1 /* 2 * Copyright (c) 1997, 2015, 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 "oops/objArrayKlass.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/arraycopynode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/compile.hpp" 34 #include "opto/connode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/machnode.hpp" 38 #include "opto/matcher.hpp" 39 #include "opto/memnode.hpp" 40 #include "opto/mulnode.hpp" 41 #include "opto/narrowptrnode.hpp" 42 #include "opto/phaseX.hpp" 43 #include "opto/regmask.hpp" 44 #include "utilities/copy.hpp" 45 46 // Portions of code courtesy of Clifford Click 47 48 // Optimization - Graph Style 49 50 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 51 52 //============================================================================= 53 uint MemNode::size_of() const { return sizeof(*this); } 54 55 const TypePtr *MemNode::adr_type() const { 56 Node* adr = in(Address); 57 if (adr == NULL) return NULL; // node is dead 58 const TypePtr* cross_check = NULL; 59 DEBUG_ONLY(cross_check = _adr_type); 60 return calculate_adr_type(adr->bottom_type(), cross_check); 61 } 62 63 #ifndef PRODUCT 64 void MemNode::dump_spec(outputStream *st) const { 65 if (in(Address) == NULL) return; // node is dead 66 #ifndef ASSERT 67 // fake the missing field 68 const TypePtr* _adr_type = NULL; 69 if (in(Address) != NULL) 70 _adr_type = in(Address)->bottom_type()->isa_ptr(); 71 #endif 72 dump_adr_type(this, _adr_type, st); 73 74 Compile* C = Compile::current(); 75 if (C->alias_type(_adr_type)->is_volatile()) { 76 st->print(" Volatile!"); 77 } 78 if (_unaligned_access) { 79 st->print(" unaligned"); 80 } 81 if (_mismatched_access) { 82 st->print(" mismatched"); 83 } 84 } 85 86 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 87 st->print(" @"); 88 if (adr_type == NULL) { 89 st->print("NULL"); 90 } else { 91 adr_type->dump_on(st); 92 Compile* C = Compile::current(); 93 Compile::AliasType* atp = NULL; 94 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 95 if (atp == NULL) 96 st->print(", idx=?\?;"); 97 else if (atp->index() == Compile::AliasIdxBot) 98 st->print(", idx=Bot;"); 99 else if (atp->index() == Compile::AliasIdxTop) 100 st->print(", idx=Top;"); 101 else if (atp->index() == Compile::AliasIdxRaw) 102 st->print(", idx=Raw;"); 103 else { 104 ciField* field = atp->field(); 105 if (field) { 106 st->print(", name="); 107 field->print_name_on(st); 108 } 109 st->print(", idx=%d;", atp->index()); 110 } 111 } 112 } 113 114 extern void print_alias_types(); 115 116 #endif 117 118 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 119 assert((t_oop != NULL), "sanity"); 120 bool is_instance = t_oop->is_known_instance_field(); 121 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 122 (load != NULL) && load->is_Load() && 123 (phase->is_IterGVN() != NULL); 124 if (!(is_instance || is_boxed_value_load)) 125 return mchain; // don't try to optimize non-instance types 126 uint instance_id = t_oop->instance_id(); 127 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 128 Node *prev = NULL; 129 Node *result = mchain; 130 while (prev != result) { 131 prev = result; 132 if (result == start_mem) 133 break; // hit one of our sentinels 134 // skip over a call which does not affect this memory slice 135 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 136 Node *proj_in = result->in(0); 137 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 138 break; // hit one of our sentinels 139 } else if (proj_in->is_Call()) { 140 // ArrayCopyNodes processed here as well 141 CallNode *call = proj_in->as_Call(); 142 if (!call->may_modify(t_oop, phase)) { // returns false for instances 143 result = call->in(TypeFunc::Memory); 144 } 145 } else if (proj_in->is_Initialize()) { 146 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 147 // Stop if this is the initialization for the object instance which 148 // contains this memory slice, otherwise skip over it. 149 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 150 break; 151 } 152 if (is_instance) { 153 result = proj_in->in(TypeFunc::Memory); 154 } else if (is_boxed_value_load) { 155 Node* klass = alloc->in(AllocateNode::KlassNode); 156 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 157 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 158 result = proj_in->in(TypeFunc::Memory); // not related allocation 159 } 160 } 161 } else if (proj_in->is_MemBar()) { 162 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase)) { 163 break; 164 } 165 result = proj_in->in(TypeFunc::Memory); 166 } else { 167 assert(false, "unexpected projection"); 168 } 169 } else if (result->is_ClearArray()) { 170 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 171 // Can not bypass initialization of the instance 172 // we are looking for. 173 break; 174 } 175 // Otherwise skip it (the call updated 'result' value). 176 } else if (result->is_MergeMem()) { 177 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 178 } 179 } 180 return result; 181 } 182 183 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 184 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 185 if (t_oop == NULL) 186 return mchain; // don't try to optimize non-oop types 187 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 188 bool is_instance = t_oop->is_known_instance_field(); 189 PhaseIterGVN *igvn = phase->is_IterGVN(); 190 if (is_instance && igvn != NULL && result->is_Phi()) { 191 PhiNode *mphi = result->as_Phi(); 192 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 193 const TypePtr *t = mphi->adr_type(); 194 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 195 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 196 t->is_oopptr()->cast_to_exactness(true) 197 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 198 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 199 // clone the Phi with our address type 200 result = mphi->split_out_instance(t_adr, igvn); 201 } else { 202 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 203 } 204 } 205 return result; 206 } 207 208 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 209 uint alias_idx = phase->C->get_alias_index(tp); 210 Node *mem = mmem; 211 #ifdef ASSERT 212 { 213 // Check that current type is consistent with the alias index used during graph construction 214 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 215 bool consistent = adr_check == NULL || adr_check->empty() || 216 phase->C->must_alias(adr_check, alias_idx ); 217 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 218 if( !consistent && adr_check != NULL && !adr_check->empty() && 219 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 220 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 221 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 222 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 223 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 224 // don't assert if it is dead code. 225 consistent = true; 226 } 227 if( !consistent ) { 228 st->print("alias_idx==%d, adr_check==", alias_idx); 229 if( adr_check == NULL ) { 230 st->print("NULL"); 231 } else { 232 adr_check->dump(); 233 } 234 st->cr(); 235 print_alias_types(); 236 assert(consistent, "adr_check must match alias idx"); 237 } 238 } 239 #endif 240 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 241 // means an array I have not precisely typed yet. Do not do any 242 // alias stuff with it any time soon. 243 const TypeOopPtr *toop = tp->isa_oopptr(); 244 if( tp->base() != Type::AnyPtr && 245 !(toop && 246 toop->klass() != NULL && 247 toop->klass()->is_java_lang_Object() && 248 toop->offset() == Type::OffsetBot) ) { 249 // compress paths and change unreachable cycles to TOP 250 // If not, we can update the input infinitely along a MergeMem cycle 251 // Equivalent code in PhiNode::Ideal 252 Node* m = phase->transform(mmem); 253 // If transformed to a MergeMem, get the desired slice 254 // Otherwise the returned node represents memory for every slice 255 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 256 // Update input if it is progress over what we have now 257 } 258 return mem; 259 } 260 261 //--------------------------Ideal_common--------------------------------------- 262 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 263 // Unhook non-raw memories from complete (macro-expanded) initializations. 264 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 265 // If our control input is a dead region, kill all below the region 266 Node *ctl = in(MemNode::Control); 267 if (ctl && remove_dead_region(phase, can_reshape)) 268 return this; 269 ctl = in(MemNode::Control); 270 // Don't bother trying to transform a dead node 271 if (ctl && ctl->is_top()) return NodeSentinel; 272 273 PhaseIterGVN *igvn = phase->is_IterGVN(); 274 // Wait if control on the worklist. 275 if (ctl && can_reshape && igvn != NULL) { 276 Node* bol = NULL; 277 Node* cmp = NULL; 278 if (ctl->in(0)->is_If()) { 279 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 280 bol = ctl->in(0)->in(1); 281 if (bol->is_Bool()) 282 cmp = ctl->in(0)->in(1)->in(1); 283 } 284 if (igvn->_worklist.member(ctl) || 285 (bol != NULL && igvn->_worklist.member(bol)) || 286 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 287 // This control path may be dead. 288 // Delay this memory node transformation until the control is processed. 289 phase->is_IterGVN()->_worklist.push(this); 290 return NodeSentinel; // caller will return NULL 291 } 292 } 293 // Ignore if memory is dead, or self-loop 294 Node *mem = in(MemNode::Memory); 295 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 296 assert(mem != this, "dead loop in MemNode::Ideal"); 297 298 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 299 // This memory slice may be dead. 300 // Delay this mem node transformation until the memory is processed. 301 phase->is_IterGVN()->_worklist.push(this); 302 return NodeSentinel; // caller will return NULL 303 } 304 305 Node *address = in(MemNode::Address); 306 const Type *t_adr = phase->type(address); 307 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 308 309 if (can_reshape && igvn != NULL && 310 (igvn->_worklist.member(address) || 311 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { 312 // The address's base and type may change when the address is processed. 313 // Delay this mem node transformation until the address is processed. 314 phase->is_IterGVN()->_worklist.push(this); 315 return NodeSentinel; // caller will return NULL 316 } 317 318 // Do NOT remove or optimize the next lines: ensure a new alias index 319 // is allocated for an oop pointer type before Escape Analysis. 320 // Note: C++ will not remove it since the call has side effect. 321 if (t_adr->isa_oopptr()) { 322 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 323 } 324 325 Node* base = NULL; 326 if (address->is_AddP()) { 327 base = address->in(AddPNode::Base); 328 } 329 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 330 !t_adr->isa_rawptr()) { 331 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 332 // Skip this node optimization if its address has TOP base. 333 return NodeSentinel; // caller will return NULL 334 } 335 336 // Avoid independent memory operations 337 Node* old_mem = mem; 338 339 // The code which unhooks non-raw memories from complete (macro-expanded) 340 // initializations was removed. After macro-expansion all stores catched 341 // by Initialize node became raw stores and there is no information 342 // which memory slices they modify. So it is unsafe to move any memory 343 // operation above these stores. Also in most cases hooked non-raw memories 344 // were already unhooked by using information from detect_ptr_independence() 345 // and find_previous_store(). 346 347 if (mem->is_MergeMem()) { 348 MergeMemNode* mmem = mem->as_MergeMem(); 349 const TypePtr *tp = t_adr->is_ptr(); 350 351 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 352 } 353 354 if (mem != old_mem) { 355 set_req(MemNode::Memory, mem); 356 if (can_reshape && old_mem->outcnt() == 0) { 357 igvn->_worklist.push(old_mem); 358 } 359 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 360 return this; 361 } 362 363 // let the subclass continue analyzing... 364 return NULL; 365 } 366 367 // Helper function for proving some simple control dominations. 368 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 369 // Already assumes that 'dom' is available at 'sub', and that 'sub' 370 // is not a constant (dominated by the method's StartNode). 371 // Used by MemNode::find_previous_store to prove that the 372 // control input of a memory operation predates (dominates) 373 // an allocation it wants to look past. 374 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 375 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 376 return false; // Conservative answer for dead code 377 378 // Check 'dom'. Skip Proj and CatchProj nodes. 379 dom = dom->find_exact_control(dom); 380 if (dom == NULL || dom->is_top()) 381 return false; // Conservative answer for dead code 382 383 if (dom == sub) { 384 // For the case when, for example, 'sub' is Initialize and the original 385 // 'dom' is Proj node of the 'sub'. 386 return false; 387 } 388 389 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 390 return true; 391 392 // 'dom' dominates 'sub' if its control edge and control edges 393 // of all its inputs dominate or equal to sub's control edge. 394 395 // Currently 'sub' is either Allocate, Initialize or Start nodes. 396 // Or Region for the check in LoadNode::Ideal(); 397 // 'sub' should have sub->in(0) != NULL. 398 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 399 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 400 401 // Get control edge of 'sub'. 402 Node* orig_sub = sub; 403 sub = sub->find_exact_control(sub->in(0)); 404 if (sub == NULL || sub->is_top()) 405 return false; // Conservative answer for dead code 406 407 assert(sub->is_CFG(), "expecting control"); 408 409 if (sub == dom) 410 return true; 411 412 if (sub->is_Start() || sub->is_Root()) 413 return false; 414 415 { 416 // Check all control edges of 'dom'. 417 418 ResourceMark rm; 419 Arena* arena = Thread::current()->resource_area(); 420 Node_List nlist(arena); 421 Unique_Node_List dom_list(arena); 422 423 dom_list.push(dom); 424 bool only_dominating_controls = false; 425 426 for (uint next = 0; next < dom_list.size(); next++) { 427 Node* n = dom_list.at(next); 428 if (n == orig_sub) 429 return false; // One of dom's inputs dominated by sub. 430 if (!n->is_CFG() && n->pinned()) { 431 // Check only own control edge for pinned non-control nodes. 432 n = n->find_exact_control(n->in(0)); 433 if (n == NULL || n->is_top()) 434 return false; // Conservative answer for dead code 435 assert(n->is_CFG(), "expecting control"); 436 dom_list.push(n); 437 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 438 only_dominating_controls = true; 439 } else if (n->is_CFG()) { 440 if (n->dominates(sub, nlist)) 441 only_dominating_controls = true; 442 else 443 return false; 444 } else { 445 // First, own control edge. 446 Node* m = n->find_exact_control(n->in(0)); 447 if (m != NULL) { 448 if (m->is_top()) 449 return false; // Conservative answer for dead code 450 dom_list.push(m); 451 } 452 // Now, the rest of edges. 453 uint cnt = n->req(); 454 for (uint i = 1; i < cnt; i++) { 455 m = n->find_exact_control(n->in(i)); 456 if (m == NULL || m->is_top()) 457 continue; 458 dom_list.push(m); 459 } 460 } 461 } 462 return only_dominating_controls; 463 } 464 } 465 466 //---------------------detect_ptr_independence--------------------------------- 467 // Used by MemNode::find_previous_store to prove that two base 468 // pointers are never equal. 469 // The pointers are accompanied by their associated allocations, 470 // if any, which have been previously discovered by the caller. 471 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 472 Node* p2, AllocateNode* a2, 473 PhaseTransform* phase) { 474 // Attempt to prove that these two pointers cannot be aliased. 475 // They may both manifestly be allocations, and they should differ. 476 // Or, if they are not both allocations, they can be distinct constants. 477 // Otherwise, one is an allocation and the other a pre-existing value. 478 if (a1 == NULL && a2 == NULL) { // neither an allocation 479 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 480 } else if (a1 != NULL && a2 != NULL) { // both allocations 481 return (a1 != a2); 482 } else if (a1 != NULL) { // one allocation a1 483 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 484 return all_controls_dominate(p2, a1); 485 } else { //(a2 != NULL) // one allocation a2 486 return all_controls_dominate(p1, a2); 487 } 488 return false; 489 } 490 491 492 // Find an arraycopy that must have set (can_see_stored_value=true) or 493 // could have set (can_see_stored_value=false) the value for this load 494 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 495 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 496 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 497 Node* mb = mem->in(0); 498 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 499 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 500 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 501 if (ac->is_clonebasic()) { 502 intptr_t offset; 503 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 504 assert(alloc != NULL && alloc->initialization()->is_complete_with_arraycopy(), "broken allocation"); 505 if (alloc == ld_alloc) { 506 return ac; 507 } 508 } 509 } 510 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 511 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 512 513 if (ac->is_arraycopy_validated() || 514 ac->is_copyof_validated() || 515 ac->is_copyofrange_validated()) { 516 Node* ld_addp = in(MemNode::Address); 517 if (ld_addp->is_AddP()) { 518 Node* ld_base = ld_addp->in(AddPNode::Address); 519 Node* ld_offs = ld_addp->in(AddPNode::Offset); 520 521 Node* dest = ac->in(ArrayCopyNode::Dest); 522 523 if (dest == ld_base) { 524 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 525 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) { 526 return ac; 527 } 528 if (!can_see_stored_value) { 529 mem = ac->in(TypeFunc::Memory); 530 } 531 } 532 } 533 } 534 } 535 return NULL; 536 } 537 538 // The logic for reordering loads and stores uses four steps: 539 // (a) Walk carefully past stores and initializations which we 540 // can prove are independent of this load. 541 // (b) Observe that the next memory state makes an exact match 542 // with self (load or store), and locate the relevant store. 543 // (c) Ensure that, if we were to wire self directly to the store, 544 // the optimizer would fold it up somehow. 545 // (d) Do the rewiring, and return, depending on some other part of 546 // the optimizer to fold up the load. 547 // This routine handles steps (a) and (b). Steps (c) and (d) are 548 // specific to loads and stores, so they are handled by the callers. 549 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 550 // 551 Node* MemNode::find_previous_store(PhaseTransform* phase) { 552 Node* ctrl = in(MemNode::Control); 553 Node* adr = in(MemNode::Address); 554 intptr_t offset = 0; 555 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 556 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 557 558 if (offset == Type::OffsetBot) 559 return NULL; // cannot unalias unless there are precise offsets 560 561 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 562 563 intptr_t size_in_bytes = memory_size(); 564 565 Node* mem = in(MemNode::Memory); // start searching here... 566 567 int cnt = 50; // Cycle limiter 568 for (;;) { // While we can dance past unrelated stores... 569 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 570 571 Node* prev = mem; 572 if (mem->is_Store()) { 573 Node* st_adr = mem->in(MemNode::Address); 574 intptr_t st_offset = 0; 575 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 576 if (st_base == NULL) 577 break; // inscrutable pointer 578 if (st_offset != offset && st_offset != Type::OffsetBot) { 579 const int MAX_STORE = BytesPerLong; 580 if (st_offset >= offset + size_in_bytes || 581 st_offset <= offset - MAX_STORE || 582 st_offset <= offset - mem->as_Store()->memory_size()) { 583 // Success: The offsets are provably independent. 584 // (You may ask, why not just test st_offset != offset and be done? 585 // The answer is that stores of different sizes can co-exist 586 // in the same sequence of RawMem effects. We sometimes initialize 587 // a whole 'tile' of array elements with a single jint or jlong.) 588 mem = mem->in(MemNode::Memory); 589 continue; // (a) advance through independent store memory 590 } 591 } 592 if (st_base != base && 593 detect_ptr_independence(base, alloc, 594 st_base, 595 AllocateNode::Ideal_allocation(st_base, phase), 596 phase)) { 597 // Success: The bases are provably independent. 598 mem = mem->in(MemNode::Memory); 599 continue; // (a) advance through independent store memory 600 } 601 602 // (b) At this point, if the bases or offsets do not agree, we lose, 603 // since we have not managed to prove 'this' and 'mem' independent. 604 if (st_base == base && st_offset == offset) { 605 return mem; // let caller handle steps (c), (d) 606 } 607 608 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 609 InitializeNode* st_init = mem->in(0)->as_Initialize(); 610 AllocateNode* st_alloc = st_init->allocation(); 611 if (st_alloc == NULL) 612 break; // something degenerated 613 bool known_identical = false; 614 bool known_independent = false; 615 if (alloc == st_alloc) 616 known_identical = true; 617 else if (alloc != NULL) 618 known_independent = true; 619 else if (all_controls_dominate(this, st_alloc)) 620 known_independent = true; 621 622 if (known_independent) { 623 // The bases are provably independent: Either they are 624 // manifestly distinct allocations, or else the control 625 // of this load dominates the store's allocation. 626 int alias_idx = phase->C->get_alias_index(adr_type()); 627 if (alias_idx == Compile::AliasIdxRaw) { 628 mem = st_alloc->in(TypeFunc::Memory); 629 } else { 630 mem = st_init->memory(alias_idx); 631 } 632 continue; // (a) advance through independent store memory 633 } 634 635 // (b) at this point, if we are not looking at a store initializing 636 // the same allocation we are loading from, we lose. 637 if (known_identical) { 638 // From caller, can_see_stored_value will consult find_captured_store. 