memnode.cpp revision 113:ba764ed4b6f2
1/* 2 * Copyright 1997-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25// Portions of code courtesy of Clifford Click 26 27// Optimization - Graph Style 28 29#include "incls/_precompiled.incl" 30#include "incls/_memnode.cpp.incl" 31 32static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 33 34//============================================================================= 35uint MemNode::size_of() const { return sizeof(*this); } 36 37const TypePtr *MemNode::adr_type() const { 38 Node* adr = in(Address); 39 const TypePtr* cross_check = NULL; 40 DEBUG_ONLY(cross_check = _adr_type); 41 return calculate_adr_type(adr->bottom_type(), cross_check); 42} 43 44#ifndef PRODUCT 45void MemNode::dump_spec(outputStream *st) const { 46 if (in(Address) == NULL) return; // node is dead 47#ifndef ASSERT 48 // fake the missing field 49 const TypePtr* _adr_type = NULL; 50 if (in(Address) != NULL) 51 _adr_type = in(Address)->bottom_type()->isa_ptr(); 52#endif 53 dump_adr_type(this, _adr_type, st); 54 55 Compile* C = Compile::current(); 56 if( C->alias_type(_adr_type)->is_volatile() ) 57 st->print(" Volatile!"); 58} 59 60void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 61 st->print(" @"); 62 if (adr_type == NULL) { 63 st->print("NULL"); 64 } else { 65 adr_type->dump_on(st); 66 Compile* C = Compile::current(); 67 Compile::AliasType* atp = NULL; 68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 69 if (atp == NULL) 70 st->print(", idx=?\?;"); 71 else if (atp->index() == Compile::AliasIdxBot) 72 st->print(", idx=Bot;"); 73 else if (atp->index() == Compile::AliasIdxTop) 74 st->print(", idx=Top;"); 75 else if (atp->index() == Compile::AliasIdxRaw) 76 st->print(", idx=Raw;"); 77 else { 78 ciField* field = atp->field(); 79 if (field) { 80 st->print(", name="); 81 field->print_name_on(st); 82 } 83 st->print(", idx=%d;", atp->index()); 84 } 85 } 86} 87 88extern void print_alias_types(); 89 90#endif 91 92Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 93 const TypeOopPtr *tinst = t_adr->isa_oopptr(); 94 if (tinst == NULL || !tinst->is_instance_field()) 95 return mchain; // don't try to optimize non-instance types 96 uint instance_id = tinst->instance_id(); 97 Node *prev = NULL; 98 Node *result = mchain; 99 while (prev != result) { 100 prev = result; 101 // skip over a call which does not affect this memory slice 102 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 103 Node *proj_in = result->in(0); 104 if (proj_in->is_Call()) { 105 CallNode *call = proj_in->as_Call(); 106 if (!call->may_modify(t_adr, phase)) { 107 result = call->in(TypeFunc::Memory); 108 } 109 } else if (proj_in->is_Initialize()) { 110 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 111 // Stop if this is the initialization for the object instance which 112 // which contains this memory slice, otherwise skip over it. 113 if (alloc != NULL && alloc->_idx != instance_id) { 114 result = proj_in->in(TypeFunc::Memory); 115 } 116 } else if (proj_in->is_MemBar()) { 117 result = proj_in->in(TypeFunc::Memory); 118 } 119 } else if (result->is_MergeMem()) { 120 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); 121 } 122 } 123 return result; 124} 125 126Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { 127 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); 128 bool is_instance = (t_oop != NULL) && t_oop->is_instance_field(); 129 PhaseIterGVN *igvn = phase->is_IterGVN(); 130 Node *result = mchain; 131 result = optimize_simple_memory_chain(result, t_adr, phase); 132 if (is_instance && igvn != NULL && result->is_Phi()) { 133 PhiNode *mphi = result->as_Phi(); 134 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 135 const TypePtr *t = mphi->adr_type(); 136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM) { 137 // clone the Phi with our address type 138 result = mphi->split_out_instance(t_adr, igvn); 139 } else { 140 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 141 } 142 } 143 return result; 144} 145 146static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 147 uint alias_idx = phase->C->get_alias_index(tp); 148 Node *mem = mmem; 149#ifdef ASSERT 150 { 151 // Check that current type is consistent with the alias index used during graph construction 152 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 153 bool consistent = adr_check == NULL || adr_check->empty() || 154 phase->C->must_alias(adr_check, alias_idx ); 155 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 156 if( !consistent && adr_check != NULL && !adr_check->empty() && 157 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 158 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 159 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 160 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 161 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 162 // don't assert if it is dead code. 163 consistent = true; 164 } 165 if( !consistent ) { 166 st->print("alias_idx==%d, adr_check==", alias_idx); 167 if( adr_check == NULL ) { 168 st->print("NULL"); 169 } else { 170 adr_check->dump(); 171 } 172 st->cr(); 173 print_alias_types(); 174 assert(consistent, "adr_check must match alias idx"); 175 } 176 } 177#endif 178 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally 179 // means an array I have not precisely typed yet. Do not do any 180 // alias stuff with it any time soon. 181 const TypeOopPtr *tinst = tp->isa_oopptr(); 182 if( tp->base() != Type::AnyPtr && 183 !(tinst && 184 tinst->klass()->is_java_lang_Object() && 185 tinst->offset() == Type::OffsetBot) ) { 186 // compress paths and change unreachable cycles to TOP 187 // If not, we can update the input infinitely along a MergeMem cycle 188 // Equivalent code in PhiNode::Ideal 189 Node* m = phase->transform(mmem); 190 // If tranformed to a MergeMem, get the desired slice 191 // Otherwise the returned node represents memory for every slice 192 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 193 // Update input if it is progress over what we have now 194 } 195 return mem; 196} 197 198//--------------------------Ideal_common--------------------------------------- 199// Look for degenerate control and memory inputs. Bypass MergeMem inputs. 200// Unhook non-raw memories from complete (macro-expanded) initializations. 201Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 202 // If our control input is a dead region, kill all below the region 203 Node *ctl = in(MemNode::Control); 204 if (ctl && remove_dead_region(phase, can_reshape)) 205 return this; 206 207 // Ignore if memory is dead, or self-loop 208 Node *mem = in(MemNode::Memory); 209 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL 210 assert( mem != this, "dead loop in MemNode::Ideal" ); 211 212 Node *address = in(MemNode::Address); 213 const Type *t_adr = phase->type( address ); 214 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL 215 216 // Avoid independent memory operations 217 Node* old_mem = mem; 218 219 // The code which unhooks non-raw memories from complete (macro-expanded) 220 // initializations was removed. After macro-expansion all stores catched 221 // by Initialize node became raw stores and there is no information 222 // which memory slices they modify. So it is unsafe to move any memory 223 // operation above these stores. Also in most cases hooked non-raw memories 224 // were already unhooked by using information from detect_ptr_independence() 225 // and find_previous_store(). 226 227 if (mem->is_MergeMem()) { 228 MergeMemNode* mmem = mem->as_MergeMem(); 229 const TypePtr *tp = t_adr->is_ptr(); 230 231 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 232 } 233 234 if (mem != old_mem) { 235 set_req(MemNode::Memory, mem); 236 return this; 237 } 238 239 // let the subclass continue analyzing... 240 return NULL; 241} 242 243// Helper function for proving some simple control dominations. 244// Attempt to prove that control input 'dom' dominates (or equals) 'sub'. 245// Already assumes that 'dom' is available at 'sub', and that 'sub' 246// is not a constant (dominated by the method's StartNode). 247// Used by MemNode::find_previous_store to prove that the 248// control input of a memory operation predates (dominates) 249// an allocation it wants to look past. 250bool MemNode::detect_dominating_control(Node* dom, Node* sub) { 251 if (dom == NULL) return false; 252 if (dom->is_Proj()) dom = dom->in(0); 253 if (dom->is_Start()) return true; // anything inside the method 254 if (dom->is_Root()) return true; // dom 'controls' a constant 255 int cnt = 20; // detect cycle or too much effort 256 while (sub != NULL) { // walk 'sub' up the chain to 'dom' 257 if (--cnt < 0) return false; // in a cycle or too complex 258 if (sub == dom) return true; 259 if (sub->is_Start()) return false; 260 if (sub->is_Root()) return false; 261 Node* up = sub->in(0); 262 if (sub == up && sub->is_Region()) { 263 for (uint i = 1; i < sub->req(); i++) { 264 Node* in = sub->in(i); 265 if (in != NULL && !in->is_top() && in != sub) { 266 up = in; break; // take any path on the way up to 'dom' 267 } 268 } 269 } 270 if (sub == up) return false; // some kind of tight cycle 271 sub = up; 272 } 273 return false; 274} 275 276//---------------------detect_ptr_independence--------------------------------- 277// Used by MemNode::find_previous_store to prove that two base 278// pointers are never equal. 279// The pointers are accompanied by their associated allocations, 280// if any, which have been previously discovered by the caller. 281bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 282 Node* p2, AllocateNode* a2, 283 PhaseTransform* phase) { 284 // Attempt to prove that these two pointers cannot be aliased. 285 // They may both manifestly be allocations, and they should differ. 286 // Or, if they are not both allocations, they can be distinct constants. 287 // Otherwise, one is an allocation and the other a pre-existing value. 288 if (a1 == NULL && a2 == NULL) { // neither an allocation 289 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 290 } else if (a1 != NULL && a2 != NULL) { // both allocations 291 return (a1 != a2); 292 } else if (a1 != NULL) { // one allocation a1 293 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 294 return detect_dominating_control(p2->in(0), a1->in(0)); 295 } else { //(a2 != NULL) // one allocation a2 296 return detect_dominating_control(p1->in(0), a2->in(0)); 297 } 298 return false; 299} 300 301 302// The logic for reordering loads and stores uses four steps: 303// (a) Walk carefully past stores and initializations which we 304// can prove are independent of this load. 305// (b) Observe that the next memory state makes an exact match 306// with self (load or store), and locate the relevant store. 307// (c) Ensure that, if we were to wire self directly to the store, 308// the optimizer would fold it up somehow. 309// (d) Do the rewiring, and return, depending on some other part of 310// the optimizer to fold up the load. 311// This routine handles steps (a) and (b). Steps (c) and (d) are 312// specific to loads and stores, so they are handled by the callers. 313// (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 314// 315Node* MemNode::find_previous_store(PhaseTransform* phase) { 316 Node* ctrl = in(MemNode::Control); 317 Node* adr = in(MemNode::Address); 318 intptr_t offset = 0; 319 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 320 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 321 322 if (offset == Type::OffsetBot) 323 return NULL; // cannot unalias unless there are precise offsets 324 325 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 326 327 intptr_t size_in_bytes = memory_size(); 328 329 Node* mem = in(MemNode::Memory); // start searching here... 330 331 int cnt = 50; // Cycle limiter 332 for (;;) { // While we can dance past unrelated stores... 333 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 334 335 if (mem->is_Store()) { 336 Node* st_adr = mem->in(MemNode::Address); 337 intptr_t st_offset = 0; 338 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 339 if (st_base == NULL) 340 break; // inscrutable pointer 341 if (st_offset != offset && st_offset != Type::OffsetBot) { 342 const int MAX_STORE = BytesPerLong; 343 if (st_offset >= offset + size_in_bytes || 344 st_offset <= offset - MAX_STORE || 345 st_offset <= offset - mem->as_Store()->memory_size()) { 346 // Success: The offsets are provably independent. 347 // (You may ask, why not just test st_offset != offset and be done? 348 // The answer is that stores of different sizes can co-exist 349 // in the same sequence of RawMem effects. We sometimes initialize 350 // a whole 'tile' of array elements with a single jint or jlong.) 351 mem = mem->in(MemNode::Memory); 352 continue; // (a) advance through independent store memory 353 } 354 } 355 if (st_base != base && 356 detect_ptr_independence(base, alloc, 357 st_base, 358 AllocateNode::Ideal_allocation(st_base, phase), 359 phase)) { 360 // Success: The bases are provably independent. 361 mem = mem->in(MemNode::Memory); 362 continue; // (a) advance through independent store memory 363 } 364 365 // (b) At this point, if the bases or offsets do not agree, we lose, 366 // since we have not managed to prove 'this' and 'mem' independent. 367 if (st_base == base && st_offset == offset) { 368 return mem; // let caller handle steps (c), (d) 369 } 370 371 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 372 InitializeNode* st_init = mem->in(0)->as_Initialize(); 373 AllocateNode* st_alloc = st_init->allocation(); 374 if (st_alloc == NULL) 375 break; // something degenerated 376 bool known_identical = false; 377 bool known_independent = false; 378 if (alloc == st_alloc) 379 known_identical = true; 380 else if (alloc != NULL) 381 known_independent = true; 382 else if (ctrl != NULL && 383 detect_dominating_control(ctrl, st_alloc->in(0))) 384 known_independent = true; 385 386 if (known_independent) { 387 // The bases are provably independent: Either they are 388 // manifestly distinct allocations, or else the control 389 // of this load dominates the store's allocation. 390 int alias_idx = phase->C->get_alias_index(adr_type()); 391 if (alias_idx == Compile::AliasIdxRaw) { 392 mem = st_alloc->in(TypeFunc::Memory); 393 } else { 394 mem = st_init->memory(alias_idx); 395 } 396 continue; // (a) advance through independent store memory 397 } 398 399 // (b) at this point, if we are not looking at a store initializing 400 // the same allocation we are loading from, we lose. 401 if (known_identical) { 402 // From caller, can_see_stored_value will consult find_captured_store. 403 return mem; // let caller handle steps (c), (d) 404 } 405 406 } else if (addr_t != NULL && addr_t->is_instance_field()) { 407 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 408 if (mem->is_Proj() && mem->in(0)->is_Call()) { 409 CallNode *call = mem->in(0)->as_Call(); 410 if (!call->may_modify(addr_t, phase)) { 411 mem = call->in(TypeFunc::Memory); 412 continue; // (a) advance through independent call memory 413 } 414 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 415 mem = mem->in(0)->in(TypeFunc::Memory); 416 continue; // (a) advance through independent MemBar memory 417 } else if (mem->is_MergeMem()) { 418 int alias_idx = phase->C->get_alias_index(adr_type()); 419 mem = mem->as_MergeMem()->memory_at(alias_idx); 420 continue; // (a) advance through independent MergeMem memory 421 } 422 } 423 424 // Unless there is an explicit 'continue', we must bail out here, 425 // because 'mem' is an inscrutable memory state (e.g., a call). 426 break; 427 } 428 429 return NULL; // bail out 430} 431 432//----------------------calculate_adr_type------------------------------------- 433// Helper function. Notices when the given type of address hits top or bottom. 434// Also, asserts a cross-check of the type against the expected address type. 435const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 436 if (t == Type::TOP) return NULL; // does not touch memory any more? 437 #ifdef PRODUCT 438 cross_check = NULL; 439 #else 440 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 441 #endif 442 const TypePtr* tp = t->isa_ptr(); 443 if (tp == NULL) { 444 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 445 return TypePtr::BOTTOM; // touches lots of memory 446 } else { 447 #ifdef ASSERT 448 // %%%% [phh] We don't check the alias index if cross_check is 449 // TypeRawPtr::BOTTOM. Needs to be investigated. 450 if (cross_check != NULL && 451 cross_check != TypePtr::BOTTOM && 452 cross_check != TypeRawPtr::BOTTOM) { 453 // Recheck the alias index, to see if it has changed (due to a bug). 454 Compile* C = Compile::current(); 455 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 456 "must stay in the original alias category"); 457 // The type of the address must be contained in the adr_type, 458 // disregarding "null"-ness. 459 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 460 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 461 assert(cross_check->meet(tp_notnull) == cross_check, 462 "real address must not escape from expected memory type"); 463 } 464 #endif 465 return tp; 466 } 467} 468 469//------------------------adr_phi_is_loop_invariant---------------------------- 470// A helper function for Ideal_DU_postCCP to check if a Phi in a counted 471// loop is loop invariant. Make a quick traversal of Phi and associated 472// CastPP nodes, looking to see if they are a closed group within the loop. 473bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 474 // The idea is that the phi-nest must boil down to only CastPP nodes 475 // with the same data. This implies that any path into the loop already 476 // includes such a CastPP, and so the original cast, whatever its input, 477 // must be covered by an equivalent cast, with an earlier control input. 478 ResourceMark rm; 479 480 // The loop entry input of the phi should be the unique dominating 481 // node for every Phi/CastPP in the loop. 482 Unique_Node_List closure; 483 closure.push(adr_phi->in(LoopNode::EntryControl)); 484 485 // Add the phi node and the cast to the worklist. 486 Unique_Node_List worklist; 487 worklist.push(adr_phi); 488 if( cast != NULL ){ 489 if( !cast->is_ConstraintCast() ) return false; 490 worklist.push(cast); 491 } 492 493 // Begin recursive walk of phi nodes. 494 while( worklist.size() ){ 495 // Take a node off the worklist 496 Node *n = worklist.pop(); 497 if( !closure.member(n) ){ 498 // Add it to the closure. 