'cache' array. 1259 if (base->is_DecodeN()) { 1260 base = base->in(1); 1261 } 1262 if (!base->in(Address)->is_AddP()) { 1263 return NULL; // Complex address 1264 } 1265 AddPNode* address = base->in(Address)->as_AddP(); 1266 Node* cache_base = address->in(AddPNode::Base); 1267 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1268 // Get ConP node which is static 'cache' field. 1269 cache_base = cache_base->in(1); 1270 } 1271 if ((cache_base != NULL) && cache_base->is_Con()) { 1272 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1273 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1274 Node* elements[4]; 1275 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1276 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1277 if (count > 0 && elements[0]->is_Con() && 1278 (count == 1 || 1279 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1280 elements[1]->in(2) == phase->intcon(shift)))) { 1281 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1282 // Fetch the box object cache[0] at the base of the array and get its value 1283 ciInstance* box = array->obj_at(0)->as_instance(); 1284 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1285 assert(ik->is_box_klass(), "sanity"); 1286 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1287 if (ik->nof_nonstatic_fields() == 1) { 1288 // This should be true nonstatic_field_at requires calling 1289 // nof_nonstatic_fields so check it anyway 1290 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1291 BasicType bt = c.basic_type(); 1292 // Only integer types have boxing cache. 1293 assert(bt == T_BOOLEAN || bt == T_CHAR || 1294 bt == T_BYTE || bt == T_SHORT || 1295 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1296 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1297 if (cache_low != (int)cache_low) { 1298 return NULL; // should not happen since cache is array indexed by value 1299 } 1300 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1301 if (offset != (int)offset) { 1302 return NULL; // should not happen since cache is array indexed by value 1303 } 1304 // Add up all the offsets making of the address of the load 1305 Node* result = elements[0]; 1306 for (int i = 1; i < count; i++) { 1307 result = phase->transform(new AddXNode(result, elements[i])); 1308 } 1309 // Remove the constant offset from the address and then 1310 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1311 // remove the scaling of the offset to recover the original index. 1312 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1313 // Peel the shift off directly but wrap it in a dummy node 1314 // since Ideal can't return existing nodes 1315 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1316 } else if (result->is_Add() && result->in(2)->is_Con() && 1317 result->in(1)->Opcode() == Op_LShiftX && 1318 result->in(1)->in(2) == phase->intcon(shift)) { 1319 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1320 // but for boxing cache access we know that X<<Z will not overflow 1321 // (there is range check) so we do this optimizatrion by hand here. 1322 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1323 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1324 } else { 1325 result = new RShiftXNode(result, phase->intcon(shift)); 1326 } 1327 #ifdef _LP64 1328 if (bt != T_LONG) { 1329 result = new ConvL2INode(phase->transform(result)); 1330 } 1331 #else 1332 if (bt == T_LONG) { 1333 result = new ConvI2LNode(phase->transform(result)); 1334 } 1335 #endif 1336 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1337 // Need to preserve unboxing load type if it is unsigned. 1338 switch(this->Opcode()) { 1339 case Op_LoadUB: 1340 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1341 break; 1342 case Op_LoadUS: 1343 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1344 break; 1345 } 1346 return result; 1347 } 1348 } 1349 } 1350 } 1351 } 1352 return NULL; 1353 } 1354 1355 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1356 Node* region = phi->in(0); 1357 if (region == NULL) { 1358 return false; // Wait stable graph 1359 } 1360 uint cnt = phi->req(); 1361 for (uint i = 1; i < cnt; i++) { 1362 Node* rc = region->in(i); 1363 if (rc == NULL || phase->type(rc) == Type::TOP) 1364 return false; // Wait stable graph 1365 Node* in = phi->in(i); 1366 if (in == NULL || phase->type(in) == Type::TOP) 1367 return false; // Wait stable graph 1368 } 1369 return true; 1370 } 1371 //------------------------------split_through_phi------------------------------ 1372 // Split instance or boxed field load through Phi. 1373 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1374 Node* mem = in(Memory); 1375 Node* address = in(Address); 1376 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1377 1378 assert((t_oop != NULL) && 1379 (t_oop->is_known_instance_field() || 1380 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1381 1382 Compile* C = phase->C; 1383 intptr_t ignore = 0; 1384 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1385 bool base_is_phi = (base != NULL) && base->is_Phi(); 1386 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1387 (base != NULL) && (base == address->in(AddPNode::Base)) && 1388 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1389 1390 if (!((mem->is_Phi() || base_is_phi) && 1391 (load_boxed_values || t_oop->is_known_instance_field()))) { 1392 return NULL; // memory is not Phi 1393 } 1394 1395 if (mem->is_Phi()) { 1396 if (!stable_phi(mem->as_Phi(), phase)) { 1397 return NULL; // Wait stable graph 1398 } 1399 uint cnt = mem->req(); 1400 // Check for loop invariant memory. 1401 if (cnt == 3) { 1402 for (uint i = 1; i < cnt; i++) { 1403 Node* in = mem->in(i); 1404 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1405 if (m == mem) { 1406 if (i == 1) { 1407 // if the first edge was a loop, check second edge too. 1408 // If both are replaceable - we are in an infinite loop 1409 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1410 if (n == mem) { 1411 break; 1412 } 1413 } 1414 set_req(Memory, mem->in(cnt - i)); 1415 return this; // made change 1416 } 1417 } 1418 } 1419 } 1420 if (base_is_phi) { 1421 if (!stable_phi(base->as_Phi(), phase)) { 1422 return NULL; // Wait stable graph 1423 } 1424 uint cnt = base->req(); 1425 // Check for loop invariant memory. 1426 if (cnt == 3) { 1427 for (uint i = 1; i < cnt; i++) { 1428 if (base->in(i) == base) { 1429 return NULL; // Wait stable graph 1430 } 1431 } 1432 } 1433 } 1434 1435 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1436 1437 // Split through Phi (see original code in loopopts.cpp). 1438 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1439 1440 // Do nothing here if Identity will find a value 1441 // (to avoid infinite chain of value phis generation). 1442 if (!phase->eqv(this, phase->apply_identity(this))) 1443 return NULL; 1444 1445 // Select Region to split through. 1446 Node* region; 1447 if (!base_is_phi) { 1448 assert(mem->is_Phi(), "sanity"); 1449 region = mem->in(0); 1450 // Skip if the region dominates some control edge of the address. 1451 if (!MemNode::all_controls_dominate(address, region)) 1452 return NULL; 1453 } else if (!mem->is_Phi()) { 1454 assert(base_is_phi, "sanity"); 1455 region = base->in(0); 1456 // Skip if the region dominates some control edge of the memory. 1457 if (!MemNode::all_controls_dominate(mem, region)) 1458 return NULL; 1459 } else if (base->in(0) != mem->in(0)) { 1460 assert(base_is_phi && mem->is_Phi(), "sanity"); 1461 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1462 region = base->in(0); 1463 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1464 region = mem->in(0); 1465 } else { 1466 return NULL; // complex graph 1467 } 1468 } else { 1469 assert(base->in(0) == mem->in(0), "sanity"); 1470 region = mem->in(0); 1471 } 1472 1473 const Type* this_type = this->bottom_type(); 1474 int this_index = C->get_alias_index(t_oop); 1475 int this_offset = t_oop->offset(); 1476 int this_iid = t_oop->instance_id(); 1477 if (!t_oop->is_known_instance() && load_boxed_values) { 1478 // Use _idx of address base for boxed values. 1479 this_iid = base->_idx; 1480 } 1481 PhaseIterGVN* igvn = phase->is_IterGVN(); 1482 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1483 for (uint i = 1; i < region->req(); i++) { 1484 Node* x; 1485 Node* the_clone = NULL; 1486 if (region->in(i) == C->top()) { 1487 x = C->top(); // Dead path? Use a dead data op 1488 } else { 1489 x = this->clone(); // Else clone up the data op 1490 the_clone = x; // Remember for possible deletion. 1491 // Alter data node to use pre-phi inputs 1492 if (this->in(0) == region) { 1493 x->set_req(0, region->in(i)); 1494 } else { 1495 x->set_req(0, NULL); 1496 } 1497 if (mem->is_Phi() && (mem->in(0) == region)) { 1498 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1499 } 1500 if (address->is_Phi() && address->in(0) == region) { 1501 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1502 } 1503 if (base_is_phi && (base->in(0) == region)) { 1504 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1505 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1506 x->set_req(Address, adr_x); 1507 } 1508 } 1509 // Check for a 'win' on some paths 1510 const Type *t = x->Value(igvn); 1511 1512 bool singleton = t->singleton(); 1513 1514 // See comments in PhaseIdealLoop::split_thru_phi(). 1515 if (singleton && t == Type::TOP) { 1516 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1517 } 1518 1519 if (singleton) { 1520 x = igvn->makecon(t); 1521 } else { 1522 // We now call Identity to try to simplify the cloned node. 1523 // Note that some Identity methods call phase->type(this). 1524 // Make sure that the type array is big enough for 1525 // our new node, even though we may throw the node away. 1526 // (This tweaking with igvn only works because x is a new node.) 1527 igvn->set_type(x, t); 1528 // If x is a TypeNode, capture any more-precise type permanently into Node 1529 // otherwise it will be not updated during igvn->transform since 1530 // igvn->type(x) is set to x->Value() already. 1531 x->raise_bottom_type(t); 1532 Node *y = igvn->apply_identity(x); 1533 if (y != x) { 1534 x = y; 1535 } else { 1536 y = igvn->hash_find_insert(x); 1537 if (y) { 1538 x = y; 1539 } else { 1540 // Else x is a new node we are keeping 1541 // We do not need register_new_node_with_optimizer 1542 // because set_type has already been called. 1543 igvn->_worklist.push(x); 1544 } 1545 } 1546 } 1547 if (x != the_clone && the_clone != NULL) { 1548 igvn->remove_dead_node(the_clone); 1549 } 1550 phi->set_req(i, x); 1551 } 1552 // Record Phi 1553 igvn->register_new_node_with_optimizer(phi); 1554 return phi; 1555 } 1556 1557 //------------------------------Ideal------------------------------------------ 1558 // If the load is from Field memory and the pointer is non-null, it might be possible to 1559 // zero out the control input. 1560 // If the offset is constant and the base is an object allocation, 1561 // try to hook me up to the exact initializing store. 1562 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1563 Node* p = MemNode::Ideal_common(phase, can_reshape); 1564 if (p) return (p == NodeSentinel) ? NULL : p; 1565 1566 Node* ctrl = in(MemNode::Control); 1567 Node* address = in(MemNode::Address); 1568 bool progress = false; 1569 1570 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1571 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1572 1573 // Skip up past a SafePoint control. Cannot do this for Stores because 1574 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1575 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1576 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1577 !addr_mark ) { 1578 ctrl = ctrl->in(0); 1579 set_req(MemNode::Control,ctrl); 1580 progress = true; 1581 } 1582 1583 intptr_t ignore = 0; 1584 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1585 if (base != NULL 1586 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1587 // Check for useless control edge in some common special cases 1588 if (in(MemNode::Control) != NULL 1589 && can_remove_control() 1590 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1591 && all_controls_dominate(base, phase->C->start())) { 1592 // A method-invariant, non-null address (constant or 'this' argument). 1593 set_req(MemNode::Control, NULL); 1594 progress = true; 1595 } 1596 } 1597 1598 Node* mem = in(MemNode::Memory); 1599 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1600 1601 if (can_reshape && (addr_t != NULL)) { 1602 // try to optimize our memory input 1603 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1604 if (opt_mem != mem) { 1605 set_req(MemNode::Memory, opt_mem); 1606 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1607 return this; 1608 } 1609 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1610 if ((t_oop != NULL) && 1611 (t_oop->is_known_instance_field() || 1612 t_oop->is_ptr_to_boxed_value())) { 1613 PhaseIterGVN *igvn = phase->is_IterGVN(); 1614 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1615 // Delay this transformation until memory Phi is processed. 1616 phase->is_IterGVN()->_worklist.push(this); 1617 return NULL; 1618 } 1619 // Split instance field load through Phi. 1620 Node* result = split_through_phi(phase); 1621 if (result != NULL) return result; 1622 1623 if (t_oop->is_ptr_to_boxed_value()) { 1624 Node* result = eliminate_autobox(phase); 1625 if (result != NULL) return result; 1626 } 1627 } 1628 } 1629 1630 // Is there a dominating load that loads the same value? Leave 1631 // anything that is not a load of a field/array element (like 1632 // barriers etc.) alone 1633 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) { 1634 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1635 Node *use = mem->fast_out(i); 1636 if (use != this && 1637 use->Opcode() == Opcode() && 1638 use->in(0) != NULL && 1639 use->in(0) != in(0) && 1640 use->in(Address) == in(Address)) { 1641 Node* ctl = in(0); 1642 for (int i = 0; i < 10 && ctl != NULL; i++) { 1643 ctl = IfNode::up_one_dom(ctl); 1644 if (ctl == use->in(0)) { 1645 set_req(0, use->in(0)); 1646 return this; 1647 } 1648 } 1649 } 1650 } 1651 } 1652 1653 // Check for prior store with a different base or offset; make Load 1654 // independent. Skip through any number of them. Bail out if the stores 1655 // are in an endless dead cycle and report no progress. This is a key 1656 // transform for Reflection. However, if after skipping through the Stores 1657 // we can't then fold up against a prior store do NOT do the transform as 1658 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1659 // array memory alive twice: once for the hoisted Load and again after the 1660 // bypassed Store. This situation only works if EVERYBODY who does 1661 // anti-dependence work knows how to bypass. I.e. we need all 1662 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1663 // the alias index stuff. So instead, peek through Stores and IFF we can 1664 // fold up, do so. 1665 Node* prev_mem = find_previous_store(phase); 1666 if (prev_mem != NULL) { 1667 Node* value = can_see_arraycopy_value(prev_mem, phase); 1668 if (value != NULL) { 1669 return value; 1670 } 1671 } 1672 // Steps (a), (b): Walk past independent stores to find an exact match. 1673 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1674 // (c) See if we can fold up on the spot, but don't fold up here. 1675 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1676 // just return a prior value, which is done by Identity calls. 1677 if (can_see_stored_value(prev_mem, phase)) { 1678 // Make ready for step (d): 1679 set_req(MemNode::Memory, prev_mem); 1680 return this; 1681 } 1682 } 1683 1684 return progress ? this : NULL; 1685 } 1686 1687 // Helper to recognize certain Klass fields which are invariant across 1688 // some group of array types (e.g., int[] or all T[] where T < Object). 1689 const Type* 1690 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1691 ciKlass* klass) const { 1692 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1693 // The field is Klass::_modifier_flags. Return its (constant) value. 1694 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1695 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1696 return TypeInt::make(klass->modifier_flags()); 1697 } 1698 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1699 // The field is Klass::_access_flags. Return its (constant) value. 1700 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1701 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1702 return TypeInt::make(klass->access_flags()); 1703 } 1704 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1705 // The field is Klass::_layout_helper. Return its constant value if known. 1706 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1707 return TypeInt::make(klass->layout_helper()); 1708 } 1709 1710 // No match. 1711 return NULL; 1712 } 1713 1714 //------------------------------Value----------------------------------------- 1715 const Type* LoadNode::Value(PhaseGVN* phase) const { 1716 // Either input is TOP ==> the result is TOP 1717 Node* mem = in(MemNode::Memory); 1718 const Type *t1 = phase->type(mem); 1719 if (t1 == Type::TOP) return Type::TOP; 1720 Node* adr = in(MemNode::Address); 1721 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1722 if (tp == NULL || tp->empty()) return Type::TOP; 1723 int off = tp->offset(); 1724 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1725 Compile* C = phase->C; 1726 1727 // Try to guess loaded type from pointer type 1728 if (tp->isa_aryptr()) { 1729 const TypeAryPtr* ary = tp->is_aryptr(); 1730 const Type* t = ary->elem(); 1731 1732 // Determine whether the reference is beyond the header or not, by comparing 1733 // the offset against the offset of the start of the array's data. 1734 // Different array types begin at slightly different offsets (12 vs. 16). 