639 return mem; // let caller handle steps (c), (d) 640 } 641 642 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 643 if (prev != mem) { 644 // Found an arraycopy but it doesn't affect that load 645 continue; 646 } 647 // Found an arraycopy that may affect that load 648 return mem; 649 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 650 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 651 if (mem->is_Proj() && mem->in(0)->is_Call()) { 652 // ArrayCopyNodes processed here as well. 653 CallNode *call = mem->in(0)->as_Call(); 654 if (!call->may_modify(addr_t, phase)) { 655 mem = call->in(TypeFunc::Memory); 656 continue; // (a) advance through independent call memory 657 } 658 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 659 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase)) { 660 break; 661 } 662 mem = mem->in(0)->in(TypeFunc::Memory); 663 continue; // (a) advance through independent MemBar memory 664 } else if (mem->is_ClearArray()) { 665 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 666 // (the call updated 'mem' value) 667 continue; // (a) advance through independent allocation memory 668 } else { 669 // Can not bypass initialization of the instance 670 // we are looking for. 671 return mem; 672 } 673 } else if (mem->is_MergeMem()) { 674 int alias_idx = phase->C->get_alias_index(adr_type()); 675 mem = mem->as_MergeMem()->memory_at(alias_idx); 676 continue; // (a) advance through independent MergeMem memory 677 } 678 } 679 680 // Unless there is an explicit 'continue', we must bail out here, 681 // because 'mem' is an inscrutable memory state (e.g., a call). 682 break; 683 } 684 685 return NULL; // bail out 686 } 687 688 //----------------------calculate_adr_type------------------------------------- 689 // Helper function. Notices when the given type of address hits top or bottom. 690 // Also, asserts a cross-check of the type against the expected address type. 691 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 692 if (t == Type::TOP) return NULL; // does not touch memory any more? 693 #ifdef PRODUCT 694 cross_check = NULL; 695 #else 696 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 697 #endif 698 const TypePtr* tp = t->isa_ptr(); 699 if (tp == NULL) { 700 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 701 return TypePtr::BOTTOM; // touches lots of memory 702 } else { 703 #ifdef ASSERT 704 // %%%% [phh] We don't check the alias index if cross_check is 705 // TypeRawPtr::BOTTOM. Needs to be investigated. 706 if (cross_check != NULL && 707 cross_check != TypePtr::BOTTOM && 708 cross_check != TypeRawPtr::BOTTOM) { 709 // Recheck the alias index, to see if it has changed (due to a bug). 710 Compile* C = Compile::current(); 711 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 712 "must stay in the original alias category"); 713 // The type of the address must be contained in the adr_type, 714 // disregarding "null"-ness. 715 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 716 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 717 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 718 "real address must not escape from expected memory type"); 719 } 720 #endif 721 return tp; 722 } 723 } 724 725 //============================================================================= 726 // Should LoadNode::Ideal() attempt to remove control edges? 727 bool LoadNode::can_remove_control() const { 728 return true; 729 } 730 uint LoadNode::size_of() const { return sizeof(*this); } 731 uint LoadNode::cmp( const Node &n ) const 732 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 733 const Type *LoadNode::bottom_type() const { return _type; } 734 uint LoadNode::ideal_reg() const { 735 return _type->ideal_reg(); 736 } 737 738 #ifndef PRODUCT 739 void LoadNode::dump_spec(outputStream *st) const { 740 MemNode::dump_spec(st); 741 if( !Verbose && !WizardMode ) { 742 // standard dump does this in Verbose and WizardMode 743 st->print(" #"); _type->dump_on(st); 744 } 745 if (!_depends_only_on_test) { 746 st->print(" (does not depend only on test)"); 747 } 748 } 749 #endif 750 751 #ifdef ASSERT 752 //----------------------------is_immutable_value------------------------------- 753 // Helper function to allow a raw load without control edge for some cases 754 bool LoadNode::is_immutable_value(Node* adr) { 755 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 756 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 757 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 758 in_bytes(JavaThread::osthread_offset()))); 759 } 760 #endif 761 762 //----------------------------LoadNode::make----------------------------------- 763 // Polymorphic factory method: 764 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, 765 ControlDependency control_dependency, bool unaligned, bool mismatched) { 766 Compile* C = gvn.C; 767 768 // sanity check the alias category against the created node type 769 assert(!(adr_type->isa_oopptr() && 770 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 771 "use LoadKlassNode instead"); 772 assert(!(adr_type->isa_aryptr() && 773 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 774 "use LoadRangeNode instead"); 775 // Check control edge of raw loads 776 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 777 // oop will be recorded in oop map if load crosses safepoint 778 rt->isa_oopptr() || is_immutable_value(adr), 779 "raw memory operations should have control edge"); 780 LoadNode* load = NULL; 781 switch (bt) { 782 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 783 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 784 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 785 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 786 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 787 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break; 788 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 789 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 790 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 791 case T_OBJECT: 792 #ifdef _LP64 793 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 794 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 795 } else 796 #endif 797 { 798 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 799 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo, control_dependency); 800 } 801 break; 802 } 803 assert(load != NULL, "LoadNode should have been created"); 804 if (unaligned) { 805 load->set_unaligned_access(); 806 } 807 if (mismatched) { 808 load->set_mismatched_access(); 809 } 810 if (load->Opcode() == Op_LoadN) { 811 Node* ld = gvn.transform(load); 812 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 813 } 814 815 return load; 816 } 817 818 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 819 ControlDependency control_dependency, bool unaligned, bool mismatched) { 820 bool require_atomic = true; 821 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 822 if (unaligned) { 823 load->set_unaligned_access(); 824 } 825 if (mismatched) { 826 load->set_mismatched_access(); 827 } 828 return load; 829 } 830 831 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, 832 ControlDependency control_dependency, bool unaligned, bool mismatched) { 833 bool require_atomic = true; 834 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 835 if (unaligned) { 836 load->set_unaligned_access(); 837 } 838 if (mismatched) { 839 load->set_mismatched_access(); 840 } 841 return load; 842 } 843 844 845 846 //------------------------------hash------------------------------------------- 847 uint LoadNode::hash() const { 848 // unroll addition of interesting fields 849 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 850 } 851 852 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 853 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 854 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 855 bool is_stable_ary = FoldStableValues && 856 (tp != NULL) && (tp->isa_aryptr() != NULL) && 857 tp->isa_aryptr()->is_stable(); 858 859 return (eliminate_boxing && non_volatile) || is_stable_ary; 860 } 861 862 return false; 863 } 864 865 // Is the value loaded previously stored by an arraycopy? If so return 866 // a load node that reads from the source array so we may be able to 867 // optimize out the ArrayCopy node later. 868 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const { 869 Node* ld_adr = in(MemNode::Address); 870 intptr_t ld_off = 0; 871 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 872 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 873 if (ac != NULL) { 874 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 875 876 Node* ld = clone(); 877 if (ac->as_ArrayCopy()->is_clonebasic()) { 878 assert(ld_alloc != NULL, "need an alloc"); 879 Node* addp = in(MemNode::Address)->clone(); 880 assert(addp->is_AddP(), "address must be addp"); 881 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern"); 882 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern"); 883 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base)); 884 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address)); 885 ld->set_req(MemNode::Address, phase->transform(addp)); 886 if (in(0) != NULL) { 887 assert(ld_alloc->in(0) != NULL, "alloc must have control"); 888 ld->set_req(0, ld_alloc->in(0)); 889 } 890 } else { 891 Node* addp = in(MemNode::Address)->clone(); 892 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 893 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)); 894 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)); 895 896 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 897 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 898 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 899 uint shift = exact_log2(type2aelembytes(ary_elem)); 900 901 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 902 #ifdef _LP64 903 diff = phase->transform(new ConvI2LNode(diff)); 904 #endif 905 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 906 907 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 908 addp->set_req(AddPNode::Offset, offset); 909 ld->set_req(MemNode::Address, phase->transform(addp)); 910 911 if (in(0) != NULL) { 912 assert(ac->in(0) != NULL, "alloc must have control"); 913 ld->set_req(0, ac->in(0)); 914 } 915 } 916 // load depends on the tests that validate the arraycopy 917 ld->as_Load()->_depends_only_on_test = Pinned; 918 return ld; 919 } 920 return NULL; 921 } 922 923 924 //---------------------------can_see_stored_value------------------------------ 925 // This routine exists to make sure this set of tests is done the same 926 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 927 // will change the graph shape in a way which makes memory alive twice at the 928 // same time (uses the Oracle model of aliasing), then some 929 // LoadXNode::Identity will fold things back to the equivalence-class model 930 // of aliasing. 931 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 932 Node* ld_adr = in(MemNode::Address); 933 intptr_t ld_off = 0; 934 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 935 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 936 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 937 // This is more general than load from boxing objects. 938 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 939 uint alias_idx = atp->index(); 940 bool final = !atp->is_rewritable(); 941 Node* result = NULL; 942 Node* current = st; 943 // Skip through chains of MemBarNodes checking the MergeMems for 944 // new states for the slice of this load. Stop once any other 945 // kind of node is encountered. Loads from final memory can skip 946 // through any kind of MemBar but normal loads shouldn't skip 947 // through MemBarAcquire since the could allow them to move out of 948 // a synchronized region. 949 while (current->is_Proj()) { 950 int opc = current->in(0)->Opcode(); 951 if ((final && (opc == Op_MemBarAcquire || 952 opc == Op_MemBarAcquireLock || 953 opc == Op_LoadFence)) || 954 opc == Op_MemBarRelease || 955 opc == Op_StoreFence || 956 opc == Op_MemBarReleaseLock || 957 opc == Op_MemBarStoreStore || 958 opc == Op_MemBarCPUOrder) { 959 Node* mem = current->in(0)->in(TypeFunc::Memory); 960 if (mem->is_MergeMem()) { 961 MergeMemNode* merge = mem->as_MergeMem(); 962 Node* new_st = merge->memory_at(alias_idx); 963 if (new_st == merge->base_memory()) { 964 // Keep searching 965 current = new_st; 966 continue; 967 } 968 // Save the new memory state for the slice and fall through 969 // to exit. 970 result = new_st; 971 } 972 } 973 break; 974 } 975 if (result != NULL) { 976 st = result; 977 } 978 } 979 980 // Loop around twice in the case Load -> Initialize -> Store. 981 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 982 for (int trip = 0; trip <= 1; trip++) { 983 984 if (st->is_Store()) { 985 Node* st_adr = st->in(MemNode::Address); 986 if (!phase->eqv(st_adr, ld_adr)) { 987 // Try harder before giving up... Match raw and non-raw pointers. 988 intptr_t st_off = 0; 989 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 990 if (alloc == NULL) return NULL; 991 if (alloc != ld_alloc) return NULL; 992 if (ld_off != st_off) return NULL; 993 // At this point we have proven something like this setup: 994 // A = Allocate(...) 995 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 996 // S = StoreQ(, AddP(, A.Parm , #Off), V) 997 // (Actually, we haven't yet proven the Q's are the same.) 998 // In other words, we are loading from a casted version of 999 // the same pointer-and-offset that we stored to. 1000 // Thus, we are able to replace L by V. 1001 } 1002 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1003 if (store_Opcode() != st->Opcode()) 1004 return NULL; 1005 return st->in(MemNode::ValueIn); 1006 } 1007 1008 // A load from a freshly-created object always returns zero. 1009 // (This can happen after LoadNode::Ideal resets the load's memory input 1010 // to find_captured_store, which returned InitializeNode::zero_memory.) 1011 if (st->is_Proj() && st->in(0)->is_Allocate() && 1012 (st->in(0) == ld_alloc) && 1013 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1014 // return a zero value for the load's basic type 1015 // (This is one of the few places where a generic PhaseTransform 1016 // can create new nodes. Think of it as lazily manifesting 1017 // virtually pre-existing constants.) 1018 return phase->zerocon(memory_type()); 1019 } 1020 1021 // A load from an initialization barrier can match a captured store. 1022 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1023 InitializeNode* init = st->in(0)->as_Initialize(); 1024 AllocateNode* alloc = init->allocation(); 1025 if ((alloc != NULL) && (alloc == ld_alloc)) { 1026 // examine a captured store value 1027 st = init->find_captured_store(ld_off, memory_size(), phase); 1028 if (st != NULL) { 1029 continue; // take one more trip around 1030 } 1031 } 1032 } 1033 1034 // Load boxed value from result of valueOf() call is input parameter. 1035 if (this->is_Load() && ld_adr->is_AddP() && 1036 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1037 intptr_t ignore = 0; 1038 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1039 if (base != NULL && base->is_Proj() && 1040 base->as_Proj()->_con == TypeFunc::Parms && 1041 base->in(0)->is_CallStaticJava() && 1042 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1043 return base->in(0)->in(TypeFunc::Parms); 1044 } 1045 } 1046 1047 break; 1048 } 1049 1050 return NULL; 1051 } 1052 1053 //----------------------is_instance_field_load_with_local_phi------------------ 1054 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1055 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1056 in(Address)->is_AddP() ) { 1057 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1058 // Only instances and boxed values. 1059 if( t_oop != NULL && 1060 (t_oop->is_ptr_to_boxed_value() || 1061 t_oop->is_known_instance_field()) && 1062 t_oop->offset() != Type::OffsetBot && 1063 t_oop->offset() != Type::OffsetTop) { 1064 return true; 1065 } 1066 } 1067 return false; 1068 } 1069 1070 //------------------------------Identity--------------------------------------- 1071 // Loads are identity if previous store is to same address 1072 Node* LoadNode::Identity(PhaseGVN* phase) { 1073 // If the previous store-maker is the right kind of Store, and the store is 1074 // to the same address, then we are equal to the value stored. 1075 Node* mem = in(Memory); 1076 Node* value = can_see_stored_value(mem, phase); 1077 if( value ) { 1078 // byte, short & char stores truncate naturally. 1079 // A load has to load the truncated value which requires 1080 // some sort of masking operation and that requires an 1081 // Ideal call instead of an Identity call. 1082 if (memory_size() < BytesPerInt) { 1083 // If the input to the store does not fit with the load's result type, 1084 // it must be truncated via an Ideal call. 1085 if (!phase->type(value)->higher_equal(phase->type(this))) 1086 return this; 1087 } 1088 // (This works even when value is a Con, but LoadNode::Value 1089 // usually runs first, producing the singleton type of the Con.) 1090 return value; 1091 } 1092 1093 // Search for an existing data phi which was generated before for the same 1094 // instance's field to avoid infinite generation of phis in a loop. 1095 Node *region = mem->in(0); 1096 if (is_instance_field_load_with_local_phi(region)) { 1097 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1098 int this_index = phase->C->get_alias_index(addr_t); 1099 int this_offset = addr_t->offset(); 1100 int this_iid = addr_t->instance_id(); 1101 if (!addr_t->is_known_instance() && 1102 addr_t->is_ptr_to_boxed_value()) { 1103 // Use _idx of address base (could be Phi node) for boxed values. 1104 intptr_t ignore = 0; 1105 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1106 this_iid = base->_idx; 1107 } 1108 const Type* this_type = bottom_type(); 1109 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1110 Node* phi = region->fast_out(i); 1111 if (phi->is_Phi() && phi != mem && 1112 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) { 1113 return phi; 1114 } 1115 } 1116 } 1117 1118 return this; 1119 } 1120 1121 // We're loading from an object which has autobox behaviour. 1122 // If this object is result of a valueOf call we'll have a phi 1123 // merging a newly allocated object and a load from the cache. 1124 // We want to replace this load with the original incoming 1125 // argument to the valueOf call. 1126 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1127 assert(phase->C->eliminate_boxing(), "sanity"); 1128 intptr_t ignore = 0; 1129 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1130 if ((base == NULL) || base->is_Phi()) { 1131 // Push the loads from the phi that comes from valueOf up 1132 // through it to allow elimination of the loads and the recovery 1133 // of the original value. It is done in split_through_phi(). 1134 return NULL; 1135 } else if (base->is_Load() || 1136 base->is_DecodeN() && base->in(1)->is_Load()) { 1137 // Eliminate the load of boxed value for integer types from the cache 1138 // array by deriving the value from the index into the array. 1139 // Capture the offset of the load and then reverse the computation. 1140 1141 // Get LoadN node which loads a boxing object from 'cache' array. 1142 if (base->is_DecodeN()) { 1143 base = base->in(1); 1144 } 1145 if (!base->in(Address)->is_AddP()) { 1146 return NULL; // Complex address 1147 } 1148 AddPNode* address = base->in(Address)->as_AddP(); 1149 Node* cache_base = address->in(AddPNode::Base); 1150 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1151 // Get ConP node which is static 'cache' field. 1152 cache_base = cache_base->in(1); 1153 } 1154 if ((cache_base != NULL) && cache_base->is_Con()) { 1155 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1156 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1157 Node* elements[4]; 1158 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1159 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1160 if ((count > 0) && elements[0]->is_Con() && 1161 ((count == 1) || 1162 (count == 2) && elements[1]->Opcode() == Op_LShiftX && 1163 elements[1]->in(2) == phase->intcon(shift))) { 1164 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1165 // Fetch the box object cache[0] at the base of the array and get its value 1166 ciInstance* box = array->obj_at(0)->as_instance(); 1167 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1168 assert(ik->is_box_klass(), "sanity"); 1169 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1170 if (ik->nof_nonstatic_fields() == 1) { 1171 // This should be true nonstatic_field_at requires calling 1172 // nof_nonstatic_fields so check it anyway 1173 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1174 BasicType bt = c.basic_type(); 1175 // Only integer types have boxing cache. 1176 assert(bt == T_BOOLEAN || bt == T_CHAR || 1177 bt == T_BYTE || bt == T_SHORT || 1178 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1179 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1180 if (cache_low != (int)cache_low) { 1181 return NULL; // should not happen since cache is array indexed by value 1182 } 1183 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1184 if (offset != (int)offset) { 1185 return NULL; // should not happen since cache is array indexed by value 1186 } 1187 // Add up all the offsets making of the address of the load 1188 Node* result = elements[0]; 1189 for (int i = 1; i < count; i++) { 1190 result = phase->transform(new AddXNode(result, elements[i])); 1191 } 1192 // Remove the constant offset from the address and then 1193 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1194 // remove the scaling of the offset to recover the original index. 