1 /* 2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/arraycopynode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/compile.hpp" 35 #include "opto/connode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/loopnode.hpp" 38 #include "opto/machnode.hpp" 39 #include "opto/matcher.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/mulnode.hpp" 42 #include "opto/narrowptrnode.hpp" 43 #include "opto/phaseX.hpp" 44 #include "opto/regmask.hpp" 45 #include "utilities/copy.hpp" 46 47 // Portions of code courtesy of Clifford Click 48 49 // Optimization - Graph Style 50 51 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 52 53 //============================================================================= 54 uint MemNode::size_of() const { return sizeof(*this); } 55 56 const TypePtr *MemNode::adr_type() const { 57 Node* adr = in(Address); 58 if (adr == NULL) return NULL; // node is dead 59 const TypePtr* cross_check = NULL; 60 DEBUG_ONLY(cross_check = _adr_type); 61 return calculate_adr_type(adr->bottom_type(), cross_check); 62 } 63 64 #ifndef PRODUCT 65 void MemNode::dump_spec(outputStream *st) const { 66 if (in(Address) == NULL) return; // node is dead 67 #ifndef ASSERT 68 // fake the missing field 69 const TypePtr* _adr_type = NULL; 70 if (in(Address) != NULL) 71 _adr_type = in(Address)->bottom_type()->isa_ptr(); 72 #endif 73 dump_adr_type(this, _adr_type, st); 74 75 Compile* C = Compile::current(); 76 if (C->alias_type(_adr_type)->is_volatile()) { 77 st->print(" Volatile!"); 78 } 79 if (_unaligned_access) { 80 st->print(" unaligned"); 81 } 82 if (_mismatched_access) { 83 st->print(" mismatched"); 84 } 85 } 86 87 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 88 st->print(" @"); 89 if (adr_type == NULL) { 90 st->print("NULL"); 91 } else { 92 adr_type->dump_on(st); 93 Compile* C = Compile::current(); 94 Compile::AliasType* atp = NULL; 95 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 96 if (atp == NULL) 97 st->print(", idx=?\?;"); 98 else if (atp->index() == Compile::AliasIdxBot) 99 st->print(", idx=Bot;"); 100 else if (atp->index() == Compile::AliasIdxTop) 101 st->print(", idx=Top;"); 102 else if (atp->index() == Compile::AliasIdxRaw) 103 st->print(", idx=Raw;"); 104 else { 105 ciField* field = atp->field(); 106 if (field) { 107 st->print(", name="); 108 field->print_name_on(st); 109 } 110 st->print(", idx=%d;", atp->index()); 111 } 112 } 113 } 114 115 extern void print_alias_types(); 116 117 #endif 118 119 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 120 assert((t_oop != NULL), "sanity"); 121 bool is_instance = t_oop->is_known_instance_field(); 122 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 123 (load != NULL) && load->is_Load() && 124 (phase->is_IterGVN() != NULL); 125 if (!(is_instance || is_boxed_value_load)) 126 return mchain; // don't try to optimize non-instance types 127 uint instance_id = t_oop->instance_id(); 128 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 129 Node *prev = NULL; 130 Node *result = mchain; 131 while (prev != result) { 132 prev = result; 133 if (result == start_mem) 134 break; // hit one of our sentinels 135 // skip over a call which does not affect this memory slice 136 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 137 Node *proj_in = result->in(0); 138 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 139 break; // hit one of our sentinels 140 } else if (proj_in->is_Call()) { 141 // ArrayCopyNodes processed here as well 142 CallNode *call = proj_in->as_Call(); 143 if (!call->may_modify(t_oop, phase)) { // returns false for instances 144 result = call->in(TypeFunc::Memory); 145 } 146 } else if (proj_in->is_Initialize()) { 147 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 148 // Stop if this is the initialization for the object instance which 149 // contains this memory slice, otherwise skip over it. 150 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 151 break; 152 } 153 if (is_instance) { 154 result = proj_in->in(TypeFunc::Memory); 155 } else if (is_boxed_value_load) { 156 Node* klass = alloc->in(AllocateNode::KlassNode); 157 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 158 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 159 result = proj_in->in(TypeFunc::Memory); // not related allocation 160 } 161 } 162 } else if (proj_in->is_MemBar()) { 163 ArrayCopyNode* ac = NULL; 164 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 165 break; 166 } 167 result = proj_in->in(TypeFunc::Memory); 168 } else { 169 assert(false, "unexpected projection"); 170 } 171 } else if (result->is_ClearArray()) { 172 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 173 // Can not bypass initialization of the instance 174 // we are looking for. 175 break; 176 } 177 // Otherwise skip it (the call updated 'result' value). 178 } else if (result->is_MergeMem()) { 179 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 180 } 181 } 182 return result; 183 } 184 185 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 186 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 187 if (t_oop == NULL) 188 return mchain; // don't try to optimize non-oop types 189 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 190 bool is_instance = t_oop->is_known_instance_field(); 191 PhaseIterGVN *igvn = phase->is_IterGVN(); 192 if (is_instance && igvn != NULL && result->is_Phi()) { 193 PhiNode *mphi = result->as_Phi(); 194 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 195 const TypePtr *t = mphi->adr_type(); 196 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 197 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 198 t->is_oopptr()->cast_to_exactness(true) 199 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 200 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 201 // clone the Phi with our address type 202 result = mphi->split_out_instance(t_adr, igvn); 203 } else { 204 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 205 } 206 } 207 return result; 208 } 209 210 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 211 uint alias_idx = phase->C->get_alias_index(tp); 212 Node *mem = mmem; 213 #ifdef ASSERT 214 { 215 // Check that current type is consistent with the alias index used during graph construction 216 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 217 bool consistent = adr_check == NULL || adr_check->empty() || 218 phase->C->must_alias(adr_check, alias_idx ); 219 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 220 if( !consistent && adr_check != NULL && !adr_check->empty() && 221 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 222 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 223 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 224 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 225 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 226 // don't assert if it is dead code. 227 consistent = true; 228 } 229 if( !consistent ) { 230 st->print("alias_idx==%d, adr_check==", alias_idx); 231 if( adr_check == NULL ) { 232 st->print("NULL"); 233 } else { 234 adr_check->dump(); 235 } 236 st->cr(); 237 print_alias_types(); 238 assert(consistent, "adr_check must match alias idx"); 239 } 240 } 241 #endif 242 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 243 // means an array I have not precisely typed yet. Do not do any 244 // alias stuff with it any time soon. 245 const TypeOopPtr *toop = tp->isa_oopptr(); 246 if( tp->base() != Type::AnyPtr && 247 !(toop && 248 toop->klass() != NULL && 249 toop->klass()->is_java_lang_Object() && 250 toop->offset() == Type::OffsetBot) ) { 251 // compress paths and change unreachable cycles to TOP 252 // If not, we can update the input infinitely along a MergeMem cycle 253 // Equivalent code in PhiNode::Ideal 254 Node* m = phase->transform(mmem); 255 // If transformed to a MergeMem, get the desired slice 256 // Otherwise the returned node represents memory for every slice 257 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 258 // Update input if it is progress over what we have now 259 } 260 return mem; 261 } 262 263 //--------------------------Ideal_common--------------------------------------- 264 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 265 // Unhook non-raw memories from complete (macro-expanded) initializations. 266 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 267 // If our control input is a dead region, kill all below the region 268 Node *ctl = in(MemNode::Control); 269 if (ctl && remove_dead_region(phase, can_reshape)) 270 return this; 271 ctl = in(MemNode::Control); 272 // Don't bother trying to transform a dead node 273 if (ctl && ctl->is_top()) return NodeSentinel; 274 275 PhaseIterGVN *igvn = phase->is_IterGVN(); 276 // Wait if control on the worklist. 277 if (ctl && can_reshape && igvn != NULL) { 278 Node* bol = NULL; 279 Node* cmp = NULL; 280 if (ctl->in(0)->is_If()) { 281 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 282 bol = ctl->in(0)->in(1); 283 if (bol->is_Bool()) 284 cmp = ctl->in(0)->in(1)->in(1); 285 } 286 if (igvn->_worklist.member(ctl) || 287 (bol != NULL && igvn->_worklist.member(bol)) || 288 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 289 // This control path may be dead. 290 // Delay this memory node transformation until the control is processed. 291 phase->is_IterGVN()->_worklist.push(this); 292 return NodeSentinel; // caller will return NULL 293 } 294 } 295 // Ignore if memory is dead, or self-loop 296 Node *mem = in(MemNode::Memory); 297 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 298 assert(mem != this, "dead loop in MemNode::Ideal"); 299 300 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 301 // This memory slice may be dead. 302 // Delay this mem node transformation until the memory is processed. 303 phase->is_IterGVN()->_worklist.push(this); 304 return NodeSentinel; // caller will return NULL 305 } 306 307 Node *address = in(MemNode::Address); 308 const Type *t_adr = phase->type(address); 309 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 310 311 if (can_reshape && igvn != NULL && 312 (igvn->_worklist.member(address) || 313 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { 314 // The address's base and type may change when the address is processed. 315 // Delay this mem node transformation until the address is processed. 316 phase->is_IterGVN()->_worklist.push(this); 317 return NodeSentinel; // caller will return NULL 318 } 319 320 // Do NOT remove or optimize the next lines: ensure a new alias index 321 // is allocated for an oop pointer type before Escape Analysis. 322 // Note: C++ will not remove it since the call has side effect. 323 if (t_adr->isa_oopptr()) { 324 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 325 } 326 327 Node* base = NULL; 328 if (address->is_AddP()) { 329 base = address->in(AddPNode::Base); 330 } 331 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 332 !t_adr->isa_rawptr()) { 333 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 334 // Skip this node optimization if its address has TOP base. 335 return NodeSentinel; // caller will return NULL 336 } 337 338 // Avoid independent memory operations 339 Node* old_mem = mem; 340 341 // The code which unhooks non-raw memories from complete (macro-expanded) 342 // initializations was removed. After macro-expansion all stores catched 343 // by Initialize node became raw stores and there is no information 344 // which memory slices they modify. So it is unsafe to move any memory 345 // operation above these stores. Also in most cases hooked non-raw memories 346 // were already unhooked by using information from detect_ptr_independence() 347 // and find_previous_store(). 348 349 if (mem->is_MergeMem()) { 350 MergeMemNode* mmem = mem->as_MergeMem(); 351 const TypePtr *tp = t_adr->is_ptr(); 352 353 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 354 } 355 356 if (mem != old_mem) { 357 set_req(MemNode::Memory, mem); 358 if (can_reshape && old_mem->outcnt() == 0) { 359 igvn->_worklist.push(old_mem); 360 } 361 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 362 return this; 363 } 364 365 // let the subclass continue analyzing... 366 return NULL; 367 } 368 369 // Helper function for proving some simple control dominations. 370 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 371 // Already assumes that 'dom' is available at 'sub', and that 'sub' 372 // is not a constant (dominated by the method's StartNode). 373 // Used by MemNode::find_previous_store to prove that the 374 // control input of a memory operation predates (dominates) 375 // an allocation it wants to look past. 376 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 377 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 378 return false; // Conservative answer for dead code 379 380 // Check 'dom'. Skip Proj and CatchProj nodes. 381 dom = dom->find_exact_control(dom); 382 if (dom == NULL || dom->is_top()) 383 return false; // Conservative answer for dead code 384 385 if (dom == sub) { 386 // For the case when, for example, 'sub' is Initialize and the original 387 // 'dom' is Proj node of the 'sub'. 388 return false; 389 } 390 391 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 392 return true; 393 394 // 'dom' dominates 'sub' if its control edge and control edges 395 // of all its inputs dominate or equal to sub's control edge. 396 397 // Currently 'sub' is either Allocate, Initialize or Start nodes. 398 // Or Region for the check in LoadNode::Ideal(); 399 // 'sub' should have sub->in(0) != NULL. 400 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 401 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 402 403 // Get control edge of 'sub'. 404 Node* orig_sub = sub; 405 sub = sub->find_exact_control(sub->in(0)); 406 if (sub == NULL || sub->is_top()) 407 return false; // Conservative answer for dead code 408 409 assert(sub->is_CFG(), "expecting control"); 410 411 if (sub == dom) 412 return true; 413 414 if (sub->is_Start() || sub->is_Root()) 415 return false; 416 417 { 418 // Check all control edges of 'dom'. 419 420 ResourceMark rm; 421 Arena* arena = Thread::current()->resource_area(); 422 Node_List nlist(arena); 423 Unique_Node_List dom_list(arena); 424 425 dom_list.push(dom); 426 bool only_dominating_controls = false; 427 428 for (uint next = 0; next < dom_list.size(); next++) { 429 Node* n = dom_list.at(next); 430 if (n == orig_sub) 431 return false; // One of dom's inputs dominated by sub. 432 if (!n->is_CFG() && n->pinned()) { 433 // Check only own control edge for pinned non-control nodes. 434 n = n->find_exact_control(n->in(0)); 435 if (n == NULL || n->is_top()) 436 return false; // Conservative answer for dead code 437 assert(n->is_CFG(), "expecting control"); 438 dom_list.push(n); 439 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 440 only_dominating_controls = true; 441 } else if (n->is_CFG()) { 442 if (n->dominates(sub, nlist)) 443 only_dominating_controls = true; 444 else 445 return false; 446 } else { 447 // First, own control edge. 448 Node* m = n->find_exact_control(n->in(0)); 449 if (m != NULL) { 450 if (m->is_top()) 451 return false; // Conservative answer for dead code 452 dom_list.push(m); 453 } 454 // Now, the rest of edges. 455 uint cnt = n->req(); 456 for (uint i = 1; i < cnt; i++) { 457 m = n->find_exact_control(n->in(i)); 458 if (m == NULL || m->is_top()) 459 continue; 460 dom_list.push(m); 461 } 462 } 463 } 464 return only_dominating_controls; 465 } 466 } 467 468 //---------------------detect_ptr_independence--------------------------------- 469 // Used by MemNode::find_previous_store to prove that two base 470 // pointers are never equal. 471 // The pointers are accompanied by their associated allocations, 472 // if any, which have been previously discovered by the caller. 473 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 474 Node* p2, AllocateNode* a2, 475 PhaseTransform* phase) { 476 // Attempt to prove that these two pointers cannot be aliased. 477 // They may both manifestly be allocations, and they should differ. 478 // Or, if they are not both allocations, they can be distinct constants. 479 // Otherwise, one is an allocation and the other a pre-existing value. 480 if (a1 == NULL && a2 == NULL) { // neither an allocation 481 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 482 } else if (a1 != NULL && a2 != NULL) { // both allocations 483 return (a1 != a2); 484 } else if (a1 != NULL) { // one allocation a1 485 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 486 return all_controls_dominate(p2, a1); 487 } else { //(a2 != NULL) // one allocation a2 488 return all_controls_dominate(p1, a2); 489 } 490 return false; 491 } 492 493 494 // Find an arraycopy that must have set (can_see_stored_value=true) or 495 // could have set (can_see_stored_value=false) the value for this load 496 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 497 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 498 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 499 Node* mb = mem->in(0); 500 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 501 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 502 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 503 if (ac->is_clonebasic()) { 504 intptr_t offset; 505 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 506 assert(alloc != NULL && (!ReduceBulkZeroing || alloc->initialization()->is_complete_with_arraycopy()), "broken allocation"); 507 if (alloc == ld_alloc) { 508 return ac; 509 } 510 } 511 } 512 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 513 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 514 515 if (ac->is_arraycopy_validated() || 516 ac->is_copyof_validated() || 517 ac->is_copyofrange_validated()) { 518 Node* ld_addp = in(MemNode::Address); 519 if (ld_addp->is_AddP()) { 520 Node* ld_base = ld_addp->in(AddPNode::Address); 521 Node* ld_offs = ld_addp->in(AddPNode::Offset); 522 523 Node* dest = ac->in(ArrayCopyNode::Dest); 524 525 if (dest == ld_base) { 526 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 527 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 528 return ac; 529 } 530 if (!can_see_stored_value) { 531 mem = ac->in(TypeFunc::Memory); 532 } 533 } 534 } 535 } 536 } 537 return NULL; 538 } 539 540 // The logic for reordering loads and stores uses four steps: 541 // (a) Walk carefully past stores and initializations which we 542 // can prove are independent of this load. 543 // (b) Observe that the next memory state makes an exact match 544 // with self (load or store), and locate the relevant store. 545 // (c) Ensure that, if we were to wire self directly to the store, 546 // the optimizer would fold it up somehow. 547 // (d) Do the rewiring, and return, depending on some other part of 548 // the optimizer to fold up the load. 549 // This routine handles steps (a) and (b). Steps (c) and (d) are 550 // specific to loads and stores, so they are handled by the callers. 551 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 552 // 553 Node* MemNode::find_previous_store(PhaseTransform* phase) { 554 Node* ctrl = in(MemNode::Control); 555 Node* adr = in(MemNode::Address); 556 intptr_t offset = 0; 557 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 558 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 559 560 if (offset == Type::OffsetBot) 561 return NULL; // cannot unalias unless there are precise offsets 562 563 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 564 565 intptr_t size_in_bytes = memory_size(); 566 567 Node* mem = in(MemNode::Memory); // start searching here... 568 569 int cnt = 50; // Cycle limiter 570 for (;;) { // While we can dance past unrelated stores... 571 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 572 573 Node* prev = mem; 574 if (mem->is_Store()) { 575 Node* st_adr = mem->in(MemNode::Address); 576 intptr_t st_offset = 0; 577 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 578 if (st_base == NULL) 579 break; // inscrutable pointer 580 if (st_offset != offset && st_offset != Type::OffsetBot) { 581 const int MAX_STORE = BytesPerLong; 582 if (st_offset >= offset + size_in_bytes || 583 st_offset <= offset - MAX_STORE || 584 st_offset <= offset - mem->as_Store()->memory_size()) { 585 // Success: The offsets are provably independent. 586 // (You may ask, why not just test st_offset != offset and be done? 587 // The answer is that stores of different sizes can co-exist 588 // in the same sequence of RawMem effects. We sometimes initialize 589 // a whole 'tile' of array elements with a single jint or jlong.) 590 mem = mem->in(MemNode::Memory); 591 continue; // (a) advance through independent store memory 592 } 593 } 594 if (st_base != base && 595 detect_ptr_independence(base, alloc, 596 st_base, 597 AllocateNode::Ideal_allocation(st_base, phase), 598 phase)) { 599 // Success: The bases are provably independent. 600 mem = mem->in(MemNode::Memory); 601 continue; // (a) advance through independent store memory 602 } 603 604 // (b) At this point, if the bases or offsets do not agree, we lose, 605 // since we have not managed to prove 'this' and 'mem' independent. 606 if (st_base == base && st_offset == offset) { 607 return mem; // let caller handle steps (c), (d) 608 } 609 610 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 611 InitializeNode* st_init = mem->in(0)->as_Initialize(); 612 AllocateNode* st_alloc = st_init->allocation(); 613 if (st_alloc == NULL) 614 break; // something degenerated 615 bool known_identical = false; 616 bool known_independent = false; 617 if (alloc == st_alloc) 618 known_identical = true; 619 else if (alloc != NULL) 620 known_independent = true; 621 else if (all_controls_dominate(this, st_alloc)) 622 known_independent = true; 623 624 if (known_independent) { 625 // The bases are provably independent: Either they are 626 // manifestly distinct allocations, or else the control 627 // of this load dominates the store's allocation. 