memnode.cpp revision 3602:da91efe96a93
1/* 2 * Copyright (c) 1997, 2012, 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/cfgnode.hpp" 32#include "opto/compile.hpp" 33#include "opto/connode.hpp" 34#include "opto/loopnode.hpp" 35#include "opto/machnode.hpp" 36#include "opto/matcher.hpp" 37#include "opto/memnode.hpp" 38#include "opto/mulnode.hpp" 39#include "opto/phaseX.hpp" 40#include "opto/regmask.hpp" 41 42// Portions of code courtesy of Clifford Click 43 44// Optimization - Graph Style 45 46static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 47 48//============================================================================= 49uint MemNode::size_of() const { return sizeof(*this); } 50 51const TypePtr *MemNode::adr_type() const { 52 Node* adr = in(Address); 53 const TypePtr* cross_check = NULL; 54 DEBUG_ONLY(cross_check = _adr_type); 55 return calculate_adr_type(adr->bottom_type(), cross_check); 56} 57 58#ifndef PRODUCT 59void MemNode::dump_spec(outputStream *st) const { 60 if (in(Address) == NULL) return; // node is dead 61#ifndef ASSERT 62 // fake the missing field 63 const TypePtr* _adr_type = NULL; 64 if (in(Address) != NULL) 65 _adr_type = in(Address)->bottom_type()->isa_ptr(); 66#endif 67 dump_adr_type(this, _adr_type, st); 68 69 Compile* C = Compile::current(); 70 if( C->alias_type(_adr_type)->is_volatile() ) 71 st->print(" Volatile!"); 72} 73 74void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 75 st->print(" @"); 76 if (adr_type == NULL) { 77 st->print("NULL"); 78 } else { 79 adr_type->dump_on(st); 80 Compile* C = Compile::current(); 81 Compile::AliasType* atp = NULL; 82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 83 if (atp == NULL) 84 st->print(", idx=?\?;"); 85 else if (atp->index() == Compile::AliasIdxBot) 86 st->print(", idx=Bot;"); 87 else if (atp->index() == Compile::AliasIdxTop) 88 st->print(", idx=Top;"); 89 else if (atp->index() == Compile::AliasIdxRaw) 90 st->print(", idx=Raw;"); 91 else { 92 ciField* field = atp->field(); 93 if (field) { 94 st->print(", name="); 95 field->print_name_on(st); 96 } 97 st->print(", idx=%d;", atp->index()); 98 } 99 } 100} 101 102extern void print_alias_types(); 103 104#endif 105 106Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 107 const TypeOopPtr *tinst = t_adr->isa_oopptr(); 108 if (tinst == NULL || !tinst->is_known_instance_field()) 109 return mchain; // don't try to optimize non-instance types 110 uint instance_id = tinst->instance_id(); 111 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 112 Node *prev = NULL; 113 Node *result = mchain; 114 while (prev != result) { 115 prev = result; 116 if (result == start_mem) 117 break; // hit one of our sentinels 118 // skip over a call which does not affect this memory slice 119 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 120 Node *proj_in = result->in(0); 121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 122 break; // hit one of our sentinels 123 } else if (proj_in->is_Call()) { 124 CallNode *call = proj_in->as_Call(); 125 if (!call->may_modify(t_adr, phase)) { 126 result = call->in(TypeFunc::Memory); 127 } 128 } else if (proj_in->is_Initialize()) { 129 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 130 // Stop if this is the initialization for the object instance which 131 // which contains this memory slice, otherwise skip over it. 132 if (alloc != NULL && alloc->_idx != instance_id) { 133 result = proj_in->in(TypeFunc::Memory); 134 } 135 } else if (proj_in->is_MemBar()) { 136 result = proj_in->in(TypeFunc::Memory); 137 } else { 138 assert(false, "unexpected projection"); 139 } 140 } else if (result->is_ClearArray()) { 141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) { 142 // Can not bypass initialization of the instance 143 // we are looking for. 144 break; 145 } 146 // Otherwise skip it (the call updated 'result' value). 147 } else if (result->is_MergeMem()) { 148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 149 } 150 } 151 return result; 152} 153 154Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 155 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); 157 PhaseIterGVN *igvn = phase->is_IterGVN(); 158 Node *result = mchain; 159 result = optimize_simple_memory_chain(result, t_adr, phase); 160 if (is_instance && igvn != NULL && result->is_Phi()) { 161 PhiNode *mphi = result->as_Phi(); 162 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 163 const TypePtr *t = mphi->adr_type(); 164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 166 t->is_oopptr()->cast_to_exactness(true) 167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 168 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 169 // clone the Phi with our address type 170 result = mphi->split_out_instance(t_adr, igvn); 171 } else { 172 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 173 } 174 } 175 return result; 176} 177 178static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 179 uint alias_idx = phase->C->get_alias_index(tp); 180 Node *mem = mmem; 181#ifdef ASSERT 182 { 183 // Check that current type is consistent with the alias index used during graph construction 184 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 185 bool consistent = adr_check == NULL || adr_check->empty() || 186 phase->C->must_alias(adr_check, alias_idx ); 187 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 188 if( !consistent && adr_check != NULL && !adr_check->empty() && 189 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 190 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 191 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 192 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 193 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 194 // don't assert if it is dead code. 195 consistent = true; 196 } 197 if( !consistent ) { 198 st->print("alias_idx==%d, adr_check==", alias_idx); 199 if( adr_check == NULL ) { 200 st->print("NULL"); 201 } else { 202 adr_check->dump(); 203 } 204 st->cr(); 205 print_alias_types(); 206 assert(consistent, "adr_check must match alias idx"); 207 } 208 } 209#endif 210 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 211 // means an array I have not precisely typed yet. Do not do any 212 // alias stuff with it any time soon. 213 const TypeOopPtr *toop = tp->isa_oopptr(); 214 if( tp->base() != Type::AnyPtr && 215 !(toop && 216 toop->klass() != NULL && 217 toop->klass()->is_java_lang_Object() && 218 toop->offset() == Type::OffsetBot) ) { 219 // compress paths and change unreachable cycles to TOP 220 // If not, we can update the input infinitely along a MergeMem cycle 221 // Equivalent code in PhiNode::Ideal 222 Node* m = phase->transform(mmem); 223 // If transformed to a MergeMem, get the desired slice 224 // Otherwise the returned node represents memory for every slice 225 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 226 // Update input if it is progress over what we have now 227 } 228 return mem; 229} 230 231//--------------------------Ideal_common--------------------------------------- 232// Look for degenerate control and memory inputs. Bypass MergeMem inputs. 233// Unhook non-raw memories from complete (macro-expanded) initializations. 234Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 235 // If our control input is a dead region, kill all below the region 236 Node *ctl = in(MemNode::Control); 237 if (ctl && remove_dead_region(phase, can_reshape)) 238 return this; 239 ctl = in(MemNode::Control); 240 // Don't bother trying to transform a dead node 241 if( ctl && ctl->is_top() ) return NodeSentinel; 242 243 PhaseIterGVN *igvn = phase->is_IterGVN(); 244 // Wait if control on the worklist. 245 if (ctl && can_reshape && igvn != NULL) { 246 Node* bol = NULL; 247 Node* cmp = NULL; 248 if (ctl->in(0)->is_If()) { 249 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 250 bol = ctl->in(0)->in(1); 251 if (bol->is_Bool()) 252 cmp = ctl->in(0)->in(1)->in(1); 253 } 254 if (igvn->_worklist.member(ctl) || 255 (bol != NULL && igvn->_worklist.member(bol)) || 256 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 257 // This control path may be dead. 258 // Delay this memory node transformation until the control is processed. 259 phase->is_IterGVN()->_worklist.push(this); 260 return NodeSentinel; // caller will return NULL 261 } 262 } 263 // Ignore if memory is dead, or self-loop 264 Node *mem = in(MemNode::Memory); 265 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 266 assert( mem != this, "dead loop in MemNode::Ideal" ); 267 268 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 269 // This memory slice may be dead. 270 // Delay this mem node transformation until the memory is processed. 271 phase->is_IterGVN()->_worklist.push(this); 272 return NodeSentinel; // caller will return NULL 273 } 274 275 Node *address = in(MemNode::Address); 276 const Type *t_adr = phase->type( address ); 277 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 278 279 if( can_reshape && igvn != NULL && 280 (igvn->_worklist.member(address) || 281 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) { 282 // The address's base and type may change when the address is processed. 283 // Delay this mem node transformation until the address is processed. 284 phase->is_IterGVN()->_worklist.push(this); 285 return NodeSentinel; // caller will return NULL 286 } 287 288 // Do NOT remove or optimize the next lines: ensure a new alias index 289 // is allocated for an oop pointer type before Escape Analysis. 290 // Note: C++ will not remove it since the call has side effect. 291 if ( t_adr->isa_oopptr() ) { 292 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 293 } 294 295#ifdef ASSERT 296 Node* base = NULL; 297 if (address->is_AddP()) 298 base = address->in(AddPNode::Base); 299 assert(base == NULL || t_adr->isa_rawptr() || 300 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); 301#endif 302 303 // Avoid independent memory operations 304 Node* old_mem = mem; 305 306 // The code which unhooks non-raw memories from complete (macro-expanded) 307 // initializations was removed. After macro-expansion all stores catched 308 // by Initialize node became raw stores and there is no information 309 // which memory slices they modify. So it is unsafe to move any memory 310 // operation above these stores. Also in most cases hooked non-raw memories 311 // were already unhooked by using information from detect_ptr_independence() 312 // and find_previous_store(). 313 314 if (mem->is_MergeMem()) { 315 MergeMemNode* mmem = mem->as_MergeMem(); 316 const TypePtr *tp = t_adr->is_ptr(); 317 318 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 319 } 320 321 if (mem != old_mem) { 322 set_req(MemNode::Memory, mem); 323 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 324 return this; 325 } 326 327 // let the subclass continue analyzing... 328 return NULL; 329} 330 331// Helper function for proving some simple control dominations. 332// Attempt to prove that all control inputs of 'dom' dominate 'sub'. 333// Already assumes that 'dom' is available at 'sub', and that 'sub' 334// is not a constant (dominated by the method's StartNode). 335// Used by MemNode::find_previous_store to prove that the 336// control input of a memory operation predates (dominates) 337// an allocation it wants to look past. 338bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 339 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 340 return false; // Conservative answer for dead code 341 342 // Check 'dom'. Skip Proj and CatchProj nodes. 343 dom = dom->find_exact_control(dom); 344 if (dom == NULL || dom->is_top()) 345 return false; // Conservative answer for dead code 346 347 if (dom == sub) { 348 // For the case when, for example, 'sub' is Initialize and the original 349 // 'dom' is Proj node of the 'sub'. 350 return false; 351 } 352 353 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 354 return true; 355 356 // 'dom' dominates 'sub' if its control edge and control edges 357 // of all its inputs dominate or equal to sub's control edge. 358 359 // Currently 'sub' is either Allocate, Initialize or Start nodes. 360 // Or Region for the check in LoadNode::Ideal(); 361 // 'sub' should have sub->in(0) != NULL. 362 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 363 sub->is_Region(), "expecting only these nodes"); 364 365 // Get control edge of 'sub'. 366 Node* orig_sub = sub; 367 sub = sub->find_exact_control(sub->in(0)); 368 if (sub == NULL || sub->is_top()) 369 return false; // Conservative answer for dead code 370 371 assert(sub->is_CFG(), "expecting control"); 372 373 if (sub == dom) 374 return true; 375 376 if (sub->is_Start() || sub->is_Root()) 377 return false; 378 379 { 380 // Check all control edges of 'dom'. 381 382 ResourceMark rm; 383 Arena* arena = Thread::current()->resource_area(); 384 Node_List nlist(arena); 385 Unique_Node_List dom_list(arena); 386 387 dom_list.push(dom); 388 bool only_dominating_controls = false; 389 390 for (uint next = 0; next < dom_list.size(); next++) { 391 Node* n = dom_list.at(next); 392 if (n == orig_sub) 393 return false; // One of dom's inputs dominated by sub. 394 if (!n->is_CFG() && n->pinned()) { 395 // Check only own control edge for pinned non-control nodes. 396 n = n->find_exact_control(n->in(0)); 397 if (n == NULL || n->is_top()) 398 return false; // Conservative answer for dead code 399 assert(n->is_CFG(), "expecting control"); 400 dom_list.push(n); 401 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 402 only_dominating_controls = true; 403 } else if (n->is_CFG()) { 404 if (n->dominates(sub, nlist)) 405 only_dominating_controls = true; 406 else 407 return false; 408 } else { 409 // First, own control edge. 410 Node* m = n->find_exact_control(n->in(0)); 411 if (m != NULL) { 412 if (m->is_top()) 413 return false; // Conservative answer for dead code 414 dom_list.push(m); 415 } 416 // Now, the rest of edges. 417 uint cnt = n->req(); 418 for (uint i = 1; i < cnt; i++) { 419 m = n->find_exact_control(n->in(i)); 420 if (m == NULL || m->is_top()) 421 continue; 422 dom_list.push(m); 423 } 424 } 425 } 426 return only_dominating_controls; 427 } 428} 429 430//---------------------detect_ptr_independence--------------------------------- 431// Used by MemNode::find_previous_store to prove that two base 432// pointers are never equal. 433// The pointers are accompanied by their associated allocations, 434// if any, which have been previously discovered by the caller. 435bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 436 Node* p2, AllocateNode* a2, 437 PhaseTransform* phase) { 438 // Attempt to prove that these two pointers cannot be aliased. 439 // They may both manifestly be allocations, and they should differ. 440 // Or, if they are not both allocations, they can be distinct constants. 441 // Otherwise, one is an allocation and the other a pre-existing value. 442 if (a1 == NULL && a2 == NULL) { // neither an allocation 443 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 444 } else if (a1 != NULL && a2 != NULL) { // both allocations 445 return (a1 != a2); 446 } else if (a1 != NULL) { // one allocation a1 447 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 448 return all_controls_dominate(p2, a1); 449 } else { //(a2 != NULL) // one allocation a2 450 return all_controls_dominate(p1, a2); 451 } 452 return false; 453} 454 455 456// The logic for reordering loads and stores uses four steps: 457// (a) Walk carefully past stores and initializations which we 458// can prove are independent of this load. 459// (b) Observe that the next memory state makes an exact match 460// with self (load or store), and locate the relevant store. 461// (c) Ensure that, if we were to wire self directly to the store, 462// the optimizer would fold it up somehow. 463// (d) Do the rewiring, and return, depending on some other part of 464// the optimizer to fold up the load. 465// This routine handles steps (a) and (b). Steps (c) and (d) are 466// specific to loads and stores, so they are handled by the callers. 467// (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 468// 469Node* MemNode::find_previous_store(PhaseTransform* phase) { 470 Node* ctrl = in(MemNode::Control); 471 Node* adr = in(MemNode::Address); 472 intptr_t offset = 0; 473 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 474 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 475 476 if (offset == Type::OffsetBot) 477 return NULL; // cannot unalias unless there are precise offsets 478 479 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 480 481 intptr_t size_in_bytes = memory_size(); 482 483 Node* mem = in(MemNode::Memory); // start searching here... 484 485 int cnt = 50; // Cycle limiter 486 for (;;) { // While we can dance past unrelated stores... 487 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 488 489 if (mem->is_Store()) { 490 Node* st_adr = mem->in(MemNode::Address); 491 intptr_t st_offset = 0; 492 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 493 if (st_base == NULL) 494 break; // inscrutable pointer 495 if (st_offset != offset && st_offset != Type::OffsetBot) { 496 const int MAX_STORE = BytesPerLong; 497 if (st_offset >= offset + size_in_bytes || 498 st_offset <= offset - MAX_STORE || 499 st_offset <= offset - mem->as_Store()->memory_size()) { 500 // Success: The offsets are provably independent. 501 // (You may ask, why not just test st_offset != offset and be done? 502 // The answer is that stores of different sizes can co-exist 503 // in the same sequence of RawMem effects. We sometimes initialize 504 // a whole 'tile' of array elements with a single jint or jlong.) 505 mem = mem->in(MemNode::Memory); 506 continue; // (a) advance through independent store memory 507 } 508 } 509 if (st_base != base && 510 detect_ptr_independence(base, alloc, 511 st_base, 512 AllocateNode::Ideal_allocation(st_base, phase), 513 phase)) { 514 // Success: The bases are provably independent. 515 mem = mem->in(MemNode::Memory); 516 continue; // (a) advance through independent store memory 517 } 518 519 // (b) At this point, if the bases or offsets do not agree, we lose, 520 // since we have not managed to prove 'this' and 'mem' independent. 521 if (st_base == base && st_offset == offset) { 522 return mem; // let caller handle steps (c), (d) 523 } 524 525 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 526 InitializeNode* st_init = mem->in(0)->as_Initialize(); 527 AllocateNode* st_alloc = st_init->allocation(); 528 if (st_alloc == NULL) 529 break; // something degenerated 530 bool known_identical = false; 531 bool known_independent = false; 532 if (alloc == st_alloc) 533 known_identical = true; 534 else if (alloc != NULL) 535 known_independent = true; 536 else if (all_controls_dominate(this, st_alloc)) 537 known_independent = true; 538 539 if (known_independent) { 540 // The bases are provably independent: Either they are 541 // manifestly distinct allocations, or else the control 542 // of this load dominates the store's allocation. 543 int alias_idx = phase->C->get_alias_index(adr_type()); 544 if (alias_idx == Compile::AliasIdxRaw) { 545 mem = st_alloc->in(TypeFunc::Memory); 546 } else { 547 mem = st_init->memory(alias_idx); 548 } 549 continue; // (a) advance through independent store memory 550 } 551 552 // (b) at this point, if we are not looking at a store initializing 553 // the same allocation we are loading from, we lose. 554 if (known_identical) { 555 // From caller, can_see_stored_value will consult find_captured_store. 556 return mem; // let caller handle steps (c), (d) 557 } 558 559 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 560 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 561 if (mem->is_Proj() && mem->in(0)->is_Call()) { 562 CallNode *call = mem->in(0)->as_Call(); 563 if (!call->may_modify(addr_t, phase)) { 564 mem = call->in(TypeFunc::Memory); 565 continue; // (a) advance through independent call memory 566 } 567 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 568 mem = mem->in(0)->in(TypeFunc::Memory); 569 continue; // (a) advance through independent MemBar memory 570 } else if (mem->is_ClearArray()) { 571 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 572 // (the call updated 'mem' value) 573 continue; // (a) advance through independent allocation memory 574 } else { 575 // Can not bypass initialization of the instance 576 // we are looking for. 577 return mem; 578 } 579 } else if (mem->is_MergeMem()) { 580 int alias_idx = phase->C->get_alias_index(adr_type()); 581 mem = mem->as_MergeMem()->memory_at(alias_idx); 582 continue; // (a) advance through independent MergeMem memory 583 } 584 } 585 586 // Unless there is an explicit 'continue', we must bail out here, 587 // because 'mem' is an inscrutable memory state (e.g., a call). 588 break; 589 } 590 591 return NULL; // bail out 592} 593 594//----------------------calculate_adr_type------------------------------------- 595// Helper function. Notices when the given type of address hits top or bottom. 