499 closure.push(n); 500 // Make a sanity check to ensure we don't waste too much time here. 501 if( closure.size() > 20) return false; 502 // This node is OK if: 503 // - it is a cast of an identical value 504 // - or it is a phi node (then we add its inputs to the worklist) 505 // Otherwise, the node is not OK, and we presume the cast is not invariant 506 if( n->is_ConstraintCast() ){ 507 worklist.push(n->in(1)); 508 } else if( n->is_Phi() ) { 509 for( uint i = 1; i < n->req(); i++ ) { 510 worklist.push(n->in(i)); 511 } 512 } else { 513 return false; 514 } 515 } 516 } 517 518 // Quit when the worklist is empty, and we've found no offending nodes. 519 return true; 520} 521 522//------------------------------Ideal_DU_postCCP------------------------------- 523// Find any cast-away of null-ness and keep its control. Null cast-aways are 524// going away in this pass and we need to make this memory op depend on the 525// gating null check. 526 527// I tried to leave the CastPP's in. This makes the graph more accurate in 528// some sense; we get to keep around the knowledge that an oop is not-null 529// after some test. Alas, the CastPP's interfere with GVN (some values are 530// the regular oop, some are the CastPP of the oop, all merge at Phi's which 531// cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 532// some of the more trivial cases in the optimizer. Removing more useless 533// Phi's started allowing Loads to illegally float above null checks. I gave 534// up on this approach. CNC 10/20/2000 535Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 536 Node *ctr = in(MemNode::Control); 537 Node *mem = in(MemNode::Memory); 538 Node *adr = in(MemNode::Address); 539 Node *skipped_cast = NULL; 540 // Need a null check? Regular static accesses do not because they are 541 // from constant addresses. Array ops are gated by the range check (which 542 // always includes a NULL check). Just check field ops. 543 if( !ctr ) { 544 // Scan upwards for the highest location we can place this memory op. 545 while( true ) { 546 switch( adr->Opcode() ) { 547 548 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 549 adr = adr->in(AddPNode::Base); 550 continue; 551 552 case Op_DecodeN: // No change to NULL-ness, so peek thru 553 adr = adr->in(1); 554 continue; 555 556 case Op_CastPP: 557 // If the CastPP is useless, just peek on through it. 558 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 559 // Remember the cast that we've peeked though. If we peek 560 // through more than one, then we end up remembering the highest 561 // one, that is, if in a loop, the one closest to the top. 562 skipped_cast = adr; 563 adr = adr->in(1); 564 continue; 565 } 566 // CastPP is going away in this pass! We need this memory op to be 567 // control-dependent on the test that is guarding the CastPP. 568 ccp->hash_delete(this); 569 set_req(MemNode::Control, adr->in(0)); 570 ccp->hash_insert(this); 571 return this; 572 573 case Op_Phi: 574 // Attempt to float above a Phi to some dominating point. 575 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 576 // If we've already peeked through a Cast (which could have set the 577 // control), we can't float above a Phi, because the skipped Cast 578 // may not be loop invariant. 579 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 580 adr = adr->in(1); 581 continue; 582 } 583 } 584 585 // Intentional fallthrough! 586 587 // No obvious dominating point. The mem op is pinned below the Phi 588 // by the Phi itself. If the Phi goes away (no true value is merged) 589 // then the mem op can float, but not indefinitely. It must be pinned 590 // behind the controls leading to the Phi. 591 case Op_CheckCastPP: 592 // These usually stick around to change address type, however a 593 // useless one can be elided and we still need to pick up a control edge 594 if (adr->in(0) == NULL) { 595 // This CheckCastPP node has NO control and is likely useless. But we 596 // need check further up the ancestor chain for a control input to keep 597 // the node in place. 4959717. 598 skipped_cast = adr; 599 adr = adr->in(1); 600 continue; 601 } 602 ccp->hash_delete(this); 603 set_req(MemNode::Control, adr->in(0)); 604 ccp->hash_insert(this); 605 return this; 606 607 // List of "safe" opcodes; those that implicitly block the memory 608 // op below any null check. 609 case Op_CastX2P: // no null checks on native pointers 610 case Op_Parm: // 'this' pointer is not null 611 case Op_LoadP: // Loading from within a klass 612 case Op_LoadN: // Loading from within a klass 613 case Op_LoadKlass: // Loading from within a klass 614 case Op_ConP: // Loading from a klass 615 case Op_CreateEx: // Sucking up the guts of an exception oop 616 case Op_Con: // Reading from TLS 617 case Op_CMoveP: // CMoveP is pinned 618 break; // No progress 619 620 case Op_Proj: // Direct call to an allocation routine 621 case Op_SCMemProj: // Memory state from store conditional ops 622#ifdef ASSERT 623 { 624 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 625 const Node* call = adr->in(0); 626 if (call->is_CallStaticJava()) { 627 const CallStaticJavaNode* call_java = call->as_CallStaticJava(); 628 const TypeTuple *r = call_java->tf()->range(); 629 assert(r->cnt() > TypeFunc::Parms, "must return value"); 630 const Type* ret_type = r->field_at(TypeFunc::Parms); 631 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 632 // We further presume that this is one of 633 // new_instance_Java, new_array_Java, or 634 // the like, but do not assert for this. 635 } else if (call->is_Allocate()) { 636 // similar case to new_instance_Java, etc. 637 } else if (!call->is_CallLeaf()) { 638 // Projections from fetch_oop (OSR) are allowed as well. 639 ShouldNotReachHere(); 640 } 641 } 642#endif 643 break; 644 default: 645 ShouldNotReachHere(); 646 } 647 break; 648 } 649 } 650 651 return NULL; // No progress 652} 653 654 655//============================================================================= 656uint LoadNode::size_of() const { return sizeof(*this); } 657uint LoadNode::cmp( const Node &n ) const 658{ return !Type::cmp( _type, ((LoadNode&)n)._type ); } 659const Type *LoadNode::bottom_type() const { return _type; } 660uint LoadNode::ideal_reg() const { 661 return Matcher::base2reg[_type->base()]; 662} 663 664#ifndef PRODUCT 665void LoadNode::dump_spec(outputStream *st) const { 666 MemNode::dump_spec(st); 667 if( !Verbose && !WizardMode ) { 668 // standard dump does this in Verbose and WizardMode 669 st->print(" #"); _type->dump_on(st); 670 } 671} 672#endif 673 674 675//----------------------------LoadNode::make----------------------------------- 676// Polymorphic factory method: 677Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { 678 Compile* C = gvn.C; 679 680 // sanity check the alias category against the created node type 681 assert(!(adr_type->isa_oopptr() && 682 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 683 "use LoadKlassNode instead"); 684 assert(!(adr_type->isa_aryptr() && 685 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 686 "use LoadRangeNode instead"); 687 switch (bt) { 688 case T_BOOLEAN: 689 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() ); 690 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() ); 691 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() ); 692 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() ); 693 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() ); 694 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt ); 695 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt ); 696 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() ); 697 case T_OBJECT: 698#ifdef _LP64 699 if (adr->bottom_type()->is_narrow()) { 700 const TypeNarrowOop* narrowtype; 701 if (rt->isa_narrowoop()) { 702 narrowtype = rt->is_narrowoop(); 703 rt = narrowtype->make_oopptr(); 704 } else { 705 narrowtype = rt->is_oopptr()->make_narrowoop(); 706 } 707 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, narrowtype)); 708 709 return new (C, 2) DecodeNNode(load, rt); 710 } else 711#endif 712 { 713 assert(!adr->bottom_type()->is_narrow(), "should have got back a narrow oop"); 714 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); 715 } 716 } 717 ShouldNotReachHere(); 718 return (LoadNode*)NULL; 719} 720 721LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { 722 bool require_atomic = true; 723 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); 724} 725 726 727 728 729//------------------------------hash------------------------------------------- 730uint LoadNode::hash() const { 731 // unroll addition of interesting fields 732 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 733} 734 735//---------------------------can_see_stored_value------------------------------ 736// This routine exists to make sure this set of tests is done the same 737// everywhere. We need to make a coordinated change: first LoadNode::Ideal 738// will change the graph shape in a way which makes memory alive twice at the 739// same time (uses the Oracle model of aliasing), then some 740// LoadXNode::Identity will fold things back to the equivalence-class model 741// of aliasing. 742Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 743 Node* ld_adr = in(MemNode::Address); 744 745 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 746 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; 747 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && 748 atp->field() != NULL && !atp->field()->is_volatile()) { 749 uint alias_idx = atp->index(); 750 bool final = atp->field()->is_final(); 751 Node* result = NULL; 752 Node* current = st; 753 // Skip through chains of MemBarNodes checking the MergeMems for 754 // new states for the slice of this load. Stop once any other 755 // kind of node is encountered. Loads from final memory can skip 756 // through any kind of MemBar but normal loads shouldn't skip 757 // through MemBarAcquire since the could allow them to move out of 758 // a synchronized region. 759 while (current->is_Proj()) { 760 int opc = current->in(0)->Opcode(); 761 if ((final && opc == Op_MemBarAcquire) || 762 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { 763 Node* mem = current->in(0)->in(TypeFunc::Memory); 764 if (mem->is_MergeMem()) { 765 MergeMemNode* merge = mem->as_MergeMem(); 766 Node* new_st = merge->memory_at(alias_idx); 767 if (new_st == merge->base_memory()) { 768 // Keep searching 769 current = merge->base_memory(); 770 continue; 771 } 772 // Save the new memory state for the slice and fall through 773 // to exit. 774 result = new_st; 775 } 776 } 777 break; 778 } 779 if (result != NULL) { 780 st = result; 781 } 782 } 783 784 785 // Loop around twice in the case Load -> Initialize -> Store. 786 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 787 for (int trip = 0; trip <= 1; trip++) { 788 789 if (st->is_Store()) { 790 Node* st_adr = st->in(MemNode::Address); 791 if (!phase->eqv(st_adr, ld_adr)) { 792 // Try harder before giving up... Match raw and non-raw pointers. 793 intptr_t st_off = 0; 794 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 795 if (alloc == NULL) return NULL; 796 intptr_t ld_off = 0; 797 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 798 if (alloc != allo2) return NULL; 799 if (ld_off != st_off) return NULL; 800 // At this point we have proven something like this setup: 801 // A = Allocate(...) 802 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 803 // S = StoreQ(, AddP(, A.Parm , #Off), V) 804 // (Actually, we haven't yet proven the Q's are the same.) 805 // In other words, we are loading from a casted version of 806 // the same pointer-and-offset that we stored to. 807 // Thus, we are able to replace L by V. 808 } 809 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 810 if (store_Opcode() != st->Opcode()) 811 return NULL; 812 return st->in(MemNode::ValueIn); 813 } 814 815 intptr_t offset = 0; // scratch 816 817 // A load from a freshly-created object always returns zero. 818 // (This can happen after LoadNode::Ideal resets the load's memory input 819 // to find_captured_store, which returned InitializeNode::zero_memory.) 820 if (st->is_Proj() && st->in(0)->is_Allocate() && 821 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && 822 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { 823 // return a zero value for the load's basic type 824 // (This is one of the few places where a generic PhaseTransform 825 // can create new nodes. Think of it as lazily manifesting 826 // virtually pre-existing constants.) 827 return phase->zerocon(memory_type()); 828 } 829 830 // A load from an initialization barrier can match a captured store. 831 if (st->is_Proj() && st->in(0)->is_Initialize()) { 832 InitializeNode* init = st->in(0)->as_Initialize(); 833 AllocateNode* alloc = init->allocation(); 834 if (alloc != NULL && 835 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { 836 // examine a captured store value 837 st = init->find_captured_store(offset, memory_size(), phase); 838 if (st != NULL) 839 continue; // take one more trip around 840 } 841 } 842 843 break; 844 } 845 846 return NULL; 847} 848 849//----------------------is_instance_field_load_with_local_phi------------------ 850bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 851 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && 852 in(MemNode::Address)->is_AddP() ) { 853 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); 854 // Only instances. 855 if( t_oop != NULL && t_oop->is_instance_field() && 856 t_oop->offset() != Type::OffsetBot && 857 t_oop->offset() != Type::OffsetTop) { 858 return true; 859 } 860 } 861 return false; 862} 863 864//------------------------------Identity--------------------------------------- 865// Loads are identity if previous store is to same address 866Node *LoadNode::Identity( PhaseTransform *phase ) { 867 // If the previous store-maker is the right kind of Store, and the store is 868 // to the same address, then we are equal to the value stored. 869 Node* mem = in(MemNode::Memory); 870 Node* value = can_see_stored_value(mem, phase); 871 if( value ) { 872 // byte, short & char stores truncate naturally. 873 // A load has to load the truncated value which requires 874 // some sort of masking operation and that requires an 875 // Ideal call instead of an Identity call. 876 if (memory_size() < BytesPerInt) { 877 // If the input to the store does not fit with the load's result type, 878 // it must be truncated via an Ideal call. 879 if (!phase->type(value)->higher_equal(phase->type(this))) 880 return this; 881 } 882 // (This works even when value is a Con, but LoadNode::Value 883 // usually runs first, producing the singleton type of the Con.) 884 return value; 885 } 886 887 // Search for an existing data phi which was generated before for the same 888 // instance's field to avoid infinite genertion of phis in a loop. 889 Node *region = mem->in(0); 890 if (is_instance_field_load_with_local_phi(region)) { 891 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); 892 int this_index = phase->C->get_alias_index(addr_t); 893 int this_offset = addr_t->offset(); 894 int this_id = addr_t->is_oopptr()->instance_id(); 895 const Type* this_type = bottom_type(); 896 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 897 Node* phi = region->fast_out(i); 898 if (phi->is_Phi() && phi != mem && 899 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { 900 return phi; 901 } 902 } 903 } 904 905 return this; 906} 907 908 909// Returns true if the AliasType refers to the field that holds the 910// cached box array. Currently only handles the IntegerCache case. 911static bool is_autobox_cache(Compile::AliasType* atp) { 912 if (atp != NULL && atp->field() != NULL) { 913 ciField* field = atp->field(); 914 ciSymbol* klass = field->holder()->name(); 915 if (field->name() == ciSymbol::cache_field_name() && 916 field->holder()->uses_default_loader() && 917 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 918 return true; 919 } 920 } 921 return false; 922} 923 924// Fetch the base value in the autobox array 925static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { 926 if (atp != NULL && atp->field() != NULL) { 927 ciField* field = atp->field(); 928 ciSymbol* klass = field->holder()->name(); 929 if (field->name() == ciSymbol::cache_field_name() && 930 field->holder()->uses_default_loader() && 931 klass == ciSymbol::java_lang_Integer_IntegerCache()) { 932 assert(field->is_constant(), "what?"); 933 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); 934 // Fetch the box object at the base of the array and get its value 935 ciInstance* box = array->obj_at(0)->as_instance(); 936 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 937 if (ik->nof_nonstatic_fields() == 1) { 938 // This should be true nonstatic_field_at requires calling 939 // nof_nonstatic_fields so check it anyway 940 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 941 cache_offset = c.as_int(); 942 } 943 return true; 944 } 945 } 946 return false; 947} 948 949// Returns true if the AliasType refers to the value field of an 950// autobox object. Currently only handles Integer. 951static bool is_autobox_object(Compile::AliasType* atp) { 952 if (atp != NULL && atp->field() != NULL) { 953 ciField* field = atp->field(); 954 ciSymbol* klass = field->holder()->name(); 955 if (field->name() == ciSymbol::value_name() && 956 field->holder()->uses_default_loader() && 957 klass == ciSymbol::java_lang_Integer()) { 958 return true; 959 } 960 } 961 return false; 962} 963 964 965// We're loading from an object which has autobox behaviour. 966// If this object is result of a valueOf call we'll have a phi 967// merging a newly allocated object and a load from the cache. 968// We want to replace this load with the original incoming 969// argument to the valueOf call. 970Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 971 Node* base = in(Address)->in(AddPNode::Base); 972 if (base->is_Phi() && base->req() == 3) { 973 AllocateNode* allocation = NULL; 974 int allocation_index = -1; 975 int load_index = -1; 976 for (uint i = 1; i < base->req(); i++) { 977 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); 978 if (allocation != NULL) { 979 allocation_index = i; 980 load_index = 3 - allocation_index; 981 break; 982 } 983 } 984 LoadNode* load = NULL; 985 if (allocation != NULL && base->in(load_index)->is_Load()) { 986 load = base->in(load_index)->as_Load(); 987 } 988 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { 989 // Push the loads from the phi that comes from valueOf up 990 // through it to allow elimination of the loads and the recovery 991 // of the original value. 