628 int alias_idx = phase->C->get_alias_index(adr_type()); 629 if (alias_idx == Compile::AliasIdxRaw) { 630 mem = st_alloc->in(TypeFunc::Memory); 631 } else { 632 mem = st_init->memory(alias_idx); 633 } 634 continue; // (a) advance through independent store memory 635 } 636 637 // (b) at this point, if we are not looking at a store initializing 638 // the same allocation we are loading from, we lose. 639 if (known_identical) { 640 // From caller, can_see_stored_value will consult find_captured_store. 641 return mem; // let caller handle steps (c), (d) 642 } 643 644 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 645 if (prev != mem) { 646 // Found an arraycopy but it doesn't affect that load 647 continue; 648 } 649 // Found an arraycopy that may affect that load 650 return mem; 651 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 652 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 653 if (mem->is_Proj() && mem->in(0)->is_Call()) { 654 // ArrayCopyNodes processed here as well. 655 CallNode *call = mem->in(0)->as_Call(); 656 if (!call->may_modify(addr_t, phase)) { 657 mem = call->in(TypeFunc::Memory); 658 continue; // (a) advance through independent call memory 659 } 660 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 661 ArrayCopyNode* ac = NULL; 662 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 663 break; 664 } 665 mem = mem->in(0)->in(TypeFunc::Memory); 666 continue; // (a) advance through independent MemBar memory 667 } else if (mem->is_ClearArray()) { 668 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 669 // (the call updated 'mem' value) 670 continue; // (a) advance through independent allocation memory 671 } else { 672 // Can not bypass initialization of the instance 673 // we are looking for. 674 return mem; 675 } 676 } else if (mem->is_MergeMem()) { 677 int alias_idx = phase->C->get_alias_index(adr_type()); 678 mem = mem->as_MergeMem()->memory_at(alias_idx); 679 continue; // (a) advance through independent MergeMem memory 680 } 681 } 682 683 // Unless there is an explicit 'continue', we must bail out here, 684 // because 'mem' is an inscrutable memory state (e.g., a call). 685 break; 686 } 687 688 return NULL; // bail out 689 } 690 691 //----------------------calculate_adr_type------------------------------------- 692 // Helper function. Notices when the given type of address hits top or bottom. 693 // Also, asserts a cross-check of the type against the expected address type. 694 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 695 if (t == Type::TOP) return NULL; // does not touch memory any more? 696 #ifdef PRODUCT 697 cross_check = NULL; 698 #else 699 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 700 #endif 701 const TypePtr* tp = t->isa_ptr(); 702 if (tp == NULL) { 703 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 704 return TypePtr::BOTTOM; // touches lots of memory 705 } else { 706 #ifdef ASSERT 707 // %%%% [phh] We don't check the alias index if cross_check is 708 // TypeRawPtr::BOTTOM. Needs to be investigated. 709 if (cross_check != NULL && 710 cross_check != TypePtr::BOTTOM && 711 cross_check != TypeRawPtr::BOTTOM) { 712 // Recheck the alias index, to see if it has changed (due to a bug). 713 Compile* C = Compile::current(); 714 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 715 "must stay in the original alias category"); 716 // The type of the address must be contained in the adr_type, 717 // disregarding "null"-ness. 718 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 719 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 720 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 721 "real address must not escape from expected memory type"); 722 } 723 #endif 724 return tp; 725 } 726 } 727 728 //============================================================================= 729 // Should LoadNode::Ideal() attempt to remove control edges? 730 bool LoadNode::can_remove_control() const { 731 return true; 732 } 733 uint LoadNode::size_of() const { return sizeof(*this); } 734 uint LoadNode::cmp( const Node &n ) const 735 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 736 const Type *LoadNode::bottom_type() const { return _type; } 737 uint LoadNode::ideal_reg() const { 738 return _type->ideal_reg(); 739 } 740 741 #ifndef PRODUCT 742 void LoadNode::dump_spec(outputStream *st) const { 743 MemNode::dump_spec(st); 744 if( !Verbose && !WizardMode ) { 745 // standard dump does this in Verbose and WizardMode 746 st->print(" #"); _type->dump_on(st); 747 } 748 if (!depends_only_on_test()) { 749 st->print(" (does not depend only on test)"); 750 } 751 } 752 #endif 753 754 #ifdef ASSERT 755 //----------------------------is_immutable_value------------------------------- 756 // Helper function to allow a raw load without control edge for some cases 757 bool LoadNode::is_immutable_value(Node* adr) { 758 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 759 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 760 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 761 in_bytes(JavaThread::osthread_offset()))); 762 } 763 #endif 764 765 //----------------------------LoadNode::make----------------------------------- 766 // Polymorphic factory method: 767 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 768 ControlDependency control_dependency, bool unaligned, bool mismatched) { 769 Compile* C = gvn.C; 770 771 // sanity check the alias category against the created node type 772 assert(!(adr_type->isa_oopptr() && 773 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 774 "use LoadKlassNode instead"); 775 assert(!(adr_type->isa_aryptr() && 776 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 777 "use LoadRangeNode instead"); 778 // Check control edge of raw loads 779 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 780 // oop will be recorded in oop map if load crosses safepoint 781 rt->isa_oopptr() || is_immutable_value(adr), 782 "raw memory operations should have control edge"); 783 LoadNode* load = NULL; 784 switch (bt) { 785 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 786 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 787 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 788 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 789 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 790 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 791 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 792 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 793 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 794 case T_OBJECT: 795 #ifdef _LP64 796 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 797 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 798 } else 799 #endif 800 { 801 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 802 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 803 } 804 break; 805 } 806 assert(load != NULL, "LoadNode should have been created"); 807 if (unaligned) { 808 load->set_unaligned_access(); 809 } 810 if (mismatched) { 811 load->set_mismatched_access(); 812 } 813 if (load->Opcode() == Op_LoadN) { 814 Node* ld = gvn.transform(load); 815 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 816 } 817 818 return load; 819 } 820 821 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 822 ControlDependency control_dependency, bool unaligned, bool mismatched) { 823 bool require_atomic = true; 824 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 825 if (unaligned) { 826 load->set_unaligned_access(); 827 } 828 if (mismatched) { 829 load->set_mismatched_access(); 830 } 831 return load; 832 } 833 834 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 835 ControlDependency control_dependency, bool unaligned, bool mismatched) { 836 bool require_atomic = true; 837 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 838 if (unaligned) { 839 load->set_unaligned_access(); 840 } 841 if (mismatched) { 842 load->set_mismatched_access(); 843 } 844 return load; 845 } 846 847 848 849 //------------------------------hash------------------------------------------- 850 uint LoadNode::hash() const { 851 // unroll addition of interesting fields 852 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 853 } 854 855 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 856 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 857 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 858 bool is_stable_ary = FoldStableValues && 859 (tp != NULL) && (tp->isa_aryptr() != NULL) && 860 tp->isa_aryptr()->is_stable(); 861 862 return (eliminate_boxing && non_volatile) || is_stable_ary; 863 } 864 865 return false; 866 } 867 868 // Is the value loaded previously stored by an arraycopy? If so return 869 // a load node that reads from the source array so we may be able to 870 // optimize out the ArrayCopy node later. 871 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const { 872 Node* ld_adr = in(MemNode::Address); 873 intptr_t ld_off = 0; 874 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 875 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 876 if (ac != NULL) { 877 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 878 879 Node* ld = clone(); 880 if (ac->as_ArrayCopy()->is_clonebasic()) { 881 assert(ld_alloc != NULL, "need an alloc"); 882 Node* addp = in(MemNode::Address)->clone(); 883 assert(addp->is_AddP(), "address must be addp"); 884 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern"); 885 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern"); 886 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base)); 887 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address)); 888 ld->set_req(MemNode::Address, phase->transform(addp)); 889 if (in(0) != NULL) { 890 assert(ld_alloc->in(0) != NULL, "alloc must have control"); 891 ld->set_req(0, ld_alloc->in(0)); 892 } 893 } else { 894 Node* addp = in(MemNode::Address)->clone(); 895 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 896 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)); 897 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)); 898 899 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 900 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 901 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 902 uint shift = exact_log2(type2aelembytes(ary_elem)); 903 904 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 905 #ifdef _LP64 906 diff = phase->transform(new ConvI2LNode(diff)); 907 #endif 908 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 909 910 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 911 addp->set_req(AddPNode::Offset, offset); 912 ld->set_req(MemNode::Address, phase->transform(addp)); 913 914 if (in(0) != NULL) { 915 assert(ac->in(0) != NULL, "alloc must have control"); 916 ld->set_req(0, ac->in(0)); 917 } 918 } 919 // load depends on the tests that validate the arraycopy 920 ld->as_Load()->_control_dependency = Pinned; 921 return ld; 922 } 923 return NULL; 924 } 925 926 927 //---------------------------can_see_stored_value------------------------------ 928 // This routine exists to make sure this set of tests is done the same 929 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 930 // will change the graph shape in a way which makes memory alive twice at the 931 // same time (uses the Oracle model of aliasing), then some 932 // LoadXNode::Identity will fold things back to the equivalence-class model 933 // of aliasing. 934 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 935 Node* ld_adr = in(MemNode::Address); 936 intptr_t ld_off = 0; 937 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 938 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 939 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 940 // This is more general than load from boxing objects. 941 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 942 uint alias_idx = atp->index(); 943 bool final = !atp->is_rewritable(); 944 Node* result = NULL; 945 Node* current = st; 946 // Skip through chains of MemBarNodes checking the MergeMems for 947 // new states for the slice of this load. Stop once any other 948 // kind of node is encountered. Loads from final memory can skip 949 // through any kind of MemBar but normal loads shouldn't skip 950 // through MemBarAcquire since the could allow them to move out of 951 // a synchronized region. 952 while (current->is_Proj()) { 953 int opc = current->in(0)->Opcode(); 954 if ((final && (opc == Op_MemBarAcquire || 955 opc == Op_MemBarAcquireLock || 956 opc == Op_LoadFence)) || 957 opc == Op_MemBarRelease || 958 opc == Op_StoreFence || 959 opc == Op_MemBarReleaseLock || 960 opc == Op_MemBarStoreStore || 961 opc == Op_MemBarCPUOrder) { 962 Node* mem = current->in(0)->in(TypeFunc::Memory); 963 if (mem->is_MergeMem()) { 964 MergeMemNode* merge = mem->as_MergeMem(); 965 Node* new_st = merge->memory_at(alias_idx); 966 if (new_st == merge->base_memory()) { 967 // Keep searching 968 current = new_st; 969 continue; 970 } 971 // Save the new memory state for the slice and fall through 972 // to exit. 973 result = new_st; 974 } 975 } 976 break; 977 } 978 if (result != NULL) { 979 st = result; 980 } 981 } 982 983 // Loop around twice in the case Load -> Initialize -> Store. 984 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 985 for (int trip = 0; trip <= 1; trip++) { 986 987 if (st->is_Store()) { 988 Node* st_adr = st->in(MemNode::Address); 989 if (!phase->eqv(st_adr, ld_adr)) { 990 // Try harder before giving up... Match raw and non-raw pointers. 991 intptr_t st_off = 0; 992 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 993 if (alloc == NULL) return NULL; 994 if (alloc != ld_alloc) return NULL; 995 if (ld_off != st_off) return NULL; 996 // At this point we have proven something like this setup: 997 // A = Allocate(...) 998 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 999 // S = StoreQ(, AddP(, A.Parm , #Off), V) 1000 // (Actually, we haven't yet proven the Q's are the same.) 1001 // In other words, we are loading from a casted version of 1002 // the same pointer-and-offset that we stored to. 1003 // Thus, we are able to replace L by V. 1004 } 1005 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1006 if (store_Opcode() != st->Opcode()) 1007 return NULL; 1008 return st->in(MemNode::ValueIn); 1009 } 1010 1011 // A load from a freshly-created object always returns zero. 1012 // (This can happen after LoadNode::Ideal resets the load's memory input 1013 // to find_captured_store, which returned InitializeNode::zero_memory.) 1014 if (st->is_Proj() && st->in(0)->is_Allocate() && 1015 (st->in(0) == ld_alloc) && 1016 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1017 // return a zero value for the load's basic type 1018 // (This is one of the few places where a generic PhaseTransform 1019 // can create new nodes. Think of it as lazily manifesting 1020 // virtually pre-existing constants.) 1021 return phase->zerocon(memory_type()); 1022 } 1023 1024 // A load from an initialization barrier can match a captured store. 1025 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1026 InitializeNode* init = st->in(0)->as_Initialize(); 1027 AllocateNode* alloc = init->allocation(); 1028 if ((alloc != NULL) && (alloc == ld_alloc)) { 1029 // examine a captured store value 1030 st = init->find_captured_store(ld_off, memory_size(), phase); 1031 if (st != NULL) { 1032 continue; // take one more trip around 1033 } 1034 } 1035 } 1036 1037 // Load boxed value from result of valueOf() call is input parameter. 1038 if (this->is_Load() && ld_adr->is_AddP() && 1039 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1040 intptr_t ignore = 0; 1041 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1042 if (base != NULL && base->is_Proj() && 1043 base->as_Proj()->_con == TypeFunc::Parms && 1044 base->in(0)->is_CallStaticJava() && 1045 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1046 return base->in(0)->in(TypeFunc::Parms); 1047 } 1048 } 1049 1050 break; 1051 } 1052 1053 return NULL; 1054 } 1055 1056 //----------------------is_instance_field_load_with_local_phi------------------ 1057 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1058 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1059 in(Address)->is_AddP() ) { 1060 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1061 // Only instances and boxed values. 1062 if( t_oop != NULL && 1063 (t_oop->is_ptr_to_boxed_value() || 1064 t_oop->is_known_instance_field()) && 1065 t_oop->offset() != Type::OffsetBot && 1066 t_oop->offset() != Type::OffsetTop) { 1067 return true; 1068 } 1069 } 1070 return false; 1071 } 1072 1073 //------------------------------Identity--------------------------------------- 1074 // Loads are identity if previous store is to same address 1075 Node* LoadNode::Identity(PhaseGVN* phase) { 1076 // If the previous store-maker is the right kind of Store, and the store is 1077 // to the same address, then we are equal to the value stored. 1078 Node* mem = in(Memory); 1079 Node* value = can_see_stored_value(mem, phase); 1080 if( value ) { 1081 // byte, short & char stores truncate naturally. 1082 // A load has to load the truncated value which requires 1083 // some sort of masking operation and that requires an 1084 // Ideal call instead of an Identity call. 1085 if (memory_size() < BytesPerInt) { 1086 // If the input to the store does not fit with the load's result type, 1087 // it must be truncated via an Ideal call. 1088 if (!phase->type(value)->higher_equal(phase->type(this))) 1089 return this; 1090 } 1091 // (This works even when value is a Con, but LoadNode::Value 1092 // usually runs first, producing the singleton type of the Con.) 1093 return value; 1094 } 1095 1096 // Search for an existing data phi which was generated before for the same 1097 // instance's field to avoid infinite generation of phis in a loop. 1098 Node *region = mem->in(0); 1099 if (is_instance_field_load_with_local_phi(region)) { 1100 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1101 int this_index = phase->C->get_alias_index(addr_t); 1102 int this_offset = addr_t->offset(); 1103 int this_iid = addr_t->instance_id(); 1104 if (!addr_t->is_known_instance() && 1105 addr_t->is_ptr_to_boxed_value()) { 1106 // Use _idx of address base (could be Phi node) for boxed values. 1107 intptr_t ignore = 0; 1108 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1109 this_iid = base->_idx; 1110 } 1111 const Type* this_type = bottom_type(); 1112 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1113 Node* phi = region->fast_out(i); 1114 if (phi->is_Phi() && phi != mem && 1115 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1116 return phi; 1117 } 1118 } 1119 } 1120 1121 return this; 1122 } 1123 1124 // Construct an equivalent unsigned load. 1125 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1126 BasicType bt = T_ILLEGAL; 1127 const Type* rt = NULL; 1128 switch (Opcode()) { 1129 case Op_LoadUB: return this; 1130 case Op_LoadUS: return this; 1131 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1132 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1133 default: 1134 assert(false, "no unsigned variant: %s", Name()); 1135 return NULL; 1136 } 1137 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1138 raw_adr_type(), rt, bt, _mo, _control_dependency, 1139 is_unaligned_access(), is_mismatched_access()); 1140 } 1141 1142 // Construct an equivalent signed load. 1143 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1144 BasicType bt = T_ILLEGAL; 1145 const Type* rt = NULL; 1146 switch (Opcode()) { 1147 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1148 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1149 case Op_LoadB: // fall through 1150 case Op_LoadS: // fall through 1151 case Op_LoadI: // fall through 1152 case Op_LoadL: return this; 1153 default: 1154 assert(false, "no signed variant: %s", Name()); 1155 return NULL; 1156 } 1157 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1158 raw_adr_type(), rt, bt, _mo, _control_dependency, 1159 is_unaligned_access(), is_mismatched_access()); 1160 } 1161 1162 // We're loading from an object which has autobox behaviour. 1163 // If this object is result of a valueOf call we'll have a phi 1164 // merging a newly allocated object and a load from the cache. 1165 // We want to replace this load with the original incoming 1166 // argument to the valueOf call. 1167 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1168 assert(phase->C->eliminate_boxing(), "sanity"); 1169 intptr_t ignore = 0; 1170 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1171 if ((base == NULL) || base->is_Phi()) { 1172 // Push the loads from the phi that comes from valueOf up 1173 // through it to allow elimination of the loads and the recovery 1174 // of the original value. It is done in split_through_phi(). 1175 return NULL; 1176 } else if (base->is_Load() || 1177 base->is_DecodeN() && base->in(1)->is_Load()) { 1178 // Eliminate the load of boxed value for integer types from the cache 1179 // array by deriving the value from the index into the array. 1180 // Capture the offset of the load and then reverse the computation. 1181 1182 // Get LoadN node which loads a boxing object from 'cache' array. 1183 if (base->is_DecodeN()) { 1184 base = base->in(1); 1185 } 1186 if (!base->in(Address)->is_AddP()) { 1187 return NULL; // Complex address 1188 } 1189 AddPNode* address = base->in(Address)->as_AddP(); 1190 Node* cache_base = address->in(AddPNode::Base); 1191 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1192 // Get ConP node which is static 'cache' field. 