596// Also, asserts a cross-check of the type against the expected address type. 597const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 598 if (t == Type::TOP) return NULL; // does not touch memory any more? 599 #ifdef PRODUCT 600 cross_check = NULL; 601 #else 602 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 603 #endif 604 const TypePtr* tp = t->isa_ptr(); 605 if (tp == NULL) { 606 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 607 return TypePtr::BOTTOM; // touches lots of memory 608 } else { 609 #ifdef ASSERT 610 // %%%% [phh] We don't check the alias index if cross_check is 611 // TypeRawPtr::BOTTOM. Needs to be investigated. 612 if (cross_check != NULL && 613 cross_check != TypePtr::BOTTOM && 614 cross_check != TypeRawPtr::BOTTOM) { 615 // Recheck the alias index, to see if it has changed (due to a bug). 616 Compile* C = Compile::current(); 617 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 618 "must stay in the original alias category"); 619 // The type of the address must be contained in the adr_type, 620 // disregarding "null"-ness. 621 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 622 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 623 assert(cross_check->meet(tp_notnull) == cross_check, 624 "real address must not escape from expected memory type"); 625 } 626 #endif 627 return tp; 628 } 629} 630 631//------------------------adr_phi_is_loop_invariant---------------------------- 632// A helper function for Ideal_DU_postCCP to check if a Phi in a counted 633// loop is loop invariant. Make a quick traversal of Phi and associated 634// CastPP nodes, looking to see if they are a closed group within the loop. 635bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 636 // The idea is that the phi-nest must boil down to only CastPP nodes 637 // with the same data. This implies that any path into the loop already 638 // includes such a CastPP, and so the original cast, whatever its input, 639 // must be covered by an equivalent cast, with an earlier control input. 640 ResourceMark rm; 641 642 // The loop entry input of the phi should be the unique dominating 643 // node for every Phi/CastPP in the loop. 644 Unique_Node_List closure; 645 closure.push(adr_phi->in(LoopNode::EntryControl)); 646 647 // Add the phi node and the cast to the worklist. 648 Unique_Node_List worklist; 649 worklist.push(adr_phi); 650 if( cast != NULL ){ 651 if( !cast->is_ConstraintCast() ) return false; 652 worklist.push(cast); 653 } 654 655 // Begin recursive walk of phi nodes. 656 while( worklist.size() ){ 657 // Take a node off the worklist 658 Node *n = worklist.pop(); 659 if( !closure.member(n) ){ 660 // Add it to the closure. 661 closure.push(n); 662 // Make a sanity check to ensure we don't waste too much time here. 663 if( closure.size() > 20) return false; 664 // This node is OK if: 665 // - it is a cast of an identical value 666 // - or it is a phi node (then we add its inputs to the worklist) 667 // Otherwise, the node is not OK, and we presume the cast is not invariant 668 if( n->is_ConstraintCast() ){ 669 worklist.push(n->in(1)); 670 } else if( n->is_Phi() ) { 671 for( uint i = 1; i < n->req(); i++ ) { 672 worklist.push(n->in(i)); 673 } 674 } else { 675 return false; 676 } 677 } 678 } 679 680 // Quit when the worklist is empty, and we've found no offending nodes. 681 return true; 682} 683 684//------------------------------Ideal_DU_postCCP------------------------------- 685// Find any cast-away of null-ness and keep its control. Null cast-aways are 686// going away in this pass and we need to make this memory op depend on the 687// gating null check. 688Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 689 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 690} 691 692// I tried to leave the CastPP's in. This makes the graph more accurate in 693// some sense; we get to keep around the knowledge that an oop is not-null 694// after some test. Alas, the CastPP's interfere with GVN (some values are 695// the regular oop, some are the CastPP of the oop, all merge at Phi's which 696// cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 697// some of the more trivial cases in the optimizer. Removing more useless 698// Phi's started allowing Loads to illegally float above null checks. I gave 699// up on this approach. CNC 10/20/2000 700// This static method may be called not from MemNode (EncodePNode calls it). 701// Only the control edge of the node 'n' might be updated. 702Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 703 Node *skipped_cast = NULL; 704 // Need a null check? Regular static accesses do not because they are 705 // from constant addresses. Array ops are gated by the range check (which 706 // always includes a NULL check). Just check field ops. 707 if( n->in(MemNode::Control) == NULL ) { 708 // Scan upwards for the highest location we can place this memory op. 709 while( true ) { 710 switch( adr->Opcode() ) { 711 712 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 713 adr = adr->in(AddPNode::Base); 714 continue; 715 716 case Op_DecodeN: // No change to NULL-ness, so peek thru 717 adr = adr->in(1); 718 continue; 719 720 case Op_EncodeP: 721 // EncodeP node's control edge could be set by this method 722 // when EncodeP node depends on CastPP node. 723 // 724 // Use its control edge for memory op because EncodeP may go away 725 // later when it is folded with following or preceding DecodeN node. 726 if (adr->in(0) == NULL) { 727 // Keep looking for cast nodes. 728 adr = adr->in(1); 729 continue; 730 } 731 ccp->hash_delete(n); 732 n->set_req(MemNode::Control, adr->in(0)); 733 ccp->hash_insert(n); 734 return n; 735 736 case Op_CastPP: 737 // If the CastPP is useless, just peek on through it. 738 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 739 // Remember the cast that we've peeked though. If we peek 740 // through more than one, then we end up remembering the highest 741 // one, that is, if in a loop, the one closest to the top. 742 skipped_cast = adr; 743 adr = adr->in(1); 744 continue; 745 } 746 // CastPP is going away in this pass! We need this memory op to be 747 // control-dependent on the test that is guarding the CastPP. 748 ccp->hash_delete(n); 749 n->set_req(MemNode::Control, adr->in(0)); 750 ccp->hash_insert(n); 751 return n; 752 753 case Op_Phi: 754 // Attempt to float above a Phi to some dominating point. 755 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 756 // If we've already peeked through a Cast (which could have set the 757 // control), we can't float above a Phi, because the skipped Cast 758 // may not be loop invariant. 759 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 760 adr = adr->in(1); 761 continue; 762 } 763 } 764 765 // Intentional fallthrough! 766 767 // No obvious dominating point. The mem op is pinned below the Phi 768 // by the Phi itself. If the Phi goes away (no true value is merged) 769 // then the mem op can float, but not indefinitely. It must be pinned 770 // behind the controls leading to the Phi. 771 case Op_CheckCastPP: 772 // These usually stick around to change address type, however a 773 // useless one can be elided and we still need to pick up a control edge 774 if (adr->in(0) == NULL) { 775 // This CheckCastPP node has NO control and is likely useless. But we 776 // need check further up the ancestor chain for a control input to keep 777 // the node in place. 4959717. 778 skipped_cast = adr; 779 adr = adr->in(1); 780 continue; 781 } 782 ccp->hash_delete(n); 783 n->set_req(MemNode::Control, adr->in(0)); 784 ccp->hash_insert(n); 785 return n; 786 787 // List of "safe" opcodes; those that implicitly block the memory 788 // op below any null check. 789 case Op_CastX2P: // no null checks on native pointers 790 case Op_Parm: // 'this' pointer is not null 791 case Op_LoadP: // Loading from within a klass 792 case Op_LoadN: // Loading from within a klass 793 case Op_LoadKlass: // Loading from within a klass 794 case Op_LoadNKlass: // Loading from within a klass 795 case Op_ConP: // Loading from a klass 796 case Op_ConN: // Loading from a klass 797 case Op_CreateEx: // Sucking up the guts of an exception oop 798 case Op_Con: // Reading from TLS 799 case Op_CMoveP: // CMoveP is pinned 800 case Op_CMoveN: // CMoveN is pinned 801 break; // No progress 802 803 case Op_Proj: // Direct call to an allocation routine 804 case Op_SCMemProj: // Memory state from store conditional ops 805#ifdef ASSERT 806 { 807 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 808 const Node* call = adr->in(0); 809 if (call->is_CallJava()) { 810 const CallJavaNode* call_java = call->as_CallJava(); 811 const TypeTuple *r = call_java->tf()->range(); 812 assert(r->cnt() > TypeFunc::Parms, "must return value"); 813 const Type* ret_type = r->field_at(TypeFunc::Parms); 814 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 815 // We further presume that this is one of 816 // new_instance_Java, new_array_Java, or 817 // the like, but do not assert for this. 818 } else if (call->is_Allocate()) { 819 // similar case to new_instance_Java, etc. 820 } else if (!call->is_CallLeaf()) { 821 // Projections from fetch_oop (OSR) are allowed as well. 822 ShouldNotReachHere(); 823 } 824 } 825#endif 826 break; 827 default: 828 ShouldNotReachHere(); 829 } 830 break; 831 } 832 } 833 834 return NULL; // No progress 835} 836 837 838//============================================================================= 839uint LoadNode::size_of() const { return sizeof(*this); } 840uint LoadNode::cmp( const Node &n ) const 841{ return !Type::cmp( _type, ((LoadNode&)n)._type ); } 842const Type *LoadNode::bottom_type() const { return _type; } 843uint LoadNode::ideal_reg() const { 844 return _type->ideal_reg(); 845} 846 847#ifndef PRODUCT 848void LoadNode::dump_spec(outputStream *st) const { 849 MemNode::dump_spec(st); 850 if( !Verbose && !WizardMode ) { 851 // standard dump does this in Verbose and WizardMode 852 st->print(" #"); _type->dump_on(st); 853 } 854} 855#endif 856 857#ifdef ASSERT 858//----------------------------is_immutable_value------------------------------- 859// Helper function to allow a raw load without control edge for some cases 860bool LoadNode::is_immutable_value(Node* adr) { 861 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 862 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 863 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 864 in_bytes(JavaThread::osthread_offset()))); 865} 866#endif 867 868//----------------------------LoadNode::make----------------------------------- 869// Polymorphic factory method: 870Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 871 Compile* C = gvn.C; 872 873 // sanity check the alias category against the created node type 874 assert(!(adr_type->isa_oopptr() && 875 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 876 "use LoadKlassNode instead"); 877 assert(!(adr_type->isa_aryptr() && 878 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 879 "use LoadRangeNode instead"); 880 // Check control edge of raw loads 881 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 882 // oop will be recorded in oop map if load crosses safepoint 883 rt->isa_oopptr() || is_immutable_value(adr), 884 "raw memory operations should have control edge"); 885 switch (bt) { 886 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); 887 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); 888 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); 889 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); 890 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); 891 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); 892 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); 893 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); 894 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); 895 case T_OBJECT: 896#ifdef _LP64 897 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 898 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); 899 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); 900 } else 901#endif 902 { 903 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 904 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 905 } 906 } 907 ShouldNotReachHere(); 908 return (LoadNode*)NULL; 909} 910 911LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 912 bool require_atomic = true; 913 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 914} 915 916 917 918 919//------------------------------hash------------------------------------------- 920uint LoadNode::hash() const { 921 // unroll addition of interesting fields 922 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 923} 924 925//---------------------------can_see_stored_value------------------------------ 926// This routine exists to make sure this set of tests is done the same 927// everywhere. We need to make a coordinated change: first LoadNode::Ideal 928// will change the graph shape in a way which makes memory alive twice at the 929// same time (uses the Oracle model of aliasing), then some 930// LoadXNode::Identity will fold things back to the equivalence-class model 931// of aliasing. 932Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 933 Node* ld_adr = in(MemNode::Address); 934 935 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 936 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 937 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 938 atp->field() != NULL && !atp->field()->is_volatile()) { 939 uint alias_idx = atp->index(); 940 bool final = atp->field()->is_final(); 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 || opc == Op_MemBarAcquireLock)) || 952 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder || 953 opc == Op_MemBarReleaseLock) { 954 Node* mem = current->in(0)->in(TypeFunc::Memory); 955 if (mem->is_MergeMem()) { 956 MergeMemNode* merge = mem->as_MergeMem(); 957 Node* new_st = merge->memory_at(alias_idx); 958 if (new_st == merge->base_memory()) { 959 // Keep searching 960 current = merge->base_memory(); 961 continue; 962 } 963 // Save the new memory state for the slice and fall through 964 // to exit. 965 result = new_st; 966 } 967 } 968 break; 969 } 970 if (result != NULL) { 971 st = result; 972 } 973 } 974 975 976 // Loop around twice in the case Load -> Initialize -> Store. 977 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 978 for (int trip = 0; trip <= 1; trip++) { 979 980 if (st->is_Store()) { 981 Node* st_adr = st->in(MemNode::Address); 982 if (!phase->eqv(st_adr, ld_adr)) { 983 // Try harder before giving up... Match raw and non-raw pointers. 984 intptr_t st_off = 0; 985 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 986 if (alloc == NULL) return NULL; 987 intptr_t ld_off = 0; 988 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 989 if (alloc != allo2) return NULL; 990 if (ld_off != st_off) return NULL; 991 // At this point we have proven something like this setup: 992 // A = Allocate(...) 993 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 994 // S = StoreQ(, AddP(, A.Parm , #Off), V) 995 // (Actually, we haven't yet proven the Q's are the same.) 996 // In other words, we are loading from a casted version of 997 // the same pointer-and-offset that we stored to. 998 // Thus, we are able to replace L by V. 999 } 1000 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1001 if (store_Opcode() != st->Opcode()) 1002 return NULL; 1003 return st->in(MemNode::ValueIn); 1004 } 1005 1006 intptr_t offset = 0; // scratch 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) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 1013 offset >= 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 && 1026 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 1027 // examine a captured store value 1028 st = init->find_captured_store(offset, memory_size(), phase); 1029 if (st != NULL) 1030 continue; // take one more trip around 1031 } 1032 } 1033 1034 break; 1035 } 1036 1037 return NULL; 1038} 1039 1040//----------------------is_instance_field_load_with_local_phi------------------ 1041bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1042 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 1043 in(MemNode::Address)->is_AddP() ) { 1044 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 1045 // Only instances. 1046 if( t_oop != NULL && t_oop->is_known_instance_field() && 1047 t_oop->offset() != Type::OffsetBot && 1048 t_oop->offset() != Type::OffsetTop) { 1049 return true; 1050 } 1051 } 1052 return false; 1053} 1054 1055//------------------------------Identity--------------------------------------- 1056// Loads are identity if previous store is to same address 1057Node *LoadNode::Identity( PhaseTransform *phase ) { 1058 // If the previous store-maker is the right kind of Store, and the store is 1059 // to the same address, then we are equal to the value stored. 1060 Node* mem = in(MemNode::Memory); 1061 Node* value = can_see_stored_value(mem, phase); 1062 if( value ) { 1063 // byte, short & char stores truncate naturally. 1064 // A load has to load the truncated value which requires 1065 // some sort of masking operation and that requires an 1066 // Ideal call instead of an Identity call. 1067 if (memory_size() < BytesPerInt) { 1068 // If the input to the store does not fit with the load's result type, 1069 // it must be truncated via an Ideal call. 1070 if (!phase->type(value)->higher_equal(phase->type(this))) 1071 return this; 1072 } 1073 // (This works even when value is a Con, but LoadNode::Value 1074 // usually runs first, producing the singleton type of the Con.) 1075 return value; 1076 } 1077 1078 // Search for an existing data phi which was generated before for the same 1079 // instance's field to avoid infinite generation of phis in a loop. 1080 Node *region = mem->in(0); 1081 if (is_instance_field_load_with_local_phi(region)) { 1082 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 1083 int this_index = phase->C->get_alias_index(addr_t); 1084 int this_offset = addr_t->offset(); 1085 int this_id = addr_t->is_oopptr()->instance_id(); 1086 const Type* this_type = bottom_type(); 1087 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1088 Node* phi = region->fast_out(i); 1089 if (phi->is_Phi() && phi != mem && 1090 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 1091 return phi; 1092 } 1093 } 1094 } 1095 1096 return this; 1097} 1098 1099 1100// Returns true if the AliasType refers to the field that holds the 1101// cached box array. Currently only handles the IntegerCache case. 1102static bool is_autobox_cache(Compile::AliasType* atp) { 1103 if (atp != NULL && atp->field() != NULL) { 1104 ciField* field = atp->field(); 1105 ciSymbol* klass = field->holder()->name(); 1106 if (field->name() == ciSymbol::cache_field_name() && 1107 field->holder()->uses_default_loader() && 1108 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1109 return true; 1110 } 1111 } 1112 return false; 1113} 1114 1115// Fetch the base value in the autobox array 1116static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 1117 if (atp != NULL && atp->field() != NULL) { 1118 ciField* field = atp->field(); 1119 ciSymbol* klass = field->holder()->name(); 1120 if (field->name() == ciSymbol::cache_field_name() && 1121 field->holder()->uses_default_loader() && 1122 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 1123 assert(field->is_constant(), "what?"); 1124 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 1125 // Fetch the box object at the base of the array and get its value 1126 ciInstance* box = array->obj_at(0)->as_instance(); 1127 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1128 if (ik->nof_nonstatic_fields() == 1) { 1129 // This should be true nonstatic_field_at requires calling 1130 // nof_nonstatic_fields so check it anyway 1131 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1132 cache_offset = c.as_int(); 1133 } 1134 return true; 1135 } 1136 } 1137 return false; 1138} 1139 1140// Returns true if the AliasType refers to the value field of an 1141// autobox object. Currently only handles Integer. 1142static bool is_autobox_object(Compile::AliasType* atp) { 1143 if (atp != NULL && atp->field() != NULL) { 1144 ciField* field = atp->field(); 1145 ciSymbol* klass = field->holder()->name(); 1146 if (field->name() == ciSymbol::value_name() && 1147 field->holder()->uses_default_loader() && 1148 klass == ciSymbol::java_lang_Integer()) { 1149 return true; 1150 } 1151 } 1152 return false; 1153} 1154 1155 1156// We're loading from an object which has autobox behaviour. 1157// If this object is result of a valueOf call we'll have a phi 1158// merging a newly allocated object and a load from the cache. 