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, 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, 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 ciEnv* env = C->env(); 1717 const TypeInstPtr* tinst = tp->is_instptr(); 1718 ciKlass* klass = tinst->klass(); 1719 assert( off != Type::OffsetBot || 1720 // arrays can be cast to Objects 1721 tp->is_oopptr()->klass()->is_java_lang_Object() || 1722 // unsafe field access may not have a constant offset 1723 C->has_unsafe_access(), 1724 "Field accesses must be precise" ); 1725 // For oop loads, we expect the _type to be precise. 1726 // Optimizations for constant objects 1727 ciObject* const_oop = tinst->const_oop(); 1728 if (const_oop != NULL && const_oop->is_instance()) { 1729 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type()); 1730 if (con_type != NULL) { 1731 return con_type; 1732 } 1733 } 1734 } else if (tp->base() == Type::KlassPtr) { 1735 assert( off != Type::OffsetBot || 1736 // arrays can be cast to Objects 1737 tp->is_klassptr()->klass()->is_java_lang_Object() || 1738 // also allow array-loading from the primary supertype 1739 // array during subtype checks 1740 Opcode() == Op_LoadKlass, 1741 "Field accesses must be precise" ); 1742 // For klass/static loads, we expect the _type to be precise 1743 } 1744 1745 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1746 if (tkls != NULL && !StressReflectiveCode) { 1747 ciKlass* klass = tkls->klass(); 1748 if (klass->is_loaded() && tkls->klass_is_exact()) { 1749 // We are loading a field from a Klass metaobject whose identity 1750 // is known at compile time (the type is "exact" or "precise"). 1751 // Check for fields we know are maintained as constants by the VM. 1752 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1753 // The field is Klass::_super_check_offset. Return its (constant) value. 1754 // (Folds up type checking code.) 1755 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1756 return TypeInt::make(klass->super_check_offset()); 1757 } 1758 // Compute index into primary_supers array 1759 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1760 // Check for overflowing; use unsigned compare to handle the negative case. 1761 if( depth < ciKlass::primary_super_limit() ) { 1762 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1763 // (Folds up type checking code.) 1764 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1765 ciKlass *ss = klass->super_of_depth(depth); 1766 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1767 } 1768 const Type* aift = load_array_final_field(tkls, klass); 1769 if (aift != NULL) return aift; 1770 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1771 // The field is Klass::_java_mirror. Return its (constant) value. 1772 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1773 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1774 return TypeInstPtr::make(klass->java_mirror()); 1775 } 1776 } 1777 1778 // We can still check if we are loading from the primary_supers array at a 1779 // shallow enough depth. Even though the klass is not exact, entries less 1780 // than or equal to its super depth are correct. 1781 if (klass->is_loaded() ) { 1782 ciType *inner = klass; 1783 while( inner->is_obj_array_klass() ) 1784 inner = inner->as_obj_array_klass()->base_element_type(); 1785 if( inner->is_instance_klass() && 1786 !inner->as_instance_klass()->flags().is_interface() ) { 1787 // Compute index into primary_supers array 1788 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1789 // Check for overflowing; use unsigned compare to handle the negative case. 1790 if( depth < ciKlass::primary_super_limit() && 1791 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1792 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1793 // (Folds up type checking code.) 1794 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1795 ciKlass *ss = klass->super_of_depth(depth); 1796 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1797 } 1798 } 1799 } 1800 1801 // If the type is enough to determine that the thing is not an array, 1802 // we can give the layout_helper a positive interval type. 1803 // This will help short-circuit some reflective code. 1804 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1805 && !klass->is_array_klass() // not directly typed as an array 1806 && !klass->is_interface() // specifically not Serializable & Cloneable 1807 && !klass->is_java_lang_Object() // not the supertype of all T[] 1808 ) { 1809 // Note: When interfaces are reliable, we can narrow the interface 1810 // test to (klass != Serializable && klass != Cloneable). 1811 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1812 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1813 // The key property of this type is that it folds up tests 1814 // for array-ness, since it proves that the layout_helper is positive. 1815 // Thus, a generic value like the basic object layout helper works fine. 1816 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1817 } 1818 } 1819 1820 // If we are loading from a freshly-allocated object, produce a zero, 1821 // if the load is provably beyond the header of the object. 1822 // (Also allow a variable load from a fresh array to produce zero.) 1823 const TypeOopPtr *tinst = tp->isa_oopptr(); 1824 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1825 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1826 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1827 Node* value = can_see_stored_value(mem,phase); 1828 if (value != NULL && value->is_Con()) { 1829 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1830 return value->bottom_type(); 1831 } 1832 } 1833 1834 if (is_instance) { 1835 // If we have an instance type and our memory input is the 1836 // programs's initial memory state, there is no matching store, 1837 // so just return a zero of the appropriate type 1838 Node *mem = in(MemNode::Memory); 1839 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1840 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1841 return Type::get_zero_type(_type->basic_type()); 1842 } 1843 } 1844 return _type; 1845 } 1846 1847 //------------------------------match_edge------------------------------------- 1848 // Do we Match on this edge index or not? Match only the address. 1849 uint LoadNode::match_edge(uint idx) const { 1850 return idx == MemNode::Address; 1851 } 1852 1853 //--------------------------LoadBNode::Ideal-------------------------------------- 1854 // 1855 // If the previous store is to the same address as this load, 1856 // and the value stored was larger than a byte, replace this load 1857 // with the value stored truncated to a byte. If no truncation is 1858 // needed, the replacement is done in LoadNode::Identity(). 1859 // 1860 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1861 Node* mem = in(MemNode::Memory); 1862 Node* value = can_see_stored_value(mem,phase); 1863 if( value && !phase->type(value)->higher_equal( _type ) ) { 1864 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1865 return new RShiftINode(result, phase->intcon(24)); 1866 } 1867 // Identity call will handle the case where truncation is not needed. 1868 return LoadNode::Ideal(phase, can_reshape); 1869 } 1870 1871 const Type* LoadBNode::Value(PhaseGVN* phase) const { 1872 Node* mem = in(MemNode::Memory); 1873 Node* value = can_see_stored_value(mem,phase); 1874 if (value != NULL && value->is_Con() && 1875 !value->bottom_type()->higher_equal(_type)) { 1876 // If the input to the store does not fit with the load's result type, 1877 // it must be truncated. We can't delay until Ideal call since 1878 // a singleton Value is needed for split_thru_phi optimization. 1879 int con = value->get_int(); 1880 return TypeInt::make((con << 24) >> 24); 1881 } 1882 return LoadNode::Value(phase); 1883 } 1884 1885 //--------------------------LoadUBNode::Ideal------------------------------------- 1886 // 1887 // If the previous store is to the same address as this load, 1888 // and the value stored was larger than a byte, replace this load 1889 // with the value stored truncated to a byte. If no truncation is 1890 // needed, the replacement is done in LoadNode::Identity(). 1891 // 1892 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1893 Node* mem = in(MemNode::Memory); 1894 Node* value = can_see_stored_value(mem, phase); 1895 if (value && !phase->type(value)->higher_equal(_type)) 1896 return new AndINode(value, phase->intcon(0xFF)); 1897 // Identity call will handle the case where truncation is not needed. 1898 return LoadNode::Ideal(phase, can_reshape); 1899 } 1900 1901 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 1902 Node* mem = in(MemNode::Memory); 1903 Node* value = can_see_stored_value(mem,phase); 1904 if (value != NULL && value->is_Con() && 1905 !value->bottom_type()->higher_equal(_type)) { 1906 // If the input to the store does not fit with the load's result type, 1907 // it must be truncated. We can't delay until Ideal call since 1908 // a singleton Value is needed for split_thru_phi optimization. 1909 int con = value->get_int(); 1910 return TypeInt::make(con & 0xFF); 1911 } 1912 return LoadNode::Value(phase); 1913 } 1914 1915 //--------------------------LoadUSNode::Ideal------------------------------------- 1916 // 1917 // If the previous store is to the same address as this load, 1918 // and the value stored was larger than a char, replace this load 1919 // with the value stored truncated to a char. If no truncation is 1920 // needed, the replacement is done in LoadNode::Identity(). 1921 // 1922 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1923 Node* mem = in(MemNode::Memory); 1924 Node* value = can_see_stored_value(mem,phase); 1925 if( value && !phase->type(value)->higher_equal( _type ) ) 1926 return new AndINode(value,phase->intcon(0xFFFF)); 1927 // Identity call will handle the case where truncation is not needed. 1928 return LoadNode::Ideal(phase, can_reshape); 1929 } 1930 1931 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 1932 Node* mem = in(MemNode::Memory); 1933 Node* value = can_see_stored_value(mem,phase); 1934 if (value != NULL && value->is_Con() && 1935 !value->bottom_type()->higher_equal(_type)) { 1936 // If the input to the store does not fit with the load's result type, 1937 // it must be truncated. We can't delay until Ideal call since 1938 // a singleton Value is needed for split_thru_phi optimization. 1939 int con = value->get_int(); 1940 return TypeInt::make(con & 0xFFFF); 1941 } 1942 return LoadNode::Value(phase); 1943 } 1944 1945 //--------------------------LoadSNode::Ideal-------------------------------------- 1946 // 1947 // If the previous store is to the same address as this load, 1948 // and the value stored was larger than a short, replace this load 1949 // with the value stored truncated to a short. If no truncation is 1950 // needed, the replacement is done in LoadNode::Identity(). 1951 // 1952 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1953 Node* mem = in(MemNode::Memory); 1954 Node* value = can_see_stored_value(mem,phase); 1955 if( value && !phase->type(value)->higher_equal( _type ) ) { 1956 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 1957 return new RShiftINode(result, phase->intcon(16)); 1958 } 1959 // Identity call will handle the case where truncation is not needed. 1960 return LoadNode::Ideal(phase, can_reshape); 1961 } 1962 1963 const Type* LoadSNode::Value(PhaseGVN* phase) const { 1964 Node* mem = in(MemNode::Memory); 1965 Node* value = can_see_stored_value(mem,phase); 1966 if (value != NULL && value->is_Con() && 1967 !value->bottom_type()->higher_equal(_type)) { 1968 // If the input to the store does not fit with the load's result type, 1969 // it must be truncated. We can't delay until Ideal call since 1970 // a singleton Value is needed for split_thru_phi optimization. 1971 int con = value->get_int(); 1972 return TypeInt::make((con << 16) >> 16); 1973 } 1974 return LoadNode::Value(phase); 1975 } 1976 1977 //============================================================================= 1978 //----------------------------LoadKlassNode::make------------------------------ 1979 // Polymorphic factory method: 1980 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 1981 // sanity check the alias category against the created node type 1982 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 1983 assert(adr_type != NULL, "expecting TypeKlassPtr"); 1984 #ifdef _LP64 1985 if (adr_type->is_ptr_to_narrowklass()) { 1986 assert(UseCompressedClassPointers, "no compressed klasses"); 1987 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 1988 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 1989 } 1990 #endif 1991 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1992 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 1993 } 1994 1995 //------------------------------Value------------------------------------------ 1996 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 1997 return klass_value_common(phase); 1998 } 1999 2000 // In most cases, LoadKlassNode does not have the control input set. If the control 2001 // input is set, it must not be removed (by LoadNode::Ideal()). 2002 bool LoadKlassNode::can_remove_control() const { 2003 return false; 2004 } 2005 2006 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2007 // Either input is TOP ==> the result is TOP 2008 const Type *t1 = phase->type( in(MemNode::Memory) ); 2009 if (t1 == Type::TOP) return Type::TOP; 2010 Node *adr = in(MemNode::Address); 2011 const Type *t2 = phase->type( adr ); 2012 if (t2 == Type::TOP) return Type::TOP; 2013 const TypePtr *tp = t2->is_ptr(); 2014 if (TypePtr::above_centerline(tp->ptr()) || 2015 tp->ptr() == TypePtr::Null) return Type::TOP; 2016 2017 // Return a more precise klass, if possible 2018 const TypeInstPtr *tinst = tp->isa_instptr(); 2019 if (tinst != NULL) { 2020 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2021 int offset = tinst->offset(); 2022 if (ik == phase->C->env()->Class_klass() 2023 && (offset == java_lang_Class::klass_offset_in_bytes() || 2024 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2025 // We are loading a special hidden field from a Class mirror object, 2026 // the field which points to the VM's Klass metaobject. 