1193 cache_base = cache_base->in(1); 1194 } 1195 if ((cache_base != NULL) && cache_base->is_Con()) { 1196 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1197 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1198 Node* elements[4]; 1199 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1200 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1201 if ((count > 0) && elements[0]->is_Con() && 1202 ((count == 1) || 1203 (count == 2) && elements[1]->Opcode() == Op_LShiftX && 1204 elements[1]->in(2) == phase->intcon(shift))) { 1205 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1206 // Fetch the box object cache[0] at the base of the array and get its value 1207 ciInstance* box = array->obj_at(0)->as_instance(); 1208 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1209 assert(ik->is_box_klass(), "sanity"); 1210 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1211 if (ik->nof_nonstatic_fields() == 1) { 1212 // This should be true nonstatic_field_at requires calling 1213 // nof_nonstatic_fields so check it anyway 1214 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1215 BasicType bt = c.basic_type(); 1216 // Only integer types have boxing cache. 1217 assert(bt == T_BOOLEAN || bt == T_CHAR || 1218 bt == T_BYTE || bt == T_SHORT || 1219 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1220 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1221 if (cache_low != (int)cache_low) { 1222 return NULL; // should not happen since cache is array indexed by value 1223 } 1224 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1225 if (offset != (int)offset) { 1226 return NULL; // should not happen since cache is array indexed by value 1227 } 1228 // Add up all the offsets making of the address of the load 1229 Node* result = elements[0]; 1230 for (int i = 1; i < count; i++) { 1231 result = phase->transform(new AddXNode(result, elements[i])); 1232 } 1233 // Remove the constant offset from the address and then 1234 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1235 // remove the scaling of the offset to recover the original index. 1236 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1237 // Peel the shift off directly but wrap it in a dummy node 1238 // since Ideal can't return existing nodes 1239 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1240 } else if (result->is_Add() && result->in(2)->is_Con() && 1241 result->in(1)->Opcode() == Op_LShiftX && 1242 result->in(1)->in(2) == phase->intcon(shift)) { 1243 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1244 // but for boxing cache access we know that X<<Z will not overflow 1245 // (there is range check) so we do this optimizatrion by hand here. 1246 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1247 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1248 } else { 1249 result = new RShiftXNode(result, phase->intcon(shift)); 1250 } 1251 #ifdef _LP64 1252 if (bt != T_LONG) { 1253 result = new ConvL2INode(phase->transform(result)); 1254 } 1255 #else 1256 if (bt == T_LONG) { 1257 result = new ConvI2LNode(phase->transform(result)); 1258 } 1259 #endif 1260 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1261 // Need to preserve unboxing load type if it is unsigned. 1262 switch(this->Opcode()) { 1263 case Op_LoadUB: 1264 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1265 break; 1266 case Op_LoadUS: 1267 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1268 break; 1269 } 1270 return result; 1271 } 1272 } 1273 } 1274 } 1275 } 1276 return NULL; 1277 } 1278 1279 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1280 Node* region = phi->in(0); 1281 if (region == NULL) { 1282 return false; // Wait stable graph 1283 } 1284 uint cnt = phi->req(); 1285 for (uint i = 1; i < cnt; i++) { 1286 Node* rc = region->in(i); 1287 if (rc == NULL || phase->type(rc) == Type::TOP) 1288 return false; // Wait stable graph 1289 Node* in = phi->in(i); 1290 if (in == NULL || phase->type(in) == Type::TOP) 1291 return false; // Wait stable graph 1292 } 1293 return true; 1294 } 1295 //------------------------------split_through_phi------------------------------ 1296 // Split instance or boxed field load through Phi. 1297 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1298 Node* mem = in(Memory); 1299 Node* address = in(Address); 1300 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1301 1302 assert((t_oop != NULL) && 1303 (t_oop->is_known_instance_field() || 1304 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1305 1306 Compile* C = phase->C; 1307 intptr_t ignore = 0; 1308 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1309 bool base_is_phi = (base != NULL) && base->is_Phi(); 1310 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1311 (base != NULL) && (base == address->in(AddPNode::Base)) && 1312 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1313 1314 if (!((mem->is_Phi() || base_is_phi) && 1315 (load_boxed_values || t_oop->is_known_instance_field()))) { 1316 return NULL; // memory is not Phi 1317 } 1318 1319 if (mem->is_Phi()) { 1320 if (!stable_phi(mem->as_Phi(), phase)) { 1321 return NULL; // Wait stable graph 1322 } 1323 uint cnt = mem->req(); 1324 // Check for loop invariant memory. 1325 if (cnt == 3) { 1326 for (uint i = 1; i < cnt; i++) { 1327 Node* in = mem->in(i); 1328 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1329 if (m == mem) { 1330 set_req(Memory, mem->in(cnt - i)); 1331 return this; // made change 1332 } 1333 } 1334 } 1335 } 1336 if (base_is_phi) { 1337 if (!stable_phi(base->as_Phi(), phase)) { 1338 return NULL; // Wait stable graph 1339 } 1340 uint cnt = base->req(); 1341 // Check for loop invariant memory. 1342 if (cnt == 3) { 1343 for (uint i = 1; i < cnt; i++) { 1344 if (base->in(i) == base) { 1345 return NULL; // Wait stable graph 1346 } 1347 } 1348 } 1349 } 1350 1351 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1352 1353 // Split through Phi (see original code in loopopts.cpp). 1354 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1355 1356 // Do nothing here if Identity will find a value 1357 // (to avoid infinite chain of value phis generation). 1358 if (!phase->eqv(this, this->Identity(phase))) 1359 return NULL; 1360 1361 // Select Region to split through. 1362 Node* region; 1363 if (!base_is_phi) { 1364 assert(mem->is_Phi(), "sanity"); 1365 region = mem->in(0); 1366 // Skip if the region dominates some control edge of the address. 1367 if (!MemNode::all_controls_dominate(address, region)) 1368 return NULL; 1369 } else if (!mem->is_Phi()) { 1370 assert(base_is_phi, "sanity"); 1371 region = base->in(0); 1372 // Skip if the region dominates some control edge of the memory. 1373 if (!MemNode::all_controls_dominate(mem, region)) 1374 return NULL; 1375 } else if (base->in(0) != mem->in(0)) { 1376 assert(base_is_phi && mem->is_Phi(), "sanity"); 1377 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1378 region = base->in(0); 1379 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1380 region = mem->in(0); 1381 } else { 1382 return NULL; // complex graph 1383 } 1384 } else { 1385 assert(base->in(0) == mem->in(0), "sanity"); 1386 region = mem->in(0); 1387 } 1388 1389 const Type* this_type = this->bottom_type(); 1390 int this_index = C->get_alias_index(t_oop); 1391 int this_offset = t_oop->offset(); 1392 int this_iid = t_oop->instance_id(); 1393 if (!t_oop->is_known_instance() && load_boxed_values) { 1394 // Use _idx of address base for boxed values. 1395 this_iid = base->_idx; 1396 } 1397 PhaseIterGVN* igvn = phase->is_IterGVN(); 1398 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset); 1399 for (uint i = 1; i < region->req(); i++) { 1400 Node* x; 1401 Node* the_clone = NULL; 1402 if (region->in(i) == C->top()) { 1403 x = C->top(); // Dead path? Use a dead data op 1404 } else { 1405 x = this->clone(); // Else clone up the data op 1406 the_clone = x; // Remember for possible deletion. 1407 // Alter data node to use pre-phi inputs 1408 if (this->in(0) == region) { 1409 x->set_req(0, region->in(i)); 1410 } else { 1411 x->set_req(0, NULL); 1412 } 1413 if (mem->is_Phi() && (mem->in(0) == region)) { 1414 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1415 } 1416 if (address->is_Phi() && address->in(0) == region) { 1417 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1418 } 1419 if (base_is_phi && (base->in(0) == region)) { 1420 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1421 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1422 x->set_req(Address, adr_x); 1423 } 1424 } 1425 // Check for a 'win' on some paths 1426 const Type *t = x->Value(igvn); 1427 1428 bool singleton = t->singleton(); 1429 1430 // See comments in PhaseIdealLoop::split_thru_phi(). 1431 if (singleton && t == Type::TOP) { 1432 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1433 } 1434 1435 if (singleton) { 1436 x = igvn->makecon(t); 1437 } else { 1438 // We now call Identity to try to simplify the cloned node. 1439 // Note that some Identity methods call phase->type(this). 1440 // Make sure that the type array is big enough for 1441 // our new node, even though we may throw the node away. 1442 // (This tweaking with igvn only works because x is a new node.) 1443 igvn->set_type(x, t); 1444 // If x is a TypeNode, capture any more-precise type permanently into Node 1445 // otherwise it will be not updated during igvn->transform since 1446 // igvn->type(x) is set to x->Value() already. 1447 x->raise_bottom_type(t); 1448 Node *y = x->Identity(igvn); 1449 if (y != x) { 1450 x = y; 1451 } else { 1452 y = igvn->hash_find_insert(x); 1453 if (y) { 1454 x = y; 1455 } else { 1456 // Else x is a new node we are keeping 1457 // We do not need register_new_node_with_optimizer 1458 // because set_type has already been called. 1459 igvn->_worklist.push(x); 1460 } 1461 } 1462 } 1463 if (x != the_clone && the_clone != NULL) { 1464 igvn->remove_dead_node(the_clone); 1465 } 1466 phi->set_req(i, x); 1467 } 1468 // Record Phi 1469 igvn->register_new_node_with_optimizer(phi); 1470 return phi; 1471 } 1472 1473 //------------------------------Ideal------------------------------------------ 1474 // If the load is from Field memory and the pointer is non-null, it might be possible to 1475 // zero out the control input. 1476 // If the offset is constant and the base is an object allocation, 1477 // try to hook me up to the exact initializing store. 1478 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1479 Node* p = MemNode::Ideal_common(phase, can_reshape); 1480 if (p) return (p == NodeSentinel) ? NULL : p; 1481 1482 Node* ctrl = in(MemNode::Control); 1483 Node* address = in(MemNode::Address); 1484 bool progress = false; 1485 1486 // Skip up past a SafePoint control. Cannot do this for Stores because 1487 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1488 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1489 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1490 ctrl = ctrl->in(0); 1491 set_req(MemNode::Control,ctrl); 1492 progress = true; 1493 } 1494 1495 intptr_t ignore = 0; 1496 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1497 if (base != NULL 1498 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1499 // Check for useless control edge in some common special cases 1500 if (in(MemNode::Control) != NULL 1501 && can_remove_control() 1502 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1503 && all_controls_dominate(base, phase->C->start())) { 1504 // A method-invariant, non-null address (constant or 'this' argument). 1505 set_req(MemNode::Control, NULL); 1506 progress = true; 1507 } 1508 } 1509 1510 Node* mem = in(MemNode::Memory); 1511 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1512 1513 if (can_reshape && (addr_t != NULL)) { 1514 // try to optimize our memory input 1515 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1516 if (opt_mem != mem) { 1517 set_req(MemNode::Memory, opt_mem); 1518 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1519 return this; 1520 } 1521 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1522 if ((t_oop != NULL) && 1523 (t_oop->is_known_instance_field() || 1524 t_oop->is_ptr_to_boxed_value())) { 1525 PhaseIterGVN *igvn = phase->is_IterGVN(); 1526 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1527 // Delay this transformation until memory Phi is processed. 1528 phase->is_IterGVN()->_worklist.push(this); 1529 return NULL; 1530 } 1531 // Split instance field load through Phi. 1532 Node* result = split_through_phi(phase); 1533 if (result != NULL) return result; 1534 1535 if (t_oop->is_ptr_to_boxed_value()) { 1536 Node* result = eliminate_autobox(phase); 1537 if (result != NULL) return result; 1538 } 1539 } 1540 } 1541 1542 // Is there a dominating load that loads the same value? Leave 1543 // anything that is not a load of a field/array element (like 1544 // barriers etc.) alone 1545 if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) { 1546 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1547 Node *use = mem->fast_out(i); 1548 if (use != this && 1549 use->Opcode() == Opcode() && 1550 use->in(0) != NULL && 1551 use->in(0) != in(0) && 1552 use->in(Address) == in(Address)) { 1553 Node* ctl = in(0); 1554 for (int i = 0; i < 10 && ctl != NULL; i++) { 1555 ctl = IfNode::up_one_dom(ctl); 1556 if (ctl == use->in(0)) { 1557 set_req(0, use->in(0)); 1558 return this; 1559 } 1560 } 1561 } 1562 } 1563 } 1564 1565 // Check for prior store with a different base or offset; make Load 1566 // independent. Skip through any number of them. Bail out if the stores 1567 // are in an endless dead cycle and report no progress. This is a key 1568 // transform for Reflection. However, if after skipping through the Stores 1569 // we can't then fold up against a prior store do NOT do the transform as 1570 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1571 // array memory alive twice: once for the hoisted Load and again after the 1572 // bypassed Store. This situation only works if EVERYBODY who does 1573 // anti-dependence work knows how to bypass. I.e. we need all 1574 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1575 // the alias index stuff. So instead, peek through Stores and IFF we can 1576 // fold up, do so. 1577 Node* prev_mem = find_previous_store(phase); 1578 if (prev_mem != NULL) { 1579 Node* value = can_see_arraycopy_value(prev_mem, phase); 1580 if (value != NULL) { 1581 return value; 1582 } 1583 } 1584 // Steps (a), (b): Walk past independent stores to find an exact match. 1585 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1586 // (c) See if we can fold up on the spot, but don't fold up here. 1587 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1588 // just return a prior value, which is done by Identity calls. 1589 if (can_see_stored_value(prev_mem, phase)) { 1590 // Make ready for step (d): 1591 set_req(MemNode::Memory, prev_mem); 1592 return this; 1593 } 1594 } 1595 1596 return progress ? this : NULL; 1597 } 1598 1599 // Helper to recognize certain Klass fields which are invariant across 1600 // some group of array types (e.g., int[] or all T[] where T < Object). 1601 const Type* 1602 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1603 ciKlass* klass) const { 1604 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1605 // The field is Klass::_modifier_flags. Return its (constant) value. 1606 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1607 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1608 return TypeInt::make(klass->modifier_flags()); 1609 } 1610 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1611 // The field is Klass::_access_flags. Return its (constant) value. 1612 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1613 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1614 return TypeInt::make(klass->access_flags()); 1615 } 1616 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1617 // The field is Klass::_layout_helper. Return its constant value if known. 1618 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1619 return TypeInt::make(klass->layout_helper()); 1620 } 1621 1622 // No match. 1623 return NULL; 1624 } 1625 1626 //------------------------------Value----------------------------------------- 1627 const Type* LoadNode::Value(PhaseGVN* phase) const { 1628 // Either input is TOP ==> the result is TOP 1629 Node* mem = in(MemNode::Memory); 1630 const Type *t1 = phase->type(mem); 1631 if (t1 == Type::TOP) return Type::TOP; 1632 Node* adr = in(MemNode::Address); 1633 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1634 if (tp == NULL || tp->empty()) return Type::TOP; 1635 int off = tp->offset(); 1636 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1637 Compile* C = phase->C; 1638 1639 // Try to guess loaded type from pointer type 1640 if (tp->isa_aryptr()) { 1641 const TypeAryPtr* ary = tp->is_aryptr(); 1642 const Type* t = ary->elem(); 1643 1644 // Determine whether the reference is beyond the header or not, by comparing 1645 // the offset against the offset of the start of the array's data. 1646 // Different array types begin at slightly different offsets (12 vs. 16). 1647 // We choose T_BYTE as an example base type that is least restrictive 1648 // as to alignment, which will therefore produce the smallest 1649 // possible base offset. 1650 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1651 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1652 1653 // Try to constant-fold a stable array element. 1654 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1655 // Make sure the reference is not into the header and the offset is constant 1656 ciObject* aobj = ary->const_oop(); 1657 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1658 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1659 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1660 stable_dimension, 1661 memory_type(), is_unsigned()); 1662 if (con_type != NULL) { 1663 return con_type; 1664 } 1665 } 1666 } 1667 1668 // Don't do this for integer types. There is only potential profit if 1669 // the element type t is lower than _type; that is, for int types, if _type is 1670 // more restrictive than t. This only happens here if one is short and the other 1671 // char (both 16 bits), and in those cases we've made an intentional decision 1672 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1673 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1674 // 1675 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1676 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1677 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1678 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1679 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1680 // In fact, that could have been the original type of p1, and p1 could have 1681 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1682 // expression (LShiftL quux 3) independently optimized to the constant 8. 1683 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1684 && (_type->isa_vect() == NULL) 1685 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1686 // t might actually be lower than _type, if _type is a unique 1687 // concrete subclass of abstract class t. 1688 if (off_beyond_header) { // is the offset beyond the header? 1689 const Type* jt = t->join_speculative(_type); 1690 // In any case, do not allow the join, per se, to empty out the type. 1691 if (jt->empty() && !t->empty()) { 1692 // This can happen if a interface-typed array narrows to a class type. 1693 jt = _type; 1694 } 1695 #ifdef ASSERT 1696 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1697 // The pointers in the autobox arrays are always non-null 1698 Node* base = adr->in(AddPNode::Base); 1699 if ((base != NULL) && base->is_DecodeN()) { 1700 // Get LoadN node which loads IntegerCache.cache field 1701 base = base->in(1); 1702 } 1703 if ((base != NULL) && base->is_Con()) { 1704 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1705 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1706 // It could be narrow oop 1707 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1708 } 1709 } 1710 } 1711 #endif 1712 return jt; 1713 } 1714 } 1715 } else if (tp->base() == Type::InstPtr) { 1716 assert( off != Type::OffsetBot || 1717 // arrays can be cast to Objects 1718 tp->is_oopptr()->klass()->is_java_lang_Object() || 1719 // unsafe field access may not have a constant offset 1720 C->has_unsafe_access(), 1721 "Field accesses must be precise" ); 1722 // For oop loads, we expect the _type to be precise. 1723 1724 // Optimize loads from constant fields. 1725 const TypeInstPtr* tinst = tp->is_instptr(); 1726 ciObject* const_oop = tinst->const_oop(); 1727 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) { 1728 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type()); 1729 if (con_type != NULL) { 1730 return con_type; 1731 } 1732 } 1733 } else if (tp->base() == Type::KlassPtr) { 1734 assert( off != Type::OffsetBot || 1735 // arrays can be cast to Objects 1736 tp->is_klassptr()->klass()->is_java_lang_Object() || 1737 // also allow array-loading from the primary supertype 1738 // array during subtype checks 1739 Opcode() == Op_LoadKlass, 1740 "Field accesses must be precise" ); 1741 // For klass/static loads, we expect the _type to be precise 1742 } 1743 1744 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1745 if (tkls != NULL && !StressReflectiveCode) { 1746 ciKlass* klass = tkls->klass(); 1747 if (klass->is_loaded() && tkls->klass_is_exact()) { 1748 // We are loading a field from a Klass metaobject whose identity 1749 // is known at compile time (the type is "exact" or "precise"). 1750 // Check for fields we know are maintained as constants by the VM. 1751 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1752 // The field is Klass::_super_check_offset. Return its (constant) value. 1753 // (Folds up type checking code.) 1754 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1755 return TypeInt::make(klass->super_check_offset()); 1756 } 1757 // Compute index into primary_supers array 1758 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1759 // Check for overflowing; use unsigned compare to handle the negative case. 1760 if( depth < ciKlass::primary_super_limit() ) { 1761 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1762 // (Folds up type checking code.) 1763 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1764 ciKlass *ss = klass->super_of_depth(depth); 1765 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1766 } 1767 const Type* aift = load_array_final_field(tkls, klass); 1768 if (aift != NULL) return aift; 1769 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1770 // The field is Klass::_java_mirror. Return its (constant) value. 