1735 // We choose T_BYTE as an example base type that is least restrictive 1736 // as to alignment, which will therefore produce the smallest 1737 // possible base offset. 1738 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1739 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1740 1741 // Try to constant-fold a stable array element. 1742 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1743 // Make sure the reference is not into the header and the offset is constant 1744 ciObject* aobj = ary->const_oop(); 1745 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1746 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1747 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1748 stable_dimension, 1749 memory_type(), is_unsigned()); 1750 if (con_type != NULL) { 1751 return con_type; 1752 } 1753 } 1754 } 1755 1756 // Don't do this for integer types. There is only potential profit if 1757 // the element type t is lower than _type; that is, for int types, if _type is 1758 // more restrictive than t. This only happens here if one is short and the other 1759 // char (both 16 bits), and in those cases we've made an intentional decision 1760 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1761 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1762 // 1763 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1764 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1765 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1766 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1767 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1768 // In fact, that could have been the original type of p1, and p1 could have 1769 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1770 // expression (LShiftL quux 3) independently optimized to the constant 8. 1771 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1772 && (_type->isa_vect() == NULL) 1773 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1774 // t might actually be lower than _type, if _type is a unique 1775 // concrete subclass of abstract class t. 1776 if (off_beyond_header) { // is the offset beyond the header? 1777 const Type* jt = t->join_speculative(_type); 1778 // In any case, do not allow the join, per se, to empty out the type. 1779 if (jt->empty() && !t->empty()) { 1780 // This can happen if a interface-typed array narrows to a class type. 1781 jt = _type; 1782 } 1783 #ifdef ASSERT 1784 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1785 // The pointers in the autobox arrays are always non-null 1786 Node* base = adr->in(AddPNode::Base); 1787 if ((base != NULL) && base->is_DecodeN()) { 1788 // Get LoadN node which loads IntegerCache.cache field 1789 base = base->in(1); 1790 } 1791 if ((base != NULL) && base->is_Con()) { 1792 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1793 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1794 // It could be narrow oop 1795 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1796 } 1797 } 1798 } 1799 #endif 1800 return jt; 1801 } 1802 } 1803 } else if (tp->base() == Type::InstPtr) { 1804 assert( off != Type::OffsetBot || 1805 // arrays can be cast to Objects 1806 tp->is_oopptr()->klass()->is_java_lang_Object() || 1807 // unsafe field access may not have a constant offset 1808 C->has_unsafe_access(), 1809 "Field accesses must be precise" ); 1810 // For oop loads, we expect the _type to be precise. 1811 1812 // Optimize loads from constant fields. 1813 const TypeInstPtr* tinst = tp->is_instptr(); 1814 ciObject* const_oop = tinst->const_oop(); 1815 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1816 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type()); 1817 if (con_type != NULL) { 1818 return con_type; 1819 } 1820 } 1821 } else if (tp->base() == Type::KlassPtr) { 1822 assert( off != Type::OffsetBot || 1823 // arrays can be cast to Objects 1824 tp->is_klassptr()->klass()->is_java_lang_Object() || 1825 // also allow array-loading from the primary supertype 1826 // array during subtype checks 1827 Opcode() == Op_LoadKlass, 1828 "Field accesses must be precise" ); 1829 // For klass/static loads, we expect the _type to be precise 1830 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) { 1831 /* With mirrors being an indirect in the Klass* 1832 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 1833 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 1834 * 1835 * So check the type and klass of the node before the LoadP. 1836 */ 1837 Node* adr2 = adr->in(MemNode::Address); 1838 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1839 if (tkls != NULL && !StressReflectiveCode) { 1840 ciKlass* klass = tkls->klass(); 1841 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1842 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1843 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1844 return TypeInstPtr::make(klass->java_mirror()); 1845 } 1846 } 1847 } 1848 1849 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1850 if (tkls != NULL && !StressReflectiveCode) { 1851 ciKlass* klass = tkls->klass(); 1852 if (klass->is_loaded() && tkls->klass_is_exact()) { 1853 // We are loading a field from a Klass metaobject whose identity 1854 // is known at compile time (the type is "exact" or "precise"). 1855 // Check for fields we know are maintained as constants by the VM. 1856 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1857 // The field is Klass::_super_check_offset. Return its (constant) value. 1858 // (Folds up type checking code.) 1859 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1860 return TypeInt::make(klass->super_check_offset()); 1861 } 1862 // Compute index into primary_supers array 1863 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1864 // Check for overflowing; use unsigned compare to handle the negative case. 1865 if( depth < ciKlass::primary_super_limit() ) { 1866 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1867 // (Folds up type checking code.) 1868 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1869 ciKlass *ss = klass->super_of_depth(depth); 1870 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1871 } 1872 const Type* aift = load_array_final_field(tkls, klass); 1873 if (aift != NULL) return aift; 1874 } 1875 1876 // We can still check if we are loading from the primary_supers array at a 1877 // shallow enough depth. Even though the klass is not exact, entries less 1878 // than or equal to its super depth are correct. 1879 if (klass->is_loaded() ) { 1880 ciType *inner = klass; 1881 while( inner->is_obj_array_klass() ) 1882 inner = inner->as_obj_array_klass()->base_element_type(); 1883 if( inner->is_instance_klass() && 1884 !inner->as_instance_klass()->flags().is_interface() ) { 1885 // Compute index into primary_supers array 1886 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1887 // Check for overflowing; use unsigned compare to handle the negative case. 1888 if( depth < ciKlass::primary_super_limit() && 1889 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1890 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1891 // (Folds up type checking code.) 1892 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1893 ciKlass *ss = klass->super_of_depth(depth); 1894 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1895 } 1896 } 1897 } 1898 1899 // If the type is enough to determine that the thing is not an array, 1900 // we can give the layout_helper a positive interval type. 1901 // This will help short-circuit some reflective code. 1902 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1903 && !klass->is_array_klass() // not directly typed as an array 1904 && !klass->is_interface() // specifically not Serializable & Cloneable 1905 && !klass->is_java_lang_Object() // not the supertype of all T[] 1906 ) { 1907 // Note: When interfaces are reliable, we can narrow the interface 1908 // test to (klass != Serializable && klass != Cloneable). 1909 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1910 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1911 // The key property of this type is that it folds up tests 1912 // for array-ness, since it proves that the layout_helper is positive. 1913 // Thus, a generic value like the basic object layout helper works fine. 1914 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1915 } 1916 } 1917 1918 // If we are loading from a freshly-allocated object, produce a zero, 1919 // if the load is provably beyond the header of the object. 1920 // (Also allow a variable load from a fresh array to produce zero.) 1921 const TypeOopPtr *tinst = tp->isa_oopptr(); 1922 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1923 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1924 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1925 Node* value = can_see_stored_value(mem,phase); 1926 if (value != NULL && value->is_Con()) { 1927 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1928 return value->bottom_type(); 1929 } 1930 } 1931 1932 if (is_instance) { 1933 // If we have an instance type and our memory input is the 1934 // programs's initial memory state, there is no matching store, 1935 // so just return a zero of the appropriate type 1936 Node *mem = in(MemNode::Memory); 1937 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1938 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1939 return Type::get_zero_type(_type->basic_type()); 1940 } 1941 } 1942 return _type; 1943 } 1944 1945 //------------------------------match_edge------------------------------------- 1946 // Do we Match on this edge index or not? Match only the address. 1947 uint LoadNode::match_edge(uint idx) const { 1948 return idx == MemNode::Address; 1949 } 1950 1951 //--------------------------LoadBNode::Ideal-------------------------------------- 1952 // 1953 // If the previous store is to the same address as this load, 1954 // and the value stored was larger than a byte, replace this load 1955 // with the value stored truncated to a byte. If no truncation is 1956 // needed, the replacement is done in LoadNode::Identity(). 1957 // 1958 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1959 Node* mem = in(MemNode::Memory); 1960 Node* value = can_see_stored_value(mem,phase); 1961 if( value && !phase->type(value)->higher_equal( _type ) ) { 1962 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1963 return new RShiftINode(result, phase->intcon(24)); 1964 } 1965 // Identity call will handle the case where truncation is not needed. 1966 return LoadNode::Ideal(phase, can_reshape); 1967 } 1968 1969 const Type* LoadBNode::Value(PhaseGVN* phase) const { 1970 Node* mem = in(MemNode::Memory); 1971 Node* value = can_see_stored_value(mem,phase); 1972 if (value != NULL && value->is_Con() && 1973 !value->bottom_type()->higher_equal(_type)) { 1974 // If the input to the store does not fit with the load's result type, 1975 // it must be truncated. We can't delay until Ideal call since 1976 // a singleton Value is needed for split_thru_phi optimization. 1977 int con = value->get_int(); 1978 return TypeInt::make((con << 24) >> 24); 1979 } 1980 return LoadNode::Value(phase); 1981 } 1982 1983 //--------------------------LoadUBNode::Ideal------------------------------------- 1984 // 1985 // If the previous store is to the same address as this load, 1986 // and the value stored was larger than a byte, replace this load 1987 // with the value stored truncated to a byte. If no truncation is 1988 // needed, the replacement is done in LoadNode::Identity(). 1989 // 1990 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1991 Node* mem = in(MemNode::Memory); 1992 Node* value = can_see_stored_value(mem, phase); 1993 if (value && !phase->type(value)->higher_equal(_type)) 1994 return new AndINode(value, phase->intcon(0xFF)); 1995 // Identity call will handle the case where truncation is not needed. 1996 return LoadNode::Ideal(phase, can_reshape); 1997 } 1998 1999 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2000 Node* mem = in(MemNode::Memory); 2001 Node* value = can_see_stored_value(mem,phase); 2002 if (value != NULL && value->is_Con() && 2003 !value->bottom_type()->higher_equal(_type)) { 2004 // If the input to the store does not fit with the load's result type, 2005 // it must be truncated. We can't delay until Ideal call since 2006 // a singleton Value is needed for split_thru_phi optimization. 2007 int con = value->get_int(); 2008 return TypeInt::make(con & 0xFF); 2009 } 2010 return LoadNode::Value(phase); 2011 } 2012 2013 //--------------------------LoadUSNode::Ideal------------------------------------- 2014 // 2015 // If the previous store is to the same address as this load, 2016 // and the value stored was larger than a char, replace this load 2017 // with the value stored truncated to a char. If no truncation is 2018 // needed, the replacement is done in LoadNode::Identity(). 2019 // 2020 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2021 Node* mem = in(MemNode::Memory); 2022 Node* value = can_see_stored_value(mem,phase); 2023 if( value && !phase->type(value)->higher_equal( _type ) ) 2024 return new AndINode(value,phase->intcon(0xFFFF)); 2025 // Identity call will handle the case where truncation is not needed. 2026 return LoadNode::Ideal(phase, can_reshape); 2027 } 2028 2029 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2030 Node* mem = in(MemNode::Memory); 2031 Node* value = can_see_stored_value(mem,phase); 2032 if (value != NULL && value->is_Con() && 2033 !value->bottom_type()->higher_equal(_type)) { 2034 // If the input to the store does not fit with the load's result type, 2035 // it must be truncated. We can't delay until Ideal call since 2036 // a singleton Value is needed for split_thru_phi optimization. 2037 int con = value->get_int(); 2038 return TypeInt::make(con & 0xFFFF); 2039 } 2040 return LoadNode::Value(phase); 2041 } 2042 2043 //--------------------------LoadSNode::Ideal-------------------------------------- 2044 // 2045 // If the previous store is to the same address as this load, 2046 // and the value stored was larger than a short, replace this load 2047 // with the value stored truncated to a short. If no truncation is 2048 // needed, the replacement is done in LoadNode::Identity(). 2049 // 2050 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2051 Node* mem = in(MemNode::Memory); 2052 Node* value = can_see_stored_value(mem,phase); 2053 if( value && !phase->type(value)->higher_equal( _type ) ) { 2054 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 2055 return new RShiftINode(result, phase->intcon(16)); 2056 } 2057 // Identity call will handle the case where truncation is not needed. 2058 return LoadNode::Ideal(phase, can_reshape); 2059 } 2060 2061 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2062 Node* mem = in(MemNode::Memory); 2063 Node* value = can_see_stored_value(mem,phase); 2064 if (value != NULL && value->is_Con() && 2065 !value->bottom_type()->higher_equal(_type)) { 2066 // If the input to the store does not fit with the load's result type, 2067 // it must be truncated. We can't delay until Ideal call since 2068 // a singleton Value is needed for split_thru_phi optimization. 2069 int con = value->get_int(); 2070 return TypeInt::make((con << 16) >> 16); 2071 } 2072 return LoadNode::Value(phase); 2073 } 2074 2075 //============================================================================= 2076 //----------------------------LoadKlassNode::make------------------------------ 2077 // Polymorphic factory method: 2078 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2079 // sanity check the alias category against the created node type 2080 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2081 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2082 #ifdef _LP64 2083 if (adr_type->is_ptr_to_narrowklass()) { 2084 assert(UseCompressedClassPointers, "no compressed klasses"); 2085 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2086 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2087 } 2088 #endif 2089 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2090 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2091 } 2092 2093 //------------------------------Value------------------------------------------ 2094 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2095 return klass_value_common(phase); 2096 } 2097 2098 // In most cases, LoadKlassNode does not have the control input set. If the control 2099 // input is set, it must not be removed (by LoadNode::Ideal()). 2100 bool LoadKlassNode::can_remove_control() const { 2101 return false; 2102 } 2103 2104 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2105 // Either input is TOP ==> the result is TOP 2106 const Type *t1 = phase->type( in(MemNode::Memory) ); 2107 if (t1 == Type::TOP) return Type::TOP; 2108 Node *adr = in(MemNode::Address); 2109 const Type *t2 = phase->type( adr ); 2110 if (t2 == Type::TOP) return Type::TOP; 2111 const TypePtr *tp = t2->is_ptr(); 2112 if (TypePtr::above_centerline(tp->ptr()) || 2113 tp->ptr() == TypePtr::Null) return Type::TOP; 2114 2115 // Return a more precise klass, if possible 2116 const TypeInstPtr *tinst = tp->isa_instptr(); 2117 if (tinst != NULL) { 2118 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2119 int offset = tinst->offset(); 2120 if (ik == phase->C->env()->Class_klass() 2121 && (offset == java_lang_Class::klass_offset_in_bytes() || 2122 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2123 // We are loading a special hidden field from a Class mirror object, 2124 // the field which points to the VM's Klass metaobject. 