2027 ciType* t = tinst->java_mirror_type(); 2028 // java_mirror_type returns non-null for compile-time Class constants. 2029 if (t != NULL) { 2030 // constant oop => constant klass 2031 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2032 if (t->is_void()) { 2033 // We cannot create a void array. Since void is a primitive type return null 2034 // klass. Users of this result need to do a null check on the returned klass. 2035 return TypePtr::NULL_PTR; 2036 } 2037 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2038 } 2039 if (!t->is_klass()) { 2040 // a primitive Class (e.g., int.class) has NULL for a klass field 2041 return TypePtr::NULL_PTR; 2042 } 2043 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2044 return TypeKlassPtr::make(t->as_klass()); 2045 } 2046 // non-constant mirror, so we can't tell what's going on 2047 } 2048 if( !ik->is_loaded() ) 2049 return _type; // Bail out if not loaded 2050 if (offset == oopDesc::klass_offset_in_bytes()) { 2051 if (tinst->klass_is_exact()) { 2052 return TypeKlassPtr::make(ik); 2053 } 2054 // See if we can become precise: no subklasses and no interface 2055 // (Note: We need to support verified interfaces.) 2056 if (!ik->is_interface() && !ik->has_subklass()) { 2057 //assert(!UseExactTypes, "this code should be useless with exact types"); 2058 // Add a dependence; if any subclass added we need to recompile 2059 if (!ik->is_final()) { 2060 // %%% should use stronger assert_unique_concrete_subtype instead 2061 phase->C->dependencies()->assert_leaf_type(ik); 2062 } 2063 // Return precise klass 2064 return TypeKlassPtr::make(ik); 2065 } 2066 2067 // Return root of possible klass 2068 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2069 } 2070 } 2071 2072 // Check for loading klass from an array 2073 const TypeAryPtr *tary = tp->isa_aryptr(); 2074 if( tary != NULL ) { 2075 ciKlass *tary_klass = tary->klass(); 2076 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2077 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2078 if (tary->klass_is_exact()) { 2079 return TypeKlassPtr::make(tary_klass); 2080 } 2081 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2082 // If the klass is an object array, we defer the question to the 2083 // array component klass. 2084 if( ak->is_obj_array_klass() ) { 2085 assert( ak->is_loaded(), "" ); 2086 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2087 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2088 ciInstanceKlass* ik = base_k->as_instance_klass(); 2089 // See if we can become precise: no subklasses and no interface 2090 if (!ik->is_interface() && !ik->has_subklass()) { 2091 //assert(!UseExactTypes, "this code should be useless with exact types"); 2092 // Add a dependence; if any subclass added we need to recompile 2093 if (!ik->is_final()) { 2094 phase->C->dependencies()->assert_leaf_type(ik); 2095 } 2096 // Return precise array klass 2097 return TypeKlassPtr::make(ak); 2098 } 2099 } 2100 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2101 } else { // Found a type-array? 2102 //assert(!UseExactTypes, "this code should be useless with exact types"); 2103 assert( ak->is_type_array_klass(), "" ); 2104 return TypeKlassPtr::make(ak); // These are always precise 2105 } 2106 } 2107 } 2108 2109 // Check for loading klass from an array klass 2110 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2111 if (tkls != NULL && !StressReflectiveCode) { 2112 ciKlass* klass = tkls->klass(); 2113 if( !klass->is_loaded() ) 2114 return _type; // Bail out if not loaded 2115 if( klass->is_obj_array_klass() && 2116 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2117 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2118 // // Always returning precise element type is incorrect, 2119 // // e.g., element type could be object and array may contain strings 2120 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2121 2122 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2123 // according to the element type's subclassing. 2124 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2125 } 2126 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2127 tkls->offset() == in_bytes(Klass::super_offset())) { 2128 ciKlass* sup = klass->as_instance_klass()->super(); 2129 // The field is Klass::_super. Return its (constant) value. 2130 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2131 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2132 } 2133 } 2134 2135 // Bailout case 2136 return LoadNode::Value(phase); 2137 } 2138 2139 //------------------------------Identity--------------------------------------- 2140 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2141 // Also feed through the klass in Allocate(...klass...)._klass. 2142 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2143 return klass_identity_common(phase); 2144 } 2145 2146 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2147 Node* x = LoadNode::Identity(phase); 2148 if (x != this) return x; 2149 2150 // Take apart the address into an oop and and offset. 2151 // Return 'this' if we cannot. 2152 Node* adr = in(MemNode::Address); 2153 intptr_t offset = 0; 2154 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2155 if (base == NULL) return this; 2156 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2157 if (toop == NULL) return this; 2158 2159 // We can fetch the klass directly through an AllocateNode. 2160 // This works even if the klass is not constant (clone or newArray). 2161 if (offset == oopDesc::klass_offset_in_bytes()) { 2162 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2163 if (allocated_klass != NULL) { 2164 return allocated_klass; 2165 } 2166 } 2167 2168 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2169 // See inline_native_Class_query for occurrences of these patterns. 2170 // Java Example: x.getClass().isAssignableFrom(y) 2171 // 2172 // This improves reflective code, often making the Class 2173 // mirror go completely dead. (Current exception: Class 2174 // mirrors may appear in debug info, but we could clean them out by 2175 // introducing a new debug info operator for Klass*.java_mirror). 2176 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2177 && offset == java_lang_Class::klass_offset_in_bytes()) { 2178 // We are loading a special hidden field from a Class mirror, 2179 // the field which points to its Klass or ArrayKlass metaobject. 2180 if (base->is_Load()) { 2181 Node* adr2 = base->in(MemNode::Address); 2182 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2183 if (tkls != NULL && !tkls->empty() 2184 && (tkls->klass()->is_instance_klass() || 2185 tkls->klass()->is_array_klass()) 2186 && adr2->is_AddP() 2187 ) { 2188 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2189 if (tkls->offset() == mirror_field) { 2190 return adr2->in(AddPNode::Base); 2191 } 2192 } 2193 } 2194 } 2195 2196 return this; 2197 } 2198 2199 2200 //------------------------------Value------------------------------------------ 2201 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2202 const Type *t = klass_value_common(phase); 2203 if (t == Type::TOP) 2204 return t; 2205 2206 return t->make_narrowklass(); 2207 } 2208 2209 //------------------------------Identity--------------------------------------- 2210 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2211 // Also feed through the klass in Allocate(...klass...)._klass. 2212 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2213 Node *x = klass_identity_common(phase); 2214 2215 const Type *t = phase->type( x ); 2216 if( t == Type::TOP ) return x; 2217 if( t->isa_narrowklass()) return x; 2218 assert (!t->isa_narrowoop(), "no narrow oop here"); 2219 2220 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2221 } 2222 2223 //------------------------------Value----------------------------------------- 2224 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2225 // Either input is TOP ==> the result is TOP 2226 const Type *t1 = phase->type( in(MemNode::Memory) ); 2227 if( t1 == Type::TOP ) return Type::TOP; 2228 Node *adr = in(MemNode::Address); 2229 const Type *t2 = phase->type( adr ); 2230 if( t2 == Type::TOP ) return Type::TOP; 2231 const TypePtr *tp = t2->is_ptr(); 2232 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2233 const TypeAryPtr *tap = tp->isa_aryptr(); 2234 if( !tap ) return _type; 2235 return tap->size(); 2236 } 2237 2238 //-------------------------------Ideal--------------------------------------- 2239 // Feed through the length in AllocateArray(...length...)._length. 2240 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2241 Node* p = MemNode::Ideal_common(phase, can_reshape); 2242 if (p) return (p == NodeSentinel) ? NULL : p; 2243 2244 // Take apart the address into an oop and and offset. 2245 // Return 'this' if we cannot. 2246 Node* adr = in(MemNode::Address); 2247 intptr_t offset = 0; 2248 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2249 if (base == NULL) return NULL; 2250 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2251 if (tary == NULL) return NULL; 2252 2253 // We can fetch the length directly through an AllocateArrayNode. 2254 // This works even if the length is not constant (clone or newArray). 2255 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2256 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2257 if (alloc != NULL) { 2258 Node* allocated_length = alloc->Ideal_length(); 2259 Node* len = alloc->make_ideal_length(tary, phase); 2260 if (allocated_length != len) { 2261 // New CastII improves on this. 2262 return len; 2263 } 2264 } 2265 } 2266 2267 return NULL; 2268 } 2269 2270 //------------------------------Identity--------------------------------------- 2271 // Feed through the length in AllocateArray(...length...)._length. 2272 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2273 Node* x = LoadINode::Identity(phase); 2274 if (x != this) return x; 2275 2276 // Take apart the address into an oop and and offset. 2277 // Return 'this' if we cannot. 2278 Node* adr = in(MemNode::Address); 2279 intptr_t offset = 0; 2280 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2281 if (base == NULL) return this; 2282 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2283 if (tary == NULL) return this; 2284 2285 // We can fetch the length directly through an AllocateArrayNode. 2286 // This works even if the length is not constant (clone or newArray). 2287 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2288 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2289 if (alloc != NULL) { 2290 Node* allocated_length = alloc->Ideal_length(); 2291 // Do not allow make_ideal_length to allocate a CastII node. 2292 Node* len = alloc->make_ideal_length(tary, phase, false); 2293 if (allocated_length == len) { 2294 // Return allocated_length only if it would not be improved by a CastII. 2295 return allocated_length; 2296 } 2297 } 2298 } 2299 2300 return this; 2301 2302 } 2303 2304 //============================================================================= 2305 //---------------------------StoreNode::make----------------------------------- 2306 // Polymorphic factory method: 2307 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2308 assert((mo == unordered || mo == release), "unexpected"); 2309 Compile* C = gvn.C; 2310 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2311 ctl != NULL, "raw memory operations should have control edge"); 2312 2313 switch (bt) { 2314 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2315 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2316 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2317 case T_CHAR: 2318 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2319 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2320 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2321 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2322 case T_METADATA: 2323 case T_ADDRESS: 2324 case T_OBJECT: 2325 #ifdef _LP64 2326 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2327 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2328 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2329 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2330 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2331 adr->bottom_type()->isa_rawptr())) { 2332 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2333 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2334 } 2335 #endif 2336 { 2337 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2338 } 2339 } 2340 ShouldNotReachHere(); 2341 return (StoreNode*)NULL; 2342 } 2343 2344 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2345 bool require_atomic = true; 2346 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2347 } 2348 2349 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2350 bool require_atomic = true; 2351 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2352 } 2353 2354 2355 //--------------------------bottom_type---------------------------------------- 2356 const Type *StoreNode::bottom_type() const { 2357 return Type::MEMORY; 2358 } 2359 2360 //------------------------------hash------------------------------------------- 2361 uint StoreNode::hash() const { 2362 // unroll addition of interesting fields 2363 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2364 2365 // Since they are not commoned, do not hash them: 2366 return NO_HASH; 2367 } 2368 2369 //------------------------------Ideal------------------------------------------ 2370 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2371 // When a store immediately follows a relevant allocation/initialization, 2372 // try to capture it into the initialization, or hoist it above. 