1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "oops/objArrayKlass.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/arraycopynode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/compile.hpp" 34 #include "opto/connode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/machnode.hpp" 38 #include "opto/matcher.hpp" 39 #include "opto/memnode.hpp" 40 #include "opto/mulnode.hpp" 41 #include "opto/narrowptrnode.hpp" 42 #include "opto/phaseX.hpp" 43 #include "opto/regmask.hpp" 44 #include "utilities/copy.hpp" 45 46 // Portions of code courtesy of Clifford Click 47 48 // Optimization - Graph Style 49 50 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 51 52 //============================================================================= 53 uint MemNode::size_of() const { return sizeof(*this); } 54 55 const TypePtr *MemNode::adr_type() const { 56 Node* adr = in(Address); 57 if (adr == NULL) return NULL; // node is dead 58 const TypePtr* cross_check = NULL; 59 DEBUG_ONLY(cross_check = _adr_type); 60 return calculate_adr_type(adr->bottom_type(), cross_check); 61 } 62 63 #ifndef PRODUCT 64 void MemNode::dump_spec(outputStream *st) const { 65 if (in(Address) == NULL) return; // node is dead 66 #ifndef ASSERT 67 // fake the missing field 68 const TypePtr* _adr_type = NULL; 69 if (in(Address) != NULL) 70 _adr_type = in(Address)->bottom_type()->isa_ptr(); 71 #endif 72 dump_adr_type(this, _adr_type, st); 73 74 Compile* C = Compile::current(); 75 if( C->alias_type(_adr_type)->is_volatile() ) 76 st->print(" Volatile!"); 77 } 78 79 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 80 st->print(" @"); 81 if (adr_type == NULL) { 82 st->print("NULL"); 83 } else { 84 adr_type->dump_on(st); 85 Compile* C = Compile::current(); 86 Compile::AliasType* atp = NULL; 87 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 88 if (atp == NULL) 89 st->print(", idx=?\?;"); 90 else if (atp->index() == Compile::AliasIdxBot) 91 st->print(", idx=Bot;"); 92 else if (atp->index() == Compile::AliasIdxTop) 93 st->print(", idx=Top;"); 94 else if (atp->index() == Compile::AliasIdxRaw) 95 st->print(", idx=Raw;"); 96 else { 97 ciField* field = atp->field(); 98 if (field) { 99 st->print(", name="); 100 field->print_name_on(st); 101 } 102 st->print(", idx=%d;", atp->index()); 103 } 104 } 105 } 106 107 extern void print_alias_types(); 108 109 #endif 110 111 static bool membar_for_arraycopy_helper(const TypeOopPtr *t_oop, MergeMemNode* mm, PhaseTransform *phase) { 112 if (mm->memory_at(Compile::AliasIdxRaw)->is_Proj()) { 113 Node* n = mm->memory_at(Compile::AliasIdxRaw)->in(0); 114 if ((n->is_ArrayCopy() && n->as_ArrayCopy()->may_modify(t_oop, phase)) || 115 (n->is_CallLeaf() && n->as_CallLeaf()->may_modify(t_oop, phase))) { 116 return true; 117 } 118 } 119 return false; 120 } 121 122 static bool membar_for_arraycopy(const TypeOopPtr *t_oop, MemBarNode* mb, PhaseTransform *phase) { 123 Node* mem = mb->in(TypeFunc::Memory); 124 if (mem->is_MergeMem()) { 125 return membar_for_arraycopy_helper(t_oop, mem->as_MergeMem(), phase); 126 } else if (mem->is_Phi()) { 127 // after macro expansion of an ArrayCopyNode we may have a Phi 128 for (uint i = 1; i < mem->req(); i++) { 129 if (mem->in(i) != NULL && mem->in(i)->is_MergeMem() && membar_for_arraycopy_helper(t_oop, mem->in(i)->as_MergeMem(), phase)) { 130 return true; 131 } 132 } 133 } 134 return false; 135 } 136 137 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 138 assert((t_oop != NULL), "sanity"); 139 bool is_instance = t_oop->is_known_instance_field(); 140 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 141 (load != NULL) && load->is_Load() && 142 (phase->is_IterGVN() != NULL); 143 if (!(is_instance || is_boxed_value_load)) 144 return mchain; // don't try to optimize non-instance types 145 uint instance_id = t_oop->instance_id(); 146 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 147 Node *prev = NULL; 148 Node *result = mchain; 149 while (prev != result) { 150 prev = result; 151 if (result == start_mem) 152 break; // hit one of our sentinels 153 // skip over a call which does not affect this memory slice 154 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 155 Node *proj_in = result->in(0); 156 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 157 break; // hit one of our sentinels 158 } else if (proj_in->is_Call()) { 159 // ArrayCopyNodes processed here as well 160 CallNode *call = proj_in->as_Call(); 161 if (!call->may_modify(t_oop, phase)) { // returns false for instances 162 result = call->in(TypeFunc::Memory); 163 } 164 } else if (proj_in->is_Initialize()) { 165 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 166 // Stop if this is the initialization for the object instance which 167 // contains this memory slice, otherwise skip over it. 168 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 169 break; 170 } 171 if (is_instance) { 172 result = proj_in->in(TypeFunc::Memory); 173 } else if (is_boxed_value_load) { 174 Node* klass = alloc->in(AllocateNode::KlassNode); 175 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 176 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 177 result = proj_in->in(TypeFunc::Memory); // not related allocation 178 } 179 } 180 } else if (proj_in->is_MemBar()) { 181 if (membar_for_arraycopy(t_oop, proj_in->as_MemBar(), phase)) { 182 break; 183 } 184 result = proj_in->in(TypeFunc::Memory); 185 } else { 186 assert(false, "unexpected projection"); 187 } 188 } else if (result->is_ClearArray()) { 189 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 190 // Can not bypass initialization of the instance 191 // we are looking for. 192 break; 193 } 194 // Otherwise skip it (the call updated 'result' value). 195 } else if (result->is_MergeMem()) { 196 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 197 } 198 } 199 return result; 200 } 201 202 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 203 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 204 if (t_oop == NULL) 205 return mchain; // don't try to optimize non-oop types 206 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 207 bool is_instance = t_oop->is_known_instance_field(); 208 PhaseIterGVN *igvn = phase->is_IterGVN(); 209 if (is_instance && igvn != NULL && result->is_Phi()) { 210 PhiNode *mphi = result->as_Phi(); 211 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 212 const TypePtr *t = mphi->adr_type(); 213 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 214 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 215 t->is_oopptr()->cast_to_exactness(true) 216 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 217 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 218 // clone the Phi with our address type 219 result = mphi->split_out_instance(t_adr, igvn); 220 } else { 221 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 222 } 223 } 224 return result; 225 } 226 227 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 228 uint alias_idx = phase->C->get_alias_index(tp); 229 Node *mem = mmem; 230 #ifdef ASSERT 231 { 232 // Check that current type is consistent with the alias index used during graph construction 233 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 234 bool consistent = adr_check == NULL || adr_check->empty() || 235 phase->C->must_alias(adr_check, alias_idx ); 236 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 237 if( !consistent && adr_check != NULL && !adr_check->empty() && 238 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 239 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 240 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 241 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 242 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 243 // don't assert if it is dead code. 244 consistent = true; 245 } 246 if( !consistent ) { 247 st->print("alias_idx==%d, adr_check==", alias_idx); 248 if( adr_check == NULL ) { 249 st->print("NULL"); 250 } else { 251 adr_check->dump(); 252 } 253 st->cr(); 254 print_alias_types(); 255 assert(consistent, "adr_check must match alias idx"); 256 } 257 } 258 #endif 259 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 260 // means an array I have not precisely typed yet. Do not do any 261 // alias stuff with it any time soon. 262 const TypeOopPtr *toop = tp->isa_oopptr(); 263 if( tp->base() != Type::AnyPtr && 264 !(toop && 265 toop->klass() != NULL && 266 toop->klass()->is_java_lang_Object() && 267 toop->offset() == Type::OffsetBot) ) { 268 // compress paths and change unreachable cycles to TOP 269 // If not, we can update the input infinitely along a MergeMem cycle 270 // Equivalent code in PhiNode::Ideal 271 Node* m = phase->transform(mmem); 272 // If transformed to a MergeMem, get the desired slice 273 // Otherwise the returned node represents memory for every slice 274 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 275 // Update input if it is progress over what we have now 276 } 277 return mem; 278 } 279 280 //--------------------------Ideal_common--------------------------------------- 281 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 282 // Unhook non-raw memories from complete (macro-expanded) initializations. 283 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 284 // If our control input is a dead region, kill all below the region 285 Node *ctl = in(MemNode::Control); 286 if (ctl && remove_dead_region(phase, can_reshape)) 287 return this; 288 ctl = in(MemNode::Control); 289 // Don't bother trying to transform a dead node 290 if (ctl && ctl->is_top()) return NodeSentinel; 291 292 PhaseIterGVN *igvn = phase->is_IterGVN(); 293 // Wait if control on the worklist. 294 if (ctl && can_reshape && igvn != NULL) { 295 Node* bol = NULL; 296 Node* cmp = NULL; 297 if (ctl->in(0)->is_If()) { 298 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 299 bol = ctl->in(0)->in(1); 300 if (bol->is_Bool()) 301 cmp = ctl->in(0)->in(1)->in(1); 302 } 303 if (igvn->_worklist.member(ctl) || 304 (bol != NULL && igvn->_worklist.member(bol)) || 305 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 306 // This control path may be dead. 307 // Delay this memory node transformation until the control is processed. 308 phase->is_IterGVN()->_worklist.push(this); 309 return NodeSentinel; // caller will return NULL 310 } 311 } 312 // Ignore if memory is dead, or self-loop 313 Node *mem = in(MemNode::Memory); 314 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 315 assert(mem != this, "dead loop in MemNode::Ideal"); 316 317 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 318 // This memory slice may be dead. 319 // Delay this mem node transformation until the memory is processed. 320 phase->is_IterGVN()->_worklist.push(this); 321 return NodeSentinel; // caller will return NULL 322 } 323 324 Node *address = in(MemNode::Address); 325 const Type *t_adr = phase->type(address); 326 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 327 328 if (can_reshape && igvn != NULL && 329 (igvn->_worklist.member(address) || 330 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { 331 // The address's base and type may change when the address is processed. 332 // Delay this mem node transformation until the address is processed. 333 phase->is_IterGVN()->_worklist.push(this); 334 return NodeSentinel; // caller will return NULL 335 } 336 337 // Do NOT remove or optimize the next lines: ensure a new alias index 338 // is allocated for an oop pointer type before Escape Analysis. 339 // Note: C++ will not remove it since the call has side effect. 340 if (t_adr->isa_oopptr()) { 341 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 342 } 343 344 Node* base = NULL; 345 if (address->is_AddP()) { 346 base = address->in(AddPNode::Base); 347 } 348 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 349 !t_adr->isa_rawptr()) { 350 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 351 // Skip this node optimization if its address has TOP base. 352 return NodeSentinel; // caller will return NULL 353 } 354 355 // Avoid independent memory operations 356 Node* old_mem = mem; 357 358 // The code which unhooks non-raw memories from complete (macro-expanded) 359 // initializations was removed. After macro-expansion all stores catched 360 // by Initialize node became raw stores and there is no information 361 // which memory slices they modify. So it is unsafe to move any memory 362 // operation above these stores. Also in most cases hooked non-raw memories 363 // were already unhooked by using information from detect_ptr_independence() 364 // and find_previous_store(). 365 366 if (mem->is_MergeMem()) { 367 MergeMemNode* mmem = mem->as_MergeMem(); 368 const TypePtr *tp = t_adr->is_ptr(); 369 370 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 371 } 372 373 if (mem != old_mem) { 374 set_req(MemNode::Memory, mem); 375 if (can_reshape && old_mem->outcnt() == 0) { 376 igvn->_worklist.push(old_mem); 377 } 378 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 379 return this; 380 } 381 382 // let the subclass continue analyzing... 383 return NULL; 384 } 385 386 // Helper function for proving some simple control dominations. 387 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 388 // Already assumes that 'dom' is available at 'sub', and that 'sub' 389 // is not a constant (dominated by the method's StartNode). 390 // Used by MemNode::find_previous_store to prove that the 391 // control input of a memory operation predates (dominates) 392 // an allocation it wants to look past. 393 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 394 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 395 return false; // Conservative answer for dead code 396 397 // Check 'dom'. Skip Proj and CatchProj nodes. 398 dom = dom->find_exact_control(dom); 399 if (dom == NULL || dom->is_top()) 400 return false; // Conservative answer for dead code 401 402 if (dom == sub) { 403 // For the case when, for example, 'sub' is Initialize and the original 404 // 'dom' is Proj node of the 'sub'. 405 return false; 406 } 407 408 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 409 return true; 410 411 // 'dom' dominates 'sub' if its control edge and control edges 412 // of all its inputs dominate or equal to sub's control edge. 413 414 // Currently 'sub' is either Allocate, Initialize or Start nodes. 415 // Or Region for the check in LoadNode::Ideal(); 416 // 'sub' should have sub->in(0) != NULL. 417 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 418 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 419 420 // Get control edge of 'sub'. 421 Node* orig_sub = sub; 422 sub = sub->find_exact_control(sub->in(0)); 423 if (sub == NULL || sub->is_top()) 424 return false; // Conservative answer for dead code 425 426 assert(sub->is_CFG(), "expecting control"); 427 428 if (sub == dom) 429 return true; 430 431 if (sub->is_Start() || sub->is_Root()) 432 return false; 433 434 { 435 // Check all control edges of 'dom'. 436 437 ResourceMark rm; 438 Arena* arena = Thread::current()->resource_area(); 439 Node_List nlist(arena); 440 Unique_Node_List dom_list(arena); 441 442 dom_list.push(dom); 443 bool only_dominating_controls = false; 444 445 for (uint next = 0; next < dom_list.size(); next++) { 446 Node* n = dom_list.at(next); 447 if (n == orig_sub) 448 return false; // One of dom's inputs dominated by sub. 449 if (!n->is_CFG() && n->pinned()) { 450 // Check only own control edge for pinned non-control nodes. 451 n = n->find_exact_control(n->in(0)); 452 if (n == NULL || n->is_top()) 453 return false; // Conservative answer for dead code 454 assert(n->is_CFG(), "expecting control"); 455 dom_list.push(n); 456 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 457 only_dominating_controls = true; 458 } else if (n->is_CFG()) { 459 if (n->dominates(sub, nlist)) 460 only_dominating_controls = true; 461 else 462 return false; 463 } else { 464 // First, own control edge. 465 Node* m = n->find_exact_control(n->in(0)); 466 if (m != NULL) { 467 if (m->is_top()) 468 return false; // Conservative answer for dead code 469 dom_list.push(m); 470 } 471 // Now, the rest of edges. 472 uint cnt = n->req(); 473 for (uint i = 1; i < cnt; i++) { 474 m = n->find_exact_control(n->in(i)); 475 if (m == NULL || m->is_top()) 476 continue; 477 dom_list.push(m); 478 } 479 } 480 } 481 return only_dominating_controls; 482 } 483 } 484 485 //---------------------detect_ptr_independence--------------------------------- 486 // Used by MemNode::find_previous_store to prove that two base 487 // pointers are never equal. 488 // The pointers are accompanied by their associated allocations, 489 // if any, which have been previously discovered by the caller. 490 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 491 Node* p2, AllocateNode* a2, 492 PhaseTransform* phase) { 493 // Attempt to prove that these two pointers cannot be aliased. 494 // They may both manifestly be allocations, and they should differ. 495 // Or, if they are not both allocations, they can be distinct constants. 496 // Otherwise, one is an allocation and the other a pre-existing value. 