2125 ciType* t = tinst->java_mirror_type(); 2126 // java_mirror_type returns non-null for compile-time Class constants. 2127 if (t != NULL) { 2128 // constant oop => constant klass 2129 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2130 if (t->is_void()) { 2131 // We cannot create a void array. Since void is a primitive type return null 2132 // klass. Users of this result need to do a null check on the returned klass. 2133 return TypePtr::NULL_PTR; 2134 } 2135 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2136 } 2137 if (!t->is_klass()) { 2138 // a primitive Class (e.g., int.class) has NULL for a klass field 2139 return TypePtr::NULL_PTR; 2140 } 2141 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2142 return TypeKlassPtr::make(t->as_klass()); 2143 } 2144 // non-constant mirror, so we can't tell what's going on 2145 } 2146 if( !ik->is_loaded() ) 2147 return _type; // Bail out if not loaded 2148 if (offset == oopDesc::klass_offset_in_bytes()) { 2149 if (tinst->klass_is_exact()) { 2150 return TypeKlassPtr::make(ik); 2151 } 2152 // See if we can become precise: no subklasses and no interface 2153 // (Note: We need to support verified interfaces.) 2154 if (!ik->is_interface() && !ik->has_subklass()) { 2155 //assert(!UseExactTypes, "this code should be useless with exact types"); 2156 // Add a dependence; if any subclass added we need to recompile 2157 if (!ik->is_final()) { 2158 // %%% should use stronger assert_unique_concrete_subtype instead 2159 phase->C->dependencies()->assert_leaf_type(ik); 2160 } 2161 // Return precise klass 2162 return TypeKlassPtr::make(ik); 2163 } 2164 2165 // Return root of possible klass 2166 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2167 } 2168 } 2169 2170 // Check for loading klass from an array 2171 const TypeAryPtr *tary = tp->isa_aryptr(); 2172 if( tary != NULL ) { 2173 ciKlass *tary_klass = tary->klass(); 2174 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2175 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2176 if (tary->klass_is_exact()) { 2177 return TypeKlassPtr::make(tary_klass); 2178 } 2179 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2180 // If the klass is an object array, we defer the question to the 2181 // array component klass. 2182 if( ak->is_obj_array_klass() ) { 2183 assert( ak->is_loaded(), "" ); 2184 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2185 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2186 ciInstanceKlass* ik = base_k->as_instance_klass(); 2187 // See if we can become precise: no subklasses and no interface 2188 if (!ik->is_interface() && !ik->has_subklass()) { 2189 //assert(!UseExactTypes, "this code should be useless with exact types"); 2190 // Add a dependence; if any subclass added we need to recompile 2191 if (!ik->is_final()) { 2192 phase->C->dependencies()->assert_leaf_type(ik); 2193 } 2194 // Return precise array klass 2195 return TypeKlassPtr::make(ak); 2196 } 2197 } 2198 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2199 } else { // Found a type-array? 2200 //assert(!UseExactTypes, "this code should be useless with exact types"); 2201 assert( ak->is_type_array_klass(), "" ); 2202 return TypeKlassPtr::make(ak); // These are always precise 2203 } 2204 } 2205 } 2206 2207 // Check for loading klass from an array klass 2208 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2209 if (tkls != NULL && !StressReflectiveCode) { 2210 ciKlass* klass = tkls->klass(); 2211 if( !klass->is_loaded() ) 2212 return _type; // Bail out if not loaded 2213 if( klass->is_obj_array_klass() && 2214 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2215 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2216 // // Always returning precise element type is incorrect, 2217 // // e.g., element type could be object and array may contain strings 2218 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2219 2220 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2221 // according to the element type's subclassing. 2222 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2223 } 2224 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2225 tkls->offset() == in_bytes(Klass::super_offset())) { 2226 ciKlass* sup = klass->as_instance_klass()->super(); 2227 // The field is Klass::_super. Return its (constant) value. 2228 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2229 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2230 } 2231 } 2232 2233 // Bailout case 2234 return LoadNode::Value(phase); 2235 } 2236 2237 //------------------------------Identity--------------------------------------- 2238 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2239 // Also feed through the klass in Allocate(...klass...)._klass. 2240 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2241 return klass_identity_common(phase); 2242 } 2243 2244 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2245 Node* x = LoadNode::Identity(phase); 2246 if (x != this) return x; 2247 2248 // Take apart the address into an oop and and offset. 2249 // Return 'this' if we cannot. 2250 Node* adr = in(MemNode::Address); 2251 intptr_t offset = 0; 2252 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2253 if (base == NULL) return this; 2254 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2255 if (toop == NULL) return this; 2256 2257 // Step over potential GC barrier for OopHandle resolve 2258 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2259 if (bs->is_gc_barrier_node(base)) { 2260 base = bs->step_over_gc_barrier(base); 2261 } 2262 2263 // We can fetch the klass directly through an AllocateNode. 2264 // This works even if the klass is not constant (clone or newArray). 2265 if (offset == oopDesc::klass_offset_in_bytes()) { 2266 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2267 if (allocated_klass != NULL) { 2268 return allocated_klass; 2269 } 2270 } 2271 2272 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2273 // See inline_native_Class_query for occurrences of these patterns. 2274 // Java Example: x.getClass().isAssignableFrom(y) 2275 // 2276 // This improves reflective code, often making the Class 2277 // mirror go completely dead. (Current exception: Class 2278 // mirrors may appear in debug info, but we could clean them out by 2279 // introducing a new debug info operator for Klass.java_mirror). 2280 2281 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2282 && offset == java_lang_Class::klass_offset_in_bytes()) { 2283 if (base->is_Load()) { 2284 Node* base2 = base->in(MemNode::Address); 2285 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2286 Node* adr2 = base2->in(MemNode::Address); 2287 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2288 if (tkls != NULL && !tkls->empty() 2289 && (tkls->klass()->is_instance_klass() || 2290 tkls->klass()->is_array_klass()) 2291 && adr2->is_AddP() 2292 ) { 2293 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2294 if (tkls->offset() == mirror_field) { 2295 return adr2->in(AddPNode::Base); 2296 } 2297 } 2298 } 2299 } 2300 } 2301 2302 return this; 2303 } 2304 2305 2306 //------------------------------Value------------------------------------------ 2307 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2308 const Type *t = klass_value_common(phase); 2309 if (t == Type::TOP) 2310 return t; 2311 2312 return t->make_narrowklass(); 2313 } 2314 2315 //------------------------------Identity--------------------------------------- 2316 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2317 // Also feed through the klass in Allocate(...klass...)._klass. 2318 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2319 Node *x = klass_identity_common(phase); 2320 2321 const Type *t = phase->type( x ); 2322 if( t == Type::TOP ) return x; 2323 if( t->isa_narrowklass()) return x; 2324 assert (!t->isa_narrowoop(), "no narrow oop here"); 2325 2326 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2327 } 2328 2329 //------------------------------Value----------------------------------------- 2330 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2331 // Either input is TOP ==> the result is TOP 2332 const Type *t1 = phase->type( in(MemNode::Memory) ); 2333 if( t1 == Type::TOP ) return Type::TOP; 2334 Node *adr = in(MemNode::Address); 2335 const Type *t2 = phase->type( adr ); 2336 if( t2 == Type::TOP ) return Type::TOP; 2337 const TypePtr *tp = t2->is_ptr(); 2338 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2339 const TypeAryPtr *tap = tp->isa_aryptr(); 2340 if( !tap ) return _type; 2341 return tap->size(); 2342 } 2343 2344 //-------------------------------Ideal--------------------------------------- 2345 // Feed through the length in AllocateArray(...length...)._length. 2346 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2347 Node* p = MemNode::Ideal_common(phase, can_reshape); 2348 if (p) return (p == NodeSentinel) ? NULL : p; 2349 2350 // Take apart the address into an oop and and offset. 2351 // Return 'this' if we cannot. 2352 Node* adr = in(MemNode::Address); 2353 intptr_t offset = 0; 2354 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2355 if (base == NULL) return NULL; 2356 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2357 if (tary == NULL) return NULL; 2358 2359 // We can fetch the length directly through an AllocateArrayNode. 2360 // This works even if the length is not constant (clone or newArray). 2361 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2362 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2363 if (alloc != NULL) { 2364 Node* allocated_length = alloc->Ideal_length(); 2365 Node* len = alloc->make_ideal_length(tary, phase); 2366 if (allocated_length != len) { 2367 // New CastII improves on this. 2368 return len; 2369 } 2370 } 2371 } 2372 2373 return NULL; 2374 } 2375 2376 //------------------------------Identity--------------------------------------- 2377 // Feed through the length in AllocateArray(...length...)._length. 2378 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2379 Node* x = LoadINode::Identity(phase); 2380 if (x != this) return x; 2381 2382 // Take apart the address into an oop and and offset. 2383 // Return 'this' if we cannot. 2384 Node* adr = in(MemNode::Address); 2385 intptr_t offset = 0; 2386 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2387 if (base == NULL) return this; 2388 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2389 if (tary == NULL) return this; 2390 2391 // We can fetch the length directly through an AllocateArrayNode. 2392 // This works even if the length is not constant (clone or newArray). 2393 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2394 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2395 if (alloc != NULL) { 2396 Node* allocated_length = alloc->Ideal_length(); 2397 // Do not allow make_ideal_length to allocate a CastII node. 2398 Node* len = alloc->make_ideal_length(tary, phase, false); 2399 if (allocated_length == len) { 2400 // Return allocated_length only if it would not be improved by a CastII. 2401 return allocated_length; 2402 } 2403 } 2404 } 2405 2406 return this; 2407 2408 } 2409 2410 //============================================================================= 2411 //---------------------------StoreNode::make----------------------------------- 2412 // Polymorphic factory method: 2413 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2414 assert((mo == unordered || mo == release), "unexpected"); 2415 Compile* C = gvn.C; 2416 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2417 ctl != NULL, "raw memory operations should have control edge"); 2418 2419 switch (bt) { 2420 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2421 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2422 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2423 case T_CHAR: 2424 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2425 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2426 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2427 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2428 case T_METADATA: 2429 case T_ADDRESS: 2430 case T_OBJECT: 2431 #ifdef _LP64 2432 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2433 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2434 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2435 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2436 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2437 adr->bottom_type()->isa_rawptr())) { 2438 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2439 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2440 } 2441 #endif 2442 { 2443 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2444 } 2445 default: 2446 ShouldNotReachHere(); 2447 return (StoreNode*)NULL; 2448 } 2449 } 2450 2451 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2452 bool require_atomic = true; 2453 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2454 } 2455 2456 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2457 bool require_atomic = true; 2458 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2459 } 2460 2461 2462 //--------------------------bottom_type---------------------------------------- 2463 const Type *StoreNode::bottom_type() const { 2464 return Type::MEMORY; 2465 } 2466 2467 //------------------------------hash------------------------------------------- 2468 uint StoreNode::hash() const { 2469 // unroll addition of interesting fields 2470 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2471 2472 // Since they are not commoned, do not hash them: 2473 return NO_HASH; 2474 } 2475 2476 //------------------------------Ideal------------------------------------------ 2477 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2478 // When a store immediately follows a relevant allocation/initialization, 2479 // try to capture it into the initialization, or hoist it above. 2480 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2481 Node* p = MemNode::Ideal_common(phase, can_reshape); 2482 if (p) return (p == NodeSentinel) ? NULL : p; 2483 2484 Node* mem = in(MemNode::Memory); 2485 Node* address = in(MemNode::Address); 2486 // Back-to-back stores to same address? Fold em up. Generally 2487 // unsafe if I have intervening uses... Also disallowed for StoreCM 2488 // since they must follow each StoreP operation. Redundant StoreCMs 2489 // are eliminated just before matching in final_graph_reshape. 2490 { 2491 Node* st = mem; 2492 // If Store 'st' has more than one use, we cannot fold 'st' away. 2493 // For example, 'st' might be the final state at a conditional 2494 // return. Or, 'st' might be used by some node which is live at 2495 // the same time 'st' is live, which might be unschedulable. So, 2496 // require exactly ONE user until such time as we clone 'mem' for 2497 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2498 // true). 2499 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2500 // Looking at a dead closed cycle of memory? 2501 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2502 assert(Opcode() == st->Opcode() || 2503 st->Opcode() == Op_StoreVector || 2504 Opcode() == Op_StoreVector || 2505 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2506 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2507 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2508 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2509 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2510 2511 if (st->in(MemNode::Address)->eqv_uncast(address) && 2512 st->as_Store()->memory_size() <= this->memory_size()) { 2513 Node* use = st->raw_out(0); 2514 phase->igvn_rehash_node_delayed(use); 2515 if (can_reshape) { 2516 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2517 } else { 2518 // It's OK to do this in the parser, since DU info is always accurate, 2519 // and the parser always refers to nodes via SafePointNode maps. 2520 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2521 } 2522 return this; 2523 } 2524 st = st->in(MemNode::Memory); 2525 } 2526 } 2527 2528 2529 // Capture an unaliased, unconditional, simple store into an initializer. 2530 // Or, if it is independent of the allocation, hoist it above the allocation. 2531 if (ReduceFieldZeroing && /*can_reshape &&*/ 2532 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2533 InitializeNode* init = mem->in(0)->as_Initialize(); 2534 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2535 if (offset > 0) { 2536 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2537 // If the InitializeNode captured me, it made a raw copy of me, 2538 // and I need to disappear. 