1771 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1772 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1773 return TypeInstPtr::make(klass->java_mirror()); 1774 } 1775 } 1776 1777 // We can still check if we are loading from the primary_supers array at a 1778 // shallow enough depth. Even though the klass is not exact, entries less 1779 // than or equal to its super depth are correct. 1780 if (klass->is_loaded() ) { 1781 ciType *inner = klass; 1782 while( inner->is_obj_array_klass() ) 1783 inner = inner->as_obj_array_klass()->base_element_type(); 1784 if( inner->is_instance_klass() && 1785 !inner->as_instance_klass()->flags().is_interface() ) { 1786 // Compute index into primary_supers array 1787 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1788 // Check for overflowing; use unsigned compare to handle the negative case. 1789 if( depth < ciKlass::primary_super_limit() && 1790 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1791 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1792 // (Folds up type checking code.) 1793 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1794 ciKlass *ss = klass->super_of_depth(depth); 1795 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1796 } 1797 } 1798 } 1799 1800 // If the type is enough to determine that the thing is not an array, 1801 // we can give the layout_helper a positive interval type. 1802 // This will help short-circuit some reflective code. 1803 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1804 && !klass->is_array_klass() // not directly typed as an array 1805 && !klass->is_interface() // specifically not Serializable & Cloneable 1806 && !klass->is_java_lang_Object() // not the supertype of all T[] 1807 ) { 1808 // Note: When interfaces are reliable, we can narrow the interface 1809 // test to (klass != Serializable && klass != Cloneable). 1810 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1811 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1812 // The key property of this type is that it folds up tests 1813 // for array-ness, since it proves that the layout_helper is positive. 1814 // Thus, a generic value like the basic object layout helper works fine. 1815 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1816 } 1817 } 1818 1819 // If we are loading from a freshly-allocated object, produce a zero, 1820 // if the load is provably beyond the header of the object. 1821 // (Also allow a variable load from a fresh array to produce zero.) 1822 const TypeOopPtr *tinst = tp->isa_oopptr(); 1823 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1824 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1825 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1826 Node* value = can_see_stored_value(mem,phase); 1827 if (value != NULL && value->is_Con()) { 1828 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1829 return value->bottom_type(); 1830 } 1831 } 1832 1833 if (is_instance) { 1834 // If we have an instance type and our memory input is the 1835 // programs's initial memory state, there is no matching store, 1836 // so just return a zero of the appropriate type 1837 Node *mem = in(MemNode::Memory); 1838 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1839 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1840 return Type::get_zero_type(_type->basic_type()); 1841 } 1842 } 1843 return _type; 1844 } 1845 1846 //------------------------------match_edge------------------------------------- 1847 // Do we Match on this edge index or not? Match only the address. 1848 uint LoadNode::match_edge(uint idx) const { 1849 return idx == MemNode::Address; 1850 } 1851 1852 //--------------------------LoadBNode::Ideal-------------------------------------- 1853 // 1854 // If the previous store is to the same address as this load, 1855 // and the value stored was larger than a byte, replace this load 1856 // with the value stored truncated to a byte. If no truncation is 1857 // needed, the replacement is done in LoadNode::Identity(). 1858 // 1859 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1860 Node* mem = in(MemNode::Memory); 1861 Node* value = can_see_stored_value(mem,phase); 1862 if( value && !phase->type(value)->higher_equal( _type ) ) { 1863 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1864 return new RShiftINode(result, phase->intcon(24)); 1865 } 1866 // Identity call will handle the case where truncation is not needed. 1867 return LoadNode::Ideal(phase, can_reshape); 1868 } 1869 1870 const Type* LoadBNode::Value(PhaseGVN* phase) const { 1871 Node* mem = in(MemNode::Memory); 1872 Node* value = can_see_stored_value(mem,phase); 1873 if (value != NULL && value->is_Con() && 1874 !value->bottom_type()->higher_equal(_type)) { 1875 // If the input to the store does not fit with the load's result type, 1876 // it must be truncated. We can't delay until Ideal call since 1877 // a singleton Value is needed for split_thru_phi optimization. 1878 int con = value->get_int(); 1879 return TypeInt::make((con << 24) >> 24); 1880 } 1881 return LoadNode::Value(phase); 1882 } 1883 1884 //--------------------------LoadUBNode::Ideal------------------------------------- 1885 // 1886 // If the previous store is to the same address as this load, 1887 // and the value stored was larger than a byte, replace this load 1888 // with the value stored truncated to a byte. If no truncation is 1889 // needed, the replacement is done in LoadNode::Identity(). 1890 // 1891 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1892 Node* mem = in(MemNode::Memory); 1893 Node* value = can_see_stored_value(mem, phase); 1894 if (value && !phase->type(value)->higher_equal(_type)) 1895 return new AndINode(value, phase->intcon(0xFF)); 1896 // Identity call will handle the case where truncation is not needed. 1897 return LoadNode::Ideal(phase, can_reshape); 1898 } 1899 1900 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 1901 Node* mem = in(MemNode::Memory); 1902 Node* value = can_see_stored_value(mem,phase); 1903 if (value != NULL && value->is_Con() && 1904 !value->bottom_type()->higher_equal(_type)) { 1905 // If the input to the store does not fit with the load's result type, 1906 // it must be truncated. We can't delay until Ideal call since 1907 // a singleton Value is needed for split_thru_phi optimization. 1908 int con = value->get_int(); 1909 return TypeInt::make(con & 0xFF); 1910 } 1911 return LoadNode::Value(phase); 1912 } 1913 1914 //--------------------------LoadUSNode::Ideal------------------------------------- 1915 // 1916 // If the previous store is to the same address as this load, 1917 // and the value stored was larger than a char, replace this load 1918 // with the value stored truncated to a char. If no truncation is 1919 // needed, the replacement is done in LoadNode::Identity(). 1920 // 1921 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1922 Node* mem = in(MemNode::Memory); 1923 Node* value = can_see_stored_value(mem,phase); 1924 if( value && !phase->type(value)->higher_equal( _type ) ) 1925 return new AndINode(value,phase->intcon(0xFFFF)); 1926 // Identity call will handle the case where truncation is not needed. 1927 return LoadNode::Ideal(phase, can_reshape); 1928 } 1929 1930 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 1931 Node* mem = in(MemNode::Memory); 1932 Node* value = can_see_stored_value(mem,phase); 1933 if (value != NULL && value->is_Con() && 1934 !value->bottom_type()->higher_equal(_type)) { 1935 // If the input to the store does not fit with the load's result type, 1936 // it must be truncated. We can't delay until Ideal call since 1937 // a singleton Value is needed for split_thru_phi optimization. 1938 int con = value->get_int(); 1939 return TypeInt::make(con & 0xFFFF); 1940 } 1941 return LoadNode::Value(phase); 1942 } 1943 1944 //--------------------------LoadSNode::Ideal-------------------------------------- 1945 // 1946 // If the previous store is to the same address as this load, 1947 // and the value stored was larger than a short, replace this load 1948 // with the value stored truncated to a short. If no truncation is 1949 // needed, the replacement is done in LoadNode::Identity(). 1950 // 1951 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1952 Node* mem = in(MemNode::Memory); 1953 Node* value = can_see_stored_value(mem,phase); 1954 if( value && !phase->type(value)->higher_equal( _type ) ) { 1955 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 1956 return new RShiftINode(result, phase->intcon(16)); 1957 } 1958 // Identity call will handle the case where truncation is not needed. 1959 return LoadNode::Ideal(phase, can_reshape); 1960 } 1961 1962 const Type* LoadSNode::Value(PhaseGVN* phase) const { 1963 Node* mem = in(MemNode::Memory); 1964 Node* value = can_see_stored_value(mem,phase); 1965 if (value != NULL && value->is_Con() && 1966 !value->bottom_type()->higher_equal(_type)) { 1967 // If the input to the store does not fit with the load's result type, 1968 // it must be truncated. We can't delay until Ideal call since 1969 // a singleton Value is needed for split_thru_phi optimization. 1970 int con = value->get_int(); 1971 return TypeInt::make((con << 16) >> 16); 1972 } 1973 return LoadNode::Value(phase); 1974 } 1975 1976 //============================================================================= 1977 //----------------------------LoadKlassNode::make------------------------------ 1978 // Polymorphic factory method: 1979 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 1980 // sanity check the alias category against the created node type 1981 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 1982 assert(adr_type != NULL, "expecting TypeKlassPtr"); 1983 #ifdef _LP64 1984 if (adr_type->is_ptr_to_narrowklass()) { 1985 assert(UseCompressedClassPointers, "no compressed klasses"); 1986 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 1987 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 1988 } 1989 #endif 1990 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1991 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 1992 } 1993 1994 //------------------------------Value------------------------------------------ 1995 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 1996 return klass_value_common(phase); 1997 } 1998 1999 // In most cases, LoadKlassNode does not have the control input set. If the control 2000 // input is set, it must not be removed (by LoadNode::Ideal()). 2001 bool LoadKlassNode::can_remove_control() const { 2002 return false; 2003 } 2004 2005 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2006 // Either input is TOP ==> the result is TOP 2007 const Type *t1 = phase->type( in(MemNode::Memory) ); 2008 if (t1 == Type::TOP) return Type::TOP; 2009 Node *adr = in(MemNode::Address); 2010 const Type *t2 = phase->type( adr ); 2011 if (t2 == Type::TOP) return Type::TOP; 2012 const TypePtr *tp = t2->is_ptr(); 2013 if (TypePtr::above_centerline(tp->ptr()) || 2014 tp->ptr() == TypePtr::Null) return Type::TOP; 2015 2016 // Return a more precise klass, if possible 2017 const TypeInstPtr *tinst = tp->isa_instptr(); 2018 if (tinst != NULL) { 2019 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2020 int offset = tinst->offset(); 2021 if (ik == phase->C->env()->Class_klass() 2022 && (offset == java_lang_Class::klass_offset_in_bytes() || 2023 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2024 // We are loading a special hidden field from a Class mirror object, 2025 // the field which points to the VM's Klass metaobject. 2026 ciType* t = tinst->java_mirror_type(); 2027 // java_mirror_type returns non-null for compile-time Class constants. 2028 if (t != NULL) { 2029 // constant oop => constant klass 2030 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2031 if (t->is_void()) { 2032 // We cannot create a void array. Since void is a primitive type return null 2033 // klass. Users of this result need to do a null check on the returned klass. 2034 return TypePtr::NULL_PTR; 2035 } 2036 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2037 } 2038 if (!t->is_klass()) { 2039 // a primitive Class (e.g., int.class) has NULL for a klass field 2040 return TypePtr::NULL_PTR; 2041 } 2042 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2043 return TypeKlassPtr::make(t->as_klass()); 2044 } 2045 // non-constant mirror, so we can't tell what's going on 2046 } 2047 if( !ik->is_loaded() ) 2048 return _type; // Bail out if not loaded 2049 if (offset == oopDesc::klass_offset_in_bytes()) { 2050 if (tinst->klass_is_exact()) { 2051 return TypeKlassPtr::make(ik); 2052 } 2053 // See if we can become precise: no subklasses and no interface 2054 // (Note: We need to support verified interfaces.) 2055 if (!ik->is_interface() && !ik->has_subklass()) { 2056 //assert(!UseExactTypes, "this code should be useless with exact types"); 2057 // Add a dependence; if any subclass added we need to recompile 2058 if (!ik->is_final()) { 2059 // %%% should use stronger assert_unique_concrete_subtype instead 2060 phase->C->dependencies()->assert_leaf_type(ik); 2061 } 2062 // Return precise klass 2063 return TypeKlassPtr::make(ik); 2064 } 2065 2066 // Return root of possible klass 2067 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2068 } 2069 } 2070 2071 // Check for loading klass from an array 2072 const TypeAryPtr *tary = tp->isa_aryptr(); 2073 if( tary != NULL ) { 2074 ciKlass *tary_klass = tary->klass(); 2075 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2076 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2077 if (tary->klass_is_exact()) { 2078 return TypeKlassPtr::make(tary_klass); 2079 } 2080 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2081 // If the klass is an object array, we defer the question to the 2082 // array component klass. 2083 if( ak->is_obj_array_klass() ) { 2084 assert( ak->is_loaded(), "" ); 2085 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2086 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2087 ciInstanceKlass* ik = base_k->as_instance_klass(); 2088 // See if we can become precise: no subklasses and no interface 2089 if (!ik->is_interface() && !ik->has_subklass()) { 2090 //assert(!UseExactTypes, "this code should be useless with exact types"); 2091 // Add a dependence; if any subclass added we need to recompile 2092 if (!ik->is_final()) { 2093 phase->C->dependencies()->assert_leaf_type(ik); 2094 } 2095 // Return precise array klass 2096 return TypeKlassPtr::make(ak); 2097 } 2098 } 2099 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2100 } else { // Found a type-array? 2101 //assert(!UseExactTypes, "this code should be useless with exact types"); 2102 assert( ak->is_type_array_klass(), "" ); 2103 return TypeKlassPtr::make(ak); // These are always precise 2104 } 2105 } 2106 } 2107 2108 // Check for loading klass from an array klass 2109 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2110 if (tkls != NULL && !StressReflectiveCode) { 2111 ciKlass* klass = tkls->klass(); 2112 if( !klass->is_loaded() ) 2113 return _type; // Bail out if not loaded 2114 if( klass->is_obj_array_klass() && 2115 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2116 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2117 // // Always returning precise element type is incorrect, 2118 // // e.g., element type could be object and array may contain strings 2119 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2120 2121 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2122 // according to the element type's subclassing. 2123 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2124 } 2125 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2126 tkls->offset() == in_bytes(Klass::super_offset())) { 2127 ciKlass* sup = klass->as_instance_klass()->super(); 2128 // The field is Klass::_super. Return its (constant) value. 2129 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2130 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2131 } 2132 } 2133 2134 // Bailout case 2135 return LoadNode::Value(phase); 2136 } 2137 2138 //------------------------------Identity--------------------------------------- 2139 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2140 // Also feed through the klass in Allocate(...klass...)._klass. 2141 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2142 return klass_identity_common(phase); 2143 } 2144 2145 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2146 Node* x = LoadNode::Identity(phase); 2147 if (x != this) return x; 2148 2149 // Take apart the address into an oop and and offset. 2150 // Return 'this' if we cannot. 2151 Node* adr = in(MemNode::Address); 2152 intptr_t offset = 0; 2153 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2154 if (base == NULL) return this; 2155 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2156 if (toop == NULL) return this; 2157 2158 // We can fetch the klass directly through an AllocateNode. 2159 // This works even if the klass is not constant (clone or newArray). 2160 if (offset == oopDesc::klass_offset_in_bytes()) { 2161 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2162 if (allocated_klass != NULL) { 2163 return allocated_klass; 2164 } 2165 } 2166 2167 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2168 // See inline_native_Class_query for occurrences of these patterns. 2169 // Java Example: x.getClass().isAssignableFrom(y) 2170 // 2171 // This improves reflective code, often making the Class 2172 // mirror go completely dead. (Current exception: Class 2173 // mirrors may appear in debug info, but we could clean them out by 2174 // introducing a new debug info operator for Klass*.java_mirror). 2175 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2176 && offset == java_lang_Class::klass_offset_in_bytes()) { 2177 // We are loading a special hidden field from a Class mirror, 2178 // the field which points to its Klass or ArrayKlass metaobject. 2179 if (base->is_Load()) { 2180 Node* adr2 = base->in(MemNode::Address); 2181 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2182 if (tkls != NULL && !tkls->empty() 2183 && (tkls->klass()->is_instance_klass() || 2184 tkls->klass()->is_array_klass()) 2185 && adr2->is_AddP() 2186 ) { 2187 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2188 if (tkls->offset() == mirror_field) { 2189 return adr2->in(AddPNode::Base); 2190 } 2191 } 2192 } 2193 } 2194 2195 return this; 2196 } 2197 2198 2199 //------------------------------Value------------------------------------------ 2200 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2201 const Type *t = klass_value_common(phase); 2202 if (t == Type::TOP) 2203 return t; 2204 2205 return t->make_narrowklass(); 2206 } 2207 2208 //------------------------------Identity--------------------------------------- 2209 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2210 // Also feed through the klass in Allocate(...klass...)._klass. 2211 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2212 Node *x = klass_identity_common(phase); 2213 2214 const Type *t = phase->type( x ); 2215 if( t == Type::TOP ) return x; 2216 if( t->isa_narrowklass()) return x; 2217 assert (!t->isa_narrowoop(), "no narrow oop here"); 2218 2219 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2220 } 2221 2222 //------------------------------Value----------------------------------------- 2223 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2224 // Either input is TOP ==> the result is TOP 2225 const Type *t1 = phase->type( in(MemNode::Memory) ); 2226 if( t1 == Type::TOP ) return Type::TOP; 2227 Node *adr = in(MemNode::Address); 2228 const Type *t2 = phase->type( adr ); 2229 if( t2 == Type::TOP ) return Type::TOP; 2230 const TypePtr *tp = t2->is_ptr(); 2231 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2232 const TypeAryPtr *tap = tp->isa_aryptr(); 2233 if( !tap ) return _type; 2234 return tap->size(); 2235 } 2236 2237 //-------------------------------Ideal--------------------------------------- 2238 // Feed through the length in AllocateArray(...length...)._length. 2239 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2240 Node* p = MemNode::Ideal_common(phase, can_reshape); 2241 if (p) return (p == NodeSentinel) ? NULL : p; 2242 2243 // Take apart the address into an oop and and offset. 2244 // Return 'this' if we cannot. 2245 Node* adr = in(MemNode::Address); 2246 intptr_t offset = 0; 2247 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2248 if (base == NULL) return NULL; 2249 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2250 if (tary == NULL) return NULL; 2251 2252 // We can fetch the length directly through an AllocateArrayNode. 2253 // This works even if the length is not constant (clone or newArray). 2254 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2255 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2256 if (alloc != NULL) { 2257 Node* allocated_length = alloc->Ideal_length(); 2258 Node* len = alloc->make_ideal_length(tary, phase); 2259 if (allocated_length != len) { 2260 // New CastII improves on this. 2261 return len; 2262 } 2263 } 2264 } 2265 2266 return NULL; 2267 } 2268 2269 //------------------------------Identity--------------------------------------- 2270 // Feed through the length in AllocateArray(...length...)._length. 2271 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2272 Node* x = LoadINode::Identity(phase); 2273 if (x != this) return x; 2274 2275 // Take apart the address into an oop and and offset. 2276 // Return 'this' if we cannot. 2277 Node* adr = in(MemNode::Address); 2278 intptr_t offset = 0; 2279 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2280 if (base == NULL) return this; 2281 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2282 if (tary == NULL) return this; 2283 2284 // We can fetch the length directly through an AllocateArrayNode. 