497 if (a1 == NULL && a2 == NULL) { // neither an allocation 498 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 499 } else if (a1 != NULL && a2 != NULL) { // both allocations 500 return (a1 != a2); 501 } else if (a1 != NULL) { // one allocation a1 502 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 503 return all_controls_dominate(p2, a1); 504 } else { //(a2 != NULL) // one allocation a2 505 return all_controls_dominate(p1, a2); 506 } 507 return false; 508 } 509 510 511 // Find an arraycopy that must have set (can_see_stored_value=true) or 512 // could have set (can_see_stored_value=false) the value for this load 513 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 514 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 515 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 516 Node* mb = mem->in(0); 517 if (mb->in(0) != NULL && mb->in(0)->is_Proj() && 518 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) { 519 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy(); 520 if (ac->is_clonebasic()) { 521 intptr_t offset; 522 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset); 523 assert(alloc != NULL && alloc->initialization()->is_complete_with_arraycopy(), "broken allocation"); 524 if (alloc == ld_alloc) { 525 return ac; 526 } 527 } 528 } 529 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) { 530 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 531 532 if (ac->is_arraycopy_validated() || 533 ac->is_copyof_validated() || 534 ac->is_copyofrange_validated()) { 535 Node* ld_addp = in(MemNode::Address); 536 if (ld_addp->is_AddP()) { 537 Node* ld_base = ld_addp->in(AddPNode::Address); 538 Node* ld_offs = ld_addp->in(AddPNode::Offset); 539 540 Node* dest = ac->in(ArrayCopyNode::Dest); 541 542 if (dest == ld_base) { 543 Node* src_pos = ac->in(ArrayCopyNode::SrcPos); 544 Node* dest_pos = ac->in(ArrayCopyNode::DestPos); 545 Node* len = ac->in(ArrayCopyNode::Length); 546 547 const TypeInt *dest_pos_t = phase->type(dest_pos)->isa_int(); 548 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 549 const TypeInt *len_t = phase->type(len)->isa_int(); 550 const TypeAryPtr* ary_t = phase->type(dest)->isa_aryptr(); 551 552 if (dest_pos_t != NULL && ld_offs_t != NULL && len_t != NULL && ary_t != NULL) { 553 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 554 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 555 uint elemsize = type2aelembytes(ary_elem); 556 557 intptr_t dest_pos_plus_len_lo = (((intptr_t)dest_pos_t->_lo) + len_t->_lo) * elemsize + header; 558 intptr_t dest_pos_plus_len_hi = (((intptr_t)dest_pos_t->_hi) + len_t->_hi) * elemsize + header; 559 intptr_t dest_pos_lo = ((intptr_t)dest_pos_t->_lo) * elemsize + header; 560 intptr_t dest_pos_hi = ((intptr_t)dest_pos_t->_hi) * elemsize + header; 561 562 if (can_see_stored_value) { 563 if (ld_offs_t->_lo >= dest_pos_hi && ld_offs_t->_hi < dest_pos_plus_len_lo) { 564 return ac; 565 } 566 } else { 567 if (ld_offs_t->_hi < dest_pos_lo || ld_offs_t->_lo >= dest_pos_plus_len_hi) { 568 mem = ac->in(TypeFunc::Memory); 569 } 570 return ac; 571 } 572 } 573 } 574 } 575 } 576 } 577 return NULL; 578 } 579 580 // The logic for reordering loads and stores uses four steps: 581 // (a) Walk carefully past stores and initializations which we 582 // can prove are independent of this load. 583 // (b) Observe that the next memory state makes an exact match 584 // with self (load or store), and locate the relevant store. 585 // (c) Ensure that, if we were to wire self directly to the store, 586 // the optimizer would fold it up somehow. 587 // (d) Do the rewiring, and return, depending on some other part of 588 // the optimizer to fold up the load. 589 // This routine handles steps (a) and (b). Steps (c) and (d) are 590 // specific to loads and stores, so they are handled by the callers. 591 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 592 // 593 Node* MemNode::find_previous_store(PhaseTransform* phase) { 594 Node* ctrl = in(MemNode::Control); 595 Node* adr = in(MemNode::Address); 596 intptr_t offset = 0; 597 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 598 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 599 600 if (offset == Type::OffsetBot) 601 return NULL; // cannot unalias unless there are precise offsets 602 603 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 604 605 intptr_t size_in_bytes = memory_size(); 606 607 Node* mem = in(MemNode::Memory); // start searching here... 608 609 int cnt = 50; // Cycle limiter 610 for (;;) { // While we can dance past unrelated stores... 611 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 612 613 Node* prev = mem; 614 if (mem->is_Store()) { 615 Node* st_adr = mem->in(MemNode::Address); 616 intptr_t st_offset = 0; 617 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 618 if (st_base == NULL) 619 break; // inscrutable pointer 620 if (st_offset != offset && st_offset != Type::OffsetBot) { 621 const int MAX_STORE = BytesPerLong; 622 if (st_offset >= offset + size_in_bytes || 623 st_offset <= offset - MAX_STORE || 624 st_offset <= offset - mem->as_Store()->memory_size()) { 625 // Success: The offsets are provably independent. 626 // (You may ask, why not just test st_offset != offset and be done? 627 // The answer is that stores of different sizes can co-exist 628 // in the same sequence of RawMem effects. We sometimes initialize 629 // a whole 'tile' of array elements with a single jint or jlong.) 630 mem = mem->in(MemNode::Memory); 631 continue; // (a) advance through independent store memory 632 } 633 } 634 if (st_base != base && 635 detect_ptr_independence(base, alloc, 636 st_base, 637 AllocateNode::Ideal_allocation(st_base, phase), 638 phase)) { 639 // Success: The bases are provably independent. 640 mem = mem->in(MemNode::Memory); 641 continue; // (a) advance through independent store memory 642 } 643 644 // (b) At this point, if the bases or offsets do not agree, we lose, 645 // since we have not managed to prove 'this' and 'mem' independent. 646 if (st_base == base && st_offset == offset) { 647 return mem; // let caller handle steps (c), (d) 648 } 649 650 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 651 InitializeNode* st_init = mem->in(0)->as_Initialize(); 652 AllocateNode* st_alloc = st_init->allocation(); 653 if (st_alloc == NULL) 654 break; // something degenerated 655 bool known_identical = false; 656 bool known_independent = false; 657 if (alloc == st_alloc) 658 known_identical = true; 659 else if (alloc != NULL) 660 known_independent = true; 661 else if (all_controls_dominate(this, st_alloc)) 662 known_independent = true; 663 664 if (known_independent) { 665 // The bases are provably independent: Either they are 666 // manifestly distinct allocations, or else the control 667 // of this load dominates the store's allocation. 668 int alias_idx = phase->C->get_alias_index(adr_type()); 669 if (alias_idx == Compile::AliasIdxRaw) { 670 mem = st_alloc->in(TypeFunc::Memory); 671 } else { 672 mem = st_init->memory(alias_idx); 673 } 674 continue; // (a) advance through independent store memory 675 } 676 677 // (b) at this point, if we are not looking at a store initializing 678 // the same allocation we are loading from, we lose. 679 if (known_identical) { 680 // From caller, can_see_stored_value will consult find_captured_store. 681 return mem; // let caller handle steps (c), (d) 682 } 683 684 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) { 685 if (prev != mem) { 686 // Found an arraycopy but it doesn't affect that load 687 continue; 688 } 689 // Found an arraycopy that may affect that load 690 return mem; 691 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 692 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 693 if (mem->is_Proj() && mem->in(0)->is_Call()) { 694 // ArrayCopyNodes processed here as well. 695 CallNode *call = mem->in(0)->as_Call(); 696 if (!call->may_modify(addr_t, phase)) { 697 mem = call->in(TypeFunc::Memory); 698 continue; // (a) advance through independent call memory 699 } 700 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 701 if (membar_for_arraycopy(addr_t, mem->in(0)->as_MemBar(), phase)) { 702 break; 703 } 704 mem = mem->in(0)->in(TypeFunc::Memory); 705 continue; // (a) advance through independent MemBar memory 706 } else if (mem->is_ClearArray()) { 707 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 708 // (the call updated 'mem' value) 709 continue; // (a) advance through independent allocation memory 710 } else { 711 // Can not bypass initialization of the instance 712 // we are looking for. 713 return mem; 714 } 715 } else if (mem->is_MergeMem()) { 716 int alias_idx = phase->C->get_alias_index(adr_type()); 717 mem = mem->as_MergeMem()->memory_at(alias_idx); 718 continue; // (a) advance through independent MergeMem memory 719 } 720 } 721 722 // Unless there is an explicit 'continue', we must bail out here, 723 // because 'mem' is an inscrutable memory state (e.g., a call). 724 break; 725 } 726 727 return NULL; // bail out 728 } 729 730 //----------------------calculate_adr_type------------------------------------- 731 // Helper function. Notices when the given type of address hits top or bottom. 732 // Also, asserts a cross-check of the type against the expected address type. 733 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 734 if (t == Type::TOP) return NULL; // does not touch memory any more? 735 #ifdef PRODUCT 736 cross_check = NULL; 737 #else 738 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 739 #endif 740 const TypePtr* tp = t->isa_ptr(); 741 if (tp == NULL) { 742 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 743 return TypePtr::BOTTOM; // touches lots of memory 744 } else { 745 #ifdef ASSERT 746 // %%%% [phh] We don't check the alias index if cross_check is 747 // TypeRawPtr::BOTTOM. Needs to be investigated. 748 if (cross_check != NULL && 749 cross_check != TypePtr::BOTTOM && 750 cross_check != TypeRawPtr::BOTTOM) { 751 // Recheck the alias index, to see if it has changed (due to a bug). 752 Compile* C = Compile::current(); 753 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 754 "must stay in the original alias category"); 755 // The type of the address must be contained in the adr_type, 756 // disregarding "null"-ness. 757 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 758 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 759 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 760 "real address must not escape from expected memory type"); 761 } 762 #endif 763 return tp; 764 } 765 } 766 767 //============================================================================= 768 // Should LoadNode::Ideal() attempt to remove control edges? 769 bool LoadNode::can_remove_control() const { 770 return true; 771 } 772 uint LoadNode::size_of() const { return sizeof(*this); } 773 uint LoadNode::cmp( const Node &n ) const 774 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 775 const Type *LoadNode::bottom_type() const { return _type; } 776 uint LoadNode::ideal_reg() const { 777 return _type->ideal_reg(); 778 } 779 780 #ifndef PRODUCT 781 void LoadNode::dump_spec(outputStream *st) const { 782 MemNode::dump_spec(st); 783 if( !Verbose && !WizardMode ) { 784 // standard dump does this in Verbose and WizardMode 785 st->print(" #"); _type->dump_on(st); 786 } 787 if (!_depends_only_on_test) { 788 st->print(" (does not depend only on test)"); 789 } 790 } 791 #endif 792 793 #ifdef ASSERT 794 //----------------------------is_immutable_value------------------------------- 795 // Helper function to allow a raw load without control edge for some cases 796 bool LoadNode::is_immutable_value(Node* adr) { 797 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 798 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 799 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 800 in_bytes(JavaThread::osthread_offset()))); 801 } 802 #endif 803 804 //----------------------------LoadNode::make----------------------------------- 805 // Polymorphic factory method: 806 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, ControlDependency control_dependency) { 807 Compile* C = gvn.C; 808 809 // sanity check the alias category against the created node type 810 assert(!(adr_type->isa_oopptr() && 811 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 812 "use LoadKlassNode instead"); 813 assert(!(adr_type->isa_aryptr() && 814 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 815 "use LoadRangeNode instead"); 816 // Check control edge of raw loads 817 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 818 // oop will be recorded in oop map if load crosses safepoint 819 rt->isa_oopptr() || is_immutable_value(adr), 820 "raw memory operations should have control edge"); 821 switch (bt) { 822 case T_BOOLEAN: return new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 823 case T_BYTE: return new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 824 case T_INT: return new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 825 case T_CHAR: return new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 826 case T_SHORT: return new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); 827 case T_LONG: return new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); 828 case T_FLOAT: return new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); 829 case T_DOUBLE: return new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); 830 case T_ADDRESS: return new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 831 case T_OBJECT: 832 #ifdef _LP64 833 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 834 Node* load = gvn.transform(new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency)); 835 return new DecodeNNode(load, load->bottom_type()->make_ptr()); 836 } else 837 #endif 838 { 839 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 840 return new LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo, control_dependency); 841 } 842 } 843 ShouldNotReachHere(); 844 return (LoadNode*)NULL; 845 } 846 847 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) { 848 bool require_atomic = true; 849 return new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic); 850 } 851 852 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) { 853 bool require_atomic = true; 854 return new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic); 855 } 856 857 858 859 //------------------------------hash------------------------------------------- 860 uint LoadNode::hash() const { 861 // unroll addition of interesting fields 862 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 863 } 864 865 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 866 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 867 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 868 bool is_stable_ary = FoldStableValues && 869 (tp != NULL) && (tp->isa_aryptr() != NULL) && 870 tp->isa_aryptr()->is_stable(); 871 872 return (eliminate_boxing && non_volatile) || is_stable_ary; 873 } 874 875 return false; 876 } 877 878 // Is the value loaded previously stored by an arraycopy? If so return 879 // a load node that reads from the source array so we may be able to 880 // optimize out the ArrayCopy node later. 881 Node* MemNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const { 882 Node* ld_adr = in(MemNode::Address); 883 intptr_t ld_off = 0; 884 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 885 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 886 if (ac != NULL) { 887 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 888 assert(is_Load(), "only for loads"); 889 890 if (ac->as_ArrayCopy()->is_clonebasic()) { 891 assert(ld_alloc != NULL, "need an alloc"); 892 Node* ld = clone(); 893 Node* addp = in(MemNode::Address)->clone(); 894 assert(addp->is_AddP(), "address must be addp"); 895 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern"); 896 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern"); 897 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base)); 898 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address)); 899 ld->set_req(MemNode::Address, phase->transform(addp)); 900 if (in(0) != NULL) { 901 assert(ld_alloc->in(0) != NULL, "alloc must have control"); 902 ld->set_req(0, ld_alloc->in(0)); 903 } 904 return ld; 905 } else { 906 Node* ld = clone(); 907 Node* addp = in(MemNode::Address)->clone(); 908 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 909 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)); 910 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)); 911 912 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 913 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type(); 914 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 915 uint shift = exact_log2(type2aelembytes(ary_elem)); 916 917 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 918 #ifdef _LP64 919 diff = phase->transform(new ConvI2LNode(diff)); 920 #endif 921 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 922 923 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 924 addp->set_req(AddPNode::Offset, offset); 925 ld->set_req(MemNode::Address, phase->transform(addp)); 926 927 if (in(0) != NULL) { 928 assert(ac->in(0) != NULL, "alloc must have control"); 929 ld->set_req(0, ac->in(0)); 930 } 931 return ld; 932 } 933 } 934 return NULL; 935 } 936 937 938 //---------------------------can_see_stored_value------------------------------ 939 // This routine exists to make sure this set of tests is done the same 940 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 941 // will change the graph shape in a way which makes memory alive twice at the 942 // same time (uses the Oracle model of aliasing), then some 943 // LoadXNode::Identity will fold things back to the equivalence-class model 944 // of aliasing. 