1195 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1196 // Peel the shift off directly but wrap it in a dummy node 1197 // since Ideal can't return existing nodes 1198 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1199 } else if (result->is_Add() && result->in(2)->is_Con() && 1200 result->in(1)->Opcode() == Op_LShiftX && 1201 result->in(1)->in(2) == phase->intcon(shift)) { 1202 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1203 // but for boxing cache access we know that X<<Z will not overflow 1204 // (there is range check) so we do this optimizatrion by hand here. 1205 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1206 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1207 } else { 1208 result = new RShiftXNode(result, phase->intcon(shift)); 1209 } 1210 #ifdef _LP64 1211 if (bt != T_LONG) { 1212 result = new ConvL2INode(phase->transform(result)); 1213 } 1214 #else 1215 if (bt == T_LONG) { 1216 result = new ConvI2LNode(phase->transform(result)); 1217 } 1218 #endif 1219 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1220 // Need to preserve unboxing load type if it is unsigned. 1221 switch(this->Opcode()) { 1222 case Op_LoadUB: 1223 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1224 break; 1225 case Op_LoadUS: 1226 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1227 break; 1228 } 1229 return result; 1230 } 1231 } 1232 } 1233 } 1234 } 1235 return NULL; 1236 } 1237 1238 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1239 Node* region = phi->in(0); 1240 if (region == NULL) { 1241 return false; // Wait stable graph 1242 } 1243 uint cnt = phi->req(); 1244 for (uint i = 1; i < cnt; i++) { 1245 Node* rc = region->in(i); 1246 if (rc == NULL || phase->type(rc) == Type::TOP) 1247 return false; // Wait stable graph 1248 Node* in = phi->in(i); 1249 if (in == NULL || phase->type(in) == Type::TOP) 1250 return false; // Wait stable graph 1251 } 1252 return true; 1253 } 1254 //------------------------------split_through_phi------------------------------ 1255 // Split instance or boxed field load through Phi. 1256 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1257 Node* mem = in(Memory); 1258 Node* address = in(Address); 1259 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1260 1261 assert((t_oop != NULL) && 1262 (t_oop->is_known_instance_field() || 1263 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1264 1265 Compile* C = phase->C; 1266 intptr_t ignore = 0; 1267 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1268 bool base_is_phi = (base != NULL) && base->is_Phi(); 1269 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1270 (base != NULL) && (base == address->in(AddPNode::Base)) && 1271 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1272 1273 if (!((mem->is_Phi() || base_is_phi) && 1274 (load_boxed_values || t_oop->is_known_instance_field()))) { 1275 return NULL; // memory is not Phi 1276 } 1277 1278 if (mem->is_Phi()) { 1279 if (!stable_phi(mem->as_Phi(), phase)) { 1280 return NULL; // Wait stable graph 1281 } 1282 uint cnt = mem->req(); 1283 // Check for loop invariant memory. 1284 if (cnt == 3) { 1285 for (uint i = 1; i < cnt; i++) { 1286 Node* in = mem->in(i); 1287 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1288 if (m == mem) { 1289 set_req(Memory, mem->in(cnt - i)); 1290 return this; // made change 1291 } 1292 } 1293 } 1294 } 1295 if (base_is_phi) { 1296 if (!stable_phi(base->as_Phi(), phase)) { 1297 return NULL; // Wait stable graph 1298 } 1299 uint cnt = base->req(); 1300 // Check for loop invariant memory. 1301 if (cnt == 3) { 1302 for (uint i = 1; i < cnt; i++) { 1303 if (base->in(i) == base) { 1304 return NULL; // Wait stable graph 1305 } 1306 } 1307 } 1308 } 1309 1310 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1311 1312 // Split through Phi (see original code in loopopts.cpp). 1313 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1314 1315 // Do nothing here if Identity will find a value 1316 // (to avoid infinite chain of value phis generation). 1317 if (!phase->eqv(this, this->Identity(phase))) 1318 return NULL; 1319 1320 // Select Region to split through. 1321 Node* region; 1322 if (!base_is_phi) { 1323 assert(mem->is_Phi(), "sanity"); 1324 region = mem->in(0); 1325 // Skip if the region dominates some control edge of the address. 1326 if (!MemNode::all_controls_dominate(address, region)) 1327 return NULL; 1328 } else if (!mem->is_Phi()) { 1329 assert(base_is_phi, "sanity"); 1330 region = base->in(0); 1331 // Skip if the region dominates some control edge of the memory. 1332 if (!MemNode::all_controls_dominate(mem, region)) 1333 return NULL; 1334 } else if (base->in(0) != mem->in(0)) { 1335 assert(base_is_phi && mem->is_Phi(), "sanity"); 1336 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1337 region = base->in(0); 1338 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1339 region = mem->in(0); 1340 } else { 1341 return NULL; // complex graph 1342 } 1343 } else { 1344 assert(base->in(0) == mem->in(0), "sanity"); 1345 region = mem->in(0); 1346 } 1347 1348 const Type* this_type = this->bottom_type(); 1349 int this_index = C->get_alias_index(t_oop); 1350 int this_offset = t_oop->offset(); 1351 int this_iid = t_oop->instance_id(); 1352 if (!t_oop->is_known_instance() && load_boxed_values) { 1353 // Use _idx of address base for boxed values. 1354 this_iid = base->_idx; 1355 } 1356 PhaseIterGVN* igvn = phase->is_IterGVN(); 1357 Node* phi = new PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1358 for (uint i = 1; i < region->req(); i++) { 1359 Node* x; 1360 Node* the_clone = NULL; 1361 if (region->in(i) == C->top()) { 1362 x = C->top(); // Dead path? Use a dead data op 1363 } else { 1364 x = this->clone(); // Else clone up the data op 1365 the_clone = x; // Remember for possible deletion. 1366 // Alter data node to use pre-phi inputs 1367 if (this->in(0) == region) { 1368 x->set_req(0, region->in(i)); 1369 } else { 1370 x->set_req(0, NULL); 1371 } 1372 if (mem->is_Phi() && (mem->in(0) == region)) { 1373 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1374 } 1375 if (address->is_Phi() && address->in(0) == region) { 1376 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1377 } 1378 if (base_is_phi && (base->in(0) == region)) { 1379 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1380 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1381 x->set_req(Address, adr_x); 1382 } 1383 } 1384 // Check for a 'win' on some paths 1385 const Type *t = x->Value(igvn); 1386 1387 bool singleton = t->singleton(); 1388 1389 // See comments in PhaseIdealLoop::split_thru_phi(). 1390 if (singleton && t == Type::TOP) { 1391 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1392 } 1393 1394 if (singleton) { 1395 x = igvn->makecon(t); 1396 } else { 1397 // We now call Identity to try to simplify the cloned node. 1398 // Note that some Identity methods call phase->type(this). 1399 // Make sure that the type array is big enough for 1400 // our new node, even though we may throw the node away. 1401 // (This tweaking with igvn only works because x is a new node.) 1402 igvn->set_type(x, t); 1403 // If x is a TypeNode, capture any more-precise type permanently into Node 1404 // otherwise it will be not updated during igvn->transform since 1405 // igvn->type(x) is set to x->Value() already. 1406 x->raise_bottom_type(t); 1407 Node *y = x->Identity(igvn); 1408 if (y != x) { 1409 x = y; 1410 } else { 1411 y = igvn->hash_find_insert(x); 1412 if (y) { 1413 x = y; 1414 } else { 1415 // Else x is a new node we are keeping 1416 // We do not need register_new_node_with_optimizer 1417 // because set_type has already been called. 1418 igvn->_worklist.push(x); 1419 } 1420 } 1421 } 1422 if (x != the_clone && the_clone != NULL) { 1423 igvn->remove_dead_node(the_clone); 1424 } 1425 phi->set_req(i, x); 1426 } 1427 // Record Phi 1428 igvn->register_new_node_with_optimizer(phi); 1429 return phi; 1430 } 1431 1432 //------------------------------Ideal------------------------------------------ 1433 // If the load is from Field memory and the pointer is non-null, it might be possible to 1434 // zero out the control input. 1435 // If the offset is constant and the base is an object allocation, 1436 // try to hook me up to the exact initializing store. 1437 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1438 Node* p = MemNode::Ideal_common(phase, can_reshape); 1439 if (p) return (p == NodeSentinel) ? NULL : p; 1440 1441 Node* ctrl = in(MemNode::Control); 1442 Node* address = in(MemNode::Address); 1443 bool progress = false; 1444 1445 // Skip up past a SafePoint control. Cannot do this for Stores because 1446 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1447 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1448 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1449 ctrl = ctrl->in(0); 1450 set_req(MemNode::Control,ctrl); 1451 progress = true; 1452 } 1453 1454 intptr_t ignore = 0; 1455 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1456 if (base != NULL 1457 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1458 // Check for useless control edge in some common special cases 1459 if (in(MemNode::Control) != NULL 1460 && can_remove_control() 1461 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1462 && all_controls_dominate(base, phase->C->start())) { 1463 // A method-invariant, non-null address (constant or 'this' argument). 1464 set_req(MemNode::Control, NULL); 1465 progress = true; 1466 } 1467 } 1468 1469 Node* mem = in(MemNode::Memory); 1470 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1471 1472 if (can_reshape && (addr_t != NULL)) { 1473 // try to optimize our memory input 1474 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1475 if (opt_mem != mem) { 1476 set_req(MemNode::Memory, opt_mem); 1477 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1478 return this; 1479 } 1480 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1481 if ((t_oop != NULL) && 1482 (t_oop->is_known_instance_field() || 1483 t_oop->is_ptr_to_boxed_value())) { 1484 PhaseIterGVN *igvn = phase->is_IterGVN(); 1485 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1486 // Delay this transformation until memory Phi is processed. 1487 phase->is_IterGVN()->_worklist.push(this); 1488 return NULL; 1489 } 1490 // Split instance field load through Phi. 1491 Node* result = split_through_phi(phase); 1492 if (result != NULL) return result; 1493 1494 if (t_oop->is_ptr_to_boxed_value()) { 1495 Node* result = eliminate_autobox(phase); 1496 if (result != NULL) return result; 1497 } 1498 } 1499 } 1500 1501 // Is there a dominating load that loads the same value? Leave 1502 // anything that is not a load of a field/array element (like 1503 // barriers etc.) alone 1504 if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) { 1505 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1506 Node *use = mem->fast_out(i); 1507 if (use != this && 1508 use->Opcode() == Opcode() && 1509 use->in(0) != NULL && 1510 use->in(0) != in(0) && 1511 use->in(Address) == in(Address)) { 1512 Node* ctl = in(0); 1513 for (int i = 0; i < 10 && ctl != NULL; i++) { 1514 ctl = IfNode::up_one_dom(ctl); 1515 if (ctl == use->in(0)) { 1516 set_req(0, use->in(0)); 1517 return this; 1518 } 1519 } 1520 } 1521 } 1522 } 1523 1524 // Check for prior store with a different base or offset; make Load 1525 // independent. Skip through any number of them. Bail out if the stores 1526 // are in an endless dead cycle and report no progress. This is a key 1527 // transform for Reflection. However, if after skipping through the Stores 1528 // we can't then fold up against a prior store do NOT do the transform as 1529 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1530 // array memory alive twice: once for the hoisted Load and again after the 1531 // bypassed Store. This situation only works if EVERYBODY who does 1532 // anti-dependence work knows how to bypass. I.e. we need all 1533 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1534 // the alias index stuff. So instead, peek through Stores and IFF we can 1535 // fold up, do so. 1536 Node* prev_mem = find_previous_store(phase); 1537 if (prev_mem != NULL) { 1538 Node* value = can_see_arraycopy_value(prev_mem, phase); 1539 if (value != NULL) { 1540 return value; 1541 } 1542 } 1543 // Steps (a), (b): Walk past independent stores to find an exact match. 1544 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1545 // (c) See if we can fold up on the spot, but don't fold up here. 1546 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1547 // just return a prior value, which is done by Identity calls. 1548 if (can_see_stored_value(prev_mem, phase)) { 1549 // Make ready for step (d): 1550 set_req(MemNode::Memory, prev_mem); 1551 return this; 1552 } 1553 } 1554 1555 return progress ? this : NULL; 1556 } 1557 1558 // Helper to recognize certain Klass fields which are invariant across 1559 // some group of array types (e.g., int[] or all T[] where T < Object). 1560 const Type* 1561 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1562 ciKlass* klass) const { 1563 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1564 // The field is Klass::_modifier_flags. Return its (constant) value. 1565 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1566 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1567 return TypeInt::make(klass->modifier_flags()); 1568 } 1569 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1570 // The field is Klass::_access_flags. Return its (constant) value. 1571 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1572 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1573 return TypeInt::make(klass->access_flags()); 1574 } 1575 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1576 // The field is Klass::_layout_helper. Return its constant value if known. 1577 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1578 return TypeInt::make(klass->layout_helper()); 1579 } 1580 1581 // No match. 1582 return NULL; 1583 } 1584 1585 static ciConstant check_mismatched_access(ciConstant con, BasicType loadbt, bool is_unsigned) { 1586 BasicType conbt = con.basic_type(); 1587 switch (conbt) { 1588 case T_BOOLEAN: conbt = T_BYTE; break; 1589 case T_ARRAY: conbt = T_OBJECT; break; 1590 } 1591 switch (loadbt) { 1592 case T_BOOLEAN: loadbt = T_BYTE; break; 1593 case T_NARROWOOP: loadbt = T_OBJECT; break; 1594 case T_ARRAY: loadbt = T_OBJECT; break; 1595 case T_ADDRESS: loadbt = T_OBJECT; break; 1596 } 1597 if (conbt == loadbt) { 1598 if (is_unsigned && conbt == T_BYTE) { 1599 // LoadB (T_BYTE) with a small mask (<=8-bit) is converted to LoadUB (T_BYTE). 1600 return ciConstant(T_INT, con.as_int() & 0xFF); 1601 } else { 1602 return con; 1603 } 1604 } 1605 if (conbt == T_SHORT && loadbt == T_CHAR) { 1606 // LoadS (T_SHORT) with a small mask (<=16-bit) is converted to LoadUS (T_CHAR). 1607 return ciConstant(T_INT, con.as_int() & 0xFFFF); 1608 } 1609 return ciConstant(); // T_ILLEGAL 1610 } 1611 1612 // Try to constant-fold a stable array element. 1613 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, bool is_unsigned_load, BasicType loadbt) { 1614 assert(ary->const_oop(), "array should be constant"); 1615 assert(ary->is_stable(), "array should be stable"); 1616 1617 // Decode the results of GraphKit::array_element_address. 1618 ciArray* aobj = ary->const_oop()->as_array(); 1619 ciConstant element_value = aobj->element_value_by_offset(off); 1620 if (element_value.basic_type() == T_ILLEGAL) { 1621 return NULL; // wrong offset 1622 } 1623 ciConstant con = check_mismatched_access(element_value, loadbt, is_unsigned_load); 1624 assert(con.basic_type() != T_ILLEGAL, "elembt=%s; loadbt=%s; unsigned=%d", 1625 type2name(element_value.basic_type()), type2name(loadbt), is_unsigned_load); 1626 1627 if (con.basic_type() != T_ILLEGAL && // not a mismatched access 1628 !con.is_null_or_zero()) { // not a default value 1629 const Type* con_type = Type::make_from_constant(con); 1630 if (con_type != NULL) { 1631 if (con_type->isa_aryptr()) { 1632 // Join with the array element type, in case it is also stable. 1633 int dim = ary->stable_dimension(); 1634 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1); 1635 } 1636 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) { 1637 con_type = con_type->make_narrowoop(); 1638 } 1639 #ifndef PRODUCT 1640 if (TraceIterativeGVN) { 1641 tty->print("FoldStableValues: array element [off=%d]: con_type=", off); 1642 con_type->dump(); tty->cr(); 1643 } 1644 #endif //PRODUCT 1645 return con_type; 1646 } 1647 } 1648 return NULL; 1649 } 1650 1651 //------------------------------Value----------------------------------------- 1652 const Type* LoadNode::Value(PhaseGVN* phase) const { 1653 // Either input is TOP ==> the result is TOP 1654 Node* mem = in(MemNode::Memory); 1655 const Type *t1 = phase->type(mem); 1656 if (t1 == Type::TOP) return Type::TOP; 1657 Node* adr = in(MemNode::Address); 1658 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1659 if (tp == NULL || tp->empty()) return Type::TOP; 1660 int off = tp->offset(); 1661 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1662 Compile* C = phase->C; 1663 1664 // Try to guess loaded type from pointer type 1665 if (tp->isa_aryptr()) { 1666 const TypeAryPtr* ary = tp->is_aryptr(); 1667 const Type* t = ary->elem(); 1668 1669 // Determine whether the reference is beyond the header or not, by comparing 1670 // the offset against the offset of the start of the array's data. 1671 // Different array types begin at slightly different offsets (12 vs. 16). 1672 // We choose T_BYTE as an example base type that is least restrictive 1673 // as to alignment, which will therefore produce the smallest 1674 // possible base offset. 1675 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1676 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1677 1678 // Try to constant-fold a stable array element. 1679 if (FoldStableValues && !is_mismatched_access() && ary->is_stable() && ary->const_oop() != NULL) { 1680 // Make sure the reference is not into the header and the offset is constant 1681 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1682 const Type* con_type = fold_stable_ary_elem(ary, off, is_unsigned(), memory_type()); 1683 if (con_type != NULL) { 1684 return con_type; 1685 } 1686 } 1687 } 1688 1689 // Don't do this for integer types. There is only potential profit if 1690 // the element type t is lower than _type; that is, for int types, if _type is 1691 // more restrictive than t. This only happens here if one is short and the other 1692 // char (both 16 bits), and in those cases we've made an intentional decision 1693 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1694 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1695 // 1696 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1697 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1698 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1699 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1700 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1701 // In fact, that could have been the original type of p1, and p1 could have 1702 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1703 // expression (LShiftL quux 3) independently optimized to the constant 8. 1704 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1705 && (_type->isa_vect() == NULL) 1706 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1707 // t might actually be lower than _type, if _type is a unique 1708 // concrete subclass of abstract class t. 1709 if (off_beyond_header) { // is the offset beyond the header? 1710 const Type* jt = t->join_speculative(_type); 1711 // In any case, do not allow the join, per se, to empty out the type. 1712 if (jt->empty() && !t->empty()) { 1713 // This can happen if a interface-typed array narrows to a class type. 1714 jt = _type; 1715 } 1716 #ifdef ASSERT 1717 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1718 // The pointers in the autobox arrays are always non-null 1719 Node* base = adr->in(AddPNode::Base); 1720 if ((base != NULL) && base->is_DecodeN()) { 1721 // Get LoadN node which loads IntegerCache.cache field 1722 base = base->in(1); 1723 } 1724 if ((base != NULL) && base->is_Con()) { 1725 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1726 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1727 // It could be narrow oop 1728 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1729 } 1730 } 1731 } 1732 #endif 1733 return jt; 1734 } 1735 } 1736 } else if (tp->base() == Type::InstPtr) { 1737 ciEnv* env = C->env(); 1738 const TypeInstPtr* tinst = tp->is_instptr(); 1739 ciKlass* klass = tinst->klass(); 1740 assert( off != Type::OffsetBot || 1741 // arrays can be cast to Objects 1742 tp->is_oopptr()->klass()->is_java_lang_Object() || 1743 // unsafe field access may not have a constant offset 1744 C->has_unsafe_access(), 1745 "Field accesses must be precise" ); 1746 // For oop loads, we expect the _type to be precise. 1747 // Optimizations for constant objects 1748 ciObject* const_oop = tinst->const_oop(); 1749 if (const_oop != NULL) { 1750 // For constant CallSites treat the target field as a compile time constant. 1751 if (const_oop->is_call_site()) { 1752 ciCallSite* call_site = const_oop->as_call_site(); 1753 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1754 if (field != NULL && field->is_call_site_target()) { 1755 ciMethodHandle* target = call_site->get_target(); 1756 if (target != NULL) { // just in case 1757 ciConstant constant(T_OBJECT, target); 1758 const Type* t; 1759 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1760 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1761 } else { 1762 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1763 } 1764 // Add a dependence for invalidation of the optimization. 1765 if (!call_site->is_constant_call_site()) { 1766 C->dependencies()->assert_call_site_target_value(call_site, target); 1767 } 1768 return t; 1769 } 1770 } 1771 } 1772 } 1773 } else if (tp->base() == Type::KlassPtr) { 1774 assert( off != Type::OffsetBot || 1775 // arrays can be cast to Objects 1776 tp->is_klassptr()->klass()->is_java_lang_Object() || 1777 // also allow array-loading from the primary supertype 1778 // array during subtype checks 1779 Opcode() == Op_LoadKlass, 1780 "Field accesses must be precise" ); 1781 // For klass/static loads, we expect the _type to be precise 1782 } 1783 1784 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1785 if (tkls != NULL && !StressReflectiveCode) { 1786 ciKlass* klass = tkls->klass(); 1787 if (klass->is_loaded() && tkls->klass_is_exact()) { 1788 // We are loading a field from a Klass metaobject whose identity 1789 // is known at compile time (the type is "exact" or "precise"). 