2539 if (moved != NULL) { 2540 // %%% hack to ensure that Ideal returns a new node: 2541 mem = MergeMemNode::make(mem); 2542 return mem; // fold me away 2543 } 2544 } 2545 } 2546 2547 return NULL; // No further progress 2548 } 2549 2550 //------------------------------Value----------------------------------------- 2551 const Type* StoreNode::Value(PhaseGVN* phase) const { 2552 // Either input is TOP ==> the result is TOP 2553 const Type *t1 = phase->type( in(MemNode::Memory) ); 2554 if( t1 == Type::TOP ) return Type::TOP; 2555 const Type *t2 = phase->type( in(MemNode::Address) ); 2556 if( t2 == Type::TOP ) return Type::TOP; 2557 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2558 if( t3 == Type::TOP ) return Type::TOP; 2559 return Type::MEMORY; 2560 } 2561 2562 //------------------------------Identity--------------------------------------- 2563 // Remove redundant stores: 2564 // Store(m, p, Load(m, p)) changes to m. 2565 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2566 Node* StoreNode::Identity(PhaseGVN* phase) { 2567 Node* mem = in(MemNode::Memory); 2568 Node* adr = in(MemNode::Address); 2569 Node* val = in(MemNode::ValueIn); 2570 2571 Node* result = this; 2572 2573 // Load then Store? Then the Store is useless 2574 if (val->is_Load() && 2575 val->in(MemNode::Address)->eqv_uncast(adr) && 2576 val->in(MemNode::Memory )->eqv_uncast(mem) && 2577 val->as_Load()->store_Opcode() == Opcode()) { 2578 result = mem; 2579 } 2580 2581 // Two stores in a row of the same value? 2582 if (result == this && 2583 mem->is_Store() && 2584 mem->in(MemNode::Address)->eqv_uncast(adr) && 2585 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2586 mem->Opcode() == Opcode()) { 2587 result = mem; 2588 } 2589 2590 // Store of zero anywhere into a freshly-allocated object? 2591 // Then the store is useless. 2592 // (It must already have been captured by the InitializeNode.) 2593 if (result == this && 2594 ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2595 // a newly allocated object is already all-zeroes everywhere 2596 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2597 result = mem; 2598 } 2599 2600 if (result == this) { 2601 // the store may also apply to zero-bits in an earlier object 2602 Node* prev_mem = find_previous_store(phase); 2603 // Steps (a), (b): Walk past independent stores to find an exact match. 2604 if (prev_mem != NULL) { 2605 Node* prev_val = can_see_stored_value(prev_mem, phase); 2606 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2607 // prev_val and val might differ by a cast; it would be good 2608 // to keep the more informative of the two. 2609 result = mem; 2610 } 2611 } 2612 } 2613 } 2614 2615 if (result != this && phase->is_IterGVN() != NULL) { 2616 MemBarNode* trailing = trailing_membar(); 2617 if (trailing != NULL) { 2618 #ifdef ASSERT 2619 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2620 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects"); 2621 #endif 2622 PhaseIterGVN* igvn = phase->is_IterGVN(); 2623 trailing->remove(igvn); 2624 } 2625 } 2626 2627 return result; 2628 } 2629 2630 //------------------------------match_edge------------------------------------- 2631 // Do we Match on this edge index or not? Match only memory & value 2632 uint StoreNode::match_edge(uint idx) const { 2633 return idx == MemNode::Address || idx == MemNode::ValueIn; 2634 } 2635 2636 //------------------------------cmp-------------------------------------------- 2637 // Do not common stores up together. They generally have to be split 2638 // back up anyways, so do not bother. 2639 uint StoreNode::cmp( const Node &n ) const { 2640 return (&n == this); // Always fail except on self 2641 } 2642 2643 //------------------------------Ideal_masked_input----------------------------- 2644 // Check for a useless mask before a partial-word store 2645 // (StoreB ... (AndI valIn conIa) ) 2646 // If (conIa & mask == mask) this simplifies to 2647 // (StoreB ... (valIn) ) 2648 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2649 Node *val = in(MemNode::ValueIn); 2650 if( val->Opcode() == Op_AndI ) { 2651 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2652 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2653 set_req(MemNode::ValueIn, val->in(1)); 2654 return this; 2655 } 2656 } 2657 return NULL; 2658 } 2659 2660 2661 //------------------------------Ideal_sign_extended_input---------------------- 2662 // Check for useless sign-extension before a partial-word store 2663 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2664 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2665 // (StoreB ... (valIn) ) 2666 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2667 Node *val = in(MemNode::ValueIn); 2668 if( val->Opcode() == Op_RShiftI ) { 2669 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2670 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2671 Node *shl = val->in(1); 2672 if( shl->Opcode() == Op_LShiftI ) { 2673 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2674 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2675 set_req(MemNode::ValueIn, shl->in(1)); 2676 return this; 2677 } 2678 } 2679 } 2680 } 2681 return NULL; 2682 } 2683 2684 //------------------------------value_never_loaded----------------------------------- 2685 // Determine whether there are any possible loads of the value stored. 2686 // For simplicity, we actually check if there are any loads from the 2687 // address stored to, not just for loads of the value stored by this node. 2688 // 2689 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2690 Node *adr = in(Address); 2691 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2692 if (adr_oop == NULL) 2693 return false; 2694 if (!adr_oop->is_known_instance_field()) 2695 return false; // if not a distinct instance, there may be aliases of the address 2696 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2697 Node *use = adr->fast_out(i); 2698 if (use->is_Load() || use->is_LoadStore()) { 2699 return false; 2700 } 2701 } 2702 return true; 2703 } 2704 2705 MemBarNode* StoreNode::trailing_membar() const { 2706 if (is_release()) { 2707 MemBarNode* trailing_mb = NULL; 2708 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2709 Node* u = fast_out(i); 2710 if (u->is_MemBar()) { 2711 if (u->as_MemBar()->trailing_store()) { 2712 assert(u->Opcode() == Op_MemBarVolatile, ""); 2713 assert(trailing_mb == NULL, "only one"); 2714 trailing_mb = u->as_MemBar(); 2715 #ifdef ASSERT 2716 Node* leading = u->as_MemBar()->leading_membar(); 2717 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2718 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2719 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2720 #endif 2721 } else { 2722 assert(u->as_MemBar()->standalone(), ""); 2723 } 2724 } 2725 } 2726 return trailing_mb; 2727 } 2728 return NULL; 2729 } 2730 2731 2732 //============================================================================= 2733 //------------------------------Ideal------------------------------------------ 2734 // If the store is from an AND mask that leaves the low bits untouched, then 2735 // we can skip the AND operation. If the store is from a sign-extension 2736 // (a left shift, then right shift) we can skip both. 2737 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2738 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2739 if( progress != NULL ) return progress; 2740 2741 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2742 if( progress != NULL ) return progress; 2743 2744 // Finally check the default case 2745 return StoreNode::Ideal(phase, can_reshape); 2746 } 2747 2748 //============================================================================= 2749 //------------------------------Ideal------------------------------------------ 2750 // If the store is from an AND mask that leaves the low bits untouched, then 2751 // we can skip the AND operation 2752 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2753 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2754 if( progress != NULL ) return progress; 2755 2756 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2757 if( progress != NULL ) return progress; 2758 2759 // Finally check the default case 2760 return StoreNode::Ideal(phase, can_reshape); 2761 } 2762 2763 //============================================================================= 2764 //------------------------------Identity--------------------------------------- 2765 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2766 // No need to card mark when storing a null ptr 2767 Node* my_store = in(MemNode::OopStore); 2768 if (my_store->is_Store()) { 2769 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2770 if( t1 == TypePtr::NULL_PTR ) { 2771 return in(MemNode::Memory); 2772 } 2773 } 2774 return this; 2775 } 2776 2777 //============================================================================= 2778 //------------------------------Ideal--------------------------------------- 2779 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2780 Node* progress = StoreNode::Ideal(phase, can_reshape); 2781 if (progress != NULL) return progress; 2782 2783 Node* my_store = in(MemNode::OopStore); 2784 if (my_store->is_MergeMem()) { 2785 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2786 set_req(MemNode::OopStore, mem); 2787 return this; 2788 } 2789 2790 return NULL; 2791 } 2792 2793 //------------------------------Value----------------------------------------- 2794 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2795 // Either input is TOP ==> the result is TOP 2796 const Type *t = phase->type( in(MemNode::Memory) ); 2797 if( t == Type::TOP ) return Type::TOP; 2798 t = phase->type( in(MemNode::Address) ); 2799 if( t == Type::TOP ) return Type::TOP; 2800 t = phase->type( in(MemNode::ValueIn) ); 2801 if( t == Type::TOP ) return Type::TOP; 2802 // If extra input is TOP ==> the result is TOP 2803 t = phase->type( in(MemNode::OopStore) ); 2804 if( t == Type::TOP ) return Type::TOP; 2805 2806 return StoreNode::Value( phase ); 2807 } 2808 2809 2810 //============================================================================= 2811 //----------------------------------SCMemProjNode------------------------------ 2812 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2813 { 2814 return bottom_type(); 2815 } 2816 2817 //============================================================================= 2818 //----------------------------------LoadStoreNode------------------------------ 2819 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2820 : Node(required), 2821 _type(rt), 2822 _adr_type(at) 2823 { 2824 init_req(MemNode::Control, c ); 2825 init_req(MemNode::Memory , mem); 2826 init_req(MemNode::Address, adr); 2827 init_req(MemNode::ValueIn, val); 2828 init_class_id(Class_LoadStore); 2829 } 2830 2831 uint LoadStoreNode::ideal_reg() const { 2832 return _type->ideal_reg(); 2833 } 2834 2835 bool LoadStoreNode::result_not_used() const { 2836 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2837 Node *x = fast_out(i); 2838 if (x->Opcode() == Op_SCMemProj) continue; 2839 return false; 2840 } 2841 return true; 2842 } 2843 2844 MemBarNode* LoadStoreNode::trailing_membar() const { 2845 MemBarNode* trailing = NULL; 2846 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2847 Node* u = fast_out(i); 2848 if (u->is_MemBar()) { 2849 if (u->as_MemBar()->trailing_load_store()) { 2850 assert(u->Opcode() == Op_MemBarAcquire, ""); 2851 assert(trailing == NULL, "only one"); 2852 trailing = u->as_MemBar(); 2853 #ifdef ASSERT 2854 Node* leading = trailing->leading_membar(); 2855 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2856 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 2857 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 2858 #endif 2859 } else { 2860 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 2861 } 2862 } 2863 } 2864 2865 return trailing; 2866 } 2867 2868 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2869 2870 //============================================================================= 2871 //----------------------------------LoadStoreConditionalNode-------------------- 2872 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2873 init_req(ExpectedIn, ex ); 2874 } 2875 2876 //============================================================================= 2877 //-------------------------------adr_type-------------------------------------- 2878 const TypePtr* ClearArrayNode::adr_type() const { 2879 Node *adr = in(3); 2880 if (adr == NULL) return NULL; // node is dead 2881 return MemNode::calculate_adr_type(adr->bottom_type()); 2882 } 2883 2884 //------------------------------match_edge------------------------------------- 2885 // Do we Match on this edge index or not? Do not match memory 2886 uint ClearArrayNode::match_edge(uint idx) const { 2887 return idx > 1; 2888 } 2889 2890 //------------------------------Identity--------------------------------------- 2891 // Clearing a zero length array does nothing 2892 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2893 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2894 } 2895 2896 //------------------------------Idealize--------------------------------------- 2897 // Clearing a short array is faster with stores 2898 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2899 // Already know this is a large node, do not try to ideal it 2900 if (!IdealizeClearArrayNode || _is_large) return NULL; 2901 2902 const int unit = BytesPerLong; 2903 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2904 if (!t) return NULL; 2905 if (!t->is_con()) return NULL; 2906 intptr_t raw_count = t->get_con(); 2907 intptr_t size = raw_count; 2908 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2909 // Clearing nothing uses the Identity call. 2910 // Negative clears are possible on dead ClearArrays 2911 // (see jck test stmt114.stmt11402.val). 2912 if (size <= 0 || size % unit != 0) return NULL; 2913 intptr_t count = size / unit; 2914 // Length too long; communicate this to matchers and assemblers. 2915 // Assemblers are responsible to produce fast hardware clears for it. 2916 if (size > InitArrayShortSize) { 2917 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 2918 } 2919 Node *mem = in(1); 2920 if( phase->type(mem)==Type::TOP ) return NULL; 2921 Node *adr = in(3); 2922 const Type* at = phase->type(adr); 2923 if( at==Type::TOP ) return NULL; 2924 const TypePtr* atp = at->isa_ptr(); 2925 // adjust atp to be the correct array element address type 2926 if (atp == NULL) atp = TypePtr::BOTTOM; 2927 else atp = atp->add_offset(Type::OffsetBot); 2928 // Get base for derived pointer purposes 2929 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2930 Node *base = adr->in(1); 2931 2932 Node *zero = phase->makecon(TypeLong::ZERO); 2933 Node *off = phase->MakeConX(BytesPerLong); 2934 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2935 count--; 2936 while( count-- ) { 2937 mem = phase->transform(mem); 2938 adr = phase->transform(new AddPNode(base,adr,off)); 2939 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2940 } 2941 return mem; 2942 } 2943 2944 //----------------------------step_through---------------------------------- 2945 // Return allocation input memory edge if it is different instance 2946 // or itself if it is the one we are looking for. 2947 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2948 Node* n = *np; 2949 assert(n->is_ClearArray(), "sanity"); 2950 intptr_t offset; 2951 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2952 // This method is called only before Allocate nodes are expanded 2953 // during macro nodes expansion. Before that ClearArray nodes are 2954 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2955 // Allocate nodes are expanded) which follows allocations. 2956 assert(alloc != NULL, "should have allocation"); 2957 if (alloc->_idx == instance_id) { 2958 // Can not bypass initialization of the instance we are looking for. 2959 return false; 2960 } 2961 // Otherwise skip it. 2962 InitializeNode* init = alloc->initialization(); 2963 if (init != NULL) 2964 *np = init->in(TypeFunc::Memory); 2965 else 2966 *np = alloc->in(TypeFunc::Memory); 2967 return true; 2968 } 2969 2970 //----------------------------clear_memory------------------------------------- 2971 // Generate code to initialize object storage to zero. 2972 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2973 intptr_t start_offset, 2974 Node* end_offset, 2975 PhaseGVN* phase) { 2976 intptr_t offset = start_offset; 2977 2978 int unit = BytesPerLong; 2979 if ((offset % unit) != 0) { 2980 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2981 adr = phase->transform(adr); 2982 const TypePtr* atp = TypeRawPtr::BOTTOM; 2983 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2984 mem = phase->transform(mem); 2985 offset += BytesPerInt; 2986 } 2987 assert((offset % unit) == 0, ""); 2988 2989 // Initialize the remaining stuff, if any, with a ClearArray. 