945 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 946 Node* ld_adr = in(MemNode::Address); 947 intptr_t ld_off = 0; 948 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 949 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 950 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 951 // This is more general than load from boxing objects. 952 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 953 uint alias_idx = atp->index(); 954 bool final = !atp->is_rewritable(); 955 Node* result = NULL; 956 Node* current = st; 957 // Skip through chains of MemBarNodes checking the MergeMems for 958 // new states for the slice of this load. Stop once any other 959 // kind of node is encountered. Loads from final memory can skip 960 // through any kind of MemBar but normal loads shouldn't skip 961 // through MemBarAcquire since the could allow them to move out of 962 // a synchronized region. 963 while (current->is_Proj()) { 964 int opc = current->in(0)->Opcode(); 965 if ((final && (opc == Op_MemBarAcquire || 966 opc == Op_MemBarAcquireLock || 967 opc == Op_LoadFence)) || 968 opc == Op_MemBarRelease || 969 opc == Op_StoreFence || 970 opc == Op_MemBarReleaseLock || 971 opc == Op_MemBarStoreStore || 972 opc == Op_MemBarCPUOrder) { 973 Node* mem = current->in(0)->in(TypeFunc::Memory); 974 if (mem->is_MergeMem()) { 975 MergeMemNode* merge = mem->as_MergeMem(); 976 Node* new_st = merge->memory_at(alias_idx); 977 if (new_st == merge->base_memory()) { 978 // Keep searching 979 current = new_st; 980 continue; 981 } 982 // Save the new memory state for the slice and fall through 983 // to exit. 984 result = new_st; 985 } 986 } 987 break; 988 } 989 if (result != NULL) { 990 st = result; 991 } 992 } 993 994 // Loop around twice in the case Load -> Initialize -> Store. 995 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 996 for (int trip = 0; trip <= 1; trip++) { 997 998 if (st->is_Store()) { 999 Node* st_adr = st->in(MemNode::Address); 1000 if (!phase->eqv(st_adr, ld_adr)) { 1001 // Try harder before giving up... Match raw and non-raw pointers. 1002 intptr_t st_off = 0; 1003 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 1004 if (alloc == NULL) return NULL; 1005 if (alloc != ld_alloc) return NULL; 1006 if (ld_off != st_off) return NULL; 1007 // At this point we have proven something like this setup: 1008 // A = Allocate(...) 1009 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 1010 // S = StoreQ(, AddP(, A.Parm , #Off), V) 1011 // (Actually, we haven't yet proven the Q's are the same.) 1012 // In other words, we are loading from a casted version of 1013 // the same pointer-and-offset that we stored to. 1014 // Thus, we are able to replace L by V. 1015 } 1016 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1017 if (store_Opcode() != st->Opcode()) 1018 return NULL; 1019 return st->in(MemNode::ValueIn); 1020 } 1021 1022 // A load from a freshly-created object always returns zero. 1023 // (This can happen after LoadNode::Ideal resets the load's memory input 1024 // to find_captured_store, which returned InitializeNode::zero_memory.) 1025 if (st->is_Proj() && st->in(0)->is_Allocate() && 1026 (st->in(0) == ld_alloc) && 1027 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1028 // return a zero value for the load's basic type 1029 // (This is one of the few places where a generic PhaseTransform 1030 // can create new nodes. Think of it as lazily manifesting 1031 // virtually pre-existing constants.) 1032 return phase->zerocon(memory_type()); 1033 } 1034 1035 // A load from an initialization barrier can match a captured store. 1036 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1037 InitializeNode* init = st->in(0)->as_Initialize(); 1038 AllocateNode* alloc = init->allocation(); 1039 if ((alloc != NULL) && (alloc == ld_alloc)) { 1040 // examine a captured store value 1041 st = init->find_captured_store(ld_off, memory_size(), phase); 1042 if (st != NULL) { 1043 continue; // take one more trip around 1044 } 1045 } 1046 } 1047 1048 // Load boxed value from result of valueOf() call is input parameter. 1049 if (this->is_Load() && ld_adr->is_AddP() && 1050 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1051 intptr_t ignore = 0; 1052 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1053 if (base != NULL && base->is_Proj() && 1054 base->as_Proj()->_con == TypeFunc::Parms && 1055 base->in(0)->is_CallStaticJava() && 1056 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1057 return base->in(0)->in(TypeFunc::Parms); 1058 } 1059 } 1060 1061 break; 1062 } 1063 1064 return NULL; 1065 } 1066 1067 //----------------------is_instance_field_load_with_local_phi------------------ 1068 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1069 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1070 in(Address)->is_AddP() ) { 1071 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1072 // Only instances and boxed values. 1073 if( t_oop != NULL && 1074 (t_oop->is_ptr_to_boxed_value() || 1075 t_oop->is_known_instance_field()) && 1076 t_oop->offset() != Type::OffsetBot && 1077 t_oop->offset() != Type::OffsetTop) { 1078 return true; 1079 } 1080 } 1081 return false; 1082 } 1083 1084 //------------------------------Identity--------------------------------------- 1085 // Loads are identity if previous store is to same address 1086 Node *LoadNode::Identity( PhaseTransform *phase ) { 1087 // If the previous store-maker is the right kind of Store, and the store is 1088 // to the same address, then we are equal to the value stored. 1089 Node* mem = in(Memory); 1090 Node* value = can_see_stored_value(mem, phase); 1091 if( value ) { 1092 // byte, short & char stores truncate naturally. 1093 // A load has to load the truncated value which requires 1094 // some sort of masking operation and that requires an 1095 // Ideal call instead of an Identity call. 1096 if (memory_size() < BytesPerInt) { 1097 // If the input to the store does not fit with the load's result type, 1098 // it must be truncated via an Ideal call. 1099 if (!phase->type(value)->higher_equal(phase->type(this))) 1100 return this; 1101 } 1102 // (This works even when value is a Con, but LoadNode::Value 1103 // usually runs first, producing the singleton type of the Con.) 1104 return value; 1105 } 1106 1107 // Search for an existing data phi which was generated before for the same 1108 // instance's field to avoid infinite generation of phis in a loop. 1109 Node *region = mem->in(0); 1110 if (is_instance_field_load_with_local_phi(region)) { 1111 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1112 int this_index = phase->C->get_alias_index(addr_t); 1113 int this_offset = addr_t->offset(); 1114 int this_iid = addr_t->instance_id(); 1115 if (!addr_t->is_known_instance() && 1116 addr_t->is_ptr_to_boxed_value()) { 1117 // Use _idx of address base (could be Phi node) for boxed values. 1118 intptr_t ignore = 0; 1119 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1120 this_iid = base->_idx; 1121 } 1122 const Type* this_type = bottom_type(); 1123 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1124 Node* phi = region->fast_out(i); 1125 if (phi->is_Phi() && phi != mem && 1126 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) { 1127 return phi; 1128 } 1129 } 1130 } 1131 1132 return this; 1133 } 1134 1135 // We're loading from an object which has autobox behaviour. 1136 // If this object is result of a valueOf call we'll have a phi 1137 // merging a newly allocated object and a load from the cache. 1138 // We want to replace this load with the original incoming 1139 // argument to the valueOf call. 1140 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1141 assert(phase->C->eliminate_boxing(), "sanity"); 1142 intptr_t ignore = 0; 1143 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1144 if ((base == NULL) || base->is_Phi()) { 1145 // Push the loads from the phi that comes from valueOf up 1146 // through it to allow elimination of the loads and the recovery 1147 // of the original value. It is done in split_through_phi(). 1148 return NULL; 1149 } else if (base->is_Load() || 1150 base->is_DecodeN() && base->in(1)->is_Load()) { 1151 // Eliminate the load of boxed value for integer types from the cache 1152 // array by deriving the value from the index into the array. 1153 // Capture the offset of the load and then reverse the computation. 1154 1155 // Get LoadN node which loads a boxing object from 'cache' array. 1156 if (base->is_DecodeN()) { 1157 base = base->in(1); 1158 } 1159 if (!base->in(Address)->is_AddP()) { 1160 return NULL; // Complex address 1161 } 1162 AddPNode* address = base->in(Address)->as_AddP(); 1163 Node* cache_base = address->in(AddPNode::Base); 1164 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1165 // Get ConP node which is static 'cache' field. 1166 cache_base = cache_base->in(1); 1167 } 1168 if ((cache_base != NULL) && cache_base->is_Con()) { 1169 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1170 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1171 Node* elements[4]; 1172 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1173 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1174 if ((count > 0) && elements[0]->is_Con() && 1175 ((count == 1) || 1176 (count == 2) && elements[1]->Opcode() == Op_LShiftX && 1177 elements[1]->in(2) == phase->intcon(shift))) { 1178 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1179 // Fetch the box object cache[0] at the base of the array and get its value 1180 ciInstance* box = array->obj_at(0)->as_instance(); 1181 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1182 assert(ik->is_box_klass(), "sanity"); 1183 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1184 if (ik->nof_nonstatic_fields() == 1) { 1185 // This should be true nonstatic_field_at requires calling 1186 // nof_nonstatic_fields so check it anyway 1187 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1188 BasicType bt = c.basic_type(); 1189 // Only integer types have boxing cache. 1190 assert(bt == T_BOOLEAN || bt == T_CHAR || 1191 bt == T_BYTE || bt == T_SHORT || 1192 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt))); 1193 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1194 if (cache_low != (int)cache_low) { 1195 return NULL; // should not happen since cache is array indexed by value 1196 } 1197 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1198 if (offset != (int)offset) { 1199 return NULL; // should not happen since cache is array indexed by value 1200 } 1201 // Add up all the offsets making of the address of the load 1202 Node* result = elements[0]; 1203 for (int i = 1; i < count; i++) { 1204 result = phase->transform(new AddXNode(result, elements[i])); 1205 } 1206 // Remove the constant offset from the address and then 1207 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1208 // remove the scaling of the offset to recover the original index. 1209 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1210 // Peel the shift off directly but wrap it in a dummy node 1211 // since Ideal can't return existing nodes 1212 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1213 } else if (result->is_Add() && result->in(2)->is_Con() && 1214 result->in(1)->Opcode() == Op_LShiftX && 1215 result->in(1)->in(2) == phase->intcon(shift)) { 1216 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1217 // but for boxing cache access we know that X<<Z will not overflow 1218 // (there is range check) so we do this optimizatrion by hand here. 1219 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1220 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1221 } else { 1222 result = new RShiftXNode(result, phase->intcon(shift)); 1223 } 1224 #ifdef _LP64 1225 if (bt != T_LONG) { 1226 result = new ConvL2INode(phase->transform(result)); 1227 } 1228 #else 1229 if (bt == T_LONG) { 1230 result = new ConvI2LNode(phase->transform(result)); 1231 } 1232 #endif 1233 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1234 // Need to preserve unboxing load type if it is unsigned. 1235 switch(this->Opcode()) { 1236 case Op_LoadUB: 1237 result = new AndINode(phase->transform(result), phase->intcon(0xFF)); 1238 break; 1239 case Op_LoadUS: 1240 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF)); 1241 break; 1242 } 1243 return result; 1244 } 1245 } 1246 } 1247 } 1248 } 1249 return NULL; 1250 } 1251 1252 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1253 Node* region = phi->in(0); 1254 if (region == NULL) { 1255 return false; // Wait stable graph 1256 } 1257 uint cnt = phi->req(); 1258 for (uint i = 1; i < cnt; i++) { 1259 Node* rc = region->in(i); 1260 if (rc == NULL || phase->type(rc) == Type::TOP) 1261 return false; // Wait stable graph 1262 Node* in = phi->in(i); 1263 if (in == NULL || phase->type(in) == Type::TOP) 1264 return false; // Wait stable graph 1265 } 1266 return true; 1267 } 1268 //------------------------------split_through_phi------------------------------ 1269 // Split instance or boxed field load through Phi. 1270 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1271 Node* mem = in(Memory); 1272 Node* address = in(Address); 1273 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1274 1275 assert((t_oop != NULL) && 1276 (t_oop->is_known_instance_field() || 1277 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1278 1279 Compile* C = phase->C; 1280 intptr_t ignore = 0; 1281 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1282 bool base_is_phi = (base != NULL) && base->is_Phi(); 1283 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1284 (base != NULL) && (base == address->in(AddPNode::Base)) && 1285 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1286 1287 if (!((mem->is_Phi() || base_is_phi) && 1288 (load_boxed_values || t_oop->is_known_instance_field()))) { 1289 return NULL; // memory is not Phi 1290 } 1291 1292 if (mem->is_Phi()) { 1293 if (!stable_phi(mem->as_Phi(), phase)) { 1294 return NULL; // Wait stable graph 1295 } 1296 uint cnt = mem->req(); 1297 // Check for loop invariant memory. 1298 if (cnt == 3) { 1299 for (uint i = 1; i < cnt; i++) { 1300 Node* in = mem->in(i); 1301 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1302 if (m == mem) { 1303 set_req(Memory, mem->in(cnt - i)); 1304 return this; // made change 1305 } 1306 } 1307 } 1308 } 1309 if (base_is_phi) { 1310 if (!stable_phi(base->as_Phi(), phase)) { 1311 return NULL; // Wait stable graph 1312 } 1313 uint cnt = base->req(); 1314 // Check for loop invariant memory. 1315 if (cnt == 3) { 1316 for (uint i = 1; i < cnt; i++) { 1317 if (base->in(i) == base) { 1318 return NULL; // Wait stable graph 1319 } 1320 } 1321 } 1322 } 1323 1324 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1325 1326 // Split through Phi (see original code in loopopts.cpp). 1327 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1328 1329 // Do nothing here if Identity will find a value 1330 // (to avoid infinite chain of value phis generation). 1331 if (!phase->eqv(this, this->Identity(phase))) 1332 return NULL; 1333 1334 // Select Region to split through. 1335 Node* region; 1336 if (!base_is_phi) { 1337 assert(mem->is_Phi(), "sanity"); 1338 region = mem->in(0); 1339 // Skip if the region dominates some control edge of the address. 1340 if (!MemNode::all_controls_dominate(address, region)) 1341 return NULL; 1342 } else if (!mem->is_Phi()) { 1343 assert(base_is_phi, "sanity"); 1344 region = base->in(0); 1345 // Skip if the region dominates some control edge of the memory. 1346 if (!MemNode::all_controls_dominate(mem, region)) 1347 return NULL; 1348 } else if (base->in(0) != mem->in(0)) { 1349 assert(base_is_phi && mem->is_Phi(), "sanity"); 1350 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1351 region = base->in(0); 1352 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1353 region = mem->in(0); 1354 } else { 1355 return NULL; // complex graph 1356 } 1357 } else { 1358 assert(base->in(0) == mem->in(0), "sanity"); 1359 region = mem->in(0); 1360 } 1361 1362 const Type* this_type = this->bottom_type(); 1363 int this_index = C->get_alias_index(t_oop); 1364 int this_offset = t_oop->offset(); 1365 int this_iid = t_oop->instance_id(); 1366 if (!t_oop->is_known_instance() && load_boxed_values) { 1367 // Use _idx of address base for boxed values. 1368 this_iid = base->_idx; 1369 } 1370 PhaseIterGVN* igvn = phase->is_IterGVN(); 1371 Node* phi = new PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1372 for (uint i = 1; i < region->req(); i++) { 1373 Node* x; 1374 Node* the_clone = NULL; 1375 if (region->in(i) == C->top()) { 1376 x = C->top(); // Dead path? Use a dead data op 1377 } else { 1378 x = this->clone(); // Else clone up the data op 1379 the_clone = x; // Remember for possible deletion. 