1790 // Check for fields we know are maintained as constants by the VM. 1791 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1792 // The field is Klass::_super_check_offset. Return its (constant) value. 1793 // (Folds up type checking code.) 1794 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1795 return TypeInt::make(klass->super_check_offset()); 1796 } 1797 // Compute index into primary_supers array 1798 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1799 // Check for overflowing; use unsigned compare to handle the negative case. 1800 if( depth < ciKlass::primary_super_limit() ) { 1801 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1802 // (Folds up type checking code.) 1803 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1804 ciKlass *ss = klass->super_of_depth(depth); 1805 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1806 } 1807 const Type* aift = load_array_final_field(tkls, klass); 1808 if (aift != NULL) return aift; 1809 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1810 // The field is Klass::_java_mirror. Return its (constant) value. 1811 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1812 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1813 return TypeInstPtr::make(klass->java_mirror()); 1814 } 1815 } 1816 1817 // We can still check if we are loading from the primary_supers array at a 1818 // shallow enough depth. Even though the klass is not exact, entries less 1819 // than or equal to its super depth are correct. 1820 if (klass->is_loaded() ) { 1821 ciType *inner = klass; 1822 while( inner->is_obj_array_klass() ) 1823 inner = inner->as_obj_array_klass()->base_element_type(); 1824 if( inner->is_instance_klass() && 1825 !inner->as_instance_klass()->flags().is_interface() ) { 1826 // Compute index into primary_supers array 1827 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1828 // Check for overflowing; use unsigned compare to handle the negative case. 1829 if( depth < ciKlass::primary_super_limit() && 1830 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1831 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1832 // (Folds up type checking code.) 1833 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1834 ciKlass *ss = klass->super_of_depth(depth); 1835 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1836 } 1837 } 1838 } 1839 1840 // If the type is enough to determine that the thing is not an array, 1841 // we can give the layout_helper a positive interval type. 1842 // This will help short-circuit some reflective code. 1843 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1844 && !klass->is_array_klass() // not directly typed as an array 1845 && !klass->is_interface() // specifically not Serializable & Cloneable 1846 && !klass->is_java_lang_Object() // not the supertype of all T[] 1847 ) { 1848 // Note: When interfaces are reliable, we can narrow the interface 1849 // test to (klass != Serializable && klass != Cloneable). 1850 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1851 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1852 // The key property of this type is that it folds up tests 1853 // for array-ness, since it proves that the layout_helper is positive. 1854 // Thus, a generic value like the basic object layout helper works fine. 1855 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1856 } 1857 } 1858 1859 // If we are loading from a freshly-allocated object, produce a zero, 1860 // if the load is provably beyond the header of the object. 1861 // (Also allow a variable load from a fresh array to produce zero.) 1862 const TypeOopPtr *tinst = tp->isa_oopptr(); 1863 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1864 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1865 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1866 Node* value = can_see_stored_value(mem,phase); 1867 if (value != NULL && value->is_Con()) { 1868 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1869 return value->bottom_type(); 1870 } 1871 } 1872 1873 if (is_instance) { 1874 // If we have an instance type and our memory input is the 1875 // programs's initial memory state, there is no matching store, 1876 // so just return a zero of the appropriate type 1877 Node *mem = in(MemNode::Memory); 1878 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1879 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1880 return Type::get_zero_type(_type->basic_type()); 1881 } 1882 } 1883 return _type; 1884 } 1885 1886 //------------------------------match_edge------------------------------------- 1887 // Do we Match on this edge index or not? Match only the address. 1888 uint LoadNode::match_edge(uint idx) const { 1889 return idx == MemNode::Address; 1890 } 1891 1892 //--------------------------LoadBNode::Ideal-------------------------------------- 1893 // 1894 // If the previous store is to the same address as this load, 1895 // and the value stored was larger than a byte, replace this load 1896 // with the value stored truncated to a byte. If no truncation is 1897 // needed, the replacement is done in LoadNode::Identity(). 1898 // 1899 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1900 Node* mem = in(MemNode::Memory); 1901 Node* value = can_see_stored_value(mem,phase); 1902 if( value && !phase->type(value)->higher_equal( _type ) ) { 1903 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1904 return new RShiftINode(result, phase->intcon(24)); 1905 } 1906 // Identity call will handle the case where truncation is not needed. 1907 return LoadNode::Ideal(phase, can_reshape); 1908 } 1909 1910 const Type* LoadBNode::Value(PhaseGVN* phase) const { 1911 Node* mem = in(MemNode::Memory); 1912 Node* value = can_see_stored_value(mem,phase); 1913 if (value != NULL && value->is_Con() && 1914 !value->bottom_type()->higher_equal(_type)) { 1915 // If the input to the store does not fit with the load's result type, 1916 // it must be truncated. We can't delay until Ideal call since 1917 // a singleton Value is needed for split_thru_phi optimization. 1918 int con = value->get_int(); 1919 return TypeInt::make((con << 24) >> 24); 1920 } 1921 return LoadNode::Value(phase); 1922 } 1923 1924 //--------------------------LoadUBNode::Ideal------------------------------------- 1925 // 1926 // If the previous store is to the same address as this load, 1927 // and the value stored was larger than a byte, replace this load 1928 // with the value stored truncated to a byte. If no truncation is 1929 // needed, the replacement is done in LoadNode::Identity(). 1930 // 1931 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1932 Node* mem = in(MemNode::Memory); 1933 Node* value = can_see_stored_value(mem, phase); 1934 if (value && !phase->type(value)->higher_equal(_type)) 1935 return new AndINode(value, phase->intcon(0xFF)); 1936 // Identity call will handle the case where truncation is not needed. 1937 return LoadNode::Ideal(phase, can_reshape); 1938 } 1939 1940 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 1941 Node* mem = in(MemNode::Memory); 1942 Node* value = can_see_stored_value(mem,phase); 1943 if (value != NULL && value->is_Con() && 1944 !value->bottom_type()->higher_equal(_type)) { 1945 // If the input to the store does not fit with the load's result type, 1946 // it must be truncated. We can't delay until Ideal call since 1947 // a singleton Value is needed for split_thru_phi optimization. 1948 int con = value->get_int(); 1949 return TypeInt::make(con & 0xFF); 1950 } 1951 return LoadNode::Value(phase); 1952 } 1953 1954 //--------------------------LoadUSNode::Ideal------------------------------------- 1955 // 1956 // If the previous store is to the same address as this load, 1957 // and the value stored was larger than a char, replace this load 1958 // with the value stored truncated to a char. If no truncation is 1959 // needed, the replacement is done in LoadNode::Identity(). 1960 // 1961 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1962 Node* mem = in(MemNode::Memory); 1963 Node* value = can_see_stored_value(mem,phase); 1964 if( value && !phase->type(value)->higher_equal( _type ) ) 1965 return new AndINode(value,phase->intcon(0xFFFF)); 1966 // Identity call will handle the case where truncation is not needed. 1967 return LoadNode::Ideal(phase, can_reshape); 1968 } 1969 1970 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 1971 Node* mem = in(MemNode::Memory); 1972 Node* value = can_see_stored_value(mem,phase); 1973 if (value != NULL && value->is_Con() && 1974 !value->bottom_type()->higher_equal(_type)) { 1975 // If the input to the store does not fit with the load's result type, 1976 // it must be truncated. We can't delay until Ideal call since 1977 // a singleton Value is needed for split_thru_phi optimization. 1978 int con = value->get_int(); 1979 return TypeInt::make(con & 0xFFFF); 1980 } 1981 return LoadNode::Value(phase); 1982 } 1983 1984 //--------------------------LoadSNode::Ideal-------------------------------------- 1985 // 1986 // If the previous store is to the same address as this load, 1987 // and the value stored was larger than a short, replace this load 1988 // with the value stored truncated to a short. If no truncation is 1989 // needed, the replacement is done in LoadNode::Identity(). 1990 // 1991 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1992 Node* mem = in(MemNode::Memory); 1993 Node* value = can_see_stored_value(mem,phase); 1994 if( value && !phase->type(value)->higher_equal( _type ) ) { 1995 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 1996 return new RShiftINode(result, phase->intcon(16)); 1997 } 1998 // Identity call will handle the case where truncation is not needed. 1999 return LoadNode::Ideal(phase, can_reshape); 2000 } 2001 2002 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2003 Node* mem = in(MemNode::Memory); 2004 Node* value = can_see_stored_value(mem,phase); 2005 if (value != NULL && value->is_Con() && 2006 !value->bottom_type()->higher_equal(_type)) { 2007 // If the input to the store does not fit with the load's result type, 2008 // it must be truncated. We can't delay until Ideal call since 2009 // a singleton Value is needed for split_thru_phi optimization. 2010 int con = value->get_int(); 2011 return TypeInt::make((con << 16) >> 16); 2012 } 2013 return LoadNode::Value(phase); 2014 } 2015 2016 //============================================================================= 2017 //----------------------------LoadKlassNode::make------------------------------ 2018 // Polymorphic factory method: 2019 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2020 // sanity check the alias category against the created node type 2021 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2022 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2023 #ifdef _LP64 2024 if (adr_type->is_ptr_to_narrowklass()) { 2025 assert(UseCompressedClassPointers, "no compressed klasses"); 2026 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2027 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2028 } 2029 #endif 2030 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2031 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2032 } 2033 2034 //------------------------------Value------------------------------------------ 2035 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2036 return klass_value_common(phase); 2037 } 2038 2039 // In most cases, LoadKlassNode does not have the control input set. If the control 2040 // input is set, it must not be removed (by LoadNode::Ideal()). 2041 bool LoadKlassNode::can_remove_control() const { 2042 return false; 2043 } 2044 2045 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2046 // Either input is TOP ==> the result is TOP 2047 const Type *t1 = phase->type( in(MemNode::Memory) ); 2048 if (t1 == Type::TOP) return Type::TOP; 2049 Node *adr = in(MemNode::Address); 2050 const Type *t2 = phase->type( adr ); 2051 if (t2 == Type::TOP) return Type::TOP; 2052 const TypePtr *tp = t2->is_ptr(); 2053 if (TypePtr::above_centerline(tp->ptr()) || 2054 tp->ptr() == TypePtr::Null) return Type::TOP; 2055 2056 // Return a more precise klass, if possible 2057 const TypeInstPtr *tinst = tp->isa_instptr(); 2058 if (tinst != NULL) { 2059 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2060 int offset = tinst->offset(); 2061 if (ik == phase->C->env()->Class_klass() 2062 && (offset == java_lang_Class::klass_offset_in_bytes() || 2063 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2064 // We are loading a special hidden field from a Class mirror object, 2065 // the field which points to the VM's Klass metaobject. 2066 ciType* t = tinst->java_mirror_type(); 2067 // java_mirror_type returns non-null for compile-time Class constants. 2068 if (t != NULL) { 2069 // constant oop => constant klass 2070 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2071 if (t->is_void()) { 2072 // We cannot create a void array. Since void is a primitive type return null 2073 // klass. Users of this result need to do a null check on the returned klass. 2074 return TypePtr::NULL_PTR; 2075 } 2076 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2077 } 2078 if (!t->is_klass()) { 2079 // a primitive Class (e.g., int.class) has NULL for a klass field 2080 return TypePtr::NULL_PTR; 2081 } 2082 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2083 return TypeKlassPtr::make(t->as_klass()); 2084 } 2085 // non-constant mirror, so we can't tell what's going on 2086 } 2087 if( !ik->is_loaded() ) 2088 return _type; // Bail out if not loaded 2089 if (offset == oopDesc::klass_offset_in_bytes()) { 2090 if (tinst->klass_is_exact()) { 2091 return TypeKlassPtr::make(ik); 2092 } 2093 // See if we can become precise: no subklasses and no interface 2094 // (Note: We need to support verified interfaces.) 2095 if (!ik->is_interface() && !ik->has_subklass()) { 2096 //assert(!UseExactTypes, "this code should be useless with exact types"); 2097 // Add a dependence; if any subclass added we need to recompile 2098 if (!ik->is_final()) { 2099 // %%% should use stronger assert_unique_concrete_subtype instead 2100 phase->C->dependencies()->assert_leaf_type(ik); 2101 } 2102 // Return precise klass 2103 return TypeKlassPtr::make(ik); 2104 } 2105 2106 // Return root of possible klass 2107 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2108 } 2109 } 2110 2111 // Check for loading klass from an array 2112 const TypeAryPtr *tary = tp->isa_aryptr(); 2113 if( tary != NULL ) { 2114 ciKlass *tary_klass = tary->klass(); 2115 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2116 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2117 if (tary->klass_is_exact()) { 2118 return TypeKlassPtr::make(tary_klass); 2119 } 2120 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2121 // If the klass is an object array, we defer the question to the 2122 // array component klass. 2123 if( ak->is_obj_array_klass() ) { 2124 assert( ak->is_loaded(), "" ); 2125 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2126 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2127 ciInstanceKlass* ik = base_k->as_instance_klass(); 2128 // See if we can become precise: no subklasses and no interface 2129 if (!ik->is_interface() && !ik->has_subklass()) { 2130 //assert(!UseExactTypes, "this code should be useless with exact types"); 2131 // Add a dependence; if any subclass added we need to recompile 2132 if (!ik->is_final()) { 2133 phase->C->dependencies()->assert_leaf_type(ik); 2134 } 2135 // Return precise array klass 2136 return TypeKlassPtr::make(ak); 2137 } 2138 } 2139 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2140 } else { // Found a type-array? 2141 //assert(!UseExactTypes, "this code should be useless with exact types"); 2142 assert( ak->is_type_array_klass(), "" ); 2143 return TypeKlassPtr::make(ak); // These are always precise 2144 } 2145 } 2146 } 2147 2148 // Check for loading klass from an array klass 2149 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2150 if (tkls != NULL && !StressReflectiveCode) { 2151 ciKlass* klass = tkls->klass(); 2152 if( !klass->is_loaded() ) 2153 return _type; // Bail out if not loaded 2154 if( klass->is_obj_array_klass() && 2155 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2156 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2157 // // Always returning precise element type is incorrect, 2158 // // e.g., element type could be object and array may contain strings 2159 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2160 2161 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2162 // according to the element type's subclassing. 2163 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2164 } 2165 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2166 tkls->offset() == in_bytes(Klass::super_offset())) { 2167 ciKlass* sup = klass->as_instance_klass()->super(); 2168 // The field is Klass::_super. Return its (constant) value. 2169 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2170 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2171 } 2172 } 2173 2174 // Bailout case 2175 return LoadNode::Value(phase); 2176 } 2177 2178 //------------------------------Identity--------------------------------------- 2179 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2180 // Also feed through the klass in Allocate(...klass...)._klass. 2181 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2182 return klass_identity_common(phase); 2183 } 2184 2185 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2186 Node* x = LoadNode::Identity(phase); 2187 if (x != this) return x; 2188 2189 // Take apart the address into an oop and and offset. 2190 // Return 'this' if we cannot. 2191 Node* adr = in(MemNode::Address); 2192 intptr_t offset = 0; 2193 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2194 if (base == NULL) return this; 2195 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2196 if (toop == NULL) return this; 2197 2198 // We can fetch the klass directly through an AllocateNode. 2199 // This works even if the klass is not constant (clone or newArray). 2200 if (offset == oopDesc::klass_offset_in_bytes()) { 2201 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2202 if (allocated_klass != NULL) { 2203 return allocated_klass; 2204 } 2205 } 2206 2207 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2208 // See inline_native_Class_query for occurrences of these patterns. 2209 // Java Example: x.getClass().isAssignableFrom(y) 2210 // 2211 // This improves reflective code, often making the Class 2212 // mirror go completely dead. (Current exception: Class 2213 // mirrors may appear in debug info, but we could clean them out by 2214 // introducing a new debug info operator for Klass*.java_mirror). 2215 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2216 && offset == java_lang_Class::klass_offset_in_bytes()) { 2217 // We are loading a special hidden field from a Class mirror, 2218 // the field which points to its Klass or ArrayKlass metaobject. 2219 if (base->is_Load()) { 2220 Node* adr2 = base->in(MemNode::Address); 2221 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2222 if (tkls != NULL && !tkls->empty() 2223 && (tkls->klass()->is_instance_klass() || 2224 tkls->klass()->is_array_klass()) 2225 && adr2->is_AddP() 2226 ) { 2227 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2228 if (tkls->offset() == mirror_field) { 2229 return adr2->in(AddPNode::Base); 2230 } 2231 } 2232 } 2233 } 2234 2235 return this; 2236 } 2237 2238 2239 //------------------------------Value------------------------------------------ 2240 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2241 const Type *t = klass_value_common(phase); 2242 if (t == Type::TOP) 2243 return t; 2244 2245 return t->make_narrowklass(); 2246 } 2247 2248 //------------------------------Identity--------------------------------------- 2249 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2250 // Also feed through the klass in Allocate(...klass...)._klass. 2251 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2252 Node *x = klass_identity_common(phase); 2253 2254 const Type *t = phase->type( x ); 2255 if( t == Type::TOP ) return x; 2256 if( t->isa_narrowklass()) return x; 2257 assert (!t->isa_narrowoop(), "no narrow oop here"); 2258 2259 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2260 } 2261 2262 //------------------------------Value----------------------------------------- 2263 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2264 // Either input is TOP ==> the result is TOP 2265 const Type *t1 = phase->type( in(MemNode::Memory) ); 2266 if( t1 == Type::TOP ) return Type::TOP; 2267 Node *adr = in(MemNode::Address); 2268 const Type *t2 = phase->type( adr ); 2269 if( t2 == Type::TOP ) return Type::TOP; 2270 const TypePtr *tp = t2->is_ptr(); 2271 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2272 const TypeAryPtr *tap = tp->isa_aryptr(); 2273 if( !tap ) return _type; 2274 return tap->size(); 2275 } 2276 2277 //-------------------------------Ideal--------------------------------------- 2278 // Feed through the length in AllocateArray(...length...)._length. 2279 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2280 Node* p = MemNode::Ideal_common(phase, can_reshape); 2281 if (p) return (p == NodeSentinel) ? NULL : p; 2282 2283 // Take apart the address into an oop and and offset. 2284 // Return 'this' if we cannot. 2285 Node* adr = in(MemNode::Address); 2286 intptr_t offset = 0; 2287 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2288 if (base == NULL) return NULL; 2289 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2290 if (tary == NULL) return NULL; 2291 2292 // We can fetch the length directly through an AllocateArrayNode. 2293 // This works even if the length is not constant (clone or newArray). 