1159// We want to replace this load with the original incoming 1160// argument to the valueOf call. 1161Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1162 Node* base = in(Address)->in(AddPNode::Base); 1163 if (base->is_Phi() && base->req() == 3) { 1164 AllocateNode* allocation = NULL; 1165 int allocation_index = -1; 1166 int load_index = -1; 1167 for (uint i = 1; i < base->req(); i++) { 1168 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 1169 if (allocation != NULL) { 1170 allocation_index = i; 1171 load_index = 3 - allocation_index; 1172 break; 1173 } 1174 } 1175 bool has_load = ( allocation != NULL && 1176 (base->in(load_index)->is_Load() || 1177 base->in(load_index)->is_DecodeN() && 1178 base->in(load_index)->in(1)->is_Load()) ); 1179 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 1180 // Push the loads from the phi that comes from valueOf up 1181 // through it to allow elimination of the loads and the recovery 1182 // of the original value. 1183 Node* mem_phi = in(Memory); 1184 Node* offset = in(Address)->in(AddPNode::Offset); 1185 Node* region = base->in(0); 1186 1187 Node* in1 = clone(); 1188 Node* in1_addr = in1->in(Address)->clone(); 1189 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 1190 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 1191 in1_addr->set_req(AddPNode::Offset, offset); 1192 in1->set_req(0, region->in(allocation_index)); 1193 in1->set_req(Address, in1_addr); 1194 in1->set_req(Memory, mem_phi->in(allocation_index)); 1195 1196 Node* in2 = clone(); 1197 Node* in2_addr = in2->in(Address)->clone(); 1198 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1199 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1200 in2_addr->set_req(AddPNode::Offset, offset); 1201 in2->set_req(0, region->in(load_index)); 1202 in2->set_req(Address, in2_addr); 1203 in2->set_req(Memory, mem_phi->in(load_index)); 1204 1205 in1_addr = phase->transform(in1_addr); 1206 in1 = phase->transform(in1); 1207 in2_addr = phase->transform(in2_addr); 1208 in2 = phase->transform(in2); 1209 1210 PhiNode* result = PhiNode::make_blank(region, this); 1211 result->set_req(allocation_index, in1); 1212 result->set_req(load_index, in2); 1213 return result; 1214 } 1215 } else if (base->is_Load() || 1216 base->is_DecodeN() && base->in(1)->is_Load()) { 1217 if (base->is_DecodeN()) { 1218 // Get LoadN node which loads cached Integer object 1219 base = base->in(1); 1220 } 1221 // Eliminate the load of Integer.value for integers from the cache 1222 // array by deriving the value from the index into the array. 1223 // Capture the offset of the load and then reverse the computation. 1224 Node* load_base = base->in(Address)->in(AddPNode::Base); 1225 if (load_base->is_DecodeN()) { 1226 // Get LoadN node which loads IntegerCache.cache field 1227 load_base = load_base->in(1); 1228 } 1229 if (load_base != NULL) { 1230 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1231 intptr_t cache_offset; 1232 int shift = -1; 1233 Node* cache = NULL; 1234 if (is_autobox_cache(atp)) { 1235 shift = exact_log2(type2aelembytes(T_OBJECT)); 1236 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1237 } 1238 if (cache != NULL && base->in(Address)->is_AddP()) { 1239 Node* elements[4]; 1240 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1241 int cache_low; 1242 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1243 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1244 // Add up all the offsets making of the address of the load 1245 Node* result = elements[0]; 1246 for (int i = 1; i < count; i++) { 1247 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1248 } 1249 // Remove the constant offset from the address and then 1250 // remove the scaling of the offset to recover the original index. 1251 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1252 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1253 // Peel the shift off directly but wrap it in a dummy node 1254 // since Ideal can't return existing nodes 1255 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1256 } else { 1257 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1258 } 1259#ifdef _LP64 1260 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1261#endif 1262 return result; 1263 } 1264 } 1265 } 1266 } 1267 return NULL; 1268} 1269 1270//------------------------------split_through_phi------------------------------ 1271// Split instance field load through Phi. 1272Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1273 Node* mem = in(MemNode::Memory); 1274 Node* address = in(MemNode::Address); 1275 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1276 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1277 1278 assert(mem->is_Phi() && (t_oop != NULL) && 1279 t_oop->is_known_instance_field(), "invalide conditions"); 1280 1281 Node *region = mem->in(0); 1282 if (region == NULL) { 1283 return NULL; // Wait stable graph 1284 } 1285 uint cnt = mem->req(); 1286 for (uint i = 1; i < cnt; i++) { 1287 Node* rc = region->in(i); 1288 if (rc == NULL || phase->type(rc) == Type::TOP) 1289 return NULL; // Wait stable graph 1290 Node *in = mem->in(i); 1291 if (in == NULL) { 1292 return NULL; // Wait stable graph 1293 } 1294 } 1295 // Check for loop invariant. 1296 if (cnt == 3) { 1297 for (uint i = 1; i < cnt; i++) { 1298 Node *in = mem->in(i); 1299 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1300 if (m == mem) { 1301 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. 1302 return this; 1303 } 1304 } 1305 } 1306 // Split through Phi (see original code in loopopts.cpp). 1307 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1308 1309 // Do nothing here if Identity will find a value 1310 // (to avoid infinite chain of value phis generation). 1311 if (!phase->eqv(this, this->Identity(phase))) 1312 return NULL; 1313 1314 // Skip the split if the region dominates some control edge of the address. 1315 if (!MemNode::all_controls_dominate(address, region)) 1316 return NULL; 1317 1318 const Type* this_type = this->bottom_type(); 1319 int this_index = phase->C->get_alias_index(addr_t); 1320 int this_offset = addr_t->offset(); 1321 int this_iid = addr_t->is_oopptr()->instance_id(); 1322 PhaseIterGVN *igvn = phase->is_IterGVN(); 1323 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1324 for (uint i = 1; i < region->req(); i++) { 1325 Node *x; 1326 Node* the_clone = NULL; 1327 if (region->in(i) == phase->C->top()) { 1328 x = phase->C->top(); // Dead path? Use a dead data op 1329 } else { 1330 x = this->clone(); // Else clone up the data op 1331 the_clone = x; // Remember for possible deletion. 1332 // Alter data node to use pre-phi inputs 1333 if (this->in(0) == region) { 1334 x->set_req(0, region->in(i)); 1335 } else { 1336 x->set_req(0, NULL); 1337 } 1338 for (uint j = 1; j < this->req(); j++) { 1339 Node *in = this->in(j); 1340 if (in->is_Phi() && in->in(0) == region) 1341 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 1342 } 1343 } 1344 // Check for a 'win' on some paths 1345 const Type *t = x->Value(igvn); 1346 1347 bool singleton = t->singleton(); 1348 1349 // See comments in PhaseIdealLoop::split_thru_phi(). 1350 if (singleton && t == Type::TOP) { 1351 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1352 } 1353 1354 if (singleton) { 1355 x = igvn->makecon(t); 1356 } else { 1357 // We now call Identity to try to simplify the cloned node. 1358 // Note that some Identity methods call phase->type(this). 1359 // Make sure that the type array is big enough for 1360 // our new node, even though we may throw the node away. 1361 // (This tweaking with igvn only works because x is a new node.) 1362 igvn->set_type(x, t); 1363 // If x is a TypeNode, capture any more-precise type permanently into Node 1364 // otherwise it will be not updated during igvn->transform since 1365 // igvn->type(x) is set to x->Value() already. 1366 x->raise_bottom_type(t); 1367 Node *y = x->Identity(igvn); 1368 if (y != x) { 1369 x = y; 1370 } else { 1371 y = igvn->hash_find(x); 1372 if (y) { 1373 x = y; 1374 } else { 1375 // Else x is a new node we are keeping 1376 // We do not need register_new_node_with_optimizer 1377 // because set_type has already been called. 1378 igvn->_worklist.push(x); 1379 } 1380 } 1381 } 1382 if (x != the_clone && the_clone != NULL) 1383 igvn->remove_dead_node(the_clone); 1384 phi->set_req(i, x); 1385 } 1386 // Record Phi 1387 igvn->register_new_node_with_optimizer(phi); 1388 return phi; 1389} 1390 1391//------------------------------Ideal------------------------------------------ 1392// If the load is from Field memory and the pointer is non-null, we can 1393// zero out the control input. 1394// If the offset is constant and the base is an object allocation, 1395// try to hook me up to the exact initializing store. 1396Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1397 Node* p = MemNode::Ideal_common(phase, can_reshape); 1398 if (p) return (p == NodeSentinel) ? NULL : p; 1399 1400 Node* ctrl = in(MemNode::Control); 1401 Node* address = in(MemNode::Address); 1402 1403 // Skip up past a SafePoint control. Cannot do this for Stores because 1404 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1405 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1406 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1407 ctrl = ctrl->in(0); 1408 set_req(MemNode::Control,ctrl); 1409 } 1410 1411 intptr_t ignore = 0; 1412 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1413 if (base != NULL 1414 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1415 // Check for useless control edge in some common special cases 1416 if (in(MemNode::Control) != NULL 1417 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1418 && all_controls_dominate(base, phase->C->start())) { 1419 // A method-invariant, non-null address (constant or 'this' argument). 1420 set_req(MemNode::Control, NULL); 1421 } 1422 1423 if (EliminateAutoBox && can_reshape) { 1424 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); 1425 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1426 if (is_autobox_object(atp)) { 1427 Node* result = eliminate_autobox(phase); 1428 if (result != NULL) return result; 1429 } 1430 } 1431 } 1432 1433 Node* mem = in(MemNode::Memory); 1434 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1435 1436 if (addr_t != NULL) { 1437 // try to optimize our memory input 1438 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1439 if (opt_mem != mem) { 1440 set_req(MemNode::Memory, opt_mem); 1441 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1442 return this; 1443 } 1444 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1445 if (can_reshape && opt_mem->is_Phi() && 1446 (t_oop != NULL) && t_oop->is_known_instance_field()) { 1447 PhaseIterGVN *igvn = phase->is_IterGVN(); 1448 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1449 // Delay this transformation until memory Phi is processed. 1450 phase->is_IterGVN()->_worklist.push(this); 1451 return NULL; 1452 } 1453 // Split instance field load through Phi. 1454 Node* result = split_through_phi(phase); 1455 if (result != NULL) return result; 1456 } 1457 } 1458 1459 // Check for prior store with a different base or offset; make Load 1460 // independent. Skip through any number of them. Bail out if the stores 1461 // are in an endless dead cycle and report no progress. This is a key 1462 // transform for Reflection. However, if after skipping through the Stores 1463 // we can't then fold up against a prior store do NOT do the transform as 1464 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1465 // array memory alive twice: once for the hoisted Load and again after the 1466 // bypassed Store. This situation only works if EVERYBODY who does 1467 // anti-dependence work knows how to bypass. I.e. we need all 1468 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1469 // the alias index stuff. So instead, peek through Stores and IFF we can 1470 // fold up, do so. 1471 Node* prev_mem = find_previous_store(phase); 1472 // Steps (a), (b): Walk past independent stores to find an exact match. 1473 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1474 // (c) See if we can fold up on the spot, but don't fold up here. 1475 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1476 // just return a prior value, which is done by Identity calls. 1477 if (can_see_stored_value(prev_mem, phase)) { 1478 // Make ready for step (d): 1479 set_req(MemNode::Memory, prev_mem); 1480 return this; 1481 } 1482 } 1483 1484 return NULL; // No further progress 1485} 1486 1487// Helper to recognize certain Klass fields which are invariant across 1488// some group of array types (e.g., int[] or all T[] where T < Object). 1489const Type* 1490LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1491 ciKlass* klass) const { 1492 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1493 // The field is Klass::_modifier_flags. Return its (constant) value. 1494 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1495 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1496 return TypeInt::make(klass->modifier_flags()); 1497 } 1498 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1499 // The field is Klass::_access_flags. Return its (constant) value. 1500 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1501 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1502 return TypeInt::make(klass->access_flags()); 1503 } 1504 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1505 // The field is Klass::_layout_helper. Return its constant value if known. 1506 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1507 return TypeInt::make(klass->layout_helper()); 1508 } 1509 1510 // No match. 1511 return NULL; 1512} 1513 1514//------------------------------Value----------------------------------------- 1515const Type *LoadNode::Value( PhaseTransform *phase ) const { 1516 // Either input is TOP ==> the result is TOP 1517 Node* mem = in(MemNode::Memory); 1518 const Type *t1 = phase->type(mem); 1519 if (t1 == Type::TOP) return Type::TOP; 1520 Node* adr = in(MemNode::Address); 1521 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1522 if (tp == NULL || tp->empty()) return Type::TOP; 1523 int off = tp->offset(); 1524 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1525 Compile* C = phase->C; 1526 1527 // Try to guess loaded type from pointer type 1528 if (tp->base() == Type::AryPtr) { 1529 const Type *t = tp->is_aryptr()->elem(); 1530 // Don't do this for integer types. There is only potential profit if 1531 // the element type t is lower than _type; that is, for int types, if _type is 1532 // more restrictive than t. This only happens here if one is short and the other 1533 // char (both 16 bits), and in those cases we've made an intentional decision 1534 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1535 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1536 // 1537 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1538 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1539 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1540 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1541 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1542 // In fact, that could have been the original type of p1, and p1 could have 1543 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1544 // expression (LShiftL quux 3) independently optimized to the constant 8. 1545 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1546 && (_type->isa_vect() == NULL) 1547 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1548 // t might actually be lower than _type, if _type is a unique 1549 // concrete subclass of abstract class t. 1550 // Make sure the reference is not into the header, by comparing 1551 // the offset against the offset of the start of the array's data. 1552 // Different array types begin at slightly different offsets (12 vs. 16). 1553 // We choose T_BYTE as an example base type that is least restrictive 1554 // as to alignment, which will therefore produce the smallest 1555 // possible base offset. 1556 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1557 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1558 const Type* jt = t->join(_type); 1559 // In any case, do not allow the join, per se, to empty out the type. 1560 if (jt->empty() && !t->empty()) { 1561 // This can happen if a interface-typed array narrows to a class type. 1562 jt = _type; 1563 } 1564 1565 if (EliminateAutoBox && adr->is_AddP()) { 1566 // The pointers in the autobox arrays are always non-null 1567 Node* base = adr->in(AddPNode::Base); 1568 if (base != NULL && 1569 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { 1570 Compile::AliasType* atp = C->alias_type(base->adr_type()); 1571 if (is_autobox_cache(atp)) { 1572 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1573 } 1574 } 1575 } 1576 return jt; 1577 } 1578 } 1579 } else if (tp->base() == Type::InstPtr) { 1580 ciEnv* env = C->env(); 1581 const TypeInstPtr* tinst = tp->is_instptr(); 1582 ciKlass* klass = tinst->klass(); 1583 assert( off != Type::OffsetBot || 1584 // arrays can be cast to Objects 1585 tp->is_oopptr()->klass()->is_java_lang_Object() || 1586 // unsafe field access may not have a constant offset 1587 C->has_unsafe_access(), 1588 "Field accesses must be precise" ); 1589 // For oop loads, we expect the _type to be precise 1590 if (klass == env->String_klass() && 1591 adr->is_AddP() && off != Type::OffsetBot) { 1592 // For constant Strings treat the final fields as compile time constants. 1593 Node* base = adr->in(AddPNode::Base); 1594 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1595 if (t != NULL && t->singleton()) { 1596 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1597 if (field != NULL && field->is_final()) { 1598 ciObject* string = t->const_oop(); 1599 ciConstant constant = string->as_instance()->field_value(field); 1600 if (constant.basic_type() == T_INT) { 1601 return TypeInt::make(constant.as_int()); 1602 } else if (constant.basic_type() == T_ARRAY) { 1603 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1604 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1605 } else { 1606 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1607 } 1608 } 1609 } 1610 } 1611 } 1612 // Optimizations for constant objects 1613 ciObject* const_oop = tinst->const_oop(); 1614 if (const_oop != NULL) { 1615 // For constant CallSites treat the target field as a compile time constant. 1616 if (const_oop->is_call_site()) { 1617 ciCallSite* call_site = const_oop->as_call_site(); 1618 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1619 if (field != NULL && field->is_call_site_target()) { 1620 ciMethodHandle* target = call_site->get_target(); 1621 if (target != NULL) { // just in case 1622 ciConstant constant(T_OBJECT, target); 1623 const Type* t; 1624 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1625 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1626 } else { 1627 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1628 } 1629 // Add a dependence for invalidation of the optimization. 1630 if (!call_site->is_constant_call_site()) { 1631 C->dependencies()->assert_call_site_target_value(call_site, target); 1632 } 1633 return t; 1634 } 1635 } 1636 } 1637 } 1638 } else if (tp->base() == Type::KlassPtr) { 1639 assert( off != Type::OffsetBot || 1640 // arrays can be cast to Objects 1641 tp->is_klassptr()->klass()->is_java_lang_Object() || 1642 // also allow array-loading from the primary supertype 1643 // array during subtype checks 1644 Opcode() == Op_LoadKlass, 1645 "Field accesses must be precise" ); 1646 // For klass/static loads, we expect the _type to be precise 1647 } 1648 1649 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1650 if (tkls != NULL && !StressReflectiveCode) { 1651 ciKlass* klass = tkls->klass(); 1652 if (klass->is_loaded() && tkls->klass_is_exact()) { 1653 // We are loading a field from a Klass metaobject whose identity 1654 // is known at compile time (the type is "exact" or "precise"). 1655 // Check for fields we know are maintained as constants by the VM. 1656 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1657 // The field is Klass::_super_check_offset. Return its (constant) value. 1658 // (Folds up type checking code.) 1659 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1660 return TypeInt::make(klass->super_check_offset()); 1661 } 1662 // Compute index into primary_supers array 1663 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1664 // Check for overflowing; use unsigned compare to handle the negative case. 1665 if( depth < ciKlass::primary_super_limit() ) { 1666 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1667 // (Folds up type checking code.) 1668 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1669 ciKlass *ss = klass->super_of_depth(depth); 1670 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1671 } 1672 const Type* aift = load_array_final_field(tkls, klass); 1673 if (aift != NULL) return aift; 1674 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) 1675 && klass->is_array_klass()) { 1676 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1677 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1678 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1679 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1680 } 1681 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1682 // The field is Klass::_java_mirror. Return its (constant) value. 