2990 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2991 } 2992 2993 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2994 Node* start_offset, 2995 Node* end_offset, 2996 PhaseGVN* phase) { 2997 if (start_offset == end_offset) { 2998 // nothing to do 2999 return mem; 3000 } 3001 3002 int unit = BytesPerLong; 3003 Node* zbase = start_offset; 3004 Node* zend = end_offset; 3005 3006 // Scale to the unit required by the CPU: 3007 if (!Matcher::init_array_count_is_in_bytes) { 3008 Node* shift = phase->intcon(exact_log2(unit)); 3009 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3010 zend = phase->transform(new URShiftXNode(zend, shift) ); 3011 } 3012 3013 // Bulk clear double-words 3014 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3015 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3016 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 3017 return phase->transform(mem); 3018 } 3019 3020 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3021 intptr_t start_offset, 3022 intptr_t end_offset, 3023 PhaseGVN* phase) { 3024 if (start_offset == end_offset) { 3025 // nothing to do 3026 return mem; 3027 } 3028 3029 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3030 intptr_t done_offset = end_offset; 3031 if ((done_offset % BytesPerLong) != 0) { 3032 done_offset -= BytesPerInt; 3033 } 3034 if (done_offset > start_offset) { 3035 mem = clear_memory(ctl, mem, dest, 3036 start_offset, phase->MakeConX(done_offset), phase); 3037 } 3038 if (done_offset < end_offset) { // emit the final 32-bit store 3039 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3040 adr = phase->transform(adr); 3041 const TypePtr* atp = TypeRawPtr::BOTTOM; 3042 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3043 mem = phase->transform(mem); 3044 done_offset += BytesPerInt; 3045 } 3046 assert(done_offset == end_offset, ""); 3047 return mem; 3048 } 3049 3050 //============================================================================= 3051 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3052 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 3053 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3054 #ifdef ASSERT 3055 , _pair_idx(0) 3056 #endif 3057 { 3058 init_class_id(Class_MemBar); 3059 Node* top = C->top(); 3060 init_req(TypeFunc::I_O,top); 3061 init_req(TypeFunc::FramePtr,top); 3062 init_req(TypeFunc::ReturnAdr,top); 3063 if (precedent != NULL) 3064 init_req(TypeFunc::Parms, precedent); 3065 } 3066 3067 //------------------------------cmp-------------------------------------------- 3068 uint MemBarNode::hash() const { return NO_HASH; } 3069 uint MemBarNode::cmp( const Node &n ) const { 3070 return (&n == this); // Always fail except on self 3071 } 3072 3073 //------------------------------make------------------------------------------- 3074 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3075 switch (opcode) { 3076 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3077 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3078 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3079 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3080 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3081 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3082 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3083 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3084 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3085 case Op_Initialize: return new InitializeNode(C, atp, pn); 3086 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3087 default: ShouldNotReachHere(); return NULL; 3088 } 3089 } 3090 3091 void MemBarNode::remove(PhaseIterGVN *igvn) { 3092 if (outcnt() != 2) { 3093 return; 3094 } 3095 if (trailing_store() || trailing_load_store()) { 3096 MemBarNode* leading = leading_membar(); 3097 if (leading != NULL) { 3098 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3099 leading->remove(igvn); 3100 } 3101 } 3102 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3103 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3104 } 3105 3106 //------------------------------Ideal------------------------------------------ 3107 // Return a node which is more "ideal" than the current node. Strip out 3108 // control copies 3109 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3110 if (remove_dead_region(phase, can_reshape)) return this; 3111 // Don't bother trying to transform a dead node 3112 if (in(0) && in(0)->is_top()) { 3113 return NULL; 3114 } 3115 3116 #if INCLUDE_ZGC 3117 if (UseZGC) { 3118 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) { 3119 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop); 3120 set_req(MemBarNode::Precedent, load_node); 3121 return this; 3122 } 3123 } 3124 #endif 3125 3126 bool progress = false; 3127 // Eliminate volatile MemBars for scalar replaced objects. 3128 if (can_reshape && req() == (Precedent+1)) { 3129 bool eliminate = false; 3130 int opc = Opcode(); 3131 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3132 // Volatile field loads and stores. 3133 Node* my_mem = in(MemBarNode::Precedent); 3134 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3135 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3136 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3137 // replace this Precedent (decodeN) with the Load instead. 3138 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3139 Node* load_node = my_mem->in(1); 3140 set_req(MemBarNode::Precedent, load_node); 3141 phase->is_IterGVN()->_worklist.push(my_mem); 3142 my_mem = load_node; 3143 } else { 3144 assert(my_mem->unique_out() == this, "sanity"); 3145 del_req(Precedent); 3146 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3147 my_mem = NULL; 3148 } 3149 progress = true; 3150 } 3151 if (my_mem != NULL && my_mem->is_Mem()) { 3152 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3153 // Check for scalar replaced object reference. 3154 if( t_oop != NULL && t_oop->is_known_instance_field() && 3155 t_oop->offset() != Type::OffsetBot && 3156 t_oop->offset() != Type::OffsetTop) { 3157 eliminate = true; 3158 } 3159 } 3160 } else if (opc == Op_MemBarRelease) { 3161 // Final field stores. 3162 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3163 if ((alloc != NULL) && alloc->is_Allocate() && 3164 alloc->as_Allocate()->does_not_escape_thread()) { 3165 // The allocated object does not escape. 3166 eliminate = true; 3167 } 3168 } 3169 if (eliminate) { 3170 // Replace MemBar projections by its inputs. 3171 PhaseIterGVN* igvn = phase->is_IterGVN(); 3172 remove(igvn); 3173 // Must return either the original node (now dead) or a new node 3174 // (Do not return a top here, since that would break the uniqueness of top.) 3175 return new ConINode(TypeInt::ZERO); 3176 } 3177 } 3178 return progress ? this : NULL; 3179 } 3180 3181 //------------------------------Value------------------------------------------ 3182 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3183 if( !in(0) ) return Type::TOP; 3184 if( phase->type(in(0)) == Type::TOP ) 3185 return Type::TOP; 3186 return TypeTuple::MEMBAR; 3187 } 3188 3189 //------------------------------match------------------------------------------ 3190 // Construct projections for memory. 3191 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3192 switch (proj->_con) { 3193 case TypeFunc::Control: 3194 case TypeFunc::Memory: 3195 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3196 } 3197 ShouldNotReachHere(); 3198 return NULL; 3199 } 3200 3201 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3202 trailing->_kind = TrailingStore; 3203 leading->_kind = LeadingStore; 3204 #ifdef ASSERT 3205 trailing->_pair_idx = leading->_idx; 3206 leading->_pair_idx = leading->_idx; 3207 #endif 3208 } 3209 3210 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3211 trailing->_kind = TrailingLoadStore; 3212 leading->_kind = LeadingLoadStore; 3213 #ifdef ASSERT 3214 trailing->_pair_idx = leading->_idx; 3215 leading->_pair_idx = leading->_idx; 3216 #endif 3217 } 3218 3219 MemBarNode* MemBarNode::trailing_membar() const { 3220 ResourceMark rm; 3221 Node* trailing = (Node*)this; 3222 VectorSet seen(Thread::current()->resource_area()); 3223 Node_Stack multis(0); 3224 do { 3225 Node* c = trailing; 3226 uint i = 0; 3227 do { 3228 trailing = NULL; 3229 for (; i < c->outcnt(); i++) { 3230 Node* next = c->raw_out(i); 3231 if (next != c && next->is_CFG()) { 3232 if (c->is_MultiBranch()) { 3233 if (multis.node() == c) { 3234 multis.set_index(i+1); 3235 } else { 3236 multis.push(c, i+1); 3237 } 3238 } 3239 trailing = next; 3240 break; 3241 } 3242 } 3243 if (trailing != NULL && !seen.test_set(trailing->_idx)) { 3244 break; 3245 } 3246 while (multis.size() > 0) { 3247 c = multis.node(); 3248 i = multis.index(); 3249 if (i < c->req()) { 3250 break; 3251 } 3252 multis.pop(); 3253 } 3254 } while (multis.size() > 0); 3255 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3256 3257 MemBarNode* mb = trailing->as_MemBar(); 3258 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3259 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3260 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3261 return mb; 3262 } 3263 3264 MemBarNode* MemBarNode::leading_membar() const { 3265 ResourceMark rm; 3266 VectorSet seen(Thread::current()->resource_area()); 3267 Node_Stack regions(0); 3268 Node* leading = in(0); 3269 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3270 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) { 3271 leading = NULL; 3272 while (regions.size() > 0 && leading == NULL) { 3273 Node* r = regions.node(); 3274 uint i = regions.index(); 3275 if (i < r->req()) { 3276 leading = r->in(i); 3277 regions.set_index(i+1); 3278 } else { 3279 regions.pop(); 3280 } 3281 } 3282 if (leading == NULL) { 3283 assert(regions.size() == 0, "all paths should have been tried"); 3284 return NULL; 3285 } 3286 } 3287 if (leading->is_Region()) { 3288 regions.push(leading, 2); 3289 leading = leading->in(1); 3290 } else { 3291 leading = leading->in(0); 3292 } 3293 } 3294 #ifdef ASSERT 3295 Unique_Node_List wq; 3296 wq.push((Node*)this); 3297 uint found = 0; 3298 for (uint i = 0; i < wq.size(); i++) { 3299 Node* n = wq.at(i); 3300 if (n->is_Region()) { 3301 for (uint j = 1; j < n->req(); j++) { 3302 Node* in = n->in(j); 3303 if (in != NULL && !in->is_top()) { 3304 wq.push(in); 3305 } 3306 } 3307 } else { 3308 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3309 assert(n == leading, "consistency check failed"); 3310 found++; 3311 } else { 3312 Node* in = n->in(0); 3313 if (in != NULL && !in->is_top()) { 3314 wq.push(in); 3315 } 3316 } 3317 } 3318 } 3319 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed"); 3320 #endif 3321 if (leading == NULL) { 3322 return NULL; 3323 } 3324 MemBarNode* mb = leading->as_MemBar(); 3325 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3326 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3327 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3328 return mb; 3329 } 3330 3331 //===========================InitializeNode==================================== 3332 // SUMMARY: 3333 // This node acts as a memory barrier on raw memory, after some raw stores. 3334 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3335 // The Initialize can 'capture' suitably constrained stores as raw inits. 3336 // It can coalesce related raw stores into larger units (called 'tiles'). 3337 // It can avoid zeroing new storage for memory units which have raw inits. 3338 // At macro-expansion, it is marked 'complete', and does not optimize further. 3339 // 3340 // EXAMPLE: 3341 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3342 // ctl = incoming control; mem* = incoming memory 3343 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3344 // First allocate uninitialized memory and fill in the header: 3345 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3346 // ctl := alloc.Control; mem* := alloc.Memory* 3347 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3348 // Then initialize to zero the non-header parts of the raw memory block: 3349 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3350 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3351 // After the initialize node executes, the object is ready for service: 3352 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3353 // Suppose its body is immediately initialized as {1,2}: 3354 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3355 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3356 // mem.SLICE(#short[*]) := store2 3357 // 3358 // DETAILS: 3359 // An InitializeNode collects and isolates object initialization after 3360 // an AllocateNode and before the next possible safepoint. As a 3361 // memory barrier (MemBarNode), it keeps critical stores from drifting 3362 // down past any safepoint or any publication of the allocation. 3363 // Before this barrier, a newly-allocated object may have uninitialized bits. 3364 // After this barrier, it may be treated as a real oop, and GC is allowed. 3365 // 3366 // The semantics of the InitializeNode include an implicit zeroing of 3367 // the new object from object header to the end of the object. 3368 // (The object header and end are determined by the AllocateNode.) 3369 // 3370 // Certain stores may be added as direct inputs to the InitializeNode. 3371 // These stores must update raw memory, and they must be to addresses 3372 // derived from the raw address produced by AllocateNode, and with 3373 // a constant offset. They must be ordered by increasing offset. 3374 // The first one is at in(RawStores), the last at in(req()-1). 3375 // Unlike most memory operations, they are not linked in a chain, 3376 // but are displayed in parallel as users of the rawmem output of 3377 // the allocation. 3378 // 3379 // (See comments in InitializeNode::capture_store, which continue 3380 // the example given above.) 3381 // 3382 // When the associated Allocate is macro-expanded, the InitializeNode 3383 // may be rewritten to optimize collected stores. A ClearArrayNode 3384 // may also be created at that point to represent any required zeroing. 3385 // The InitializeNode is then marked 'complete', prohibiting further 3386 // capturing of nearby memory operations. 3387 // 3388 // During macro-expansion, all captured initializations which store 3389 // constant values of 32 bits or smaller are coalesced (if advantageous) 3390 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3391 // initialized in fewer memory operations. Memory words which are 3392 // covered by neither tiles nor non-constant stores are pre-zeroed 3393 // by explicit stores of zero. (The code shape happens to do all 3394 // zeroing first, then all other stores, with both sequences occurring 3395 // in order of ascending offsets.) 3396 // 3397 // Alternatively, code may be inserted between an AllocateNode and its 3398 // InitializeNode, to perform arbitrary initialization of the new object. 3399 // E.g., the object copying intrinsics insert complex data transfers here. 3400 // The initialization must then be marked as 'complete' disable the 3401 // built-in zeroing semantics and the collection of initializing stores. 3402 // 3403 // While an InitializeNode is incomplete, reads from the memory state 3404 // produced by it are optimizable if they match the control edge and 3405 // new oop address associated with the allocation/initialization. 3406 // They return a stored value (if the offset matches) or else zero. 3407 // A write to the memory state, if it matches control and address, 3408 // and if it is to a constant offset, may be 'captured' by the 3409 // InitializeNode. It is cloned as a raw memory operation and rewired 3410 // inside the initialization, to the raw oop produced by the allocation. 3411 // Operations on addresses which are provably distinct (e.g., to 3412 // other AllocateNodes) are allowed to bypass the initialization. 3413 // 3414 // The effect of all this is to consolidate object initialization 3415 // (both arrays and non-arrays, both piecewise and bulk) into a 3416 // single location, where it can be optimized as a unit. 3417 // 3418 // Only stores with an offset less than TrackedInitializationLimit words 3419 // will be considered for capture by an InitializeNode. This puts a 3420 // reasonable limit on the complexity of optimized initializations. 3421 3422 //---------------------------InitializeNode------------------------------------ 3423 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3424 : _is_complete(Incomplete), _does_not_escape(false), 3425 MemBarNode(C, adr_type, rawoop) 3426 { 3427 init_class_id(Class_Initialize); 3428 3429 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3430 assert(in(RawAddress) == rawoop, "proper init"); 3431 // Note: allocation() can be NULL, for secondary initialization barriers 3432 } 3433 3434 // Since this node is not matched, it will be processed by the 3435 // register allocator. Declare that there are no constraints 3436 // on the allocation of the RawAddress edge. 3437 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3438 // This edge should be set to top, by the set_complete. But be conservative. 3439 if (idx == InitializeNode::RawAddress) 3440 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3441 return RegMask::Empty; 3442 } 3443 3444 Node* InitializeNode::memory(uint alias_idx) { 3445 Node* mem = in(Memory); 3446 if (mem->is_MergeMem()) { 3447 return mem->as_MergeMem()->memory_at(alias_idx); 3448 } else { 3449 // incoming raw memory is not split 3450 return mem; 3451 } 3452 } 3453 3454 bool InitializeNode::is_non_zero() { 3455 if (is_complete()) return false; 3456 remove_extra_zeroes(); 3457 return (req() > RawStores); 3458 } 3459 3460 void InitializeNode::set_complete(PhaseGVN* phase) { 3461 assert(!is_complete(), "caller responsibility"); 3462 _is_complete = Complete; 3463 3464 // After this node is complete, it contains a bunch of 3465 // raw-memory initializations. There is no need for 3466 // it to have anything to do with non-raw memory effects. 3467 // Therefore, tell all non-raw users to re-optimize themselves, 3468 // after skipping the memory effects of this initialization. 3469 PhaseIterGVN* igvn = phase->is_IterGVN(); 3470 if (igvn) igvn->add_users_to_worklist(this); 3471 } 3472 3473 // convenience function 3474 // return false if the init contains any stores already 3475 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3476 InitializeNode* init = initialization(); 3477 if (init == NULL || init->is_complete()) return false; 3478 init->remove_extra_zeroes(); 3479 // for now, if this allocation has already collected any inits, bail: 3480 if (init->is_non_zero()) return false; 3481 init->set_complete(phase); 3482 return true; 3483 } 3484 3485 void InitializeNode::remove_extra_zeroes() { 3486 if (req() == RawStores) return; 3487 Node* zmem = zero_memory(); 3488 uint fill = RawStores; 3489 for (uint i = fill; i < req(); i++) { 3490 Node* n = in(i); 3491 if (n->is_top() || n == zmem) continue; // skip 3492 if (fill < i) set_req(fill, n); // compact 3493 ++fill; 3494 } 3495 // delete any empty spaces created: 3496 while (fill < req()) { 3497 del_req(fill); 3498 } 3499 } 3500 3501 // Helper for remembering which stores go with which offsets. 