2373 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2374 Node* p = MemNode::Ideal_common(phase, can_reshape); 2375 if (p) return (p == NodeSentinel) ? NULL : p; 2376 2377 Node* mem = in(MemNode::Memory); 2378 Node* address = in(MemNode::Address); 2379 // Back-to-back stores to same address? Fold em up. Generally 2380 // unsafe if I have intervening uses... Also disallowed for StoreCM 2381 // since they must follow each StoreP operation. Redundant StoreCMs 2382 // are eliminated just before matching in final_graph_reshape. 2383 { 2384 Node* st = mem; 2385 // If Store 'st' has more than one use, we cannot fold 'st' away. 2386 // For example, 'st' might be the final state at a conditional 2387 // return. Or, 'st' might be used by some node which is live at 2388 // the same time 'st' is live, which might be unschedulable. So, 2389 // require exactly ONE user until such time as we clone 'mem' for 2390 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2391 // true). 2392 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2393 // Looking at a dead closed cycle of memory? 2394 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2395 assert(Opcode() == st->Opcode() || 2396 st->Opcode() == Op_StoreVector || 2397 Opcode() == Op_StoreVector || 2398 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2399 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2400 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2401 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2402 2403 if (st->in(MemNode::Address)->eqv_uncast(address) && 2404 st->as_Store()->memory_size() <= this->memory_size()) { 2405 Node* use = st->raw_out(0); 2406 phase->igvn_rehash_node_delayed(use); 2407 if (can_reshape) { 2408 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2409 } else { 2410 // It's OK to do this in the parser, since DU info is always accurate, 2411 // and the parser always refers to nodes via SafePointNode maps. 2412 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2413 } 2414 return this; 2415 } 2416 st = st->in(MemNode::Memory); 2417 } 2418 } 2419 2420 2421 // Capture an unaliased, unconditional, simple store into an initializer. 2422 // Or, if it is independent of the allocation, hoist it above the allocation. 2423 if (ReduceFieldZeroing && /*can_reshape &&*/ 2424 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2425 InitializeNode* init = mem->in(0)->as_Initialize(); 2426 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2427 if (offset > 0) { 2428 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2429 // If the InitializeNode captured me, it made a raw copy of me, 2430 // and I need to disappear. 2431 if (moved != NULL) { 2432 // %%% hack to ensure that Ideal returns a new node: 2433 mem = MergeMemNode::make(mem); 2434 return mem; // fold me away 2435 } 2436 } 2437 } 2438 2439 return NULL; // No further progress 2440 } 2441 2442 //------------------------------Value----------------------------------------- 2443 const Type* StoreNode::Value(PhaseGVN* phase) const { 2444 // Either input is TOP ==> the result is TOP 2445 const Type *t1 = phase->type( in(MemNode::Memory) ); 2446 if( t1 == Type::TOP ) return Type::TOP; 2447 const Type *t2 = phase->type( in(MemNode::Address) ); 2448 if( t2 == Type::TOP ) return Type::TOP; 2449 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2450 if( t3 == Type::TOP ) return Type::TOP; 2451 return Type::MEMORY; 2452 } 2453 2454 //------------------------------Identity--------------------------------------- 2455 // Remove redundant stores: 2456 // Store(m, p, Load(m, p)) changes to m. 2457 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2458 Node* StoreNode::Identity(PhaseGVN* phase) { 2459 Node* mem = in(MemNode::Memory); 2460 Node* adr = in(MemNode::Address); 2461 Node* val = in(MemNode::ValueIn); 2462 2463 // Load then Store? Then the Store is useless 2464 if (val->is_Load() && 2465 val->in(MemNode::Address)->eqv_uncast(adr) && 2466 val->in(MemNode::Memory )->eqv_uncast(mem) && 2467 val->as_Load()->store_Opcode() == Opcode()) { 2468 return mem; 2469 } 2470 2471 // Two stores in a row of the same value? 2472 if (mem->is_Store() && 2473 mem->in(MemNode::Address)->eqv_uncast(adr) && 2474 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2475 mem->Opcode() == Opcode()) { 2476 return mem; 2477 } 2478 2479 // Store of zero anywhere into a freshly-allocated object? 2480 // Then the store is useless. 2481 // (It must already have been captured by the InitializeNode.) 2482 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2483 // a newly allocated object is already all-zeroes everywhere 2484 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2485 return mem; 2486 } 2487 2488 // the store may also apply to zero-bits in an earlier object 2489 Node* prev_mem = find_previous_store(phase); 2490 // Steps (a), (b): Walk past independent stores to find an exact match. 2491 if (prev_mem != NULL) { 2492 Node* prev_val = can_see_stored_value(prev_mem, phase); 2493 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2494 // prev_val and val might differ by a cast; it would be good 2495 // to keep the more informative of the two. 2496 return mem; 2497 } 2498 } 2499 } 2500 2501 return this; 2502 } 2503 2504 //------------------------------match_edge------------------------------------- 2505 // Do we Match on this edge index or not? Match only memory & value 2506 uint StoreNode::match_edge(uint idx) const { 2507 return idx == MemNode::Address || idx == MemNode::ValueIn; 2508 } 2509 2510 //------------------------------cmp-------------------------------------------- 2511 // Do not common stores up together. They generally have to be split 2512 // back up anyways, so do not bother. 2513 uint StoreNode::cmp( const Node &n ) const { 2514 return (&n == this); // Always fail except on self 2515 } 2516 2517 //------------------------------Ideal_masked_input----------------------------- 2518 // Check for a useless mask before a partial-word store 2519 // (StoreB ... (AndI valIn conIa) ) 2520 // If (conIa & mask == mask) this simplifies to 2521 // (StoreB ... (valIn) ) 2522 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2523 Node *val = in(MemNode::ValueIn); 2524 if( val->Opcode() == Op_AndI ) { 2525 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2526 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2527 set_req(MemNode::ValueIn, val->in(1)); 2528 return this; 2529 } 2530 } 2531 return NULL; 2532 } 2533 2534 2535 //------------------------------Ideal_sign_extended_input---------------------- 2536 // Check for useless sign-extension before a partial-word store 2537 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2538 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2539 // (StoreB ... (valIn) ) 2540 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2541 Node *val = in(MemNode::ValueIn); 2542 if( val->Opcode() == Op_RShiftI ) { 2543 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2544 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2545 Node *shl = val->in(1); 2546 if( shl->Opcode() == Op_LShiftI ) { 2547 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2548 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2549 set_req(MemNode::ValueIn, shl->in(1)); 2550 return this; 2551 } 2552 } 2553 } 2554 } 2555 return NULL; 2556 } 2557 2558 //------------------------------value_never_loaded----------------------------------- 2559 // Determine whether there are any possible loads of the value stored. 2560 // For simplicity, we actually check if there are any loads from the 2561 // address stored to, not just for loads of the value stored by this node. 2562 // 2563 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2564 Node *adr = in(Address); 2565 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2566 if (adr_oop == NULL) 2567 return false; 2568 if (!adr_oop->is_known_instance_field()) 2569 return false; // if not a distinct instance, there may be aliases of the address 2570 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2571 Node *use = adr->fast_out(i); 2572 if (use->is_Load() || use->is_LoadStore()) { 2573 return false; 2574 } 2575 } 2576 return true; 2577 } 2578 2579 //============================================================================= 2580 //------------------------------Ideal------------------------------------------ 2581 // If the store is from an AND mask that leaves the low bits untouched, then 2582 // we can skip the AND operation. If the store is from a sign-extension 2583 // (a left shift, then right shift) we can skip both. 2584 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2585 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2586 if( progress != NULL ) return progress; 2587 2588 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2589 if( progress != NULL ) return progress; 2590 2591 // Finally check the default case 2592 return StoreNode::Ideal(phase, can_reshape); 2593 } 2594 2595 //============================================================================= 2596 //------------------------------Ideal------------------------------------------ 2597 // If the store is from an AND mask that leaves the low bits untouched, then 2598 // we can skip the AND operation 2599 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2600 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2601 if( progress != NULL ) return progress; 2602 2603 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2604 if( progress != NULL ) return progress; 2605 2606 // Finally check the default case 2607 return StoreNode::Ideal(phase, can_reshape); 2608 } 2609 2610 //============================================================================= 2611 //------------------------------Identity--------------------------------------- 2612 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2613 // No need to card mark when storing a null ptr 2614 Node* my_store = in(MemNode::OopStore); 2615 if (my_store->is_Store()) { 2616 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2617 if( t1 == TypePtr::NULL_PTR ) { 2618 return in(MemNode::Memory); 2619 } 2620 } 2621 return this; 2622 } 2623 2624 //============================================================================= 2625 //------------------------------Ideal--------------------------------------- 2626 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2627 Node* progress = StoreNode::Ideal(phase, can_reshape); 2628 if (progress != NULL) return progress; 2629 2630 Node* my_store = in(MemNode::OopStore); 2631 if (my_store->is_MergeMem()) { 2632 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2633 set_req(MemNode::OopStore, mem); 2634 return this; 2635 } 2636 2637 return NULL; 2638 } 2639 2640 //------------------------------Value----------------------------------------- 2641 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2642 // Either input is TOP ==> the result is TOP 2643 const Type *t = phase->type( in(MemNode::Memory) ); 2644 if( t == Type::TOP ) return Type::TOP; 2645 t = phase->type( in(MemNode::Address) ); 2646 if( t == Type::TOP ) return Type::TOP; 2647 t = phase->type( in(MemNode::ValueIn) ); 2648 if( t == Type::TOP ) return Type::TOP; 2649 // If extra input is TOP ==> the result is TOP 2650 t = phase->type( in(MemNode::OopStore) ); 2651 if( t == Type::TOP ) return Type::TOP; 2652 2653 return StoreNode::Value( phase ); 2654 } 2655 2656 2657 //============================================================================= 2658 //----------------------------------SCMemProjNode------------------------------ 2659 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2660 { 2661 return bottom_type(); 2662 } 2663 2664 //============================================================================= 2665 //----------------------------------LoadStoreNode------------------------------ 2666 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2667 : Node(required), 2668 _type(rt), 2669 _adr_type(at) 2670 { 2671 init_req(MemNode::Control, c ); 2672 init_req(MemNode::Memory , mem); 2673 init_req(MemNode::Address, adr); 2674 init_req(MemNode::ValueIn, val); 2675 init_class_id(Class_LoadStore); 2676 } 2677 2678 uint LoadStoreNode::ideal_reg() const { 2679 return _type->ideal_reg(); 2680 } 2681 2682 bool LoadStoreNode::result_not_used() const { 2683 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2684 Node *x = fast_out(i); 2685 if (x->Opcode() == Op_SCMemProj) continue; 2686 return false; 2687 } 2688 return true; 2689 } 2690 2691 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2692 2693 //============================================================================= 2694 //----------------------------------LoadStoreConditionalNode-------------------- 2695 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2696 init_req(ExpectedIn, ex ); 2697 } 2698 2699 //============================================================================= 2700 //-------------------------------adr_type-------------------------------------- 2701 const TypePtr* ClearArrayNode::adr_type() const { 2702 Node *adr = in(3); 2703 if (adr == NULL) return NULL; // node is dead 2704 return MemNode::calculate_adr_type(adr->bottom_type()); 2705 } 2706 2707 //------------------------------match_edge------------------------------------- 2708 // Do we Match on this edge index or not? Do not match memory 2709 uint ClearArrayNode::match_edge(uint idx) const { 2710 return idx > 1; 2711 } 2712 2713 //------------------------------Identity--------------------------------------- 2714 // Clearing a zero length array does nothing 2715 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2716 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2717 } 2718 2719 //------------------------------Idealize--------------------------------------- 2720 // Clearing a short array is faster with stores 2721 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2722 // Already know this is a large node, do not try to ideal it 2723 if (_is_large) return NULL; 2724 2725 const int unit = BytesPerLong; 2726 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2727 if (!t) return NULL; 2728 if (!t->is_con()) return NULL; 2729 intptr_t raw_count = t->get_con(); 2730 intptr_t size = raw_count; 2731 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2732 // Clearing nothing uses the Identity call. 2733 // Negative clears are possible on dead ClearArrays 2734 // (see jck test stmt114.stmt11402.val). 