2294 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2295 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2296 if (alloc != NULL) { 2297 Node* allocated_length = alloc->Ideal_length(); 2298 Node* len = alloc->make_ideal_length(tary, phase); 2299 if (allocated_length != len) { 2300 // New CastII improves on this. 2301 return len; 2302 } 2303 } 2304 } 2305 2306 return NULL; 2307 } 2308 2309 //------------------------------Identity--------------------------------------- 2310 // Feed through the length in AllocateArray(...length...)._length. 2311 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2312 Node* x = LoadINode::Identity(phase); 2313 if (x != this) return x; 2314 2315 // Take apart the address into an oop and and offset. 2316 // Return 'this' if we cannot. 2317 Node* adr = in(MemNode::Address); 2318 intptr_t offset = 0; 2319 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2320 if (base == NULL) return this; 2321 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2322 if (tary == NULL) return this; 2323 2324 // We can fetch the length directly through an AllocateArrayNode. 2325 // This works even if the length is not constant (clone or newArray). 2326 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2327 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2328 if (alloc != NULL) { 2329 Node* allocated_length = alloc->Ideal_length(); 2330 // Do not allow make_ideal_length to allocate a CastII node. 2331 Node* len = alloc->make_ideal_length(tary, phase, false); 2332 if (allocated_length == len) { 2333 // Return allocated_length only if it would not be improved by a CastII. 2334 return allocated_length; 2335 } 2336 } 2337 } 2338 2339 return this; 2340 2341 } 2342 2343 //============================================================================= 2344 //---------------------------StoreNode::make----------------------------------- 2345 // Polymorphic factory method: 2346 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2347 assert((mo == unordered || mo == release), "unexpected"); 2348 Compile* C = gvn.C; 2349 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2350 ctl != NULL, "raw memory operations should have control edge"); 2351 2352 switch (bt) { 2353 case T_BOOLEAN: 2354 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2355 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2356 case T_CHAR: 2357 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2358 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2359 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2360 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2361 case T_METADATA: 2362 case T_ADDRESS: 2363 case T_OBJECT: 2364 #ifdef _LP64 2365 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2366 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2367 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2368 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2369 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2370 adr->bottom_type()->isa_rawptr())) { 2371 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2372 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2373 } 2374 #endif 2375 { 2376 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2377 } 2378 } 2379 ShouldNotReachHere(); 2380 return (StoreNode*)NULL; 2381 } 2382 2383 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2384 bool require_atomic = true; 2385 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2386 } 2387 2388 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2389 bool require_atomic = true; 2390 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2391 } 2392 2393 2394 //--------------------------bottom_type---------------------------------------- 2395 const Type *StoreNode::bottom_type() const { 2396 return Type::MEMORY; 2397 } 2398 2399 //------------------------------hash------------------------------------------- 2400 uint StoreNode::hash() const { 2401 // unroll addition of interesting fields 2402 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2403 2404 // Since they are not commoned, do not hash them: 2405 return NO_HASH; 2406 } 2407 2408 //------------------------------Ideal------------------------------------------ 2409 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2410 // When a store immediately follows a relevant allocation/initialization, 2411 // try to capture it into the initialization, or hoist it above. 2412 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2413 Node* p = MemNode::Ideal_common(phase, can_reshape); 2414 if (p) return (p == NodeSentinel) ? NULL : p; 2415 2416 Node* mem = in(MemNode::Memory); 2417 Node* address = in(MemNode::Address); 2418 // Back-to-back stores to same address? Fold em up. Generally 2419 // unsafe if I have intervening uses... Also disallowed for StoreCM 2420 // since they must follow each StoreP operation. Redundant StoreCMs 2421 // are eliminated just before matching in final_graph_reshape. 2422 { 2423 Node* st = mem; 2424 // If Store 'st' has more than one use, we cannot fold 'st' away. 2425 // For example, 'st' might be the final state at a conditional 2426 // return. Or, 'st' might be used by some node which is live at 2427 // the same time 'st' is live, which might be unschedulable. So, 2428 // require exactly ONE user until such time as we clone 'mem' for 2429 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2430 // true). 2431 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2432 // Looking at a dead closed cycle of memory? 2433 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2434 assert(Opcode() == st->Opcode() || 2435 st->Opcode() == Op_StoreVector || 2436 Opcode() == Op_StoreVector || 2437 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2438 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2439 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2440 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2441 2442 if (st->in(MemNode::Address)->eqv_uncast(address) && 2443 st->as_Store()->memory_size() <= this->memory_size()) { 2444 Node* use = st->raw_out(0); 2445 phase->igvn_rehash_node_delayed(use); 2446 if (can_reshape) { 2447 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2448 } else { 2449 // It's OK to do this in the parser, since DU info is always accurate, 2450 // and the parser always refers to nodes via SafePointNode maps. 2451 use->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2452 } 2453 return this; 2454 } 2455 st = st->in(MemNode::Memory); 2456 } 2457 } 2458 2459 2460 // Capture an unaliased, unconditional, simple store into an initializer. 2461 // Or, if it is independent of the allocation, hoist it above the allocation. 2462 if (ReduceFieldZeroing && /*can_reshape &&*/ 2463 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2464 InitializeNode* init = mem->in(0)->as_Initialize(); 2465 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2466 if (offset > 0) { 2467 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2468 // If the InitializeNode captured me, it made a raw copy of me, 2469 // and I need to disappear. 2470 if (moved != NULL) { 2471 // %%% hack to ensure that Ideal returns a new node: 2472 mem = MergeMemNode::make(mem); 2473 return mem; // fold me away 2474 } 2475 } 2476 } 2477 2478 return NULL; // No further progress 2479 } 2480 2481 //------------------------------Value----------------------------------------- 2482 const Type* StoreNode::Value(PhaseGVN* phase) const { 2483 // Either input is TOP ==> the result is TOP 2484 const Type *t1 = phase->type( in(MemNode::Memory) ); 2485 if( t1 == Type::TOP ) return Type::TOP; 2486 const Type *t2 = phase->type( in(MemNode::Address) ); 2487 if( t2 == Type::TOP ) return Type::TOP; 2488 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2489 if( t3 == Type::TOP ) return Type::TOP; 2490 return Type::MEMORY; 2491 } 2492 2493 //------------------------------Identity--------------------------------------- 2494 // Remove redundant stores: 2495 // Store(m, p, Load(m, p)) changes to m. 2496 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2497 Node* StoreNode::Identity(PhaseGVN* phase) { 2498 Node* mem = in(MemNode::Memory); 2499 Node* adr = in(MemNode::Address); 2500 Node* val = in(MemNode::ValueIn); 2501 2502 // Load then Store? Then the Store is useless 2503 if (val->is_Load() && 2504 val->in(MemNode::Address)->eqv_uncast(adr) && 2505 val->in(MemNode::Memory )->eqv_uncast(mem) && 2506 val->as_Load()->store_Opcode() == Opcode()) { 2507 return mem; 2508 } 2509 2510 // Two stores in a row of the same value? 2511 if (mem->is_Store() && 2512 mem->in(MemNode::Address)->eqv_uncast(adr) && 2513 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2514 mem->Opcode() == Opcode()) { 2515 return mem; 2516 } 2517 2518 // Store of zero anywhere into a freshly-allocated object? 2519 // Then the store is useless. 2520 // (It must already have been captured by the InitializeNode.) 2521 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2522 // a newly allocated object is already all-zeroes everywhere 2523 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2524 return mem; 2525 } 2526 2527 // the store may also apply to zero-bits in an earlier object 2528 Node* prev_mem = find_previous_store(phase); 2529 // Steps (a), (b): Walk past independent stores to find an exact match. 2530 if (prev_mem != NULL) { 2531 Node* prev_val = can_see_stored_value(prev_mem, phase); 2532 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2533 // prev_val and val might differ by a cast; it would be good 2534 // to keep the more informative of the two. 2535 return mem; 2536 } 2537 } 2538 } 2539 2540 return this; 2541 } 2542 2543 //------------------------------match_edge------------------------------------- 2544 // Do we Match on this edge index or not? Match only memory & value 2545 uint StoreNode::match_edge(uint idx) const { 2546 return idx == MemNode::Address || idx == MemNode::ValueIn; 2547 } 2548 2549 //------------------------------cmp-------------------------------------------- 2550 // Do not common stores up together. They generally have to be split 2551 // back up anyways, so do not bother. 2552 uint StoreNode::cmp( const Node &n ) const { 2553 return (&n == this); // Always fail except on self 2554 } 2555 2556 //------------------------------Ideal_masked_input----------------------------- 2557 // Check for a useless mask before a partial-word store 2558 // (StoreB ... (AndI valIn conIa) ) 2559 // If (conIa & mask == mask) this simplifies to 2560 // (StoreB ... (valIn) ) 2561 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2562 Node *val = in(MemNode::ValueIn); 2563 if( val->Opcode() == Op_AndI ) { 2564 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2565 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2566 set_req(MemNode::ValueIn, val->in(1)); 2567 return this; 2568 } 2569 } 2570 return NULL; 2571 } 2572 2573 2574 //------------------------------Ideal_sign_extended_input---------------------- 2575 // Check for useless sign-extension before a partial-word store 2576 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2577 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2578 // (StoreB ... (valIn) ) 2579 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2580 Node *val = in(MemNode::ValueIn); 2581 if( val->Opcode() == Op_RShiftI ) { 2582 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2583 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2584 Node *shl = val->in(1); 2585 if( shl->Opcode() == Op_LShiftI ) { 2586 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2587 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2588 set_req(MemNode::ValueIn, shl->in(1)); 2589 return this; 2590 } 2591 } 2592 } 2593 } 2594 return NULL; 2595 } 2596 2597 //------------------------------value_never_loaded----------------------------------- 2598 // Determine whether there are any possible loads of the value stored. 2599 // For simplicity, we actually check if there are any loads from the 2600 // address stored to, not just for loads of the value stored by this node. 2601 // 2602 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2603 Node *adr = in(Address); 2604 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2605 if (adr_oop == NULL) 2606 return false; 2607 if (!adr_oop->is_known_instance_field()) 2608 return false; // if not a distinct instance, there may be aliases of the address 2609 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2610 Node *use = adr->fast_out(i); 2611 if (use->is_Load() || use->is_LoadStore()) { 2612 return false; 2613 } 2614 } 2615 return true; 2616 } 2617 2618 //============================================================================= 2619 //------------------------------Ideal------------------------------------------ 2620 // If the store is from an AND mask that leaves the low bits untouched, then 2621 // we can skip the AND operation. If the store is from a sign-extension 2622 // (a left shift, then right shift) we can skip both. 2623 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2624 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2625 if( progress != NULL ) return progress; 2626 2627 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2628 if( progress != NULL ) return progress; 2629 2630 // Finally check the default case 2631 return StoreNode::Ideal(phase, can_reshape); 2632 } 2633 2634 //============================================================================= 2635 //------------------------------Ideal------------------------------------------ 2636 // If the store is from an AND mask that leaves the low bits untouched, then 2637 // we can skip the AND operation 2638 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2639 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2640 if( progress != NULL ) return progress; 2641 2642 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2643 if( progress != NULL ) return progress; 2644 2645 // Finally check the default case 2646 return StoreNode::Ideal(phase, can_reshape); 2647 } 2648 2649 //============================================================================= 2650 //------------------------------Identity--------------------------------------- 2651 Node* StoreCMNode::Identity(PhaseGVN* phase) { 2652 // No need to card mark when storing a null ptr 2653 Node* my_store = in(MemNode::OopStore); 2654 if (my_store->is_Store()) { 2655 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2656 if( t1 == TypePtr::NULL_PTR ) { 2657 return in(MemNode::Memory); 2658 } 2659 } 2660 return this; 2661 } 2662 2663 //============================================================================= 2664 //------------------------------Ideal--------------------------------------- 2665 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2666 Node* progress = StoreNode::Ideal(phase, can_reshape); 2667 if (progress != NULL) return progress; 2668 2669 Node* my_store = in(MemNode::OopStore); 2670 if (my_store->is_MergeMem()) { 2671 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2672 set_req(MemNode::OopStore, mem); 2673 return this; 2674 } 2675 2676 return NULL; 2677 } 2678 2679 //------------------------------Value----------------------------------------- 2680 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 2681 // Either input is TOP ==> the result is TOP 2682 const Type *t = phase->type( in(MemNode::Memory) ); 2683 if( t == Type::TOP ) return Type::TOP; 2684 t = phase->type( in(MemNode::Address) ); 2685 if( t == Type::TOP ) return Type::TOP; 2686 t = phase->type( in(MemNode::ValueIn) ); 2687 if( t == Type::TOP ) return Type::TOP; 2688 // If extra input is TOP ==> the result is TOP 2689 t = phase->type( in(MemNode::OopStore) ); 2690 if( t == Type::TOP ) return Type::TOP; 2691 2692 return StoreNode::Value( phase ); 2693 } 2694 2695 2696 //============================================================================= 2697 //----------------------------------SCMemProjNode------------------------------ 2698 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 2699 { 2700 return bottom_type(); 2701 } 2702 2703 //============================================================================= 2704 //----------------------------------LoadStoreNode------------------------------ 2705 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2706 : Node(required), 2707 _type(rt), 2708 _adr_type(at) 2709 { 2710 init_req(MemNode::Control, c ); 2711 init_req(MemNode::Memory , mem); 2712 init_req(MemNode::Address, adr); 2713 init_req(MemNode::ValueIn, val); 2714 init_class_id(Class_LoadStore); 2715 } 2716 2717 uint LoadStoreNode::ideal_reg() const { 2718 return _type->ideal_reg(); 2719 } 2720 2721 bool LoadStoreNode::result_not_used() const { 2722 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2723 Node *x = fast_out(i); 2724 if (x->Opcode() == Op_SCMemProj) continue; 2725 return false; 2726 } 2727 return true; 2728 } 2729 2730 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2731 2732 //============================================================================= 2733 //----------------------------------LoadStoreConditionalNode-------------------- 2734 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2735 init_req(ExpectedIn, ex ); 2736 } 2737 2738 //============================================================================= 2739 //-------------------------------adr_type-------------------------------------- 2740 const TypePtr* ClearArrayNode::adr_type() const { 2741 Node *adr = in(3); 2742 if (adr == NULL) return NULL; // node is dead 2743 return MemNode::calculate_adr_type(adr->bottom_type()); 2744 } 2745 2746 //------------------------------match_edge------------------------------------- 2747 // Do we Match on this edge index or not? Do not match memory 2748 uint ClearArrayNode::match_edge(uint idx) const { 2749 return idx > 1; 2750 } 2751 2752 //------------------------------Identity--------------------------------------- 2753 // Clearing a zero length array does nothing 2754 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 2755 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2756 } 2757 2758 //------------------------------Idealize--------------------------------------- 2759 // Clearing a short array is faster with stores 2760 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2761 // Already know this is a large node, do not try to ideal it 2762 if (_is_large) return NULL; 2763 2764 const int unit = BytesPerLong; 2765 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2766 if (!t) return NULL; 2767 if (!t->is_con()) return NULL; 2768 intptr_t raw_count = t->get_con(); 2769 intptr_t size = raw_count; 2770 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2771 // Clearing nothing uses the Identity call. 2772 // Negative clears are possible on dead ClearArrays 2773 // (see jck test stmt114.stmt11402.val). 2774 if (size <= 0 || size % unit != 0) return NULL; 2775 intptr_t count = size / unit; 2776 // Length too long; communicate this to matchers and assemblers. 2777 // Assemblers are responsible to produce fast hardware clears for it. 2778 if (size > InitArrayShortSize) { 2779 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 2780 } 2781 Node *mem = in(1); 2782 if( phase->type(mem)==Type::TOP ) return NULL; 2783 Node *adr = in(3); 2784 const Type* at = phase->type(adr); 2785 if( at==Type::TOP ) return NULL; 2786 const TypePtr* atp = at->isa_ptr(); 2787 // adjust atp to be the correct array element address type 2788 if (atp == NULL) atp = TypePtr::BOTTOM; 2789 else atp = atp->add_offset(Type::OffsetBot); 2790 // Get base for derived pointer purposes 2791 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2792 Node *base = adr->in(1); 2793 2794 Node *zero = phase->makecon(TypeLong::ZERO); 2795 Node *off = phase->MakeConX(BytesPerLong); 2796 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2797 count--; 2798 while( count-- ) { 2799 mem = phase->transform(mem); 2800 adr = phase->transform(new AddPNode(base,adr,off)); 2801 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2802 } 2803 return mem; 2804 } 2805 2806 //----------------------------step_through---------------------------------- 2807 // Return allocation input memory edge if it is different instance 2808 // or itself if it is the one we are looking for. 2809 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2810 Node* n = *np; 2811 assert(n->is_ClearArray(), "sanity"); 2812 intptr_t offset; 2813 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2814 // This method is called only before Allocate nodes are expanded 2815 // during macro nodes expansion. Before that ClearArray nodes are 2816 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2817 // Allocate nodes are expanded) which follows allocations. 2818 assert(alloc != NULL, "should have allocation"); 2819 if (alloc->_idx == instance_id) { 2820 // Can not bypass initialization of the instance we are looking for. 2821 return false; 2822 } 2823 // Otherwise skip it. 2824 InitializeNode* init = alloc->initialization(); 2825 if (init != NULL) 2826 *np = init->in(TypeFunc::Memory); 2827 else 2828 *np = alloc->in(TypeFunc::Memory); 2829 return true; 2830 } 2831 2832 //----------------------------clear_memory------------------------------------- 2833 // Generate code to initialize object storage to zero. 2834 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2835 intptr_t start_offset, 2836 Node* end_offset, 2837 PhaseGVN* phase) { 2838 intptr_t offset = start_offset; 2839 2840 int unit = BytesPerLong; 2841 if ((offset % unit) != 0) { 2842 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2843 adr = phase->transform(adr); 2844 const TypePtr* atp = TypeRawPtr::BOTTOM; 2845 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2846 mem = phase->transform(mem); 2847 offset += BytesPerInt; 2848 } 2849 assert((offset % unit) == 0, ""); 2850 2851 // Initialize the remaining stuff, if any, with a ClearArray. 