1683 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1684 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1685 return TypeInstPtr::make(klass->java_mirror()); 1686 } 1687 } 1688 1689 // We can still check if we are loading from the primary_supers array at a 1690 // shallow enough depth. Even though the klass is not exact, entries less 1691 // than or equal to its super depth are correct. 1692 if (klass->is_loaded() ) { 1693 ciType *inner = klass; 1694 while( inner->is_obj_array_klass() ) 1695 inner = inner->as_obj_array_klass()->base_element_type(); 1696 if( inner->is_instance_klass() && 1697 !inner->as_instance_klass()->flags().is_interface() ) { 1698 // Compute index into primary_supers array 1699 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1700 // Check for overflowing; use unsigned compare to handle the negative case. 1701 if( depth < ciKlass::primary_super_limit() && 1702 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1703 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1704 // (Folds up type checking code.) 1705 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1706 ciKlass *ss = klass->super_of_depth(depth); 1707 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1708 } 1709 } 1710 } 1711 1712 // If the type is enough to determine that the thing is not an array, 1713 // we can give the layout_helper a positive interval type. 1714 // This will help short-circuit some reflective code. 1715 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1716 && !klass->is_array_klass() // not directly typed as an array 1717 && !klass->is_interface() // specifically not Serializable & Cloneable 1718 && !klass->is_java_lang_Object() // not the supertype of all T[] 1719 ) { 1720 // Note: When interfaces are reliable, we can narrow the interface 1721 // test to (klass != Serializable && klass != Cloneable). 1722 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1723 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1724 // The key property of this type is that it folds up tests 1725 // for array-ness, since it proves that the layout_helper is positive. 1726 // Thus, a generic value like the basic object layout helper works fine. 1727 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1728 } 1729 } 1730 1731 // If we are loading from a freshly-allocated object, produce a zero, 1732 // if the load is provably beyond the header of the object. 1733 // (Also allow a variable load from a fresh array to produce zero.) 1734 const TypeOopPtr *tinst = tp->isa_oopptr(); 1735 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1736 if (ReduceFieldZeroing || is_instance) { 1737 Node* value = can_see_stored_value(mem,phase); 1738 if (value != NULL && value->is_Con()) { 1739 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1740 return value->bottom_type(); 1741 } 1742 } 1743 1744 if (is_instance) { 1745 // If we have an instance type and our memory input is the 1746 // programs's initial memory state, there is no matching store, 1747 // so just return a zero of the appropriate type 1748 Node *mem = in(MemNode::Memory); 1749 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1750 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1751 return Type::get_zero_type(_type->basic_type()); 1752 } 1753 } 1754 return _type; 1755} 1756 1757//------------------------------match_edge------------------------------------- 1758// Do we Match on this edge index or not? Match only the address. 1759uint LoadNode::match_edge(uint idx) const { 1760 return idx == MemNode::Address; 1761} 1762 1763//--------------------------LoadBNode::Ideal-------------------------------------- 1764// 1765// If the previous store is to the same address as this load, 1766// and the value stored was larger than a byte, replace this load 1767// with the value stored truncated to a byte. If no truncation is 1768// needed, the replacement is done in LoadNode::Identity(). 1769// 1770Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1771 Node* mem = in(MemNode::Memory); 1772 Node* value = can_see_stored_value(mem,phase); 1773 if( value && !phase->type(value)->higher_equal( _type ) ) { 1774 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1775 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1776 } 1777 // Identity call will handle the case where truncation is not needed. 1778 return LoadNode::Ideal(phase, can_reshape); 1779} 1780 1781const Type* LoadBNode::Value(PhaseTransform *phase) const { 1782 Node* mem = in(MemNode::Memory); 1783 Node* value = can_see_stored_value(mem,phase); 1784 if (value != NULL && value->is_Con() && 1785 !value->bottom_type()->higher_equal(_type)) { 1786 // If the input to the store does not fit with the load's result type, 1787 // it must be truncated. We can't delay until Ideal call since 1788 // a singleton Value is needed for split_thru_phi optimization. 1789 int con = value->get_int(); 1790 return TypeInt::make((con << 24) >> 24); 1791 } 1792 return LoadNode::Value(phase); 1793} 1794 1795//--------------------------LoadUBNode::Ideal------------------------------------- 1796// 1797// If the previous store is to the same address as this load, 1798// and the value stored was larger than a byte, replace this load 1799// with the value stored truncated to a byte. If no truncation is 1800// needed, the replacement is done in LoadNode::Identity(). 1801// 1802Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1803 Node* mem = in(MemNode::Memory); 1804 Node* value = can_see_stored_value(mem, phase); 1805 if (value && !phase->type(value)->higher_equal(_type)) 1806 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); 1807 // Identity call will handle the case where truncation is not needed. 1808 return LoadNode::Ideal(phase, can_reshape); 1809} 1810 1811const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1812 Node* mem = in(MemNode::Memory); 1813 Node* value = can_see_stored_value(mem,phase); 1814 if (value != NULL && value->is_Con() && 1815 !value->bottom_type()->higher_equal(_type)) { 1816 // If the input to the store does not fit with the load's result type, 1817 // it must be truncated. We can't delay until Ideal call since 1818 // a singleton Value is needed for split_thru_phi optimization. 1819 int con = value->get_int(); 1820 return TypeInt::make(con & 0xFF); 1821 } 1822 return LoadNode::Value(phase); 1823} 1824 1825//--------------------------LoadUSNode::Ideal------------------------------------- 1826// 1827// If the previous store is to the same address as this load, 1828// and the value stored was larger than a char, replace this load 1829// with the value stored truncated to a char. If no truncation is 1830// needed, the replacement is done in LoadNode::Identity(). 1831// 1832Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1833 Node* mem = in(MemNode::Memory); 1834 Node* value = can_see_stored_value(mem,phase); 1835 if( value && !phase->type(value)->higher_equal( _type ) ) 1836 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1837 // Identity call will handle the case where truncation is not needed. 1838 return LoadNode::Ideal(phase, can_reshape); 1839} 1840 1841const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1842 Node* mem = in(MemNode::Memory); 1843 Node* value = can_see_stored_value(mem,phase); 1844 if (value != NULL && value->is_Con() && 1845 !value->bottom_type()->higher_equal(_type)) { 1846 // If the input to the store does not fit with the load's result type, 1847 // it must be truncated. We can't delay until Ideal call since 1848 // a singleton Value is needed for split_thru_phi optimization. 1849 int con = value->get_int(); 1850 return TypeInt::make(con & 0xFFFF); 1851 } 1852 return LoadNode::Value(phase); 1853} 1854 1855//--------------------------LoadSNode::Ideal-------------------------------------- 1856// 1857// If the previous store is to the same address as this load, 1858// and the value stored was larger than a short, replace this load 1859// with the value stored truncated to a short. If no truncation is 1860// needed, the replacement is done in LoadNode::Identity(). 1861// 1862Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1863 Node* mem = in(MemNode::Memory); 1864 Node* value = can_see_stored_value(mem,phase); 1865 if( value && !phase->type(value)->higher_equal( _type ) ) { 1866 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1867 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1868 } 1869 // Identity call will handle the case where truncation is not needed. 1870 return LoadNode::Ideal(phase, can_reshape); 1871} 1872 1873const Type* LoadSNode::Value(PhaseTransform *phase) const { 1874 Node* mem = in(MemNode::Memory); 1875 Node* value = can_see_stored_value(mem,phase); 1876 if (value != NULL && value->is_Con() && 1877 !value->bottom_type()->higher_equal(_type)) { 1878 // If the input to the store does not fit with the load's result type, 1879 // it must be truncated. We can't delay until Ideal call since 1880 // a singleton Value is needed for split_thru_phi optimization. 1881 int con = value->get_int(); 1882 return TypeInt::make((con << 16) >> 16); 1883 } 1884 return LoadNode::Value(phase); 1885} 1886 1887//============================================================================= 1888//----------------------------LoadKlassNode::make------------------------------ 1889// Polymorphic factory method: 1890Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 1891 Compile* C = gvn.C; 1892 Node *ctl = NULL; 1893 // sanity check the alias category against the created node type 1894 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 1895 assert(adr_type != NULL, "expecting TypeKlassPtr"); 1896#ifdef _LP64 1897 if (adr_type->is_ptr_to_narrowoop()) { 1898 assert(UseCompressedKlassPointers, "no compressed klasses"); 1899 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); 1900 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); 1901 } 1902#endif 1903 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 1904 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); 1905} 1906 1907//------------------------------Value------------------------------------------ 1908const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1909 return klass_value_common(phase); 1910} 1911 1912const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 1913 // Either input is TOP ==> the result is TOP 1914 const Type *t1 = phase->type( in(MemNode::Memory) ); 1915 if (t1 == Type::TOP) return Type::TOP; 1916 Node *adr = in(MemNode::Address); 1917 const Type *t2 = phase->type( adr ); 1918 if (t2 == Type::TOP) return Type::TOP; 1919 const TypePtr *tp = t2->is_ptr(); 1920 if (TypePtr::above_centerline(tp->ptr()) || 1921 tp->ptr() == TypePtr::Null) return Type::TOP; 1922 1923 // Return a more precise klass, if possible 1924 const TypeInstPtr *tinst = tp->isa_instptr(); 1925 if (tinst != NULL) { 1926 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1927 int offset = tinst->offset(); 1928 if (ik == phase->C->env()->Class_klass() 1929 && (offset == java_lang_Class::klass_offset_in_bytes() || 1930 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1931 // We are loading a special hidden field from a Class mirror object, 1932 // the field which points to the VM's Klass metaobject. 1933 ciType* t = tinst->java_mirror_type(); 1934 // java_mirror_type returns non-null for compile-time Class constants. 1935 if (t != NULL) { 1936 // constant oop => constant klass 1937 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1938 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1939 } 1940 if (!t->is_klass()) { 1941 // a primitive Class (e.g., int.class) has NULL for a klass field 1942 return TypePtr::NULL_PTR; 1943 } 1944 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1945 return TypeKlassPtr::make(t->as_klass()); 1946 } 1947 // non-constant mirror, so we can't tell what's going on 1948 } 1949 if( !ik->is_loaded() ) 1950 return _type; // Bail out if not loaded 1951 if (offset == oopDesc::klass_offset_in_bytes()) { 1952 if (tinst->klass_is_exact()) { 1953 return TypeKlassPtr::make(ik); 1954 } 1955 // See if we can become precise: no subklasses and no interface 1956 // (Note: We need to support verified interfaces.) 1957 if (!ik->is_interface() && !ik->has_subklass()) { 1958 //assert(!UseExactTypes, "this code should be useless with exact types"); 1959 // Add a dependence; if any subclass added we need to recompile 1960 if (!ik->is_final()) { 1961 // %%% should use stronger assert_unique_concrete_subtype instead 1962 phase->C->dependencies()->assert_leaf_type(ik); 1963 } 1964 // Return precise klass 1965 return TypeKlassPtr::make(ik); 1966 } 1967 1968 // Return root of possible klass 1969 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1970 } 1971 } 1972 1973 // Check for loading klass from an array 1974 const TypeAryPtr *tary = tp->isa_aryptr(); 1975 if( tary != NULL ) { 1976 ciKlass *tary_klass = tary->klass(); 1977 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1978 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1979 if (tary->klass_is_exact()) { 1980 return TypeKlassPtr::make(tary_klass); 1981 } 1982 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1983 // If the klass is an object array, we defer the question to the 1984 // array component klass. 1985 if( ak->is_obj_array_klass() ) { 1986 assert( ak->is_loaded(), "" ); 1987 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1988 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1989 ciInstanceKlass* ik = base_k->as_instance_klass(); 1990 // See if we can become precise: no subklasses and no interface 1991 if (!ik->is_interface() && !ik->has_subklass()) { 1992 //assert(!UseExactTypes, "this code should be useless with exact types"); 1993 // Add a dependence; if any subclass added we need to recompile 1994 if (!ik->is_final()) { 1995 phase->C->dependencies()->assert_leaf_type(ik); 1996 } 1997 // Return precise array klass 1998 return TypeKlassPtr::make(ak); 1999 } 2000 } 2001 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2002 } else { // Found a type-array? 2003 //assert(!UseExactTypes, "this code should be useless with exact types"); 2004 assert( ak->is_type_array_klass(), "" ); 2005 return TypeKlassPtr::make(ak); // These are always precise 2006 } 2007 } 2008 } 2009 2010 // Check for loading klass from an array klass 2011 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2012 if (tkls != NULL && !StressReflectiveCode) { 2013 ciKlass* klass = tkls->klass(); 2014 if( !klass->is_loaded() ) 2015 return _type; // Bail out if not loaded 2016 if( klass->is_obj_array_klass() && 2017 tkls->offset() == in_bytes(objArrayKlass::element_klass_offset())) { 2018 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2019 // // Always returning precise element type is incorrect, 2020 // // e.g., element type could be object and array may contain strings 2021 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2022 2023 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2024 // according to the element type's subclassing. 2025 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2026 } 2027 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2028 tkls->offset() == in_bytes(Klass::super_offset())) { 2029 ciKlass* sup = klass->as_instance_klass()->super(); 2030 // The field is Klass::_super. Return its (constant) value. 2031 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2032 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2033 } 2034 } 2035 2036 // Bailout case 2037 return LoadNode::Value(phase); 2038} 2039 2040//------------------------------Identity--------------------------------------- 2041// To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2042// Also feed through the klass in Allocate(...klass...)._klass. 2043Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2044 return klass_identity_common(phase); 2045} 2046 2047Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2048 Node* x = LoadNode::Identity(phase); 2049 if (x != this) return x; 2050 2051 // Take apart the address into an oop and and offset. 2052 // Return 'this' if we cannot. 2053 Node* adr = in(MemNode::Address); 2054 intptr_t offset = 0; 2055 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2056 if (base == NULL) return this; 2057 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2058 if (toop == NULL) return this; 2059 2060 // We can fetch the klass directly through an AllocateNode. 2061 // This works even if the klass is not constant (clone or newArray). 2062 if (offset == oopDesc::klass_offset_in_bytes()) { 2063 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2064 if (allocated_klass != NULL) { 2065 return allocated_klass; 2066 } 2067 } 2068 2069 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2070 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 2071 // See inline_native_Class_query for occurrences of these patterns. 2072 // Java Example: x.getClass().isAssignableFrom(y) 2073 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 2074 // 2075 // This improves reflective code, often making the Class 2076 // mirror go completely dead. (Current exception: Class 2077 // mirrors may appear in debug info, but we could clean them out by 2078 // introducing a new debug info operator for Klass*.java_mirror). 2079 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2080 && (offset == java_lang_Class::klass_offset_in_bytes() || 2081 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2082 // We are loading a special hidden field from a Class mirror, 2083 // the field which points to its Klass or arrayKlass metaobject. 2084 if (base->is_Load()) { 2085 Node* adr2 = base->in(MemNode::Address); 2086 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2087 if (tkls != NULL && !tkls->empty() 2088 && (tkls->klass()->is_instance_klass() || 2089 tkls->klass()->is_array_klass()) 2090 && adr2->is_AddP() 2091 ) { 2092 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2093 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2094 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 2095 } 2096 if (tkls->offset() == mirror_field) { 2097 return adr2->in(AddPNode::Base); 2098 } 2099 } 2100 } 2101 } 2102 2103 return this; 2104} 2105 2106 2107//------------------------------Value------------------------------------------ 2108const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2109 const Type *t = klass_value_common(phase); 2110 if (t == Type::TOP) 2111 return t; 2112 2113 return t->make_narrowoop(); 2114} 2115 2116//------------------------------Identity--------------------------------------- 2117// To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2118// Also feed through the klass in Allocate(...klass...)._klass. 2119Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2120 Node *x = klass_identity_common(phase); 2121 2122 const Type *t = phase->type( x ); 2123 if( t == Type::TOP ) return x; 2124 if( t->isa_narrowoop()) return x; 2125 2126 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); 2127} 2128 2129//------------------------------Value----------------------------------------- 2130const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2131 // Either input is TOP ==> the result is TOP 2132 const Type *t1 = phase->type( in(MemNode::Memory) ); 2133 if( t1 == Type::TOP ) return Type::TOP; 2134 Node *adr = in(MemNode::Address); 2135 const Type *t2 = phase->type( adr ); 2136 if( t2 == Type::TOP ) return Type::TOP; 2137 const TypePtr *tp = t2->is_ptr(); 2138 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2139 const TypeAryPtr *tap = tp->isa_aryptr(); 2140 if( !tap ) return _type; 2141 return tap->size(); 2142} 2143 2144//-------------------------------Ideal--------------------------------------- 2145// Feed through the length in AllocateArray(...length...)._length. 2146Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2147 Node* p = MemNode::Ideal_common(phase, can_reshape); 2148 if (p) return (p == NodeSentinel) ? NULL : p; 2149 2150 // Take apart the address into an oop and and offset. 2151 // Return 'this' if we cannot. 2152 Node* adr = in(MemNode::Address); 2153 intptr_t offset = 0; 2154 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2155 if (base == NULL) return NULL; 2156 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2157 if (tary == NULL) return NULL; 2158 2159 // We can fetch the length directly through an AllocateArrayNode. 2160 // This works even if the length is not constant (clone or newArray). 2161 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2162 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2163 if (alloc != NULL) { 2164 Node* allocated_length = alloc->Ideal_length(); 2165 Node* len = alloc->make_ideal_length(tary, phase); 2166 if (allocated_length != len) { 2167 // New CastII improves on this. 2168 return len; 2169 } 2170 } 2171 } 2172 2173 return NULL; 2174} 2175 2176//------------------------------Identity--------------------------------------- 2177// Feed through the length in AllocateArray(...length...)