2285 // This works even if the length is not constant (clone or newArray). 2286 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2287 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2288 if (alloc != NULL) { 2289 Node* allocated_length = alloc->Ideal_length(); 2290 // Do not allow make_ideal_length to allocate a CastII node. 2291 Node* len = alloc->make_ideal_length(tary, phase, false); 2292 if (allocated_length == len) { 2293 // Return allocated_length only if it would not be improved by a CastII. 2294 return allocated_length; 2295 } 2296 } 2297 } 2298 2299 return this; 2300 2301 } 2302 2303 //============================================================================= 2304 //---------------------------StoreNode::make----------------------------------- 2305 // Polymorphic factory method: 2306 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2307 assert((mo == unordered || mo == release), "unexpected"); 2308 Compile* C = gvn.C; 2309 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2310 ctl != NULL, "raw memory operations should have control edge"); 2311 2312 switch (bt) { 2313 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2314 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2315 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2316 case T_CHAR: 2317 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2318 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2319 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2320 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2321 case T_METADATA: 2322 case T_ADDRESS: 2323 case T_OBJECT: 2324 #ifdef _LP64 2325 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2326 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2327 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2328 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2329 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2330 adr->bottom_type()->isa_rawptr())) { 2331 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2332 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2333 } 2334 #endif 2335 { 2336 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2337 } 2338 } 2339 ShouldNotReachHere(); 2340 return (StoreNode*)NULL; 2341 } 2342 2343 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2344 bool require_atomic = true; 2345 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2346 } 2347 2348 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2349 bool require_atomic = true; 2350 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2351 } 2352 2353 2354 //--------------------------bottom_type---------------------------------------- 2355 const Type *StoreNode::bottom_type() const { 2356 return Type::MEMORY; 2357 } 2358 2359 //------------------------------hash------------------------------------------- 2360 uint StoreNode::hash() const { 2361 // unroll addition of interesting fields 2362 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2363 2364 // Since they are not commoned, do not hash them: 2365 return NO_HASH; 2366 } 2367 2368 //------------------------------Ideal------------------------------------------ 2369 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2370 // When a store immediately follows a relevant allocation/initialization, 2371 // try to capture it into the initialization, or hoist it above. 2372 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2373 Node* p = MemNode::Ideal_common(phase, can_reshape); 2374 if (p) return (p == NodeSentinel) ? NULL : p; 2375 2376 Node* mem = in(MemNode::Memory); 2377 Node* address = in(MemNode::Address); 2378 // Back-to-back stores to same address? Fold em up. Generally 2379 // unsafe if I have intervening uses... Also disallowed for StoreCM 2380 // since they must follow each StoreP operation. Redundant StoreCMs 2381 // are eliminated just before matching in final_graph_reshape. 2382 { 2383 Node* st = mem; 2384 // If Store 'st' has more than one use, we cannot fold 'st' away. 2385 // For example, 'st' might be the final state at a conditional 2386 // return. Or, 'st' might be used by some node which is live at 2387 // the same time 'st' is live, which might be unschedulable. So, 2388 // require exactly ONE user until such time as we clone 'mem' for 2389 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2390 // true). 2391 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2392 // Looking at a dead closed cycle of memory? 2393 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2394 assert(Opcode() == st->Opcode() || 2395 st->Opcode() == Op_StoreVector || 2396 Opcode() == Op_StoreVector || 2397 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2398 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2399 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2400 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2401 2402 if (st->in(MemNode::Address)->eqv_uncast(address) && 2403 st->as_Store()->memory_size() <= this->memory_size()) { 2404 Node* use = st->raw_out(0); 2405 phase->igvn_rehash_node_delayed(use); 2406 if (can_reshape) { 2407 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2408 } else { 2409 // It's OK to do this in the parser, since DU info is always accurate, 2410 // and the parser always refers to nodes via SafePointNode maps. 2411 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2412 } 2413 return this; 2414 } 2415 st = st->in(MemNode::Memory); 2416 } 2417 } 2418 2419 2420 // Capture an unaliased, unconditional, simple store into an initializer. 2421 // Or, if it is independent of the allocation, hoist it above the allocation. 2422 if (ReduceFieldZeroing && /*can_reshape &&*/ 2423 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2424 InitializeNode* init = mem->in(0)->as_Initialize(); 2425 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2426 if (offset > 0) { 2427 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2428 // If the InitializeNode captured me, it made a raw copy of me, 2429 // and I need to disappear. 2430 if (moved != NULL) { 2431 // %%% hack to ensure that Ideal returns a new node: 2432 mem = MergeMemNode::make(mem); 2433 return mem; // fold me away 2434 } 2435 } 2436 } 2437 2438 return NULL; // No further progress 2439 } 2440 2441 //------------------------------Value----------------------------------------- 2442 const Type* StoreNode::Value(PhaseGVN* phase) const { 2443 // Either input is TOP ==> the result is TOP 2444 const Type *t1 = phase->type( in(MemNode::Memory) ); 2445 if( t1 == Type::TOP ) return Type::TOP; 2446 const Type *t2 = phase->type( in(MemNode::Address) ); 2447 if( t2 == Type::TOP ) return Type::TOP; 2448 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2449 if( t3 == Type::TOP ) return Type::TOP; 2450 return Type::MEMORY; 2451 } 2452 2453 //------------------------------Identity--------------------------------------- 2454 // Remove redundant stores: 2455 // Store(m, p, Load(m, p)) changes to m. 2456 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2457 Node* StoreNode::Identity(PhaseGVN* phase) { 2458 Node* mem = in(MemNode::Memory); 2459 Node* adr = in(MemNode::Address); 2460 Node* val = in(MemNode::ValueIn); 2461 2462 // Load then Store? Then the Store is useless 2463 if (val->is_Load() && 2464 val->in(MemNode::Address)->eqv_uncast(adr) && 2465 val->in(MemNode::Memory )->eqv_uncast(mem) && 2466 val->as_Load()->store_Opcode() == Opcode()) { 2467 return mem; 2468 } 2469 2470 // Two stores in a row of the same value? 2471 if (mem->is_Store() && 2472 mem->in(MemNode::Address)->eqv_uncast(adr) && 2473 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2474 mem->Opcode() == Opcode()) { 2475 return mem; 2476 } 2477 2478 // Store of zero anywhere into a freshly-allocated object? 2479 // Then the store is useless. 2480 // (It must already have been captured by the InitializeNode.) 2481 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2482 // a newly allocated object is already all-zeroes everywhere 2483 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2484 return mem; 2485 } 2486 2487 // the store may also apply to zero-bits in an earlier object 2488 Node* prev_mem = find_previous_store(phase); 2489 // Steps (a), (b): Walk past independent stores to find an exact match. 2490 if (prev_mem != NULL) { 2491 Node* prev_val = can_see_stored_value(prev_mem, phase); 2492 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2493 // prev_val and val might differ by a cast; it would be good 2494 // to keep the more informative of the two. 2495 return mem; 2496 } 2497 } 2498 } 2499 2500 return this; 2501 } 2502 2503 //------------------------------match_edge------------------------------------- 2504 // Do we Match on this edge index or not? Match only memory & value 2505 uint StoreNode::match_edge(uint idx) const { 2506 return idx == MemNode::Address || idx == MemNode::ValueIn; 2507 } 2508 2509 //------------------------------cmp-------------------------------------------- 2510 // Do not common stores up together. They generally have to be split 2511 // back up anyways, so do not bother. 2512 uint StoreNode::cmp( const Node &n ) const { 2513 return (&n == this); // Always fail except on self 2514 } 2515 2516 //------------------------------Ideal_masked_input----------------------------- 2517 // Check for a useless mask before a partial-word store 2518 // (StoreB ... (AndI valIn conIa) ) 2519 // If (conIa & mask == mask) this simplifies to 2520 // (StoreB ... (valIn) ) 2521 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2522 Node *val = in(MemNode::ValueIn); 2523 if( val->Opcode() == Op_AndI ) { 2524 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2525 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2526 set_req(MemNode::ValueIn, val->in(1)); 2527 return this; 2528 } 2529 } 2530 return NULL; 2531 } 2532 2533 2534 //------------------------------Ideal_sign_extended_input---------------------- 2535 // Check for useless sign-extension before a partial-word store 2536 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2537 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2538 // (StoreB ... (valIn) ) 2539 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2540 Node *val = in(MemNode::ValueIn); 2541 if( val->Opcode() == Op_RShiftI ) { 2542 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2543 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2544 Node *shl = val->in(1); 2545 if( shl->Opcode() == Op_LShiftI ) { 2546 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2547 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2548 set_req(MemNode::ValueIn, shl->in(1)); 2549 return this; 2550 } 2551 } 2552 } 2553 } 2554 return NULL; 2555 } 2556 2557 //------------------------------value_never_loaded----------------------------------- 2558 // Determine whether there are any possible loads of the value stored. 2559 // For simplicity, we actually check if there are any loads from the 2560 // address stored to, not just for loads of the value stored by this node. 2561 // 2562 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2563 Node *adr = in(Address); 2564 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2565 if (adr_oop == NULL) 2566 return false; 2567 if (!adr_oop->is_known_instance_field()) 2568 return false; // if not a distinct instance, there may be aliases of the address 2569 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2570 Node *use = adr->fast_out(i); 2571 if (use->is_Load() || use->is_LoadStore()) { 2572 return false; 2573 } 2574 } 2575 return true; 2576 } 2577 2578 //============================================================================= 2579 //------------------------------Ideal------------------------------------------ 2580 // If the store is from an AND mask that leaves the low bits untouched, then 2581 // we can skip the AND operation. If the store is from a sign-extension 2582 // (a left shift, then right shift) we can skip both. 2583 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2584 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2585 if( progress != NULL ) return progress; 2586 2587 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2588 if( progress != NULL ) return progress; 2589 2590 // Finally check the default case 2591 return StoreNode::Ideal(phase, can_reshape); 2592 } 2593 2594 //============================================================================= 2595 //------------------------------Ideal------------------------------------------ 2596 // If the store is from an AND mask that leaves the low bits untouched, then 2597 // we can skip the AND operation 2598 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2599 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2600 if( progress != NULL ) return progress; 2601 2602 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2603 if( progress != NULL ) return progress; 2604 2605 // Finally check the default case 2606 return StoreNode::Ideal(phase, can_reshape); 2607 } 2608 2609 //============================================================================= 2610 //------------------------------Identity--------------------------------------- 2611 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2612 // No need to card mark when storing a null ptr 2613 Node* my_store = in(MemNode::OopStore); 2614 if (my_store->is_Store()) { 2615 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2616 if( t1 == TypePtr::NULL_PTR ) { 2617 return in(MemNode::Memory); 2618 } 2619 } 2620 return this; 2621 } 2622 2623 //============================================================================= 2624 //------------------------------Ideal--------------------------------------- 2625 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2626 Node* progress = StoreNode::Ideal(phase, can_reshape); 2627 if (progress != NULL) return progress; 2628 2629 Node* my_store = in(MemNode::OopStore); 2630 if (my_store->is_MergeMem()) { 2631 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2632 set_req(MemNode::OopStore, mem); 2633 return this; 2634 } 2635 2636 return NULL; 2637 } 2638 2639 //------------------------------Value----------------------------------------- 2640 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2641 // Either input is TOP ==> the result is TOP 2642 const Type *t = phase->type( in(MemNode::Memory) ); 2643 if( t == Type::TOP ) return Type::TOP; 2644 t = phase->type( in(MemNode::Address) ); 2645 if( t == Type::TOP ) return Type::TOP; 2646 t = phase->type( in(MemNode::ValueIn) ); 2647 if( t == Type::TOP ) return Type::TOP; 2648 // If extra input is TOP ==> the result is TOP 2649 t = phase->type( in(MemNode::OopStore) ); 2650 if( t == Type::TOP ) return Type::TOP; 2651 2652 return StoreNode::Value( phase ); 2653 } 2654 2655 2656 //============================================================================= 2657 //----------------------------------SCMemProjNode------------------------------ 2658 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2659 { 2660 return bottom_type(); 2661 } 2662 2663 //============================================================================= 2664 //----------------------------------LoadStoreNode------------------------------ 2665 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2666 : Node(required), 2667 _type(rt), 2668 _adr_type(at) 2669 { 2670 init_req(MemNode::Control, c ); 2671 init_req(MemNode::Memory , mem); 2672 init_req(MemNode::Address, adr); 2673 init_req(MemNode::ValueIn, val); 2674 init_class_id(Class_LoadStore); 2675 } 2676 2677 uint LoadStoreNode::ideal_reg() const { 2678 return _type->ideal_reg(); 2679 } 2680 2681 bool LoadStoreNode::result_not_used() const { 2682 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2683 Node *x = fast_out(i); 2684 if (x->Opcode() == Op_SCMemProj) continue; 2685 return false; 2686 } 2687 return true; 2688 } 2689 2690 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2691 2692 //============================================================================= 2693 //----------------------------------LoadStoreConditionalNode-------------------- 2694 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2695 init_req(ExpectedIn, ex ); 2696 } 2697 2698 //============================================================================= 2699 //-------------------------------adr_type-------------------------------------- 2700 const TypePtr* ClearArrayNode::adr_type() const { 2701 Node *adr = in(3); 2702 if (adr == NULL) return NULL; // node is dead 2703 return MemNode::calculate_adr_type(adr->bottom_type()); 2704 } 2705 2706 //------------------------------match_edge------------------------------------- 2707 // Do we Match on this edge index or not? Do not match memory 2708 uint ClearArrayNode::match_edge(uint idx) const { 2709 return idx > 1; 2710 } 2711 2712 //------------------------------Identity--------------------------------------- 2713 // Clearing a zero length array does nothing 2714 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2715 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2716 } 2717 2718 //------------------------------Idealize--------------------------------------- 2719 // Clearing a short array is faster with stores 2720 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2721 // Already know this is a large node, do not try to ideal it 2722 if (!IdealizeClearArrayNode || _is_large) return NULL; 2723 2724 const int unit = BytesPerLong; 2725 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2726 if (!t) return NULL; 2727 if (!t->is_con()) return NULL; 2728 intptr_t raw_count = t->get_con(); 2729 intptr_t size = raw_count; 2730 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2731 // Clearing nothing uses the Identity call. 2732 // Negative clears are possible on dead ClearArrays 2733 // (see jck test stmt114.stmt11402.val). 2734 if (size <= 0 || size % unit != 0) return NULL; 2735 intptr_t count = size / unit; 2736 // Length too long; communicate this to matchers and assemblers. 2737 // Assemblers are responsible to produce fast hardware clears for it. 2738 if (size > InitArrayShortSize) { 2739 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 2740 } 2741 Node *mem = in(1); 2742 if( phase->type(mem)==Type::TOP ) return NULL; 2743 Node *adr = in(3); 2744 const Type* at = phase->type(adr); 2745 if( at==Type::TOP ) return NULL; 2746 const TypePtr* atp = at->isa_ptr(); 2747 // adjust atp to be the correct array element address type 2748 if (atp == NULL) atp = TypePtr::BOTTOM; 2749 else atp = atp->add_offset(Type::OffsetBot); 2750 // Get base for derived pointer purposes 2751 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2752 Node *base = adr->in(1); 2753 2754 Node *zero = phase->makecon(TypeLong::ZERO); 2755 Node *off = phase->MakeConX(BytesPerLong); 2756 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2757 count--; 2758 while( count-- ) { 2759 mem = phase->transform(mem); 2760 adr = phase->transform(new AddPNode(base,adr,off)); 2761 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2762 } 2763 return mem; 2764 } 2765 2766 //----------------------------step_through---------------------------------- 2767 // Return allocation input memory edge if it is different instance 2768 // or itself if it is the one we are looking for. 2769 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2770 Node* n = *np; 2771 assert(n->is_ClearArray(), "sanity"); 2772 intptr_t offset; 2773 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2774 // This method is called only before Allocate nodes are expanded 2775 // during macro nodes expansion. Before that ClearArray nodes are 2776 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2777 // Allocate nodes are expanded) which follows allocations. 2778 assert(alloc != NULL, "should have allocation"); 2779 if (alloc->_idx == instance_id) { 2780 // Can not bypass initialization of the instance we are looking for. 2781 return false; 2782 } 2783 // Otherwise skip it. 2784 InitializeNode* init = alloc->initialization(); 2785 if (init != NULL) 2786 *np = init->in(TypeFunc::Memory); 2787 else 2788 *np = alloc->in(TypeFunc::Memory); 2789 return true; 2790 } 2791 2792 //----------------------------clear_memory------------------------------------- 2793 // Generate code to initialize object storage to zero. 2794 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2795 intptr_t start_offset, 2796 Node* end_offset, 2797 PhaseGVN* phase) { 2798 intptr_t offset = start_offset; 2799 2800 int unit = BytesPerLong; 2801 if ((offset % unit) != 0) { 2802 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2803 adr = phase->transform(adr); 2804 const TypePtr* atp = TypeRawPtr::BOTTOM; 2805 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2806 mem = phase->transform(mem); 2807 offset += BytesPerInt; 2808 } 2809 assert((offset % unit) == 0, ""); 2810 2811 // Initialize the remaining stuff, if any, with a ClearArray. 