992 Node* mem_phi = in(Memory); 993 Node* offset = in(Address)->in(AddPNode::Offset); 994 995 Node* in1 = clone(); 996 Node* in1_addr = in1->in(Address)->clone(); 997 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); 998 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); 999 in1_addr->set_req(AddPNode::Offset, offset); 1000 in1->set_req(0, base->in(allocation_index)); 1001 in1->set_req(Address, in1_addr); 1002 in1->set_req(Memory, mem_phi->in(allocation_index)); 1003 1004 Node* in2 = clone(); 1005 Node* in2_addr = in2->in(Address)->clone(); 1006 in2_addr->set_req(AddPNode::Base, base->in(load_index)); 1007 in2_addr->set_req(AddPNode::Address, base->in(load_index)); 1008 in2_addr->set_req(AddPNode::Offset, offset); 1009 in2->set_req(0, base->in(load_index)); 1010 in2->set_req(Address, in2_addr); 1011 in2->set_req(Memory, mem_phi->in(load_index)); 1012 1013 in1_addr = phase->transform(in1_addr); 1014 in1 = phase->transform(in1); 1015 in2_addr = phase->transform(in2_addr); 1016 in2 = phase->transform(in2); 1017 1018 PhiNode* result = PhiNode::make_blank(base->in(0), this); 1019 result->set_req(allocation_index, in1); 1020 result->set_req(load_index, in2); 1021 return result; 1022 } 1023 } else if (base->is_Load()) { 1024 // Eliminate the load of Integer.value for integers from the cache 1025 // array by deriving the value from the index into the array. 1026 // Capture the offset of the load and then reverse the computation. 1027 Node* load_base = base->in(Address)->in(AddPNode::Base); 1028 if (load_base != NULL) { 1029 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); 1030 intptr_t cache_offset; 1031 int shift = -1; 1032 Node* cache = NULL; 1033 if (is_autobox_cache(atp)) { 1034 shift = exact_log2(type2aelembytes(T_OBJECT)); 1035 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); 1036 } 1037 if (cache != NULL && base->in(Address)->is_AddP()) { 1038 Node* elements[4]; 1039 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 1040 int cache_low; 1041 if (count > 0 && fetch_autobox_base(atp, cache_low)) { 1042 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); 1043 // Add up all the offsets making of the address of the load 1044 Node* result = elements[0]; 1045 for (int i = 1; i < count; i++) { 1046 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); 1047 } 1048 // Remove the constant offset from the address and then 1049 // remove the scaling of the offset to recover the original index. 1050 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); 1051 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1052 // Peel the shift off directly but wrap it in a dummy node 1053 // since Ideal can't return existing nodes 1054 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); 1055 } else { 1056 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); 1057 } 1058#ifdef _LP64 1059 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); 1060#endif 1061 return result; 1062 } 1063 } 1064 } 1065 } 1066 return NULL; 1067} 1068 1069 1070//------------------------------Ideal------------------------------------------ 1071// If the load is from Field memory and the pointer is non-null, we can 1072// zero out the control input. 1073// If the offset is constant and the base is an object allocation, 1074// try to hook me up to the exact initializing store. 1075Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1076 Node* p = MemNode::Ideal_common(phase, can_reshape); 1077 if (p) return (p == NodeSentinel) ? NULL : p; 1078 1079 Node* ctrl = in(MemNode::Control); 1080 Node* address = in(MemNode::Address); 1081 1082 // Skip up past a SafePoint control. Cannot do this for Stores because 1083 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1084 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1085 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1086 ctrl = ctrl->in(0); 1087 set_req(MemNode::Control,ctrl); 1088 } 1089 1090 // Check for useless control edge in some common special cases 1091 if (in(MemNode::Control) != NULL) { 1092 intptr_t ignore = 0; 1093 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1094 if (base != NULL 1095 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1096 && detect_dominating_control(base->in(0), phase->C->start())) { 1097 // A method-invariant, non-null address (constant or 'this' argument). 1098 set_req(MemNode::Control, NULL); 1099 } 1100 } 1101 1102 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) { 1103 Node* base = in(Address)->in(AddPNode::Base); 1104 if (base != NULL) { 1105 Compile::AliasType* atp = phase->C->alias_type(adr_type()); 1106 if (is_autobox_object(atp)) { 1107 Node* result = eliminate_autobox(phase); 1108 if (result != NULL) return result; 1109 } 1110 } 1111 } 1112 1113 Node* mem = in(MemNode::Memory); 1114 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1115 1116 if (addr_t != NULL) { 1117 // try to optimize our memory input 1118 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); 1119 if (opt_mem != mem) { 1120 set_req(MemNode::Memory, opt_mem); 1121 return this; 1122 } 1123 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1124 if (can_reshape && opt_mem->is_Phi() && 1125 (t_oop != NULL) && t_oop->is_instance_field()) { 1126 assert(t_oop->offset() != Type::OffsetBot && t_oop->offset() != Type::OffsetTop, ""); 1127 Node *region = opt_mem->in(0); 1128 uint cnt = opt_mem->req(); 1129 for( uint i = 1; i < cnt; i++ ) { 1130 Node *in = opt_mem->in(i); 1131 if( in == NULL ) { 1132 region = NULL; // Wait stable graph 1133 break; 1134 } 1135 } 1136 if (region != NULL) { 1137 // Check for loop invariant. 1138 if (cnt == 3) { 1139 for( uint i = 1; i < cnt; i++ ) { 1140 Node *in = opt_mem->in(i); 1141 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); 1142 if (m == opt_mem) { 1143 set_req(MemNode::Memory, opt_mem->in(cnt - i)); // Skip this phi. 1144 return this; 1145 } 1146 } 1147 } 1148 // Split through Phi (see original code in loopopts.cpp). 1149 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); 1150 1151 // Do nothing here if Identity will find a value 1152 // (to avoid infinite chain of value phis generation). 1153 if ( !phase->eqv(this, this->Identity(phase)) ) 1154 return NULL; 1155 1156 const Type* this_type = this->bottom_type(); 1157 int this_index = phase->C->get_alias_index(addr_t); 1158 int this_offset = addr_t->offset(); 1159 int this_iid = addr_t->is_oopptr()->instance_id(); 1160 int wins = 0; 1161 PhaseIterGVN *igvn = phase->is_IterGVN(); 1162 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1163 for( uint i = 1; i < region->req(); i++ ) { 1164 Node *x; 1165 Node* the_clone = NULL; 1166 if( region->in(i) == phase->C->top() ) { 1167 x = phase->C->top(); // Dead path? Use a dead data op 1168 } else { 1169 x = this->clone(); // Else clone up the data op 1170 the_clone = x; // Remember for possible deletion. 1171 // Alter data node to use pre-phi inputs 1172 if( this->in(0) == region ) { 1173 x->set_req( 0, region->in(i) ); 1174 } else { 1175 x->set_req( 0, NULL ); 1176 } 1177 for( uint j = 1; j < this->req(); j++ ) { 1178 Node *in = this->in(j); 1179 if( in->is_Phi() && in->in(0) == region ) 1180 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 1181 } 1182 } 1183 // Check for a 'win' on some paths 1184 const Type *t = x->Value(igvn); 1185 1186 bool singleton = t->singleton(); 1187 1188 // See comments in PhaseIdealLoop::split_thru_phi(). 1189 if( singleton && t == Type::TOP ) { 1190 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1191 } 1192 1193 if( singleton ) { 1194 wins++; 1195 x = igvn->makecon(t); 1196 } else { 1197 // We now call Identity to try to simplify the cloned node. 1198 // Note that some Identity methods call phase->type(this). 1199 // Make sure that the type array is big enough for 1200 // our new node, even though we may throw the node away. 1201 // (This tweaking with igvn only works because x is a new node.) 1202 igvn->set_type(x, t); 1203 Node *y = x->Identity(igvn); 1204 if( y != x ) { 1205 wins++; 1206 x = y; 1207 } else { 1208 y = igvn->hash_find(x); 1209 if( y ) { 1210 wins++; 1211 x = y; 1212 } else { 1213 // Else x is a new node we are keeping 1214 // We do not need register_new_node_with_optimizer 1215 // because set_type has already been called. 1216 igvn->_worklist.push(x); 1217 } 1218 } 1219 } 1220 if (x != the_clone && the_clone != NULL) 1221 igvn->remove_dead_node(the_clone); 1222 phi->set_req(i, x); 1223 } 1224 if( wins > 0 ) { 1225 // Record Phi 1226 igvn->register_new_node_with_optimizer(phi); 1227 return phi; 1228 } else { 1229 igvn->remove_dead_node(phi); 1230 } 1231 } 1232 } 1233 } 1234 1235 // Check for prior store with a different base or offset; make Load 1236 // independent. Skip through any number of them. Bail out if the stores 1237 // are in an endless dead cycle and report no progress. This is a key 1238 // transform for Reflection. However, if after skipping through the Stores 1239 // we can't then fold up against a prior store do NOT do the transform as 1240 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1241 // array memory alive twice: once for the hoisted Load and again after the 1242 // bypassed Store. This situation only works if EVERYBODY who does 1243 // anti-dependence work knows how to bypass. I.e. we need all 1244 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1245 // the alias index stuff. So instead, peek through Stores and IFF we can 1246 // fold up, do so. 1247 Node* prev_mem = find_previous_store(phase); 1248 // Steps (a), (b): Walk past independent stores to find an exact match. 1249 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1250 // (c) See if we can fold up on the spot, but don't fold up here. 1251 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or 1252 // just return a prior value, which is done by Identity calls. 1253 if (can_see_stored_value(prev_mem, phase)) { 1254 // Make ready for step (d): 1255 set_req(MemNode::Memory, prev_mem); 1256 return this; 1257 } 1258 } 1259 1260 return NULL; // No further progress 1261} 1262 1263// Helper to recognize certain Klass fields which are invariant across 1264// some group of array types (e.g., int[] or all T[] where T < Object). 1265const Type* 1266LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1267 ciKlass* klass) const { 1268 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1269 // The field is Klass::_modifier_flags. Return its (constant) value. 1270 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1271 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1272 return TypeInt::make(klass->modifier_flags()); 1273 } 1274 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { 1275 // The field is Klass::_access_flags. Return its (constant) value. 1276 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1277 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1278 return TypeInt::make(klass->access_flags()); 1279 } 1280 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { 1281 // The field is Klass::_layout_helper. Return its constant value if known. 1282 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1283 return TypeInt::make(klass->layout_helper()); 1284 } 1285 1286 // No match. 1287 return NULL; 1288} 1289 1290//------------------------------Value----------------------------------------- 1291const Type *LoadNode::Value( PhaseTransform *phase ) const { 1292 // Either input is TOP ==> the result is TOP 1293 Node* mem = in(MemNode::Memory); 1294 const Type *t1 = phase->type(mem); 1295 if (t1 == Type::TOP) return Type::TOP; 1296 Node* adr = in(MemNode::Address); 1297 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1298 if (tp == NULL || tp->empty()) return Type::TOP; 1299 int off = tp->offset(); 1300 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1301 1302 // Try to guess loaded type from pointer type 1303 if (tp->base() == Type::AryPtr) { 1304 const Type *t = tp->is_aryptr()->elem(); 1305 // Don't do this for integer types. There is only potential profit if 1306 // the element type t is lower than _type; that is, for int types, if _type is 1307 // more restrictive than t. This only happens here if one is short and the other 1308 // char (both 16 bits), and in those cases we've made an intentional decision 1309 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1310 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1311 // 1312 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1313 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1314 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1315 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1316 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1317 // In fact, that could have been the original type of p1, and p1 could have 1318 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1319 // expression (LShiftL quux 3) independently optimized to the constant 8. 1320 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1321 && Opcode() != Op_LoadKlass) { 1322 // t might actually be lower than _type, if _type is a unique 1323 // concrete subclass of abstract class t. 1324 // Make sure the reference is not into the header, by comparing 1325 // the offset against the offset of the start of the array's data. 1326 // Different array types begin at slightly different offsets (12 vs. 16). 1327 // We choose T_BYTE as an example base type that is least restrictive 1328 // as to alignment, which will therefore produce the smallest 1329 // possible base offset. 1330 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1331 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? 1332 const Type* jt = t->join(_type); 1333 // In any case, do not allow the join, per se, to empty out the type. 1334 if (jt->empty() && !t->empty()) { 1335 // This can happen if a interface-typed array narrows to a class type. 1336 jt = _type; 1337 } 1338 1339 if (EliminateAutoBox) { 1340 // The pointers in the autobox arrays are always non-null 1341 Node* base = in(Address)->in(AddPNode::Base); 1342 if (base != NULL) { 1343 Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); 1344 if (is_autobox_cache(atp)) { 1345 return jt->join(TypePtr::NOTNULL)->is_ptr(); 1346 } 1347 } 1348 } 1349 return jt; 1350 } 1351 } 1352 } else if (tp->base() == Type::InstPtr) { 1353 assert( off != Type::OffsetBot || 1354 // arrays can be cast to Objects 1355 tp->is_oopptr()->klass()->is_java_lang_Object() || 1356 // unsafe field access may not have a constant offset 1357 phase->C->has_unsafe_access(), 1358 "Field accesses must be precise" ); 1359 // For oop loads, we expect the _type to be precise 1360 } else if (tp->base() == Type::KlassPtr) { 1361 assert( off != Type::OffsetBot || 1362 // arrays can be cast to Objects 1363 tp->is_klassptr()->klass()->is_java_lang_Object() || 1364 // also allow array-loading from the primary supertype 1365 // array during subtype checks 1366 Opcode() == Op_LoadKlass, 1367 "Field accesses must be precise" ); 1368 // For klass/static loads, we expect the _type to be precise 1369 } 1370 1371 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1372 if (tkls != NULL && !StressReflectiveCode) { 1373 ciKlass* klass = tkls->klass(); 1374 if (klass->is_loaded() && tkls->klass_is_exact()) { 1375 // We are loading a field from a Klass metaobject whose identity 1376 // is known at compile time (the type is "exact" or "precise"). 1377 // Check for fields we know are maintained as constants by the VM. 1378 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { 1379 // The field is Klass::_super_check_offset. Return its (constant) value. 1380 // (Folds up type checking code.) 1381 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1382 return TypeInt::make(klass->super_check_offset()); 1383 } 1384 // Compute index into primary_supers array 1385 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1386 // Check for overflowing; use unsigned compare to handle the negative case. 1387 if( depth < ciKlass::primary_super_limit() ) { 1388 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1389 // (Folds up type checking code.) 1390 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1391 ciKlass *ss = klass->super_of_depth(depth); 1392 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1393 } 1394 const Type* aift = load_array_final_field(tkls, klass); 1395 if (aift != NULL) return aift; 1396 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) 1397 && klass->is_array_klass()) { 1398 // The field is arrayKlass::_component_mirror. Return its (constant) value. 1399 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1400 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1401 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1402 } 1403 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { 1404 // The field is Klass::_java_mirror. Return its (constant) value. 1405 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1406 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1407 return TypeInstPtr::make(klass->java_mirror()); 1408 } 1409 } 1410 1411 // We can still check if we are loading from the primary_supers array at a 1412 // shallow enough depth. Even though the klass is not exact, entries less 1413 // than or equal to its super depth are correct. 1414 if (klass->is_loaded() ) { 1415 ciType *inner = klass->klass(); 1416 while( inner->is_obj_array_klass() ) 1417 inner = inner->as_obj_array_klass()->base_element_type(); 1418 if( inner->is_instance_klass() && 1419 !inner->as_instance_klass()->flags().is_interface() ) { 1420 // Compute index into primary_supers array 1421 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); 1422 // Check for overflowing; use unsigned compare to handle the negative case. 1423 if( depth < ciKlass::primary_super_limit() && 1424 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1425 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1426 // (Folds up type checking code.) 1427 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1428 ciKlass *ss = klass->super_of_depth(depth); 1429 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1430 } 1431 } 1432 } 1433 1434 // If the type is enough to determine that the thing is not an array, 1435 // we can give the layout_helper a positive interval type. 1436 // This will help short-circuit some reflective code. 1437 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) 1438 && !klass->is_array_klass() // not directly typed as an array 1439 && !klass->is_interface() // specifically not Serializable & Cloneable 1440 && !klass->is_java_lang_Object() // not the supertype of all T[] 1441 ) { 1442 // Note: When interfaces are reliable, we can narrow the interface 1443 // test to (klass != Serializable && klass != Cloneable). 