3502 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3503 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3504 intptr_t offset = -1; 3505 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3506 phase, offset); 3507 if (base == NULL) return -1; // something is dead, 3508 if (offset < 0) return -1; // dead, dead 3509 return offset; 3510 } 3511 3512 // Helper for proving that an initialization expression is 3513 // "simple enough" to be folded into an object initialization. 3514 // Attempts to prove that a store's initial value 'n' can be captured 3515 // within the initialization without creating a vicious cycle, such as: 3516 // { Foo p = new Foo(); p.next = p; } 3517 // True for constants and parameters and small combinations thereof. 3518 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3519 if (n == NULL) return true; // (can this really happen?) 3520 if (n->is_Proj()) n = n->in(0); 3521 if (n == this) return false; // found a cycle 3522 if (n->is_Con()) return true; 3523 if (n->is_Start()) return true; // params, etc., are OK 3524 if (n->is_Root()) return true; // even better 3525 3526 Node* ctl = n->in(0); 3527 if (ctl != NULL && !ctl->is_top()) { 3528 if (ctl->is_Proj()) ctl = ctl->in(0); 3529 if (ctl == this) return false; 3530 3531 // If we already know that the enclosing memory op is pinned right after 3532 // the init, then any control flow that the store has picked up 3533 // must have preceded the init, or else be equal to the init. 3534 // Even after loop optimizations (which might change control edges) 3535 // a store is never pinned *before* the availability of its inputs. 3536 if (!MemNode::all_controls_dominate(n, this)) 3537 return false; // failed to prove a good control 3538 } 3539 3540 // Check data edges for possible dependencies on 'this'. 3541 if ((count += 1) > 20) return false; // complexity limit 3542 for (uint i = 1; i < n->req(); i++) { 3543 Node* m = n->in(i); 3544 if (m == NULL || m == n || m->is_top()) continue; 3545 uint first_i = n->find_edge(m); 3546 if (i != first_i) continue; // process duplicate edge just once 3547 if (!detect_init_independence(m, count)) { 3548 return false; 3549 } 3550 } 3551 3552 return true; 3553 } 3554 3555 // Here are all the checks a Store must pass before it can be moved into 3556 // an initialization. Returns zero if a check fails. 3557 // On success, returns the (constant) offset to which the store applies, 3558 // within the initialized memory. 3559 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3560 const int FAIL = 0; 3561 if (st->req() != MemNode::ValueIn + 1) 3562 return FAIL; // an inscrutable StoreNode (card mark?) 3563 Node* ctl = st->in(MemNode::Control); 3564 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3565 return FAIL; // must be unconditional after the initialization 3566 Node* mem = st->in(MemNode::Memory); 3567 if (!(mem->is_Proj() && mem->in(0) == this)) 3568 return FAIL; // must not be preceded by other stores 3569 Node* adr = st->in(MemNode::Address); 3570 intptr_t offset; 3571 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3572 if (alloc == NULL) 3573 return FAIL; // inscrutable address 3574 if (alloc != allocation()) 3575 return FAIL; // wrong allocation! (store needs to float up) 3576 int size_in_bytes = st->memory_size(); 3577 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) { 3578 return FAIL; // mismatched access 3579 } 3580 Node* val = st->in(MemNode::ValueIn); 3581 int complexity_count = 0; 3582 if (!detect_init_independence(val, complexity_count)) 3583 return FAIL; // stored value must be 'simple enough' 3584 3585 // The Store can be captured only if nothing after the allocation 3586 // and before the Store is using the memory location that the store 3587 // overwrites. 3588 bool failed = false; 3589 // If is_complete_with_arraycopy() is true the shape of the graph is 3590 // well defined and is safe so no need for extra checks. 3591 if (!is_complete_with_arraycopy()) { 3592 // We are going to look at each use of the memory state following 3593 // the allocation to make sure nothing reads the memory that the 3594 // Store writes. 3595 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3596 int alias_idx = phase->C->get_alias_index(t_adr); 3597 ResourceMark rm; 3598 Unique_Node_List mems; 3599 mems.push(mem); 3600 Node* unique_merge = NULL; 3601 for (uint next = 0; next < mems.size(); ++next) { 3602 Node *m = mems.at(next); 3603 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3604 Node *n = m->fast_out(j); 3605 if (n->outcnt() == 0) { 3606 continue; 3607 } 3608 if (n == st) { 3609 continue; 3610 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3611 // If the control of this use is different from the control 3612 // of the Store which is right after the InitializeNode then 3613 // this node cannot be between the InitializeNode and the 3614 // Store. 3615 continue; 3616 } else if (n->is_MergeMem()) { 3617 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3618 // We can hit a MergeMemNode (that will likely go away 3619 // later) that is a direct use of the memory state 3620 // following the InitializeNode on the same slice as the 3621 // store node that we'd like to capture. We need to check 3622 // the uses of the MergeMemNode. 3623 mems.push(n); 3624 } 3625 } else if (n->is_Mem()) { 3626 Node* other_adr = n->in(MemNode::Address); 3627 if (other_adr == adr) { 3628 failed = true; 3629 break; 3630 } else { 3631 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3632 if (other_t_adr != NULL) { 3633 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3634 if (other_alias_idx == alias_idx) { 3635 // A load from the same memory slice as the store right 3636 // after the InitializeNode. We check the control of the 3637 // object/array that is loaded from. If it's the same as 3638 // the store control then we cannot capture the store. 3639 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3640 Node* base = other_adr; 3641 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3642 base = base->in(AddPNode::Base); 3643 if (base != NULL) { 3644 base = base->uncast(); 3645 if (base->is_Proj() && base->in(0) == alloc) { 3646 failed = true; 3647 break; 3648 } 3649 } 3650 } 3651 } 3652 } 3653 } else { 3654 failed = true; 3655 break; 3656 } 3657 } 3658 } 3659 } 3660 if (failed) { 3661 if (!can_reshape) { 3662 // We decided we couldn't capture the store during parsing. We 3663 // should try again during the next IGVN once the graph is 3664 // cleaner. 3665 phase->C->record_for_igvn(st); 3666 } 3667 return FAIL; 3668 } 3669 3670 return offset; // success 3671 } 3672 3673 // Find the captured store in(i) which corresponds to the range 3674 // [start..start+size) in the initialized object. 3675 // If there is one, return its index i. If there isn't, return the 3676 // negative of the index where it should be inserted. 3677 // Return 0 if the queried range overlaps an initialization boundary 3678 // or if dead code is encountered. 3679 // If size_in_bytes is zero, do not bother with overlap checks. 3680 int InitializeNode::captured_store_insertion_point(intptr_t start, 3681 int size_in_bytes, 3682 PhaseTransform* phase) { 3683 const int FAIL = 0, MAX_STORE = BytesPerLong; 3684 3685 if (is_complete()) 3686 return FAIL; // arraycopy got here first; punt 3687 3688 assert(allocation() != NULL, "must be present"); 3689 3690 // no negatives, no header fields: 3691 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3692 3693 // after a certain size, we bail out on tracking all the stores: 3694 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3695 if (start >= ti_limit) return FAIL; 3696 3697 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3698 if (i >= limit) return -(int)i; // not found; here is where to put it 3699 3700 Node* st = in(i); 3701 intptr_t st_off = get_store_offset(st, phase); 3702 if (st_off < 0) { 3703 if (st != zero_memory()) { 3704 return FAIL; // bail out if there is dead garbage 3705 } 3706 } else if (st_off > start) { 3707 // ...we are done, since stores are ordered 3708 if (st_off < start + size_in_bytes) { 3709 return FAIL; // the next store overlaps 3710 } 3711 return -(int)i; // not found; here is where to put it 3712 } else if (st_off < start) { 3713 if (size_in_bytes != 0 && 3714 start < st_off + MAX_STORE && 3715 start < st_off + st->as_Store()->memory_size()) { 3716 return FAIL; // the previous store overlaps 3717 } 3718 } else { 3719 if (size_in_bytes != 0 && 3720 st->as_Store()->memory_size() != size_in_bytes) { 3721 return FAIL; // mismatched store size 3722 } 3723 return i; 3724 } 3725 3726 ++i; 3727 } 3728 } 3729 3730 // Look for a captured store which initializes at the offset 'start' 3731 // with the given size. If there is no such store, and no other 3732 // initialization interferes, then return zero_memory (the memory 3733 // projection of the AllocateNode). 3734 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3735 PhaseTransform* phase) { 3736 assert(stores_are_sane(phase), ""); 3737 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3738 if (i == 0) { 3739 return NULL; // something is dead 3740 } else if (i < 0) { 3741 return zero_memory(); // just primordial zero bits here 3742 } else { 3743 Node* st = in(i); // here is the store at this position 3744 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3745 return st; 3746 } 3747 } 3748 3749 // Create, as a raw pointer, an address within my new object at 'offset'. 3750 Node* InitializeNode::make_raw_address(intptr_t offset, 3751 PhaseTransform* phase) { 3752 Node* addr = in(RawAddress); 3753 if (offset != 0) { 3754 Compile* C = phase->C; 3755 addr = phase->transform( new AddPNode(C->top(), addr, 3756 phase->MakeConX(offset)) ); 3757 } 3758 return addr; 3759 } 3760 3761 // Clone the given store, converting it into a raw store 3762 // initializing a field or element of my new object. 3763 // Caller is responsible for retiring the original store, 3764 // with subsume_node or the like. 3765 // 3766 // From the example above InitializeNode::InitializeNode, 3767 // here are the old stores to be captured: 3768 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3769 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3770 // 3771 // Here is the changed code; note the extra edges on init: 3772 // alloc = (Allocate ...) 3773 // rawoop = alloc.RawAddress 3774 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3775 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3776 // init = (Initialize alloc.Control alloc.Memory rawoop 3777 // rawstore1 rawstore2) 3778 // 3779 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3780 PhaseTransform* phase, bool can_reshape) { 3781 assert(stores_are_sane(phase), ""); 3782 3783 if (start < 0) return NULL; 3784 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3785 3786 Compile* C = phase->C; 3787 int size_in_bytes = st->memory_size(); 3788 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3789 if (i == 0) return NULL; // bail out 3790 Node* prev_mem = NULL; // raw memory for the captured store 3791 if (i > 0) { 3792 prev_mem = in(i); // there is a pre-existing store under this one 3793 set_req(i, C->top()); // temporarily disconnect it 3794 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3795 } else { 3796 i = -i; // no pre-existing store 3797 prev_mem = zero_memory(); // a slice of the newly allocated object 3798 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3799 set_req(--i, C->top()); // reuse this edge; it has been folded away 3800 else 3801 ins_req(i, C->top()); // build a new edge 3802 } 3803 Node* new_st = st->clone(); 3804 new_st->set_req(MemNode::Control, in(Control)); 3805 new_st->set_req(MemNode::Memory, prev_mem); 3806 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3807 new_st = phase->transform(new_st); 3808 3809 // At this point, new_st might have swallowed a pre-existing store 3810 // at the same offset, or perhaps new_st might have disappeared, 3811 // if it redundantly stored the same value (or zero to fresh memory). 3812 3813 // In any case, wire it in: 3814 phase->igvn_rehash_node_delayed(this); 3815 set_req(i, new_st); 3816 3817 // The caller may now kill the old guy. 3818 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3819 assert(check_st == new_st || check_st == NULL, "must be findable"); 3820 assert(!is_complete(), ""); 3821 return new_st; 3822 } 3823 3824 static bool store_constant(jlong* tiles, int num_tiles, 3825 intptr_t st_off, int st_size, 3826 jlong con) { 3827 if ((st_off & (st_size-1)) != 0) 3828 return false; // strange store offset (assume size==2**N) 3829 address addr = (address)tiles + st_off; 3830 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3831 switch (st_size) { 3832 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3833 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3834 case sizeof(jint): *(jint*) addr = (jint) con; break; 3835 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3836 default: return false; // strange store size (detect size!=2**N here) 3837 } 3838 return true; // return success to caller 3839 } 3840 3841 // Coalesce subword constants into int constants and possibly 3842 // into long constants. The goal, if the CPU permits, 3843 // is to initialize the object with a small number of 64-bit tiles. 3844 // Also, convert floating-point constants to bit patterns. 3845 // Non-constants are not relevant to this pass. 3846 // 3847 // In terms of the running example on InitializeNode::InitializeNode 3848 // and InitializeNode::capture_store, here is the transformation 3849 // of rawstore1 and rawstore2 into rawstore12: 3850 // alloc = (Allocate ...) 3851 // rawoop = alloc.RawAddress 3852 // tile12 = 0x00010002 3853 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3854 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3855 // 3856 void 3857 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3858 Node* size_in_bytes, 3859 PhaseGVN* phase) { 3860 Compile* C = phase->C; 3861 3862 assert(stores_are_sane(phase), ""); 3863 // Note: After this pass, they are not completely sane, 3864 // since there may be some overlaps. 3865 3866 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3867 3868 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3869 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3870 size_limit = MIN2(size_limit, ti_limit); 3871 size_limit = align_up(size_limit, BytesPerLong); 3872 int num_tiles = size_limit / BytesPerLong; 3873 3874 // allocate space for the tile map: 3875 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3876 jlong tiles_buf[small_len]; 3877 Node* nodes_buf[small_len]; 3878 jlong inits_buf[small_len]; 3879 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3880 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3881 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3882 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3883 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3884 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3885 // tiles: exact bitwise model of all primitive constants 3886 // nodes: last constant-storing node subsumed into the tiles model 3887 // inits: which bytes (in each tile) are touched by any initializations 3888 3889 //// Pass A: Fill in the tile model with any relevant stores. 3890 3891 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3892 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3893 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3894 Node* zmem = zero_memory(); // initially zero memory state 3895 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3896 Node* st = in(i); 3897 intptr_t st_off = get_store_offset(st, phase); 3898 3899 // Figure out the store's offset and constant value: 3900 if (st_off < header_size) continue; //skip (ignore header) 3901 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3902 int st_size = st->as_Store()->memory_size(); 3903 if (st_off + st_size > size_limit) break; 3904 3905 // Record which bytes are touched, whether by constant or not. 3906 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3907 continue; // skip (strange store size) 3908 3909 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3910 if (!val->singleton()) continue; //skip (non-con store) 3911 BasicType type = val->basic_type(); 3912 3913 jlong con = 0; 3914 switch (type) { 3915 case T_INT: con = val->is_int()->get_con(); break; 3916 case T_LONG: con = val->is_long()->get_con(); break; 3917 case T_FLOAT: con = jint_cast(val->getf()); break; 3918 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3919 default: continue; //skip (odd store type) 3920 } 3921 3922 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3923 st->Opcode() == Op_StoreL) { 3924 continue; // This StoreL is already optimal. 3925 } 3926 3927 // Store down the constant. 3928 store_constant(tiles, num_tiles, st_off, st_size, con); 3929 3930 intptr_t j = st_off >> LogBytesPerLong; 3931 3932 if (type == T_INT && st_size == BytesPerInt 3933 && (st_off & BytesPerInt) == BytesPerInt) { 3934 jlong lcon = tiles[j]; 3935 if (!Matcher::isSimpleConstant64(lcon) && 3936 st->Opcode() == Op_StoreI) { 3937 // This StoreI is already optimal by itself. 