2852 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2853 } 2854 2855 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2856 Node* start_offset, 2857 Node* end_offset, 2858 PhaseGVN* phase) { 2859 if (start_offset == end_offset) { 2860 // nothing to do 2861 return mem; 2862 } 2863 2864 int unit = BytesPerLong; 2865 Node* zbase = start_offset; 2866 Node* zend = end_offset; 2867 2868 // Scale to the unit required by the CPU: 2869 if (!Matcher::init_array_count_is_in_bytes) { 2870 Node* shift = phase->intcon(exact_log2(unit)); 2871 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2872 zend = phase->transform(new URShiftXNode(zend, shift) ); 2873 } 2874 2875 // Bulk clear double-words 2876 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2877 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2878 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 2879 return phase->transform(mem); 2880 } 2881 2882 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2883 intptr_t start_offset, 2884 intptr_t end_offset, 2885 PhaseGVN* phase) { 2886 if (start_offset == end_offset) { 2887 // nothing to do 2888 return mem; 2889 } 2890 2891 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2892 intptr_t done_offset = end_offset; 2893 if ((done_offset % BytesPerLong) != 0) { 2894 done_offset -= BytesPerInt; 2895 } 2896 if (done_offset > start_offset) { 2897 mem = clear_memory(ctl, mem, dest, 2898 start_offset, phase->MakeConX(done_offset), phase); 2899 } 2900 if (done_offset < end_offset) { // emit the final 32-bit store 2901 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 2902 adr = phase->transform(adr); 2903 const TypePtr* atp = TypeRawPtr::BOTTOM; 2904 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2905 mem = phase->transform(mem); 2906 done_offset += BytesPerInt; 2907 } 2908 assert(done_offset == end_offset, ""); 2909 return mem; 2910 } 2911 2912 //============================================================================= 2913 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2914 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2915 _adr_type(C->get_adr_type(alias_idx)) 2916 { 2917 init_class_id(Class_MemBar); 2918 Node* top = C->top(); 2919 init_req(TypeFunc::I_O,top); 2920 init_req(TypeFunc::FramePtr,top); 2921 init_req(TypeFunc::ReturnAdr,top); 2922 if (precedent != NULL) 2923 init_req(TypeFunc::Parms, precedent); 2924 } 2925 2926 //------------------------------cmp-------------------------------------------- 2927 uint MemBarNode::hash() const { return NO_HASH; } 2928 uint MemBarNode::cmp( const Node &n ) const { 2929 return (&n == this); // Always fail except on self 2930 } 2931 2932 //------------------------------make------------------------------------------- 2933 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2934 switch (opcode) { 2935 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 2936 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 2937 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 2938 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 2939 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 2940 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 2941 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 2942 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 2943 case Op_Initialize: return new InitializeNode(C, atp, pn); 2944 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 2945 default: ShouldNotReachHere(); return NULL; 2946 } 2947 } 2948 2949 //------------------------------Ideal------------------------------------------ 2950 // Return a node which is more "ideal" than the current node. Strip out 2951 // control copies 2952 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2953 if (remove_dead_region(phase, can_reshape)) return this; 2954 // Don't bother trying to transform a dead node 2955 if (in(0) && in(0)->is_top()) { 2956 return NULL; 2957 } 2958 2959 bool progress = false; 2960 // Eliminate volatile MemBars for scalar replaced objects. 2961 if (can_reshape && req() == (Precedent+1)) { 2962 bool eliminate = false; 2963 int opc = Opcode(); 2964 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2965 // Volatile field loads and stores. 2966 Node* my_mem = in(MemBarNode::Precedent); 2967 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2968 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2969 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2970 // replace this Precedent (decodeN) with the Load instead. 2971 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2972 Node* load_node = my_mem->in(1); 2973 set_req(MemBarNode::Precedent, load_node); 2974 phase->is_IterGVN()->_worklist.push(my_mem); 2975 my_mem = load_node; 2976 } else { 2977 assert(my_mem->unique_out() == this, "sanity"); 2978 del_req(Precedent); 2979 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2980 my_mem = NULL; 2981 } 2982 progress = true; 2983 } 2984 if (my_mem != NULL && my_mem->is_Mem()) { 2985 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2986 // Check for scalar replaced object reference. 2987 if( t_oop != NULL && t_oop->is_known_instance_field() && 2988 t_oop->offset() != Type::OffsetBot && 2989 t_oop->offset() != Type::OffsetTop) { 2990 eliminate = true; 2991 } 2992 } 2993 } else if (opc == Op_MemBarRelease) { 2994 // Final field stores. 2995 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 2996 if ((alloc != NULL) && alloc->is_Allocate() && 2997 alloc->as_Allocate()->does_not_escape_thread()) { 2998 // The allocated object does not escape. 2999 eliminate = true; 3000 } 3001 } 3002 if (eliminate) { 3003 // Replace MemBar projections by its inputs. 3004 PhaseIterGVN* igvn = phase->is_IterGVN(); 3005 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3006 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3007 // Must return either the original node (now dead) or a new node 3008 // (Do not return a top here, since that would break the uniqueness of top.) 3009 return new ConINode(TypeInt::ZERO); 3010 } 3011 } 3012 return progress ? this : NULL; 3013 } 3014 3015 //------------------------------Value------------------------------------------ 3016 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3017 if( !in(0) ) return Type::TOP; 3018 if( phase->type(in(0)) == Type::TOP ) 3019 return Type::TOP; 3020 return TypeTuple::MEMBAR; 3021 } 3022 3023 //------------------------------match------------------------------------------ 3024 // Construct projections for memory. 3025 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3026 switch (proj->_con) { 3027 case TypeFunc::Control: 3028 case TypeFunc::Memory: 3029 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3030 } 3031 ShouldNotReachHere(); 3032 return NULL; 3033 } 3034 3035 //===========================InitializeNode==================================== 3036 // SUMMARY: 3037 // This node acts as a memory barrier on raw memory, after some raw stores. 3038 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3039 // The Initialize can 'capture' suitably constrained stores as raw inits. 3040 // It can coalesce related raw stores into larger units (called 'tiles'). 3041 // It can avoid zeroing new storage for memory units which have raw inits. 3042 // At macro-expansion, it is marked 'complete', and does not optimize further. 3043 // 3044 // EXAMPLE: 3045 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3046 // ctl = incoming control; mem* = incoming memory 3047 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3048 // First allocate uninitialized memory and fill in the header: 3049 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3050 // ctl := alloc.Control; mem* := alloc.Memory* 3051 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3052 // Then initialize to zero the non-header parts of the raw memory block: 3053 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3054 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3055 // After the initialize node executes, the object is ready for service: 3056 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3057 // Suppose its body is immediately initialized as {1,2}: 3058 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3059 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3060 // mem.SLICE(#short[*]) := store2 3061 // 3062 // DETAILS: 3063 // An InitializeNode collects and isolates object initialization after 3064 // an AllocateNode and before the next possible safepoint. As a 3065 // memory barrier (MemBarNode), it keeps critical stores from drifting 3066 // down past any safepoint or any publication of the allocation. 3067 // Before this barrier, a newly-allocated object may have uninitialized bits. 3068 // After this barrier, it may be treated as a real oop, and GC is allowed. 3069 // 3070 // The semantics of the InitializeNode include an implicit zeroing of 3071 // the new object from object header to the end of the object. 3072 // (The object header and end are determined by the AllocateNode.) 3073 // 3074 // Certain stores may be added as direct inputs to the InitializeNode. 3075 // These stores must update raw memory, and they must be to addresses 3076 // derived from the raw address produced by AllocateNode, and with 3077 // a constant offset. They must be ordered by increasing offset. 3078 // The first one is at in(RawStores), the last at in(req()-1). 3079 // Unlike most memory operations, they are not linked in a chain, 3080 // but are displayed in parallel as users of the rawmem output of 3081 // the allocation. 3082 // 3083 // (See comments in InitializeNode::capture_store, which continue 3084 // the example given above.) 3085 // 3086 // When the associated Allocate is macro-expanded, the InitializeNode 3087 // may be rewritten to optimize collected stores. A ClearArrayNode 3088 // may also be created at that point to represent any required zeroing. 3089 // The InitializeNode is then marked 'complete', prohibiting further 3090 // capturing of nearby memory operations. 3091 // 3092 // During macro-expansion, all captured initializations which store 3093 // constant values of 32 bits or smaller are coalesced (if advantageous) 3094 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3095 // initialized in fewer memory operations. Memory words which are 3096 // covered by neither tiles nor non-constant stores are pre-zeroed 3097 // by explicit stores of zero. (The code shape happens to do all 3098 // zeroing first, then all other stores, with both sequences occurring 3099 // in order of ascending offsets.) 3100 // 3101 // Alternatively, code may be inserted between an AllocateNode and its 3102 // InitializeNode, to perform arbitrary initialization of the new object. 3103 // E.g., the object copying intrinsics insert complex data transfers here. 3104 // The initialization must then be marked as 'complete' disable the 3105 // built-in zeroing semantics and the collection of initializing stores. 3106 // 3107 // While an InitializeNode is incomplete, reads from the memory state 3108 // produced by it are optimizable if they match the control edge and 3109 // new oop address associated with the allocation/initialization. 3110 // They return a stored value (if the offset matches) or else zero. 3111 // A write to the memory state, if it matches control and address, 3112 // and if it is to a constant offset, may be 'captured' by the 3113 // InitializeNode. It is cloned as a raw memory operation and rewired 3114 // inside the initialization, to the raw oop produced by the allocation. 3115 // Operations on addresses which are provably distinct (e.g., to 3116 // other AllocateNodes) are allowed to bypass the initialization. 3117 // 3118 // The effect of all this is to consolidate object initialization 3119 // (both arrays and non-arrays, both piecewise and bulk) into a 3120 // single location, where it can be optimized as a unit. 3121 // 3122 // Only stores with an offset less than TrackedInitializationLimit words 3123 // will be considered for capture by an InitializeNode. This puts a 3124 // reasonable limit on the complexity of optimized initializations. 3125 3126 //---------------------------InitializeNode------------------------------------ 3127 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3128 : _is_complete(Incomplete), _does_not_escape(false), 3129 MemBarNode(C, adr_type, rawoop) 3130 { 3131 init_class_id(Class_Initialize); 3132 3133 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3134 assert(in(RawAddress) == rawoop, "proper init"); 3135 // Note: allocation() can be NULL, for secondary initialization barriers 3136 } 3137 3138 // Since this node is not matched, it will be processed by the 3139 // register allocator. Declare that there are no constraints 3140 // on the allocation of the RawAddress edge. 3141 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3142 // This edge should be set to top, by the set_complete. But be conservative. 3143 if (idx == InitializeNode::RawAddress) 3144 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3145 return RegMask::Empty; 3146 } 3147 3148 Node* InitializeNode::memory(uint alias_idx) { 3149 Node* mem = in(Memory); 3150 if (mem->is_MergeMem()) { 3151 return mem->as_MergeMem()->memory_at(alias_idx); 3152 } else { 3153 // incoming raw memory is not split 3154 return mem; 3155 } 3156 } 3157 3158 bool InitializeNode::is_non_zero() { 3159 if (is_complete()) return false; 3160 remove_extra_zeroes(); 3161 return (req() > RawStores); 3162 } 3163 3164 void InitializeNode::set_complete(PhaseGVN* phase) { 3165 assert(!is_complete(), "caller responsibility"); 3166 _is_complete = Complete; 3167 3168 // After this node is complete, it contains a bunch of 3169 // raw-memory initializations. There is no need for 3170 // it to have anything to do with non-raw memory effects. 3171 // Therefore, tell all non-raw users to re-optimize themselves, 3172 // after skipping the memory effects of this initialization. 3173 PhaseIterGVN* igvn = phase->is_IterGVN(); 3174 if (igvn) igvn->add_users_to_worklist(this); 3175 } 3176 3177 // convenience function 3178 // return false if the init contains any stores already 3179 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3180 InitializeNode* init = initialization(); 3181 if (init == NULL || init->is_complete()) return false; 3182 init->remove_extra_zeroes(); 3183 // for now, if this allocation has already collected any inits, bail: 3184 if (init->is_non_zero()) return false; 3185 init->set_complete(phase); 3186 return true; 3187 } 3188 3189 void InitializeNode::remove_extra_zeroes() { 3190 if (req() == RawStores) return; 3191 Node* zmem = zero_memory(); 3192 uint fill = RawStores; 3193 for (uint i = fill; i < req(); i++) { 3194 Node* n = in(i); 3195 if (n->is_top() || n == zmem) continue; // skip 3196 if (fill < i) set_req(fill, n); // compact 3197 ++fill; 3198 } 3199 // delete any empty spaces created: 3200 while (fill < req()) { 3201 del_req(fill); 3202 } 3203 } 3204 3205 // Helper for remembering which stores go with which offsets. 3206 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3207 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3208 intptr_t offset = -1; 3209 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3210 phase, offset); 3211 if (base == NULL) return -1; // something is dead, 3212 if (offset < 0) return -1; // dead, dead 3213 return offset; 3214 } 3215 3216 // Helper for proving that an initialization expression is 3217 // "simple enough" to be folded into an object initialization. 3218 // Attempts to prove that a store's initial value 'n' can be captured 3219 // within the initialization without creating a vicious cycle, such as: 3220 // { Foo p = new Foo(); p.next = p; } 3221 // True for constants and parameters and small combinations thereof. 3222 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3223 if (n == NULL) return true; // (can this really happen?) 3224 if (n->is_Proj()) n = n->in(0); 3225 if (n == this) return false; // found a cycle 3226 if (n->is_Con()) return true; 3227 if (n->is_Start()) return true; // params, etc., are OK 3228 if (n->is_Root()) return true; // even better 3229 3230 Node* ctl = n->in(0); 3231 if (ctl != NULL && !ctl->is_top()) { 3232 if (ctl->is_Proj()) ctl = ctl->in(0); 3233 if (ctl == this) return false; 3234 3235 // If we already know that the enclosing memory op is pinned right after 3236 // the init, then any control flow that the store has picked up 3237 // must have preceded the init, or else be equal to the init. 3238 // Even after loop optimizations (which might change control edges) 3239 // a store is never pinned *before* the availability of its inputs. 3240 if (!MemNode::all_controls_dominate(n, this)) 3241 return false; // failed to prove a good control 3242 } 3243 3244 // Check data edges for possible dependencies on 'this'. 3245 if ((count += 1) > 20) return false; // complexity limit 3246 for (uint i = 1; i < n->req(); i++) { 3247 Node* m = n->in(i); 3248 if (m == NULL || m == n || m->is_top()) continue; 3249 uint first_i = n->find_edge(m); 3250 if (i != first_i) continue; // process duplicate edge just once 3251 if (!detect_init_independence(m, count)) { 3252 return false; 3253 } 3254 } 3255 3256 return true; 3257 } 3258 3259 // Here are all the checks a Store must pass before it can be moved into 3260 // an initialization. Returns zero if a check fails. 3261 // On success, returns the (constant) offset to which the store applies, 3262 // within the initialized memory. 3263 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3264 const int FAIL = 0; 3265 if (st->is_unaligned_access()) { 3266 return FAIL; 3267 } 3268 if (st->req() != MemNode::ValueIn + 1) 3269 return FAIL; // an inscrutable StoreNode (card mark?) 3270 Node* ctl = st->in(MemNode::Control); 3271 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3272 return FAIL; // must be unconditional after the initialization 3273 Node* mem = st->in(MemNode::Memory); 3274 if (!(mem->is_Proj() && mem->in(0) == this)) 3275 return FAIL; // must not be preceded by other stores 3276 Node* adr = st->in(MemNode::Address); 3277 intptr_t offset; 3278 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3279 if (alloc == NULL) 3280 return FAIL; // inscrutable address 3281 if (alloc != allocation()) 3282 return FAIL; // wrong allocation! (store needs to float up) 3283 Node* val = st->in(MemNode::ValueIn); 3284 int complexity_count = 0; 3285 if (!detect_init_independence(val, complexity_count)) 3286 return FAIL; // stored value must be 'simple enough' 3287 3288 // The Store can be captured only if nothing after the allocation 3289 // and before the Store is using the memory location that the store 3290 // overwrites. 3291 bool failed = false; 3292 // If is_complete_with_arraycopy() is true the shape of the graph is 3293 // well defined and is safe so no need for extra checks. 3294 if (!is_complete_with_arraycopy()) { 3295 // We are going to look at each use of the memory state following 3296 // the allocation to make sure nothing reads the memory that the 3297 // Store writes. 3298 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3299 int alias_idx = phase->C->get_alias_index(t_adr); 3300 ResourceMark rm; 3301 Unique_Node_List mems; 3302 mems.push(mem); 3303 Node* unique_merge = NULL; 3304 for (uint next = 0; next < mems.size(); ++next) { 3305 Node *m = mems.at(next); 3306 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3307 Node *n = m->fast_out(j); 3308 if (n->outcnt() == 0) { 3309 continue; 3310 } 3311 if (n == st) { 3312 continue; 3313 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3314 // If the control of this use is different from the control 3315 // of the Store which is right after the InitializeNode then 3316 // this node cannot be between the InitializeNode and the 3317 // Store. 3318 continue; 3319 } else if (n->is_MergeMem()) { 3320 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3321 // We can hit a MergeMemNode (that will likely go away 3322 // later) that is a direct use of the memory state 3323 // following the InitializeNode on the same slice as the 3324 // store node that we'd like to capture. We need to check 3325 // the uses of the MergeMemNode. 3326 mems.push(n); 3327 } 3328 } else if (n->is_Mem()) { 3329 Node* other_adr = n->in(MemNode::Address); 3330 if (other_adr == adr) { 3331 failed = true; 3332 break; 3333 } else { 3334 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3335 if (other_t_adr != NULL) { 3336 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3337 if (other_alias_idx == alias_idx) { 3338 // A load from the same memory slice as the store right 3339 // after the InitializeNode. We check the control of the 3340 // object/array that is loaded from. If it's the same as 3341 // the store control then we cannot capture the store. 3342 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3343 Node* base = other_adr; 3344 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3345 base = base->in(AddPNode::Base); 3346 if (base != NULL) { 3347 base = base->uncast(); 3348 if (base->is_Proj() && base->in(0) == alloc) { 3349 failed = true; 3350 break; 3351 } 3352 } 3353 } 3354 } 3355 } 3356 } else { 3357 failed = true; 3358 break; 3359 } 3360 } 3361 } 3362 } 3363 if (failed) { 3364 if (!can_reshape) { 3365 // We decided we couldn't capture the store during parsing. We 3366 // should try again during the next IGVN once the graph is 3367 // cleaner. 3368 phase->C->record_for_igvn(st); 3369 } 3370 return FAIL; 3371 } 3372 3373 return offset; // success 3374 } 3375 3376 // Find the captured store in(i) which corresponds to the range 3377 // [start..start+size) in the initialized object. 3378 // If there is one, return its index i. If there isn't, return the 3379 // negative of the index where it should be inserted. 3380 // Return 0 if the queried range overlaps an initialization boundary 3381 // or if dead code is encountered. 3382 // If size_in_bytes is zero, do not bother with overlap checks. 