2735 if (size <= 0 || size % unit != 0) return NULL; 2736 intptr_t count = size / unit; 2737 // Length too long; communicate this to matchers and assemblers. 2738 // Assemblers are responsible to produce fast hardware clears for it. 2739 if (size > InitArrayShortSize) { 2740 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 2741 } 2742 Node *mem = in(1); 2743 if( phase->type(mem)==Type::TOP ) return NULL; 2744 Node *adr = in(3); 2745 const Type* at = phase->type(adr); 2746 if( at==Type::TOP ) return NULL; 2747 const TypePtr* atp = at->isa_ptr(); 2748 // adjust atp to be the correct array element address type 2749 if (atp == NULL) atp = TypePtr::BOTTOM; 2750 else atp = atp->add_offset(Type::OffsetBot); 2751 // Get base for derived pointer purposes 2752 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2753 Node *base = adr->in(1); 2754 2755 Node *zero = phase->makecon(TypeLong::ZERO); 2756 Node *off = phase->MakeConX(BytesPerLong); 2757 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2758 count--; 2759 while( count-- ) { 2760 mem = phase->transform(mem); 2761 adr = phase->transform(new AddPNode(base,adr,off)); 2762 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2763 } 2764 return mem; 2765 } 2766 2767 //----------------------------step_through---------------------------------- 2768 // Return allocation input memory edge if it is different instance 2769 // or itself if it is the one we are looking for. 2770 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2771 Node* n = *np; 2772 assert(n->is_ClearArray(), "sanity"); 2773 intptr_t offset; 2774 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2775 // This method is called only before Allocate nodes are expanded 2776 // during macro nodes expansion. Before that ClearArray nodes are 2777 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2778 // Allocate nodes are expanded) which follows allocations. 2779 assert(alloc != NULL, "should have allocation"); 2780 if (alloc->_idx == instance_id) { 2781 // Can not bypass initialization of the instance we are looking for. 2782 return false; 2783 } 2784 // Otherwise skip it. 2785 InitializeNode* init = alloc->initialization(); 2786 if (init != NULL) 2787 *np = init->in(TypeFunc::Memory); 2788 else 2789 *np = alloc->in(TypeFunc::Memory); 2790 return true; 2791 } 2792 2793 //----------------------------clear_memory------------------------------------- 2794 // Generate code to initialize object storage to zero. 2795 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2796 intptr_t start_offset, 2797 Node* end_offset, 2798 PhaseGVN* phase) { 2799 intptr_t offset = start_offset; 2800 2801 int unit = BytesPerLong; 2802 if ((offset % unit) != 0) { 2803 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2804 adr = phase->transform(adr); 2805 const TypePtr* atp = TypeRawPtr::BOTTOM; 2806 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2807 mem = phase->transform(mem); 2808 offset += BytesPerInt; 2809 } 2810 assert((offset % unit) == 0, ""); 2811 2812 // Initialize the remaining stuff, if any, with a ClearArray. 2813 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2814 } 2815 2816 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2817 Node* start_offset, 2818 Node* end_offset, 2819 PhaseGVN* phase) { 2820 if (start_offset == end_offset) { 2821 // nothing to do 2822 return mem; 2823 } 2824 2825 int unit = BytesPerLong; 2826 Node* zbase = start_offset; 2827 Node* zend = end_offset; 2828 2829 // Scale to the unit required by the CPU: 2830 if (!Matcher::init_array_count_is_in_bytes) { 2831 Node* shift = phase->intcon(exact_log2(unit)); 2832 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2833 zend = phase->transform(new URShiftXNode(zend, shift) ); 2834 } 2835 2836 // Bulk clear double-words 2837 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2838 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2839 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 2840 return phase->transform(mem); 2841 } 2842 2843 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2844 intptr_t start_offset, 2845 intptr_t end_offset, 2846 PhaseGVN* phase) { 2847 if (start_offset == end_offset) { 2848 // nothing to do 2849 return mem; 2850 } 2851 2852 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2853 intptr_t done_offset = end_offset; 2854 if ((done_offset % BytesPerLong) != 0) { 2855 done_offset -= BytesPerInt; 2856 } 2857 if (done_offset > start_offset) { 2858 mem = clear_memory(ctl, mem, dest, 2859 start_offset, phase->MakeConX(done_offset), phase); 2860 } 2861 if (done_offset < end_offset) { // emit the final 32-bit store 2862 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 2863 adr = phase->transform(adr); 2864 const TypePtr* atp = TypeRawPtr::BOTTOM; 2865 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2866 mem = phase->transform(mem); 2867 done_offset += BytesPerInt; 2868 } 2869 assert(done_offset == end_offset, ""); 2870 return mem; 2871 } 2872 2873 //============================================================================= 2874 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2875 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2876 _adr_type(C->get_adr_type(alias_idx)) 2877 { 2878 init_class_id(Class_MemBar); 2879 Node* top = C->top(); 2880 init_req(TypeFunc::I_O,top); 2881 init_req(TypeFunc::FramePtr,top); 2882 init_req(TypeFunc::ReturnAdr,top); 2883 if (precedent != NULL) 2884 init_req(TypeFunc::Parms, precedent); 2885 } 2886 2887 //------------------------------cmp-------------------------------------------- 2888 uint MemBarNode::hash() const { return NO_HASH; } 2889 uint MemBarNode::cmp( const Node &n ) const { 2890 return (&n == this); // Always fail except on self 2891 } 2892 2893 //------------------------------make------------------------------------------- 2894 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2895 switch (opcode) { 2896 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 2897 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 2898 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 2899 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 2900 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 2901 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 2902 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 2903 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 2904 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 2905 case Op_Initialize: return new InitializeNode(C, atp, pn); 2906 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 2907 default: ShouldNotReachHere(); return NULL; 2908 } 2909 } 2910 2911 //------------------------------Ideal------------------------------------------ 2912 // Return a node which is more "ideal" than the current node. Strip out 2913 // control copies 2914 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2915 if (remove_dead_region(phase, can_reshape)) return this; 2916 // Don't bother trying to transform a dead node 2917 if (in(0) && in(0)->is_top()) { 2918 return NULL; 2919 } 2920 2921 bool progress = false; 2922 // Eliminate volatile MemBars for scalar replaced objects. 2923 if (can_reshape && req() == (Precedent+1)) { 2924 bool eliminate = false; 2925 int opc = Opcode(); 2926 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2927 // Volatile field loads and stores. 2928 Node* my_mem = in(MemBarNode::Precedent); 2929 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2930 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2931 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2932 // replace this Precedent (decodeN) with the Load instead. 2933 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2934 Node* load_node = my_mem->in(1); 2935 set_req(MemBarNode::Precedent, load_node); 2936 phase->is_IterGVN()->_worklist.push(my_mem); 2937 my_mem = load_node; 2938 } else { 2939 assert(my_mem->unique_out() == this, "sanity"); 2940 del_req(Precedent); 2941 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2942 my_mem = NULL; 2943 } 2944 progress = true; 2945 } 2946 if (my_mem != NULL && my_mem->is_Mem()) { 2947 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2948 // Check for scalar replaced object reference. 2949 if( t_oop != NULL && t_oop->is_known_instance_field() && 2950 t_oop->offset() != Type::OffsetBot && 2951 t_oop->offset() != Type::OffsetTop) { 2952 eliminate = true; 2953 } 2954 } 2955 } else if (opc == Op_MemBarRelease) { 2956 // Final field stores. 2957 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 2958 if ((alloc != NULL) && alloc->is_Allocate() && 2959 alloc->as_Allocate()->does_not_escape_thread()) { 2960 // The allocated object does not escape. 2961 eliminate = true; 2962 } 2963 } 2964 if (eliminate) { 2965 // Replace MemBar projections by its inputs. 2966 PhaseIterGVN* igvn = phase->is_IterGVN(); 2967 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2968 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2969 // Must return either the original node (now dead) or a new node 2970 // (Do not return a top here, since that would break the uniqueness of top.) 2971 return new ConINode(TypeInt::ZERO); 2972 } 2973 } 2974 return progress ? this : NULL; 2975 } 2976 2977 //------------------------------Value------------------------------------------ 2978 const Type* MemBarNode::Value(PhaseGVN* phase) const { 2979 if( !in(0) ) return Type::TOP; 2980 if( phase->type(in(0)) == Type::TOP ) 2981 return Type::TOP; 2982 return TypeTuple::MEMBAR; 2983 } 2984 2985 //------------------------------match------------------------------------------ 2986 // Construct projections for memory. 2987 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2988 switch (proj->_con) { 2989 case TypeFunc::Control: 2990 case TypeFunc::Memory: 2991 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2992 } 2993 ShouldNotReachHere(); 2994 return NULL; 2995 } 2996 2997 //===========================InitializeNode==================================== 2998 // SUMMARY: 2999 // This node acts as a memory barrier on raw memory, after some raw stores. 3000 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3001 // The Initialize can 'capture' suitably constrained stores as raw inits. 3002 // It can coalesce related raw stores into larger units (called 'tiles'). 3003 // It can avoid zeroing new storage for memory units which have raw inits. 3004 // At macro-expansion, it is marked 'complete', and does not optimize further. 3005 // 3006 // EXAMPLE: 3007 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3008 // ctl = incoming control; mem* = incoming memory 3009 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3010 // First allocate uninitialized memory and fill in the header: 3011 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3012 // ctl := alloc.Control; mem* := alloc.Memory* 3013 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3014 // Then initialize to zero the non-header parts of the raw memory block: 3015 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3016 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3017 // After the initialize node executes, the object is ready for service: 3018 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3019 // Suppose its body is immediately initialized as {1,2}: 3020 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3021 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3022 // mem.SLICE(#short[*]) := store2 3023 // 3024 // DETAILS: 3025 // An InitializeNode collects and isolates object initialization after 3026 // an AllocateNode and before the next possible safepoint. As a 3027 // memory barrier (MemBarNode), it keeps critical stores from drifting 3028 // down past any safepoint or any publication of the allocation. 3029 // Before this barrier, a newly-allocated object may have uninitialized bits. 3030 // After this barrier, it may be treated as a real oop, and GC is allowed. 3031 // 3032 // The semantics of the InitializeNode include an implicit zeroing of 3033 // the new object from object header to the end of the object. 3034 // (The object header and end are determined by the AllocateNode.) 3035 // 3036 // Certain stores may be added as direct inputs to the InitializeNode. 3037 // These stores must update raw memory, and they must be to addresses 3038 // derived from the raw address produced by AllocateNode, and with 3039 // a constant offset. They must be ordered by increasing offset. 3040 // The first one is at in(RawStores), the last at in(req()-1). 3041 // Unlike most memory operations, they are not linked in a chain, 3042 // but are displayed in parallel as users of the rawmem output of 3043 // the allocation. 3044 // 3045 // (See comments in InitializeNode::capture_store, which continue 3046 // the example given above.) 3047 // 3048 // When the associated Allocate is macro-expanded, the InitializeNode 3049 // may be rewritten to optimize collected stores. A ClearArrayNode 3050 // may also be created at that point to represent any required zeroing. 3051 // The InitializeNode is then marked 'complete', prohibiting further 3052 // capturing of nearby memory operations. 3053 // 3054 // During macro-expansion, all captured initializations which store 3055 // constant values of 32 bits or smaller are coalesced (if advantageous) 3056 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3057 // initialized in fewer memory operations. Memory words which are 3058 // covered by neither tiles nor non-constant stores are pre-zeroed 3059 // by explicit stores of zero. (The code shape happens to do all 3060 // zeroing first, then all other stores, with both sequences occurring 3061 // in order of ascending offsets.) 3062 // 3063 // Alternatively, code may be inserted between an AllocateNode and its 3064 // InitializeNode, to perform arbitrary initialization of the new object. 