1444 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1445 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1446 // The key property of this type is that it folds up tests 1447 // for array-ness, since it proves that the layout_helper is positive. 1448 // Thus, a generic value like the basic object layout helper works fine. 1449 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1450 } 1451 } 1452 1453 // If we are loading from a freshly-allocated object, produce a zero, 1454 // if the load is provably beyond the header of the object. 1455 // (Also allow a variable load from a fresh array to produce zero.) 1456 if (ReduceFieldZeroing) { 1457 Node* value = can_see_stored_value(mem,phase); 1458 if (value != NULL && value->is_Con()) 1459 return value->bottom_type(); 1460 } 1461 1462 const TypeOopPtr *tinst = tp->isa_oopptr(); 1463 if (tinst != NULL && tinst->is_instance_field()) { 1464 // If we have an instance type and our memory input is the 1465 // programs's initial memory state, there is no matching store, 1466 // so just return a zero of the appropriate type 1467 Node *mem = in(MemNode::Memory); 1468 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1469 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1470 return Type::get_zero_type(_type->basic_type()); 1471 } 1472 } 1473 return _type; 1474} 1475 1476//------------------------------match_edge------------------------------------- 1477// Do we Match on this edge index or not? Match only the address. 1478uint LoadNode::match_edge(uint idx) const { 1479 return idx == MemNode::Address; 1480} 1481 1482//--------------------------LoadBNode::Ideal-------------------------------------- 1483// 1484// If the previous store is to the same address as this load, 1485// and the value stored was larger than a byte, replace this load 1486// with the value stored truncated to a byte. If no truncation is 1487// needed, the replacement is done in LoadNode::Identity(). 1488// 1489Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1490 Node* mem = in(MemNode::Memory); 1491 Node* value = can_see_stored_value(mem,phase); 1492 if( value && !phase->type(value)->higher_equal( _type ) ) { 1493 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); 1494 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); 1495 } 1496 // Identity call will handle the case where truncation is not needed. 1497 return LoadNode::Ideal(phase, can_reshape); 1498} 1499 1500//--------------------------LoadCNode::Ideal-------------------------------------- 1501// 1502// If the previous store is to the same address as this load, 1503// and the value stored was larger than a char, replace this load 1504// with the value stored truncated to a char. If no truncation is 1505// needed, the replacement is done in LoadNode::Identity(). 1506// 1507Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1508 Node* mem = in(MemNode::Memory); 1509 Node* value = can_see_stored_value(mem,phase); 1510 if( value && !phase->type(value)->higher_equal( _type ) ) 1511 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); 1512 // Identity call will handle the case where truncation is not needed. 1513 return LoadNode::Ideal(phase, can_reshape); 1514} 1515 1516//--------------------------LoadSNode::Ideal-------------------------------------- 1517// 1518// If the previous store is to the same address as this load, 1519// and the value stored was larger than a short, replace this load 1520// with the value stored truncated to a short. If no truncation is 1521// needed, the replacement is done in LoadNode::Identity(). 1522// 1523Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1524 Node* mem = in(MemNode::Memory); 1525 Node* value = can_see_stored_value(mem,phase); 1526 if( value && !phase->type(value)->higher_equal( _type ) ) { 1527 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); 1528 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); 1529 } 1530 // Identity call will handle the case where truncation is not needed. 1531 return LoadNode::Ideal(phase, can_reshape); 1532} 1533 1534//============================================================================= 1535//------------------------------Value------------------------------------------ 1536const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 1537 // Either input is TOP ==> the result is TOP 1538 const Type *t1 = phase->type( in(MemNode::Memory) ); 1539 if (t1 == Type::TOP) return Type::TOP; 1540 Node *adr = in(MemNode::Address); 1541 const Type *t2 = phase->type( adr ); 1542 if (t2 == Type::TOP) return Type::TOP; 1543 const TypePtr *tp = t2->is_ptr(); 1544 if (TypePtr::above_centerline(tp->ptr()) || 1545 tp->ptr() == TypePtr::Null) return Type::TOP; 1546 1547 // Return a more precise klass, if possible 1548 const TypeInstPtr *tinst = tp->isa_instptr(); 1549 if (tinst != NULL) { 1550 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1551 int offset = tinst->offset(); 1552 if (ik == phase->C->env()->Class_klass() 1553 && (offset == java_lang_Class::klass_offset_in_bytes() || 1554 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1555 // We are loading a special hidden field from a Class mirror object, 1556 // the field which points to the VM's Klass metaobject. 1557 ciType* t = tinst->java_mirror_type(); 1558 // java_mirror_type returns non-null for compile-time Class constants. 1559 if (t != NULL) { 1560 // constant oop => constant klass 1561 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1562 return TypeKlassPtr::make(ciArrayKlass::make(t)); 1563 } 1564 if (!t->is_klass()) { 1565 // a primitive Class (e.g., int.class) has NULL for a klass field 1566 return TypePtr::NULL_PTR; 1567 } 1568 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 1569 return TypeKlassPtr::make(t->as_klass()); 1570 } 1571 // non-constant mirror, so we can't tell what's going on 1572 } 1573 if( !ik->is_loaded() ) 1574 return _type; // Bail out if not loaded 1575 if (offset == oopDesc::klass_offset_in_bytes()) { 1576 if (tinst->klass_is_exact()) { 1577 return TypeKlassPtr::make(ik); 1578 } 1579 // See if we can become precise: no subklasses and no interface 1580 // (Note: We need to support verified interfaces.) 1581 if (!ik->is_interface() && !ik->has_subklass()) { 1582 //assert(!UseExactTypes, "this code should be useless with exact types"); 1583 // Add a dependence; if any subclass added we need to recompile 1584 if (!ik->is_final()) { 1585 // %%% should use stronger assert_unique_concrete_subtype instead 1586 phase->C->dependencies()->assert_leaf_type(ik); 1587 } 1588 // Return precise klass 1589 return TypeKlassPtr::make(ik); 1590 } 1591 1592 // Return root of possible klass 1593 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 1594 } 1595 } 1596 1597 // Check for loading klass from an array 1598 const TypeAryPtr *tary = tp->isa_aryptr(); 1599 if( tary != NULL ) { 1600 ciKlass *tary_klass = tary->klass(); 1601 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 1602 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 1603 if (tary->klass_is_exact()) { 1604 return TypeKlassPtr::make(tary_klass); 1605 } 1606 ciArrayKlass *ak = tary->klass()->as_array_klass(); 1607 // If the klass is an object array, we defer the question to the 1608 // array component klass. 1609 if( ak->is_obj_array_klass() ) { 1610 assert( ak->is_loaded(), "" ); 1611 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 1612 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 1613 ciInstanceKlass* ik = base_k->as_instance_klass(); 1614 // See if we can become precise: no subklasses and no interface 1615 if (!ik->is_interface() && !ik->has_subklass()) { 1616 //assert(!UseExactTypes, "this code should be useless with exact types"); 1617 // Add a dependence; if any subclass added we need to recompile 1618 if (!ik->is_final()) { 1619 phase->C->dependencies()->assert_leaf_type(ik); 1620 } 1621 // Return precise array klass 1622 return TypeKlassPtr::make(ak); 1623 } 1624 } 1625 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 1626 } else { // Found a type-array? 1627 //assert(!UseExactTypes, "this code should be useless with exact types"); 1628 assert( ak->is_type_array_klass(), "" ); 1629 return TypeKlassPtr::make(ak); // These are always precise 1630 } 1631 } 1632 } 1633 1634 // Check for loading klass from an array klass 1635 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1636 if (tkls != NULL && !StressReflectiveCode) { 1637 ciKlass* klass = tkls->klass(); 1638 if( !klass->is_loaded() ) 1639 return _type; // Bail out if not loaded 1640 if( klass->is_obj_array_klass() && 1641 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { 1642 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 1643 // // Always returning precise element type is incorrect, 1644 // // e.g., element type could be object and array may contain strings 1645 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 1646 1647 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 1648 // according to the element type's subclassing. 1649 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 1650 } 1651 if( klass->is_instance_klass() && tkls->klass_is_exact() && 1652 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { 1653 ciKlass* sup = klass->as_instance_klass()->super(); 1654 // The field is Klass::_super. Return its (constant) value. 1655 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 1656 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 1657 } 1658 } 1659 1660 // Bailout case 1661 return LoadNode::Value(phase); 1662} 1663 1664//------------------------------Identity--------------------------------------- 1665// To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 1666// Also feed through the klass in Allocate(...klass...)._klass. 1667Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 1668 Node* x = LoadNode::Identity(phase); 1669 if (x != this) return x; 1670 1671 // Take apart the address into an oop and and offset. 1672 // Return 'this' if we cannot. 1673 Node* adr = in(MemNode::Address); 1674 intptr_t offset = 0; 1675 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1676 if (base == NULL) return this; 1677 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 1678 if (toop == NULL) return this; 1679 1680 // We can fetch the klass directly through an AllocateNode. 1681 // This works even if the klass is not constant (clone or newArray). 1682 if (offset == oopDesc::klass_offset_in_bytes()) { 1683 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 1684 if (allocated_klass != NULL) { 1685 return allocated_klass; 1686 } 1687 } 1688 1689 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. 1690 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. 1691 // See inline_native_Class_query for occurrences of these patterns. 1692 // Java Example: x.getClass().isAssignableFrom(y) 1693 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 1694 // 1695 // This improves reflective code, often making the Class 1696 // mirror go completely dead. (Current exception: Class 1697 // mirrors may appear in debug info, but we could clean them out by 1698 // introducing a new debug info operator for klassOop.java_mirror). 1699 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 1700 && (offset == java_lang_Class::klass_offset_in_bytes() || 1701 offset == java_lang_Class::array_klass_offset_in_bytes())) { 1702 // We are loading a special hidden field from a Class mirror, 1703 // the field which points to its Klass or arrayKlass metaobject. 1704 if (base->is_Load()) { 1705 Node* adr2 = base->in(MemNode::Address); 1706 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 1707 if (tkls != NULL && !tkls->empty() 1708 && (tkls->klass()->is_instance_klass() || 1709 tkls->klass()->is_array_klass()) 1710 && adr2->is_AddP() 1711 ) { 1712 int mirror_field = Klass::java_mirror_offset_in_bytes(); 1713 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 1714 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); 1715 } 1716 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { 1717 return adr2->in(AddPNode::Base); 1718 } 1719 } 1720 } 1721 } 1722 1723 return this; 1724} 1725 1726//------------------------------Value----------------------------------------- 1727const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 1728 // Either input is TOP ==> the result is TOP 1729 const Type *t1 = phase->type( in(MemNode::Memory) ); 1730 if( t1 == Type::TOP ) return Type::TOP; 1731 Node *adr = in(MemNode::Address); 1732 const Type *t2 = phase->type( adr ); 1733 if( t2 == Type::TOP ) return Type::TOP; 1734 const TypePtr *tp = t2->is_ptr(); 1735 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 1736 const TypeAryPtr *tap = tp->isa_aryptr(); 1737 if( !tap ) return _type; 1738 return tap->size(); 1739} 1740 1741//------------------------------Identity--------------------------------------- 1742// Feed through the length in AllocateArray(...length...)._length. 1743Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 1744 Node* x = LoadINode::Identity(phase); 1745 if (x != this) return x; 1746 1747 // Take apart the address into an oop and and offset. 1748 // Return 'this' if we cannot. 1749 Node* adr = in(MemNode::Address); 1750 intptr_t offset = 0; 1751 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 1752 if (base == NULL) return this; 1753 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 1754 if (tary == NULL) return this; 1755 1756 // We can fetch the length directly through an AllocateArrayNode. 1757 // This works even if the length is not constant (clone or newArray). 1758 if (offset == arrayOopDesc::length_offset_in_bytes()) { 1759 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase); 1760 if (allocated_length != NULL) { 1761 return allocated_length; 1762 } 1763 } 1764 1765 return this; 1766 1767} 1768//============================================================================= 1769//---------------------------StoreNode::make----------------------------------- 1770// Polymorphic factory method: 1771StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { 1772 Compile* C = gvn.C; 1773 1774 switch (bt) { 1775 case T_BOOLEAN: 1776 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); 1777 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); 1778 case T_CHAR: 1779 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); 1780 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); 1781 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); 1782 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); 1783 case T_ADDRESS: 1784 case T_OBJECT: 1785#ifdef _LP64 1786 if (adr->bottom_type()->is_narrow() || 1787 (UseCompressedOops && val->bottom_type()->isa_klassptr() && 1788 adr->bottom_type()->isa_rawptr())) { 1789 const TypePtr* type = val->bottom_type()->is_ptr(); 1790 Node* cp; 1791 if (type->isa_oopptr()) { 1792 const TypeNarrowOop* etype = type->is_oopptr()->make_narrowoop(); 1793 cp = gvn.transform(new (C, 2) EncodePNode(val, etype)); 1794 } else if (type == TypePtr::NULL_PTR) { 1795 cp = gvn.transform(new (C, 1) ConNNode(TypeNarrowOop::NULL_PTR)); 1796 } else { 1797 ShouldNotReachHere(); 1798 } 1799 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp); 1800 } else 1801#endif 1802 { 1803 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); 1804 } 1805 } 1806 ShouldNotReachHere(); 1807 return (StoreNode*)NULL; 1808} 1809 1810StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { 1811 bool require_atomic = true; 1812 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); 1813} 1814 1815 1816//--------------------------bottom_type---------------------------------------- 1817const Type *StoreNode::bottom_type() const { 1818 return Type::MEMORY; 1819} 1820 1821//------------------------------hash------------------------------------------- 1822uint StoreNode::hash() const { 1823 // unroll addition of interesting fields 1824 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 1825 1826 // Since they are not commoned, do not hash them: 1827 return NO_HASH; 1828} 1829 1830//------------------------------Ideal------------------------------------------ 1831// Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 1832// When a store immediately follows a relevant allocation/initialization, 1833// try to capture it into the initialization, or hoist it above. 1834Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1835 Node* p = MemNode::Ideal_common(phase, can_reshape); 1836 if (p) return (p == NodeSentinel) ? NULL : p; 1837 1838 Node* mem = in(MemNode::Memory); 1839 Node* address = in(MemNode::Address); 1840 1841 // Back-to-back stores to same address? Fold em up. 1842 // Generally unsafe if I have intervening uses... 1843 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { 1844 // Looking at a dead closed cycle of memory? 1845 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 1846 1847 assert(Opcode() == mem->Opcode() || 1848 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 1849 "no mismatched stores, except on raw memory"); 1850 1851 if (mem->outcnt() == 1 && // check for intervening uses 1852 mem->as_Store()->memory_size() <= this->memory_size()) { 1853 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 1854 // For example, 'mem' might be the final state at a conditional return. 1855 // Or, 'mem' might be used by some node which is live at the same time 1856 // 'this' is live, which might be unschedulable. So, require exactly 1857 // ONE user, the 'this' store, until such time as we clone 'mem' for 1858 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 1859 if (can_reshape) { // (%%% is this an anachronism?) 1860 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 1861 phase->is_IterGVN()); 1862 } else { 1863 // It's OK to do this in the parser, since DU info is always accurate, 1864 // and the parser always refers to nodes via SafePointNode maps. 1865 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 1866 } 1867 return this; 1868 } 1869 } 1870 1871 // Capture an unaliased, unconditional, simple store into an initializer. 1872 // Or, if it is independent of the allocation, hoist it above the allocation. 1873 if (ReduceFieldZeroing && /*can_reshape &&*/ 1874 mem->is_Proj() && mem->in(0)->is_Initialize()) { 1875 InitializeNode* init = mem->in(0)->as_Initialize(); 1876 intptr_t offset = init->can_capture_store(this, phase); 1877 if (offset > 0) { 1878 Node* moved = init->capture_store(this, offset, phase); 1879 // If the InitializeNode captured me, it made a raw copy of me, 1880 // and I need to disappear. 