3383 int InitializeNode::captured_store_insertion_point(intptr_t start, 3384 int size_in_bytes, 3385 PhaseTransform* phase) { 3386 const int FAIL = 0, MAX_STORE = BytesPerLong; 3387 3388 if (is_complete()) 3389 return FAIL; // arraycopy got here first; punt 3390 3391 assert(allocation() != NULL, "must be present"); 3392 3393 // no negatives, no header fields: 3394 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3395 3396 // after a certain size, we bail out on tracking all the stores: 3397 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3398 if (start >= ti_limit) return FAIL; 3399 3400 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3401 if (i >= limit) return -(int)i; // not found; here is where to put it 3402 3403 Node* st = in(i); 3404 intptr_t st_off = get_store_offset(st, phase); 3405 if (st_off < 0) { 3406 if (st != zero_memory()) { 3407 return FAIL; // bail out if there is dead garbage 3408 } 3409 } else if (st_off > start) { 3410 // ...we are done, since stores are ordered 3411 if (st_off < start + size_in_bytes) { 3412 return FAIL; // the next store overlaps 3413 } 3414 return -(int)i; // not found; here is where to put it 3415 } else if (st_off < start) { 3416 if (size_in_bytes != 0 && 3417 start < st_off + MAX_STORE && 3418 start < st_off + st->as_Store()->memory_size()) { 3419 return FAIL; // the previous store overlaps 3420 } 3421 } else { 3422 if (size_in_bytes != 0 && 3423 st->as_Store()->memory_size() != size_in_bytes) { 3424 return FAIL; // mismatched store size 3425 } 3426 return i; 3427 } 3428 3429 ++i; 3430 } 3431 } 3432 3433 // Look for a captured store which initializes at the offset 'start' 3434 // with the given size. If there is no such store, and no other 3435 // initialization interferes, then return zero_memory (the memory 3436 // projection of the AllocateNode). 3437 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3438 PhaseTransform* phase) { 3439 assert(stores_are_sane(phase), ""); 3440 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3441 if (i == 0) { 3442 return NULL; // something is dead 3443 } else if (i < 0) { 3444 return zero_memory(); // just primordial zero bits here 3445 } else { 3446 Node* st = in(i); // here is the store at this position 3447 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3448 return st; 3449 } 3450 } 3451 3452 // Create, as a raw pointer, an address within my new object at 'offset'. 3453 Node* InitializeNode::make_raw_address(intptr_t offset, 3454 PhaseTransform* phase) { 3455 Node* addr = in(RawAddress); 3456 if (offset != 0) { 3457 Compile* C = phase->C; 3458 addr = phase->transform( new AddPNode(C->top(), addr, 3459 phase->MakeConX(offset)) ); 3460 } 3461 return addr; 3462 } 3463 3464 // Clone the given store, converting it into a raw store 3465 // initializing a field or element of my new object. 3466 // Caller is responsible for retiring the original store, 3467 // with subsume_node or the like. 3468 // 3469 // From the example above InitializeNode::InitializeNode, 3470 // here are the old stores to be captured: 3471 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3472 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3473 // 3474 // Here is the changed code; note the extra edges on init: 3475 // alloc = (Allocate ...) 3476 // rawoop = alloc.RawAddress 3477 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3478 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3479 // init = (Initialize alloc.Control alloc.Memory rawoop 3480 // rawstore1 rawstore2) 3481 // 3482 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3483 PhaseTransform* phase, bool can_reshape) { 3484 assert(stores_are_sane(phase), ""); 3485 3486 if (start < 0) return NULL; 3487 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3488 3489 Compile* C = phase->C; 3490 int size_in_bytes = st->memory_size(); 3491 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3492 if (i == 0) return NULL; // bail out 3493 Node* prev_mem = NULL; // raw memory for the captured store 3494 if (i > 0) { 3495 prev_mem = in(i); // there is a pre-existing store under this one 3496 set_req(i, C->top()); // temporarily disconnect it 3497 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3498 } else { 3499 i = -i; // no pre-existing store 3500 prev_mem = zero_memory(); // a slice of the newly allocated object 3501 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3502 set_req(--i, C->top()); // reuse this edge; it has been folded away 3503 else 3504 ins_req(i, C->top()); // build a new edge 3505 } 3506 Node* new_st = st->clone(); 3507 new_st->set_req(MemNode::Control, in(Control)); 3508 new_st->set_req(MemNode::Memory, prev_mem); 3509 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3510 new_st = phase->transform(new_st); 3511 3512 // At this point, new_st might have swallowed a pre-existing store 3513 // at the same offset, or perhaps new_st might have disappeared, 3514 // if it redundantly stored the same value (or zero to fresh memory). 3515 3516 // In any case, wire it in: 3517 phase->igvn_rehash_node_delayed(this); 3518 set_req(i, new_st); 3519 3520 // The caller may now kill the old guy. 3521 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3522 assert(check_st == new_st || check_st == NULL, "must be findable"); 3523 assert(!is_complete(), ""); 3524 return new_st; 3525 } 3526 3527 static bool store_constant(jlong* tiles, int num_tiles, 3528 intptr_t st_off, int st_size, 3529 jlong con) { 3530 if ((st_off & (st_size-1)) != 0) 3531 return false; // strange store offset (assume size==2**N) 3532 address addr = (address)tiles + st_off; 3533 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3534 switch (st_size) { 3535 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3536 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3537 case sizeof(jint): *(jint*) addr = (jint) con; break; 3538 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3539 default: return false; // strange store size (detect size!=2**N here) 3540 } 3541 return true; // return success to caller 3542 } 3543 3544 // Coalesce subword constants into int constants and possibly 3545 // into long constants. The goal, if the CPU permits, 3546 // is to initialize the object with a small number of 64-bit tiles. 3547 // Also, convert floating-point constants to bit patterns. 3548 // Non-constants are not relevant to this pass. 3549 // 3550 // In terms of the running example on InitializeNode::InitializeNode 3551 // and InitializeNode::capture_store, here is the transformation 3552 // of rawstore1 and rawstore2 into rawstore12: 3553 // alloc = (Allocate ...) 3554 // rawoop = alloc.RawAddress 3555 // tile12 = 0x00010002 3556 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3557 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3558 // 3559 void 3560 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3561 Node* size_in_bytes, 3562 PhaseGVN* phase) { 3563 Compile* C = phase->C; 3564 3565 assert(stores_are_sane(phase), ""); 3566 // Note: After this pass, they are not completely sane, 3567 // since there may be some overlaps. 3568 3569 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3570 3571 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3572 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3573 size_limit = MIN2(size_limit, ti_limit); 3574 size_limit = align_size_up(size_limit, BytesPerLong); 3575 int num_tiles = size_limit / BytesPerLong; 3576 3577 // allocate space for the tile map: 3578 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3579 jlong tiles_buf[small_len]; 3580 Node* nodes_buf[small_len]; 3581 jlong inits_buf[small_len]; 3582 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3583 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3584 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3585 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3586 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3587 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3588 // tiles: exact bitwise model of all primitive constants 3589 // nodes: last constant-storing node subsumed into the tiles model 3590 // inits: which bytes (in each tile) are touched by any initializations 3591 3592 //// Pass A: Fill in the tile model with any relevant stores. 3593 3594 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3595 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3596 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3597 Node* zmem = zero_memory(); // initially zero memory state 3598 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3599 Node* st = in(i); 3600 intptr_t st_off = get_store_offset(st, phase); 3601 3602 // Figure out the store's offset and constant value: 3603 if (st_off < header_size) continue; //skip (ignore header) 3604 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3605 int st_size = st->as_Store()->memory_size(); 3606 if (st_off + st_size > size_limit) break; 3607 3608 // Record which bytes are touched, whether by constant or not. 3609 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3610 continue; // skip (strange store size) 3611 3612 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3613 if (!val->singleton()) continue; //skip (non-con store) 3614 BasicType type = val->basic_type(); 3615 3616 jlong con = 0; 3617 switch (type) { 3618 case T_INT: con = val->is_int()->get_con(); break; 3619 case T_LONG: con = val->is_long()->get_con(); break; 3620 case T_FLOAT: con = jint_cast(val->getf()); break; 3621 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3622 default: continue; //skip (odd store type) 3623 } 3624 3625 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3626 st->Opcode() == Op_StoreL) { 3627 continue; // This StoreL is already optimal. 3628 } 3629 3630 // Store down the constant. 3631 store_constant(tiles, num_tiles, st_off, st_size, con); 3632 3633 intptr_t j = st_off >> LogBytesPerLong; 3634 3635 if (type == T_INT && st_size == BytesPerInt 3636 && (st_off & BytesPerInt) == BytesPerInt) { 3637 jlong lcon = tiles[j]; 3638 if (!Matcher::isSimpleConstant64(lcon) && 3639 st->Opcode() == Op_StoreI) { 3640 // This StoreI is already optimal by itself. 3641 jint* intcon = (jint*) &tiles[j]; 3642 intcon[1] = 0; // undo the store_constant() 3643 3644 // If the previous store is also optimal by itself, back up and 3645 // undo the action of the previous loop iteration... if we can. 3646 // But if we can't, just let the previous half take care of itself. 3647 st = nodes[j]; 3648 st_off -= BytesPerInt; 3649 con = intcon[0]; 3650 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3651 assert(st_off >= header_size, "still ignoring header"); 3652 assert(get_store_offset(st, phase) == st_off, "must be"); 3653 assert(in(i-1) == zmem, "must be"); 3654 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3655 assert(con == tcon->is_int()->get_con(), "must be"); 3656 // Undo the effects of the previous loop trip, which swallowed st: 3657 intcon[0] = 0; // undo store_constant() 3658 set_req(i-1, st); // undo set_req(i, zmem) 3659 nodes[j] = NULL; // undo nodes[j] = st 3660 --old_subword; // undo ++old_subword 3661 } 3662 continue; // This StoreI is already optimal. 3663 } 3664 } 3665 3666 // This store is not needed. 3667 set_req(i, zmem); 3668 nodes[j] = st; // record for the moment 3669 if (st_size < BytesPerLong) // something has changed 3670 ++old_subword; // includes int/float, but who's counting... 3671 else ++old_long; 3672 } 3673 3674 if ((old_subword + old_long) == 0) 3675 return; // nothing more to do 3676 3677 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3678 // Be sure to insert them before overlapping non-constant stores. 3679 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3680 for (int j = 0; j < num_tiles; j++) { 3681 jlong con = tiles[j]; 3682 jlong init = inits[j]; 3683 if (con == 0) continue; 3684 jint con0, con1; // split the constant, address-wise 3685 jint init0, init1; // split the init map, address-wise 3686 { union { jlong con; jint intcon[2]; } u; 3687 u.con = con; 3688 con0 = u.intcon[0]; 3689 con1 = u.intcon[1]; 3690 u.con = init; 3691 init0 = u.intcon[0]; 3692 init1 = u.intcon[1]; 3693 } 3694 3695 Node* old = nodes[j]; 3696 assert(old != NULL, "need the prior store"); 3697 intptr_t offset = (j * BytesPerLong); 3698 3699 bool split = !Matcher::isSimpleConstant64(con); 3700 3701 if (offset < header_size) { 3702 assert(offset + BytesPerInt >= header_size, "second int counts"); 3703 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3704 split = true; // only the second word counts 3705 // Example: int a[] = { 42 ... } 3706 } else if (con0 == 0 && init0 == -1) { 3707 split = true; // first word is covered by full inits 3708 // Example: int a[] = { ... foo(), 42 ... } 3709 } else if (con1 == 0 && init1 == -1) { 3710 split = true; // second word is covered by full inits 3711 // Example: int a[] = { ... 42, foo() ... } 3712 } 3713 3714 // Here's a case where init0 is neither 0 nor -1: 3715 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3716 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3717 // In this case the tile is not split; it is (jlong)42. 3718 // The big tile is stored down, and then the foo() value is inserted. 3719 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3720 3721 Node* ctl = old->in(MemNode::Control); 3722 Node* adr = make_raw_address(offset, phase); 3723 const TypePtr* atp = TypeRawPtr::BOTTOM; 3724 3725 // One or two coalesced stores to plop down. 3726 Node* st[2]; 3727 intptr_t off[2]; 3728 int nst = 0; 3729 if (!split) { 3730 ++new_long; 3731 off[nst] = offset; 3732 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3733 phase->longcon(con), T_LONG, MemNode::unordered); 3734 } else { 3735 // Omit either if it is a zero. 3736 if (con0 != 0) { 3737 ++new_int; 3738 off[nst] = offset; 3739 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3740 phase->intcon(con0), T_INT, MemNode::unordered); 3741 } 3742 if (con1 != 0) { 3743 ++new_int; 3744 offset += BytesPerInt; 3745 adr = make_raw_address(offset, phase); 3746 off[nst] = offset; 3747 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3748 phase->intcon(con1), T_INT, MemNode::unordered); 3749 } 3750 } 3751 3752 // Insert second store first, then the first before the second. 3753 // Insert each one just before any overlapping non-constant stores. 3754 while (nst > 0) { 3755 Node* st1 = st[--nst]; 3756 C->copy_node_notes_to(st1, old); 3757 st1 = phase->transform(st1); 3758 offset = off[nst]; 3759 assert(offset >= header_size, "do not smash header"); 3760 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3761 guarantee(ins_idx != 0, "must re-insert constant store"); 3762 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3763 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3764 set_req(--ins_idx, st1); 3765 else 3766 ins_req(ins_idx, st1); 3767 } 3768 } 3769 3770 if (PrintCompilation && WizardMode) 3771 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3772 old_subword, old_long, new_int, new_long); 3773 if (C->log() != NULL) 3774 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3775 old_subword, old_long, new_int, new_long); 3776 3777 // Clean up any remaining occurrences of zmem: 3778 remove_extra_zeroes(); 3779 } 3780 3781 // Explore forward from in(start) to find the first fully initialized 3782 // word, and return its offset. Skip groups of subword stores which 3783 // together initialize full words. If in(start) is itself part of a 3784 // fully initialized word, return the offset of in(start). If there 3785 // are no following full-word stores, or if something is fishy, return 3786 // a negative value. 3787 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3788 int int_map = 0; 3789 intptr_t int_map_off = 0; 3790 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3791 3792 for (uint i = start, limit = req(); i < limit; i++) { 3793 Node* st = in(i); 3794 3795 intptr_t st_off = get_store_offset(st, phase); 3796 if (st_off < 0) break; // return conservative answer 3797 3798 int st_size = st->as_Store()->memory_size(); 3799 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3800 return st_off; // we found a complete word init 3801 } 3802 3803 // update the map: 3804 3805 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3806 if (this_int_off != int_map_off) { 3807 // reset the map: 3808 int_map = 0; 3809 int_map_off = this_int_off; 3810 } 3811 3812 int subword_off = st_off - this_int_off; 3813 int_map |= right_n_bits(st_size) << subword_off; 3814 if ((int_map & FULL_MAP) == FULL_MAP) { 3815 return this_int_off; // we found a complete word init 3816 } 3817 3818 // Did this store hit or cross the word boundary? 3819 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3820 if (next_int_off == this_int_off + BytesPerInt) { 3821 // We passed the current int, without fully initializing it. 3822 int_map_off = next_int_off; 3823 int_map >>= BytesPerInt; 3824 } else if (next_int_off > this_int_off + BytesPerInt) { 3825 // We passed the current and next int. 3826 return this_int_off + BytesPerInt; 3827 } 3828 } 3829 3830 return -1; 3831 } 3832 3833 3834 // Called when the associated AllocateNode is expanded into CFG. 3835 // At this point, we may perform additional optimizations. 3836 // Linearize the stores by ascending offset, to make memory 3837 // activity as coherent as possible. 3838 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3839 intptr_t header_size, 3840 Node* size_in_bytes, 3841 PhaseGVN* phase) { 3842 assert(!is_complete(), "not already complete"); 3843 assert(stores_are_sane(phase), ""); 3844 assert(allocation() != NULL, "must be present"); 3845 3846 remove_extra_zeroes(); 3847 3848 if (ReduceFieldZeroing || ReduceBulkZeroing) 3849 // reduce instruction count for common initialization patterns 3850 coalesce_subword_stores(header_size, size_in_bytes, phase); 3851 3852 Node* zmem = zero_memory(); // initially zero memory state 3853 Node* inits = zmem; // accumulating a linearized chain of inits 3854 #ifdef ASSERT 3855 intptr_t first_offset = allocation()->minimum_header_size(); 3856 intptr_t last_init_off = first_offset; // previous init offset 3857 intptr_t last_init_end = first_offset; // previous init offset+size 3858 intptr_t last_tile_end = first_offset; // previous tile offset+size 3859 #endif 3860 intptr_t zeroes_done = header_size; 3861 3862 bool do_zeroing = true; // we might give up if inits are very sparse 3863 int big_init_gaps = 0; // how many large gaps have we seen? 3864 3865 if (UseTLAB && ZeroTLAB) do_zeroing = false; 3866 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3867 3868 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3869 Node* st = in(i); 3870 intptr_t st_off = get_store_offset(st, phase); 3871 if (st_off < 0) 3872 break; // unknown junk in the inits 3873 if (st->in(MemNode::Memory) != zmem) 3874 break; // complicated store chains somehow in list 3875 3876 int st_size = st->as_Store()->memory_size(); 3877 intptr_t next_init_off = st_off + st_size; 3878 3879 if (do_zeroing && zeroes_done < next_init_off) { 3880 // See if this store needs a zero before it or under it. 3881 intptr_t zeroes_needed = st_off; 3882 3883 if (st_size < BytesPerInt) { 3884 // Look for subword stores which only partially initialize words. 3885 // If we find some, we must lay down some word-level zeroes first, 3886 // underneath the subword stores. 3887 // 3888 // Examples: 3889 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3890 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3891 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3892 // 3893 // Note: coalesce_subword_stores may have already done this, 3894 // if it was prompted by constant non-zero subword initializers. 3895 // But this case can still arise with non-constant stores. 3896 3897 intptr_t next_full_store = find_next_fullword_store(i, phase); 3898 3899 // In the examples above: 3900 // in(i) p q r s x y z 3901 // st_off 12 13 14 15 12 13 14 3902 // st_size 1 1 1 1 1 1 1 3903 // next_full_s. 12 16 16 16 16 16 16 3904 // z's_done 12 16 16 16 12 16 12 3905 // z's_needed 12 16 16 16 16 16 16 3906 // zsize 0 0 0 0 4 0 4 3907 if (next_full_store < 0) { 3908 // Conservative tack: Zero to end of current word. 3909 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3910 } else { 3911 // Zero to beginning of next fully initialized word. 3912 // Or, don't zero at all, if we are already in that word. 3913 assert(next_full_store >= zeroes_needed, "must go forward"); 3914 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3915 zeroes_needed = next_full_store; 3916 } 3917 } 3918 3919 if (zeroes_needed > zeroes_done) { 3920 intptr_t zsize = zeroes_needed - zeroes_done; 3921 // Do some incremental zeroing on rawmem, in parallel with inits. 3922 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3923 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3924 zeroes_done, zeroes_needed, 3925 phase); 3926 zeroes_done = zeroes_needed; 3927 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 3928 do_zeroing = false; // leave the hole, next time 3929 } 3930 } 3931 3932 // Collect the store and move on: 3933 st->set_req(MemNode::Memory, inits); 3934 inits = st; // put it on the linearized chain 3935 set_req(i, zmem); // unhook from previous position 3936 3937 if (zeroes_done == st_off) 3938 zeroes_done = next_init_off; 3939 3940 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3941 3942 #ifdef ASSERT 3943 // Various order invariants. Weaker than stores_are_sane because 3944 // a large constant tile can be filled in by smaller non-constant stores. 3945 assert(st_off >= last_init_off, "inits do not reverse"); 3946 last_init_off = st_off; 3947 const Type* val = NULL; 3948 if (st_size >= BytesPerInt && 3949 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3950 (int)val->basic_type() < (int)T_OBJECT) { 3951 assert(st_off >= last_tile_end, "tiles do not overlap"); 3952 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3953 last_tile_end = MAX2(last_tile_end, next_init_off); 3954 } else { 3955 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3956 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3957 assert(st_off >= last_init_end, "inits do not overlap"); 3958 last_init_end = next_init_off; // it's a non-tile 3959 } 3960 #endif //ASSERT 3961 } 3962 3963 remove_extra_zeroes(); // clear out all the zmems left over 3964 add_req(inits); 3965 3966 if (!