3065 // E.g., the object copying intrinsics insert complex data transfers here. 3066 // The initialization must then be marked as 'complete' disable the 3067 // built-in zeroing semantics and the collection of initializing stores. 3068 // 3069 // While an InitializeNode is incomplete, reads from the memory state 3070 // produced by it are optimizable if they match the control edge and 3071 // new oop address associated with the allocation/initialization. 3072 // They return a stored value (if the offset matches) or else zero. 3073 // A write to the memory state, if it matches control and address, 3074 // and if it is to a constant offset, may be 'captured' by the 3075 // InitializeNode. It is cloned as a raw memory operation and rewired 3076 // inside the initialization, to the raw oop produced by the allocation. 3077 // Operations on addresses which are provably distinct (e.g., to 3078 // other AllocateNodes) are allowed to bypass the initialization. 3079 // 3080 // The effect of all this is to consolidate object initialization 3081 // (both arrays and non-arrays, both piecewise and bulk) into a 3082 // single location, where it can be optimized as a unit. 3083 // 3084 // Only stores with an offset less than TrackedInitializationLimit words 3085 // will be considered for capture by an InitializeNode. This puts a 3086 // reasonable limit on the complexity of optimized initializations. 3087 3088 //---------------------------InitializeNode------------------------------------ 3089 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3090 : _is_complete(Incomplete), _does_not_escape(false), 3091 MemBarNode(C, adr_type, rawoop) 3092 { 3093 init_class_id(Class_Initialize); 3094 3095 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3096 assert(in(RawAddress) == rawoop, "proper init"); 3097 // Note: allocation() can be NULL, for secondary initialization barriers 3098 } 3099 3100 // Since this node is not matched, it will be processed by the 3101 // register allocator. Declare that there are no constraints 3102 // on the allocation of the RawAddress edge. 3103 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3104 // This edge should be set to top, by the set_complete. But be conservative. 3105 if (idx == InitializeNode::RawAddress) 3106 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3107 return RegMask::Empty; 3108 } 3109 3110 Node* InitializeNode::memory(uint alias_idx) { 3111 Node* mem = in(Memory); 3112 if (mem->is_MergeMem()) { 3113 return mem->as_MergeMem()->memory_at(alias_idx); 3114 } else { 3115 // incoming raw memory is not split 3116 return mem; 3117 } 3118 } 3119 3120 bool InitializeNode::is_non_zero() { 3121 if (is_complete()) return false; 3122 remove_extra_zeroes(); 3123 return (req() > RawStores); 3124 } 3125 3126 void InitializeNode::set_complete(PhaseGVN* phase) { 3127 assert(!is_complete(), "caller responsibility"); 3128 _is_complete = Complete; 3129 3130 // After this node is complete, it contains a bunch of 3131 // raw-memory initializations. There is no need for 3132 // it to have anything to do with non-raw memory effects. 3133 // Therefore, tell all non-raw users to re-optimize themselves, 3134 // after skipping the memory effects of this initialization. 3135 PhaseIterGVN* igvn = phase->is_IterGVN(); 3136 if (igvn) igvn->add_users_to_worklist(this); 3137 } 3138 3139 // convenience function 3140 // return false if the init contains any stores already 3141 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3142 InitializeNode* init = initialization(); 3143 if (init == NULL || init->is_complete()) return false; 3144 init->remove_extra_zeroes(); 3145 // for now, if this allocation has already collected any inits, bail: 3146 if (init->is_non_zero()) return false; 3147 init->set_complete(phase); 3148 return true; 3149 } 3150 3151 void InitializeNode::remove_extra_zeroes() { 3152 if (req() == RawStores) return; 3153 Node* zmem = zero_memory(); 3154 uint fill = RawStores; 3155 for (uint i = fill; i < req(); i++) { 3156 Node* n = in(i); 3157 if (n->is_top() || n == zmem) continue; // skip 3158 if (fill < i) set_req(fill, n); // compact 3159 ++fill; 3160 } 3161 // delete any empty spaces created: 3162 while (fill < req()) { 3163 del_req(fill); 3164 } 3165 } 3166 3167 // Helper for remembering which stores go with which offsets. 3168 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3169 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3170 intptr_t offset = -1; 3171 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3172 phase, offset); 3173 if (base == NULL) return -1; // something is dead, 3174 if (offset < 0) return -1; // dead, dead 3175 return offset; 3176 } 3177 3178 // Helper for proving that an initialization expression is 3179 // "simple enough" to be folded into an object initialization. 3180 // Attempts to prove that a store's initial value 'n' can be captured 3181 // within the initialization without creating a vicious cycle, such as: 3182 // { Foo p = new Foo(); p.next = p; } 3183 // True for constants and parameters and small combinations thereof. 3184 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3185 if (n == NULL) return true; // (can this really happen?) 3186 if (n->is_Proj()) n = n->in(0); 3187 if (n == this) return false; // found a cycle 3188 if (n->is_Con()) return true; 3189 if (n->is_Start()) return true; // params, etc., are OK 3190 if (n->is_Root()) return true; // even better 3191 3192 Node* ctl = n->in(0); 3193 if (ctl != NULL && !ctl->is_top()) { 3194 if (ctl->is_Proj()) ctl = ctl->in(0); 3195 if (ctl == this) return false; 3196 3197 // If we already know that the enclosing memory op is pinned right after 3198 // the init, then any control flow that the store has picked up 3199 // must have preceded the init, or else be equal to the init. 3200 // Even after loop optimizations (which might change control edges) 3201 // a store is never pinned *before* the availability of its inputs. 3202 if (!MemNode::all_controls_dominate(n, this)) 3203 return false; // failed to prove a good control 3204 } 3205 3206 // Check data edges for possible dependencies on 'this'. 3207 if ((count += 1) > 20) return false; // complexity limit 3208 for (uint i = 1; i < n->req(); i++) { 3209 Node* m = n->in(i); 3210 if (m == NULL || m == n || m->is_top()) continue; 3211 uint first_i = n->find_edge(m); 3212 if (i != first_i) continue; // process duplicate edge just once 3213 if (!detect_init_independence(m, count)) { 3214 return false; 3215 } 3216 } 3217 3218 return true; 3219 } 3220 3221 // Here are all the checks a Store must pass before it can be moved into 3222 // an initialization. Returns zero if a check fails. 3223 // On success, returns the (constant) offset to which the store applies, 3224 // within the initialized memory. 3225 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3226 const int FAIL = 0; 3227 if (st->is_unaligned_access()) { 3228 return FAIL; 3229 } 3230 if (st->req() != MemNode::ValueIn + 1) 3231 return FAIL; // an inscrutable StoreNode (card mark?) 3232 Node* ctl = st->in(MemNode::Control); 3233 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3234 return FAIL; // must be unconditional after the initialization 3235 Node* mem = st->in(MemNode::Memory); 3236 if (!(mem->is_Proj() && mem->in(0) == this)) 3237 return FAIL; // must not be preceded by other stores 3238 Node* adr = st->in(MemNode::Address); 3239 intptr_t offset; 3240 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3241 if (alloc == NULL) 3242 return FAIL; // inscrutable address 3243 if (alloc != allocation()) 3244 return FAIL; // wrong allocation! (store needs to float up) 3245 Node* val = st->in(MemNode::ValueIn); 3246 int complexity_count = 0; 3247 if (!detect_init_independence(val, complexity_count)) 3248 return FAIL; // stored value must be 'simple enough' 3249 3250 // The Store can be captured only if nothing after the allocation 3251 // and before the Store is using the memory location that the store 3252 // overwrites. 3253 bool failed = false; 3254 // If is_complete_with_arraycopy() is true the shape of the graph is 3255 // well defined and is safe so no need for extra checks. 3256 if (!is_complete_with_arraycopy()) { 3257 // We are going to look at each use of the memory state following 3258 // the allocation to make sure nothing reads the memory that the 3259 // Store writes. 3260 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3261 int alias_idx = phase->C->get_alias_index(t_adr); 3262 ResourceMark rm; 3263 Unique_Node_List mems; 3264 mems.push(mem); 3265 Node* unique_merge = NULL; 3266 for (uint next = 0; next < mems.size(); ++next) { 3267 Node *m = mems.at(next); 3268 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3269 Node *n = m->fast_out(j); 3270 if (n->outcnt() == 0) { 3271 continue; 3272 } 3273 if (n == st) { 3274 continue; 3275 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3276 // If the control of this use is different from the control 3277 // of the Store which is right after the InitializeNode then 3278 // this node cannot be between the InitializeNode and the 3279 // Store. 3280 continue; 3281 } else if (n->is_MergeMem()) { 3282 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3283 // We can hit a MergeMemNode (that will likely go away 3284 // later) that is a direct use of the memory state 3285 // following the InitializeNode on the same slice as the 3286 // store node that we'd like to capture. We need to check 3287 // the uses of the MergeMemNode. 3288 mems.push(n); 3289 } 3290 } else if (n->is_Mem()) { 3291 Node* other_adr = n->in(MemNode::Address); 3292 if (other_adr == adr) { 3293 failed = true; 3294 break; 3295 } else { 3296 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3297 if (other_t_adr != NULL) { 3298 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3299 if (other_alias_idx == alias_idx) { 3300 // A load from the same memory slice as the store right 3301 // after the InitializeNode. We check the control of the 3302 // object/array that is loaded from. If it's the same as 3303 // the store control then we cannot capture the store. 3304 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3305 Node* base = other_adr; 3306 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3307 base = base->in(AddPNode::Base); 3308 if (base != NULL) { 3309 base = base->uncast(); 3310 if (base->is_Proj() && base->in(0) == alloc) { 3311 failed = true; 3312 break; 3313 } 3314 } 3315 } 3316 } 3317 } 3318 } else { 3319 failed = true; 3320 break; 3321 } 3322 } 3323 } 3324 } 3325 if (failed) { 3326 if (!can_reshape) { 3327 // We decided we couldn't capture the store during parsing. We 3328 // should try again during the next IGVN once the graph is 3329 // cleaner. 3330 phase->C->record_for_igvn(st); 3331 } 3332 return FAIL; 3333 } 3334 3335 return offset; // success 3336 } 3337 3338 // Find the captured store in(i) which corresponds to the range 3339 // [start..start+size) in the initialized object. 3340 // If there is one, return its index i. If there isn't, return the 3341 // negative of the index where it should be inserted. 3342 // Return 0 if the queried range overlaps an initialization boundary 3343 // or if dead code is encountered. 3344 // If size_in_bytes is zero, do not bother with overlap checks. 3345 int InitializeNode::captured_store_insertion_point(intptr_t start, 3346 int size_in_bytes, 3347 PhaseTransform* phase) { 3348 const int FAIL = 0, MAX_STORE = BytesPerLong; 3349 3350 if (is_complete()) 3351 return FAIL; // arraycopy got here first; punt 3352 3353 assert(allocation() != NULL, "must be present"); 3354 3355 // no negatives, no header fields: 3356 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3357 3358 // after a certain size, we bail out on tracking all the stores: 3359 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3360 if (start >= ti_limit) return FAIL; 3361 3362 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3363 if (i >= limit) return -(int)i; // not found; here is where to put it 3364 3365 Node* st = in(i); 3366 intptr_t st_off = get_store_offset(st, phase); 3367 if (st_off < 0) { 3368 if (st != zero_memory()) { 3369 return FAIL; // bail out if there is dead garbage 3370 } 3371 } else if (st_off > start) { 3372 // ...we are done, since stores are ordered 3373 if (st_off < start + size_in_bytes) { 3374 return FAIL; // the next store overlaps 3375 } 3376 return -(int)i; // not found; here is where to put it 3377 } else if (st_off < start) { 3378 if (size_in_bytes != 0 && 3379 start < st_off + MAX_STORE && 3380 start < st_off + st->as_Store()->memory_size()) { 3381 return FAIL; // the previous store overlaps 3382 } 3383 } else { 3384 if (size_in_bytes != 0 && 3385 st->as_Store()->memory_size() != size_in_bytes) { 3386 return FAIL; // mismatched store size 3387 } 3388 return i; 3389 } 3390 3391 ++i; 3392 } 3393 } 3394 3395 // Look for a captured store which initializes at the offset 'start' 3396 // with the given size. If there is no such store, and no other 3397 // initialization interferes, then return zero_memory (the memory 3398 // projection of the AllocateNode). 3399 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3400 PhaseTransform* phase) { 3401 assert(stores_are_sane(phase), ""); 3402 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3403 if (i == 0) { 3404 return NULL; // something is dead 3405 } else if (i < 0) { 3406 return zero_memory(); // just primordial zero bits here 3407 } else { 3408 Node* st = in(i); // here is the store at this position 3409 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3410 return st; 3411 } 3412 } 3413 3414 // Create, as a raw pointer, an address within my new object at 'offset'. 3415 Node* InitializeNode::make_raw_address(intptr_t offset, 3416 PhaseTransform* phase) { 3417 Node* addr = in(RawAddress); 3418 if (offset != 0) { 3419 Compile* C = phase->C; 3420 addr = phase->transform( new AddPNode(C->top(), addr, 3421 phase->MakeConX(offset)) ); 3422 } 3423 return addr; 3424 } 3425 3426 // Clone the given store, converting it into a raw store 3427 // initializing a field or element of my new object. 