2812 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2813 } 2814 2815 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2816 Node* start_offset, 2817 Node* end_offset, 2818 PhaseGVN* phase) { 2819 if (start_offset == end_offset) { 2820 // nothing to do 2821 return mem; 2822 } 2823 2824 int unit = BytesPerLong; 2825 Node* zbase = start_offset; 2826 Node* zend = end_offset; 2827 2828 // Scale to the unit required by the CPU: 2829 if (!Matcher::init_array_count_is_in_bytes) { 2830 Node* shift = phase->intcon(exact_log2(unit)); 2831 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2832 zend = phase->transform(new URShiftXNode(zend, shift) ); 2833 } 2834 2835 // Bulk clear double-words 2836 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2837 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2838 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 2839 return phase->transform(mem); 2840 } 2841 2842 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2843 intptr_t start_offset, 2844 intptr_t end_offset, 2845 PhaseGVN* phase) { 2846 if (start_offset == end_offset) { 2847 // nothing to do 2848 return mem; 2849 } 2850 2851 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2852 intptr_t done_offset = end_offset; 2853 if ((done_offset % BytesPerLong) != 0) { 2854 done_offset -= BytesPerInt; 2855 } 2856 if (done_offset > start_offset) { 2857 mem = clear_memory(ctl, mem, dest, 2858 start_offset, phase->MakeConX(done_offset), phase); 2859 } 2860 if (done_offset < end_offset) { // emit the final 32-bit store 2861 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 2862 adr = phase->transform(adr); 2863 const TypePtr* atp = TypeRawPtr::BOTTOM; 2864 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2865 mem = phase->transform(mem); 2866 done_offset += BytesPerInt; 2867 } 2868 assert(done_offset == end_offset, ""); 2869 return mem; 2870 } 2871 2872 //============================================================================= 2873 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2874 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2875 _adr_type(C->get_adr_type(alias_idx)) 2876 { 2877 init_class_id(Class_MemBar); 2878 Node* top = C->top(); 2879 init_req(TypeFunc::I_O,top); 2880 init_req(TypeFunc::FramePtr,top); 2881 init_req(TypeFunc::ReturnAdr,top); 2882 if (precedent != NULL) 2883 init_req(TypeFunc::Parms, precedent); 2884 } 2885 2886 //------------------------------cmp-------------------------------------------- 2887 uint MemBarNode::hash() const { return NO_HASH; } 2888 uint MemBarNode::cmp( const Node &n ) const { 2889 return (&n == this); // Always fail except on self 2890 } 2891 2892 //------------------------------make------------------------------------------- 2893 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2894 switch (opcode) { 2895 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 2896 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 2897 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 2898 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 2899 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 2900 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 2901 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 2902 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 2903 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 2904 case Op_Initialize: return new InitializeNode(C, atp, pn); 2905 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 2906 default: ShouldNotReachHere(); return NULL; 2907 } 2908 } 2909 2910 //------------------------------Ideal------------------------------------------ 2911 // Return a node which is more "ideal" than the current node. Strip out 2912 // control copies 2913 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2914 if (remove_dead_region(phase, can_reshape)) return this; 2915 // Don't bother trying to transform a dead node 2916 if (in(0) && in(0)->is_top()) { 2917 return NULL; 2918 } 2919 2920 bool progress = false; 2921 // Eliminate volatile MemBars for scalar replaced objects. 2922 if (can_reshape && req() == (Precedent+1)) { 2923 bool eliminate = false; 2924 int opc = Opcode(); 2925 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2926 // Volatile field loads and stores. 2927 Node* my_mem = in(MemBarNode::Precedent); 2928 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2929 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2930 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2931 // replace this Precedent (decodeN) with the Load instead. 2932 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2933 Node* load_node = my_mem->in(1); 2934 set_req(MemBarNode::Precedent, load_node); 2935 phase->is_IterGVN()->_worklist.push(my_mem); 2936 my_mem = load_node; 2937 } else { 2938 assert(my_mem->unique_out() == this, "sanity"); 2939 del_req(Precedent); 2940 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2941 my_mem = NULL; 2942 } 2943 progress = true; 2944 } 2945 if (my_mem != NULL && my_mem->is_Mem()) { 2946 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2947 // Check for scalar replaced object reference. 2948 if( t_oop != NULL && t_oop->is_known_instance_field() && 2949 t_oop->offset() != Type::OffsetBot && 2950 t_oop->offset() != Type::OffsetTop) { 2951 eliminate = true; 2952 } 2953 } 2954 } else if (opc == Op_MemBarRelease) { 2955 // Final field stores. 2956 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 2957 if ((alloc != NULL) && alloc->is_Allocate() && 2958 alloc->as_Allocate()->does_not_escape_thread()) { 2959 // The allocated object does not escape. 2960 eliminate = true; 2961 } 2962 } 2963 if (eliminate) { 2964 // Replace MemBar projections by its inputs. 2965 PhaseIterGVN* igvn = phase->is_IterGVN(); 2966 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2967 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2968 // Must return either the original node (now dead) or a new node 2969 // (Do not return a top here, since that would break the uniqueness of top.) 2970 return new ConINode(TypeInt::ZERO); 2971 } 2972 } 2973 return progress ? this : NULL; 2974 } 2975 2976 //------------------------------Value------------------------------------------ 2977 const Type* MemBarNode::Value(PhaseGVN* phase) const { 2978 if( !in(0) ) return Type::TOP; 2979 if( phase->type(in(0)) == Type::TOP ) 2980 return Type::TOP; 2981 return TypeTuple::MEMBAR; 2982 } 2983 2984 //------------------------------match------------------------------------------ 2985 // Construct projections for memory. 2986 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2987 switch (proj->_con) { 2988 case TypeFunc::Control: 2989 case TypeFunc::Memory: 2990 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2991 } 2992 ShouldNotReachHere(); 2993 return NULL; 2994 } 2995 2996 //===========================InitializeNode==================================== 2997 // SUMMARY: 2998 // This node acts as a memory barrier on raw memory, after some raw stores. 2999 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3000 // The Initialize can 'capture' suitably constrained stores as raw inits. 3001 // It can coalesce related raw stores into larger units (called 'tiles'). 3002 // It can avoid zeroing new storage for memory units which have raw inits. 3003 // At macro-expansion, it is marked 'complete', and does not optimize further. 3004 // 3005 // EXAMPLE: 3006 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3007 // ctl = incoming control; mem* = incoming memory 3008 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3009 // First allocate uninitialized memory and fill in the header: 3010 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3011 // ctl := alloc.Control; mem* := alloc.Memory* 3012 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3013 // Then initialize to zero the non-header parts of the raw memory block: 3014 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3015 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3016 // After the initialize node executes, the object is ready for service: 3017 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3018 // Suppose its body is immediately initialized as {1,2}: 3019 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3020 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3021 // mem.SLICE(#short[*]) := store2 3022 // 3023 // DETAILS: 3024 // An InitializeNode collects and isolates object initialization after 3025 // an AllocateNode and before the next possible safepoint. As a 3026 // memory barrier (MemBarNode), it keeps critical stores from drifting 3027 // down past any safepoint or any publication of the allocation. 3028 // Before this barrier, a newly-allocated object may have uninitialized bits. 3029 // After this barrier, it may be treated as a real oop, and GC is allowed. 3030 // 3031 // The semantics of the InitializeNode include an implicit zeroing of 3032 // the new object from object header to the end of the object. 3033 // (The object header and end are determined by the AllocateNode.) 3034 // 3035 // Certain stores may be added as direct inputs to the InitializeNode. 3036 // These stores must update raw memory, and they must be to addresses 3037 // derived from the raw address produced by AllocateNode, and with 3038 // a constant offset. They must be ordered by increasing offset. 3039 // The first one is at in(RawStores), the last at in(req()-1). 3040 // Unlike most memory operations, they are not linked in a chain, 3041 // but are displayed in parallel as users of the rawmem output of 3042 // the allocation. 3043 // 3044 // (See comments in InitializeNode::capture_store, which continue 3045 // the example given above.) 3046 // 3047 // When the associated Allocate is macro-expanded, the InitializeNode 3048 // may be rewritten to optimize collected stores. A ClearArrayNode 3049 // may also be created at that point to represent any required zeroing. 3050 // The InitializeNode is then marked 'complete', prohibiting further 3051 // capturing of nearby memory operations. 3052 // 3053 // During macro-expansion, all captured initializations which store 3054 // constant values of 32 bits or smaller are coalesced (if advantageous) 3055 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3056 // initialized in fewer memory operations. Memory words which are 3057 // covered by neither tiles nor non-constant stores are pre-zeroed 3058 // by explicit stores of zero. (The code shape happens to do all 3059 // zeroing first, then all other stores, with both sequences occurring 3060 // in order of ascending offsets.) 3061 // 3062 // Alternatively, code may be inserted between an AllocateNode and its 3063 // InitializeNode, to perform arbitrary initialization of the new object. 3064 // E.g., the object copying intrinsics insert complex data transfers here. 3065 // The initialization must then be marked as 'complete' disable the 3066 // built-in zeroing semantics and the collection of initializing stores. 3067 // 3068 // While an InitializeNode is incomplete, reads from the memory state 3069 // produced by it are optimizable if they match the control edge and 3070 // new oop address associated with the allocation/initialization. 3071 // They return a stored value (if the offset matches) or else zero. 3072 // A write to the memory state, if it matches control and address, 3073 // and if it is to a constant offset, may be 'captured' by the 3074 // InitializeNode. It is cloned as a raw memory operation and rewired 3075 // inside the initialization, to the raw oop produced by the allocation. 3076 // Operations on addresses which are provably distinct (e.g., to 3077 // other AllocateNodes) are allowed to bypass the initialization. 3078 // 3079 // The effect of all this is to consolidate object initialization 3080 // (both arrays and non-arrays, both piecewise and bulk) into a 3081 // single location, where it can be optimized as a unit. 3082 // 3083 // Only stores with an offset less than TrackedInitializationLimit words 3084 // will be considered for capture by an InitializeNode. This puts a 3085 // reasonable limit on the complexity of optimized initializations. 3086 3087 //---------------------------InitializeNode------------------------------------ 3088 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3089 : _is_complete(Incomplete), _does_not_escape(false), 3090 MemBarNode(C, adr_type, rawoop) 3091 { 3092 init_class_id(Class_Initialize); 3093 3094 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3095 assert(in(RawAddress) == rawoop, "proper init"); 3096 // Note: allocation() can be NULL, for secondary initialization barriers 3097 } 3098 3099 // Since this node is not matched, it will be processed by the 3100 // register allocator. Declare that there are no constraints 3101 // on the allocation of the RawAddress edge. 3102 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3103 // This edge should be set to top, by the set_complete. But be conservative. 3104 if (idx == InitializeNode::RawAddress) 3105 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3106 return RegMask::Empty; 3107 } 3108 3109 Node* InitializeNode::memory(uint alias_idx) { 3110 Node* mem = in(Memory); 3111 if (mem->is_MergeMem()) { 3112 return mem->as_MergeMem()->memory_at(alias_idx); 3113 } else { 3114 // incoming raw memory is not split 3115 return mem; 3116 } 3117 } 3118 3119 bool InitializeNode::is_non_zero() { 3120 if (is_complete()) return false; 3121 remove_extra_zeroes(); 3122 return (req() > RawStores); 3123 } 3124 3125 void InitializeNode::set_complete(PhaseGVN* phase) { 3126 assert(!is_complete(), "caller responsibility"); 3127 _is_complete = Complete; 3128 3129 // After this node is complete, it contains a bunch of 3130 // raw-memory initializations. There is no need for 3131 // it to have anything to do with non-raw memory effects. 3132 // Therefore, tell all non-raw users to re-optimize themselves, 3133 // after skipping the memory effects of this initialization. 3134 PhaseIterGVN* igvn = phase->is_IterGVN(); 3135 if (igvn) igvn->add_users_to_worklist(this); 3136 } 3137 3138 // convenience function 3139 // return false if the init contains any stores already 3140 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3141 InitializeNode* init = initialization(); 3142 if (init == NULL || init->is_complete()) return false; 3143 init->remove_extra_zeroes(); 3144 // for now, if this allocation has already collected any inits, bail: 3145 if (init->is_non_zero()) return false; 3146 init->set_complete(phase); 3147 return true; 3148 } 3149 3150 void InitializeNode::remove_extra_zeroes() { 3151 if (req() == RawStores) return; 3152 Node* zmem = zero_memory(); 3153 uint fill = RawStores; 3154 for (uint i = fill; i < req(); i++) { 3155 Node* n = in(i); 3156 if (n->is_top() || n == zmem) continue; // skip 3157 if (fill < i) set_req(fill, n); // compact 3158 ++fill; 3159 } 3160 // delete any empty spaces created: 3161 while (fill < req()) { 3162 del_req(fill); 3163 } 3164 } 3165 3166 // Helper for remembering which stores go with which offsets. 3167 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3168 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3169 intptr_t offset = -1; 3170 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3171 phase, offset); 3172 if (base == NULL) return -1; // something is dead, 3173 if (offset < 0) return -1; // dead, dead 3174 return offset; 3175 } 3176 3177 // Helper for proving that an initialization expression is 3178 // "simple enough" to be folded into an object initialization. 3179 // Attempts to prove that a store's initial value 'n' can be captured 3180 // within the initialization without creating a vicious cycle, such as: 3181 // { Foo p = new Foo(); p.next = p; } 3182 // True for constants and parameters and small combinations thereof. 3183 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3184 if (n == NULL) return true; // (can this really happen?) 3185 if (n->is_Proj()) n = n->in(0); 3186 if (n == this) return false; // found a cycle 3187 if (n->is_Con()) return true; 3188 if (n->is_Start()) return true; // params, etc., are OK 3189 if (n->is_Root()) return true; // even better 3190 3191 Node* ctl = n->in(0); 3192 if (ctl != NULL && !ctl->is_top()) { 3193 if (ctl->is_Proj()) ctl = ctl->in(0); 3194 if (ctl == this) return false; 3195 3196 // If we already know that the enclosing memory op is pinned right after 3197 // the init, then any control flow that the store has picked up 3198 // must have preceded the init, or else be equal to the init. 3199 // Even after loop optimizations (which might change control edges) 3200 // a store is never pinned *before* the availability of its inputs. 3201 if (!MemNode::all_controls_dominate(n, this)) 3202 return false; // failed to prove a good control 3203 } 3204 3205 // Check data edges for possible dependencies on 'this'. 3206 if ((count += 1) > 20) return false; // complexity limit 3207 for (uint i = 1; i < n->req(); i++) { 3208 Node* m = n->in(i); 3209 if (m == NULL || m == n || m->is_top()) continue; 3210 uint first_i = n->find_edge(m); 3211 if (i != first_i) continue; // process duplicate edge just once 3212 if (!detect_init_independence(m, count)) { 3213 return false; 3214 } 3215 } 3216 3217 return true; 3218 } 3219 3220 // Here are all the checks a Store must pass before it can be moved into 3221 // an initialization. Returns zero if a check fails. 3222 // On success, returns the (constant) offset to which the store applies, 3223 // within the initialized memory. 3224 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3225 const int FAIL = 0; 3226 if (st->is_unaligned_access()) { 3227 return FAIL; 3228 } 3229 if (st->req() != MemNode::ValueIn + 1) 3230 return FAIL; // an inscrutable StoreNode (card mark?) 3231 Node* ctl = st->in(MemNode::Control); 3232 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3233 return FAIL; // must be unconditional after the initialization 3234 Node* mem = st->in(MemNode::Memory); 3235 if (!(mem->is_Proj() && mem->in(0) == this)) 3236 return FAIL; // must not be preceded by other stores 3237 Node* adr = st->in(MemNode::Address); 3238 intptr_t offset; 3239 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3240 if (alloc == NULL) 3241 return FAIL; // inscrutable address 3242 if (alloc != allocation()) 3243 return FAIL; // wrong allocation! (store needs to float up) 3244 Node* val = st->in(MemNode::ValueIn); 3245 int complexity_count = 0; 3246 if (!detect_init_independence(val, complexity_count)) 3247 return FAIL; // stored value must be 'simple enough' 3248 3249 // The Store can be captured only if nothing after the allocation 3250 // and before the Store is using the memory location that the store 3251 // overwrites. 3252 bool failed = false; 3253 // If is_complete_with_arraycopy() is true the shape of the graph is 3254 // well defined and is safe so no need for extra checks. 3255 if (!is_complete_with_arraycopy()) { 3256 // We are going to look at each use of the memory state following 3257 // the allocation to make sure nothing reads the memory that the 3258 // Store writes. 3259 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3260 int alias_idx = phase->C->get_alias_index(t_adr); 3261 ResourceMark rm; 3262 Unique_Node_List mems; 3263 mems.push(mem); 3264 Node* unique_merge = NULL; 3265 for (uint next = 0; next < mems.size(); ++next) { 3266 Node *m = mems.at(next); 3267 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3268 Node *n = m->fast_out(j); 3269 if (n->outcnt() == 0) { 3270 continue; 3271 } 3272 if (n == st) { 3273 continue; 3274 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3275 // If the control of this use is different from the control 3276 // of the Store which is right after the InitializeNode then 3277 // this node cannot be between the InitializeNode and the 3278 // Store. 3279 continue; 3280 } else if (n->is_MergeMem()) { 3281 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3282 // We can hit a MergeMemNode (that will likely go away 3283 // later) that is a direct use of the memory state 3284 // following the InitializeNode on the same slice as the 3285 // store node that we'd like to capture. We need to check 3286 // the uses of the MergeMemNode. 3287 mems.push(n); 3288 } 3289 } else if (n->is_Mem()) { 3290 Node* other_adr = n->in(MemNode::Address); 3291 if (other_adr == adr) { 3292 failed = true; 3293 break; 3294 } else { 3295 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3296 if (other_t_adr != NULL) { 3297 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3298 if (other_alias_idx == alias_idx) { 3299 // A load from the same memory slice as the store right 3300 // after the InitializeNode. We check the control of the 3301 // object/array that is loaded from. If it's the same as 3302 // the store control then we cannot capture the store. 3303 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3304 Node* base = other_adr; 3305 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3306 base = base->in(AddPNode::Base); 3307 if (base != NULL) { 3308 base = base->uncast(); 3309 if (base->is_Proj() && base->in(0) == alloc) { 3310 failed = true; 3311 break; 3312 } 3313 } 3314 } 3315 } 3316 } 3317 } else { 3318 failed = true; 3319 break; 3320 } 3321 } 3322 } 3323 } 3324 if (failed) { 3325 if (!can_reshape) { 3326 // We decided we couldn't capture the store during parsing. We 3327 // should try again during the next IGVN once the graph is 3328 // cleaner. 3329 phase->C->record_for_igvn(st); 3330 } 3331 return FAIL; 3332 } 3333 3334 return offset; // success 3335 } 3336 3337 // Find the captured store in(i) which corresponds to the range 3338 // [start..start+size) in the initialized object. 3339 // If there is one, return its index i. If there isn't, return the 3340 // negative of the index where it should be inserted. 3341 // Return 0 if the queried range overlaps an initialization boundary 3342 // or if dead code is encountered. 3343 // If size_in_bytes is zero, do not bother with overlap checks. 3344 int InitializeNode::captured_store_insertion_point(intptr_t start, 3345 int size_in_bytes, 3346 PhaseTransform* phase) { 3347 const int FAIL = 0, MAX_STORE = BytesPerLong; 3348 3349 if (is_complete()) 3350 return FAIL; // arraycopy got here first; punt 3351 3352 assert(allocation() != NULL, "must be present"); 3353 3354 // no negatives, no header fields: 3355 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3356 3357 // after a certain size, we bail out on tracking all the stores: 3358 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3359 if (start >= ti_limit) return FAIL; 3360 3361 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3362 if (i >= limit) return -(int)i; // not found; here is where to put it 3363 3364 Node* st = in(i); 3365 intptr_t st_off = get_store_offset(st, phase); 3366 if (st_off < 0) { 3367 if (st != zero_memory()) { 3368 return FAIL; // bail out if there is dead garbage 3369 } 3370 } else if (st_off > start) { 3371 // ...we are done, since stores are ordered 3372 if (st_off < start + size_in_bytes) { 3373 return FAIL; // the next store overlaps 3374 } 3375 return -(int)i; // not found; here is where to put it 3376 } else if (st_off < start) { 3377 if (size_in_bytes != 0 && 3378 start < st_off + MAX_STORE && 3379 start < st_off + st->as_Store()->memory_size()) { 3380 return FAIL; // the previous store overlaps 3381 } 3382 } else { 3383 if (size_in_bytes != 0 && 3384 st->as_Store()->memory_size() != size_in_bytes) { 3385 return FAIL; // mismatched store size 3386 } 3387 return i; 3388 } 3389 3390 ++i; 3391 } 3392 } 3393 3394 // Look for a captured store which initializes at the offset 'start' 3395 // with the given size. If there is no such store, and no other 3396 // initialization interferes, then return zero_memory (the memory 3397 // projection of the AllocateNode). 