1881 if (moved != NULL) { 1882 // %%% hack to ensure that Ideal returns a new node: 1883 mem = MergeMemNode::make(phase->C, mem); 1884 return mem; // fold me away 1885 } 1886 } 1887 } 1888 1889 return NULL; // No further progress 1890} 1891 1892//------------------------------Value----------------------------------------- 1893const Type *StoreNode::Value( PhaseTransform *phase ) const { 1894 // Either input is TOP ==> the result is TOP 1895 const Type *t1 = phase->type( in(MemNode::Memory) ); 1896 if( t1 == Type::TOP ) return Type::TOP; 1897 const Type *t2 = phase->type( in(MemNode::Address) ); 1898 if( t2 == Type::TOP ) return Type::TOP; 1899 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 1900 if( t3 == Type::TOP ) return Type::TOP; 1901 return Type::MEMORY; 1902} 1903 1904//------------------------------Identity--------------------------------------- 1905// Remove redundant stores: 1906// Store(m, p, Load(m, p)) changes to m. 1907// Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 1908Node *StoreNode::Identity( PhaseTransform *phase ) { 1909 Node* mem = in(MemNode::Memory); 1910 Node* adr = in(MemNode::Address); 1911 Node* val = in(MemNode::ValueIn); 1912 1913 // Load then Store? Then the Store is useless 1914 if (val->is_Load() && 1915 phase->eqv_uncast( val->in(MemNode::Address), adr ) && 1916 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && 1917 val->as_Load()->store_Opcode() == Opcode()) { 1918 return mem; 1919 } 1920 1921 // Two stores in a row of the same value? 1922 if (mem->is_Store() && 1923 phase->eqv_uncast( mem->in(MemNode::Address), adr ) && 1924 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && 1925 mem->Opcode() == Opcode()) { 1926 return mem; 1927 } 1928 1929 // Store of zero anywhere into a freshly-allocated object? 1930 // Then the store is useless. 1931 // (It must already have been captured by the InitializeNode.) 1932 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 1933 // a newly allocated object is already all-zeroes everywhere 1934 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 1935 return mem; 1936 } 1937 1938 // the store may also apply to zero-bits in an earlier object 1939 Node* prev_mem = find_previous_store(phase); 1940 // Steps (a), (b): Walk past independent stores to find an exact match. 1941 if (prev_mem != NULL) { 1942 Node* prev_val = can_see_stored_value(prev_mem, phase); 1943 if (prev_val != NULL && phase->eqv(prev_val, val)) { 1944 // prev_val and val might differ by a cast; it would be good 1945 // to keep the more informative of the two. 1946 return mem; 1947 } 1948 } 1949 } 1950 1951 return this; 1952} 1953 1954//------------------------------match_edge------------------------------------- 1955// Do we Match on this edge index or not? Match only memory & value 1956uint StoreNode::match_edge(uint idx) const { 1957 return idx == MemNode::Address || idx == MemNode::ValueIn; 1958} 1959 1960//------------------------------cmp-------------------------------------------- 1961// Do not common stores up together. They generally have to be split 1962// back up anyways, so do not bother. 1963uint StoreNode::cmp( const Node &n ) const { 1964 return (&n == this); // Always fail except on self 1965} 1966 1967//------------------------------Ideal_masked_input----------------------------- 1968// Check for a useless mask before a partial-word store 1969// (StoreB ... (AndI valIn conIa) ) 1970// If (conIa & mask == mask) this simplifies to 1971// (StoreB ... (valIn) ) 1972Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 1973 Node *val = in(MemNode::ValueIn); 1974 if( val->Opcode() == Op_AndI ) { 1975 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 1976 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 1977 set_req(MemNode::ValueIn, val->in(1)); 1978 return this; 1979 } 1980 } 1981 return NULL; 1982} 1983 1984 1985//------------------------------Ideal_sign_extended_input---------------------- 1986// Check for useless sign-extension before a partial-word store 1987// (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 1988// If (conIL == conIR && conIR <= num_bits) this simplifies to 1989// (StoreB ... (valIn) ) 1990Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 1991 Node *val = in(MemNode::ValueIn); 1992 if( val->Opcode() == Op_RShiftI ) { 1993 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 1994 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 1995 Node *shl = val->in(1); 1996 if( shl->Opcode() == Op_LShiftI ) { 1997 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 1998 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 1999 set_req(MemNode::ValueIn, shl->in(1)); 2000 return this; 2001 } 2002 } 2003 } 2004 } 2005 return NULL; 2006} 2007 2008//------------------------------value_never_loaded----------------------------------- 2009// Determine whether there are any possible loads of the value stored. 2010// For simplicity, we actually check if there are any loads from the 2011// address stored to, not just for loads of the value stored by this node. 2012// 2013bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2014 Node *adr = in(Address); 2015 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2016 if (adr_oop == NULL) 2017 return false; 2018 if (!adr_oop->is_instance_field()) 2019 return false; // if not a distinct instance, there may be aliases of the address 2020 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2021 Node *use = adr->fast_out(i); 2022 int opc = use->Opcode(); 2023 if (use->is_Load() || use->is_LoadStore()) { 2024 return false; 2025 } 2026 } 2027 return true; 2028} 2029 2030//============================================================================= 2031//------------------------------Ideal------------------------------------------ 2032// If the store is from an AND mask that leaves the low bits untouched, then 2033// we can skip the AND operation. If the store is from a sign-extension 2034// (a left shift, then right shift) we can skip both. 2035Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2036 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2037 if( progress != NULL ) return progress; 2038 2039 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2040 if( progress != NULL ) return progress; 2041 2042 // Finally check the default case 2043 return StoreNode::Ideal(phase, can_reshape); 2044} 2045 2046//============================================================================= 2047//------------------------------Ideal------------------------------------------ 2048// If the store is from an AND mask that leaves the low bits untouched, then 2049// we can skip the AND operation 2050Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2051 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2052 if( progress != NULL ) return progress; 2053 2054 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2055 if( progress != NULL ) return progress; 2056 2057 // Finally check the default case 2058 return StoreNode::Ideal(phase, can_reshape); 2059} 2060 2061//============================================================================= 2062//------------------------------Identity--------------------------------------- 2063Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2064 // No need to card mark when storing a null ptr 2065 Node* my_store = in(MemNode::OopStore); 2066 if (my_store->is_Store()) { 2067 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2068 if( t1 == TypePtr::NULL_PTR ) { 2069 return in(MemNode::Memory); 2070 } 2071 } 2072 return this; 2073} 2074 2075//------------------------------Value----------------------------------------- 2076const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2077 // Either input is TOP ==> the result is TOP 2078 const Type *t = phase->type( in(MemNode::Memory) ); 2079 if( t == Type::TOP ) return Type::TOP; 2080 t = phase->type( in(MemNode::Address) ); 2081 if( t == Type::TOP ) return Type::TOP; 2082 t = phase->type( in(MemNode::ValueIn) ); 2083 if( t == Type::TOP ) return Type::TOP; 2084 // If extra input is TOP ==> the result is TOP 2085 t = phase->type( in(MemNode::OopStore) ); 2086 if( t == Type::TOP ) return Type::TOP; 2087 2088 return StoreNode::Value( phase ); 2089} 2090 2091 2092//============================================================================= 2093//----------------------------------SCMemProjNode------------------------------ 2094const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2095{ 2096 return bottom_type(); 2097} 2098 2099//============================================================================= 2100LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { 2101 init_req(MemNode::Control, c ); 2102 init_req(MemNode::Memory , mem); 2103 init_req(MemNode::Address, adr); 2104 init_req(MemNode::ValueIn, val); 2105 init_req( ExpectedIn, ex ); 2106 init_class_id(Class_LoadStore); 2107 2108} 2109 2110//============================================================================= 2111//-------------------------------adr_type-------------------------------------- 2112// Do we Match on this edge index or not? Do not match memory 2113const TypePtr* ClearArrayNode::adr_type() const { 2114 Node *adr = in(3); 2115 return MemNode::calculate_adr_type(adr->bottom_type()); 2116} 2117 2118//------------------------------match_edge------------------------------------- 2119// Do we Match on this edge index or not? Do not match memory 2120uint ClearArrayNode::match_edge(uint idx) const { 2121 return idx > 1; 2122} 2123 2124//------------------------------Identity--------------------------------------- 2125// Clearing a zero length array does nothing 2126Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2127 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2128} 2129 2130//------------------------------Idealize--------------------------------------- 2131// Clearing a short array is faster with stores 2132Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2133 const int unit = BytesPerLong; 2134 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2135 if (!t) return NULL; 2136 if (!t->is_con()) return NULL; 2137 intptr_t raw_count = t->get_con(); 2138 intptr_t size = raw_count; 2139 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2140 // Clearing nothing uses the Identity call. 2141 // Negative clears are possible on dead ClearArrays 2142 // (see jck test stmt114.stmt11402.val). 2143 if (size <= 0 || size % unit != 0) return NULL; 2144 intptr_t count = size / unit; 2145 // Length too long; use fast hardware clear 2146 if (size > Matcher::init_array_short_size) return NULL; 2147 Node *mem = in(1); 2148 if( phase->type(mem)==Type::TOP ) return NULL; 2149 Node *adr = in(3); 2150 const Type* at = phase->type(adr); 2151 if( at==Type::TOP ) return NULL; 2152 const TypePtr* atp = at->isa_ptr(); 2153 // adjust atp to be the correct array element address type 2154 if (atp == NULL) atp = TypePtr::BOTTOM; 2155 else atp = atp->add_offset(Type::OffsetBot); 2156 // Get base for derived pointer purposes 2157 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2158 Node *base = adr->in(1); 2159 2160 Node *zero = phase->makecon(TypeLong::ZERO); 2161 Node *off = phase->MakeConX(BytesPerLong); 2162 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2163 count--; 2164 while( count-- ) { 2165 mem = phase->transform(mem); 2166 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); 2167 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); 2168 } 2169 return mem; 2170} 2171 2172//----------------------------clear_memory------------------------------------- 2173// Generate code to initialize object storage to zero. 2174Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2175 intptr_t start_offset, 2176 Node* end_offset, 2177 PhaseGVN* phase) { 2178 Compile* C = phase->C; 2179 intptr_t offset = start_offset; 2180 2181 int unit = BytesPerLong; 2182 if ((offset % unit) != 0) { 2183 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); 2184 adr = phase->transform(adr); 2185 const TypePtr* atp = TypeRawPtr::BOTTOM; 2186 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2187 mem = phase->transform(mem); 2188 offset += BytesPerInt; 2189 } 2190 assert((offset % unit) == 0, ""); 2191 2192 // Initialize the remaining stuff, if any, with a ClearArray. 2193 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2194} 2195 2196Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2197 Node* start_offset, 2198 Node* end_offset, 2199 PhaseGVN* phase) { 2200 if (start_offset == end_offset) { 2201 // nothing to do 2202 return mem; 2203 } 2204 2205 Compile* C = phase->C; 2206 int unit = BytesPerLong; 2207 Node* zbase = start_offset; 2208 Node* zend = end_offset; 2209 2210 // Scale to the unit required by the CPU: 2211 if (!Matcher::init_array_count_is_in_bytes) { 2212 Node* shift = phase->intcon(exact_log2(unit)); 2213 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); 2214 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); 2215 } 2216 2217 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); 2218 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); 2219 2220 // Bulk clear double-words 2221 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); 2222 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); 2223 return phase->transform(mem); 2224} 2225 2226Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2227 intptr_t start_offset, 2228 intptr_t end_offset, 2229 PhaseGVN* phase) { 2230 if (start_offset == end_offset) { 2231 // nothing to do 2232 return mem; 2233 } 2234 2235 Compile* C = phase->C; 2236 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2237 intptr_t done_offset = end_offset; 2238 if ((done_offset % BytesPerLong) != 0) { 2239 done_offset -= BytesPerInt; 2240 } 2241 if (done_offset > start_offset) { 2242 mem = clear_memory(ctl, mem, dest, 2243 start_offset, phase->MakeConX(done_offset), phase); 2244 } 2245 if (done_offset < end_offset) { // emit the final 32-bit store 2246 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); 2247 adr = phase->transform(adr); 2248 const TypePtr* atp = TypeRawPtr::BOTTOM; 2249 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); 2250 mem = phase->transform(mem); 2251 done_offset += BytesPerInt; 2252 } 2253 assert(done_offset == end_offset, ""); 2254 return mem; 2255} 2256 2257//============================================================================= 2258// Do we match on this edge? No memory edges 2259uint StrCompNode::match_edge(uint idx) const { 2260 return idx == 5 || idx == 6; 2261} 2262 2263//------------------------------Ideal------------------------------------------ 2264// Return a node which is more "ideal" than the current node. Strip out 2265// control copies 2266Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2267 return remove_dead_region(phase, can_reshape) ? this : NULL; 2268} 2269 2270 2271//============================================================================= 2272MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2273 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2274 _adr_type(C->get_adr_type(alias_idx)) 2275{ 2276 init_class_id(Class_MemBar); 2277 Node* top = C->top(); 2278 init_req(TypeFunc::I_O,top); 2279 init_req(TypeFunc::FramePtr,top); 2280 init_req(TypeFunc::ReturnAdr,top); 2281 if (precedent != NULL) 2282 init_req(TypeFunc::Parms, precedent); 2283} 2284 2285//------------------------------cmp-------------------------------------------- 2286uint MemBarNode::hash() const { return NO_HASH; } 2287uint MemBarNode::cmp( const Node &n ) const { 2288 return (&n == this); // Always fail except on self 2289} 2290 2291//------------------------------make------------------------------------------- 2292MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2293 int len = Precedent + (pn == NULL? 0: 1); 2294 switch (opcode) { 2295 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); 2296 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); 2297 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); 2298 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); 2299 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); 2300 default: ShouldNotReachHere(); return NULL; 2301 } 2302} 2303 2304//------------------------------Ideal------------------------------------------ 2305// Return a node which is more "ideal" than the current node. Strip out 2306// control copies 2307Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2308 if (remove_dead_region(phase, can_reshape)) return this; 2309 return NULL; 2310} 2311 2312//------------------------------Value------------------------------------------ 2313const Type *MemBarNode::Value( PhaseTransform *phase ) const { 2314 if( !in(0) ) return Type::TOP; 2315 if( phase->type(in(0)) == Type::TOP ) 2316 return Type::TOP; 2317 return TypeTuple::MEMBAR; 2318} 2319 2320//------------------------------match------------------------------------------ 2321// Construct projections for memory. 2322Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 2323 switch (proj->_con) { 2324 case TypeFunc::Control: 2325 case TypeFunc::Memory: 2326 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 2327 } 2328 ShouldNotReachHere(); 2329 return NULL; 2330} 2331 2332//===========================InitializeNode==================================== 2333// SUMMARY: 2334// This node acts as a memory barrier on raw memory, after some raw stores. 2335// The 'cooked' oop value feeds from the Initialize, not the Allocation. 2336// The Initialize can 'capture' suitably constrained stores as raw inits. 2337// It can coalesce related raw stores into larger units (called 'tiles'). 2338// It can avoid zeroing new storage for memory units which have raw inits. 2339// At macro-expansion, it is marked 'complete', and does not optimize further. 2340// 2341// EXAMPLE: 2342// The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 2343// ctl = incoming control; mem* = incoming memory 2344// (Note: A star * on a memory edge denotes I/O and other standard edges.) 2345// First allocate uninitialized memory and fill in the header: 2346// alloc = (Allocate ctl mem* 16 #short[].klass ...) 2347// ctl := alloc.Control; mem* := alloc.Memory* 2348// rawmem = alloc.Memory; rawoop = alloc.RawAddress 2349// Then initialize to zero the non-header parts of the raw memory block: 2350// init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 2351// ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 2352// After the initialize node executes, the object is ready for service: 2353// oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 2354// Suppose its body is immediately initialized as {1,2}: 2355// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2356// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2357// mem.SLICE(#short[*]) := store2 2358// 2359// DETAILS: 2360// An InitializeNode collects and isolates object initialization after 2361// an AllocateNode and before the next possible safepoint. As a 2362// memory barrier (MemBarNode), it keeps critical stores from drifting 2363// down past any safepoint or any publication of the allocation. 