._length. 2178Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2179 Node* x = LoadINode::Identity(phase); 2180 if (x != this) return x; 2181 2182 // Take apart the address into an oop and and offset. 2183 // Return 'this' if we cannot. 2184 Node* adr = in(MemNode::Address); 2185 intptr_t offset = 0; 2186 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2187 if (base == NULL) return this; 2188 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2189 if (tary == NULL) return this; 2190 2191 // We can fetch the length directly through an AllocateArrayNode. 2192 // This works even if the length is not constant (clone or newArray). 2193 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2194 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2195 if (alloc != NULL) { 2196 Node* allocated_length = alloc->Ideal_length(); 2197 // Do not allow make_ideal_length to allocate a CastII node. 2198 Node* len = alloc->make_ideal_length(tary, phase, false); 2199 if (allocated_length == len) { 2200 // Return allocated_length only if it would not be improved by a CastII. 2201 return allocated_length; 2202 } 2203 } 2204 } 2205 2206 return this; 2207 2208} 2209 2210//============================================================================= 2211//---------------------------StoreNode::make----------------------------------- 2212// Polymorphic factory method: 2213StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 2214 Compile* C = gvn.C; 2215 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2216 ctl != NULL, "raw memory operations should have control edge"); 2217 2218 switch (bt) { 2219 case T_BOOLEAN: 2220 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 2221 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 2222 case T_CHAR: 2223 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 2224 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 2225 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 2226 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 2227 case T_METADATA: 2228 case T_ADDRESS: 2229 case T_OBJECT: 2230#ifdef _LP64 2231 if (adr->bottom_type()->is_ptr_to_narrowoop() || 2232 (UseCompressedKlassPointers && val->bottom_type()->isa_klassptr() && 2233 adr->bottom_type()->isa_rawptr())) { 2234 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); 2235 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); 2236 } else 2237#endif 2238 { 2239 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 2240 } 2241 } 2242 ShouldNotReachHere(); 2243 return (StoreNode*)NULL; 2244} 2245 2246StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 2247 bool require_atomic = true; 2248 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 2249} 2250 2251 2252//--------------------------bottom_type---------------------------------------- 2253const Type *StoreNode::bottom_type() const { 2254 return Type::MEMORY; 2255} 2256 2257//------------------------------hash------------------------------------------- 2258uint StoreNode::hash() const { 2259 // unroll addition of interesting fields 2260 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2261 2262 // Since they are not commoned, do not hash them: 2263 return NO_HASH; 2264} 2265 2266//------------------------------Ideal------------------------------------------ 2267// Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2268// When a store immediately follows a relevant allocation/initialization, 2269// try to capture it into the initialization, or hoist it above. 2270Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2271 Node* p = MemNode::Ideal_common(phase, can_reshape); 2272 if (p) return (p == NodeSentinel) ? NULL : p; 2273 2274 Node* mem = in(MemNode::Memory); 2275 Node* address = in(MemNode::Address); 2276 2277 // Back-to-back stores to same address? Fold em up. Generally 2278 // unsafe if I have intervening uses... Also disallowed for StoreCM 2279 // since they must follow each StoreP operation. Redundant StoreCMs 2280 // are eliminated just before matching in final_graph_reshape. 2281 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && 2282 mem->Opcode() != Op_StoreCM) { 2283 // Looking at a dead closed cycle of memory? 2284 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2285 2286 assert(Opcode() == mem->Opcode() || 2287 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2288 "no mismatched stores, except on raw memory"); 2289 2290 if (mem->outcnt() == 1 && // check for intervening uses 2291 mem->as_Store()->memory_size() <= this->memory_size()) { 2292 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2293 // For example, 'mem' might be the final state at a conditional return. 2294 // Or, 'mem' might be used by some node which is live at the same time 2295 // 'this' is live, which might be unschedulable. So, require exactly 2296 // ONE user, the 'this' store, until such time as we clone 'mem' for 2297 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2298 if (can_reshape) { // (%%% is this an anachronism?) 2299 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2300 phase->is_IterGVN()); 2301 } else { 2302 // It's OK to do this in the parser, since DU info is always accurate, 2303 // and the parser always refers to nodes via SafePointNode maps. 2304 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2305 } 2306 return this; 2307 } 2308 } 2309 2310 // Capture an unaliased, unconditional, simple store into an initializer. 2311 // Or, if it is independent of the allocation, hoist it above the allocation. 2312 if (ReduceFieldZeroing && /*can_reshape &&*/ 2313 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2314 InitializeNode* init = mem->in(0)->as_Initialize(); 2315 intptr_t offset = init->can_capture_store(this, phase); 2316 if (offset > 0) { 2317 Node* moved = init->capture_store(this, offset, phase); 2318 // If the InitializeNode captured me, it made a raw copy of me, 2319 // and I need to disappear. 2320 if (moved != NULL) { 2321 // %%% hack to ensure that Ideal returns a new node: 2322 mem = MergeMemNode::make(phase->C, mem); 2323 return mem; // fold me away 2324 } 2325 } 2326 } 2327 2328 return NULL; // No further progress 2329} 2330 2331//------------------------------Value----------------------------------------- 2332const Type *StoreNode::Value( PhaseTransform *phase ) const { 2333 // Either input is TOP ==> the result is TOP 2334 const Type *t1 = phase->type( in(MemNode::Memory) ); 2335 if( t1 == Type::TOP ) return Type::TOP; 2336 const Type *t2 = phase->type( in(MemNode::Address) ); 2337 if( t2 == Type::TOP ) return Type::TOP; 2338 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2339 if( t3 == Type::TOP ) return Type::TOP; 2340 return Type::MEMORY; 2341} 2342 2343//------------------------------Identity--------------------------------------- 2344// Remove redundant stores: 2345// Store(m, p, Load(m, p)) changes to m. 2346// Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2347Node *StoreNode::Identity( PhaseTransform *phase ) { 2348 Node* mem = in(MemNode::Memory); 2349 Node* adr = in(MemNode::Address); 2350 Node* val = in(MemNode::ValueIn); 2351 2352 // Load then Store? Then the Store is useless 2353 if (val->is_Load() && 2354 val->in(MemNode::Address)->eqv_uncast(adr) && 2355 val->in(MemNode::Memory )->eqv_uncast(mem) && 2356 val->as_Load()->store_Opcode() == Opcode()) { 2357 return mem; 2358 } 2359 2360 // Two stores in a row of the same value? 2361 if (mem->is_Store() && 2362 mem->in(MemNode::Address)->eqv_uncast(adr) && 2363 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2364 mem->Opcode() == Opcode()) { 2365 return mem; 2366 } 2367 2368 // Store of zero anywhere into a freshly-allocated object? 2369 // Then the store is useless. 2370 // (It must already have been captured by the InitializeNode.) 2371 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2372 // a newly allocated object is already all-zeroes everywhere 2373 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2374 return mem; 2375 } 2376 2377 // the store may also apply to zero-bits in an earlier object 2378 Node* prev_mem = find_previous_store(phase); 2379 // Steps (a), (b): Walk past independent stores to find an exact match. 2380 if (prev_mem != NULL) { 2381 Node* prev_val = can_see_stored_value(prev_mem, phase); 2382 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2383 // prev_val and val might differ by a cast; it would be good 2384 // to keep the more informative of the two. 2385 return mem; 2386 } 2387 } 2388 } 2389 2390 return this; 2391} 2392 2393//------------------------------match_edge------------------------------------- 2394// Do we Match on this edge index or not? Match only memory & value 2395uint StoreNode::match_edge(uint idx) const { 2396 return idx == MemNode::Address || idx == MemNode::ValueIn; 2397} 2398 2399//------------------------------cmp-------------------------------------------- 2400// Do not common stores up together. They generally have to be split 2401// back up anyways, so do not bother. 2402uint StoreNode::cmp( const Node &n ) const { 2403 return (&n == this); // Always fail except on self 2404} 2405 2406//------------------------------Ideal_masked_input----------------------------- 2407// Check for a useless mask before a partial-word store 2408// (StoreB ... (AndI valIn conIa) ) 2409// If (conIa & mask == mask) this simplifies to 2410// (StoreB ... (valIn) ) 2411Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2412 Node *val = in(MemNode::ValueIn); 2413 if( val->Opcode() == Op_AndI ) { 2414 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2415 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2416 set_req(MemNode::ValueIn, val->in(1)); 2417 return this; 2418 } 2419 } 2420 return NULL; 2421} 2422 2423 2424//------------------------------Ideal_sign_extended_input---------------------- 2425// Check for useless sign-extension before a partial-word store 2426// (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2427// If (conIL == conIR && conIR <= num_bits) this simplifies to 2428// (StoreB ... (valIn) ) 2429Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2430 Node *val = in(MemNode::ValueIn); 2431 if( val->Opcode() == Op_RShiftI ) { 2432 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2433 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2434 Node *shl = val->in(1); 2435 if( shl->Opcode() == Op_LShiftI ) { 2436 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2437 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2438 set_req(MemNode::ValueIn, shl->in(1)); 2439 return this; 2440 } 2441 } 2442 } 2443 } 2444 return NULL; 2445} 2446 2447//------------------------------value_never_loaded----------------------------------- 2448// Determine whether there are any possible loads of the value stored. 2449// For simplicity, we actually check if there are any loads from the 2450// address stored to, not just for loads of the value stored by this node. 2451// 2452bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2453 Node *adr = in(Address); 2454 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2455 if (adr_oop == NULL) 2456 return false; 2457 if (!adr_oop->is_known_instance_field()) 2458 return false; // if not a distinct instance, there may be aliases of the address 2459 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2460 Node *use = adr->fast_out(i); 2461 int opc = use->Opcode(); 2462 if (use->is_Load() || use->is_LoadStore()) { 2463 return false; 2464 } 2465 } 2466 return true; 2467} 2468 2469//============================================================================= 2470//------------------------------Ideal------------------------------------------ 2471// If the store is from an AND mask that leaves the low bits untouched, then 2472// we can skip the AND operation. If the store is from a sign-extension 2473// (a left shift, then right shift) we can skip both. 2474Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2475 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2476 if( progress != NULL ) return progress; 2477 2478 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2479 if( progress != NULL ) return progress; 2480 2481 // Finally check the default case 2482 return StoreNode::Ideal(phase, can_reshape); 2483} 2484 2485//============================================================================= 2486//------------------------------Ideal------------------------------------------ 2487// If the store is from an AND mask that leaves the low bits untouched, then 2488// we can skip the AND operation 2489Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2490 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2491 if( progress != NULL ) return progress; 2492 2493 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2494 if( progress != NULL ) return progress; 2495 2496 // Finally check the default case 2497 return StoreNode::Ideal(phase, can_reshape); 2498} 2499 2500//============================================================================= 2501//------------------------------Identity--------------------------------------- 2502Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2503 // No need to card mark when storing a null ptr 2504 Node* my_store = in(MemNode::OopStore); 2505 if (my_store->is_Store()) { 2506 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2507 if( t1 == TypePtr::NULL_PTR ) { 2508 return in(MemNode::Memory); 2509 } 2510 } 2511 return this; 2512} 2513 2514//============================================================================= 2515//------------------------------Ideal--------------------------------------- 2516Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2517 Node* progress = StoreNode::Ideal(phase, can_reshape); 2518 if (progress != NULL) return progress; 2519 2520 Node* my_store = in(MemNode::OopStore); 2521 if (my_store->is_MergeMem()) { 2522 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2523 set_req(MemNode::OopStore, mem); 2524 return this; 2525 } 2526 2527 return NULL; 2528} 2529 2530//------------------------------Value----------------------------------------- 2531const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2532 // Either input is TOP ==> the result is TOP 2533 const Type *t = phase->type( in(MemNode::Memory) ); 2534 if( t == Type::TOP ) return Type::TOP; 2535 t = phase->type( in(MemNode::Address) ); 2536 if( t == Type::TOP ) return Type::TOP; 2537 t = phase->type( in(MemNode::ValueIn) ); 2538 if( t == Type::TOP ) return Type::TOP; 2539 // If extra input is TOP ==> the result is TOP 2540 t = phase->type( in(MemNode::OopStore) ); 2541 if( t == Type::TOP ) return Type::TOP; 2542 2543 return StoreNode::Value( phase ); 2544} 2545 2546 2547//============================================================================= 2548//----------------------------------SCMemProjNode------------------------------ 2549const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2550{ 2551 return bottom_type(); 2552} 2553 2554//============================================================================= 2555LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2556 init_req(MemNode::Control, c ); 2557 init_req(MemNode::Memory , mem); 2558 init_req(MemNode::Address, adr); 2559 init_req(MemNode::ValueIn, val); 2560 init_req( ExpectedIn, ex ); 2561 init_class_id(Class_LoadStore); 2562 2563} 2564 2565//============================================================================= 2566//-------------------------------adr_type-------------------------------------- 2567// Do we Match on this edge index or not? Do not match memory 2568const TypePtr* ClearArrayNode::adr_type() const { 2569 Node *adr = in(3); 2570 return MemNode::calculate_adr_type(adr->bottom_type()); 2571} 2572 2573//------------------------------match_edge------------------------------------- 2574// Do we Match on this edge index or not? Do not match memory 2575uint ClearArrayNode::match_edge(uint idx) const { 2576 return idx > 1; 2577} 2578 2579//------------------------------Identity--------------------------------------- 2580// Clearing a zero length array does nothing 2581Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2582 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2583} 2584 2585//------------------------------Idealize--------------------------------------- 2586// Clearing a short array is faster with stores 2587Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2588 const int unit = BytesPerLong; 2589 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2590 if (!t) return NULL; 2591 if (!t->is_con()) return NULL; 2592 intptr_t raw_count = t->get_con(); 2593 intptr_t size = raw_count; 2594 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2595 // Clearing nothing uses the Identity call. 2596 // Negative clears are possible on dead ClearArrays 2597 // (see jck test stmt114.stmt11402.val). 2598 if (size <= 0 || size % unit != 0) return NULL; 2599 intptr_t count = size / unit; 2600 // Length too long; use fast hardware clear 2601 if (size > Matcher::init_array_short_size) return NULL; 2602 Node *mem = in(1); 2603 if( phase->type(mem)==Type::TOP ) return NULL; 2604 Node *adr = in(3); 2605 const Type* at = phase->type(adr); 2606 if( at==Type::TOP ) return NULL; 2607 const TypePtr* atp = at->isa_ptr(); 2608 // adjust atp to be the correct array element address type 2609 if (atp == NULL) atp = TypePtr::BOTTOM; 2610 else atp = atp->add_offset(Type::OffsetBot); 2611 // Get base for derived pointer purposes 2612 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2613 Node *base = adr->in(1); 2614 2615 Node *zero = phase->makecon(TypeLong::ZERO); 2616 Node *off = phase->MakeConX(BytesPerLong); 2617 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2618 count--; 2619 while( count-- ) { 2620 mem = phase->transform(mem); 2621 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2622 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2623 } 2624 return mem; 2625} 2626 2627//----------------------------step_through---------------------------------- 2628// Return allocation input memory edge if it is different instance 2629// or itself if it is the one we are looking for. 2630bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2631 Node* n = *np; 2632 assert(n->is_ClearArray(), "sanity"); 2633 intptr_t offset; 2634 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2635 // This method is called only before Allocate nodes are expanded during 2636 // macro nodes expansion. Before that ClearArray nodes are only generated 2637 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2638 assert(alloc != NULL, "should have allocation"); 2639 if (alloc->_idx == instance_id) { 2640 // Can not bypass initialization of the instance we are looking for. 2641 return false; 2642 } 2643 // Otherwise skip it. 2644 InitializeNode* init = alloc->initialization(); 2645 if (init != NULL) 2646 *np = init->in(TypeFunc::Memory); 2647 else 2648 *np = alloc->in(TypeFunc::Memory); 2649 return true; 2650} 2651 2652//----------------------------clear_memory------------------------------------- 2653// Generate code to initialize object storage to zero. 2654Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2655 intptr_t start_offset, 2656 Node* end_offset, 2657 PhaseGVN* phase) { 2658 Compile* C = phase->C; 2659 intptr_t offset = start_offset; 2660 2661 int unit = BytesPerLong; 2662 if ((offset % unit) != 0) { 2663 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2664 adr = phase->transform(adr); 2665 const TypePtr* atp = TypeRawPtr::BOTTOM; 2666 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2667 mem = phase->transform(mem); 2668 offset += BytesPerInt; 2669 } 2670 assert((offset % unit) == 0, ""); 2671 2672 // Initialize the remaining stuff, if any, with a ClearArray. 