3428 // Caller is responsible for retiring the original store, 3429 // with subsume_node or the like. 3430 // 3431 // From the example above InitializeNode::InitializeNode, 3432 // here are the old stores to be captured: 3433 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3434 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3435 // 3436 // Here is the changed code; note the extra edges on init: 3437 // alloc = (Allocate ...) 3438 // rawoop = alloc.RawAddress 3439 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3440 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3441 // init = (Initialize alloc.Control alloc.Memory rawoop 3442 // rawstore1 rawstore2) 3443 // 3444 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3445 PhaseTransform* phase, bool can_reshape) { 3446 assert(stores_are_sane(phase), ""); 3447 3448 if (start < 0) return NULL; 3449 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3450 3451 Compile* C = phase->C; 3452 int size_in_bytes = st->memory_size(); 3453 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3454 if (i == 0) return NULL; // bail out 3455 Node* prev_mem = NULL; // raw memory for the captured store 3456 if (i > 0) { 3457 prev_mem = in(i); // there is a pre-existing store under this one 3458 set_req(i, C->top()); // temporarily disconnect it 3459 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3460 } else { 3461 i = -i; // no pre-existing store 3462 prev_mem = zero_memory(); // a slice of the newly allocated object 3463 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3464 set_req(--i, C->top()); // reuse this edge; it has been folded away 3465 else 3466 ins_req(i, C->top()); // build a new edge 3467 } 3468 Node* new_st = st->clone(); 3469 new_st->set_req(MemNode::Control, in(Control)); 3470 new_st->set_req(MemNode::Memory, prev_mem); 3471 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3472 new_st = phase->transform(new_st); 3473 3474 // At this point, new_st might have swallowed a pre-existing store 3475 // at the same offset, or perhaps new_st might have disappeared, 3476 // if it redundantly stored the same value (or zero to fresh memory). 3477 3478 // In any case, wire it in: 3479 phase->igvn_rehash_node_delayed(this); 3480 set_req(i, new_st); 3481 3482 // The caller may now kill the old guy. 3483 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3484 assert(check_st == new_st || check_st == NULL, "must be findable"); 3485 assert(!is_complete(), ""); 3486 return new_st; 3487 } 3488 3489 static bool store_constant(jlong* tiles, int num_tiles, 3490 intptr_t st_off, int st_size, 3491 jlong con) { 3492 if ((st_off & (st_size-1)) != 0) 3493 return false; // strange store offset (assume size==2**N) 3494 address addr = (address)tiles + st_off; 3495 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3496 switch (st_size) { 3497 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3498 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3499 case sizeof(jint): *(jint*) addr = (jint) con; break; 3500 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3501 default: return false; // strange store size (detect size!=2**N here) 3502 } 3503 return true; // return success to caller 3504 } 3505 3506 // Coalesce subword constants into int constants and possibly 3507 // into long constants. The goal, if the CPU permits, 3508 // is to initialize the object with a small number of 64-bit tiles. 3509 // Also, convert floating-point constants to bit patterns. 3510 // Non-constants are not relevant to this pass. 3511 // 3512 // In terms of the running example on InitializeNode::InitializeNode 3513 // and InitializeNode::capture_store, here is the transformation 3514 // of rawstore1 and rawstore2 into rawstore12: 3515 // alloc = (Allocate ...) 3516 // rawoop = alloc.RawAddress 3517 // tile12 = 0x00010002 3518 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3519 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3520 // 3521 void 3522 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3523 Node* size_in_bytes, 3524 PhaseGVN* phase) { 3525 Compile* C = phase->C; 3526 3527 assert(stores_are_sane(phase), ""); 3528 // Note: After this pass, they are not completely sane, 3529 // since there may be some overlaps. 3530 3531 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3532 3533 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3534 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3535 size_limit = MIN2(size_limit, ti_limit); 3536 size_limit = align_size_up(size_limit, BytesPerLong); 3537 int num_tiles = size_limit / BytesPerLong; 3538 3539 // allocate space for the tile map: 3540 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3541 jlong tiles_buf[small_len]; 3542 Node* nodes_buf[small_len]; 3543 jlong inits_buf[small_len]; 3544 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3545 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3546 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3547 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3548 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3549 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3550 // tiles: exact bitwise model of all primitive constants 3551 // nodes: last constant-storing node subsumed into the tiles model 3552 // inits: which bytes (in each tile) are touched by any initializations 3553 3554 //// Pass A: Fill in the tile model with any relevant stores. 3555 3556 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3557 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3558 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3559 Node* zmem = zero_memory(); // initially zero memory state 3560 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3561 Node* st = in(i); 3562 intptr_t st_off = get_store_offset(st, phase); 3563 3564 // Figure out the store's offset and constant value: 3565 if (st_off < header_size) continue; //skip (ignore header) 3566 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3567 int st_size = st->as_Store()->memory_size(); 3568 if (st_off + st_size > size_limit) break; 3569 3570 // Record which bytes are touched, whether by constant or not. 3571 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3572 continue; // skip (strange store size) 3573 3574 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3575 if (!val->singleton()) continue; //skip (non-con store) 3576 BasicType type = val->basic_type(); 3577 3578 jlong con = 0; 3579 switch (type) { 3580 case T_INT: con = val->is_int()->get_con(); break; 3581 case T_LONG: con = val->is_long()->get_con(); break; 3582 case T_FLOAT: con = jint_cast(val->getf()); break; 3583 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3584 default: continue; //skip (odd store type) 3585 } 3586 3587 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3588 st->Opcode() == Op_StoreL) { 3589 continue; // This StoreL is already optimal. 3590 } 3591 3592 // Store down the constant. 3593 store_constant(tiles, num_tiles, st_off, st_size, con); 3594 3595 intptr_t j = st_off >> LogBytesPerLong; 3596 3597 if (type == T_INT && st_size == BytesPerInt 3598 && (st_off & BytesPerInt) == BytesPerInt) { 3599 jlong lcon = tiles[j]; 3600 if (!Matcher::isSimpleConstant64(lcon) && 3601 st->Opcode() == Op_StoreI) { 3602 // This StoreI is already optimal by itself. 3603 jint* intcon = (jint*) &tiles[j]; 3604 intcon[1] = 0; // undo the store_constant() 3605 3606 // If the previous store is also optimal by itself, back up and 3607 // undo the action of the previous loop iteration... if we can. 3608 // But if we can't, just let the previous half take care of itself. 3609 st = nodes[j]; 3610 st_off -= BytesPerInt; 3611 con = intcon[0]; 3612 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3613 assert(st_off >= header_size, "still ignoring header"); 3614 assert(get_store_offset(st, phase) == st_off, "must be"); 3615 assert(in(i-1) == zmem, "must be"); 3616 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3617 assert(con == tcon->is_int()->get_con(), "must be"); 3618 // Undo the effects of the previous loop trip, which swallowed st: 3619 intcon[0] = 0; // undo store_constant() 3620 set_req(i-1, st); // undo set_req(i, zmem) 3621 nodes[j] = NULL; // undo nodes[j] = st 3622 --old_subword; // undo ++old_subword 3623 } 3624 continue; // This StoreI is already optimal. 3625 } 3626 } 3627 3628 // This store is not needed. 3629 set_req(i, zmem); 3630 nodes[j] = st; // record for the moment 3631 if (st_size < BytesPerLong) // something has changed 3632 ++old_subword; // includes int/float, but who's counting... 3633 else ++old_long; 3634 } 3635 3636 if ((old_subword + old_long) == 0) 3637 return; // nothing more to do 3638 3639 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3640 // Be sure to insert them before overlapping non-constant stores. 3641 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3642 for (int j = 0; j < num_tiles; j++) { 3643 jlong con = tiles[j]; 3644 jlong init = inits[j]; 3645 if (con == 0) continue; 3646 jint con0, con1; // split the constant, address-wise 3647 jint init0, init1; // split the init map, address-wise 3648 { union { jlong con; jint intcon[2]; } u; 3649 u.con = con; 3650 con0 = u.intcon[0]; 3651 con1 = u.intcon[1]; 3652 u.con = init; 3653 init0 = u.intcon[0]; 3654 init1 = u.intcon[1]; 3655 } 3656 3657 Node* old = nodes[j]; 3658 assert(old != NULL, "need the prior store"); 3659 intptr_t offset = (j * BytesPerLong); 3660 3661 bool split = !Matcher::isSimpleConstant64(con); 3662 3663 if (offset < header_size) { 3664 assert(offset + BytesPerInt >= header_size, "second int counts"); 3665 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3666 split = true; // only the second word counts 3667 // Example: int a[] = { 42 ... } 3668 } else if (con0 == 0 && init0 == -1) { 3669 split = true; // first word is covered by full inits 3670 // Example: int a[] = { ... foo(), 42 ... } 3671 } else if (con1 == 0 && init1 == -1) { 3672 split = true; // second word is covered by full inits 3673 // Example: int a[] = { ... 42, foo() ... } 3674 } 3675 3676 // Here's a case where init0 is neither 0 nor -1: 3677 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3678 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3679 // In this case the tile is not split; it is (jlong)42. 3680 // The big tile is stored down, and then the foo() value is inserted. 3681 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3682 3683 Node* ctl = old->in(MemNode::Control); 3684 Node* adr = make_raw_address(offset, phase); 3685 const TypePtr* atp = TypeRawPtr::BOTTOM; 3686 3687 // One or two coalesced stores to plop down. 3688 Node* st[2]; 3689 intptr_t off[2]; 3690 int nst = 0; 3691 if (!split) { 3692 ++new_long; 3693 off[nst] = offset; 3694 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3695 phase->longcon(con), T_LONG, MemNode::unordered); 3696 } else { 3697 // Omit either if it is a zero. 3698 if (con0 != 0) { 3699 ++new_int; 3700 off[nst] = offset; 3701 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3702 phase->intcon(con0), T_INT, MemNode::unordered); 3703 } 3704 if (con1 != 0) { 3705 ++new_int; 3706 offset += BytesPerInt; 3707 adr = make_raw_address(offset, phase); 3708 off[nst] = offset; 3709 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3710 phase->intcon(con1), T_INT, MemNode::unordered); 3711 } 3712 } 3713 3714 // Insert second store first, then the first before the second. 3715 // Insert each one just before any overlapping non-constant stores. 3716 while (nst > 0) { 3717 Node* st1 = st[--nst]; 3718 C->copy_node_notes_to(st1, old); 3719 st1 = phase->transform(st1); 3720 offset = off[nst]; 3721 assert(offset >= header_size, "do not smash header"); 3722 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3723 guarantee(ins_idx != 0, "must re-insert constant store"); 3724 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3725 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3726 set_req(--ins_idx, st1); 3727 else 3728 ins_req(ins_idx, st1); 3729 } 3730 } 3731 3732 if (PrintCompilation && WizardMode) 3733 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3734 old_subword, old_long, new_int, new_long); 3735 if (C->log() != NULL) 3736 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3737 old_subword, old_long, new_int, new_long); 3738 3739 // Clean up any remaining occurrences of zmem: 3740 remove_extra_zeroes(); 3741 } 3742 3743 // Explore forward from in(start) to find the first fully initialized 3744 // word, and return its offset. Skip groups of subword stores which 3745 // together initialize full words. If in(start) is itself part of a 3746 // fully initialized word, return the offset of in(start). If there 3747 // are no following full-word stores, or if something is fishy, return 3748 // a negative value. 3749 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3750 int int_map = 0; 3751 intptr_t int_map_off = 0; 3752 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3753 3754 for (uint i = start, limit = req(); i < limit; i++) { 3755 Node* st = in(i); 3756 3757 intptr_t st_off = get_store_offset(st, phase); 3758 if (st_off < 0) break; // return conservative answer 3759 3760 int st_size = st->as_Store()->memory_size(); 3761 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3762 return st_off; // we found a complete word init 3763 } 3764 3765 // update the map: 3766 3767 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3768 if (this_int_off != int_map_off) { 3769 // reset the map: 3770 int_map = 0; 3771 int_map_off = this_int_off; 3772 } 3773 3774 int subword_off = st_off - this_int_off; 3775 int_map |= right_n_bits(st_size) << subword_off; 3776 if ((int_map & FULL_MAP) == FULL_MAP) { 3777 return this_int_off; // we found a complete word init 3778 } 3779 3780 // Did this store hit or cross the word boundary? 