2364// Before this barrier, a newly-allocated object may have uninitialized bits. 2365// After this barrier, it may be treated as a real oop, and GC is allowed. 2366// 2367// The semantics of the InitializeNode include an implicit zeroing of 2368// the new object from object header to the end of the object. 2369// (The object header and end are determined by the AllocateNode.) 2370// 2371// Certain stores may be added as direct inputs to the InitializeNode. 2372// These stores must update raw memory, and they must be to addresses 2373// derived from the raw address produced by AllocateNode, and with 2374// a constant offset. They must be ordered by increasing offset. 2375// The first one is at in(RawStores), the last at in(req()-1). 2376// Unlike most memory operations, they are not linked in a chain, 2377// but are displayed in parallel as users of the rawmem output of 2378// the allocation. 2379// 2380// (See comments in InitializeNode::capture_store, which continue 2381// the example given above.) 2382// 2383// When the associated Allocate is macro-expanded, the InitializeNode 2384// may be rewritten to optimize collected stores. A ClearArrayNode 2385// may also be created at that point to represent any required zeroing. 2386// The InitializeNode is then marked 'complete', prohibiting further 2387// capturing of nearby memory operations. 2388// 2389// During macro-expansion, all captured initializations which store 2390// constant values of 32 bits or smaller are coalesced (if advantagous) 2391// into larger 'tiles' 32 or 64 bits. This allows an object to be 2392// initialized in fewer memory operations. Memory words which are 2393// covered by neither tiles nor non-constant stores are pre-zeroed 2394// by explicit stores of zero. (The code shape happens to do all 2395// zeroing first, then all other stores, with both sequences occurring 2396// in order of ascending offsets.) 2397// 2398// Alternatively, code may be inserted between an AllocateNode and its 2399// InitializeNode, to perform arbitrary initialization of the new object. 2400// E.g., the object copying intrinsics insert complex data transfers here. 2401// The initialization must then be marked as 'complete' disable the 2402// built-in zeroing semantics and the collection of initializing stores. 2403// 2404// While an InitializeNode is incomplete, reads from the memory state 2405// produced by it are optimizable if they match the control edge and 2406// new oop address associated with the allocation/initialization. 2407// They return a stored value (if the offset matches) or else zero. 2408// A write to the memory state, if it matches control and address, 2409// and if it is to a constant offset, may be 'captured' by the 2410// InitializeNode. It is cloned as a raw memory operation and rewired 2411// inside the initialization, to the raw oop produced by the allocation. 2412// Operations on addresses which are provably distinct (e.g., to 2413// other AllocateNodes) are allowed to bypass the initialization. 2414// 2415// The effect of all this is to consolidate object initialization 2416// (both arrays and non-arrays, both piecewise and bulk) into a 2417// single location, where it can be optimized as a unit. 2418// 2419// Only stores with an offset less than TrackedInitializationLimit words 2420// will be considered for capture by an InitializeNode. This puts a 2421// reasonable limit on the complexity of optimized initializations. 2422 2423//---------------------------InitializeNode------------------------------------ 2424InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 2425 : _is_complete(false), 2426 MemBarNode(C, adr_type, rawoop) 2427{ 2428 init_class_id(Class_Initialize); 2429 2430 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 2431 assert(in(RawAddress) == rawoop, "proper init"); 2432 // Note: allocation() can be NULL, for secondary initialization barriers 2433} 2434 2435// Since this node is not matched, it will be processed by the 2436// register allocator. Declare that there are no constraints 2437// on the allocation of the RawAddress edge. 2438const RegMask &InitializeNode::in_RegMask(uint idx) const { 2439 // This edge should be set to top, by the set_complete. But be conservative. 2440 if (idx == InitializeNode::RawAddress) 2441 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 2442 return RegMask::Empty; 2443} 2444 2445Node* InitializeNode::memory(uint alias_idx) { 2446 Node* mem = in(Memory); 2447 if (mem->is_MergeMem()) { 2448 return mem->as_MergeMem()->memory_at(alias_idx); 2449 } else { 2450 // incoming raw memory is not split 2451 return mem; 2452 } 2453} 2454 2455bool InitializeNode::is_non_zero() { 2456 if (is_complete()) return false; 2457 remove_extra_zeroes(); 2458 return (req() > RawStores); 2459} 2460 2461void InitializeNode::set_complete(PhaseGVN* phase) { 2462 assert(!is_complete(), "caller responsibility"); 2463 _is_complete = true; 2464 2465 // After this node is complete, it contains a bunch of 2466 // raw-memory initializations. There is no need for 2467 // it to have anything to do with non-raw memory effects. 2468 // Therefore, tell all non-raw users to re-optimize themselves, 2469 // after skipping the memory effects of this initialization. 2470 PhaseIterGVN* igvn = phase->is_IterGVN(); 2471 if (igvn) igvn->add_users_to_worklist(this); 2472} 2473 2474// convenience function 2475// return false if the init contains any stores already 2476bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 2477 InitializeNode* init = initialization(); 2478 if (init == NULL || init->is_complete()) return false; 2479 init->remove_extra_zeroes(); 2480 // for now, if this allocation has already collected any inits, bail: 2481 if (init->is_non_zero()) return false; 2482 init->set_complete(phase); 2483 return true; 2484} 2485 2486void InitializeNode::remove_extra_zeroes() { 2487 if (req() == RawStores) return; 2488 Node* zmem = zero_memory(); 2489 uint fill = RawStores; 2490 for (uint i = fill; i < req(); i++) { 2491 Node* n = in(i); 2492 if (n->is_top() || n == zmem) continue; // skip 2493 if (fill < i) set_req(fill, n); // compact 2494 ++fill; 2495 } 2496 // delete any empty spaces created: 2497 while (fill < req()) { 2498 del_req(fill); 2499 } 2500} 2501 2502// Helper for remembering which stores go with which offsets. 2503intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 2504 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 2505 intptr_t offset = -1; 2506 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 2507 phase, offset); 2508 if (base == NULL) return -1; // something is dead, 2509 if (offset < 0) return -1; // dead, dead 2510 return offset; 2511} 2512 2513// Helper for proving that an initialization expression is 2514// "simple enough" to be folded into an object initialization. 2515// Attempts to prove that a store's initial value 'n' can be captured 2516// within the initialization without creating a vicious cycle, such as: 2517// { Foo p = new Foo(); p.next = p; } 2518// True for constants and parameters and small combinations thereof. 2519bool InitializeNode::detect_init_independence(Node* n, 2520 bool st_is_pinned, 2521 int& count) { 2522 if (n == NULL) return true; // (can this really happen?) 2523 if (n->is_Proj()) n = n->in(0); 2524 if (n == this) return false; // found a cycle 2525 if (n->is_Con()) return true; 2526 if (n->is_Start()) return true; // params, etc., are OK 2527 if (n->is_Root()) return true; // even better 2528 2529 Node* ctl = n->in(0); 2530 if (ctl != NULL && !ctl->is_top()) { 2531 if (ctl->is_Proj()) ctl = ctl->in(0); 2532 if (ctl == this) return false; 2533 2534 // If we already know that the enclosing memory op is pinned right after 2535 // the init, then any control flow that the store has picked up 2536 // must have preceded the init, or else be equal to the init. 2537 // Even after loop optimizations (which might change control edges) 2538 // a store is never pinned *before* the availability of its inputs. 2539 if (!MemNode::detect_dominating_control(ctl, this->in(0))) 2540 return false; // failed to prove a good control 2541 2542 } 2543 2544 // Check data edges for possible dependencies on 'this'. 2545 if ((count += 1) > 20) return false; // complexity limit 2546 for (uint i = 1; i < n->req(); i++) { 2547 Node* m = n->in(i); 2548 if (m == NULL || m == n || m->is_top()) continue; 2549 uint first_i = n->find_edge(m); 2550 if (i != first_i) continue; // process duplicate edge just once 2551 if (!detect_init_independence(m, st_is_pinned, count)) { 2552 return false; 2553 } 2554 } 2555 2556 return true; 2557} 2558 2559// Here are all the checks a Store must pass before it can be moved into 2560// an initialization. Returns zero if a check fails. 2561// On success, returns the (constant) offset to which the store applies, 2562// within the initialized memory. 2563intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { 2564 const int FAIL = 0; 2565 if (st->req() != MemNode::ValueIn + 1) 2566 return FAIL; // an inscrutable StoreNode (card mark?) 2567 Node* ctl = st->in(MemNode::Control); 2568 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 2569 return FAIL; // must be unconditional after the initialization 2570 Node* mem = st->in(MemNode::Memory); 2571 if (!(mem->is_Proj() && mem->in(0) == this)) 2572 return FAIL; // must not be preceded by other stores 2573 Node* adr = st->in(MemNode::Address); 2574 intptr_t offset; 2575 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 2576 if (alloc == NULL) 2577 return FAIL; // inscrutable address 2578 if (alloc != allocation()) 2579 return FAIL; // wrong allocation! (store needs to float up) 2580 Node* val = st->in(MemNode::ValueIn); 2581 int complexity_count = 0; 2582 if (!detect_init_independence(val, true, complexity_count)) 2583 return FAIL; // stored value must be 'simple enough' 2584 2585 return offset; // success 2586} 2587 2588// Find the captured store in(i) which corresponds to the range 2589// [start..start+size) in the initialized object. 2590// If there is one, return its index i. If there isn't, return the 2591// negative of the index where it should be inserted. 2592// Return 0 if the queried range overlaps an initialization boundary 2593// or if dead code is encountered. 2594// If size_in_bytes is zero, do not bother with overlap checks. 2595int InitializeNode::captured_store_insertion_point(intptr_t start, 2596 int size_in_bytes, 2597 PhaseTransform* phase) { 2598 const int FAIL = 0, MAX_STORE = BytesPerLong; 2599 2600 if (is_complete()) 2601 return FAIL; // arraycopy got here first; punt 2602 2603 assert(allocation() != NULL, "must be present"); 2604 2605 // no negatives, no header fields: 2606 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 2607 2608 // after a certain size, we bail out on tracking all the stores: 2609 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2610 if (start >= ti_limit) return FAIL; 2611 2612 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 2613 if (i >= limit) return -(int)i; // not found; here is where to put it 2614 2615 Node* st = in(i); 2616 intptr_t st_off = get_store_offset(st, phase); 2617 if (st_off < 0) { 2618 if (st != zero_memory()) { 2619 return FAIL; // bail out if there is dead garbage 2620 } 2621 } else if (st_off > start) { 2622 // ...we are done, since stores are ordered 2623 if (st_off < start + size_in_bytes) { 2624 return FAIL; // the next store overlaps 2625 } 2626 return -(int)i; // not found; here is where to put it 2627 } else if (st_off < start) { 2628 if (size_in_bytes != 0 && 2629 start < st_off + MAX_STORE && 2630 start < st_off + st->as_Store()->memory_size()) { 2631 return FAIL; // the previous store overlaps 2632 } 2633 } else { 2634 if (size_in_bytes != 0 && 2635 st->as_Store()->memory_size() != size_in_bytes) { 2636 return FAIL; // mismatched store size 2637 } 2638 return i; 2639 } 2640 2641 ++i; 2642 } 2643} 2644 2645// Look for a captured store which initializes at the offset 'start' 2646// with the given size. If there is no such store, and no other 2647// initialization interferes, then return zero_memory (the memory 2648// projection of the AllocateNode). 2649Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 2650 PhaseTransform* phase) { 2651 assert(stores_are_sane(phase), ""); 2652 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2653 if (i == 0) { 2654 return NULL; // something is dead 2655 } else if (i < 0) { 2656 return zero_memory(); // just primordial zero bits here 2657 } else { 2658 Node* st = in(i); // here is the store at this position 2659 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 2660 return st; 2661 } 2662} 2663 2664// Create, as a raw pointer, an address within my new object at 'offset'. 2665Node* InitializeNode::make_raw_address(intptr_t offset, 2666 PhaseTransform* phase) { 2667 Node* addr = in(RawAddress); 2668 if (offset != 0) { 2669 Compile* C = phase->C; 2670 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, 2671 phase->MakeConX(offset)) ); 2672 } 2673 return addr; 2674} 2675 2676// Clone the given store, converting it into a raw store 2677// initializing a field or element of my new object. 2678// Caller is responsible for retiring the original store, 2679// with subsume_node or the like. 2680// 2681// From the example above InitializeNode::InitializeNode, 2682// here are the old stores to be captured: 2683// store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 2684// store2 = (StoreC init.Control store1 (+ oop 14) 2) 2685// 2686// Here is the changed code; note the extra edges on init: 2687// alloc = (Allocate ...) 2688// rawoop = alloc.RawAddress 2689// rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 2690// rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 2691// init = (Initialize alloc.Control alloc.Memory rawoop 2692// rawstore1 rawstore2) 2693// 2694Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 2695 PhaseTransform* phase) { 2696 assert(stores_are_sane(phase), ""); 2697 2698 if (start < 0) return NULL; 2699 assert(can_capture_store(st, phase) == start, "sanity"); 2700 2701 Compile* C = phase->C; 2702 int size_in_bytes = st->memory_size(); 2703 int i = captured_store_insertion_point(start, size_in_bytes, phase); 2704 if (i == 0) return NULL; // bail out 2705 Node* prev_mem = NULL; // raw memory for the captured store 2706 if (i > 0) { 2707 prev_mem = in(i); // there is a pre-existing store under this one 2708 set_req(i, C->top()); // temporarily disconnect it 2709 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 2710 } else { 2711 i = -i; // no pre-existing store 2712 prev_mem = zero_memory(); // a slice of the newly allocated object 2713 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 2714 set_req(--i, C->top()); // reuse this edge; it has been folded away 2715 else 2716 ins_req(i, C->top()); // build a new edge 2717 } 2718 Node* new_st = st->clone(); 2719 new_st->set_req(MemNode::Control, in(Control)); 2720 new_st->set_req(MemNode::Memory, prev_mem); 2721 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 2722 new_st = phase->transform(new_st); 2723 2724 // At this point, new_st might have swallowed a pre-existing store 2725 // at the same offset, or perhaps new_st might have disappeared, 2726 // if it redundantly stored the same value (or zero to fresh memory). 2727 2728 // In any case, wire it in: 2729 set_req(i, new_st); 2730 2731 // The caller may now kill the old guy. 2732 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 2733 assert(check_st == new_st || check_st == NULL, "must be findable"); 2734 assert(!is_complete(), ""); 2735 return new_st; 2736} 2737 2738static bool store_constant(jlong* tiles, int num_tiles, 2739 intptr_t st_off, int st_size, 2740 jlong con) { 2741 if ((st_off & (st_size-1)) != 0) 2742 return false; // strange store offset (assume size==2**N) 2743 address addr = (address)tiles + st_off; 2744 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 2745 switch (st_size) { 2746 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 2747 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 2748 case sizeof(jint): *(jint*) addr = (jint) con; break; 2749 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 2750 default: return false; // strange store size (detect size!=2**N here) 2751 } 2752 return true; // return success to caller 2753} 2754 2755// Coalesce subword constants into int constants and possibly 2756// into long constants. The goal, if the CPU permits, 2757// is to initialize the object with a small number of 64-bit tiles. 2758// Also, convert floating-point constants to bit patterns. 2759// Non-constants are not relevant to this pass. 2760// 2761// In terms of the running example on InitializeNode::InitializeNode 2762// and InitializeNode::capture_store, here is the transformation 2763// of rawstore1 and rawstore2 into rawstore12: 2764// alloc = (Allocate ...) 2765// rawoop = alloc.RawAddress 2766// tile12 = 0x00010002 2767// rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 2768// init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 2769// 2770void 2771InitializeNode::coalesce_subword_stores(intptr_t header_size, 2772 Node* size_in_bytes, 2773 PhaseGVN* phase) { 2774 Compile* C = phase->C; 2775 2776 assert(stores_are_sane(phase), ""); 2777 // Note: After this pass, they are not completely sane, 2778 // since there may be some overlaps. 2779 2780 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 2781 2782 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 2783 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 2784 size_limit = MIN2(size_limit, ti_limit); 2785 size_limit = align_size_up(size_limit, BytesPerLong); 2786 int num_tiles = size_limit / BytesPerLong; 2787 2788 // allocate space for the tile map: 2789 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 2790 jlong tiles_buf[small_len]; 2791 Node* nodes_buf[small_len]; 2792 jlong inits_buf[small_len]; 2793 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 2794 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2795 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 2796 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 2797 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 2798 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 2799 // tiles: exact bitwise model of all primitive constants 2800 // nodes: last constant-storing node subsumed into the tiles model 2801 // inits: which bytes (in each tile) are touched by any initializations 2802 2803 //// Pass A: Fill in the tile model with any relevant stores. 