3938 jint* intcon = (jint*) &tiles[j]; 3939 intcon[1] = 0; // undo the store_constant() 3940 3941 // If the previous store is also optimal by itself, back up and 3942 // undo the action of the previous loop iteration... if we can. 3943 // But if we can't, just let the previous half take care of itself. 3944 st = nodes[j]; 3945 st_off -= BytesPerInt; 3946 con = intcon[0]; 3947 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3948 assert(st_off >= header_size, "still ignoring header"); 3949 assert(get_store_offset(st, phase) == st_off, "must be"); 3950 assert(in(i-1) == zmem, "must be"); 3951 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3952 assert(con == tcon->is_int()->get_con(), "must be"); 3953 // Undo the effects of the previous loop trip, which swallowed st: 3954 intcon[0] = 0; // undo store_constant() 3955 set_req(i-1, st); // undo set_req(i, zmem) 3956 nodes[j] = NULL; // undo nodes[j] = st 3957 --old_subword; // undo ++old_subword 3958 } 3959 continue; // This StoreI is already optimal. 3960 } 3961 } 3962 3963 // This store is not needed. 3964 set_req(i, zmem); 3965 nodes[j] = st; // record for the moment 3966 if (st_size < BytesPerLong) // something has changed 3967 ++old_subword; // includes int/float, but who's counting... 3968 else ++old_long; 3969 } 3970 3971 if ((old_subword + old_long) == 0) 3972 return; // nothing more to do 3973 3974 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3975 // Be sure to insert them before overlapping non-constant stores. 3976 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3977 for (int j = 0; j < num_tiles; j++) { 3978 jlong con = tiles[j]; 3979 jlong init = inits[j]; 3980 if (con == 0) continue; 3981 jint con0, con1; // split the constant, address-wise 3982 jint init0, init1; // split the init map, address-wise 3983 { union { jlong con; jint intcon[2]; } u; 3984 u.con = con; 3985 con0 = u.intcon[0]; 3986 con1 = u.intcon[1]; 3987 u.con = init; 3988 init0 = u.intcon[0]; 3989 init1 = u.intcon[1]; 3990 } 3991 3992 Node* old = nodes[j]; 3993 assert(old != NULL, "need the prior store"); 3994 intptr_t offset = (j * BytesPerLong); 3995 3996 bool split = !Matcher::isSimpleConstant64(con); 3997 3998 if (offset < header_size) { 3999 assert(offset + BytesPerInt >= header_size, "second int counts"); 4000 assert(*(jint*)&tiles[j] == 0, "junk in header"); 4001 split = true; // only the second word counts 4002 // Example: int a[] = { 42 ... } 4003 } else if (con0 == 0 && init0 == -1) { 4004 split = true; // first word is covered by full inits 4005 // Example: int a[] = { ... foo(), 42 ... } 4006 } else if (con1 == 0 && init1 == -1) { 4007 split = true; // second word is covered by full inits 4008 // Example: int a[] = { ... 42, foo() ... } 4009 } 4010 4011 // Here's a case where init0 is neither 0 nor -1: 4012 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4013 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4014 // In this case the tile is not split; it is (jlong)42. 4015 // The big tile is stored down, and then the foo() value is inserted. 4016 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4017 4018 Node* ctl = old->in(MemNode::Control); 4019 Node* adr = make_raw_address(offset, phase); 4020 const TypePtr* atp = TypeRawPtr::BOTTOM; 4021 4022 // One or two coalesced stores to plop down. 4023 Node* st[2]; 4024 intptr_t off[2]; 4025 int nst = 0; 4026 if (!split) { 4027 ++new_long; 4028 off[nst] = offset; 4029 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4030 phase->longcon(con), T_LONG, MemNode::unordered); 4031 } else { 4032 // Omit either if it is a zero. 4033 if (con0 != 0) { 4034 ++new_int; 4035 off[nst] = offset; 4036 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4037 phase->intcon(con0), T_INT, MemNode::unordered); 4038 } 4039 if (con1 != 0) { 4040 ++new_int; 4041 offset += BytesPerInt; 4042 adr = make_raw_address(offset, phase); 4043 off[nst] = offset; 4044 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4045 phase->intcon(con1), T_INT, MemNode::unordered); 4046 } 4047 } 4048 4049 // Insert second store first, then the first before the second. 4050 // Insert each one just before any overlapping non-constant stores. 4051 while (nst > 0) { 4052 Node* st1 = st[--nst]; 4053 C->copy_node_notes_to(st1, old); 4054 st1 = phase->transform(st1); 4055 offset = off[nst]; 4056 assert(offset >= header_size, "do not smash header"); 4057 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4058 guarantee(ins_idx != 0, "must re-insert constant store"); 4059 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4060 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4061 set_req(--ins_idx, st1); 4062 else 4063 ins_req(ins_idx, st1); 4064 } 4065 } 4066 4067 if (PrintCompilation && WizardMode) 4068 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4069 old_subword, old_long, new_int, new_long); 4070 if (C->log() != NULL) 4071 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4072 old_subword, old_long, new_int, new_long); 4073 4074 // Clean up any remaining occurrences of zmem: 4075 remove_extra_zeroes(); 4076 } 4077 4078 // Explore forward from in(start) to find the first fully initialized 4079 // word, and return its offset. Skip groups of subword stores which 4080 // together initialize full words. If in(start) is itself part of a 4081 // fully initialized word, return the offset of in(start). If there 4082 // are no following full-word stores, or if something is fishy, return 4083 // a negative value. 4084 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4085 int int_map = 0; 4086 intptr_t int_map_off = 0; 4087 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4088 4089 for (uint i = start, limit = req(); i < limit; i++) { 4090 Node* st = in(i); 4091 4092 intptr_t st_off = get_store_offset(st, phase); 4093 if (st_off < 0) break; // return conservative answer 4094 4095 int st_size = st->as_Store()->memory_size(); 4096 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4097 return st_off; // we found a complete word init 4098 } 4099 4100 // update the map: 4101 4102 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4103 if (this_int_off != int_map_off) { 4104 // reset the map: 4105 int_map = 0; 4106 int_map_off = this_int_off; 4107 } 4108 4109 int subword_off = st_off - this_int_off; 4110 int_map |= right_n_bits(st_size) << subword_off; 4111 if ((int_map & FULL_MAP) == FULL_MAP) { 4112 return this_int_off; // we found a complete word init 4113 } 4114 4115 // Did this store hit or cross the word boundary? 4116 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4117 if (next_int_off == this_int_off + BytesPerInt) { 4118 // We passed the current int, without fully initializing it. 4119 int_map_off = next_int_off; 4120 int_map >>= BytesPerInt; 4121 } else if (next_int_off > this_int_off + BytesPerInt) { 4122 // We passed the current and next int. 4123 return this_int_off + BytesPerInt; 4124 } 4125 } 4126 4127 return -1; 4128 } 4129 4130 4131 // Called when the associated AllocateNode is expanded into CFG. 4132 // At this point, we may perform additional optimizations. 4133 // Linearize the stores by ascending offset, to make memory 4134 // activity as coherent as possible. 4135 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4136 intptr_t header_size, 4137 Node* size_in_bytes, 4138 PhaseGVN* phase) { 4139 assert(!is_complete(), "not already complete"); 4140 assert(stores_are_sane(phase), ""); 4141 assert(allocation() != NULL, "must be present"); 4142 4143 remove_extra_zeroes(); 4144 4145 if (ReduceFieldZeroing || ReduceBulkZeroing) 4146 // reduce instruction count for common initialization patterns 4147 coalesce_subword_stores(header_size, size_in_bytes, phase); 4148 4149 Node* zmem = zero_memory(); // initially zero memory state 4150 Node* inits = zmem; // accumulating a linearized chain of inits 4151 #ifdef ASSERT 4152 intptr_t first_offset = allocation()->minimum_header_size(); 4153 intptr_t last_init_off = first_offset; // previous init offset 4154 intptr_t last_init_end = first_offset; // previous init offset+size 4155 intptr_t last_tile_end = first_offset; // previous tile offset+size 4156 #endif 4157 intptr_t zeroes_done = header_size; 4158 4159 bool do_zeroing = true; // we might give up if inits are very sparse 4160 int big_init_gaps = 0; // how many large gaps have we seen? 4161 4162 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4163 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4164 4165 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4166 Node* st = in(i); 4167 intptr_t st_off = get_store_offset(st, phase); 4168 if (st_off < 0) 4169 break; // unknown junk in the inits 4170 if (st->in(MemNode::Memory) != zmem) 4171 break; // complicated store chains somehow in list 4172 4173 int st_size = st->as_Store()->memory_size(); 4174 intptr_t next_init_off = st_off + st_size; 4175 4176 if (do_zeroing && zeroes_done < next_init_off) { 4177 // See if this store needs a zero before it or under it. 4178 intptr_t zeroes_needed = st_off; 4179 4180 if (st_size < BytesPerInt) { 4181 // Look for subword stores which only partially initialize words. 4182 // If we find some, we must lay down some word-level zeroes first, 4183 // underneath the subword stores. 4184 // 4185 // Examples: 4186 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4187 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4188 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4189 // 4190 // Note: coalesce_subword_stores may have already done this, 4191 // if it was prompted by constant non-zero subword initializers. 4192 // But this case can still arise with non-constant stores. 4193 4194 intptr_t next_full_store = find_next_fullword_store(i, phase); 4195 4196 // In the examples above: 4197 // in(i) p q r s x y z 4198 // st_off 12 13 14 15 12 13 14 4199 // st_size 1 1 1 1 1 1 1 4200 // next_full_s. 12 16 16 16 16 16 16 4201 // z's_done 12 16 16 16 12 16 12 4202 // z's_needed 12 16 16 16 16 16 16 4203 // zsize 0 0 0 0 4 0 4 4204 if (next_full_store < 0) { 4205 // Conservative tack: Zero to end of current word. 4206 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4207 } else { 4208 // Zero to beginning of next fully initialized word. 4209 // Or, don't zero at all, if we are already in that word. 4210 assert(next_full_store >= zeroes_needed, "must go forward"); 4211 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4212 zeroes_needed = next_full_store; 4213 } 4214 } 4215 4216 if (zeroes_needed > zeroes_done) { 4217 intptr_t zsize = zeroes_needed - zeroes_done; 4218 // Do some incremental zeroing on rawmem, in parallel with inits. 4219 zeroes_done = align_down(zeroes_done, BytesPerInt); 4220 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4221 zeroes_done, zeroes_needed, 4222 phase); 4223 zeroes_done = zeroes_needed; 4224 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4225 do_zeroing = false; // leave the hole, next time 4226 } 4227 } 4228 4229 // Collect the store and move on: 4230 st->set_req(MemNode::Memory, inits); 4231 inits = st; // put it on the linearized chain 4232 set_req(i, zmem); // unhook from previous position 4233 4234 if (zeroes_done == st_off) 4235 zeroes_done = next_init_off; 4236 4237 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4238 4239 #ifdef ASSERT 4240 // Various order invariants. Weaker than stores_are_sane because 4241 // a large constant tile can be filled in by smaller non-constant stores. 4242 assert(st_off >= last_init_off, "inits do not reverse"); 4243 last_init_off = st_off; 4244 const Type* val = NULL; 4245 if (st_size >= BytesPerInt && 4246 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4247 (int)val->basic_type() < (int)T_OBJECT) { 4248 assert(st_off >= last_tile_end, "tiles do not overlap"); 4249 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4250 last_tile_end = MAX2(last_tile_end, next_init_off); 4251 } else { 4252 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4253 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4254 assert(st_off >= last_init_end, "inits do not overlap"); 4255 last_init_end = next_init_off; // it's a non-tile 4256 } 4257 #endif //ASSERT 4258 } 4259 4260 remove_extra_zeroes(); // clear out all the zmems left over 4261 add_req(inits); 4262 4263 if (!(UseTLAB && ZeroTLAB)) { 4264 // If anything remains to be zeroed, zero it all now. 4265 zeroes_done = align_down(zeroes_done, BytesPerInt); 4266 // if it is the last unused 4 bytes of an instance, forget about it 4267 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4268 if (zeroes_done + BytesPerLong >= size_limit) { 4269 AllocateNode* alloc = allocation(); 4270 assert(alloc != NULL, "must be present"); 4271 if (alloc != NULL && alloc->Opcode() == Op_Allocate) { 4272 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4273 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 4274 if (zeroes_done == k->layout_helper()) 4275 zeroes_done = size_limit; 4276 } 4277 } 4278 if (zeroes_done < size_limit) { 4279 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4280 zeroes_done, size_in_bytes, phase); 4281 } 4282 } 4283 4284 set_complete(phase); 4285 return rawmem; 4286 } 4287 4288 4289 #ifdef ASSERT 4290 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 4291 if (is_complete()) 4292 return true; // stores could be anything at this point 4293 assert(allocation() != NULL, "must be present"); 4294 intptr_t last_off = allocation()->minimum_header_size(); 4295 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4296 Node* st = in(i); 4297 intptr_t st_off = get_store_offset(st, phase); 4298 if (st_off < 0) continue; // ignore dead garbage 4299 if (last_off > st_off) { 4300 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4301 this->dump(2); 4302 assert(false, "ascending store offsets"); 4303 return false; 4304 } 4305 last_off = st_off + st->as_Store()->memory_size(); 4306 } 4307 return true; 4308 } 4309 #endif //ASSERT 4310 4311 4312 4313 4314 //============================MergeMemNode===================================== 4315 // 4316 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4317 // contributing store or call operations. Each contributor provides the memory 4318 // state for a particular "alias type" (see Compile::alias_type). For example, 4319 // if a MergeMem has an input X for alias category #6, then any memory reference 4320 // to alias category #6 may use X as its memory state input, as an exact equivalent 4321 // to using the MergeMem as a whole. 4322 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4323 // 4324 // (Here, the <N> notation gives the index of the relevant adr_type.) 4325 // 4326 // In one special case (and more cases in the future), alias categories overlap. 4327 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4328 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4329 // it is exactly equivalent to that state W: 4330 // MergeMem(<Bot>: W) <==> W 4331 // 4332 // Usually, the merge has more than one input. In that case, where inputs 4333 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4334 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4335 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4336 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4337 // 4338 // A merge can take a "wide" memory state as one of its narrow inputs. 4339 // This simply means that the merge observes out only the relevant parts of 4340 // the wide input. That is, wide memory states arriving at narrow merge inputs 4341 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4342 // 4343 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4344 // and that memory slices "leak through": 4345 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4346 // 4347 // But, in such a cascade, repeated memory slices can "block the leak": 4348 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4349 // 4350 // In the last example, Y is not part of the combined memory state of the 4351 // outermost MergeMem. The system must, of course, prevent unschedulable 4352 // memory states from arising, so you can be sure that the state Y is somehow 4353 // a precursor to state Y'. 4354 // 4355 // 4356 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4357 // of each MergeMemNode array are exactly the numerical alias indexes, including 4358 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4359 // Compile::alias_type (and kin) produce and manage these indexes. 4360 // 4361 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4362 // (Note that this provides quick access to the top node inside MergeMem methods, 4363 // without the need to reach out via TLS to Compile::current.) 4364 // 4365 // As a consequence of what was just described, a MergeMem that represents a full 4366 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4367 // containing all alias categories. 4368 // 4369 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4370 // 4371 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4372 // a memory state for the alias type <N>, or else the top node, meaning that 4373 // there is no particular input for that alias type. Note that the length of 4374 // a MergeMem is variable, and may be extended at any time to accommodate new 4375 // memory states at larger alias indexes. When merges grow, they are of course 4376 // filled with "top" in the unused in() positions. 