2673 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2674} 2675 2676Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2677 Node* start_offset, 2678 Node* end_offset, 2679 PhaseGVN* phase) { 2680 if (start_offset == end_offset) { 2681 // nothing to do 2682 return mem; 2683 } 2684 2685 Compile* C = phase->C; 2686 int unit = BytesPerLong; 2687 Node* zbase = start_offset; 2688 Node* zend = end_offset; 2689 2690 // Scale to the unit required by the CPU: 2691 if (!Matcher::init_array_count_is_in_bytes) { 2692 Node* shift = phase->intcon(exact_log2(unit)); 2693 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2694 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2695 } 2696 2697 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2698 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2699 2700 // Bulk clear double-words 2701 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2702 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2703 return phase->transform(mem); 2704} 2705 2706Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2707 intptr_t start_offset, 2708 intptr_t end_offset, 2709 PhaseGVN* phase) { 2710 if (start_offset == end_offset) { 2711 // nothing to do 2712 return mem; 2713 } 2714 2715 Compile* C = phase->C; 2716 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2717 intptr_t done_offset = end_offset; 2718 if ((done_offset % BytesPerLong) != 0) { 2719 done_offset -= BytesPerInt; 2720 } 2721 if (done_offset > start_offset) { 2722 mem = clear_memory(ctl, mem, dest, 2723 start_offset, phase->MakeConX(done_offset), phase); 2724 } 2725 if (done_offset < end_offset) { // emit the final 32-bit store 2726 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2727 adr = phase->transform(adr); 2728 const TypePtr* atp = TypeRawPtr::BOTTOM; 2729 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2730 mem = phase->transform(mem); 2731 done_offset += BytesPerInt; 2732 } 2733 assert(done_offset == end_offset, ""); 2734 return mem; 2735} 2736 2737//============================================================================= 2738// Do not match memory edge. 2739uint StrIntrinsicNode::match_edge(uint idx) const { 2740 return idx == 2 || idx == 3; 2741} 2742 2743//------------------------------Ideal------------------------------------------ 2744// Return a node which is more "ideal" than the current node. Strip out 2745// control copies 2746Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2747 if (remove_dead_region(phase, can_reshape)) return this; 2748 // Don't bother trying to transform a dead node 2749 if (in(0) && in(0)->is_top()) return NULL; 2750 2751 if (can_reshape) { 2752 Node* mem = phase->transform(in(MemNode::Memory)); 2753 // If transformed to a MergeMem, get the desired slice 2754 uint alias_idx = phase->C->get_alias_index(adr_type()); 2755 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; 2756 if (mem != in(MemNode::Memory)) { 2757 set_req(MemNode::Memory, mem); 2758 return this; 2759 } 2760 } 2761 return NULL; 2762} 2763 2764//------------------------------Value------------------------------------------ 2765const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { 2766 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; 2767 return bottom_type(); 2768} 2769 2770//============================================================================= 2771MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2772 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2773 _adr_type(C->get_adr_type(alias_idx)) 2774{ 2775 init_class_id(Class_MemBar); 2776 Node* top = C->top(); 2777 init_req(TypeFunc::I_O,top); 2778 init_req(TypeFunc::FramePtr,top); 2779 init_req(TypeFunc::ReturnAdr,top); 2780 if (precedent != NULL) 2781 init_req(TypeFunc::Parms, precedent); 2782} 2783 2784//------------------------------cmp-------------------------------------------- 2785uint MemBarNode::hash() const { return NO_HASH; } 2786uint MemBarNode::cmp( const Node &n ) const { 2787 return (&n == this); // Always fail except on self 2788} 2789 2790//------------------------------make------------------------------------------- 2791MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2792 int len = Precedent + (pn == NULL? 0: 1); 2793 switch (opcode) { 2794 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2795 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2796 case Op_MemBarAcquireLock: return new(C, len) MemBarAcquireLockNode(C, atp, pn); 2797 case Op_MemBarReleaseLock: return new(C, len) MemBarReleaseLockNode(C, atp, pn); 2798 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2799 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2800 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2801 case Op_MemBarStoreStore: return new(C, len) MemBarStoreStoreNode(C, atp, pn); 2802 default: ShouldNotReachHere(); return NULL; 2803 } 2804} 2805 2806//------------------------------Ideal------------------------------------------ 2807// Return a node which is more "ideal" than the current node. Strip out 2808// control copies 2809Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2810 if (remove_dead_region(phase, can_reshape)) return this; 2811 // Don't bother trying to transform a dead node 2812 if (in(0) && in(0)->is_top()) return NULL; 2813 2814 // Eliminate volatile MemBars for scalar replaced objects. 2815 if (can_reshape && req() == (Precedent+1) && 2816 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { 2817 // Volatile field loads and stores. 2818 Node* my_mem = in(MemBarNode::Precedent); 2819 if (my_mem != NULL && my_mem->is_Mem()) { 2820 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2821 // Check for scalar replaced object reference. 2822 if( t_oop != NULL && t_oop->is_known_instance_field() && 2823 t_oop->offset() != Type::OffsetBot && 2824 t_oop->offset() != Type::OffsetTop) { 2825 // Replace MemBar projections by its inputs. 2826 PhaseIterGVN* igvn = phase->is_IterGVN(); 2827 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2828 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2829 // Must return either the original node (now dead) or a new node 2830 // (Do not return a top here, since that would break the uniqueness of top.) 2831 return new (phase->C, 1) ConINode(TypeInt::ZERO); 2832 } 2833 } 2834 } 2835 return NULL; 2836} 2837 2838//------------------------------Value------------------------------------------ 2839const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2840 if( !in(0) ) return Type::TOP; 2841 if( phase->type(in(0)) == Type::TOP ) 2842 return Type::TOP; 2843 return TypeTuple::MEMBAR; 2844} 2845 2846//------------------------------match------------------------------------------ 2847// Construct projections for memory. 2848Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2849 switch (proj->_con) { 2850 case TypeFunc::Control: 2851 case TypeFunc::Memory: 2852 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2853 } 2854 ShouldNotReachHere(); 2855 return NULL; 2856} 2857 2858//===========================InitializeNode==================================== 2859// SUMMARY: 2860// This node acts as a memory barrier on raw memory, after some raw stores. 2861// The 'cooked' oop value feeds from the Initialize, not the Allocation. 2862// The Initialize can 'capture' suitably constrained stores as raw inits. 2863// It can coalesce related raw stores into larger units (called 'tiles'). 2864// It can avoid zeroing new storage for memory units which have raw inits. 2865// At macro-expansion, it is marked 'complete', and does not optimize further. 2866// 2867// EXAMPLE: 2868// The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2869// ctl = incoming control; mem* = incoming memory 2870// (Note: A star * on a memory edge denotes I/O and other standard edges.) 2871// First allocate uninitialized memory and fill in the header: 2872// alloc = (Allocate ctl mem* 16 #short[].klass ...) 2873// ctl := alloc.Control; mem* := alloc.Memory* 2874// rawmem = alloc.Memory; rawoop = alloc.RawAddress 2875// Then initialize to zero the non-header parts of the raw memory block: 2876// init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2877// ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2878// After the initialize node executes, the object is ready for service: 2879// oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2880// Suppose its body is immediately initialized as {1,2}: 2881// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2882// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2883// mem.SLICE(#short[*]) := store2 2884// 2885// DETAILS: 2886// An InitializeNode collects and isolates object initialization after 2887// an AllocateNode and before the next possible safepoint. As a 2888// memory barrier (MemBarNode), it keeps critical stores from drifting 2889// down past any safepoint or any publication of the allocation. 2890// Before this barrier, a newly-allocated object may have uninitialized bits. 2891// After this barrier, it may be treated as a real oop, and GC is allowed. 2892// 2893// The semantics of the InitializeNode include an implicit zeroing of 2894// the new object from object header to the end of the object. 2895// (The object header and end are determined by the AllocateNode.) 2896// 2897// Certain stores may be added as direct inputs to the InitializeNode. 2898// These stores must update raw memory, and they must be to addresses 2899// derived from the raw address produced by AllocateNode, and with 2900// a constant offset. They must be ordered by increasing offset. 2901// The first one is at in(RawStores), the last at in(req()-1). 2902// Unlike most memory operations, they are not linked in a chain, 2903// but are displayed in parallel as users of the rawmem output of 2904// the allocation. 2905// 2906// (See comments in InitializeNode::capture_store, which continue 2907// the example given above.) 2908// 2909// When the associated Allocate is macro-expanded, the InitializeNode 2910// may be rewritten to optimize collected stores. A ClearArrayNode 2911// may also be created at that point to represent any required zeroing. 2912// The InitializeNode is then marked 'complete', prohibiting further 2913// capturing of nearby memory operations. 2914// 2915// During macro-expansion, all captured initializations which store 2916// constant values of 32 bits or smaller are coalesced (if advantageous) 2917// into larger 'tiles' 32 or 64 bits. This allows an object to be 2918// initialized in fewer memory operations. Memory words which are 2919// covered by neither tiles nor non-constant stores are pre-zeroed 2920// by explicit stores of zero. (The code shape happens to do all 2921// zeroing first, then all other stores, with both sequences occurring 2922// in order of ascending offsets.) 2923// 2924// Alternatively, code may be inserted between an AllocateNode and its 2925// InitializeNode, to perform arbitrary initialization of the new object. 2926// E.g., the object copying intrinsics insert complex data transfers here. 2927// The initialization must then be marked as 'complete' disable the 2928// built-in zeroing semantics and the collection of initializing stores. 2929// 2930// While an InitializeNode is incomplete, reads from the memory state 2931// produced by it are optimizable if they match the control edge and 2932// new oop address associated with the allocation/initialization. 2933// They return a stored value (if the offset matches) or else zero. 2934// A write to the memory state, if it matches control and address, 2935// and if it is to a constant offset, may be 'captured' by the 2936// InitializeNode. It is cloned as a raw memory operation and rewired 2937// inside the initialization, to the raw oop produced by the allocation. 2938// Operations on addresses which are provably distinct (e.g., to 2939// other AllocateNodes) are allowed to bypass the initialization. 2940// 2941// The effect of all this is to consolidate object initialization 2942// (both arrays and non-arrays, both piecewise and bulk) into a 2943// single location, where it can be optimized as a unit. 2944// 2945// Only stores with an offset less than TrackedInitializationLimit words 2946// will be considered for capture by an InitializeNode. This puts a 2947// reasonable limit on the complexity of optimized initializations. 2948 2949//---------------------------InitializeNode------------------------------------ 2950InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2951 : _is_complete(Incomplete), _does_not_escape(false), 2952 MemBarNode(C, adr_type, rawoop) 2953{ 2954 init_class_id(Class_Initialize); 2955 2956 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2957 assert(in(RawAddress) == rawoop, "proper init"); 2958 // Note: allocation() can be NULL, for secondary initialization barriers 2959} 2960 2961// Since this node is not matched, it will be processed by the 2962// register allocator. Declare that there are no constraints 2963// on the allocation of the RawAddress edge. 2964const RegMask &InitializeNode::in_RegMask(uint idx) const { 2965 // This edge should be set to top, by the set_complete. But be conservative. 2966 if (idx == InitializeNode::RawAddress) 2967 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2968 return RegMask::Empty; 2969} 2970 2971Node* InitializeNode::memory(uint alias_idx) { 2972 Node* mem = in(Memory); 2973 if (mem->is_MergeMem()) { 2974 return mem->as_MergeMem()->memory_at(alias_idx); 2975 } else { 2976 // incoming raw memory is not split 2977 return mem; 2978 } 2979} 2980 2981bool InitializeNode::is_non_zero() { 2982 if (is_complete()) return false; 2983 remove_extra_zeroes(); 2984 return (req() > RawStores); 2985} 2986 2987void InitializeNode::set_complete(PhaseGVN* phase) { 2988 assert(!is_complete(), "caller responsibility"); 2989 _is_complete = Complete; 2990 2991 // After this node is complete, it contains a bunch of 2992 // raw-memory initializations. There is no need for 2993 // it to have anything to do with non-raw memory effects. 2994 // Therefore, tell all non-raw users to re-optimize themselves, 2995 // after skipping the memory effects of this initialization. 2996 PhaseIterGVN* igvn = phase->is_IterGVN(); 2997 if (igvn) igvn->add_users_to_worklist(this); 2998} 2999 3000// convenience function 3001// return false if the init contains any stores already 3002bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3003 InitializeNode* init = initialization(); 3004 if (init == NULL || init->is_complete()) return false; 3005 init->remove_extra_zeroes(); 3006 // for now, if this allocation has already collected any inits, bail: 3007 if (init->is_non_zero()) return false; 3008 init->set_complete(phase); 3009 return true; 3010} 3011 3012void InitializeNode::remove_extra_zeroes() { 3013 if (req() == RawStores) return; 3014 Node* zmem = zero_memory(); 3015 uint fill = RawStores; 3016 for (uint i = fill; i < req(); i++) { 3017 Node* n = in(i); 3018 if (n->is_top() || n == zmem) continue; // skip 3019 if (fill < i) set_req(fill, n); // compact 3020 ++fill; 3021 } 3022 // delete any empty spaces created: 3023 while (fill < req()) { 3024 del_req(fill); 3025 } 3026} 3027 3028// Helper for remembering which stores go with which offsets. 3029intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3030 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3031 intptr_t offset = -1; 3032 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3033 phase, offset); 3034 if (base == NULL) return -1; // something is dead, 3035 if (offset < 0) return -1; // dead, dead 3036 return offset; 3037} 3038 3039// Helper for proving that an initialization expression is 3040// "simple enough" to be folded into an object initialization. 3041// Attempts to prove that a store's initial value 'n' can be captured 3042// within the initialization without creating a vicious cycle, such as: 3043// { Foo p = new Foo(); p.next = p; } 3044// True for constants and parameters and small combinations thereof. 3045bool InitializeNode::detect_init_independence(Node* n, 3046 bool st_is_pinned, 3047 int& count) { 3048 if (n == NULL) return true; // (can this really happen?) 3049 if (n->is_Proj()) n = n->in(0); 3050 if (n == this) return false; // found a cycle 3051 if (n->is_Con()) return true; 3052 if (n->is_Start()) return true; // params, etc., are OK 3053 if (n->is_Root()) return true; // even better 3054 3055 Node* ctl = n->in(0); 3056 if (ctl != NULL && !ctl->is_top()) { 3057 if (ctl->is_Proj()) ctl = ctl->in(0); 3058 if (ctl == this) return false; 3059 3060 // If we already know that the enclosing memory op is pinned right after 3061 // the init, then any control flow that the store has picked up 3062 // must have preceded the init, or else be equal to the init. 3063 // Even after loop optimizations (which might change control edges) 3064 // a store is never pinned *before* the availability of its inputs. 3065 if (!MemNode::all_controls_dominate(n, this)) 3066 return false; // failed to prove a good control 3067 3068 } 3069 3070 // Check data edges for possible dependencies on 'this'. 3071 if ((count += 1) > 20) return false; // complexity limit 3072 for (uint i = 1; i < n->req(); i++) { 3073 Node* m = n->in(i); 3074 if (m == NULL || m == n || m->is_top()) continue; 3075 uint first_i = n->find_edge(m); 3076 if (i != first_i) continue; // process duplicate edge just once 3077 if (!detect_init_independence(m, st_is_pinned, count)) { 3078 return false; 3079 } 3080 } 3081 3082 return true; 3083} 3084 3085// Here are all the checks a Store must pass before it can be moved into 3086// an initialization. Returns zero if a check fails. 3087// On success, returns the (constant) offset to which the store applies, 3088// within the initialized memory. 3089intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 3090 const int FAIL = 0; 3091 if (st->req() != MemNode::ValueIn + 1) 3092 return FAIL; // an inscrutable StoreNode (card mark?) 3093 Node* ctl = st->in(MemNode::Control); 3094 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3095 return FAIL; // must be unconditional after the initialization 3096 Node* mem = st->in(MemNode::Memory); 3097 if (!(mem->is_Proj() && mem->in(0) == this)) 3098 return FAIL; // must not be preceded by other stores 3099 Node* adr = st->in(MemNode::Address); 3100 intptr_t offset; 3101 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3102 if (alloc == NULL) 3103 return FAIL; // inscrutable address 3104 if (alloc != allocation()) 3105 return FAIL; // wrong allocation! (store needs to float up) 3106 Node* val = st->in(MemNode::ValueIn); 3107 int complexity_count = 0; 3108 if (!detect_init_independence(val, true, complexity_count)) 3109 return FAIL; // stored value must be 'simple enough' 3110 3111 return offset; // success 3112} 3113 3114// Find the captured store in(i) which corresponds to the range 3115// [start..start+size) in the initialized object. 3116// If there is one, return its index i. If there isn't, return the 3117// negative of the index where it should be inserted. 3118// Return 0 if the queried range overlaps an initialization boundary 3119// or if dead code is encountered. 3120// If size_in_bytes is zero, do not bother with overlap checks. 3121int InitializeNode::captured_store_insertion_point(intptr_t start, 3122 int size_in_bytes, 3123 PhaseTransform* phase) { 3124 const int FAIL = 0, MAX_STORE = BytesPerLong; 3125 3126 if (is_complete()) 3127 return FAIL; // arraycopy got here first; punt 3128 3129 assert(allocation() != NULL, "must be present"); 3130 3131 // no negatives, no header fields: 3132 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3133 3134 // after a certain size, we bail out on tracking all the stores: 3135 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3136 if (start >= ti_limit) return FAIL; 3137 3138 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3139 if (i >= limit) return -(int)i; // not found; here is where to put it 3140 3141 Node* st = in(i); 3142 intptr_t st_off = get_store_offset(st, phase); 3143 if (st_off < 0) { 3144 if (st != zero_memory()) { 3145 return FAIL; // bail out if there is dead garbage 3146 } 3147 } else if (st_off > start) { 3148 // ...we are done, since stores are ordered 3149 if (st_off < start + size_in_bytes) { 3150 return FAIL; // the next store overlaps 3151 } 3152 return -(int)i; // not found; here is where to put it 3153 } else if (st_off < start) { 3154 if (size_in_bytes != 0 && 3155 start < st_off + MAX_STORE && 3156 start < st_off + st->as_Store()->memory_size()) { 3157 return FAIL; // the previous store overlaps 3158 } 3159 } else { 3160 if (size_in_bytes != 0 && 3161 st->as_Store()->memory_size() != size_in_bytes) { 3162 return FAIL; // mismatched store size 3163 } 3164 return i; 3165 } 3166 3167 ++i; 3168 } 3169} 3170 3171// Look for a captured store which initializes at the offset 'start' 3172// with the given size. If there is no such store, and no other 3173// initialization interferes, then return zero_memory (the memory 3174// projection of the AllocateNode). 3175Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3176 PhaseTransform* phase) { 3177 assert(stores_are_sane(phase), ""); 3178 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3179 if (i == 0) { 3180 return NULL; // something is dead 3181 } else if (i < 0) { 3182 return zero_memory(); // just primordial zero bits here 3183 } else { 3184 Node* st = in(i); // here is the store at this position 3185 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3186 return st; 3187 } 3188} 3189 3190// Create, as a raw pointer, an address within my new object at 'offset'. 3191Node* InitializeNode::make_raw_address(intptr_t offset, 3192 PhaseTransform* phase) { 3193 Node* addr = in(RawAddress); 3194 if (offset != 0) { 3195 Compile* C = phase->C; 3196 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 3197 phase->MakeConX(offset)) ); 3198 } 3199 return addr; 3200} 3201 3202// Clone the given store, converting it into a raw store 3203// initializing a field or element of my new object. 