2804 2805 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 2806 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 2807 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 2808 Node* zmem = zero_memory(); // initially zero memory state 2809 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 2810 Node* st = in(i); 2811 intptr_t st_off = get_store_offset(st, phase); 2812 2813 // Figure out the store's offset and constant value: 2814 if (st_off < header_size) continue; //skip (ignore header) 2815 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 2816 int st_size = st->as_Store()->memory_size(); 2817 if (st_off + st_size > size_limit) break; 2818 2819 // Record which bytes are touched, whether by constant or not. 2820 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 2821 continue; // skip (strange store size) 2822 2823 const Type* val = phase->type(st->in(MemNode::ValueIn)); 2824 if (!val->singleton()) continue; //skip (non-con store) 2825 BasicType type = val->basic_type(); 2826 2827 jlong con = 0; 2828 switch (type) { 2829 case T_INT: con = val->is_int()->get_con(); break; 2830 case T_LONG: con = val->is_long()->get_con(); break; 2831 case T_FLOAT: con = jint_cast(val->getf()); break; 2832 case T_DOUBLE: con = jlong_cast(val->getd()); break; 2833 default: continue; //skip (odd store type) 2834 } 2835 2836 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 2837 st->Opcode() == Op_StoreL) { 2838 continue; // This StoreL is already optimal. 2839 } 2840 2841 // Store down the constant. 2842 store_constant(tiles, num_tiles, st_off, st_size, con); 2843 2844 intptr_t j = st_off >> LogBytesPerLong; 2845 2846 if (type == T_INT && st_size == BytesPerInt 2847 && (st_off & BytesPerInt) == BytesPerInt) { 2848 jlong lcon = tiles[j]; 2849 if (!Matcher::isSimpleConstant64(lcon) && 2850 st->Opcode() == Op_StoreI) { 2851 // This StoreI is already optimal by itself. 2852 jint* intcon = (jint*) &tiles[j]; 2853 intcon[1] = 0; // undo the store_constant() 2854 2855 // If the previous store is also optimal by itself, back up and 2856 // undo the action of the previous loop iteration... if we can. 2857 // But if we can't, just let the previous half take care of itself. 2858 st = nodes[j]; 2859 st_off -= BytesPerInt; 2860 con = intcon[0]; 2861 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 2862 assert(st_off >= header_size, "still ignoring header"); 2863 assert(get_store_offset(st, phase) == st_off, "must be"); 2864 assert(in(i-1) == zmem, "must be"); 2865 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 2866 assert(con == tcon->is_int()->get_con(), "must be"); 2867 // Undo the effects of the previous loop trip, which swallowed st: 2868 intcon[0] = 0; // undo store_constant() 2869 set_req(i-1, st); // undo set_req(i, zmem) 2870 nodes[j] = NULL; // undo nodes[j] = st 2871 --old_subword; // undo ++old_subword 2872 } 2873 continue; // This StoreI is already optimal. 2874 } 2875 } 2876 2877 // This store is not needed. 2878 set_req(i, zmem); 2879 nodes[j] = st; // record for the moment 2880 if (st_size < BytesPerLong) // something has changed 2881 ++old_subword; // includes int/float, but who's counting... 2882 else ++old_long; 2883 } 2884 2885 if ((old_subword + old_long) == 0) 2886 return; // nothing more to do 2887 2888 //// Pass B: Convert any non-zero tiles into optimal constant stores. 2889 // Be sure to insert them before overlapping non-constant stores. 2890 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 2891 for (int j = 0; j < num_tiles; j++) { 2892 jlong con = tiles[j]; 2893 jlong init = inits[j]; 2894 if (con == 0) continue; 2895 jint con0, con1; // split the constant, address-wise 2896 jint init0, init1; // split the init map, address-wise 2897 { union { jlong con; jint intcon[2]; } u; 2898 u.con = con; 2899 con0 = u.intcon[0]; 2900 con1 = u.intcon[1]; 2901 u.con = init; 2902 init0 = u.intcon[0]; 2903 init1 = u.intcon[1]; 2904 } 2905 2906 Node* old = nodes[j]; 2907 assert(old != NULL, "need the prior store"); 2908 intptr_t offset = (j * BytesPerLong); 2909 2910 bool split = !Matcher::isSimpleConstant64(con); 2911 2912 if (offset < header_size) { 2913 assert(offset + BytesPerInt >= header_size, "second int counts"); 2914 assert(*(jint*)&tiles[j] == 0, "junk in header"); 2915 split = true; // only the second word counts 2916 // Example: int a[] = { 42 ... } 2917 } else if (con0 == 0 && init0 == -1) { 2918 split = true; // first word is covered by full inits 2919 // Example: int a[] = { ... foo(), 42 ... } 2920 } else if (con1 == 0 && init1 == -1) { 2921 split = true; // second word is covered by full inits 2922 // Example: int a[] = { ... 42, foo() ... } 2923 } 2924 2925 // Here's a case where init0 is neither 0 nor -1: 2926 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 2927 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 2928 // In this case the tile is not split; it is (jlong)42. 2929 // The big tile is stored down, and then the foo() value is inserted. 2930 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 2931 2932 Node* ctl = old->in(MemNode::Control); 2933 Node* adr = make_raw_address(offset, phase); 2934 const TypePtr* atp = TypeRawPtr::BOTTOM; 2935 2936 // One or two coalesced stores to plop down. 2937 Node* st[2]; 2938 intptr_t off[2]; 2939 int nst = 0; 2940 if (!split) { 2941 ++new_long; 2942 off[nst] = offset; 2943 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 2944 phase->longcon(con), T_LONG); 2945 } else { 2946 // Omit either if it is a zero. 2947 if (con0 != 0) { 2948 ++new_int; 2949 off[nst] = offset; 2950 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 2951 phase->intcon(con0), T_INT); 2952 } 2953 if (con1 != 0) { 2954 ++new_int; 2955 offset += BytesPerInt; 2956 adr = make_raw_address(offset, phase); 2957 off[nst] = offset; 2958 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 2959 phase->intcon(con1), T_INT); 2960 } 2961 } 2962 2963 // Insert second store first, then the first before the second. 2964 // Insert each one just before any overlapping non-constant stores. 2965 while (nst > 0) { 2966 Node* st1 = st[--nst]; 2967 C->copy_node_notes_to(st1, old); 2968 st1 = phase->transform(st1); 2969 offset = off[nst]; 2970 assert(offset >= header_size, "do not smash header"); 2971 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 2972 guarantee(ins_idx != 0, "must re-insert constant store"); 2973 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 2974 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 2975 set_req(--ins_idx, st1); 2976 else 2977 ins_req(ins_idx, st1); 2978 } 2979 } 2980 2981 if (PrintCompilation && WizardMode) 2982 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 2983 old_subword, old_long, new_int, new_long); 2984 if (C->log() != NULL) 2985 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 2986 old_subword, old_long, new_int, new_long); 2987 2988 // Clean up any remaining occurrences of zmem: 2989 remove_extra_zeroes(); 2990} 2991 2992// Explore forward from in(start) to find the first fully initialized 2993// word, and return its offset. Skip groups of subword stores which 2994// together initialize full words. If in(start) is itself part of a 2995// fully initialized word, return the offset of in(start). If there 2996// are no following full-word stores, or if something is fishy, return 2997// a negative value. 2998intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 2999 int int_map = 0; 3000 intptr_t int_map_off = 0; 3001 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3002 3003 for (uint i = start, limit = req(); i < limit; i++) { 3004 Node* st = in(i); 3005 3006 intptr_t st_off = get_store_offset(st, phase); 3007 if (st_off < 0) break; // return conservative answer 3008 3009 int st_size = st->as_Store()->memory_size(); 3010 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3011 return st_off; // we found a complete word init 3012 } 3013 3014 // update the map: 3015 3016 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3017 if (this_int_off != int_map_off) { 3018 // reset the map: 3019 int_map = 0; 3020 int_map_off = this_int_off; 3021 } 3022 3023 int subword_off = st_off - this_int_off; 3024 int_map |= right_n_bits(st_size) << subword_off; 3025 if ((int_map & FULL_MAP) == FULL_MAP) { 3026 return this_int_off; // we found a complete word init 3027 } 3028 3029 // Did this store hit or cross the word boundary? 3030 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3031 if (next_int_off == this_int_off + BytesPerInt) { 3032 // We passed the current int, without fully initializing it. 3033 int_map_off = next_int_off; 3034 int_map >>= BytesPerInt; 3035 } else if (next_int_off > this_int_off + BytesPerInt) { 3036 // We passed the current and next int. 3037 return this_int_off + BytesPerInt; 3038 } 3039 } 3040 3041 return -1; 3042} 3043 3044 3045// Called when the associated AllocateNode is expanded into CFG. 3046// At this point, we may perform additional optimizations. 3047// Linearize the stores by ascending offset, to make memory 3048// activity as coherent as possible. 3049Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3050 intptr_t header_size, 3051 Node* size_in_bytes, 3052 PhaseGVN* phase) { 3053 assert(!is_complete(), "not already complete"); 3054 assert(stores_are_sane(phase), ""); 3055 assert(allocation() != NULL, "must be present"); 3056 3057 remove_extra_zeroes(); 3058 3059 if (ReduceFieldZeroing || ReduceBulkZeroing) 3060 // reduce instruction count for common initialization patterns 3061 coalesce_subword_stores(header_size, size_in_bytes, phase); 3062 3063 Node* zmem = zero_memory(); // initially zero memory state 3064 Node* inits = zmem; // accumulating a linearized chain of inits 3065 #ifdef ASSERT 3066 intptr_t first_offset = allocation()->minimum_header_size(); 3067 intptr_t last_init_off = first_offset; // previous init offset 3068 intptr_t last_init_end = first_offset; // previous init offset+size 3069 intptr_t last_tile_end = first_offset; // previous tile offset+size 3070 #endif 3071 intptr_t zeroes_done = header_size; 3072 3073 bool do_zeroing = true; // we might give up if inits are very sparse 3074 int big_init_gaps = 0; // how many large gaps have we seen? 3075 3076 if (ZeroTLAB) do_zeroing = false; 3077 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3078 3079 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3080 Node* st = in(i); 3081 intptr_t st_off = get_store_offset(st, phase); 3082 if (st_off < 0) 3083 break; // unknown junk in the inits 3084 if (st->in(MemNode::Memory) != zmem) 3085 break; // complicated store chains somehow in list 3086 3087 int st_size = st->as_Store()->memory_size(); 3088 intptr_t next_init_off = st_off + st_size; 3089 3090 if (do_zeroing && zeroes_done < next_init_off) { 3091 // See if this store needs a zero before it or under it. 3092 intptr_t zeroes_needed = st_off; 3093 3094 if (st_size < BytesPerInt) { 3095 // Look for subword stores which only partially initialize words. 3096 // If we find some, we must lay down some word-level zeroes first, 3097 // underneath the subword stores. 3098 // 3099 // Examples: 3100 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3101 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3102 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3103 // 3104 // Note: coalesce_subword_stores may have already done this, 3105 // if it was prompted by constant non-zero subword initializers. 3106 // But this case can still arise with non-constant stores. 3107 3108 intptr_t next_full_store = find_next_fullword_store(i, phase); 3109 3110 // In the examples above: 3111 // in(i) p q r s x y z 3112 // st_off 12 13 14 15 12 13 14 3113 // st_size 1 1 1 1 1 1 1 3114 // next_full_s. 12 16 16 16 16 16 16 3115 // z's_done 12 16 16 16 12 16 12 3116 // z's_needed 12 16 16 16 16 16 16 3117 // zsize 0 0 0 0 4 0 4 3118 if (next_full_store < 0) { 3119 // Conservative tack: Zero to end of current word. 3120 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3121 } else { 3122 // Zero to beginning of next fully initialized word. 3123 // Or, don't zero at all, if we are already in that word. 3124 assert(next_full_store >= zeroes_needed, "must go forward"); 3125 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3126 zeroes_needed = next_full_store; 3127 } 3128 } 3129 3130 if (zeroes_needed > zeroes_done) { 3131 intptr_t zsize = zeroes_needed - zeroes_done; 3132 // Do some incremental zeroing on rawmem, in parallel with inits. 3133 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3134 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3135 zeroes_done, zeroes_needed, 3136 phase); 3137 zeroes_done = zeroes_needed; 3138 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3139 do_zeroing = false; // leave the hole, next time 3140 } 3141 } 3142 3143 // Collect the store and move on: 3144 st->set_req(MemNode::Memory, inits); 3145 inits = st; // put it on the linearized chain 3146 set_req(i, zmem); // unhook from previous position 3147 3148 if (zeroes_done == st_off) 3149 zeroes_done = next_init_off; 3150 3151 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3152 3153 #ifdef ASSERT 3154 // Various order invariants. Weaker than stores_are_sane because 3155 // a large constant tile can be filled in by smaller non-constant stores. 3156 assert(st_off >= last_init_off, "inits do not reverse"); 3157 last_init_off = st_off; 3158 const Type* val = NULL; 3159 if (st_size >= BytesPerInt && 3160 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3161 (int)val->basic_type() < (int)T_OBJECT) { 3162 assert(st_off >= last_tile_end, "tiles do not overlap"); 3163 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3164 last_tile_end = MAX2(last_tile_end, next_init_off); 3165 } else { 3166 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3167 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3168 assert(st_off >= last_init_end, "inits do not overlap"); 3169 last_init_end = next_init_off; // it's a non-tile 3170 } 3171 #endif //ASSERT 3172 } 3173 3174 remove_extra_zeroes(); // clear out all the zmems left over 3175 add_req(inits); 3176 3177 if (!ZeroTLAB) { 3178 // If anything remains to be zeroed, zero it all now. 3179 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3180 // if it is the last unused 4 bytes of an instance, forget about it 3181 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3182 if (zeroes_done + BytesPerLong >= size_limit) { 3183 assert(allocation() != NULL, ""); 3184 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3185 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3186 if (zeroes_done == k->layout_helper()) 3187 zeroes_done = size_limit; 3188 } 3189 if (zeroes_done < size_limit) { 3190 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3191 zeroes_done, size_in_bytes, phase); 3192 } 3193 } 3194 3195 set_complete(phase); 3196 return rawmem; 3197} 3198 3199 3200#ifdef ASSERT 3201bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3202 if (is_complete()) 3203 return true; // stores could be anything at this point 3204 assert(allocation() != NULL, "must be present"); 3205 intptr_t last_off = allocation()->minimum_header_size(); 3206 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3207 Node* st = in(i); 3208 intptr_t st_off = get_store_offset(st, phase); 3209 if (st_off < 0) continue; // ignore dead garbage 3210 if (last_off > st_off) { 3211 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); 3212 this->dump(2); 3213 assert(false, "ascending store offsets"); 3214 return false; 3215 } 3216 last_off = st_off + st->as_Store()->memory_size(); 3217 } 3218 return true; 3219} 3220#endif //ASSERT 3221 3222 3223 3224 3225//============================MergeMemNode===================================== 3226// 3227// SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 3228// contributing store or call operations. Each contributor provides the memory 3229// state for a particular "alias type" (see Compile::alias_type). For example, 3230// if a MergeMem has an input X for alias category #6, then any memory reference 3231// to alias category #6 may use X as its memory state input, as an exact equivalent 3232// to using the MergeMem as a whole. 3233// Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 3234// 3235// (Here, the <N> notation gives the index of the relevant adr_type.) 3236// 3237// In one special case (and more cases in the future), alias categories overlap. 3238// The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 3239// states. Therefore, if a MergeMem has only one contributing input W for Bot, 3240// it is exactly equivalent to that state W: 3241// MergeMem(<Bot>: W) <==> W 3242// 3243// Usually, the merge has more than one input. In that case, where inputs 3244// overlap (i.e., one is Bot), the narrower alias type determines the memory 3245// state for that type, and the wider alias type (Bot) fills in everywhere else: 3246// Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 3247// Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 3248// 3249// A merge can take a "wide" memory state as one of its narrow inputs. 3250// This simply means that the merge observes out only the relevant parts of 3251// the wide input. That is, wide memory states arriving at narrow merge inputs 3252// are implicitly "filtered" or "sliced" as necessary. (This is rare.) 3253// 3254// These rules imply that MergeMem nodes may cascade (via their <Bot> links), 3255// and that memory slices "leak through": 3256// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 3257// 3258// But, in such a cascade, repeated memory slices can "block the leak": 3259// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 3260// 3261// In the last example, Y is not part of the combined memory state of the 3262// outermost MergeMem. The system must, of course, prevent unschedulable 3263// memory states from arising, so you can be sure that the state Y is somehow 3264// a precursor to state Y'. 3265// 3266// 3267// REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 3268// of each MergeMemNode array are exactly the numerical alias indexes, including 3269// but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 3270// Compile::alias_type (and kin) produce and manage these indexes. 