4377 // 4378 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4379 // (Top was chosen because it works smoothly with passes like GCM.) 4380 // 4381 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4382 // the type of random VM bits like TLS references.) Since it is always the 4383 // first non-Bot memory slice, some low-level loops use it to initialize an 4384 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4385 // 4386 // 4387 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4388 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4389 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4390 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4391 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4392 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4393 // 4394 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4395 // really that different from the other memory inputs. An abbreviation called 4396 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4397 // 4398 // 4399 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4400 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4401 // that "emerges though" the base memory will be marked as excluding the alias types 4402 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4403 // 4404 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4405 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4406 // 4407 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4408 // (It is currently unimplemented.) As you can see, the resulting merge is 4409 // actually a disjoint union of memory states, rather than an overlay. 4410 // 4411 4412 //------------------------------MergeMemNode----------------------------------- 4413 Node* MergeMemNode::make_empty_memory() { 4414 Node* empty_memory = (Node*) Compile::current()->top(); 4415 assert(empty_memory->is_top(), "correct sentinel identity"); 4416 return empty_memory; 4417 } 4418 4419 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4420 init_class_id(Class_MergeMem); 4421 // all inputs are nullified in Node::Node(int) 4422 // set_input(0, NULL); // no control input 4423 4424 // Initialize the edges uniformly to top, for starters. 4425 Node* empty_mem = make_empty_memory(); 4426 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4427 init_req(i,empty_mem); 4428 } 4429 assert(empty_memory() == empty_mem, ""); 4430 4431 if( new_base != NULL && new_base->is_MergeMem() ) { 4432 MergeMemNode* mdef = new_base->as_MergeMem(); 4433 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4434 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4435 mms.set_memory(mms.memory2()); 4436 } 4437 assert(base_memory() == mdef->base_memory(), ""); 4438 } else { 4439 set_base_memory(new_base); 4440 } 4441 } 4442 4443 // Make a new, untransformed MergeMem with the same base as 'mem'. 4444 // If mem is itself a MergeMem, populate the result with the same edges. 4445 MergeMemNode* MergeMemNode::make(Node* mem) { 4446 return new MergeMemNode(mem); 4447 } 4448 4449 //------------------------------cmp-------------------------------------------- 4450 uint MergeMemNode::hash() const { return NO_HASH; } 4451 uint MergeMemNode::cmp( const Node &n ) const { 4452 return (&n == this); // Always fail except on self 4453 } 4454 4455 //------------------------------Identity--------------------------------------- 4456 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4457 // Identity if this merge point does not record any interesting memory 4458 // disambiguations. 4459 Node* base_mem = base_memory(); 4460 Node* empty_mem = empty_memory(); 4461 if (base_mem != empty_mem) { // Memory path is not dead? 4462 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4463 Node* mem = in(i); 4464 if (mem != empty_mem && mem != base_mem) { 4465 return this; // Many memory splits; no change 4466 } 4467 } 4468 } 4469 return base_mem; // No memory splits; ID on the one true input 4470 } 4471 4472 //------------------------------Ideal------------------------------------------ 4473 // This method is invoked recursively on chains of MergeMem nodes 4474 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4475 // Remove chain'd MergeMems 4476 // 4477 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4478 // relative to the "in(Bot)". Since we are patching both at the same time, 4479 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4480 // but rewrite each "in(i)" relative to the new "in(Bot)". 4481 Node *progress = NULL; 4482 4483 4484 Node* old_base = base_memory(); 4485 Node* empty_mem = empty_memory(); 4486 if (old_base == empty_mem) 4487 return NULL; // Dead memory path. 4488 4489 MergeMemNode* old_mbase; 4490 if (old_base != NULL && old_base->is_MergeMem()) 4491 old_mbase = old_base->as_MergeMem(); 4492 else 4493 old_mbase = NULL; 4494 Node* new_base = old_base; 4495 4496 // simplify stacked MergeMems in base memory 4497 if (old_mbase) new_base = old_mbase->base_memory(); 4498 4499 // the base memory might contribute new slices beyond my req() 4500 if (old_mbase) grow_to_match(old_mbase); 4501 4502 // Look carefully at the base node if it is a phi. 4503 PhiNode* phi_base; 4504 if (new_base != NULL && new_base->is_Phi()) 4505 phi_base = new_base->as_Phi(); 4506 else 4507 phi_base = NULL; 4508 4509 Node* phi_reg = NULL; 4510 uint phi_len = (uint)-1; 4511 if (phi_base != NULL && !phi_base->is_copy()) { 4512 // do not examine phi if degraded to a copy 4513 phi_reg = phi_base->region(); 4514 phi_len = phi_base->req(); 4515 // see if the phi is unfinished 4516 for (uint i = 1; i < phi_len; i++) { 4517 if (phi_base->in(i) == NULL) { 4518 // incomplete phi; do not look at it yet! 4519 phi_reg = NULL; 4520 phi_len = (uint)-1; 4521 break; 4522 } 4523 } 4524 } 4525 4526 // Note: We do not call verify_sparse on entry, because inputs 4527 // can normalize to the base_memory via subsume_node or similar 4528 // mechanisms. This method repairs that damage. 4529 4530 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4531 4532 // Look at each slice. 4533 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4534 Node* old_in = in(i); 4535 // calculate the old memory value 4536 Node* old_mem = old_in; 4537 if (old_mem == empty_mem) old_mem = old_base; 4538 assert(old_mem == memory_at(i), ""); 4539 4540 // maybe update (reslice) the old memory value 4541 4542 // simplify stacked MergeMems 4543 Node* new_mem = old_mem; 4544 MergeMemNode* old_mmem; 4545 if (old_mem != NULL && old_mem->is_MergeMem()) 4546 old_mmem = old_mem->as_MergeMem(); 4547 else 4548 old_mmem = NULL; 4549 if (old_mmem == this) { 4550 // This can happen if loops break up and safepoints disappear. 4551 // A merge of BotPtr (default) with a RawPtr memory derived from a 4552 // safepoint can be rewritten to a merge of the same BotPtr with 4553 // the BotPtr phi coming into the loop. If that phi disappears 4554 // also, we can end up with a self-loop of the mergemem. 4555 // In general, if loops degenerate and memory effects disappear, 4556 // a mergemem can be left looking at itself. This simply means 4557 // that the mergemem's default should be used, since there is 4558 // no longer any apparent effect on this slice. 4559 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4560 // from start. Update the input to TOP. 4561 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4562 } 4563 else if (old_mmem != NULL) { 4564 new_mem = old_mmem->memory_at(i); 4565 } 4566 // else preceding memory was not a MergeMem 4567 4568 // replace equivalent phis (unfortunately, they do not GVN together) 4569 if (new_mem != NULL && new_mem != new_base && 4570 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4571 if (new_mem->is_Phi()) { 4572 PhiNode* phi_mem = new_mem->as_Phi(); 4573 for (uint i = 1; i < phi_len; i++) { 4574 if (phi_base->in(i) != phi_mem->in(i)) { 4575 phi_mem = NULL; 4576 break; 4577 } 4578 } 4579 if (phi_mem != NULL) { 4580 // equivalent phi nodes; revert to the def 4581 new_mem = new_base; 4582 } 4583 } 4584 } 4585 4586 // maybe store down a new value 4587 Node* new_in = new_mem; 4588 if (new_in == new_base) new_in = empty_mem; 4589 4590 if (new_in != old_in) { 4591 // Warning: Do not combine this "if" with the previous "if" 4592 // A memory slice might have be be rewritten even if it is semantically 4593 // unchanged, if the base_memory value has changed. 4594 set_req(i, new_in); 4595 progress = this; // Report progress 4596 } 4597 } 4598 4599 if (new_base != old_base) { 4600 set_req(Compile::AliasIdxBot, new_base); 4601 // Don't use set_base_memory(new_base), because we need to update du. 4602 assert(base_memory() == new_base, ""); 4603 progress = this; 4604 } 4605 4606 if( base_memory() == this ) { 4607 // a self cycle indicates this memory path is dead 4608 set_req(Compile::AliasIdxBot, empty_mem); 4609 } 4610 4611 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4612 // Recursion must occur after the self cycle check above 4613 if( base_memory()->is_MergeMem() ) { 4614 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4615 Node *m = phase->transform(new_mbase); // Rollup any cycles 4616 if( m != NULL && 4617 (m->is_top() || 4618 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4619 // propagate rollup of dead cycle to self 4620 set_req(Compile::AliasIdxBot, empty_mem); 4621 } 4622 } 4623 4624 if( base_memory() == empty_mem ) { 4625 progress = this; 4626 // Cut inputs during Parse phase only. 4627 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4628 if( !can_reshape ) { 4629 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4630 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4631 } 4632 } 4633 } 4634 4635 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4636 // Check if PhiNode::Ideal's "Split phis through memory merges" 4637 // transform should be attempted. Look for this->phi->this cycle. 4638 uint merge_width = req(); 4639 if (merge_width > Compile::AliasIdxRaw) { 4640 PhiNode* phi = base_memory()->as_Phi(); 4641 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4642 if (phi->in(i) == this) { 4643 phase->is_IterGVN()->_worklist.push(phi); 4644 break; 4645 } 4646 } 4647 } 4648 } 4649 4650 assert(progress || verify_sparse(), "please, no dups of base"); 4651 return progress; 4652 } 4653 4654 //-------------------------set_base_memory------------------------------------- 4655 void MergeMemNode::set_base_memory(Node *new_base) { 4656 Node* empty_mem = empty_memory(); 4657 set_req(Compile::AliasIdxBot, new_base); 4658 assert(memory_at(req()) == new_base, "must set default memory"); 4659 // Clear out other occurrences of new_base: 4660 if (new_base != empty_mem) { 4661 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4662 if (in(i) == new_base) set_req(i, empty_mem); 4663 } 4664 } 4665 } 4666 4667 //------------------------------out_RegMask------------------------------------ 4668 const RegMask &MergeMemNode::out_RegMask() const { 4669 return RegMask::Empty; 4670 } 4671 4672 //------------------------------dump_spec-------------------------------------- 4673 #ifndef PRODUCT 4674 void MergeMemNode::dump_spec(outputStream *st) const { 4675 st->print(" {"); 4676 Node* base_mem = base_memory(); 4677 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4678 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4679 if (mem == base_mem) { st->print(" -"); continue; } 4680 st->print( " N%d:", mem->_idx ); 4681 Compile::current()->get_adr_type(i)->dump_on(st); 4682 } 4683 st->print(" }"); 4684 } 4685 #endif // !PRODUCT 4686 4687 4688 #ifdef ASSERT 4689 static bool might_be_same(Node* a, Node* b) { 4690 if (a == b) return true; 4691 if (!(a->is_Phi() || b->is_Phi())) return false; 4692 // phis shift around during optimization 4693 return true; // pretty stupid... 4694 } 4695 4696 // verify a narrow slice (either incoming or outgoing) 4697 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4698 if (!VerifyAliases) return; // don't bother to verify unless requested 4699 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4700 if (Node::in_dump()) return; // muzzle asserts when printing 4701 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4702 assert(n != NULL, ""); 4703 // Elide intervening MergeMem's 4704 while (n->is_MergeMem()) { 4705 n = n->as_MergeMem()->memory_at(alias_idx); 4706 } 4707 Compile* C = Compile::current(); 4708 const TypePtr* n_adr_type = n->adr_type(); 4709 if (n == m->empty_memory()) { 4710 // Implicit copy of base_memory() 4711 } else if (n_adr_type != TypePtr::BOTTOM) { 4712 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4713 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4714 } else { 4715 // A few places like make_runtime_call "know" that VM calls are narrow, 4716 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4717 bool expected_wide_mem = false; 4718 if (n == m->base_memory()) { 4719 expected_wide_mem = true; 4720 } else if (alias_idx == Compile::AliasIdxRaw || 4721 n == m->memory_at(Compile::AliasIdxRaw)) { 4722 expected_wide_mem = true; 4723 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4724 // memory can "leak through" calls on channels that 4725 // are write-once. Allow this also. 4726 expected_wide_mem = true; 4727 } 4728 assert(expected_wide_mem, "expected narrow slice replacement"); 4729 } 4730 } 4731 #else // !ASSERT 4732 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4733 #endif 4734 4735 4736 //-----------------------------memory_at--------------------------------------- 4737 Node* MergeMemNode::memory_at(uint alias_idx) const { 4738 assert(alias_idx >= Compile::AliasIdxRaw || 4739 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4740 "must avoid base_memory and AliasIdxTop"); 4741 4742 // Otherwise, it is a narrow slice. 4743 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4744 Compile *C = Compile::current(); 4745 if (is_empty_memory(n)) { 4746 // the array is sparse; empty slots are the "top" node 4747 n = base_memory(); 4748 assert(Node::in_dump() 4749 || n == NULL || n->bottom_type() == Type::TOP 4750 || n->adr_type() == NULL // address is TOP 4751 || n->adr_type() == TypePtr::BOTTOM 4752 || n->adr_type() == TypeRawPtr::BOTTOM 4753 || Compile::current()->AliasLevel() == 0, 4754 "must be a wide memory"); 4755 // AliasLevel == 0 if we are organizing the memory states manually. 4756 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4757 } else { 4758 // make sure the stored slice is sane 4759 #ifdef ASSERT 4760 if (VMError::is_error_reported() || Node::in_dump()) { 4761 } else if (might_be_same(n, base_memory())) { 4762 // Give it a pass: It is a mostly harmless repetition of the base. 4763 // This can arise normally from node subsumption during optimization. 4764 } else { 4765 verify_memory_slice(this, alias_idx, n); 4766 } 4767 #endif 4768 } 4769 return n; 4770 } 4771 4772 //---------------------------set_memory_at------------------------------------- 4773 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4774 verify_memory_slice(this, alias_idx, n); 4775 Node* empty_mem = empty_memory(); 4776 if (n == base_memory()) n = empty_mem; // collapse default 4777 uint need_req = alias_idx+1; 4778 if (req() < need_req) { 4779 if (n == empty_mem) return; // already the default, so do not grow me 4780 // grow the sparse array 4781 do { 4782 add_req(empty_mem); 4783 } while (req() < need_req); 4784 } 4785 set_req( alias_idx, n ); 4786 } 4787 4788 4789 4790 //--------------------------iteration_setup------------------------------------ 4791 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4792 if (other != NULL) { 4793 grow_to_match(other); 4794 // invariant: the finite support of mm2 is within mm->req() 4795 #ifdef ASSERT 4796 for (uint i = req(); i < other->req(); i++) { 4797 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4798 } 4799 #endif 4800 } 4801 // Replace spurious copies of base_memory by top. 4802 Node* base_mem = base_memory(); 4803 if (base_mem != NULL && !base_mem->is_top()) { 4804 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4805 if (in(i) == base_mem) 4806 set_req(i, empty_memory()); 4807 } 4808 } 4809 } 4810 4811 //---------------------------grow_to_match------------------------------------- 4812 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4813 Node* empty_mem = empty_memory(); 4814 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4815 // look for the finite support of the other memory 4816 for (uint i = other->req(); --i >= req(); ) { 4817 if (other->in(i) != empty_mem) { 4818 uint new_len = i+1; 4819 while (req() < new_len) add_req(empty_mem); 4820 break; 4821 } 4822 } 4823 } 4824 4825 //---------------------------verify_sparse------------------------------------- 4826 #ifndef PRODUCT 4827 bool MergeMemNode::verify_sparse() const { 4828 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4829 Node* base_mem = base_memory(); 4830 // The following can happen in degenerate cases, since empty==top. 4831 if (is_empty_memory(base_mem)) return true; 4832 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4833 assert(in(i) != NULL, "sane slice"); 4834 if (in(i) == base_mem) return false; // should have been the sentinel value! 4835 } 4836 return true; 4837 } 4838 4839 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4840 Node* n; 4841 n = mm->in(idx); 4842 if (mem == n) return true; // might be empty_memory() 4843 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4844 if (mem == n) return true; 4845 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4846 if (mem == n) return true; 4847 if (n == NULL) break; 4848 } 4849 return false; 4850 } 4851 #endif // !PRODUCT