3781 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3782 if (next_int_off == this_int_off + BytesPerInt) { 3783 // We passed the current int, without fully initializing it. 3784 int_map_off = next_int_off; 3785 int_map >>= BytesPerInt; 3786 } else if (next_int_off > this_int_off + BytesPerInt) { 3787 // We passed the current and next int. 3788 return this_int_off + BytesPerInt; 3789 } 3790 } 3791 3792 return -1; 3793 } 3794 3795 3796 // Called when the associated AllocateNode is expanded into CFG. 3797 // At this point, we may perform additional optimizations. 3798 // Linearize the stores by ascending offset, to make memory 3799 // activity as coherent as possible. 3800 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3801 intptr_t header_size, 3802 Node* size_in_bytes, 3803 PhaseGVN* phase) { 3804 assert(!is_complete(), "not already complete"); 3805 assert(stores_are_sane(phase), ""); 3806 assert(allocation() != NULL, "must be present"); 3807 3808 remove_extra_zeroes(); 3809 3810 if (ReduceFieldZeroing || ReduceBulkZeroing) 3811 // reduce instruction count for common initialization patterns 3812 coalesce_subword_stores(header_size, size_in_bytes, phase); 3813 3814 Node* zmem = zero_memory(); // initially zero memory state 3815 Node* inits = zmem; // accumulating a linearized chain of inits 3816 #ifdef ASSERT 3817 intptr_t first_offset = allocation()->minimum_header_size(); 3818 intptr_t last_init_off = first_offset; // previous init offset 3819 intptr_t last_init_end = first_offset; // previous init offset+size 3820 intptr_t last_tile_end = first_offset; // previous tile offset+size 3821 #endif 3822 intptr_t zeroes_done = header_size; 3823 3824 bool do_zeroing = true; // we might give up if inits are very sparse 3825 int big_init_gaps = 0; // how many large gaps have we seen? 3826 3827 if (UseTLAB && ZeroTLAB) do_zeroing = false; 3828 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3829 3830 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3831 Node* st = in(i); 3832 intptr_t st_off = get_store_offset(st, phase); 3833 if (st_off < 0) 3834 break; // unknown junk in the inits 3835 if (st->in(MemNode::Memory) != zmem) 3836 break; // complicated store chains somehow in list 3837 3838 int st_size = st->as_Store()->memory_size(); 3839 intptr_t next_init_off = st_off + st_size; 3840 3841 if (do_zeroing && zeroes_done < next_init_off) { 3842 // See if this store needs a zero before it or under it. 3843 intptr_t zeroes_needed = st_off; 3844 3845 if (st_size < BytesPerInt) { 3846 // Look for subword stores which only partially initialize words. 3847 // If we find some, we must lay down some word-level zeroes first, 3848 // underneath the subword stores. 3849 // 3850 // Examples: 3851 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3852 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3853 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3854 // 3855 // Note: coalesce_subword_stores may have already done this, 3856 // if it was prompted by constant non-zero subword initializers. 3857 // But this case can still arise with non-constant stores. 3858 3859 intptr_t next_full_store = find_next_fullword_store(i, phase); 3860 3861 // In the examples above: 3862 // in(i) p q r s x y z 3863 // st_off 12 13 14 15 12 13 14 3864 // st_size 1 1 1 1 1 1 1 3865 // next_full_s. 12 16 16 16 16 16 16 3866 // z's_done 12 16 16 16 12 16 12 3867 // z's_needed 12 16 16 16 16 16 16 3868 // zsize 0 0 0 0 4 0 4 3869 if (next_full_store < 0) { 3870 // Conservative tack: Zero to end of current word. 3871 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3872 } else { 3873 // Zero to beginning of next fully initialized word. 3874 // Or, don't zero at all, if we are already in that word. 3875 assert(next_full_store >= zeroes_needed, "must go forward"); 3876 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3877 zeroes_needed = next_full_store; 3878 } 3879 } 3880 3881 if (zeroes_needed > zeroes_done) { 3882 intptr_t zsize = zeroes_needed - zeroes_done; 3883 // Do some incremental zeroing on rawmem, in parallel with inits. 3884 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3885 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3886 zeroes_done, zeroes_needed, 3887 phase); 3888 zeroes_done = zeroes_needed; 3889 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 3890 do_zeroing = false; // leave the hole, next time 3891 } 3892 } 3893 3894 // Collect the store and move on: 3895 st->set_req(MemNode::Memory, inits); 3896 inits = st; // put it on the linearized chain 3897 set_req(i, zmem); // unhook from previous position 3898 3899 if (zeroes_done == st_off) 3900 zeroes_done = next_init_off; 3901 3902 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3903 3904 #ifdef ASSERT 3905 // Various order invariants. Weaker than stores_are_sane because 3906 // a large constant tile can be filled in by smaller non-constant stores. 3907 assert(st_off >= last_init_off, "inits do not reverse"); 3908 last_init_off = st_off; 3909 const Type* val = NULL; 3910 if (st_size >= BytesPerInt && 3911 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3912 (int)val->basic_type() < (int)T_OBJECT) { 3913 assert(st_off >= last_tile_end, "tiles do not overlap"); 3914 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3915 last_tile_end = MAX2(last_tile_end, next_init_off); 3916 } else { 3917 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3918 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3919 assert(st_off >= last_init_end, "inits do not overlap"); 3920 last_init_end = next_init_off; // it's a non-tile 3921 } 3922 #endif //ASSERT 3923 } 3924 3925 remove_extra_zeroes(); // clear out all the zmems left over 3926 add_req(inits); 3927 3928 if (!(UseTLAB && ZeroTLAB)) { 3929 // If anything remains to be zeroed, zero it all now. 3930 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3931 // if it is the last unused 4 bytes of an instance, forget about it 3932 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3933 if (zeroes_done + BytesPerLong >= size_limit) { 3934 assert(allocation() != NULL, ""); 3935 if (allocation()->Opcode() == Op_Allocate) { 3936 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3937 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3938 if (zeroes_done == k->layout_helper()) 3939 zeroes_done = size_limit; 3940 } 3941 } 3942 if (zeroes_done < size_limit) { 3943 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3944 zeroes_done, size_in_bytes, phase); 3945 } 3946 } 3947 3948 set_complete(phase); 3949 return rawmem; 3950 } 3951 3952 3953 #ifdef ASSERT 3954 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3955 if (is_complete()) 3956 return true; // stores could be anything at this point 3957 assert(allocation() != NULL, "must be present"); 3958 intptr_t last_off = allocation()->minimum_header_size(); 3959 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3960 Node* st = in(i); 3961 intptr_t st_off = get_store_offset(st, phase); 3962 if (st_off < 0) continue; // ignore dead garbage 3963 if (last_off > st_off) { 3964 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 3965 this->dump(2); 3966 assert(false, "ascending store offsets"); 3967 return false; 3968 } 3969 last_off = st_off + st->as_Store()->memory_size(); 3970 } 3971 return true; 3972 } 3973 #endif //ASSERT 3974 3975 3976 3977 3978 //============================MergeMemNode===================================== 3979 // 3980 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3981 // contributing store or call operations. Each contributor provides the memory 3982 // state for a particular "alias type" (see Compile::alias_type). For example, 3983 // if a MergeMem has an input X for alias category #6, then any memory reference 3984 // to alias category #6 may use X as its memory state input, as an exact equivalent 3985 // to using the MergeMem as a whole. 3986 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3987 // 3988 // (Here, the <N> notation gives the index of the relevant adr_type.) 3989 // 3990 // In one special case (and more cases in the future), alias categories overlap. 3991 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3992 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 3993 // it is exactly equivalent to that state W: 3994 // MergeMem(<Bot>: W) <==> W 3995 // 3996 // Usually, the merge has more than one input. In that case, where inputs 3997 // overlap (i.e., one is Bot), the narrower alias type determines the memory 3998 // state for that type, and the wider alias type (Bot) fills in everywhere else: 3999 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4000 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4001 // 4002 // A merge can take a "wide" memory state as one of its narrow inputs. 4003 // This simply means that the merge observes out only the relevant parts of 4004 // the wide input. That is, wide memory states arriving at narrow merge inputs 4005 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4006 // 4007 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4008 // and that memory slices "leak through": 4009 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4010 // 4011 // But, in such a cascade, repeated memory slices can "block the leak": 4012 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4013 // 4014 // In the last example, Y is not part of the combined memory state of the 4015 // outermost MergeMem. The system must, of course, prevent unschedulable 4016 // memory states from arising, so you can be sure that the state Y is somehow 4017 // a precursor to state Y'. 4018 // 4019 // 4020 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4021 // of each MergeMemNode array are exactly the numerical alias indexes, including 4022 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4023 // Compile::alias_type (and kin) produce and manage these indexes. 4024 // 4025 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4026 // (Note that this provides quick access to the top node inside MergeMem methods, 4027 // without the need to reach out via TLS to Compile::current.) 4028 // 4029 // As a consequence of what was just described, a MergeMem that represents a full 4030 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4031 // containing all alias categories. 4032 // 4033 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4034 // 4035 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4036 // a memory state for the alias type <N>, or else the top node, meaning that 4037 // there is no particular input for that alias type. Note that the length of 4038 // a MergeMem is variable, and may be extended at any time to accommodate new 4039 // memory states at larger alias indexes. When merges grow, they are of course 4040 // filled with "top" in the unused in() positions. 4041 // 4042 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4043 // (Top was chosen because it works smoothly with passes like GCM.) 4044 // 4045 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4046 // the type of random VM bits like TLS references.) Since it is always the 4047 // first non-Bot memory slice, some low-level loops use it to initialize an 4048 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4049 // 4050 // 4051 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4052 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4053 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4054 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4055 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4056 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4057 // 4058 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4059 // really that different from the other memory inputs. An abbreviation called 4060 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4061 // 4062 // 4063 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4064 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4065 // that "emerges though" the base memory will be marked as excluding the alias types 4066 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4067 // 4068 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4069 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4070 // 4071 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4072 // (It is currently unimplemented.) As you can see, the resulting merge is 4073 // actually a disjoint union of memory states, rather than an overlay. 4074 // 4075 4076 //------------------------------MergeMemNode----------------------------------- 4077 Node* MergeMemNode::make_empty_memory() { 4078 Node* empty_memory = (Node*) Compile::current()->top(); 4079 assert(empty_memory->is_top(), "correct sentinel identity"); 4080 return empty_memory; 4081 } 4082 4083 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4084 init_class_id(Class_MergeMem); 4085 // all inputs are nullified in Node::Node(int) 4086 // set_input(0, NULL); // no control input 4087 4088 // Initialize the edges uniformly to top, for starters. 4089 Node* empty_mem = make_empty_memory(); 4090 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4091 init_req(i,empty_mem); 4092 } 4093 assert(empty_memory() == empty_mem, ""); 4094 4095 if( new_base != NULL && new_base->is_MergeMem() ) { 4096 MergeMemNode* mdef = new_base->as_MergeMem(); 4097 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4098 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4099 mms.set_memory(mms.memory2()); 4100 } 4101 assert(base_memory() == mdef->base_memory(), ""); 4102 } else { 4103 set_base_memory(new_base); 4104 } 4105 } 4106 4107 // Make a new, untransformed MergeMem with the same base as 'mem'. 4108 // If mem is itself a MergeMem, populate the result with the same edges. 4109 MergeMemNode* MergeMemNode::make(Node* mem) { 4110 return new MergeMemNode(mem); 4111 } 4112 4113 //------------------------------cmp-------------------------------------------- 4114 uint MergeMemNode::hash() const { return NO_HASH; } 4115 uint MergeMemNode::cmp( const Node &n ) const { 4116 return (&n == this); // Always fail except on self 4117 } 4118 4119 //------------------------------Identity--------------------------------------- 4120 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4121 // Identity if this merge point does not record any interesting memory 4122 // disambiguations. 