1380 // Alter data node to use pre-phi inputs 1381 if (this->in(0) == region) { 1382 x->set_req(0, region->in(i)); 1383 } else { 1384 x->set_req(0, NULL); 1385 } 1386 if (mem->is_Phi() && (mem->in(0) == region)) { 1387 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1388 } 1389 if (address->is_Phi() && address->in(0) == region) { 1390 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1391 } 1392 if (base_is_phi && (base->in(0) == region)) { 1393 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1394 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1395 x->set_req(Address, adr_x); 1396 } 1397 } 1398 // Check for a 'win' on some paths 1399 const Type *t = x->Value(igvn); 1400 1401 bool singleton = t->singleton(); 1402 1403 // See comments in PhaseIdealLoop::split_thru_phi(). 1404 if (singleton && t == Type::TOP) { 1405 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1406 } 1407 1408 if (singleton) { 1409 x = igvn->makecon(t); 1410 } else { 1411 // We now call Identity to try to simplify the cloned node. 1412 // Note that some Identity methods call phase->type(this). 1413 // Make sure that the type array is big enough for 1414 // our new node, even though we may throw the node away. 1415 // (This tweaking with igvn only works because x is a new node.) 1416 igvn->set_type(x, t); 1417 // If x is a TypeNode, capture any more-precise type permanently into Node 1418 // otherwise it will be not updated during igvn->transform since 1419 // igvn->type(x) is set to x->Value() already. 1420 x->raise_bottom_type(t); 1421 Node *y = x->Identity(igvn); 1422 if (y != x) { 1423 x = y; 1424 } else { 1425 y = igvn->hash_find_insert(x); 1426 if (y) { 1427 x = y; 1428 } else { 1429 // Else x is a new node we are keeping 1430 // We do not need register_new_node_with_optimizer 1431 // because set_type has already been called. 1432 igvn->_worklist.push(x); 1433 } 1434 } 1435 } 1436 if (x != the_clone && the_clone != NULL) { 1437 igvn->remove_dead_node(the_clone); 1438 } 1439 phi->set_req(i, x); 1440 } 1441 // Record Phi 1442 igvn->register_new_node_with_optimizer(phi); 1443 return phi; 1444 } 1445 1446 //------------------------------Ideal------------------------------------------ 1447 // If the load is from Field memory and the pointer is non-null, it might be possible to 1448 // zero out the control input. 1449 // If the offset is constant and the base is an object allocation, 1450 // try to hook me up to the exact initializing store. 1451 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1452 Node* p = MemNode::Ideal_common(phase, can_reshape); 1453 if (p) return (p == NodeSentinel) ? NULL : p; 1454 1455 Node* ctrl = in(MemNode::Control); 1456 Node* address = in(MemNode::Address); 1457 bool progress = false; 1458 1459 // Skip up past a SafePoint control. Cannot do this for Stores because 1460 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1461 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1462 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1463 ctrl = ctrl->in(0); 1464 set_req(MemNode::Control,ctrl); 1465 progress = true; 1466 } 1467 1468 intptr_t ignore = 0; 1469 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1470 if (base != NULL 1471 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1472 // Check for useless control edge in some common special cases 1473 if (in(MemNode::Control) != NULL 1474 && can_remove_control() 1475 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1476 && all_controls_dominate(base, phase->C->start())) { 1477 // A method-invariant, non-null address (constant or 'this' argument). 1478 set_req(MemNode::Control, NULL); 1479 progress = true; 1480 } 1481 } 1482 1483 Node* mem = in(MemNode::Memory); 1484 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1485 1486 if (can_reshape && (addr_t != NULL)) { 1487 // try to optimize our memory input 1488 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1489 if (opt_mem != mem) { 1490 set_req(MemNode::Memory, opt_mem); 1491 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1492 return this; 1493 } 1494 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1495 if ((t_oop != NULL) && 1496 (t_oop->is_known_instance_field() || 1497 t_oop->is_ptr_to_boxed_value())) { 1498 PhaseIterGVN *igvn = phase->is_IterGVN(); 1499 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1500 // Delay this transformation until memory Phi is processed. 1501 phase->is_IterGVN()->_worklist.push(this); 1502 return NULL; 1503 } 1504 // Split instance field load through Phi. 1505 Node* result = split_through_phi(phase); 1506 if (result != NULL) return result; 1507 1508 if (t_oop->is_ptr_to_boxed_value()) { 1509 Node* result = eliminate_autobox(phase); 1510 if (result != NULL) return result; 1511 } 1512 } 1513 } 1514 1515 // Is there a dominating load that loads the same value? Leave 1516 // anything that is not a load of a field/array element (like 1517 // barriers etc.) alone 1518 if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) { 1519 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1520 Node *use = mem->fast_out(i); 1521 if (use != this && 1522 use->Opcode() == Opcode() && 1523 use->in(0) != NULL && 1524 use->in(0) != in(0) && 1525 use->in(Address) == in(Address)) { 1526 Node* ctl = in(0); 1527 for (int i = 0; i < 10 && ctl != NULL; i++) { 1528 ctl = IfNode::up_one_dom(ctl); 1529 if (ctl == use->in(0)) { 1530 set_req(0, use->in(0)); 1531 return this; 1532 } 1533 } 1534 } 1535 } 1536 } 1537 1538 // Check for prior store with a different base or offset; make Load 1539 // independent. Skip through any number of them. Bail out if the stores 1540 // are in an endless dead cycle and report no progress. This is a key 1541 // transform for Reflection. However, if after skipping through the Stores 1542 // we can't then fold up against a prior store do NOT do the transform as 1543 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1544 // array memory alive twice: once for the hoisted Load and again after the 1545 // bypassed Store. This situation only works if EVERYBODY who does 1546 // anti-dependence work knows how to bypass. I.e. we need all 1547 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1548 // the alias index stuff. So instead, peek through Stores and IFF we can 1549 // fold up, do so. 1550 Node* prev_mem = find_previous_store(phase); 1551 if (prev_mem != NULL) { 1552 Node* value = can_see_arraycopy_value(prev_mem, phase); 1553 if (value != NULL) { 1554 return value; 1555 } 1556 } 1557 // Steps (a), (b): Walk past independent stores to find an exact match. 1558 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1559 // (c) See if we can fold up on the spot, but don't fold up here. 1560 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1561 // just return a prior value, which is done by Identity calls. 1562 if (can_see_stored_value(prev_mem, phase)) { 1563 // Make ready for step (d): 1564 set_req(MemNode::Memory, prev_mem); 1565 return this; 1566 } 1567 } 1568 1569 return progress ? this : NULL; 1570 } 1571 1572 // Helper to recognize certain Klass fields which are invariant across 1573 // some group of array types (e.g., int[] or all T[] where T < Object). 1574 const Type* 1575 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1576 ciKlass* klass) const { 1577 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1578 // The field is Klass::_modifier_flags. Return its (constant) value. 1579 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1580 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1581 return TypeInt::make(klass->modifier_flags()); 1582 } 1583 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1584 // The field is Klass::_access_flags. Return its (constant) value. 1585 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1586 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1587 return TypeInt::make(klass->access_flags()); 1588 } 1589 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1590 // The field is Klass::_layout_helper. Return its constant value if known. 1591 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1592 return TypeInt::make(klass->layout_helper()); 1593 } 1594 1595 // No match. 1596 return NULL; 1597 } 1598 1599 // Try to constant-fold a stable array element. 1600 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) { 1601 assert(ary->const_oop(), "array should be constant"); 1602 assert(ary->is_stable(), "array should be stable"); 1603 1604 // Decode the results of GraphKit::array_element_address. 1605 ciArray* aobj = ary->const_oop()->as_array(); 1606 ciConstant con = aobj->element_value_by_offset(off); 1607 1608 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) { 1609 const Type* con_type = Type::make_from_constant(con); 1610 if (con_type != NULL) { 1611 if (con_type->isa_aryptr()) { 1612 // Join with the array element type, in case it is also stable. 1613 int dim = ary->stable_dimension(); 1614 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1); 1615 } 1616 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) { 1617 con_type = con_type->make_narrowoop(); 1618 } 1619 #ifndef PRODUCT 1620 if (TraceIterativeGVN) { 1621 tty->print("FoldStableValues: array element [off=%d]: con_type=", off); 1622 con_type->dump(); tty->cr(); 1623 } 1624 #endif //PRODUCT 1625 return con_type; 1626 } 1627 } 1628 return NULL; 1629 } 1630 1631 //------------------------------Value----------------------------------------- 1632 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1633 // Either input is TOP ==> the result is TOP 1634 Node* mem = in(MemNode::Memory); 1635 const Type *t1 = phase->type(mem); 1636 if (t1 == Type::TOP) return Type::TOP; 1637 Node* adr = in(MemNode::Address); 1638 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1639 if (tp == NULL || tp->empty()) return Type::TOP; 1640 int off = tp->offset(); 1641 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1642 Compile* C = phase->C; 1643 1644 // Try to guess loaded type from pointer type 1645 if (tp->isa_aryptr()) { 1646 const TypeAryPtr* ary = tp->is_aryptr(); 1647 const Type* t = ary->elem(); 1648 1649 // Determine whether the reference is beyond the header or not, by comparing 1650 // the offset against the offset of the start of the array's data. 1651 // Different array types begin at slightly different offsets (12 vs. 16). 1652 // We choose T_BYTE as an example base type that is least restrictive 1653 // as to alignment, which will therefore produce the smallest 1654 // possible base offset. 1655 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1656 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1657 1658 // Try to constant-fold a stable array element. 1659 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) { 1660 // Make sure the reference is not into the header and the offset is constant 1661 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1662 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type()); 1663 if (con_type != NULL) { 1664 return con_type; 1665 } 1666 } 1667 } 1668 1669 // Don't do this for integer types. There is only potential profit if 1670 // the element type t is lower than _type; that is, for int types, if _type is 1671 // more restrictive than t. This only happens here if one is short and the other 1672 // char (both 16 bits), and in those cases we've made an intentional decision 1673 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1674 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1675 // 1676 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1677 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1678 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1679 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1680 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1681 // In fact, that could have been the original type of p1, and p1 could have 1682 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1683 // expression (LShiftL quux 3) independently optimized to the constant 8. 1684 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1685 && (_type->isa_vect() == NULL) 1686 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1687 // t might actually be lower than _type, if _type is a unique 1688 // concrete subclass of abstract class t. 1689 if (off_beyond_header) { // is the offset beyond the header? 1690 const Type* jt = t->join_speculative(_type); 1691 // In any case, do not allow the join, per se, to empty out the type. 1692 if (jt->empty() && !t->empty()) { 1693 // This can happen if a interface-typed array narrows to a class type. 1694 jt = _type; 1695 } 1696 #ifdef ASSERT 1697 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1698 // The pointers in the autobox arrays are always non-null 1699 Node* base = adr->in(AddPNode::Base); 1700 if ((base != NULL) && base->is_DecodeN()) { 1701 // Get LoadN node which loads IntegerCache.cache field 1702 base = base->in(1); 1703 } 1704 if ((base != NULL) && base->is_Con()) { 1705 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1706 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1707 // It could be narrow oop 1708 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1709 } 1710 } 1711 } 1712 #endif 1713 return jt; 1714 } 1715 } 1716 } else if (tp->base() == Type::InstPtr) { 1717 ciEnv* env = C->env(); 1718 const TypeInstPtr* tinst = tp->is_instptr(); 1719 ciKlass* klass = tinst->klass(); 1720 assert( off != Type::OffsetBot || 1721 // arrays can be cast to Objects 1722 tp->is_oopptr()->klass()->is_java_lang_Object() || 1723 // unsafe field access may not have a constant offset 1724 C->has_unsafe_access(), 1725 "Field accesses must be precise" ); 1726 // For oop loads, we expect the _type to be precise 1727 if (klass == env->String_klass() && 1728 adr->is_AddP() && off != Type::OffsetBot) { 1729 // For constant Strings treat the final fields as compile time constants. 1730 Node* base = adr->in(AddPNode::Base); 1731 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1732 if (t != NULL && t->singleton()) { 1733 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1734 if (field != NULL && field->is_final()) { 1735 ciObject* string = t->const_oop(); 1736 ciConstant constant = string->as_instance()->field_value(field); 1737 if (constant.basic_type() == T_INT) { 1738 return TypeInt::make(constant.as_int()); 1739 } else if (constant.basic_type() == T_ARRAY) { 1740 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1741 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1742 } else { 1743 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1744 } 1745 } 1746 } 1747 } 1748 } 1749 // Optimizations for constant objects 1750 ciObject* const_oop = tinst->const_oop(); 1751 if (const_oop != NULL) { 1752 // For constant Boxed value treat the target field as a compile time constant. 1753 if (tinst->is_ptr_to_boxed_value()) { 1754 return tinst->get_const_boxed_value(); 1755 } else 1756 // For constant CallSites treat the target field as a compile time constant. 1757 if (const_oop->is_call_site()) { 1758 ciCallSite* call_site = const_oop->as_call_site(); 1759 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1760 if (field != NULL && field->is_call_site_target()) { 1761 ciMethodHandle* target = call_site->get_target(); 1762 if (target != NULL) { // just in case 1763 ciConstant constant(T_OBJECT, target); 1764 const Type* t; 1765 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1766 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1767 } else { 1768 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1769 } 1770 // Add a dependence for invalidation of the optimization. 1771 if (!call_site->is_constant_call_site()) { 1772 C->dependencies()->assert_call_site_target_value(call_site, target); 1773 } 1774 return t; 1775 } 1776 } 1777 } 1778 } 1779 } else if (tp->base() == Type::KlassPtr) { 1780 assert( off != Type::OffsetBot || 1781 // arrays can be cast to Objects 1782 tp->is_klassptr()->klass()->is_java_lang_Object() || 1783 // also allow array-loading from the primary supertype 1784 // array during subtype checks 1785 Opcode() == Op_LoadKlass, 1786 "Field accesses must be precise" ); 1787 // For klass/static loads, we expect the _type to be precise 1788 } 1789 1790 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1791 if (tkls != NULL && !StressReflectiveCode) { 1792 ciKlass* klass = tkls->klass(); 1793 if (klass->is_loaded() && tkls->klass_is_exact()) { 1794 // We are loading a field from a Klass metaobject whose identity 1795 // is known at compile time (the type is "exact" or "precise"). 1796 // Check for fields we know are maintained as constants by the VM. 1797 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1798 // The field is Klass::_super_check_offset. Return its (constant) value. 