3204// Caller is responsible for retiring the original store, 3205// with subsume_node or the like. 3206// 3207// From the example above InitializeNode::InitializeNode, 3208// here are the old stores to be captured: 3209// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3210// store2 = (StoreC init.Control store1 (+ oop 14) 2) 3211// 3212// Here is the changed code; note the extra edges on init: 3213// alloc = (Allocate ...) 3214// rawoop = alloc.RawAddress 3215// rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3216// rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3217// init = (Initialize alloc.Control alloc.Memory rawoop 3218// rawstore1 rawstore2) 3219// 3220Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3221 PhaseTransform* phase) { 3222 assert(stores_are_sane(phase), ""); 3223 3224 if (start < 0) return NULL; 3225 assert(can_capture_store(st, phase) == start, "sanity"); 3226 3227 Compile* C = phase->C; 3228 int size_in_bytes = st->memory_size(); 3229 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3230 if (i == 0) return NULL; // bail out 3231 Node* prev_mem = NULL; // raw memory for the captured store 3232 if (i > 0) { 3233 prev_mem = in(i); // there is a pre-existing store under this one 3234 set_req(i, C->top()); // temporarily disconnect it 3235 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3236 } else { 3237 i = -i; // no pre-existing store 3238 prev_mem = zero_memory(); // a slice of the newly allocated object 3239 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3240 set_req(--i, C->top()); // reuse this edge; it has been folded away 3241 else 3242 ins_req(i, C->top()); // build a new edge 3243 } 3244 Node* new_st = st->clone(); 3245 new_st->set_req(MemNode::Control, in(Control)); 3246 new_st->set_req(MemNode::Memory, prev_mem); 3247 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3248 new_st = phase->transform(new_st); 3249 3250 // At this point, new_st might have swallowed a pre-existing store 3251 // at the same offset, or perhaps new_st might have disappeared, 3252 // if it redundantly stored the same value (or zero to fresh memory). 3253 3254 // In any case, wire it in: 3255 set_req(i, new_st); 3256 3257 // The caller may now kill the old guy. 3258 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3259 assert(check_st == new_st || check_st == NULL, "must be findable"); 3260 assert(!is_complete(), ""); 3261 return new_st; 3262} 3263 3264static bool store_constant(jlong* tiles, int num_tiles, 3265 intptr_t st_off, int st_size, 3266 jlong con) { 3267 if ((st_off & (st_size-1)) != 0) 3268 return false; // strange store offset (assume size==2**N) 3269 address addr = (address)tiles + st_off; 3270 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3271 switch (st_size) { 3272 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3273 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3274 case sizeof(jint): *(jint*) addr = (jint) con; break; 3275 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3276 default: return false; // strange store size (detect size!=2**N here) 3277 } 3278 return true; // return success to caller 3279} 3280 3281// Coalesce subword constants into int constants and possibly 3282// into long constants. The goal, if the CPU permits, 3283// is to initialize the object with a small number of 64-bit tiles. 3284// Also, convert floating-point constants to bit patterns. 3285// Non-constants are not relevant to this pass. 3286// 3287// In terms of the running example on InitializeNode::InitializeNode 3288// and InitializeNode::capture_store, here is the transformation 3289// of rawstore1 and rawstore2 into rawstore12: 3290// alloc = (Allocate ...) 3291// rawoop = alloc.RawAddress 3292// tile12 = 0x00010002 3293// rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3294// init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3295// 3296void 3297InitializeNode::coalesce_subword_stores(intptr_t header_size, 3298 Node* size_in_bytes, 3299 PhaseGVN* phase) { 3300 Compile* C = phase->C; 3301 3302 assert(stores_are_sane(phase), ""); 3303 // Note: After this pass, they are not completely sane, 3304 // since there may be some overlaps. 3305 3306 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3307 3308 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3309 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3310 size_limit = MIN2(size_limit, ti_limit); 3311 size_limit = align_size_up(size_limit, BytesPerLong); 3312 int num_tiles = size_limit / BytesPerLong; 3313 3314 // allocate space for the tile map: 3315 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3316 jlong tiles_buf[small_len]; 3317 Node* nodes_buf[small_len]; 3318 jlong inits_buf[small_len]; 3319 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3320 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3321 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3322 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3323 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3324 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3325 // tiles: exact bitwise model of all primitive constants 3326 // nodes: last constant-storing node subsumed into the tiles model 3327 // inits: which bytes (in each tile) are touched by any initializations 3328 3329 //// Pass A: Fill in the tile model with any relevant stores. 3330 3331 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3332 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3333 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3334 Node* zmem = zero_memory(); // initially zero memory state 3335 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3336 Node* st = in(i); 3337 intptr_t st_off = get_store_offset(st, phase); 3338 3339 // Figure out the store's offset and constant value: 3340 if (st_off < header_size) continue; //skip (ignore header) 3341 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3342 int st_size = st->as_Store()->memory_size(); 3343 if (st_off + st_size > size_limit) break; 3344 3345 // Record which bytes are touched, whether by constant or not. 3346 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3347 continue; // skip (strange store size) 3348 3349 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3350 if (!val->singleton()) continue; //skip (non-con store) 3351 BasicType type = val->basic_type(); 3352 3353 jlong con = 0; 3354 switch (type) { 3355 case T_INT: con = val->is_int()->get_con(); break; 3356 case T_LONG: con = val->is_long()->get_con(); break; 3357 case T_FLOAT: con = jint_cast(val->getf()); break; 3358 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3359 default: continue; //skip (odd store type) 3360 } 3361 3362 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3363 st->Opcode() == Op_StoreL) { 3364 continue; // This StoreL is already optimal. 3365 } 3366 3367 // Store down the constant. 3368 store_constant(tiles, num_tiles, st_off, st_size, con); 3369 3370 intptr_t j = st_off >> LogBytesPerLong; 3371 3372 if (type == T_INT && st_size == BytesPerInt 3373 && (st_off & BytesPerInt) == BytesPerInt) { 3374 jlong lcon = tiles[j]; 3375 if (!Matcher::isSimpleConstant64(lcon) && 3376 st->Opcode() == Op_StoreI) { 3377 // This StoreI is already optimal by itself. 3378 jint* intcon = (jint*) &tiles[j]; 3379 intcon[1] = 0; // undo the store_constant() 3380 3381 // If the previous store is also optimal by itself, back up and 3382 // undo the action of the previous loop iteration... if we can. 3383 // But if we can't, just let the previous half take care of itself. 3384 st = nodes[j]; 3385 st_off -= BytesPerInt; 3386 con = intcon[0]; 3387 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3388 assert(st_off >= header_size, "still ignoring header"); 3389 assert(get_store_offset(st, phase) == st_off, "must be"); 3390 assert(in(i-1) == zmem, "must be"); 3391 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3392 assert(con == tcon->is_int()->get_con(), "must be"); 3393 // Undo the effects of the previous loop trip, which swallowed st: 3394 intcon[0] = 0; // undo store_constant() 3395 set_req(i-1, st); // undo set_req(i, zmem) 3396 nodes[j] = NULL; // undo nodes[j] = st 3397 --old_subword; // undo ++old_subword 3398 } 3399 continue; // This StoreI is already optimal. 3400 } 3401 } 3402 3403 // This store is not needed. 3404 set_req(i, zmem); 3405 nodes[j] = st; // record for the moment 3406 if (st_size < BytesPerLong) // something has changed 3407 ++old_subword; // includes int/float, but who's counting... 3408 else ++old_long; 3409 } 3410 3411 if ((old_subword + old_long) == 0) 3412 return; // nothing more to do 3413 3414 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3415 // Be sure to insert them before overlapping non-constant stores. 3416 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3417 for (int j = 0; j < num_tiles; j++) { 3418 jlong con = tiles[j]; 3419 jlong init = inits[j]; 3420 if (con == 0) continue; 3421 jint con0, con1; // split the constant, address-wise 3422 jint init0, init1; // split the init map, address-wise 3423 { union { jlong con; jint intcon[2]; } u; 3424 u.con = con; 3425 con0 = u.intcon[0]; 3426 con1 = u.intcon[1]; 3427 u.con = init; 3428 init0 = u.intcon[0]; 3429 init1 = u.intcon[1]; 3430 } 3431 3432 Node* old = nodes[j]; 3433 assert(old != NULL, "need the prior store"); 3434 intptr_t offset = (j * BytesPerLong); 3435 3436 bool split = !Matcher::isSimpleConstant64(con); 3437 3438 if (offset < header_size) { 3439 assert(offset + BytesPerInt >= header_size, "second int counts"); 3440 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3441 split = true; // only the second word counts 3442 // Example: int a[] = { 42 ... } 3443 } else if (con0 == 0 && init0 == -1) { 3444 split = true; // first word is covered by full inits 3445 // Example: int a[] = { ... foo(), 42 ... } 3446 } else if (con1 == 0 && init1 == -1) { 3447 split = true; // second word is covered by full inits 3448 // Example: int a[] = { ... 42, foo() ... } 3449 } 3450 3451 // Here's a case where init0 is neither 0 nor -1: 3452 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3453 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3454 // In this case the tile is not split; it is (jlong)42. 3455 // The big tile is stored down, and then the foo() value is inserted. 3456 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3457 3458 Node* ctl = old->in(MemNode::Control); 3459 Node* adr = make_raw_address(offset, phase); 3460 const TypePtr* atp = TypeRawPtr::BOTTOM; 3461 3462 // One or two coalesced stores to plop down. 3463 Node* st[2]; 3464 intptr_t off[2]; 3465 int nst = 0; 3466 if (!split) { 3467 ++new_long; 3468 off[nst] = offset; 3469 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3470 phase->longcon(con), T_LONG); 3471 } else { 3472 // Omit either if it is a zero. 3473 if (con0 != 0) { 3474 ++new_int; 3475 off[nst] = offset; 3476 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3477 phase->intcon(con0), T_INT); 3478 } 3479 if (con1 != 0) { 3480 ++new_int; 3481 offset += BytesPerInt; 3482 adr = make_raw_address(offset, phase); 3483 off[nst] = offset; 3484 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3485 phase->intcon(con1), T_INT); 3486 } 3487 } 3488 3489 // Insert second store first, then the first before the second. 3490 // Insert each one just before any overlapping non-constant stores. 3491 while (nst > 0) { 3492 Node* st1 = st[--nst]; 3493 C->copy_node_notes_to(st1, old); 3494 st1 = phase->transform(st1); 3495 offset = off[nst]; 3496 assert(offset >= header_size, "do not smash header"); 3497 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3498 guarantee(ins_idx != 0, "must re-insert constant store"); 3499 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3500 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3501 set_req(--ins_idx, st1); 3502 else 3503 ins_req(ins_idx, st1); 3504 } 3505 } 3506 3507 if (PrintCompilation && WizardMode) 3508 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3509 old_subword, old_long, new_int, new_long); 3510 if (C->log() != NULL) 3511 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3512 old_subword, old_long, new_int, new_long); 3513 3514 // Clean up any remaining occurrences of zmem: 3515 remove_extra_zeroes(); 3516} 3517 3518// Explore forward from in(start) to find the first fully initialized 3519// word, and return its offset. Skip groups of subword stores which 3520// together initialize full words. If in(start) is itself part of a 3521// fully initialized word, return the offset of in(start). If there 3522// are no following full-word stores, or if something is fishy, return 3523// a negative value. 3524intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3525 int int_map = 0; 3526 intptr_t int_map_off = 0; 3527 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3528 3529 for (uint i = start, limit = req(); i < limit; i++) { 3530 Node* st = in(i); 3531 3532 intptr_t st_off = get_store_offset(st, phase); 3533 if (st_off < 0) break; // return conservative answer 3534 3535 int st_size = st->as_Store()->memory_size(); 3536 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3537 return st_off; // we found a complete word init 3538 } 3539 3540 // update the map: 3541 3542 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3543 if (this_int_off != int_map_off) { 3544 // reset the map: 3545 int_map = 0; 3546 int_map_off = this_int_off; 3547 } 3548 3549 int subword_off = st_off - this_int_off; 3550 int_map |= right_n_bits(st_size) << subword_off; 3551 if ((int_map & FULL_MAP) == FULL_MAP) { 3552 return this_int_off; // we found a complete word init 3553 } 3554 3555 // Did this store hit or cross the word boundary? 3556 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3557 if (next_int_off == this_int_off + BytesPerInt) { 3558 // We passed the current int, without fully initializing it. 3559 int_map_off = next_int_off; 3560 int_map >>= BytesPerInt; 3561 } else if (next_int_off > this_int_off + BytesPerInt) { 3562 // We passed the current and next int. 3563 return this_int_off + BytesPerInt; 3564 } 3565 } 3566 3567 return -1; 3568} 3569 3570 3571// Called when the associated AllocateNode is expanded into CFG. 3572// At this point, we may perform additional optimizations. 3573// Linearize the stores by ascending offset, to make memory 3574// activity as coherent as possible. 3575Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3576 intptr_t header_size, 3577 Node* size_in_bytes, 3578 PhaseGVN* phase) { 3579 assert(!is_complete(), "not already complete"); 3580 assert(stores_are_sane(phase), ""); 3581 assert(allocation() != NULL, "must be present"); 3582 3583 remove_extra_zeroes(); 3584 3585 if (ReduceFieldZeroing || ReduceBulkZeroing) 3586 // reduce instruction count for common initialization patterns 3587 coalesce_subword_stores(header_size, size_in_bytes, phase); 3588 3589 Node* zmem = zero_memory(); // initially zero memory state 3590 Node* inits = zmem; // accumulating a linearized chain of inits 3591 #ifdef ASSERT 3592 intptr_t first_offset = allocation()->minimum_header_size(); 3593 intptr_t last_init_off = first_offset; // previous init offset 3594 intptr_t last_init_end = first_offset; // previous init offset+size 3595 intptr_t last_tile_end = first_offset; // previous tile offset+size 3596 #endif 3597 intptr_t zeroes_done = header_size; 3598 3599 bool do_zeroing = true; // we might give up if inits are very sparse 3600 int big_init_gaps = 0; // how many large gaps have we seen? 3601 3602 if (ZeroTLAB) do_zeroing = false; 3603 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3604 3605 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3606 Node* st = in(i); 3607 intptr_t st_off = get_store_offset(st, phase); 3608 if (st_off < 0) 3609 break; // unknown junk in the inits 3610 if (st->in(MemNode::Memory) != zmem) 3611 break; // complicated store chains somehow in list 3612 3613 int st_size = st->as_Store()->memory_size(); 3614 intptr_t next_init_off = st_off + st_size; 3615 3616 if (do_zeroing && zeroes_done < next_init_off) { 3617 // See if this store needs a zero before it or under it. 3618 intptr_t zeroes_needed = st_off; 3619 3620 if (st_size < BytesPerInt) { 3621 // Look for subword stores which only partially initialize words. 3622 // If we find some, we must lay down some word-level zeroes first, 3623 // underneath the subword stores. 3624 // 3625 // Examples: 3626 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3627 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3628 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3629 // 3630 // Note: coalesce_subword_stores may have already done this, 3631 // if it was prompted by constant non-zero subword initializers. 3632 // But this case can still arise with non-constant stores. 3633 3634 intptr_t next_full_store = find_next_fullword_store(i, phase); 3635 3636 // In the examples above: 3637 // in(i) p q r s x y z 3638 // st_off 12 13 14 15 12 13 14 3639 // st_size 1 1 1 1 1 1 1 3640 // next_full_s. 12 16 16 16 16 16 16 3641 // z's_done 12 16 16 16 12 16 12 3642 // z's_needed 12 16 16 16 16 16 16 3643 // zsize 0 0 0 0 4 0 4 3644 if (next_full_store < 0) { 3645 // Conservative tack: Zero to end of current word. 3646 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3647 } else { 3648 // Zero to beginning of next fully initialized word. 3649 // Or, don't zero at all, if we are already in that word. 3650 assert(next_full_store >= zeroes_needed, "must go forward"); 3651 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3652 zeroes_needed = next_full_store; 3653 } 3654 } 3655 3656 if (zeroes_needed > zeroes_done) { 3657 intptr_t zsize = zeroes_needed - zeroes_done; 3658 // Do some incremental zeroing on rawmem, in parallel with inits. 3659 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3660 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3661 zeroes_done, zeroes_needed, 3662 phase); 3663 zeroes_done = zeroes_needed; 3664 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3665 do_zeroing = false; // leave the hole, next time 3666 } 3667 } 3668 3669 // Collect the store and move on: 3670 st->set_req(MemNode::Memory, inits); 3671 inits = st; // put it on the linearized chain 3672 set_req(i, zmem); // unhook from previous position 3673 3674 if (zeroes_done == st_off) 3675 zeroes_done = next_init_off; 3676 3677 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3678 3679 #ifdef ASSERT 3680 // Various order invariants. Weaker than stores_are_sane because 3681 // a large constant tile can be filled in by smaller non-constant stores. 3682 assert(st_off >= last_init_off, "inits do not reverse"); 3683 last_init_off = st_off; 3684 const Type* val = NULL; 3685 if (st_size >= BytesPerInt && 3686 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3687 (int)val->basic_type() < (int)T_OBJECT) { 3688 assert(st_off >= last_tile_end, "tiles do not overlap"); 3689 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3690 last_tile_end = MAX2(last_tile_end, next_init_off); 3691 } else { 3692 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3693 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3694 assert(st_off >= last_init_end, "inits do not overlap"); 3695 last_init_end = next_init_off; // it's a non-tile 3696 } 3697 #endif //ASSERT 3698 } 3699 3700 remove_extra_zeroes(); // clear out all the zmems left over 3701 add_req(inits); 3702 3703 if (!ZeroTLAB) { 3704 // If anything remains to be zeroed, zero it all now. 3705 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3706 // if it is the last unused 4 bytes of an instance, forget about it 3707 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3708 if (zeroes_done + BytesPerLong >= size_limit) { 3709 assert(allocation() != NULL, ""); 3710 if (allocation()->Opcode() == Op_Allocate) { 3711 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3712 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3713 if (zeroes_done == k->layout_helper()) 3714 zeroes_done = size_limit; 3715 } 3716 } 3717 if (zeroes_done < size_limit) { 3718 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3719 zeroes_done, size_in_bytes, phase); 3720 } 3721 } 3722 3723 set_complete(phase); 3724 return rawmem; 3725} 3726 3727 3728#ifdef ASSERT 3729bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3730 if (is_complete()) 3731 return true; // stores could be anything at this point 3732 assert(allocation() != NULL, "must be present"); 3733 intptr_t last_off = allocation()->minimum_header_size(); 3734 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3735 Node* st = in(i); 3736 intptr_t st_off = get_store_offset(st, phase); 3737 if (st_off < 0) continue; // ignore dead garbage 3738 if (last_off > st_off) { 3739 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3740 this->dump(2); 3741 assert(false, "ascending store offsets"); 3742 return false; 3743 } 3744 last_off = st_off + st->as_Store()->memory_size(); 3745 } 3746 return true; 3747} 3748#endif //ASSERT 3749 3750 3751 3752 3753//============================MergeMemNode===================================== 3754// 3755// SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3756// contributing store or call operations. Each contributor provides the memory 3757// state for a particular "alias type" (see Compile::alias_type). For example, 3758// if a MergeMem has an input X for alias category #6, then any memory reference 3759// to alias category #6 may use X as its memory state input, as an exact equivalent 3760// to using the MergeMem as a whole. 