3271// 3272// By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 3273// (Note that this provides quick access to the top node inside MergeMem methods, 3274// without the need to reach out via TLS to Compile::current.) 3275// 3276// As a consequence of what was just described, a MergeMem that represents a full 3277// memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 3278// containing all alias categories. 3279// 3280// MergeMem nodes never (?) have control inputs, so in(0) is NULL. 3281// 3282// All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 3283// a memory state for the alias type <N>, or else the top node, meaning that 3284// there is no particular input for that alias type. Note that the length of 3285// a MergeMem is variable, and may be extended at any time to accommodate new 3286// memory states at larger alias indexes. When merges grow, they are of course 3287// filled with "top" in the unused in() positions. 3288// 3289// This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 3290// (Top was chosen because it works smoothly with passes like GCM.) 3291// 3292// For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 3293// the type of random VM bits like TLS references.) Since it is always the 3294// first non-Bot memory slice, some low-level loops use it to initialize an 3295// index variable: for (i = AliasIdxRaw; i < req(); i++). 3296// 3297// 3298// ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 3299// the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 3300// the memory state for alias type <N>, or (if there is no particular slice at <N>, 3301// it returns the base memory. To prevent bugs, memory_at does not accept <Top> 3302// or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 3303// MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 3304// 3305// %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 3306// really that different from the other memory inputs. An abbreviation called 3307// "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 3308// 3309// 3310// PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 3311// partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 3312// that "emerges though" the base memory will be marked as excluding the alias types 3313// of the other (narrow-memory) copies which "emerged through" the narrow edges: 3314// 3315// Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 3316// ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 3317// 3318// This strange "subtraction" effect is necessary to ensure IGVN convergence. 3319// (It is currently unimplemented.) As you can see, the resulting merge is 3320// actually a disjoint union of memory states, rather than an overlay. 3321// 3322 3323//------------------------------MergeMemNode----------------------------------- 3324Node* MergeMemNode::make_empty_memory() { 3325 Node* empty_memory = (Node*) Compile::current()->top(); 3326 assert(empty_memory->is_top(), "correct sentinel identity"); 3327 return empty_memory; 3328} 3329 3330MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 3331 init_class_id(Class_MergeMem); 3332 // all inputs are nullified in Node::Node(int) 3333 // set_input(0, NULL); // no control input 3334 3335 // Initialize the edges uniformly to top, for starters. 3336 Node* empty_mem = make_empty_memory(); 3337 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 3338 init_req(i,empty_mem); 3339 } 3340 assert(empty_memory() == empty_mem, ""); 3341 3342 if( new_base != NULL && new_base->is_MergeMem() ) { 3343 MergeMemNode* mdef = new_base->as_MergeMem(); 3344 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 3345 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 3346 mms.set_memory(mms.memory2()); 3347 } 3348 assert(base_memory() == mdef->base_memory(), ""); 3349 } else { 3350 set_base_memory(new_base); 3351 } 3352} 3353 3354// Make a new, untransformed MergeMem with the same base as 'mem'. 3355// If mem is itself a MergeMem, populate the result with the same edges. 3356MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 3357 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); 3358} 3359 3360//------------------------------cmp-------------------------------------------- 3361uint MergeMemNode::hash() const { return NO_HASH; } 3362uint MergeMemNode::cmp( const Node &n ) const { 3363 return (&n == this); // Always fail except on self 3364} 3365 3366//------------------------------Identity--------------------------------------- 3367Node* MergeMemNode::Identity(PhaseTransform *phase) { 3368 // Identity if this merge point does not record any interesting memory 3369 // disambiguations. 3370 Node* base_mem = base_memory(); 3371 Node* empty_mem = empty_memory(); 3372 if (base_mem != empty_mem) { // Memory path is not dead? 3373 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3374 Node* mem = in(i); 3375 if (mem != empty_mem && mem != base_mem) { 3376 return this; // Many memory splits; no change 3377 } 3378 } 3379 } 3380 return base_mem; // No memory splits; ID on the one true input 3381} 3382 3383//------------------------------Ideal------------------------------------------ 3384// This method is invoked recursively on chains of MergeMem nodes 3385Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3386 // Remove chain'd MergeMems 3387 // 3388 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 3389 // relative to the "in(Bot)". Since we are patching both at the same time, 3390 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 3391 // but rewrite each "in(i)" relative to the new "in(Bot)". 3392 Node *progress = NULL; 3393 3394 3395 Node* old_base = base_memory(); 3396 Node* empty_mem = empty_memory(); 3397 if (old_base == empty_mem) 3398 return NULL; // Dead memory path. 3399 3400 MergeMemNode* old_mbase; 3401 if (old_base != NULL && old_base->is_MergeMem()) 3402 old_mbase = old_base->as_MergeMem(); 3403 else 3404 old_mbase = NULL; 3405 Node* new_base = old_base; 3406 3407 // simplify stacked MergeMems in base memory 3408 if (old_mbase) new_base = old_mbase->base_memory(); 3409 3410 // the base memory might contribute new slices beyond my req() 3411 if (old_mbase) grow_to_match(old_mbase); 3412 3413 // Look carefully at the base node if it is a phi. 3414 PhiNode* phi_base; 3415 if (new_base != NULL && new_base->is_Phi()) 3416 phi_base = new_base->as_Phi(); 3417 else 3418 phi_base = NULL; 3419 3420 Node* phi_reg = NULL; 3421 uint phi_len = (uint)-1; 3422 if (phi_base != NULL && !phi_base->is_copy()) { 3423 // do not examine phi if degraded to a copy 3424 phi_reg = phi_base->region(); 3425 phi_len = phi_base->req(); 3426 // see if the phi is unfinished 3427 for (uint i = 1; i < phi_len; i++) { 3428 if (phi_base->in(i) == NULL) { 3429 // incomplete phi; do not look at it yet! 3430 phi_reg = NULL; 3431 phi_len = (uint)-1; 3432 break; 3433 } 3434 } 3435 } 3436 3437 // Note: We do not call verify_sparse on entry, because inputs 3438 // can normalize to the base_memory via subsume_node or similar 3439 // mechanisms. This method repairs that damage. 3440 3441 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 3442 3443 // Look at each slice. 3444 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3445 Node* old_in = in(i); 3446 // calculate the old memory value 3447 Node* old_mem = old_in; 3448 if (old_mem == empty_mem) old_mem = old_base; 3449 assert(old_mem == memory_at(i), ""); 3450 3451 // maybe update (reslice) the old memory value 3452 3453 // simplify stacked MergeMems 3454 Node* new_mem = old_mem; 3455 MergeMemNode* old_mmem; 3456 if (old_mem != NULL && old_mem->is_MergeMem()) 3457 old_mmem = old_mem->as_MergeMem(); 3458 else 3459 old_mmem = NULL; 3460 if (old_mmem == this) { 3461 // This can happen if loops break up and safepoints disappear. 3462 // A merge of BotPtr (default) with a RawPtr memory derived from a 3463 // safepoint can be rewritten to a merge of the same BotPtr with 3464 // the BotPtr phi coming into the loop. If that phi disappears 3465 // also, we can end up with a self-loop of the mergemem. 3466 // In general, if loops degenerate and memory effects disappear, 3467 // a mergemem can be left looking at itself. This simply means 3468 // that the mergemem's default should be used, since there is 3469 // no longer any apparent effect on this slice. 3470 // Note: If a memory slice is a MergeMem cycle, it is unreachable 3471 // from start. Update the input to TOP. 3472 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 3473 } 3474 else if (old_mmem != NULL) { 3475 new_mem = old_mmem->memory_at(i); 3476 } 3477 // else preceeding memory was not a MergeMem 3478 3479 // replace equivalent phis (unfortunately, they do not GVN together) 3480 if (new_mem != NULL && new_mem != new_base && 3481 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 3482 if (new_mem->is_Phi()) { 3483 PhiNode* phi_mem = new_mem->as_Phi(); 3484 for (uint i = 1; i < phi_len; i++) { 3485 if (phi_base->in(i) != phi_mem->in(i)) { 3486 phi_mem = NULL; 3487 break; 3488 } 3489 } 3490 if (phi_mem != NULL) { 3491 // equivalent phi nodes; revert to the def 3492 new_mem = new_base; 3493 } 3494 } 3495 } 3496 3497 // maybe store down a new value 3498 Node* new_in = new_mem; 3499 if (new_in == new_base) new_in = empty_mem; 3500 3501 if (new_in != old_in) { 3502 // Warning: Do not combine this "if" with the previous "if" 3503 // A memory slice might have be be rewritten even if it is semantically 3504 // unchanged, if the base_memory value has changed. 3505 set_req(i, new_in); 3506 progress = this; // Report progress 3507 } 3508 } 3509 3510 if (new_base != old_base) { 3511 set_req(Compile::AliasIdxBot, new_base); 3512 // Don't use set_base_memory(new_base), because we need to update du. 3513 assert(base_memory() == new_base, ""); 3514 progress = this; 3515 } 3516 3517 if( base_memory() == this ) { 3518 // a self cycle indicates this memory path is dead 3519 set_req(Compile::AliasIdxBot, empty_mem); 3520 } 3521 3522 // Resolve external cycles by calling Ideal on a MergeMem base_memory 3523 // Recursion must occur after the self cycle check above 3524 if( base_memory()->is_MergeMem() ) { 3525 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 3526 Node *m = phase->transform(new_mbase); // Rollup any cycles 3527 if( m != NULL && (m->is_top() || 3528 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 3529 // propagate rollup of dead cycle to self 3530 set_req(Compile::AliasIdxBot, empty_mem); 3531 } 3532 } 3533 3534 if( base_memory() == empty_mem ) { 3535 progress = this; 3536 // Cut inputs during Parse phase only. 3537 // During Optimize phase a dead MergeMem node will be subsumed by Top. 3538 if( !can_reshape ) { 3539 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3540 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 3541 } 3542 } 3543 } 3544 3545 if( !progress && base_memory()->is_Phi() && can_reshape ) { 3546 // Check if PhiNode::Ideal's "Split phis through memory merges" 3547 // transform should be attempted. Look for this->phi->this cycle. 3548 uint merge_width = req(); 3549 if (merge_width > Compile::AliasIdxRaw) { 3550 PhiNode* phi = base_memory()->as_Phi(); 3551 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 3552 if (phi->in(i) == this) { 3553 phase->is_IterGVN()->_worklist.push(phi); 3554 break; 3555 } 3556 } 3557 } 3558 } 3559 3560 assert(progress || verify_sparse(), "please, no dups of base"); 3561 return progress; 3562} 3563 3564//-------------------------set_base_memory------------------------------------- 3565void MergeMemNode::set_base_memory(Node *new_base) { 3566 Node* empty_mem = empty_memory(); 3567 set_req(Compile::AliasIdxBot, new_base); 3568 assert(memory_at(req()) == new_base, "must set default memory"); 3569 // Clear out other occurrences of new_base: 3570 if (new_base != empty_mem) { 3571 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3572 if (in(i) == new_base) set_req(i, empty_mem); 3573 } 3574 } 3575} 3576 3577//------------------------------out_RegMask------------------------------------ 3578const RegMask &MergeMemNode::out_RegMask() const { 3579 return RegMask::Empty; 3580} 3581 3582//------------------------------dump_spec-------------------------------------- 3583#ifndef PRODUCT 3584void MergeMemNode::dump_spec(outputStream *st) const { 3585 st->print(" {"); 3586 Node* base_mem = base_memory(); 3587 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 3588 Node* mem = memory_at(i); 3589 if (mem == base_mem) { st->print(" -"); continue; } 3590 st->print( " N%d:", mem->_idx ); 3591 Compile::current()->get_adr_type(i)->dump_on(st); 3592 } 3593 st->print(" }"); 3594} 3595#endif // !PRODUCT 3596 3597 3598#ifdef ASSERT 3599static bool might_be_same(Node* a, Node* b) { 3600 if (a == b) return true; 3601 if (!(a->is_Phi() || b->is_Phi())) return false; 3602 // phis shift around during optimization 3603 return true; // pretty stupid... 3604} 3605 3606// verify a narrow slice (either incoming or outgoing) 3607static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 3608 if (!VerifyAliases) return; // don't bother to verify unless requested 3609 if (is_error_reported()) return; // muzzle asserts when debugging an error 3610 if (Node::in_dump()) return; // muzzle asserts when printing 3611 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 3612 assert(n != NULL, ""); 3613 // Elide intervening MergeMem's 3614 while (n->is_MergeMem()) { 3615 n = n->as_MergeMem()->memory_at(alias_idx); 3616 } 3617 Compile* C = Compile::current(); 3618 const TypePtr* n_adr_type = n->adr_type(); 3619 if (n == m->empty_memory()) { 3620 // Implicit copy of base_memory() 3621 } else if (n_adr_type != TypePtr::BOTTOM) { 3622 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 3623 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 3624 } else { 3625 // A few places like make_runtime_call "know" that VM calls are narrow, 3626 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 3627 bool expected_wide_mem = false; 3628 if (n == m->base_memory()) { 3629 expected_wide_mem = true; 3630 } else if (alias_idx == Compile::AliasIdxRaw || 3631 n == m->memory_at(Compile::AliasIdxRaw)) { 3632 expected_wide_mem = true; 3633 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 3634 // memory can "leak through" calls on channels that 3635 // are write-once. Allow this also. 3636 expected_wide_mem = true; 3637 } 3638 assert(expected_wide_mem, "expected narrow slice replacement"); 3639 } 3640} 3641#else // !ASSERT 3642#define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op 3643#endif 3644 3645 3646//-----------------------------memory_at--------------------------------------- 3647Node* MergeMemNode::memory_at(uint alias_idx) const { 3648 assert(alias_idx >= Compile::AliasIdxRaw || 3649 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 3650 "must avoid base_memory and AliasIdxTop"); 3651 3652 // Otherwise, it is a narrow slice. 3653 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 3654 Compile *C = Compile::current(); 3655 if (is_empty_memory(n)) { 3656 // the array is sparse; empty slots are the "top" node 3657 n = base_memory(); 3658 assert(Node::in_dump() 3659 || n == NULL || n->bottom_type() == Type::TOP 3660 || n->adr_type() == TypePtr::BOTTOM 3661 || n->adr_type() == TypeRawPtr::BOTTOM 3662 || Compile::current()->AliasLevel() == 0, 3663 "must be a wide memory"); 3664 // AliasLevel == 0 if we are organizing the memory states manually. 3665 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 3666 } else { 3667 // make sure the stored slice is sane 3668 #ifdef ASSERT 3669 if (is_error_reported() || Node::in_dump()) { 3670 } else if (might_be_same(n, base_memory())) { 3671 // Give it a pass: It is a mostly harmless repetition of the base. 3672 // This can arise normally from node subsumption during optimization. 3673 } else { 3674 verify_memory_slice(this, alias_idx, n); 3675 } 3676 #endif 3677 } 3678 return n; 3679} 3680 3681//---------------------------set_memory_at------------------------------------- 3682void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 3683 verify_memory_slice(this, alias_idx, n); 3684 Node* empty_mem = empty_memory(); 3685 if (n == base_memory()) n = empty_mem; // collapse default 3686 uint need_req = alias_idx+1; 3687 if (req() < need_req) { 3688 if (n == empty_mem) return; // already the default, so do not grow me 3689 // grow the sparse array 3690 do { 3691 add_req(empty_mem); 3692 } while (req() < need_req); 3693 } 3694 set_req( alias_idx, n ); 3695} 3696 3697 3698 3699//--------------------------iteration_setup------------------------------------ 3700void MergeMemNode::iteration_setup(const MergeMemNode* other) { 3701 if (other != NULL) { 3702 grow_to_match(other); 3703 // invariant: the finite support of mm2 is within mm->req() 3704 #ifdef ASSERT 3705 for (uint i = req(); i < other->req(); i++) { 3706 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 3707 } 3708 #endif 3709 } 3710 // Replace spurious copies of base_memory by top. 3711 Node* base_mem = base_memory(); 3712 if (base_mem != NULL && !base_mem->is_top()) { 3713 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 3714 if (in(i) == base_mem) 3715 set_req(i, empty_memory()); 3716 } 3717 } 3718} 3719 3720//---------------------------grow_to_match------------------------------------- 3721void MergeMemNode::grow_to_match(const MergeMemNode* other) { 3722 Node* empty_mem = empty_memory(); 3723 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 3724 // look for the finite support of the other memory 3725 for (uint i = other->req(); --i >= req(); ) { 3726 if (other->in(i) != empty_mem) { 3727 uint new_len = i+1; 3728 while (req() < new_len) add_req(empty_mem); 3729 break; 3730 } 3731 } 3732} 3733 3734//---------------------------verify_sparse------------------------------------- 3735#ifndef PRODUCT 3736bool MergeMemNode::verify_sparse() const { 3737 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 3738 Node* base_mem = base_memory(); 3739 // The following can happen in degenerate cases, since empty==top. 3740 if (is_empty_memory(base_mem)) return true; 3741 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 3742 assert(in(i) != NULL, "sane slice"); 3743 if (in(i) == base_mem) return false; // should have been the sentinel value! 3744 } 3745 return true; 3746} 3747 3748bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 3749 Node* n; 3750 n = mm->in(idx); 3751 if (mem == n) return true; // might be empty_memory() 3752 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 3753 if (mem == n) return true; 3754 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 3755 if (mem == n) return true; 3756 if (n == NULL) break; 3757 } 3758 return false; 3759} 3760#endif // !PRODUCT 3761