1799 // (Folds up type checking code.) 1800 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1801 return TypeInt::make(klass->super_check_offset()); 1802 } 1803 // Compute index into primary_supers array 1804 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1805 // Check for overflowing; use unsigned compare to handle the negative case. 1806 if( depth < ciKlass::primary_super_limit() ) { 1807 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1808 // (Folds up type checking code.) 1809 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1810 ciKlass *ss = klass->super_of_depth(depth); 1811 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1812 } 1813 const Type* aift = load_array_final_field(tkls, klass); 1814 if (aift != NULL) return aift; 1815 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1816 // The field is Klass::_java_mirror. Return its (constant) value. 1817 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1818 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1819 return TypeInstPtr::make(klass->java_mirror()); 1820 } 1821 } 1822 1823 // We can still check if we are loading from the primary_supers array at a 1824 // shallow enough depth. Even though the klass is not exact, entries less 1825 // than or equal to its super depth are correct. 1826 if (klass->is_loaded() ) { 1827 ciType *inner = klass; 1828 while( inner->is_obj_array_klass() ) 1829 inner = inner->as_obj_array_klass()->base_element_type(); 1830 if( inner->is_instance_klass() && 1831 !inner->as_instance_klass()->flags().is_interface() ) { 1832 // Compute index into primary_supers array 1833 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1834 // Check for overflowing; use unsigned compare to handle the negative case. 1835 if( depth < ciKlass::primary_super_limit() && 1836 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1837 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1838 // (Folds up type checking code.) 1839 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1840 ciKlass *ss = klass->super_of_depth(depth); 1841 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1842 } 1843 } 1844 } 1845 1846 // If the type is enough to determine that the thing is not an array, 1847 // we can give the layout_helper a positive interval type. 1848 // This will help short-circuit some reflective code. 1849 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1850 && !klass->is_array_klass() // not directly typed as an array 1851 && !klass->is_interface() // specifically not Serializable & Cloneable 1852 && !klass->is_java_lang_Object() // not the supertype of all T[] 1853 ) { 1854 // Note: When interfaces are reliable, we can narrow the interface 1855 // test to (klass != Serializable && klass != Cloneable). 1856 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1857 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1858 // The key property of this type is that it folds up tests 1859 // for array-ness, since it proves that the layout_helper is positive. 1860 // Thus, a generic value like the basic object layout helper works fine. 1861 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1862 } 1863 } 1864 1865 // If we are loading from a freshly-allocated object, produce a zero, 1866 // if the load is provably beyond the header of the object. 1867 // (Also allow a variable load from a fresh array to produce zero.) 1868 const TypeOopPtr *tinst = tp->isa_oopptr(); 1869 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1870 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1871 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1872 Node* value = can_see_stored_value(mem,phase); 1873 if (value != NULL && value->is_Con()) { 1874 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1875 return value->bottom_type(); 1876 } 1877 } 1878 1879 if (is_instance) { 1880 // If we have an instance type and our memory input is the 1881 // programs's initial memory state, there is no matching store, 1882 // so just return a zero of the appropriate type 1883 Node *mem = in(MemNode::Memory); 1884 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1885 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1886 return Type::get_zero_type(_type->basic_type()); 1887 } 1888 } 1889 return _type; 1890 } 1891 1892 //------------------------------match_edge------------------------------------- 1893 // Do we Match on this edge index or not? Match only the address. 1894 uint LoadNode::match_edge(uint idx) const { 1895 return idx == MemNode::Address; 1896 } 1897 1898 //--------------------------LoadBNode::Ideal-------------------------------------- 1899 // 1900 // If the previous store is to the same address as this load, 1901 // and the value stored was larger than a byte, replace this load 1902 // with the value stored truncated to a byte. If no truncation is 1903 // needed, the replacement is done in LoadNode::Identity(). 1904 // 1905 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1906 Node* mem = in(MemNode::Memory); 1907 Node* value = can_see_stored_value(mem,phase); 1908 if( value && !phase->type(value)->higher_equal( _type ) ) { 1909 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1910 return new RShiftINode(result, phase->intcon(24)); 1911 } 1912 // Identity call will handle the case where truncation is not needed. 1913 return LoadNode::Ideal(phase, can_reshape); 1914 } 1915 1916 const Type* LoadBNode::Value(PhaseTransform *phase) const { 1917 Node* mem = in(MemNode::Memory); 1918 Node* value = can_see_stored_value(mem,phase); 1919 if (value != NULL && value->is_Con() && 1920 !value->bottom_type()->higher_equal(_type)) { 1921 // If the input to the store does not fit with the load's result type, 1922 // it must be truncated. We can't delay until Ideal call since 1923 // a singleton Value is needed for split_thru_phi optimization. 1924 int con = value->get_int(); 1925 return TypeInt::make((con << 24) >> 24); 1926 } 1927 return LoadNode::Value(phase); 1928 } 1929 1930 //--------------------------LoadUBNode::Ideal------------------------------------- 1931 // 1932 // If the previous store is to the same address as this load, 1933 // and the value stored was larger than a byte, replace this load 1934 // with the value stored truncated to a byte. If no truncation is 1935 // needed, the replacement is done in LoadNode::Identity(). 1936 // 1937 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1938 Node* mem = in(MemNode::Memory); 1939 Node* value = can_see_stored_value(mem, phase); 1940 if (value && !phase->type(value)->higher_equal(_type)) 1941 return new AndINode(value, phase->intcon(0xFF)); 1942 // Identity call will handle the case where truncation is not needed. 1943 return LoadNode::Ideal(phase, can_reshape); 1944 } 1945 1946 const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1947 Node* mem = in(MemNode::Memory); 1948 Node* value = can_see_stored_value(mem,phase); 1949 if (value != NULL && value->is_Con() && 1950 !value->bottom_type()->higher_equal(_type)) { 1951 // If the input to the store does not fit with the load's result type, 1952 // it must be truncated. We can't delay until Ideal call since 1953 // a singleton Value is needed for split_thru_phi optimization. 1954 int con = value->get_int(); 1955 return TypeInt::make(con & 0xFF); 1956 } 1957 return LoadNode::Value(phase); 1958 } 1959 1960 //--------------------------LoadUSNode::Ideal------------------------------------- 1961 // 1962 // If the previous store is to the same address as this load, 1963 // and the value stored was larger than a char, replace this load 1964 // with the value stored truncated to a char. If no truncation is 1965 // needed, the replacement is done in LoadNode::Identity(). 1966 // 1967 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1968 Node* mem = in(MemNode::Memory); 1969 Node* value = can_see_stored_value(mem,phase); 1970 if( value && !phase->type(value)->higher_equal( _type ) ) 1971 return new AndINode(value,phase->intcon(0xFFFF)); 1972 // Identity call will handle the case where truncation is not needed. 1973 return LoadNode::Ideal(phase, can_reshape); 1974 } 1975 1976 const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1977 Node* mem = in(MemNode::Memory); 1978 Node* value = can_see_stored_value(mem,phase); 1979 if (value != NULL && value->is_Con() && 1980 !value->bottom_type()->higher_equal(_type)) { 1981 // If the input to the store does not fit with the load's result type, 1982 // it must be truncated. We can't delay until Ideal call since 1983 // a singleton Value is needed for split_thru_phi optimization. 1984 int con = value->get_int(); 1985 return TypeInt::make(con & 0xFFFF); 1986 } 1987 return LoadNode::Value(phase); 1988 } 1989 1990 //--------------------------LoadSNode::Ideal-------------------------------------- 1991 // 1992 // If the previous store is to the same address as this load, 1993 // and the value stored was larger than a short, replace this load 1994 // with the value stored truncated to a short. If no truncation is 1995 // needed, the replacement is done in LoadNode::Identity(). 1996 // 1997 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1998 Node* mem = in(MemNode::Memory); 1999 Node* value = can_see_stored_value(mem,phase); 2000 if( value && !phase->type(value)->higher_equal( _type ) ) { 2001 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 2002 return new RShiftINode(result, phase->intcon(16)); 2003 } 2004 // Identity call will handle the case where truncation is not needed. 2005 return LoadNode::Ideal(phase, can_reshape); 2006 } 2007 2008 const Type* LoadSNode::Value(PhaseTransform *phase) const { 2009 Node* mem = in(MemNode::Memory); 2010 Node* value = can_see_stored_value(mem,phase); 2011 if (value != NULL && value->is_Con() && 2012 !value->bottom_type()->higher_equal(_type)) { 2013 // If the input to the store does not fit with the load's result type, 2014 // it must be truncated. We can't delay until Ideal call since 2015 // a singleton Value is needed for split_thru_phi optimization. 2016 int con = value->get_int(); 2017 return TypeInt::make((con << 16) >> 16); 2018 } 2019 return LoadNode::Value(phase); 2020 } 2021 2022 //============================================================================= 2023 //----------------------------LoadKlassNode::make------------------------------ 2024 // Polymorphic factory method: 2025 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2026 // sanity check the alias category against the created node type 2027 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2028 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2029 #ifdef _LP64 2030 if (adr_type->is_ptr_to_narrowklass()) { 2031 assert(UseCompressedClassPointers, "no compressed klasses"); 2032 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2033 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2034 } 2035 #endif 2036 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2037 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2038 } 2039 2040 //------------------------------Value------------------------------------------ 2041 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 2042 return klass_value_common(phase); 2043 } 2044 2045 // In most cases, LoadKlassNode does not have the control input set. If the control 2046 // input is set, it must not be removed (by LoadNode::Ideal()). 2047 bool LoadKlassNode::can_remove_control() const { 2048 return false; 2049 } 2050 2051 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 2052 // Either input is TOP ==> the result is TOP 2053 const Type *t1 = phase->type( in(MemNode::Memory) ); 2054 if (t1 == Type::TOP) return Type::TOP; 2055 Node *adr = in(MemNode::Address); 2056 const Type *t2 = phase->type( adr ); 2057 if (t2 == Type::TOP) return Type::TOP; 2058 const TypePtr *tp = t2->is_ptr(); 2059 if (TypePtr::above_centerline(tp->ptr()) || 2060 tp->ptr() == TypePtr::Null) return Type::TOP; 2061 2062 // Return a more precise klass, if possible 2063 const TypeInstPtr *tinst = tp->isa_instptr(); 2064 if (tinst != NULL) { 2065 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2066 int offset = tinst->offset(); 2067 if (ik == phase->C->env()->Class_klass() 2068 && (offset == java_lang_Class::klass_offset_in_bytes() || 2069 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2070 // We are loading a special hidden field from a Class mirror object, 2071 // the field which points to the VM's Klass metaobject. 2072 ciType* t = tinst->java_mirror_type(); 2073 // java_mirror_type returns non-null for compile-time Class constants. 2074 if (t != NULL) { 2075 // constant oop => constant klass 2076 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2077 if (t->is_void()) { 2078 // We cannot create a void array. Since void is a primitive type return null 2079 // klass. Users of this result need to do a null check on the returned klass. 2080 return TypePtr::NULL_PTR; 2081 } 2082 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2083 } 2084 if (!t->is_klass()) { 2085 // a primitive Class (e.g., int.class) has NULL for a klass field 2086 return TypePtr::NULL_PTR; 2087 } 2088 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2089 return TypeKlassPtr::make(t->as_klass()); 2090 } 2091 // non-constant mirror, so we can't tell what's going on 2092 } 2093 if( !ik->is_loaded() ) 2094 return _type; // Bail out if not loaded 2095 if (offset == oopDesc::klass_offset_in_bytes()) { 2096 if (tinst->klass_is_exact()) { 2097 return TypeKlassPtr::make(ik); 2098 } 2099 // See if we can become precise: no subklasses and no interface 2100 // (Note: We need to support verified interfaces.) 2101 if (!ik->is_interface() && !ik->has_subklass()) { 2102 //assert(!UseExactTypes, "this code should be useless with exact types"); 2103 // Add a dependence; if any subclass added we need to recompile 2104 if (!ik->is_final()) { 2105 // %%% should use stronger assert_unique_concrete_subtype instead 2106 phase->C->dependencies()->assert_leaf_type(ik); 2107 } 2108 // Return precise klass 2109 return TypeKlassPtr::make(ik); 2110 } 2111 2112 // Return root of possible klass 2113 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2114 } 2115 } 2116 2117 // Check for loading klass from an array 2118 const TypeAryPtr *tary = tp->isa_aryptr(); 2119 if( tary != NULL ) { 2120 ciKlass *tary_klass = tary->klass(); 2121 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2122 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2123 if (tary->klass_is_exact()) { 2124 return TypeKlassPtr::make(tary_klass); 2125 } 2126 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2127 // If the klass is an object array, we defer the question to the 2128 // array component klass. 2129 if( ak->is_obj_array_klass() ) { 2130 assert( ak->is_loaded(), "" ); 2131 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2132 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2133 ciInstanceKlass* ik = base_k->as_instance_klass(); 2134 // See if we can become precise: no subklasses and no interface 2135 if (!ik->is_interface() && !ik->has_subklass()) { 2136 //assert(!UseExactTypes, "this code should be useless with exact types"); 2137 // Add a dependence; if any subclass added we need to recompile 2138 if (!ik->is_final()) { 2139 phase->C->dependencies()->assert_leaf_type(ik); 2140 } 2141 // Return precise array klass 2142 return TypeKlassPtr::make(ak); 2143 } 2144 } 2145 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2146 } else { // Found a type-array? 2147 //assert(!UseExactTypes, "this code should be useless with exact types"); 2148 assert( ak->is_type_array_klass(), "" ); 2149 return TypeKlassPtr::make(ak); // These are always precise 2150 } 2151 } 2152 } 2153 2154 // Check for loading klass from an array klass 2155 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2156 if (tkls != NULL && !StressReflectiveCode) { 2157 ciKlass* klass = tkls->klass(); 2158 if( !klass->is_loaded() ) 2159 return _type; // Bail out if not loaded 2160 if( klass->is_obj_array_klass() && 2161 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2162 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2163 // // Always returning precise element type is incorrect, 2164 // // e.g., element type could be object and array may contain strings 2165 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2166 2167 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2168 // according to the element type's subclassing. 2169 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2170 } 2171 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2172 tkls->offset() == in_bytes(Klass::super_offset())) { 2173 ciKlass* sup = klass->as_instance_klass()->super(); 2174 // The field is Klass::_super. Return its (constant) value. 2175 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2176 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2177 } 2178 } 2179 2180 // Bailout case 2181 return LoadNode::Value(phase); 2182 } 2183 2184 //------------------------------Identity--------------------------------------- 2185 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2186 // Also feed through the klass in Allocate(...klass...)._klass. 2187 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2188 return klass_identity_common(phase); 2189 } 2190 2191 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2192 Node* x = LoadNode::Identity(phase); 2193 if (x != this) return x; 2194 2195 // Take apart the address into an oop and and offset. 