3398 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3399 PhaseTransform* phase) { 3400 assert(stores_are_sane(phase), ""); 3401 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3402 if (i == 0) { 3403 return NULL; // something is dead 3404 } else if (i < 0) { 3405 return zero_memory(); // just primordial zero bits here 3406 } else { 3407 Node* st = in(i); // here is the store at this position 3408 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3409 return st; 3410 } 3411 } 3412 3413 // Create, as a raw pointer, an address within my new object at 'offset'. 3414 Node* InitializeNode::make_raw_address(intptr_t offset, 3415 PhaseTransform* phase) { 3416 Node* addr = in(RawAddress); 3417 if (offset != 0) { 3418 Compile* C = phase->C; 3419 addr = phase->transform( new AddPNode(C->top(), addr, 3420 phase->MakeConX(offset)) ); 3421 } 3422 return addr; 3423 } 3424 3425 // Clone the given store, converting it into a raw store 3426 // initializing a field or element of my new object. 3427 // Caller is responsible for retiring the original store, 3428 // with subsume_node or the like. 3429 // 3430 // From the example above InitializeNode::InitializeNode, 3431 // here are the old stores to be captured: 3432 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3433 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3434 // 3435 // Here is the changed code; note the extra edges on init: 3436 // alloc = (Allocate ...) 3437 // rawoop = alloc.RawAddress 3438 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3439 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3440 // init = (Initialize alloc.Control alloc.Memory rawoop 3441 // rawstore1 rawstore2) 3442 // 3443 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3444 PhaseTransform* phase, bool can_reshape) { 3445 assert(stores_are_sane(phase), ""); 3446 3447 if (start < 0) return NULL; 3448 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3449 3450 Compile* C = phase->C; 3451 int size_in_bytes = st->memory_size(); 3452 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3453 if (i == 0) return NULL; // bail out 3454 Node* prev_mem = NULL; // raw memory for the captured store 3455 if (i > 0) { 3456 prev_mem = in(i); // there is a pre-existing store under this one 3457 set_req(i, C->top()); // temporarily disconnect it 3458 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3459 } else { 3460 i = -i; // no pre-existing store 3461 prev_mem = zero_memory(); // a slice of the newly allocated object 3462 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3463 set_req(--i, C->top()); // reuse this edge; it has been folded away 3464 else 3465 ins_req(i, C->top()); // build a new edge 3466 } 3467 Node* new_st = st->clone(); 3468 new_st->set_req(MemNode::Control, in(Control)); 3469 new_st->set_req(MemNode::Memory, prev_mem); 3470 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3471 new_st = phase->transform(new_st); 3472 3473 // At this point, new_st might have swallowed a pre-existing store 3474 // at the same offset, or perhaps new_st might have disappeared, 3475 // if it redundantly stored the same value (or zero to fresh memory). 3476 3477 // In any case, wire it in: 3478 phase->igvn_rehash_node_delayed(this); 3479 set_req(i, new_st); 3480 3481 // The caller may now kill the old guy. 3482 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3483 assert(check_st == new_st || check_st == NULL, "must be findable"); 3484 assert(!is_complete(), ""); 3485 return new_st; 3486 } 3487 3488 static bool store_constant(jlong* tiles, int num_tiles, 3489 intptr_t st_off, int st_size, 3490 jlong con) { 3491 if ((st_off & (st_size-1)) != 0) 3492 return false; // strange store offset (assume size==2**N) 3493 address addr = (address)tiles + st_off; 3494 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3495 switch (st_size) { 3496 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3497 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3498 case sizeof(jint): *(jint*) addr = (jint) con; break; 3499 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3500 default: return false; // strange store size (detect size!=2**N here) 3501 } 3502 return true; // return success to caller 3503 } 3504 3505 // Coalesce subword constants into int constants and possibly 3506 // into long constants. The goal, if the CPU permits, 3507 // is to initialize the object with a small number of 64-bit tiles. 3508 // Also, convert floating-point constants to bit patterns. 3509 // Non-constants are not relevant to this pass. 3510 // 3511 // In terms of the running example on InitializeNode::InitializeNode 3512 // and InitializeNode::capture_store, here is the transformation 3513 // of rawstore1 and rawstore2 into rawstore12: 3514 // alloc = (Allocate ...) 3515 // rawoop = alloc.RawAddress 3516 // tile12 = 0x00010002 3517 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3518 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3519 // 3520 void 3521 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3522 Node* size_in_bytes, 3523 PhaseGVN* phase) { 3524 Compile* C = phase->C; 3525 3526 assert(stores_are_sane(phase), ""); 3527 // Note: After this pass, they are not completely sane, 3528 // since there may be some overlaps. 3529 3530 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3531 3532 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3533 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3534 size_limit = MIN2(size_limit, ti_limit); 3535 size_limit = align_size_up(size_limit, BytesPerLong); 3536 int num_tiles = size_limit / BytesPerLong; 3537 3538 // allocate space for the tile map: 3539 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3540 jlong tiles_buf[small_len]; 3541 Node* nodes_buf[small_len]; 3542 jlong inits_buf[small_len]; 3543 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3544 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3545 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3546 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3547 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3548 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3549 // tiles: exact bitwise model of all primitive constants 3550 // nodes: last constant-storing node subsumed into the tiles model 3551 // inits: which bytes (in each tile) are touched by any initializations 3552 3553 //// Pass A: Fill in the tile model with any relevant stores. 3554 3555 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3556 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3557 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3558 Node* zmem = zero_memory(); // initially zero memory state 3559 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3560 Node* st = in(i); 3561 intptr_t st_off = get_store_offset(st, phase); 3562 3563 // Figure out the store's offset and constant value: 3564 if (st_off < header_size) continue; //skip (ignore header) 3565 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3566 int st_size = st->as_Store()->memory_size(); 3567 if (st_off + st_size > size_limit) break; 3568 3569 // Record which bytes are touched, whether by constant or not. 3570 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3571 continue; // skip (strange store size) 3572 3573 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3574 if (!val->singleton()) continue; //skip (non-con store) 3575 BasicType type = val->basic_type(); 3576 3577 jlong con = 0; 3578 switch (type) { 3579 case T_INT: con = val->is_int()->get_con(); break; 3580 case T_LONG: con = val->is_long()->get_con(); break; 3581 case T_FLOAT: con = jint_cast(val->getf()); break; 3582 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3583 default: continue; //skip (odd store type) 3584 } 3585 3586 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3587 st->Opcode() == Op_StoreL) { 3588 continue; // This StoreL is already optimal. 3589 } 3590 3591 // Store down the constant. 3592 store_constant(tiles, num_tiles, st_off, st_size, con); 3593 3594 intptr_t j = st_off >> LogBytesPerLong; 3595 3596 if (type == T_INT && st_size == BytesPerInt 3597 && (st_off & BytesPerInt) == BytesPerInt) { 3598 jlong lcon = tiles[j]; 3599 if (!Matcher::isSimpleConstant64(lcon) && 3600 st->Opcode() == Op_StoreI) { 3601 // This StoreI is already optimal by itself. 3602 jint* intcon = (jint*) &tiles[j]; 3603 intcon[1] = 0; // undo the store_constant() 3604 3605 // If the previous store is also optimal by itself, back up and 3606 // undo the action of the previous loop iteration... if we can. 3607 // But if we can't, just let the previous half take care of itself. 3608 st = nodes[j]; 3609 st_off -= BytesPerInt; 3610 con = intcon[0]; 3611 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3612 assert(st_off >= header_size, "still ignoring header"); 3613 assert(get_store_offset(st, phase) == st_off, "must be"); 3614 assert(in(i-1) == zmem, "must be"); 3615 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3616 assert(con == tcon->is_int()->get_con(), "must be"); 3617 // Undo the effects of the previous loop trip, which swallowed st: 3618 intcon[0] = 0; // undo store_constant() 3619 set_req(i-1, st); // undo set_req(i, zmem) 3620 nodes[j] = NULL; // undo nodes[j] = st 3621 --old_subword; // undo ++old_subword 3622 } 3623 continue; // This StoreI is already optimal. 3624 } 3625 } 3626 3627 // This store is not needed. 3628 set_req(i, zmem); 3629 nodes[j] = st; // record for the moment 3630 if (st_size < BytesPerLong) // something has changed 3631 ++old_subword; // includes int/float, but who's counting... 3632 else ++old_long; 3633 } 3634 3635 if ((old_subword + old_long) == 0) 3636 return; // nothing more to do 3637 3638 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3639 // Be sure to insert them before overlapping non-constant stores. 3640 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3641 for (int j = 0; j < num_tiles; j++) { 3642 jlong con = tiles[j]; 3643 jlong init = inits[j]; 3644 if (con == 0) continue; 3645 jint con0, con1; // split the constant, address-wise 3646 jint init0, init1; // split the init map, address-wise 3647 { union { jlong con; jint intcon[2]; } u; 3648 u.con = con; 3649 con0 = u.intcon[0]; 3650 con1 = u.intcon[1]; 3651 u.con = init; 3652 init0 = u.intcon[0]; 3653 init1 = u.intcon[1]; 3654 } 3655 3656 Node* old = nodes[j]; 3657 assert(old != NULL, "need the prior store"); 3658 intptr_t offset = (j * BytesPerLong); 3659 3660 bool split = !Matcher::isSimpleConstant64(con); 3661 3662 if (offset < header_size) { 3663 assert(offset + BytesPerInt >= header_size, "second int counts"); 3664 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3665 split = true; // only the second word counts 3666 // Example: int a[] = { 42 ... } 3667 } else if (con0 == 0 && init0 == -1) { 3668 split = true; // first word is covered by full inits 3669 // Example: int a[] = { ... foo(), 42 ... } 3670 } else if (con1 == 0 && init1 == -1) { 3671 split = true; // second word is covered by full inits 3672 // Example: int a[] = { ... 42, foo() ... } 3673 } 3674 3675 // Here's a case where init0 is neither 0 nor -1: 3676 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3677 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3678 // In this case the tile is not split; it is (jlong)42. 3679 // The big tile is stored down, and then the foo() value is inserted. 3680 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3681 3682 Node* ctl = old->in(MemNode::Control); 3683 Node* adr = make_raw_address(offset, phase); 3684 const TypePtr* atp = TypeRawPtr::BOTTOM; 3685 3686 // One or two coalesced stores to plop down. 3687 Node* st[2]; 3688 intptr_t off[2]; 3689 int nst = 0; 3690 if (!split) { 3691 ++new_long; 3692 off[nst] = offset; 3693 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3694 phase->longcon(con), T_LONG, MemNode::unordered); 3695 } else { 3696 // Omit either if it is a zero. 3697 if (con0 != 0) { 3698 ++new_int; 3699 off[nst] = offset; 3700 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3701 phase->intcon(con0), T_INT, MemNode::unordered); 3702 } 3703 if (con1 != 0) { 3704 ++new_int; 3705 offset += BytesPerInt; 3706 adr = make_raw_address(offset, phase); 3707 off[nst] = offset; 3708 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3709 phase->intcon(con1), T_INT, MemNode::unordered); 3710 } 3711 } 3712 3713 // Insert second store first, then the first before the second. 3714 // Insert each one just before any overlapping non-constant stores. 3715 while (nst > 0) { 3716 Node* st1 = st[--nst]; 3717 C->copy_node_notes_to(st1, old); 3718 st1 = phase->transform(st1); 3719 offset = off[nst]; 3720 assert(offset >= header_size, "do not smash header"); 3721 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3722 guarantee(ins_idx != 0, "must re-insert constant store"); 3723 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3724 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3725 set_req(--ins_idx, st1); 3726 else 3727 ins_req(ins_idx, st1); 3728 } 3729 } 3730 3731 if (PrintCompilation && WizardMode) 3732 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3733 old_subword, old_long, new_int, new_long); 3734 if (C->log() != NULL) 3735 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3736 old_subword, old_long, new_int, new_long); 3737 3738 // Clean up any remaining occurrences of zmem: 3739 remove_extra_zeroes(); 3740 } 3741 3742 // Explore forward from in(start) to find the first fully initialized 3743 // word, and return its offset. Skip groups of subword stores which 3744 // together initialize full words. If in(start) is itself part of a 3745 // fully initialized word, return the offset of in(start). If there 3746 // are no following full-word stores, or if something is fishy, return 3747 // a negative value. 3748 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3749 int int_map = 0; 3750 intptr_t int_map_off = 0; 3751 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3752 3753 for (uint i = start, limit = req(); i < limit; i++) { 3754 Node* st = in(i); 3755 3756 intptr_t st_off = get_store_offset(st, phase); 3757 if (st_off < 0) break; // return conservative answer 3758 3759 int st_size = st->as_Store()->memory_size(); 3760 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3761 return st_off; // we found a complete word init 3762 } 3763 3764 // update the map: 3765 3766 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3767 if (this_int_off != int_map_off) { 3768 // reset the map: 3769 int_map = 0; 3770 int_map_off = this_int_off; 3771 } 3772 3773 int subword_off = st_off - this_int_off; 3774 int_map |= right_n_bits(st_size) << subword_off; 3775 if ((int_map & FULL_MAP) == FULL_MAP) { 3776 return this_int_off; // we found a complete word init 3777 } 3778 3779 // Did this store hit or cross the word boundary? 3780 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3781 if (next_int_off == this_int_off + BytesPerInt) { 3782 // We passed the current int, without fully initializing it. 3783 int_map_off = next_int_off; 3784 int_map >>= BytesPerInt; 3785 } else if (next_int_off > this_int_off + BytesPerInt) { 3786 // We passed the current and next int. 3787 return this_int_off + BytesPerInt; 3788 } 3789 } 3790 3791 return -1; 3792 } 3793 3794 3795 // Called when the associated AllocateNode is expanded into CFG. 3796 // At this point, we may perform additional optimizations. 3797 // Linearize the stores by ascending offset, to make memory 3798 // activity as coherent as possible. 3799 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3800 intptr_t header_size, 3801 Node* size_in_bytes, 3802 PhaseGVN* phase) { 3803 assert(!is_complete(), "not already complete"); 3804 assert(stores_are_sane(phase), ""); 3805 assert(allocation() != NULL, "must be present"); 3806 3807 remove_extra_zeroes(); 3808 3809 if (ReduceFieldZeroing || ReduceBulkZeroing) 3810 // reduce instruction count for common initialization patterns 3811 coalesce_subword_stores(header_size, size_in_bytes, phase); 3812 3813 Node* zmem = zero_memory(); // initially zero memory state 3814 Node* inits = zmem; // accumulating a linearized chain of inits 3815 #ifdef ASSERT 3816 intptr_t first_offset = allocation()->minimum_header_size(); 3817 intptr_t last_init_off = first_offset; // previous init offset 3818 intptr_t last_init_end = first_offset; // previous init offset+size 3819 intptr_t last_tile_end = first_offset; // previous tile offset+size 3820 #endif 3821 intptr_t zeroes_done = header_size; 3822 3823 bool do_zeroing = true; // we might give up if inits are very sparse 3824 int big_init_gaps = 0; // how many large gaps have we seen? 3825 3826 if (UseTLAB && ZeroTLAB) do_zeroing = false; 3827 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3828 3829 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3830 Node* st = in(i); 3831 intptr_t st_off = get_store_offset(st, phase); 3832 if (st_off < 0) 3833 break; // unknown junk in the inits 3834 if (st->in(MemNode::Memory) != zmem) 3835 break; // complicated store chains somehow in list 3836 3837 int st_size = st->as_Store()->memory_size(); 3838 intptr_t next_init_off = st_off + st_size; 3839 3840 if (do_zeroing && zeroes_done < next_init_off) { 3841 // See if this store needs a zero before it or under it. 3842 intptr_t zeroes_needed = st_off; 3843 3844 if (st_size < BytesPerInt) { 3845 // Look for subword stores which only partially initialize words. 3846 // If we find some, we must lay down some word-level zeroes first, 3847 // underneath the subword stores. 3848 // 3849 // Examples: 3850 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3851 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3852 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3853 // 3854 // Note: coalesce_subword_stores may have already done this, 3855 // if it was prompted by constant non-zero subword initializers. 3856 // But this case can still arise with non-constant stores. 3857 3858 intptr_t next_full_store = find_next_fullword_store(i, phase); 3859 3860 // In the examples above: 3861 // in(i) p q r s x y z 3862 // st_off 12 13 14 15 12 13 14 3863 // st_size 1 1 1 1 1 1 1 3864 // next_full_s. 12 16 16 16 16 16 16 3865 // z's_done 12 16 16 16 12 16 12 3866 // z's_needed 12 16 16 16 16 16 16 3867 // zsize 0 0 0 0 4 0 4 3868 if (next_full_store < 0) { 3869 // Conservative tack: Zero to end of current word. 3870 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3871 } else { 3872 // Zero to beginning of next fully initialized word. 3873 // Or, don't zero at all, if we are already in that word. 3874 assert(next_full_store >= zeroes_needed, "must go forward"); 3875 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3876 zeroes_needed = next_full_store; 3877 } 3878 } 3879 3880 if (zeroes_needed > zeroes_done) { 3881 intptr_t zsize = zeroes_needed - zeroes_done; 3882 // Do some incremental zeroing on rawmem, in parallel with inits. 3883 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3884 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3885 zeroes_done, zeroes_needed, 3886 phase); 3887 zeroes_done = zeroes_needed; 3888 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 3889 do_zeroing = false; // leave the hole, next time 3890 } 3891 } 3892 3893 // Collect the store and move on: 3894 st->set_req(MemNode::Memory, inits); 3895 inits = st; // put it on the linearized chain 3896 set_req(i, zmem); // unhook from previous position 3897 3898 if (zeroes_done == st_off) 3899 zeroes_done = next_init_off; 3900 3901 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3902 3903 #ifdef ASSERT 3904 // Various order invariants. Weaker than stores_are_sane because 3905 // a large constant tile can be filled in by smaller non-constant stores. 3906 assert(st_off >= last_init_off, "inits do not reverse"); 3907 last_init_off = st_off; 3908 const Type* val = NULL; 3909 if (st_size >= BytesPerInt && 3910 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3911 (int)val->basic_type() < (int)T_OBJECT) { 3912 assert(st_off >= last_tile_end, "tiles do not overlap"); 3913 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3914 last_tile_end = MAX2(last_tile_end, next_init_off); 3915 } else { 3916 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3917 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3918 assert(st_off >= last_init_end, "inits do not overlap"); 3919 last_init_end = next_init_off; // it's a non-tile 3920 } 3921 #endif //ASSERT 3922 } 3923 3924 remove_extra_zeroes(); // clear out all the zmems left over 3925 add_req(inits); 3926 3927 if (!