4123 Node* base_mem = base_memory(); 4124 Node* empty_mem = empty_memory(); 4125 if (base_mem != empty_mem) { // Memory path is not dead? 4126 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4127 Node* mem = in(i); 4128 if (mem != empty_mem && mem != base_mem) { 4129 return this; // Many memory splits; no change 4130 } 4131 } 4132 } 4133 return base_mem; // No memory splits; ID on the one true input 4134 } 4135 4136 //------------------------------Ideal------------------------------------------ 4137 // This method is invoked recursively on chains of MergeMem nodes 4138 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4139 // Remove chain'd MergeMems 4140 // 4141 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4142 // relative to the "in(Bot)". Since we are patching both at the same time, 4143 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4144 // but rewrite each "in(i)" relative to the new "in(Bot)". 4145 Node *progress = NULL; 4146 4147 4148 Node* old_base = base_memory(); 4149 Node* empty_mem = empty_memory(); 4150 if (old_base == empty_mem) 4151 return NULL; // Dead memory path. 4152 4153 MergeMemNode* old_mbase; 4154 if (old_base != NULL && old_base->is_MergeMem()) 4155 old_mbase = old_base->as_MergeMem(); 4156 else 4157 old_mbase = NULL; 4158 Node* new_base = old_base; 4159 4160 // simplify stacked MergeMems in base memory 4161 if (old_mbase) new_base = old_mbase->base_memory(); 4162 4163 // the base memory might contribute new slices beyond my req() 4164 if (old_mbase) grow_to_match(old_mbase); 4165 4166 // Look carefully at the base node if it is a phi. 4167 PhiNode* phi_base; 4168 if (new_base != NULL && new_base->is_Phi()) 4169 phi_base = new_base->as_Phi(); 4170 else 4171 phi_base = NULL; 4172 4173 Node* phi_reg = NULL; 4174 uint phi_len = (uint)-1; 4175 if (phi_base != NULL && !phi_base->is_copy()) { 4176 // do not examine phi if degraded to a copy 4177 phi_reg = phi_base->region(); 4178 phi_len = phi_base->req(); 4179 // see if the phi is unfinished 4180 for (uint i = 1; i < phi_len; i++) { 4181 if (phi_base->in(i) == NULL) { 4182 // incomplete phi; do not look at it yet! 4183 phi_reg = NULL; 4184 phi_len = (uint)-1; 4185 break; 4186 } 4187 } 4188 } 4189 4190 // Note: We do not call verify_sparse on entry, because inputs 4191 // can normalize to the base_memory via subsume_node or similar 4192 // mechanisms. This method repairs that damage. 4193 4194 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4195 4196 // Look at each slice. 4197 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4198 Node* old_in = in(i); 4199 // calculate the old memory value 4200 Node* old_mem = old_in; 4201 if (old_mem == empty_mem) old_mem = old_base; 4202 assert(old_mem == memory_at(i), ""); 4203 4204 // maybe update (reslice) the old memory value 4205 4206 // simplify stacked MergeMems 4207 Node* new_mem = old_mem; 4208 MergeMemNode* old_mmem; 4209 if (old_mem != NULL && old_mem->is_MergeMem()) 4210 old_mmem = old_mem->as_MergeMem(); 4211 else 4212 old_mmem = NULL; 4213 if (old_mmem == this) { 4214 // This can happen if loops break up and safepoints disappear. 4215 // A merge of BotPtr (default) with a RawPtr memory derived from a 4216 // safepoint can be rewritten to a merge of the same BotPtr with 4217 // the BotPtr phi coming into the loop. If that phi disappears 4218 // also, we can end up with a self-loop of the mergemem. 4219 // In general, if loops degenerate and memory effects disappear, 4220 // a mergemem can be left looking at itself. This simply means 4221 // that the mergemem's default should be used, since there is 4222 // no longer any apparent effect on this slice. 4223 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4224 // from start. Update the input to TOP. 4225 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4226 } 4227 else if (old_mmem != NULL) { 4228 new_mem = old_mmem->memory_at(i); 4229 } 4230 // else preceding memory was not a MergeMem 4231 4232 // replace equivalent phis (unfortunately, they do not GVN together) 4233 if (new_mem != NULL && new_mem != new_base && 4234 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4235 if (new_mem->is_Phi()) { 4236 PhiNode* phi_mem = new_mem->as_Phi(); 4237 for (uint i = 1; i < phi_len; i++) { 4238 if (phi_base->in(i) != phi_mem->in(i)) { 4239 phi_mem = NULL; 4240 break; 4241 } 4242 } 4243 if (phi_mem != NULL) { 4244 // equivalent phi nodes; revert to the def 4245 new_mem = new_base; 4246 } 4247 } 4248 } 4249 4250 // maybe store down a new value 4251 Node* new_in = new_mem; 4252 if (new_in == new_base) new_in = empty_mem; 4253 4254 if (new_in != old_in) { 4255 // Warning: Do not combine this "if" with the previous "if" 4256 // A memory slice might have be be rewritten even if it is semantically 4257 // unchanged, if the base_memory value has changed. 4258 set_req(i, new_in); 4259 progress = this; // Report progress 4260 } 4261 } 4262 4263 if (new_base != old_base) { 4264 set_req(Compile::AliasIdxBot, new_base); 4265 // Don't use set_base_memory(new_base), because we need to update du. 4266 assert(base_memory() == new_base, ""); 4267 progress = this; 4268 } 4269 4270 if( base_memory() == this ) { 4271 // a self cycle indicates this memory path is dead 4272 set_req(Compile::AliasIdxBot, empty_mem); 4273 } 4274 4275 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4276 // Recursion must occur after the self cycle check above 4277 if( base_memory()->is_MergeMem() ) { 4278 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4279 Node *m = phase->transform(new_mbase); // Rollup any cycles 4280 if( m != NULL && (m->is_top() || 4281 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4282 // propagate rollup of dead cycle to self 4283 set_req(Compile::AliasIdxBot, empty_mem); 4284 } 4285 } 4286 4287 if( base_memory() == empty_mem ) { 4288 progress = this; 4289 // Cut inputs during Parse phase only. 4290 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4291 if( !can_reshape ) { 4292 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4293 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4294 } 4295 } 4296 } 4297 4298 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4299 // Check if PhiNode::Ideal's "Split phis through memory merges" 4300 // transform should be attempted. Look for this->phi->this cycle. 4301 uint merge_width = req(); 4302 if (merge_width > Compile::AliasIdxRaw) { 4303 PhiNode* phi = base_memory()->as_Phi(); 4304 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4305 if (phi->in(i) == this) { 4306 phase->is_IterGVN()->_worklist.push(phi); 4307 break; 4308 } 4309 } 4310 } 4311 } 4312 4313 assert(progress || verify_sparse(), "please, no dups of base"); 4314 return progress; 4315 } 4316 4317 //-------------------------set_base_memory------------------------------------- 4318 void MergeMemNode::set_base_memory(Node *new_base) { 4319 Node* empty_mem = empty_memory(); 4320 set_req(Compile::AliasIdxBot, new_base); 4321 assert(memory_at(req()) == new_base, "must set default memory"); 4322 // Clear out other occurrences of new_base: 4323 if (new_base != empty_mem) { 4324 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4325 if (in(i) == new_base) set_req(i, empty_mem); 4326 } 4327 } 4328 } 4329 4330 //------------------------------out_RegMask------------------------------------ 4331 const RegMask &MergeMemNode::out_RegMask() const { 4332 return RegMask::Empty; 4333 } 4334 4335 //------------------------------dump_spec-------------------------------------- 4336 #ifndef PRODUCT 4337 void MergeMemNode::dump_spec(outputStream *st) const { 4338 st->print(" {"); 4339 Node* base_mem = base_memory(); 4340 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4341 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4342 if (mem == base_mem) { st->print(" -"); continue; } 4343 st->print( " N%d:", mem->_idx ); 4344 Compile::current()->get_adr_type(i)->dump_on(st); 4345 } 4346 st->print(" }"); 4347 } 4348 #endif // !PRODUCT 4349 4350 4351 #ifdef ASSERT 4352 static bool might_be_same(Node* a, Node* b) { 4353 if (a == b) return true; 4354 if (!(a->is_Phi() || b->is_Phi())) return false; 4355 // phis shift around during optimization 4356 return true; // pretty stupid... 4357 } 4358 4359 // verify a narrow slice (either incoming or outgoing) 4360 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4361 if (!VerifyAliases) return; // don't bother to verify unless requested 4362 if (is_error_reported()) return; // muzzle asserts when debugging an error 4363 if (Node::in_dump()) return; // muzzle asserts when printing 4364 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4365 assert(n != NULL, ""); 4366 // Elide intervening MergeMem's 4367 while (n->is_MergeMem()) { 4368 n = n->as_MergeMem()->memory_at(alias_idx); 4369 } 4370 Compile* C = Compile::current(); 4371 const TypePtr* n_adr_type = n->adr_type(); 4372 if (n == m->empty_memory()) { 4373 // Implicit copy of base_memory() 4374 } else if (n_adr_type != TypePtr::BOTTOM) { 4375 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4376 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4377 } else { 4378 // A few places like make_runtime_call "know" that VM calls are narrow, 4379 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4380 bool expected_wide_mem = false; 4381 if (n == m->base_memory()) { 4382 expected_wide_mem = true; 4383 } else if (alias_idx == Compile::AliasIdxRaw || 4384 n == m->memory_at(Compile::AliasIdxRaw)) { 4385 expected_wide_mem = true; 4386 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4387 // memory can "leak through" calls on channels that 4388 // are write-once. Allow this also. 4389 expected_wide_mem = true; 4390 } 4391 assert(expected_wide_mem, "expected narrow slice replacement"); 4392 } 4393 } 4394 #else // !ASSERT 4395 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4396 #endif 4397 4398 4399 //-----------------------------memory_at--------------------------------------- 4400 Node* MergeMemNode::memory_at(uint alias_idx) const { 4401 assert(alias_idx >= Compile::AliasIdxRaw || 4402 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4403 "must avoid base_memory and AliasIdxTop"); 4404 4405 // Otherwise, it is a narrow slice. 4406 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4407 Compile *C = Compile::current(); 4408 if (is_empty_memory(n)) { 4409 // the array is sparse; empty slots are the "top" node 4410 n = base_memory(); 4411 assert(Node::in_dump() 4412 || n == NULL || n->bottom_type() == Type::TOP 4413 || n->adr_type() == NULL // address is TOP 4414 || n->adr_type() == TypePtr::BOTTOM 4415 || n->adr_type() == TypeRawPtr::BOTTOM 4416 || Compile::current()->AliasLevel() == 0, 4417 "must be a wide memory"); 4418 // AliasLevel == 0 if we are organizing the memory states manually. 4419 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4420 } else { 4421 // make sure the stored slice is sane 4422 #ifdef ASSERT 4423 if (is_error_reported() || Node::in_dump()) { 4424 } else if (might_be_same(n, base_memory())) { 4425 // Give it a pass: It is a mostly harmless repetition of the base. 4426 // This can arise normally from node subsumption during optimization. 4427 } else { 4428 verify_memory_slice(this, alias_idx, n); 4429 } 4430 #endif 4431 } 4432 return n; 4433 } 4434 4435 //---------------------------set_memory_at------------------------------------- 4436 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4437 verify_memory_slice(this, alias_idx, n); 4438 Node* empty_mem = empty_memory(); 4439 if (n == base_memory()) n = empty_mem; // collapse default 4440 uint need_req = alias_idx+1; 4441 if (req() < need_req) { 4442 if (n == empty_mem) return; // already the default, so do not grow me 4443 // grow the sparse array 4444 do { 4445 add_req(empty_mem); 4446 } while (req() < need_req); 4447 } 4448 set_req( alias_idx, n ); 4449 } 4450 4451 4452 4453 //--------------------------iteration_setup------------------------------------ 4454 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4455 if (other != NULL) { 4456 grow_to_match(other); 4457 // invariant: the finite support of mm2 is within mm->req() 4458 #ifdef ASSERT 4459 for (uint i = req(); i < other->req(); i++) { 4460 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4461 } 4462 #endif 4463 } 4464 // Replace spurious copies of base_memory by top. 4465 Node* base_mem = base_memory(); 4466 if (base_mem != NULL && !base_mem->is_top()) { 4467 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4468 if (in(i) == base_mem) 4469 set_req(i, empty_memory()); 4470 } 4471 } 4472 } 4473 4474 //---------------------------grow_to_match------------------------------------- 4475 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4476 Node* empty_mem = empty_memory(); 4477 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4478 // look for the finite support of the other memory 4479 for (uint i = other->req(); --i >= req(); ) { 4480 if (other->in(i) != empty_mem) { 4481 uint new_len = i+1; 4482 while (req() < new_len) add_req(empty_mem); 4483 break; 4484 } 4485 } 4486 } 4487 4488 //---------------------------verify_sparse------------------------------------- 4489 #ifndef PRODUCT 4490 bool MergeMemNode::verify_sparse() const { 4491 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4492 Node* base_mem = base_memory(); 4493 // The following can happen in degenerate cases, since empty==top. 4494 if (is_empty_memory(base_mem)) return true; 4495 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4496 assert(in(i) != NULL, "sane slice"); 4497 if (in(i) == base_mem) return false; // should have been the sentinel value! 4498 } 4499 return true; 4500 } 4501 4502 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4503 Node* n; 4504 n = mm->in(idx); 4505 if (mem == n) return true; // might be empty_memory() 4506 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4507 if (mem == n) return true; 4508 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4509 if (mem == n) return true; 4510 if (n == NULL) break; 4511 } 4512 return false; 4513 } 4514 #endif // !PRODUCT