(UseTLAB && ZeroTLAB)) { 3967 // If anything remains to be zeroed, zero it all now. 3968 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3969 // if it is the last unused 4 bytes of an instance, forget about it 3970 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3971 if (zeroes_done + BytesPerLong >= size_limit) { 3972 assert(allocation() != NULL, ""); 3973 if (allocation()->Opcode() == Op_Allocate) { 3974 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3975 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3976 if (zeroes_done == k->layout_helper()) 3977 zeroes_done = size_limit; 3978 } 3979 } 3980 if (zeroes_done < size_limit) { 3981 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3982 zeroes_done, size_in_bytes, phase); 3983 } 3984 } 3985 3986 set_complete(phase); 3987 return rawmem; 3988 } 3989 3990 3991 #ifdef ASSERT 3992 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3993 if (is_complete()) 3994 return true; // stores could be anything at this point 3995 assert(allocation() != NULL, "must be present"); 3996 intptr_t last_off = allocation()->minimum_header_size(); 3997 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3998 Node* st = in(i); 3999 intptr_t st_off = get_store_offset(st, phase); 4000 if (st_off < 0) continue; // ignore dead garbage 4001 if (last_off > st_off) { 4002 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4003 this->dump(2); 4004 assert(false, "ascending store offsets"); 4005 return false; 4006 } 4007 last_off = st_off + st->as_Store()->memory_size(); 4008 } 4009 return true; 4010 } 4011 #endif //ASSERT 4012 4013 4014 4015 4016 //============================MergeMemNode===================================== 4017 // 4018 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4019 // contributing store or call operations. Each contributor provides the memory 4020 // state for a particular "alias type" (see Compile::alias_type). For example, 4021 // if a MergeMem has an input X for alias category #6, then any memory reference 4022 // to alias category #6 may use X as its memory state input, as an exact equivalent 4023 // to using the MergeMem as a whole. 4024 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4025 // 4026 // (Here, the <N> notation gives the index of the relevant adr_type.) 4027 // 4028 // In one special case (and more cases in the future), alias categories overlap. 4029 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4030 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4031 // it is exactly equivalent to that state W: 4032 // MergeMem(<Bot>: W) <==> W 4033 // 4034 // Usually, the merge has more than one input. In that case, where inputs 4035 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4036 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4037 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4038 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4039 // 4040 // A merge can take a "wide" memory state as one of its narrow inputs. 4041 // This simply means that the merge observes out only the relevant parts of 4042 // the wide input. That is, wide memory states arriving at narrow merge inputs 4043 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4044 // 4045 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4046 // and that memory slices "leak through": 4047 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4048 // 4049 // But, in such a cascade, repeated memory slices can "block the leak": 4050 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4051 // 4052 // In the last example, Y is not part of the combined memory state of the 4053 // outermost MergeMem. The system must, of course, prevent unschedulable 4054 // memory states from arising, so you can be sure that the state Y is somehow 4055 // a precursor to state Y'. 4056 // 4057 // 4058 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4059 // of each MergeMemNode array are exactly the numerical alias indexes, including 4060 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4061 // Compile::alias_type (and kin) produce and manage these indexes. 4062 // 4063 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4064 // (Note that this provides quick access to the top node inside MergeMem methods, 4065 // without the need to reach out via TLS to Compile::current.) 4066 // 4067 // As a consequence of what was just described, a MergeMem that represents a full 4068 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4069 // containing all alias categories. 4070 // 4071 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4072 // 4073 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4074 // a memory state for the alias type <N>, or else the top node, meaning that 4075 // there is no particular input for that alias type. Note that the length of 4076 // a MergeMem is variable, and may be extended at any time to accommodate new 4077 // memory states at larger alias indexes. When merges grow, they are of course 4078 // filled with "top" in the unused in() positions. 4079 // 4080 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4081 // (Top was chosen because it works smoothly with passes like GCM.) 4082 // 4083 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4084 // the type of random VM bits like TLS references.) Since it is always the 4085 // first non-Bot memory slice, some low-level loops use it to initialize an 4086 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4087 // 4088 // 4089 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4090 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4091 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4092 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4093 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4094 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4095 // 4096 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4097 // really that different from the other memory inputs. An abbreviation called 4098 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4099 // 4100 // 4101 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4102 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4103 // that "emerges though" the base memory will be marked as excluding the alias types 4104 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4105 // 4106 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4107 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4108 // 4109 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4110 // (It is currently unimplemented.) As you can see, the resulting merge is 4111 // actually a disjoint union of memory states, rather than an overlay. 4112 // 4113 4114 //------------------------------MergeMemNode----------------------------------- 4115 Node* MergeMemNode::make_empty_memory() { 4116 Node* empty_memory = (Node*) Compile::current()->top(); 4117 assert(empty_memory->is_top(), "correct sentinel identity"); 4118 return empty_memory; 4119 } 4120 4121 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4122 init_class_id(Class_MergeMem); 4123 // all inputs are nullified in Node::Node(int) 4124 // set_input(0, NULL); // no control input 4125 4126 // Initialize the edges uniformly to top, for starters. 4127 Node* empty_mem = make_empty_memory(); 4128 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4129 init_req(i,empty_mem); 4130 } 4131 assert(empty_memory() == empty_mem, ""); 4132 4133 if( new_base != NULL && new_base->is_MergeMem() ) { 4134 MergeMemNode* mdef = new_base->as_MergeMem(); 4135 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4136 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4137 mms.set_memory(mms.memory2()); 4138 } 4139 assert(base_memory() == mdef->base_memory(), ""); 4140 } else { 4141 set_base_memory(new_base); 4142 } 4143 } 4144 4145 // Make a new, untransformed MergeMem with the same base as 'mem'. 4146 // If mem is itself a MergeMem, populate the result with the same edges. 4147 MergeMemNode* MergeMemNode::make(Node* mem) { 4148 return new MergeMemNode(mem); 4149 } 4150 4151 //------------------------------cmp-------------------------------------------- 4152 uint MergeMemNode::hash() const { return NO_HASH; } 4153 uint MergeMemNode::cmp( const Node &n ) const { 4154 return (&n == this); // Always fail except on self 4155 } 4156 4157 //------------------------------Identity--------------------------------------- 4158 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4159 // Identity if this merge point does not record any interesting memory 4160 // disambiguations. 4161 Node* base_mem = base_memory(); 4162 Node* empty_mem = empty_memory(); 4163 if (base_mem != empty_mem) { // Memory path is not dead? 4164 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4165 Node* mem = in(i); 4166 if (mem != empty_mem && mem != base_mem) { 4167 return this; // Many memory splits; no change 4168 } 4169 } 4170 } 4171 return base_mem; // No memory splits; ID on the one true input 4172 } 4173 4174 //------------------------------Ideal------------------------------------------ 4175 // This method is invoked recursively on chains of MergeMem nodes 4176 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4177 // Remove chain'd MergeMems 4178 // 4179 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4180 // relative to the "in(Bot)". Since we are patching both at the same time, 4181 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4182 // but rewrite each "in(i)" relative to the new "in(Bot)". 4183 Node *progress = NULL; 4184 4185 4186 Node* old_base = base_memory(); 4187 Node* empty_mem = empty_memory(); 4188 if (old_base == empty_mem) 4189 return NULL; // Dead memory path. 4190 4191 MergeMemNode* old_mbase; 4192 if (old_base != NULL && old_base->is_MergeMem()) 4193 old_mbase = old_base->as_MergeMem(); 4194 else 4195 old_mbase = NULL; 4196 Node* new_base = old_base; 4197 4198 // simplify stacked MergeMems in base memory 4199 if (old_mbase) new_base = old_mbase->base_memory(); 4200 4201 // the base memory might contribute new slices beyond my req() 4202 if (old_mbase) grow_to_match(old_mbase); 4203 4204 // Look carefully at the base node if it is a phi. 4205 PhiNode* phi_base; 4206 if (new_base != NULL && new_base->is_Phi()) 4207 phi_base = new_base->as_Phi(); 4208 else 4209 phi_base = NULL; 4210 4211 Node* phi_reg = NULL; 4212 uint phi_len = (uint)-1; 4213 if (phi_base != NULL && !phi_base->is_copy()) { 4214 // do not examine phi if degraded to a copy 4215 phi_reg = phi_base->region(); 4216 phi_len = phi_base->req(); 4217 // see if the phi is unfinished 4218 for (uint i = 1; i < phi_len; i++) { 4219 if (phi_base->in(i) == NULL) { 4220 // incomplete phi; do not look at it yet! 4221 phi_reg = NULL; 4222 phi_len = (uint)-1; 4223 break; 4224 } 4225 } 4226 } 4227 4228 // Note: We do not call verify_sparse on entry, because inputs 4229 // can normalize to the base_memory via subsume_node or similar 4230 // mechanisms. This method repairs that damage. 4231 4232 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4233 4234 // Look at each slice. 4235 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4236 Node* old_in = in(i); 4237 // calculate the old memory value 4238 Node* old_mem = old_in; 4239 if (old_mem == empty_mem) old_mem = old_base; 4240 assert(old_mem == memory_at(i), ""); 4241 4242 // maybe update (reslice) the old memory value 4243 4244 // simplify stacked MergeMems 4245 Node* new_mem = old_mem; 4246 MergeMemNode* old_mmem; 4247 if (old_mem != NULL && old_mem->is_MergeMem()) 4248 old_mmem = old_mem->as_MergeMem(); 4249 else 4250 old_mmem = NULL; 4251 if (old_mmem == this) { 4252 // This can happen if loops break up and safepoints disappear. 4253 // A merge of BotPtr (default) with a RawPtr memory derived from a 4254 // safepoint can be rewritten to a merge of the same BotPtr with 4255 // the BotPtr phi coming into the loop. If that phi disappears 4256 // also, we can end up with a self-loop of the mergemem. 4257 // In general, if loops degenerate and memory effects disappear, 4258 // a mergemem can be left looking at itself. This simply means 4259 // that the mergemem's default should be used, since there is 4260 // no longer any apparent effect on this slice. 4261 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4262 // from start. Update the input to TOP. 4263 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4264 } 4265 else if (old_mmem != NULL) { 4266 new_mem = old_mmem->memory_at(i); 4267 } 4268 // else preceding memory was not a MergeMem 4269 4270 // replace equivalent phis (unfortunately, they do not GVN together) 4271 if (new_mem != NULL && new_mem != new_base && 4272 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4273 if (new_mem->is_Phi()) { 4274 PhiNode* phi_mem = new_mem->as_Phi(); 4275 for (uint i = 1; i < phi_len; i++) { 4276 if (phi_base->in(i) != phi_mem->in(i)) { 4277 phi_mem = NULL; 4278 break; 4279 } 4280 } 4281 if (phi_mem != NULL) { 4282 // equivalent phi nodes; revert to the def 4283 new_mem = new_base; 4284 } 4285 } 4286 } 4287 4288 // maybe store down a new value 4289 Node* new_in = new_mem; 4290 if (new_in == new_base) new_in = empty_mem; 4291 4292 if (new_in != old_in) { 4293 // Warning: Do not combine this "if" with the previous "if" 4294 // A memory slice might have be be rewritten even if it is semantically 4295 // unchanged, if the base_memory value has changed. 4296 set_req(i, new_in); 4297 progress = this; // Report progress 4298 } 4299 } 4300 4301 if (new_base != old_base) { 4302 set_req(Compile::AliasIdxBot, new_base); 4303 // Don't use set_base_memory(new_base), because we need to update du. 4304 assert(base_memory() == new_base, ""); 4305 progress = this; 4306 } 4307 4308 if( base_memory() == this ) { 4309 // a self cycle indicates this memory path is dead 4310 set_req(Compile::AliasIdxBot, empty_mem); 4311 } 4312 4313 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4314 // Recursion must occur after the self cycle check above 4315 if( base_memory()->is_MergeMem() ) { 4316 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4317 Node *m = phase->transform(new_mbase); // Rollup any cycles 4318 if( m != NULL && (m->is_top() || 4319 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4320 // propagate rollup of dead cycle to self 4321 set_req(Compile::AliasIdxBot, empty_mem); 4322 } 4323 } 4324 4325 if( base_memory() == empty_mem ) { 4326 progress = this; 4327 // Cut inputs during Parse phase only. 4328 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4329 if( !can_reshape ) { 4330 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4331 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4332 } 4333 } 4334 } 4335 4336 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4337 // Check if PhiNode::Ideal's "Split phis through memory merges" 4338 // transform should be attempted. Look for this->phi->this cycle. 4339 uint merge_width = req(); 4340 if (merge_width > Compile::AliasIdxRaw) { 4341 PhiNode* phi = base_memory()->as_Phi(); 4342 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4343 if (phi->in(i) == this) { 4344 phase->is_IterGVN()->_worklist.push(phi); 4345 break; 4346 } 4347 } 4348 } 4349 } 4350 4351 assert(progress || verify_sparse(), "please, no dups of base"); 4352 return progress; 4353 } 4354 4355 //-------------------------set_base_memory------------------------------------- 4356 void MergeMemNode::set_base_memory(Node *new_base) { 4357 Node* empty_mem = empty_memory(); 4358 set_req(Compile::AliasIdxBot, new_base); 4359 assert(memory_at(req()) == new_base, "must set default memory"); 4360 // Clear out other occurrences of new_base: 4361 if (new_base != empty_mem) { 4362 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4363 if (in(i) == new_base) set_req(i, empty_mem); 4364 } 4365 } 4366 } 4367 4368 //------------------------------out_RegMask------------------------------------ 4369 const RegMask &MergeMemNode::out_RegMask() const { 4370 return RegMask::Empty; 4371 } 4372 4373 //------------------------------dump_spec-------------------------------------- 4374 #ifndef PRODUCT 4375 void MergeMemNode::dump_spec(outputStream *st) const { 4376 st->print(" {"); 4377 Node* base_mem = base_memory(); 4378 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4379 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4380 if (mem == base_mem) { st->print(" -"); continue; } 4381 st->print( " N%d:", mem->_idx ); 4382 Compile::current()->get_adr_type(i)->dump_on(st); 4383 } 4384 st->print(" }"); 4385 } 4386 #endif // !PRODUCT 4387 4388 4389 #ifdef ASSERT 4390 static bool might_be_same(Node* a, Node* b) { 4391 if (a == b) return true; 4392 if (!(a->is_Phi() || b->is_Phi())) return false; 4393 // phis shift around during optimization 4394 return true; // pretty stupid... 4395 } 4396 4397 // verify a narrow slice (either incoming or outgoing) 4398 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4399 if (!VerifyAliases) return; // don't bother to verify unless requested 4400 if (is_error_reported()) return; // muzzle asserts when debugging an error 4401 if (Node::in_dump()) return; // muzzle asserts when printing 4402 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4403 assert(n != NULL, ""); 4404 // Elide intervening MergeMem's 4405 while (n->is_MergeMem()) { 4406 n = n->as_MergeMem()->memory_at(alias_idx); 4407 } 4408 Compile* C = Compile::current(); 4409 const TypePtr* n_adr_type = n->adr_type(); 4410 if (n == m->empty_memory()) { 4411 // Implicit copy of base_memory() 4412 } else if (n_adr_type != TypePtr::BOTTOM) { 4413 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4414 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4415 } else { 4416 // A few places like make_runtime_call "know" that VM calls are narrow, 4417 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4418 bool expected_wide_mem = false; 4419 if (n == m->base_memory()) { 4420 expected_wide_mem = true; 4421 } else if (alias_idx == Compile::AliasIdxRaw || 4422 n == m->memory_at(Compile::AliasIdxRaw)) { 4423 expected_wide_mem = true; 4424 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4425 // memory can "leak through" calls on channels that 4426 // are write-once. Allow this also. 4427 expected_wide_mem = true; 4428 } 4429 assert(expected_wide_mem, "expected narrow slice replacement"); 4430 } 4431 } 4432 #else // !ASSERT 4433 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4434 #endif 4435 4436 4437 //-----------------------------memory_at--------------------------------------- 4438 Node* MergeMemNode::memory_at(uint alias_idx) const { 4439 assert(alias_idx >= Compile::AliasIdxRaw || 4440 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4441 "must avoid base_memory and AliasIdxTop"); 4442 4443 // Otherwise, it is a narrow slice. 4444 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4445 Compile *C = Compile::current(); 4446 if (is_empty_memory(n)) { 4447 // the array is sparse; empty slots are the "top" node 4448 n = base_memory(); 4449 assert(Node::in_dump() 4450 || n == NULL || n->bottom_type() == Type::TOP 4451 || n->adr_type() == NULL // address is TOP 4452 || n->adr_type() == TypePtr::BOTTOM 4453 || n->adr_type() == TypeRawPtr::BOTTOM 4454 || Compile::current()->AliasLevel() == 0, 4455 "must be a wide memory"); 4456 // AliasLevel == 0 if we are organizing the memory states manually. 4457 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4458 } else { 4459 // make sure the stored slice is sane 4460 #ifdef ASSERT 4461 if (is_error_reported() || Node::in_dump()) { 4462 } else if (might_be_same(n, base_memory())) { 4463 // Give it a pass: It is a mostly harmless repetition of the base. 4464 // This can arise normally from node subsumption during optimization. 4465 } else { 4466 verify_memory_slice(this, alias_idx, n); 4467 } 4468 #endif 4469 } 4470 return n; 4471 } 4472 4473 //---------------------------set_memory_at------------------------------------- 4474 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4475 verify_memory_slice(this, alias_idx, n); 4476 Node* empty_mem = empty_memory(); 4477 if (n == base_memory()) n = empty_mem; // collapse default 4478 uint need_req = alias_idx+1; 4479 if (req() < need_req) { 4480 if (n == empty_mem) return; // already the default, so do not grow me 4481 // grow the sparse array 4482 do { 4483 add_req(empty_mem); 4484 } while (req() < need_req); 4485 } 4486 set_req( alias_idx, n ); 4487 } 4488 4489 4490 4491 //--------------------------iteration_setup------------------------------------ 4492 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4493 if (other != NULL) { 4494 grow_to_match(other); 4495 // invariant: the finite support of mm2 is within mm->req() 4496 #ifdef ASSERT 4497 for (uint i = req(); i < other->req(); i++) { 4498 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4499 } 4500 #endif 4501 } 4502 // Replace spurious copies of base_memory by top. 4503 Node* base_mem = base_memory(); 4504 if (base_mem != NULL && !base_mem->is_top()) { 4505 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4506 if (in(i) == base_mem) 4507 set_req(i, empty_memory()); 4508 } 4509 } 4510 } 4511 4512 //---------------------------grow_to_match------------------------------------- 4513 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4514 Node* empty_mem = empty_memory(); 4515 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4516 // look for the finite support of the other memory 4517 for (uint i = other->req(); --i >= req(); ) { 4518 if (other->in(i) != empty_mem) { 4519 uint new_len = i+1; 4520 while (req() < new_len) add_req(empty_mem); 4521 break; 4522 } 4523 } 4524 } 4525 4526 //---------------------------verify_sparse------------------------------------- 4527 #ifndef PRODUCT 4528 bool MergeMemNode::verify_sparse() const { 4529 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4530 Node* base_mem = base_memory(); 4531 // The following can happen in degenerate cases, since empty==top. 4532 if (is_empty_memory(base_mem)) return true; 4533 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4534 assert(in(i) != NULL, "sane slice"); 4535 if (in(i) == base_mem) return false; // should have been the sentinel value! 4536 } 4537 return true; 4538 } 4539 4540 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4541 Node* n; 4542 n = mm->in(idx); 4543 if (mem == n) return true; // might be empty_memory() 4544 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4545 if (mem == n) return true; 4546 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4547 if (mem == n) return true; 4548 if (n == NULL) break; 4549 } 4550 return false; 4551 } 4552 #endif // !PRODUCT