2196 // Return 'this' if we cannot. 2197 Node* adr = in(MemNode::Address); 2198 intptr_t offset = 0; 2199 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2200 if (base == NULL) return this; 2201 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2202 if (toop == NULL) return this; 2203 2204 // We can fetch the klass directly through an AllocateNode. 2205 // This works even if the klass is not constant (clone or newArray). 2206 if (offset == oopDesc::klass_offset_in_bytes()) { 2207 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2208 if (allocated_klass != NULL) { 2209 return allocated_klass; 2210 } 2211 } 2212 2213 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2214 // See inline_native_Class_query for occurrences of these patterns. 2215 // Java Example: x.getClass().isAssignableFrom(y) 2216 // 2217 // This improves reflective code, often making the Class 2218 // mirror go completely dead. (Current exception: Class 2219 // mirrors may appear in debug info, but we could clean them out by 2220 // introducing a new debug info operator for Klass*.java_mirror). 2221 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2222 && offset == java_lang_Class::klass_offset_in_bytes()) { 2223 // We are loading a special hidden field from a Class mirror, 2224 // the field which points to its Klass or ArrayKlass metaobject. 2225 if (base->is_Load()) { 2226 Node* adr2 = base->in(MemNode::Address); 2227 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2228 if (tkls != NULL && !tkls->empty() 2229 && (tkls->klass()->is_instance_klass() || 2230 tkls->klass()->is_array_klass()) 2231 && adr2->is_AddP() 2232 ) { 2233 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2234 if (tkls->offset() == mirror_field) { 2235 return adr2->in(AddPNode::Base); 2236 } 2237 } 2238 } 2239 } 2240 2241 return this; 2242 } 2243 2244 2245 //------------------------------Value------------------------------------------ 2246 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2247 const Type *t = klass_value_common(phase); 2248 if (t == Type::TOP) 2249 return t; 2250 2251 return t->make_narrowklass(); 2252 } 2253 2254 //------------------------------Identity--------------------------------------- 2255 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2256 // Also feed through the klass in Allocate(...klass...)._klass. 2257 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2258 Node *x = klass_identity_common(phase); 2259 2260 const Type *t = phase->type( x ); 2261 if( t == Type::TOP ) return x; 2262 if( t->isa_narrowklass()) return x; 2263 assert (!t->isa_narrowoop(), "no narrow oop here"); 2264 2265 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2266 } 2267 2268 //------------------------------Value----------------------------------------- 2269 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2270 // Either input is TOP ==> the result is TOP 2271 const Type *t1 = phase->type( in(MemNode::Memory) ); 2272 if( t1 == Type::TOP ) return Type::TOP; 2273 Node *adr = in(MemNode::Address); 2274 const Type *t2 = phase->type( adr ); 2275 if( t2 == Type::TOP ) return Type::TOP; 2276 const TypePtr *tp = t2->is_ptr(); 2277 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2278 const TypeAryPtr *tap = tp->isa_aryptr(); 2279 if( !tap ) return _type; 2280 return tap->size(); 2281 } 2282 2283 //-------------------------------Ideal--------------------------------------- 2284 // Feed through the length in AllocateArray(...length...)._length. 2285 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2286 Node* p = MemNode::Ideal_common(phase, can_reshape); 2287 if (p) return (p == NodeSentinel) ? NULL : p; 2288 2289 // Take apart the address into an oop and and offset. 2290 // Return 'this' if we cannot. 2291 Node* adr = in(MemNode::Address); 2292 intptr_t offset = 0; 2293 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2294 if (base == NULL) return NULL; 2295 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2296 if (tary == NULL) return NULL; 2297 2298 // We can fetch the length directly through an AllocateArrayNode. 2299 // This works even if the length is not constant (clone or newArray). 2300 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2301 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2302 if (alloc != NULL) { 2303 Node* allocated_length = alloc->Ideal_length(); 2304 Node* len = alloc->make_ideal_length(tary, phase); 2305 if (allocated_length != len) { 2306 // New CastII improves on this. 2307 return len; 2308 } 2309 } 2310 } 2311 2312 return NULL; 2313 } 2314 2315 //------------------------------Identity--------------------------------------- 2316 // Feed through the length in AllocateArray(...length...)._length. 2317 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2318 Node* x = LoadINode::Identity(phase); 2319 if (x != this) return x; 2320 2321 // Take apart the address into an oop and and offset. 2322 // Return 'this' if we cannot. 2323 Node* adr = in(MemNode::Address); 2324 intptr_t offset = 0; 2325 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2326 if (base == NULL) return this; 2327 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2328 if (tary == NULL) return this; 2329 2330 // We can fetch the length directly through an AllocateArrayNode. 2331 // This works even if the length is not constant (clone or newArray). 2332 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2333 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2334 if (alloc != NULL) { 2335 Node* allocated_length = alloc->Ideal_length(); 2336 // Do not allow make_ideal_length to allocate a CastII node. 2337 Node* len = alloc->make_ideal_length(tary, phase, false); 2338 if (allocated_length == len) { 2339 // Return allocated_length only if it would not be improved by a CastII. 2340 return allocated_length; 2341 } 2342 } 2343 } 2344 2345 return this; 2346 2347 } 2348 2349 //============================================================================= 2350 //---------------------------StoreNode::make----------------------------------- 2351 // Polymorphic factory method: 2352 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2353 assert((mo == unordered || mo == release), "unexpected"); 2354 Compile* C = gvn.C; 2355 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2356 ctl != NULL, "raw memory operations should have control edge"); 2357 2358 switch (bt) { 2359 case T_BOOLEAN: 2360 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2361 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2362 case T_CHAR: 2363 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2364 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2365 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2366 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2367 case T_METADATA: 2368 case T_ADDRESS: 2369 case T_OBJECT: 2370 #ifdef _LP64 2371 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2372 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2373 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2374 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2375 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2376 adr->bottom_type()->isa_rawptr())) { 2377 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2378 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2379 } 2380 #endif 2381 { 2382 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2383 } 2384 } 2385 ShouldNotReachHere(); 2386 return (StoreNode*)NULL; 2387 } 2388 2389 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2390 bool require_atomic = true; 2391 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2392 } 2393 2394 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2395 bool require_atomic = true; 2396 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2397 } 2398 2399 2400 //--------------------------bottom_type---------------------------------------- 2401 const Type *StoreNode::bottom_type() const { 2402 return Type::MEMORY; 2403 } 2404 2405 //------------------------------hash------------------------------------------- 2406 uint StoreNode::hash() const { 2407 // unroll addition of interesting fields 2408 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2409 2410 // Since they are not commoned, do not hash them: 2411 return NO_HASH; 2412 } 2413 2414 //------------------------------Ideal------------------------------------------ 2415 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2416 // When a store immediately follows a relevant allocation/initialization, 2417 // try to capture it into the initialization, or hoist it above. 2418 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2419 Node* p = MemNode::Ideal_common(phase, can_reshape); 2420 if (p) return (p == NodeSentinel) ? NULL : p; 2421 2422 Node* mem = in(MemNode::Memory); 2423 Node* address = in(MemNode::Address); 2424 // Back-to-back stores to same address? Fold em up. Generally 2425 // unsafe if I have intervening uses... Also disallowed for StoreCM 2426 // since they must follow each StoreP operation. Redundant StoreCMs 2427 // are eliminated just before matching in final_graph_reshape. 2428 { 2429 bool improved = false; 2430 Node* st = mem; 2431 Node* prev = this; 2432 // If anybody other than the previous Store on the memory chain 2433 // uses 'st', we cannot fold 'st' away. For example, 'st' 2434 // might be the final state at a conditional return. Or, 'st' 2435 // might be used by some node which is live at the same time 2436 // 'st' is live, which might be unschedulable. So, require 2437 // exactly ONE user, the 'prev' store, until such time as we clone 2438 // 'mem' for each of 'mem's uses (thus making the 2439 // exactly-1-user-rule hold true). 2440 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2441 // Looking at a dead closed cycle of memory? 2442 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2443 assert(Opcode() == st->Opcode() || 2444 st->Opcode() == Op_StoreVector || 2445 Opcode() == Op_StoreVector || 2446 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2447 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI), // expanded ClearArrayNode 2448 err_msg_res("no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()])); 2449 2450 if (st->in(MemNode::Address)->eqv_uncast(address) && 2451 st->as_Store()->memory_size() <= this->memory_size()) { 2452 phase->igvn_rehash_node_delayed(prev); 2453 if (can_reshape) { 2454 prev->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN()); 2455 } else { 2456 // It's OK to do this in the parser, since DU info is always accurate, 2457 // and the parser always refers to nodes via SafePointNode maps. 2458 prev->set_req(MemNode::Memory, st->in(MemNode::Memory)); 2459 } 2460 improved = (prev == this); 2461 } 2462 prev = st; 2463 st = st->in(MemNode::Memory); 2464 } 2465 if (improved) { 2466 return this; 2467 } 2468 } 2469 2470 2471 // Capture an unaliased, unconditional, simple store into an initializer. 2472 // Or, if it is independent of the allocation, hoist it above the allocation. 2473 if (ReduceFieldZeroing && /*can_reshape &&*/ 2474 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2475 InitializeNode* init = mem->in(0)->as_Initialize(); 2476 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2477 if (offset > 0) { 2478 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2479 // If the InitializeNode captured me, it made a raw copy of me, 2480 // and I need to disappear. 2481 if (moved != NULL) { 2482 // %%% hack to ensure that Ideal returns a new node: 2483 mem = MergeMemNode::make(mem); 2484 return mem; // fold me away 2485 } 2486 } 2487 } 2488 2489 return NULL; // No further progress 2490 } 2491 2492 //------------------------------Value----------------------------------------- 2493 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2494 // Either input is TOP ==> the result is TOP 2495 const Type *t1 = phase->type( in(MemNode::Memory) ); 2496 if( t1 == Type::TOP ) return Type::TOP; 2497 const Type *t2 = phase->type( in(MemNode::Address) ); 2498 if( t2 == Type::TOP ) return Type::TOP; 2499 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2500 if( t3 == Type::TOP ) return Type::TOP; 2501 return Type::MEMORY; 2502 } 2503 2504 //------------------------------Identity--------------------------------------- 2505 // Remove redundant stores: 2506 // Store(m, p, Load(m, p)) changes to m. 2507 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2508 Node *StoreNode::Identity( PhaseTransform *phase ) { 2509 Node* mem = in(MemNode::Memory); 2510 Node* adr = in(MemNode::Address); 2511 Node* val = in(MemNode::ValueIn); 2512 2513 // Load then Store? Then the Store is useless 2514 if (val->is_Load() && 2515 val->in(MemNode::Address)->eqv_uncast(adr) && 2516 val->in(MemNode::Memory )->eqv_uncast(mem) && 2517 val->as_Load()->store_Opcode() == Opcode()) { 2518 return mem; 2519 } 2520 2521 // Two stores in a row of the same value? 2522 if (mem->is_Store() && 2523 mem->in(MemNode::Address)->eqv_uncast(adr) && 2524 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2525 mem->Opcode() == Opcode()) { 2526 return mem; 2527 } 2528 2529 // Store of zero anywhere into a freshly-allocated object? 2530 // Then the store is useless. 2531 // (It must already have been captured by the InitializeNode.) 2532 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2533 // a newly allocated object is already all-zeroes everywhere 2534 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2535 return mem; 2536 } 2537 2538 // the store may also apply to zero-bits in an earlier object 2539 Node* prev_mem = find_previous_store(phase); 2540 // Steps (a), (b): Walk past independent stores to find an exact match. 2541 if (prev_mem != NULL) { 2542 Node* prev_val = can_see_stored_value(prev_mem, phase); 2543 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2544 // prev_val and val might differ by a cast; it would be good 2545 // to keep the more informative of the two. 2546 return mem; 2547 } 2548 } 2549 } 2550 2551 return this; 2552 } 2553 2554 //------------------------------match_edge------------------------------------- 2555 // Do we Match on this edge index or not? Match only memory & value 2556 uint StoreNode::match_edge(uint idx) const { 2557 return idx == MemNode::Address || idx == MemNode::ValueIn; 2558 } 2559 2560 //------------------------------cmp-------------------------------------------- 2561 // Do not common stores up together. They generally have to be split 2562 // back up anyways, so do not bother. 2563 uint StoreNode::cmp( const Node &n ) const { 2564 return (&n == this); // Always fail except on self 2565 } 2566 2567 //------------------------------Ideal_masked_input----------------------------- 2568 // Check for a useless mask before a partial-word store 2569 // (StoreB ... (AndI valIn conIa) ) 2570 // If (conIa & mask == mask) this simplifies to 2571 // (StoreB ... (valIn) ) 2572 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2573 Node *val = in(MemNode::ValueIn); 2574 if( val->Opcode() == Op_AndI ) { 2575 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2576 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2577 set_req(MemNode::ValueIn, val->in(1)); 2578 return this; 2579 } 2580 } 2581 return NULL; 2582 } 2583 2584 2585 //------------------------------Ideal_sign_extended_input---------------------- 2586 // Check for useless sign-extension before a partial-word store 2587 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2588 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2589 // (StoreB ... (valIn) ) 2590 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2591 Node *val = in(MemNode::ValueIn); 2592 if( val->Opcode() == Op_RShiftI ) { 2593 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2594 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2595 Node *shl = val->in(1); 2596 if( shl->Opcode() == Op_LShiftI ) { 2597 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2598 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2599 set_req(MemNode::ValueIn, shl->in(1)); 2600 return this; 2601 } 2602 } 2603 } 2604 } 2605 return NULL; 2606 } 2607 2608 //------------------------------value_never_loaded----------------------------------- 2609 // Determine whether there are any possible loads of the value stored. 