(UseTLAB && ZeroTLAB)) { 3928 // If anything remains to be zeroed, zero it all now. 3929 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3930 // if it is the last unused 4 bytes of an instance, forget about it 3931 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3932 if (zeroes_done + BytesPerLong >= size_limit) { 3933 assert(allocation() != NULL, ""); 3934 if (allocation()->Opcode() == Op_Allocate) { 3935 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3936 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3937 if (zeroes_done == k->layout_helper()) 3938 zeroes_done = size_limit; 3939 } 3940 } 3941 if (zeroes_done < size_limit) { 3942 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3943 zeroes_done, size_in_bytes, phase); 3944 } 3945 } 3946 3947 set_complete(phase); 3948 return rawmem; 3949 } 3950 3951 3952 #ifdef ASSERT 3953 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3954 if (is_complete()) 3955 return true; // stores could be anything at this point 3956 assert(allocation() != NULL, "must be present"); 3957 intptr_t last_off = allocation()->minimum_header_size(); 3958 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3959 Node* st = in(i); 3960 intptr_t st_off = get_store_offset(st, phase); 3961 if (st_off < 0) continue; // ignore dead garbage 3962 if (last_off > st_off) { 3963 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 3964 this->dump(2); 3965 assert(false, "ascending store offsets"); 3966 return false; 3967 } 3968 last_off = st_off + st->as_Store()->memory_size(); 3969 } 3970 return true; 3971 } 3972 #endif //ASSERT 3973 3974 3975 3976 3977 //============================MergeMemNode===================================== 3978 // 3979 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3980 // contributing store or call operations. Each contributor provides the memory 3981 // state for a particular "alias type" (see Compile::alias_type). For example, 3982 // if a MergeMem has an input X for alias category #6, then any memory reference 3983 // to alias category #6 may use X as its memory state input, as an exact equivalent 3984 // to using the MergeMem as a whole. 3985 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3986 // 3987 // (Here, the <N> notation gives the index of the relevant adr_type.) 3988 // 3989 // In one special case (and more cases in the future), alias categories overlap. 3990 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3991 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3992 // it is exactly equivalent to that state W: 3993 // MergeMem(<Bot>: W) <==> W 3994 // 3995 // Usually, the merge has more than one input. In that case, where inputs 3996 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3997 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3998 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3999 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4000 // 4001 // A merge can take a "wide" memory state as one of its narrow inputs. 4002 // This simply means that the merge observes out only the relevant parts of 4003 // the wide input. That is, wide memory states arriving at narrow merge inputs 4004 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4005 // 4006 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4007 // and that memory slices "leak through": 4008 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4009 // 4010 // But, in such a cascade, repeated memory slices can "block the leak": 4011 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4012 // 4013 // In the last example, Y is not part of the combined memory state of the 4014 // outermost MergeMem. The system must, of course, prevent unschedulable 4015 // memory states from arising, so you can be sure that the state Y is somehow 4016 // a precursor to state Y'. 4017 // 4018 // 4019 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4020 // of each MergeMemNode array are exactly the numerical alias indexes, including 4021 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4022 // Compile::alias_type (and kin) produce and manage these indexes. 4023 // 4024 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4025 // (Note that this provides quick access to the top node inside MergeMem methods, 4026 // without the need to reach out via TLS to Compile::current.) 4027 // 4028 // As a consequence of what was just described, a MergeMem that represents a full 4029 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4030 // containing all alias categories. 4031 // 4032 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4033 // 4034 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4035 // a memory state for the alias type <N>, or else the top node, meaning that 4036 // there is no particular input for that alias type. Note that the length of 4037 // a MergeMem is variable, and may be extended at any time to accommodate new 4038 // memory states at larger alias indexes. When merges grow, they are of course 4039 // filled with "top" in the unused in() positions. 4040 // 4041 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4042 // (Top was chosen because it works smoothly with passes like GCM.) 4043 // 4044 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4045 // the type of random VM bits like TLS references.) Since it is always the 4046 // first non-Bot memory slice, some low-level loops use it to initialize an 4047 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4048 // 4049 // 4050 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4051 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4052 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4053 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4054 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4055 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4056 // 4057 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4058 // really that different from the other memory inputs. An abbreviation called 4059 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4060 // 4061 // 4062 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4063 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4064 // that "emerges though" the base memory will be marked as excluding the alias types 4065 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4066 // 4067 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4068 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4069 // 4070 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4071 // (It is currently unimplemented.) As you can see, the resulting merge is 4072 // actually a disjoint union of memory states, rather than an overlay. 4073 // 4074 4075 //------------------------------MergeMemNode----------------------------------- 4076 Node* MergeMemNode::make_empty_memory() { 4077 Node* empty_memory = (Node*) Compile::current()->top(); 4078 assert(empty_memory->is_top(), "correct sentinel identity"); 4079 return empty_memory; 4080 } 4081 4082 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4083 init_class_id(Class_MergeMem); 4084 // all inputs are nullified in Node::Node(int) 4085 // set_input(0, NULL); // no control input 4086 4087 // Initialize the edges uniformly to top, for starters. 4088 Node* empty_mem = make_empty_memory(); 4089 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4090 init_req(i,empty_mem); 4091 } 4092 assert(empty_memory() == empty_mem, ""); 4093 4094 if( new_base != NULL && new_base->is_MergeMem() ) { 4095 MergeMemNode* mdef = new_base->as_MergeMem(); 4096 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4097 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4098 mms.set_memory(mms.memory2()); 4099 } 4100 assert(base_memory() == mdef->base_memory(), ""); 4101 } else { 4102 set_base_memory(new_base); 4103 } 4104 } 4105 4106 // Make a new, untransformed MergeMem with the same base as 'mem'. 4107 // If mem is itself a MergeMem, populate the result with the same edges. 4108 MergeMemNode* MergeMemNode::make(Node* mem) { 4109 return new MergeMemNode(mem); 4110 } 4111 4112 //------------------------------cmp-------------------------------------------- 4113 uint MergeMemNode::hash() const { return NO_HASH; } 4114 uint MergeMemNode::cmp( const Node &n ) const { 4115 return (&n == this); // Always fail except on self 4116 } 4117 4118 //------------------------------Identity--------------------------------------- 4119 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4120 // Identity if this merge point does not record any interesting memory 4121 // disambiguations. 4122 Node* base_mem = base_memory(); 4123 Node* empty_mem = empty_memory(); 4124 if (base_mem != empty_mem) { // Memory path is not dead? 4125 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4126 Node* mem = in(i); 4127 if (mem != empty_mem && mem != base_mem) { 4128 return this; // Many memory splits; no change 4129 } 4130 } 4131 } 4132 return base_mem; // No memory splits; ID on the one true input 4133 } 4134 4135 //------------------------------Ideal------------------------------------------ 4136 // This method is invoked recursively on chains of MergeMem nodes 4137 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4138 // Remove chain'd MergeMems 4139 // 4140 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4141 // relative to the "in(Bot)". Since we are patching both at the same time, 4142 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4143 // but rewrite each "in(i)" relative to the new "in(Bot)". 4144 Node *progress = NULL; 4145 4146 4147 Node* old_base = base_memory(); 4148 Node* empty_mem = empty_memory(); 4149 if (old_base == empty_mem) 4150 return NULL; // Dead memory path. 4151 4152 MergeMemNode* old_mbase; 4153 if (old_base != NULL && old_base->is_MergeMem()) 4154 old_mbase = old_base->as_MergeMem(); 4155 else 4156 old_mbase = NULL; 4157 Node* new_base = old_base; 4158 4159 // simplify stacked MergeMems in base memory 4160 if (old_mbase) new_base = old_mbase->base_memory(); 4161 4162 // the base memory might contribute new slices beyond my req() 4163 if (old_mbase) grow_to_match(old_mbase); 4164 4165 // Look carefully at the base node if it is a phi. 4166 PhiNode* phi_base; 4167 if (new_base != NULL && new_base->is_Phi()) 4168 phi_base = new_base->as_Phi(); 4169 else 4170 phi_base = NULL; 4171 4172 Node* phi_reg = NULL; 4173 uint phi_len = (uint)-1; 4174 if (phi_base != NULL && !phi_base->is_copy()) { 4175 // do not examine phi if degraded to a copy 4176 phi_reg = phi_base->region(); 4177 phi_len = phi_base->req(); 4178 // see if the phi is unfinished 4179 for (uint i = 1; i < phi_len; i++) { 4180 if (phi_base->in(i) == NULL) { 4181 // incomplete phi; do not look at it yet! 4182 phi_reg = NULL; 4183 phi_len = (uint)-1; 4184 break; 4185 } 4186 } 4187 } 4188 4189 // Note: We do not call verify_sparse on entry, because inputs 4190 // can normalize to the base_memory via subsume_node or similar 4191 // mechanisms. This method repairs that damage. 4192 4193 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4194 4195 // Look at each slice. 4196 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4197 Node* old_in = in(i); 4198 // calculate the old memory value 4199 Node* old_mem = old_in; 4200 if (old_mem == empty_mem) old_mem = old_base; 4201 assert(old_mem == memory_at(i), ""); 4202 4203 // maybe update (reslice) the old memory value 4204 4205 // simplify stacked MergeMems 4206 Node* new_mem = old_mem; 4207 MergeMemNode* old_mmem; 4208 if (old_mem != NULL && old_mem->is_MergeMem()) 4209 old_mmem = old_mem->as_MergeMem(); 4210 else 4211 old_mmem = NULL; 4212 if (old_mmem == this) { 4213 // This can happen if loops break up and safepoints disappear. 4214 // A merge of BotPtr (default) with a RawPtr memory derived from a 4215 // safepoint can be rewritten to a merge of the same BotPtr with 4216 // the BotPtr phi coming into the loop. If that phi disappears 4217 // also, we can end up with a self-loop of the mergemem. 4218 // In general, if loops degenerate and memory effects disappear, 4219 // a mergemem can be left looking at itself. This simply means 4220 // that the mergemem's default should be used, since there is 4221 // no longer any apparent effect on this slice. 4222 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4223 // from start. Update the input to TOP. 4224 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4225 } 4226 else if (old_mmem != NULL) { 4227 new_mem = old_mmem->memory_at(i); 4228 } 4229 // else preceding memory was not a MergeMem 4230 4231 // replace equivalent phis (unfortunately, they do not GVN together) 4232 if (new_mem != NULL && new_mem != new_base && 4233 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4234 if (new_mem->is_Phi()) { 4235 PhiNode* phi_mem = new_mem->as_Phi(); 4236 for (uint i = 1; i < phi_len; i++) { 4237 if (phi_base->in(i) != phi_mem->in(i)) { 4238 phi_mem = NULL; 4239 break; 4240 } 4241 } 4242 if (phi_mem != NULL) { 4243 // equivalent phi nodes; revert to the def 4244 new_mem = new_base; 4245 } 4246 } 4247 } 4248 4249 // maybe store down a new value 4250 Node* new_in = new_mem; 4251 if (new_in == new_base) new_in = empty_mem; 4252 4253 if (new_in != old_in) { 4254 // Warning: Do not combine this "if" with the previous "if" 4255 // A memory slice might have be be rewritten even if it is semantically 4256 // unchanged, if the base_memory value has changed. 4257 set_req(i, new_in); 4258 progress = this; // Report progress 4259 } 4260 } 4261 4262 if (new_base != old_base) { 4263 set_req(Compile::AliasIdxBot, new_base); 4264 // Don't use set_base_memory(new_base), because we need to update du. 4265 assert(base_memory() == new_base, ""); 4266 progress = this; 4267 } 4268 4269 if( base_memory() == this ) { 4270 // a self cycle indicates this memory path is dead 4271 set_req(Compile::AliasIdxBot, empty_mem); 4272 } 4273 4274 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4275 // Recursion must occur after the self cycle check above 4276 if( base_memory()->is_MergeMem() ) { 4277 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4278 Node *m = phase->transform(new_mbase); // Rollup any cycles 4279 if( m != NULL && (m->is_top() || 4280 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4281 // propagate rollup of dead cycle to self 4282 set_req(Compile::AliasIdxBot, empty_mem); 4283 } 4284 } 4285 4286 if( base_memory() == empty_mem ) { 4287 progress = this; 4288 // Cut inputs during Parse phase only. 4289 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4290 if( !can_reshape ) { 4291 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4292 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4293 } 4294 } 4295 } 4296 4297 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4298 // Check if PhiNode::Ideal's "Split phis through memory merges" 4299 // transform should be attempted. Look for this->phi->this cycle. 4300 uint merge_width = req(); 4301 if (merge_width > Compile::AliasIdxRaw) { 4302 PhiNode* phi = base_memory()->as_Phi(); 4303 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4304 if (phi->in(i) == this) { 4305 phase->is_IterGVN()->_worklist.push(phi); 4306 break; 4307 } 4308 } 4309 } 4310 } 4311 4312 assert(progress || verify_sparse(), "please, no dups of base"); 4313 return progress; 4314 } 4315 4316 //-------------------------set_base_memory------------------------------------- 4317 void MergeMemNode::set_base_memory(Node *new_base) { 4318 Node* empty_mem = empty_memory(); 4319 set_req(Compile::AliasIdxBot, new_base); 4320 assert(memory_at(req()) == new_base, "must set default memory"); 4321 // Clear out other occurrences of new_base: 4322 if (new_base != empty_mem) { 4323 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4324 if (in(i) == new_base) set_req(i, empty_mem); 4325 } 4326 } 4327 } 4328 4329 //------------------------------out_RegMask------------------------------------ 4330 const RegMask &MergeMemNode::out_RegMask() const { 4331 return RegMask::Empty; 4332 } 4333 4334 //------------------------------dump_spec-------------------------------------- 4335 #ifndef PRODUCT 4336 void MergeMemNode::dump_spec(outputStream *st) const { 4337 st->print(" {"); 4338 Node* base_mem = base_memory(); 4339 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4340 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4341 if (mem == base_mem) { st->print(" -"); continue; } 4342 st->print( " N%d:", mem->_idx ); 4343 Compile::current()->get_adr_type(i)->dump_on(st); 4344 } 4345 st->print(" }"); 4346 } 4347 #endif // !PRODUCT 4348 4349 4350 #ifdef ASSERT 4351 static bool might_be_same(Node* a, Node* b) { 4352 if (a == b) return true; 4353 if (!(a->is_Phi() || b->is_Phi())) return false; 4354 // phis shift around during optimization 4355 return true; // pretty stupid... 4356 } 4357 4358 // verify a narrow slice (either incoming or outgoing) 4359 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4360 if (!VerifyAliases) return; // don't bother to verify unless requested 4361 if (is_error_reported()) return; // muzzle asserts when debugging an error 4362 if (Node::in_dump()) return; // muzzle asserts when printing 4363 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4364 assert(n != NULL, ""); 4365 // Elide intervening MergeMem's 4366 while (n->is_MergeMem()) { 4367 n = n->as_MergeMem()->memory_at(alias_idx); 4368 } 4369 Compile* C = Compile::current(); 4370 const TypePtr* n_adr_type = n->adr_type(); 4371 if (n == m->empty_memory()) { 4372 // Implicit copy of base_memory() 4373 } else if (n_adr_type != TypePtr::BOTTOM) { 4374 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4375 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4376 } else { 4377 // A few places like make_runtime_call "know" that VM calls are narrow, 4378 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4379 bool expected_wide_mem = false; 4380 if (n == m->base_memory()) { 4381 expected_wide_mem = true; 4382 } else if (alias_idx == Compile::AliasIdxRaw || 4383 n == m->memory_at(Compile::AliasIdxRaw)) { 4384 expected_wide_mem = true; 4385 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4386 // memory can "leak through" calls on channels that 4387 // are write-once. Allow this also. 4388 expected_wide_mem = true; 4389 } 4390 assert(expected_wide_mem, "expected narrow slice replacement"); 4391 } 4392 } 4393 #else // !ASSERT 4394 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4395 #endif 4396 4397 4398 //-----------------------------memory_at--------------------------------------- 4399 Node* MergeMemNode::memory_at(uint alias_idx) const { 4400 assert(alias_idx >= Compile::AliasIdxRaw || 4401 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4402 "must avoid base_memory and AliasIdxTop"); 4403 4404 // Otherwise, it is a narrow slice. 4405 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4406 Compile *C = Compile::current(); 4407 if (is_empty_memory(n)) { 4408 // the array is sparse; empty slots are the "top" node 4409 n = base_memory(); 4410 assert(Node::in_dump() 4411 || n == NULL || n->bottom_type() == Type::TOP 4412 || n->adr_type() == NULL // address is TOP 4413 || n->adr_type() == TypePtr::BOTTOM 4414 || n->adr_type() == TypeRawPtr::BOTTOM 4415 || Compile::current()->AliasLevel() == 0, 4416 "must be a wide memory"); 4417 // AliasLevel == 0 if we are organizing the memory states manually. 4418 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4419 } else { 4420 // make sure the stored slice is sane 4421 #ifdef ASSERT 4422 if (is_error_reported() || Node::in_dump()) { 4423 } else if (might_be_same(n, base_memory())) { 4424 // Give it a pass: It is a mostly harmless repetition of the base. 4425 // This can arise normally from node subsumption during optimization. 4426 } else { 4427 verify_memory_slice(this, alias_idx, n); 4428 } 4429 #endif 4430 } 4431 return n; 4432 } 4433 4434 //---------------------------set_memory_at------------------------------------- 4435 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4436 verify_memory_slice(this, alias_idx, n); 4437 Node* empty_mem = empty_memory(); 4438 if (n == base_memory()) n = empty_mem; // collapse default 4439 uint need_req = alias_idx+1; 4440 if (req() < need_req) { 4441 if (n == empty_mem) return; // already the default, so do not grow me 4442 // grow the sparse array 4443 do { 4444 add_req(empty_mem); 4445 } while (req() < need_req); 4446 } 4447 set_req( alias_idx, n ); 4448 } 4449 4450 4451 4452 //--------------------------iteration_setup------------------------------------ 4453 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4454 if (other != NULL) { 4455 grow_to_match(other); 4456 // invariant: the finite support of mm2 is within mm->req() 4457 #ifdef ASSERT 4458 for (uint i = req(); i < other->req(); i++) { 4459 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4460 } 4461 #endif 4462 } 4463 // Replace spurious copies of base_memory by top. 4464 Node* base_mem = base_memory(); 4465 if (base_mem != NULL && !base_mem->is_top()) { 4466 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4467 if (in(i) == base_mem) 4468 set_req(i, empty_memory()); 4469 } 4470 } 4471 } 4472 4473 //---------------------------grow_to_match------------------------------------- 4474 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4475 Node* empty_mem = empty_memory(); 4476 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4477 // look for the finite support of the other memory 4478 for (uint i = other->req(); --i >= req(); ) { 4479 if (other->in(i) != empty_mem) { 4480 uint new_len = i+1; 4481 while (req() < new_len) add_req(empty_mem); 4482 break; 4483 } 4484 } 4485 } 4486 4487 //---------------------------verify_sparse------------------------------------- 4488 #ifndef PRODUCT 4489 bool MergeMemNode::verify_sparse() const { 4490 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4491 Node* base_mem = base_memory(); 4492 // The following can happen in degenerate cases, since empty==top. 4493 if (is_empty_memory(base_mem)) return true; 4494 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4495 assert(in(i) != NULL, "sane slice"); 4496 if (in(i) == base_mem) return false; // should have been the sentinel value! 4497 } 4498 return true; 4499 } 4500 4501 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4502 Node* n; 4503 n = mm->in(idx); 4504 if (mem == n) return true; // might be empty_memory() 4505 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4506 if (mem == n) return true; 4507 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4508 if (mem == n) return true; 4509 if (n == NULL) break; 4510 } 4511 return false; 4512 } 4513 #endif // !PRODUCT