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