2610 // For simplicity, we actually check if there are any loads from the 2611 // address stored to, not just for loads of the value stored by this node. 2612 // 2613 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2614 Node *adr = in(Address); 2615 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2616 if (adr_oop == NULL) 2617 return false; 2618 if (!adr_oop->is_known_instance_field()) 2619 return false; // if not a distinct instance, there may be aliases of the address 2620 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2621 Node *use = adr->fast_out(i); 2622 if (use->is_Load() || use->is_LoadStore()) { 2623 return false; 2624 } 2625 } 2626 return true; 2627 } 2628 2629 //============================================================================= 2630 //------------------------------Ideal------------------------------------------ 2631 // If the store is from an AND mask that leaves the low bits untouched, then 2632 // we can skip the AND operation. If the store is from a sign-extension 2633 // (a left shift, then right shift) we can skip both. 2634 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2635 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2636 if( progress != NULL ) return progress; 2637 2638 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2639 if( progress != NULL ) return progress; 2640 2641 // Finally check the default case 2642 return StoreNode::Ideal(phase, can_reshape); 2643 } 2644 2645 //============================================================================= 2646 //------------------------------Ideal------------------------------------------ 2647 // If the store is from an AND mask that leaves the low bits untouched, then 2648 // we can skip the AND operation 2649 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2650 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2651 if( progress != NULL ) return progress; 2652 2653 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2654 if( progress != NULL ) return progress; 2655 2656 // Finally check the default case 2657 return StoreNode::Ideal(phase, can_reshape); 2658 } 2659 2660 //============================================================================= 2661 //------------------------------Identity--------------------------------------- 2662 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2663 // No need to card mark when storing a null ptr 2664 Node* my_store = in(MemNode::OopStore); 2665 if (my_store->is_Store()) { 2666 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2667 if( t1 == TypePtr::NULL_PTR ) { 2668 return in(MemNode::Memory); 2669 } 2670 } 2671 return this; 2672 } 2673 2674 //============================================================================= 2675 //------------------------------Ideal--------------------------------------- 2676 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2677 Node* progress = StoreNode::Ideal(phase, can_reshape); 2678 if (progress != NULL) return progress; 2679 2680 Node* my_store = in(MemNode::OopStore); 2681 if (my_store->is_MergeMem()) { 2682 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2683 set_req(MemNode::OopStore, mem); 2684 return this; 2685 } 2686 2687 return NULL; 2688 } 2689 2690 //------------------------------Value----------------------------------------- 2691 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2692 // Either input is TOP ==> the result is TOP 2693 const Type *t = phase->type( in(MemNode::Memory) ); 2694 if( t == Type::TOP ) return Type::TOP; 2695 t = phase->type( in(MemNode::Address) ); 2696 if( t == Type::TOP ) return Type::TOP; 2697 t = phase->type( in(MemNode::ValueIn) ); 2698 if( t == Type::TOP ) return Type::TOP; 2699 // If extra input is TOP ==> the result is TOP 2700 t = phase->type( in(MemNode::OopStore) ); 2701 if( t == Type::TOP ) return Type::TOP; 2702 2703 return StoreNode::Value( phase ); 2704 } 2705 2706 2707 //============================================================================= 2708 //----------------------------------SCMemProjNode------------------------------ 2709 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2710 { 2711 return bottom_type(); 2712 } 2713 2714 //============================================================================= 2715 //----------------------------------LoadStoreNode------------------------------ 2716 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2717 : Node(required), 2718 _type(rt), 2719 _adr_type(at) 2720 { 2721 init_req(MemNode::Control, c ); 2722 init_req(MemNode::Memory , mem); 2723 init_req(MemNode::Address, adr); 2724 init_req(MemNode::ValueIn, val); 2725 init_class_id(Class_LoadStore); 2726 } 2727 2728 uint LoadStoreNode::ideal_reg() const { 2729 return _type->ideal_reg(); 2730 } 2731 2732 bool LoadStoreNode::result_not_used() const { 2733 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2734 Node *x = fast_out(i); 2735 if (x->Opcode() == Op_SCMemProj) continue; 2736 return false; 2737 } 2738 return true; 2739 } 2740 2741 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2742 2743 //============================================================================= 2744 //----------------------------------LoadStoreConditionalNode-------------------- 2745 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2746 init_req(ExpectedIn, ex ); 2747 } 2748 2749 //============================================================================= 2750 //-------------------------------adr_type-------------------------------------- 2751 const TypePtr* ClearArrayNode::adr_type() const { 2752 Node *adr = in(3); 2753 if (adr == NULL) return NULL; // node is dead 2754 return MemNode::calculate_adr_type(adr->bottom_type()); 2755 } 2756 2757 //------------------------------match_edge------------------------------------- 2758 // Do we Match on this edge index or not? Do not match memory 2759 uint ClearArrayNode::match_edge(uint idx) const { 2760 return idx > 1; 2761 } 2762 2763 //------------------------------Identity--------------------------------------- 2764 // Clearing a zero length array does nothing 2765 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2766 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2767 } 2768 2769 //------------------------------Idealize--------------------------------------- 2770 // Clearing a short array is faster with stores 2771 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2772 const int unit = BytesPerLong; 2773 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2774 if (!t) return NULL; 2775 if (!t->is_con()) return NULL; 2776 intptr_t raw_count = t->get_con(); 2777 intptr_t size = raw_count; 2778 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2779 // Clearing nothing uses the Identity call. 2780 // Negative clears are possible on dead ClearArrays 2781 // (see jck test stmt114.stmt11402.val). 2782 if (size <= 0 || size % unit != 0) return NULL; 2783 intptr_t count = size / unit; 2784 // Length too long; use fast hardware clear 2785 if (size > Matcher::init_array_short_size) return NULL; 2786 Node *mem = in(1); 2787 if( phase->type(mem)==Type::TOP ) return NULL; 2788 Node *adr = in(3); 2789 const Type* at = phase->type(adr); 2790 if( at==Type::TOP ) return NULL; 2791 const TypePtr* atp = at->isa_ptr(); 2792 // adjust atp to be the correct array element address type 2793 if (atp == NULL) atp = TypePtr::BOTTOM; 2794 else atp = atp->add_offset(Type::OffsetBot); 2795 // Get base for derived pointer purposes 2796 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2797 Node *base = adr->in(1); 2798 2799 Node *zero = phase->makecon(TypeLong::ZERO); 2800 Node *off = phase->MakeConX(BytesPerLong); 2801 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2802 count--; 2803 while( count-- ) { 2804 mem = phase->transform(mem); 2805 adr = phase->transform(new AddPNode(base,adr,off)); 2806 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2807 } 2808 return mem; 2809 } 2810 2811 //----------------------------step_through---------------------------------- 2812 // Return allocation input memory edge if it is different instance 2813 // or itself if it is the one we are looking for. 2814 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2815 Node* n = *np; 2816 assert(n->is_ClearArray(), "sanity"); 2817 intptr_t offset; 2818 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2819 // This method is called only before Allocate nodes are expanded 2820 // during macro nodes expansion. Before that ClearArray nodes are 2821 // only generated in PhaseMacroExpand::generate_arraycopy() (before 2822 // Allocate nodes are expanded) which follows allocations. 2823 assert(alloc != NULL, "should have allocation"); 2824 if (alloc->_idx == instance_id) { 2825 // Can not bypass initialization of the instance we are looking for. 2826 return false; 2827 } 2828 // Otherwise skip it. 2829 InitializeNode* init = alloc->initialization(); 2830 if (init != NULL) 2831 *np = init->in(TypeFunc::Memory); 2832 else 2833 *np = alloc->in(TypeFunc::Memory); 2834 return true; 2835 } 2836 2837 //----------------------------clear_memory------------------------------------- 2838 // Generate code to initialize object storage to zero. 2839 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2840 intptr_t start_offset, 2841 Node* end_offset, 2842 PhaseGVN* phase) { 2843 intptr_t offset = start_offset; 2844 2845 int unit = BytesPerLong; 2846 if ((offset % unit) != 0) { 2847 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2848 adr = phase->transform(adr); 2849 const TypePtr* atp = TypeRawPtr::BOTTOM; 2850 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2851 mem = phase->transform(mem); 2852 offset += BytesPerInt; 2853 } 2854 assert((offset % unit) == 0, ""); 2855 2856 // Initialize the remaining stuff, if any, with a ClearArray. 2857 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2858 } 2859 2860 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2861 Node* start_offset, 2862 Node* end_offset, 2863 PhaseGVN* phase) { 2864 if (start_offset == end_offset) { 2865 // nothing to do 2866 return mem; 2867 } 2868 2869 int unit = BytesPerLong; 2870 Node* zbase = start_offset; 2871 Node* zend = end_offset; 2872 2873 // Scale to the unit required by the CPU: 2874 if (!Matcher::init_array_count_is_in_bytes) { 2875 Node* shift = phase->intcon(exact_log2(unit)); 2876 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2877 zend = phase->transform(new URShiftXNode(zend, shift) ); 2878 } 2879 2880 // Bulk clear double-words 2881 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2882 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2883 mem = new ClearArrayNode(ctl, mem, zsize, adr); 2884 return phase->transform(mem); 2885 } 2886 2887 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2888 intptr_t start_offset, 2889 intptr_t end_offset, 2890 PhaseGVN* phase) { 2891 if (start_offset == end_offset) { 2892 // nothing to do 2893 return mem; 2894 } 2895 2896 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2897 intptr_t done_offset = end_offset; 2898 if ((done_offset % BytesPerLong) != 0) { 2899 done_offset -= BytesPerInt; 2900 } 2901 if (done_offset > start_offset) { 2902 mem = clear_memory(ctl, mem, dest, 2903 start_offset, phase->MakeConX(done_offset), phase); 2904 } 2905 if (done_offset < end_offset) { // emit the final 32-bit store 2906 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 2907 adr = phase->transform(adr); 2908 const TypePtr* atp = TypeRawPtr::BOTTOM; 2909 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2910 mem = phase->transform(mem); 2911 done_offset += BytesPerInt; 2912 } 2913 assert(done_offset == end_offset, ""); 2914 return mem; 2915 } 2916 2917 //============================================================================= 2918 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2919 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2920 _adr_type(C->get_adr_type(alias_idx)) 2921 { 2922 init_class_id(Class_MemBar); 2923 Node* top = C->top(); 2924 init_req(TypeFunc::I_O,top); 2925 init_req(TypeFunc::FramePtr,top); 2926 init_req(TypeFunc::ReturnAdr,top); 2927 if (precedent != NULL) 2928 init_req(TypeFunc::Parms, precedent); 2929 } 2930 2931 //------------------------------cmp-------------------------------------------- 2932 uint MemBarNode::hash() const { return NO_HASH; } 2933 uint MemBarNode::cmp( const Node &n ) const { 2934 return (&n == this); // Always fail except on self 2935 } 2936 2937 //------------------------------make------------------------------------------- 2938 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2939 switch (opcode) { 2940 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 2941 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 2942 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 2943 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 2944 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 2945 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 2946 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 2947 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 2948 case Op_Initialize: return new InitializeNode(C, atp, pn); 2949 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 2950 default: ShouldNotReachHere(); return NULL; 2951 } 2952 } 2953 2954 //------------------------------Ideal------------------------------------------ 2955 // Return a node which is more "ideal" than the current node. Strip out 2956 // control copies 2957 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2958 if (remove_dead_region(phase, can_reshape)) return this; 2959 // Don't bother trying to transform a dead node 2960 if (in(0) && in(0)->is_top()) { 2961 return NULL; 2962 } 2963 2964 bool progress = false; 2965 // Eliminate volatile MemBars for scalar replaced objects. 2966 if (can_reshape && req() == (Precedent+1)) { 2967 bool eliminate = false; 2968 int opc = Opcode(); 2969 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2970 // Volatile field loads and stores. 2971 Node* my_mem = in(MemBarNode::Precedent); 2972 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2973 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2974 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2975 // replace this Precedent (decodeN) with the Load instead. 2976 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2977 Node* load_node = my_mem->in(1); 2978 set_req(MemBarNode::Precedent, load_node); 2979 phase->is_IterGVN()->_worklist.push(my_mem); 2980 my_mem = load_node; 2981 } else { 2982 assert(my_mem->unique_out() == this, "sanity"); 2983 del_req(Precedent); 2984 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2985 my_mem = NULL; 2986 } 2987 progress = true; 2988 } 2989 if (my_mem != NULL && my_mem->is_Mem()) { 2990 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2991 // Check for scalar replaced object reference. 2992 if( t_oop != NULL && t_oop->is_known_instance_field() && 2993 t_oop->offset() != Type::OffsetBot && 2994 t_oop->offset() != Type::OffsetTop) { 2995 eliminate = true; 2996 } 2997 } 2998 } else if (opc == Op_MemBarRelease) { 2999 // Final field stores. 3000 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 3001 if ((alloc != NULL) && alloc->is_Allocate() && 3002 alloc->as_Allocate()->_is_non_escaping) { 3003 // The allocated object does not escape. 3004 eliminate = true; 3005 } 3006 } 3007 if (eliminate) { 3008 // Replace MemBar projections by its inputs. 3009 PhaseIterGVN* igvn = phase->is_IterGVN(); 3010 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3011 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3012 // Must return either the original node (now dead) or a new node 3013 // (Do not return a top here, since that would break the uniqueness of top.) 3014 return new ConINode(TypeInt::ZERO); 3015 } 3016 } 3017 return progress ? this : NULL; 3018 } 3019 3020 //------------------------------Value------------------------------------------ 3021 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 3022 if( !in(0) ) return Type::TOP; 3023 if( phase->type(in(0)) == Type::TOP ) 3024 return Type::TOP; 3025 return TypeTuple::MEMBAR; 3026 } 3027 3028 //------------------------------match------------------------------------------ 3029 // Construct projections for memory. 3030 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3031 switch (proj->_con) { 3032 case TypeFunc::Control: 3033 case TypeFunc::Memory: 3034 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3035 } 3036 ShouldNotReachHere(); 3037 return NULL; 3038 } 3039 3040 //===========================InitializeNode==================================== 3041 // SUMMARY: 3042 // This node acts as a memory barrier on raw memory, after some raw stores. 3043 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3044 // The Initialize can 'capture' suitably constrained stores as raw inits. 3045 // It can coalesce related raw stores into larger units (called 'tiles'). 3046 // It can avoid zeroing new storage for memory units which have raw inits. 3047 // At macro-expansion, it is marked 'complete', and does not optimize further. 3048 // 3049 // EXAMPLE: 3050 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3051 // ctl = incoming control; mem* = incoming memory 3052 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3053 // First allocate uninitialized memory and fill in the header: 3054 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3055 // ctl := alloc.Control; mem* := alloc.Memory* 3056 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3057 // Then initialize to zero the non-header parts of the raw memory block: 3058 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3059 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3060 // After the initialize node executes, the object is ready for service: 3061 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3062 // Suppose its body is immediately initialized as {1,2}: 3063 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3064 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3065 // mem.SLICE(#short[*]) := store2 3066 // 3067 // DETAILS: 3068 // An InitializeNode collects and isolates object initialization after 3069 // an AllocateNode and before the next possible safepoint. As a 3070 // memory barrier (MemBarNode), it keeps critical stores from drifting 3071 // down past any safepoint or any publication of the allocation. 