3761// Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3762// 3763// (Here, the <N> notation gives the index of the relevant adr_type.) 3764// 3765// In one special case (and more cases in the future), alias categories overlap. 3766// The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3767// states. Therefore, if a MergeMem has only one contributing input W for Bot, 3768// it is exactly equivalent to that state W: 3769// MergeMem(<Bot>: W) <==> W 3770// 3771// Usually, the merge has more than one input. In that case, where inputs 3772// overlap (i.e., one is Bot), the narrower alias type determines the memory 3773// state for that type, and the wider alias type (Bot) fills in everywhere else: 3774// Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3775// Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3776// 3777// A merge can take a "wide" memory state as one of its narrow inputs. 3778// This simply means that the merge observes out only the relevant parts of 3779// the wide input. That is, wide memory states arriving at narrow merge inputs 3780// are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3781// 3782// These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3783// and that memory slices "leak through": 3784// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3785// 3786// But, in such a cascade, repeated memory slices can "block the leak": 3787// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3788// 3789// In the last example, Y is not part of the combined memory state of the 3790// outermost MergeMem. The system must, of course, prevent unschedulable 3791// memory states from arising, so you can be sure that the state Y is somehow 3792// a precursor to state Y'. 3793// 3794// 3795// REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3796// of each MergeMemNode array are exactly the numerical alias indexes, including 3797// but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3798// Compile::alias_type (and kin) produce and manage these indexes. 3799// 3800// By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3801// (Note that this provides quick access to the top node inside MergeMem methods, 3802// without the need to reach out via TLS to Compile::current.) 3803// 3804// As a consequence of what was just described, a MergeMem that represents a full 3805// memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3806// containing all alias categories. 3807// 3808// MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3809// 3810// All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3811// a memory state for the alias type <N>, or else the top node, meaning that 3812// there is no particular input for that alias type. Note that the length of 3813// a MergeMem is variable, and may be extended at any time to accommodate new 3814// memory states at larger alias indexes. When merges grow, they are of course 3815// filled with "top" in the unused in() positions. 3816// 3817// This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3818// (Top was chosen because it works smoothly with passes like GCM.) 3819// 3820// For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3821// the type of random VM bits like TLS references.) Since it is always the 3822// first non-Bot memory slice, some low-level loops use it to initialize an 3823// index variable: for (i = AliasIdxRaw; i < req(); i++). 3824// 3825// 3826// ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3827// the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3828// the memory state for alias type <N>, or (if there is no particular slice at <N>, 3829// it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3830// or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3831// MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3832// 3833// %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3834// really that different from the other memory inputs. An abbreviation called 3835// "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3836// 3837// 3838// PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3839// partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3840// that "emerges though" the base memory will be marked as excluding the alias types 3841// of the other (narrow-memory) copies which "emerged through" the narrow edges: 3842// 3843// Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3844// ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3845// 3846// This strange "subtraction" effect is necessary to ensure IGVN convergence. 3847// (It is currently unimplemented.) As you can see, the resulting merge is 3848// actually a disjoint union of memory states, rather than an overlay. 3849// 3850 3851//------------------------------MergeMemNode----------------------------------- 3852Node* MergeMemNode::make_empty_memory() { 3853 Node* empty_memory = (Node*) Compile::current()->top(); 3854 assert(empty_memory->is_top(), "correct sentinel identity"); 3855 return empty_memory; 3856} 3857 3858MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3859 init_class_id(Class_MergeMem); 3860 // all inputs are nullified in Node::Node(int) 3861 // set_input(0, NULL); // no control input 3862 3863 // Initialize the edges uniformly to top, for starters. 3864 Node* empty_mem = make_empty_memory(); 3865 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3866 init_req(i,empty_mem); 3867 } 3868 assert(empty_memory() == empty_mem, ""); 3869 3870 if( new_base != NULL && new_base->is_MergeMem() ) { 3871 MergeMemNode* mdef = new_base->as_MergeMem(); 3872 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3873 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3874 mms.set_memory(mms.memory2()); 3875 } 3876 assert(base_memory() == mdef->base_memory(), ""); 3877 } else { 3878 set_base_memory(new_base); 3879 } 3880} 3881 3882// Make a new, untransformed MergeMem with the same base as 'mem'. 3883// If mem is itself a MergeMem, populate the result with the same edges. 3884MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3885 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3886} 3887 3888//------------------------------cmp-------------------------------------------- 3889uint MergeMemNode::hash() const { return NO_HASH; } 3890uint MergeMemNode::cmp( const Node &n ) const { 3891 return (&n == this); // Always fail except on self 3892} 3893 3894//------------------------------Identity--------------------------------------- 3895Node* MergeMemNode::Identity(PhaseTransform *phase) { 3896 // Identity if this merge point does not record any interesting memory 3897 // disambiguations. 3898 Node* base_mem = base_memory(); 3899 Node* empty_mem = empty_memory(); 3900 if (base_mem != empty_mem) { // Memory path is not dead? 3901 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3902 Node* mem = in(i); 3903 if (mem != empty_mem && mem != base_mem) { 3904 return this; // Many memory splits; no change 3905 } 3906 } 3907 } 3908 return base_mem; // No memory splits; ID on the one true input 3909} 3910 3911//------------------------------Ideal------------------------------------------ 3912// This method is invoked recursively on chains of MergeMem nodes 3913Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3914 // Remove chain'd MergeMems 3915 // 3916 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3917 // relative to the "in(Bot)". Since we are patching both at the same time, 3918 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3919 // but rewrite each "in(i)" relative to the new "in(Bot)". 3920 Node *progress = NULL; 3921 3922 3923 Node* old_base = base_memory(); 3924 Node* empty_mem = empty_memory(); 3925 if (old_base == empty_mem) 3926 return NULL; // Dead memory path. 3927 3928 MergeMemNode* old_mbase; 3929 if (old_base != NULL && old_base->is_MergeMem()) 3930 old_mbase = old_base->as_MergeMem(); 3931 else 3932 old_mbase = NULL; 3933 Node* new_base = old_base; 3934 3935 // simplify stacked MergeMems in base memory 3936 if (old_mbase) new_base = old_mbase->base_memory(); 3937 3938 // the base memory might contribute new slices beyond my req() 3939 if (old_mbase) grow_to_match(old_mbase); 3940 3941 // Look carefully at the base node if it is a phi. 3942 PhiNode* phi_base; 3943 if (new_base != NULL && new_base->is_Phi()) 3944 phi_base = new_base->as_Phi(); 3945 else 3946 phi_base = NULL; 3947 3948 Node* phi_reg = NULL; 3949 uint phi_len = (uint)-1; 3950 if (phi_base != NULL && !phi_base->is_copy()) { 3951 // do not examine phi if degraded to a copy 3952 phi_reg = phi_base->region(); 3953 phi_len = phi_base->req(); 3954 // see if the phi is unfinished 3955 for (uint i = 1; i < phi_len; i++) { 3956 if (phi_base->in(i) == NULL) { 3957 // incomplete phi; do not look at it yet! 3958 phi_reg = NULL; 3959 phi_len = (uint)-1; 3960 break; 3961 } 3962 } 3963 } 3964 3965 // Note: We do not call verify_sparse on entry, because inputs 3966 // can normalize to the base_memory via subsume_node or similar 3967 // mechanisms. This method repairs that damage. 3968 3969 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3970 3971 // Look at each slice. 3972 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3973 Node* old_in = in(i); 3974 // calculate the old memory value 3975 Node* old_mem = old_in; 3976 if (old_mem == empty_mem) old_mem = old_base; 3977 assert(old_mem == memory_at(i), ""); 3978 3979 // maybe update (reslice) the old memory value 3980 3981 // simplify stacked MergeMems 3982 Node* new_mem = old_mem; 3983 MergeMemNode* old_mmem; 3984 if (old_mem != NULL && old_mem->is_MergeMem()) 3985 old_mmem = old_mem->as_MergeMem(); 3986 else 3987 old_mmem = NULL; 3988 if (old_mmem == this) { 3989 // This can happen if loops break up and safepoints disappear. 3990 // A merge of BotPtr (default) with a RawPtr memory derived from a 3991 // safepoint can be rewritten to a merge of the same BotPtr with 3992 // the BotPtr phi coming into the loop. If that phi disappears 3993 // also, we can end up with a self-loop of the mergemem. 3994 // In general, if loops degenerate and memory effects disappear, 3995 // a mergemem can be left looking at itself. This simply means 3996 // that the mergemem's default should be used, since there is 3997 // no longer any apparent effect on this slice. 3998 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3999 // from start. Update the input to TOP. 4000 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4001 } 4002 else if (old_mmem != NULL) { 4003 new_mem = old_mmem->memory_at(i); 4004 } 4005 // else preceding memory was not a MergeMem 4006 4007 // replace equivalent phis (unfortunately, they do not GVN together) 4008 if (new_mem != NULL && new_mem != new_base && 4009 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4010 if (new_mem->is_Phi()) { 4011 PhiNode* phi_mem = new_mem->as_Phi(); 4012 for (uint i = 1; i < phi_len; i++) { 4013 if (phi_base->in(i) != phi_mem->in(i)) { 4014 phi_mem = NULL; 4015 break; 4016 } 4017 } 4018 if (phi_mem != NULL) { 4019 // equivalent phi nodes; revert to the def 4020 new_mem = new_base; 4021 } 4022 } 4023 } 4024 4025 // maybe store down a new value 4026 Node* new_in = new_mem; 4027 if (new_in == new_base) new_in = empty_mem; 4028 4029 if (new_in != old_in) { 4030 // Warning: Do not combine this "if" with the previous "if" 4031 // A memory slice might have be be rewritten even if it is semantically 4032 // unchanged, if the base_memory value has changed. 4033 set_req(i, new_in); 4034 progress = this; // Report progress 4035 } 4036 } 4037 4038 if (new_base != old_base) { 4039 set_req(Compile::AliasIdxBot, new_base); 4040 // Don't use set_base_memory(new_base), because we need to update du. 4041 assert(base_memory() == new_base, ""); 4042 progress = this; 4043 } 4044 4045 if( base_memory() == this ) { 4046 // a self cycle indicates this memory path is dead 4047 set_req(Compile::AliasIdxBot, empty_mem); 4048 } 4049 4050 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4051 // Recursion must occur after the self cycle check above 4052 if( base_memory()->is_MergeMem() ) { 4053 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4054 Node *m = phase->transform(new_mbase); // Rollup any cycles 4055 if( m != NULL && (m->is_top() || 4056 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4057 // propagate rollup of dead cycle to self 4058 set_req(Compile::AliasIdxBot, empty_mem); 4059 } 4060 } 4061 4062 if( base_memory() == empty_mem ) { 4063 progress = this; 4064 // Cut inputs during Parse phase only. 4065 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4066 if( !can_reshape ) { 4067 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4068 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4069 } 4070 } 4071 } 4072 4073 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4074 // Check if PhiNode::Ideal's "Split phis through memory merges" 4075 // transform should be attempted. Look for this->phi->this cycle. 4076 uint merge_width = req(); 4077 if (merge_width > Compile::AliasIdxRaw) { 4078 PhiNode* phi = base_memory()->as_Phi(); 4079 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4080 if (phi->in(i) == this) { 4081 phase->is_IterGVN()->_worklist.push(phi); 4082 break; 4083 } 4084 } 4085 } 4086 } 4087 4088 assert(progress || verify_sparse(), "please, no dups of base"); 4089 return progress; 4090} 4091 4092//-------------------------set_base_memory------------------------------------- 4093void MergeMemNode::set_base_memory(Node *new_base) { 4094 Node* empty_mem = empty_memory(); 4095 set_req(Compile::AliasIdxBot, new_base); 4096 assert(memory_at(req()) == new_base, "must set default memory"); 4097 // Clear out other occurrences of new_base: 4098 if (new_base != empty_mem) { 4099 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4100 if (in(i) == new_base) set_req(i, empty_mem); 4101 } 4102 } 4103} 4104 4105//------------------------------out_RegMask------------------------------------ 4106const RegMask &MergeMemNode::out_RegMask() const { 4107 return RegMask::Empty; 4108} 4109 4110//------------------------------dump_spec-------------------------------------- 4111#ifndef PRODUCT 4112void MergeMemNode::dump_spec(outputStream *st) const { 4113 st->print(" {"); 4114 Node* base_mem = base_memory(); 4115 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4116 Node* mem = memory_at(i); 4117 if (mem == base_mem) { st->print(" -"); continue; } 4118 st->print( " N%d:", mem->_idx ); 4119 Compile::current()->get_adr_type(i)->dump_on(st); 4120 } 4121 st->print(" }"); 4122} 4123#endif // !PRODUCT 4124 4125 4126#ifdef ASSERT 4127static bool might_be_same(Node* a, Node* b) { 4128 if (a == b) return true; 4129 if (!(a->is_Phi() || b->is_Phi())) return false; 4130 // phis shift around during optimization 4131 return true; // pretty stupid... 4132} 4133 4134// verify a narrow slice (either incoming or outgoing) 4135static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4136 if (!VerifyAliases) return; // don't bother to verify unless requested 4137 if (is_error_reported()) return; // muzzle asserts when debugging an error 4138 if (Node::in_dump()) return; // muzzle asserts when printing 4139 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4140 assert(n != NULL, ""); 4141 // Elide intervening MergeMem's 4142 while (n->is_MergeMem()) { 4143 n = n->as_MergeMem()->memory_at(alias_idx); 4144 } 4145 Compile* C = Compile::current(); 4146 const TypePtr* n_adr_type = n->adr_type(); 4147 if (n == m->empty_memory()) { 4148 // Implicit copy of base_memory() 4149 } else if (n_adr_type != TypePtr::BOTTOM) { 4150 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4151 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4152 } else { 4153 // A few places like make_runtime_call "know" that VM calls are narrow, 4154 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4155 bool expected_wide_mem = false; 4156 if (n == m->base_memory()) { 4157 expected_wide_mem = true; 4158 } else if (alias_idx == Compile::AliasIdxRaw || 4159 n == m->memory_at(Compile::AliasIdxRaw)) { 4160 expected_wide_mem = true; 4161 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4162 // memory can "leak through" calls on channels that 4163 // are write-once. Allow this also. 4164 expected_wide_mem = true; 4165 } 4166 assert(expected_wide_mem, "expected narrow slice replacement"); 4167 } 4168} 4169#else // !ASSERT 4170#define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 4171#endif 4172 4173 4174//-----------------------------memory_at--------------------------------------- 4175Node* MergeMemNode::memory_at(uint alias_idx) const { 4176 assert(alias_idx >= Compile::AliasIdxRaw || 4177 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4178 "must avoid base_memory and AliasIdxTop"); 4179 4180 // Otherwise, it is a narrow slice. 4181 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4182 Compile *C = Compile::current(); 4183 if (is_empty_memory(n)) { 4184 // the array is sparse; empty slots are the "top" node 4185 n = base_memory(); 4186 assert(Node::in_dump() 4187 || n == NULL || n->bottom_type() == Type::TOP 4188 || n->adr_type() == NULL // address is TOP 4189 || n->adr_type() == TypePtr::BOTTOM 4190 || n->adr_type() == TypeRawPtr::BOTTOM 4191 || Compile::current()->AliasLevel() == 0, 4192 "must be a wide memory"); 4193 // AliasLevel == 0 if we are organizing the memory states manually. 4194 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4195 } else { 4196 // make sure the stored slice is sane 4197 #ifdef ASSERT 4198 if (is_error_reported() || Node::in_dump()) { 4199 } else if (might_be_same(n, base_memory())) { 4200 // Give it a pass: It is a mostly harmless repetition of the base. 4201 // This can arise normally from node subsumption during optimization. 4202 } else { 4203 verify_memory_slice(this, alias_idx, n); 4204 } 4205 #endif 4206 } 4207 return n; 4208} 4209 4210//---------------------------set_memory_at------------------------------------- 4211void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4212 verify_memory_slice(this, alias_idx, n); 4213 Node* empty_mem = empty_memory(); 4214 if (n == base_memory()) n = empty_mem; // collapse default 4215 uint need_req = alias_idx+1; 4216 if (req() < need_req) { 4217 if (n == empty_mem) return; // already the default, so do not grow me 4218 // grow the sparse array 4219 do { 4220 add_req(empty_mem); 4221 } while (req() < need_req); 4222 } 4223 set_req( alias_idx, n ); 4224} 4225 4226 4227 4228//--------------------------iteration_setup------------------------------------ 4229void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4230 if (other != NULL) { 4231 grow_to_match(other); 4232 // invariant: the finite support of mm2 is within mm->req() 4233 #ifdef ASSERT 4234 for (uint i = req(); i < other->req(); i++) { 4235 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4236 } 4237 #endif 4238 } 4239 // Replace spurious copies of base_memory by top. 4240 Node* base_mem = base_memory(); 4241 if (base_mem != NULL && !base_mem->is_top()) { 4242 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4243 if (in(i) == base_mem) 4244 set_req(i, empty_memory()); 4245 } 4246 } 4247} 4248 4249//---------------------------grow_to_match------------------------------------- 4250void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4251 Node* empty_mem = empty_memory(); 4252 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4253 // look for the finite support of the other memory 4254 for (uint i = other->req(); --i >= req(); ) { 4255 if (other->in(i) != empty_mem) { 4256 uint new_len = i+1; 4257 while (req() < new_len) add_req(empty_mem); 4258 break; 4259 } 4260 } 4261} 4262 4263//---------------------------verify_sparse------------------------------------- 4264#ifndef PRODUCT 4265bool MergeMemNode::verify_sparse() const { 4266 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4267 Node* base_mem = base_memory(); 4268 // The following can happen in degenerate cases, since empty==top. 4269 if (is_empty_memory(base_mem)) return true; 4270 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4271 assert(in(i) != NULL, "sane slice"); 4272 if (in(i) == base_mem) return false; // should have been the sentinel value! 4273 } 4274 return true; 4275} 4276 4277bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4278 Node* n; 4279 n = mm->in(idx); 4280 if (mem == n) return true; // might be empty_memory() 4281 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4282 if (mem == n) return true; 4283 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4284 if (mem == n) return true; 4285 if (n == NULL) break; 4286 } 4287 return false; 4288} 4289#endif // !PRODUCT 4290