3072 // Before this barrier, a newly-allocated object may have uninitialized bits. 3073 // After this barrier, it may be treated as a real oop, and GC is allowed. 3074 // 3075 // The semantics of the InitializeNode include an implicit zeroing of 3076 // the new object from object header to the end of the object. 3077 // (The object header and end are determined by the AllocateNode.) 3078 // 3079 // Certain stores may be added as direct inputs to the InitializeNode. 3080 // These stores must update raw memory, and they must be to addresses 3081 // derived from the raw address produced by AllocateNode, and with 3082 // a constant offset. They must be ordered by increasing offset. 3083 // The first one is at in(RawStores), the last at in(req()-1). 3084 // Unlike most memory operations, they are not linked in a chain, 3085 // but are displayed in parallel as users of the rawmem output of 3086 // the allocation. 3087 // 3088 // (See comments in InitializeNode::capture_store, which continue 3089 // the example given above.) 3090 // 3091 // When the associated Allocate is macro-expanded, the InitializeNode 3092 // may be rewritten to optimize collected stores. A ClearArrayNode 3093 // may also be created at that point to represent any required zeroing. 3094 // The InitializeNode is then marked 'complete', prohibiting further 3095 // capturing of nearby memory operations. 3096 // 3097 // During macro-expansion, all captured initializations which store 3098 // constant values of 32 bits or smaller are coalesced (if advantageous) 3099 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3100 // initialized in fewer memory operations. Memory words which are 3101 // covered by neither tiles nor non-constant stores are pre-zeroed 3102 // by explicit stores of zero. (The code shape happens to do all 3103 // zeroing first, then all other stores, with both sequences occurring 3104 // in order of ascending offsets.) 3105 // 3106 // Alternatively, code may be inserted between an AllocateNode and its 3107 // InitializeNode, to perform arbitrary initialization of the new object. 3108 // E.g., the object copying intrinsics insert complex data transfers here. 3109 // The initialization must then be marked as 'complete' disable the 3110 // built-in zeroing semantics and the collection of initializing stores. 3111 // 3112 // While an InitializeNode is incomplete, reads from the memory state 3113 // produced by it are optimizable if they match the control edge and 3114 // new oop address associated with the allocation/initialization. 3115 // They return a stored value (if the offset matches) or else zero. 3116 // A write to the memory state, if it matches control and address, 3117 // and if it is to a constant offset, may be 'captured' by the 3118 // InitializeNode. It is cloned as a raw memory operation and rewired 3119 // inside the initialization, to the raw oop produced by the allocation. 3120 // Operations on addresses which are provably distinct (e.g., to 3121 // other AllocateNodes) are allowed to bypass the initialization. 3122 // 3123 // The effect of all this is to consolidate object initialization 3124 // (both arrays and non-arrays, both piecewise and bulk) into a 3125 // single location, where it can be optimized as a unit. 3126 // 3127 // Only stores with an offset less than TrackedInitializationLimit words 3128 // will be considered for capture by an InitializeNode. This puts a 3129 // reasonable limit on the complexity of optimized initializations. 3130 3131 //---------------------------InitializeNode------------------------------------ 3132 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3133 : _is_complete(Incomplete), _does_not_escape(false), 3134 MemBarNode(C, adr_type, rawoop) 3135 { 3136 init_class_id(Class_Initialize); 3137 3138 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3139 assert(in(RawAddress) == rawoop, "proper init"); 3140 // Note: allocation() can be NULL, for secondary initialization barriers 3141 } 3142 3143 // Since this node is not matched, it will be processed by the 3144 // register allocator. Declare that there are no constraints 3145 // on the allocation of the RawAddress edge. 3146 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3147 // This edge should be set to top, by the set_complete. But be conservative. 3148 if (idx == InitializeNode::RawAddress) 3149 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3150 return RegMask::Empty; 3151 } 3152 3153 Node* InitializeNode::memory(uint alias_idx) { 3154 Node* mem = in(Memory); 3155 if (mem->is_MergeMem()) { 3156 return mem->as_MergeMem()->memory_at(alias_idx); 3157 } else { 3158 // incoming raw memory is not split 3159 return mem; 3160 } 3161 } 3162 3163 bool InitializeNode::is_non_zero() { 3164 if (is_complete()) return false; 3165 remove_extra_zeroes(); 3166 return (req() > RawStores); 3167 } 3168 3169 void InitializeNode::set_complete(PhaseGVN* phase) { 3170 assert(!is_complete(), "caller responsibility"); 3171 _is_complete = Complete; 3172 3173 // After this node is complete, it contains a bunch of 3174 // raw-memory initializations. There is no need for 3175 // it to have anything to do with non-raw memory effects. 3176 // Therefore, tell all non-raw users to re-optimize themselves, 3177 // after skipping the memory effects of this initialization. 3178 PhaseIterGVN* igvn = phase->is_IterGVN(); 3179 if (igvn) igvn->add_users_to_worklist(this); 3180 } 3181 3182 // convenience function 3183 // return false if the init contains any stores already 3184 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3185 InitializeNode* init = initialization(); 3186 if (init == NULL || init->is_complete()) return false; 3187 init->remove_extra_zeroes(); 3188 // for now, if this allocation has already collected any inits, bail: 3189 if (init->is_non_zero()) return false; 3190 init->set_complete(phase); 3191 return true; 3192 } 3193 3194 void InitializeNode::remove_extra_zeroes() { 3195 if (req() == RawStores) return; 3196 Node* zmem = zero_memory(); 3197 uint fill = RawStores; 3198 for (uint i = fill; i < req(); i++) { 3199 Node* n = in(i); 3200 if (n->is_top() || n == zmem) continue; // skip 3201 if (fill < i) set_req(fill, n); // compact 3202 ++fill; 3203 } 3204 // delete any empty spaces created: 3205 while (fill < req()) { 3206 del_req(fill); 3207 } 3208 } 3209 3210 // Helper for remembering which stores go with which offsets. 3211 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3212 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3213 intptr_t offset = -1; 3214 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3215 phase, offset); 3216 if (base == NULL) return -1; // something is dead, 3217 if (offset < 0) return -1; // dead, dead 3218 return offset; 3219 } 3220 3221 // Helper for proving that an initialization expression is 3222 // "simple enough" to be folded into an object initialization. 3223 // Attempts to prove that a store's initial value 'n' can be captured 3224 // within the initialization without creating a vicious cycle, such as: 3225 // { Foo p = new Foo(); p.next = p; } 3226 // True for constants and parameters and small combinations thereof. 3227 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3228 if (n == NULL) return true; // (can this really happen?) 3229 if (n->is_Proj()) n = n->in(0); 3230 if (n == this) return false; // found a cycle 3231 if (n->is_Con()) return true; 3232 if (n->is_Start()) return true; // params, etc., are OK 3233 if (n->is_Root()) return true; // even better 3234 3235 Node* ctl = n->in(0); 3236 if (ctl != NULL && !ctl->is_top()) { 3237 if (ctl->is_Proj()) ctl = ctl->in(0); 3238 if (ctl == this) return false; 3239 3240 // If we already know that the enclosing memory op is pinned right after 3241 // the init, then any control flow that the store has picked up 3242 // must have preceded the init, or else be equal to the init. 3243 // Even after loop optimizations (which might change control edges) 3244 // a store is never pinned *before* the availability of its inputs. 3245 if (!MemNode::all_controls_dominate(n, this)) 3246 return false; // failed to prove a good control 3247 } 3248 3249 // Check data edges for possible dependencies on 'this'. 3250 if ((count += 1) > 20) return false; // complexity limit 3251 for (uint i = 1; i < n->req(); i++) { 3252 Node* m = n->in(i); 3253 if (m == NULL || m == n || m->is_top()) continue; 3254 uint first_i = n->find_edge(m); 3255 if (i != first_i) continue; // process duplicate edge just once 3256 if (!detect_init_independence(m, count)) { 3257 return false; 3258 } 3259 } 3260 3261 return true; 3262 } 3263 3264 // Here are all the checks a Store must pass before it can be moved into 3265 // an initialization. Returns zero if a check fails. 3266 // On success, returns the (constant) offset to which the store applies, 3267 // within the initialized memory. 3268 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3269 const int FAIL = 0; 3270 if (st->req() != MemNode::ValueIn + 1) 3271 return FAIL; // an inscrutable StoreNode (card mark?) 3272 Node* ctl = st->in(MemNode::Control); 3273 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3274 return FAIL; // must be unconditional after the initialization 3275 Node* mem = st->in(MemNode::Memory); 3276 if (!(mem->is_Proj() && mem->in(0) == this)) 3277 return FAIL; // must not be preceded by other stores 3278 Node* adr = st->in(MemNode::Address); 3279 intptr_t offset; 3280 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3281 if (alloc == NULL) 3282 return FAIL; // inscrutable address 3283 if (alloc != allocation()) 3284 return FAIL; // wrong allocation! (store needs to float up) 3285 Node* val = st->in(MemNode::ValueIn); 3286 int complexity_count = 0; 3287 if (!detect_init_independence(val, complexity_count)) 3288 return FAIL; // stored value must be 'simple enough' 3289 3290 // The Store can be captured only if nothing after the allocation 3291 // and before the Store is using the memory location that the store 3292 // overwrites. 3293 bool failed = false; 3294 // If is_complete_with_arraycopy() is true the shape of the graph is 3295 // well defined and is safe so no need for extra checks. 3296 if (!is_complete_with_arraycopy()) { 3297 // We are going to look at each use of the memory state following 3298 // the allocation to make sure nothing reads the memory that the 3299 // Store writes. 3300 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3301 int alias_idx = phase->C->get_alias_index(t_adr); 3302 ResourceMark rm; 3303 Unique_Node_List mems; 3304 mems.push(mem); 3305 Node* unique_merge = NULL; 3306 for (uint next = 0; next < mems.size(); ++next) { 3307 Node *m = mems.at(next); 3308 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3309 Node *n = m->fast_out(j); 3310 if (n->outcnt() == 0) { 3311 continue; 3312 } 3313 if (n == st) { 3314 continue; 3315 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3316 // If the control of this use is different from the control 3317 // of the Store which is right after the InitializeNode then 3318 // this node cannot be between the InitializeNode and the 3319 // Store. 3320 continue; 3321 } else if (n->is_MergeMem()) { 3322 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3323 // We can hit a MergeMemNode (that will likely go away 3324 // later) that is a direct use of the memory state 3325 // following the InitializeNode on the same slice as the 3326 // store node that we'd like to capture. We need to check 3327 // the uses of the MergeMemNode. 3328 mems.push(n); 3329 } 3330 } else if (n->is_Mem()) { 3331 Node* other_adr = n->in(MemNode::Address); 3332 if (other_adr == adr) { 3333 failed = true; 3334 break; 3335 } else { 3336 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3337 if (other_t_adr != NULL) { 3338 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3339 if (other_alias_idx == alias_idx) { 3340 // A load from the same memory slice as the store right 3341 // after the InitializeNode. We check the control of the 3342 // object/array that is loaded from. If it's the same as 3343 // the store control then we cannot capture the store. 3344 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3345 Node* base = other_adr; 3346 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); 3347 base = base->in(AddPNode::Base); 3348 if (base != NULL) { 3349 base = base->uncast(); 3350 if (base->is_Proj() && base->in(0) == alloc) { 3351 failed = true; 3352 break; 3353 } 3354 } 3355 } 3356 } 3357 } 3358 } else { 3359 failed = true; 3360 break; 3361 } 3362 } 3363 } 3364 } 3365 if (failed) { 3366 if (!can_reshape) { 3367 // We decided we couldn't capture the store during parsing. We 3368 // should try again during the next IGVN once the graph is 3369 // cleaner. 3370 phase->C->record_for_igvn(st); 3371 } 3372 return FAIL; 3373 } 3374 3375 return offset; // success 3376 } 3377 3378 // Find the captured store in(i) which corresponds to the range 3379 // [start..start+size) in the initialized object. 3380 // If there is one, return its index i. If there isn't, return the 3381 // negative of the index where it should be inserted. 3382 // Return 0 if the queried range overlaps an initialization boundary 3383 // or if dead code is encountered. 3384 // If size_in_bytes is zero, do not bother with overlap checks. 3385 int InitializeNode::captured_store_insertion_point(intptr_t start, 3386 int size_in_bytes, 3387 PhaseTransform* phase) { 3388 const int FAIL = 0, MAX_STORE = BytesPerLong; 3389 3390 if (is_complete()) 3391 return FAIL; // arraycopy got here first; punt 3392 3393 assert(allocation() != NULL, "must be present"); 3394 3395 // no negatives, no header fields: 3396 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3397 3398 // after a certain size, we bail out on tracking all the stores: 3399 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3400 if (start >= ti_limit) return FAIL; 3401 3402 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3403 if (i >= limit) return -(int)i; // not found; here is where to put it 3404 3405 Node* st = in(i); 3406 intptr_t st_off = get_store_offset(st, phase); 3407 if (st_off < 0) { 3408 if (st != zero_memory()) { 3409 return FAIL; // bail out if there is dead garbage 3410 } 3411 } else if (st_off > start) { 3412 // ...we are done, since stores are ordered 3413 if (st_off < start + size_in_bytes) { 3414 return FAIL; // the next store overlaps 3415 } 3416 return -(int)i; // not found; here is where to put it 3417 } else if (st_off < start) { 3418 if (size_in_bytes != 0 && 3419 start < st_off + MAX_STORE && 3420 start < st_off + st->as_Store()->memory_size()) { 3421 return FAIL; // the previous store overlaps 3422 } 3423 } else { 3424 if (size_in_bytes != 0 && 3425 st->as_Store()->memory_size() != size_in_bytes) { 3426 return FAIL; // mismatched store size 3427 } 3428 return i; 3429 } 3430 3431 ++i; 3432 } 3433 } 3434 3435 // Look for a captured store which initializes at the offset 'start' 3436 // with the given size. If there is no such store, and no other 3437 // initialization interferes, then return zero_memory (the memory 3438 // projection of the AllocateNode). 3439 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3440 PhaseTransform* phase) { 3441 assert(stores_are_sane(phase), ""); 3442 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3443 if (i == 0) { 3444 return NULL; // something is dead 3445 } else if (i < 0) { 3446 return zero_memory(); // just primordial zero bits here 3447 } else { 3448 Node* st = in(i); // here is the store at this position 3449 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3450 return st; 3451 } 3452 } 3453 3454 // Create, as a raw pointer, an address within my new object at 'offset'. 3455 Node* InitializeNode::make_raw_address(intptr_t offset, 3456 PhaseTransform* phase) { 3457 Node* addr = in(RawAddress); 3458 if (offset != 0) { 3459 Compile* C = phase->C; 3460 addr = phase->transform( new AddPNode(C->top(), addr, 3461 phase->MakeConX(offset)) ); 3462 } 3463 return addr; 3464 } 3465 3466 // Clone the given store, converting it into a raw store 3467 // initializing a field or element of my new object. 3468 // Caller is responsible for retiring the original store, 3469 // with subsume_node or the like. 3470 // 3471 // From the example above InitializeNode::InitializeNode, 3472 // here are the old stores to be captured: 3473 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3474 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3475 // 3476 // Here is the changed code; note the extra edges on init: 3477 // alloc = (Allocate ...) 3478 // rawoop = alloc.RawAddress 3479 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3480 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3481 // init = (Initialize alloc.Control alloc.Memory rawoop 3482 // rawstore1 rawstore2) 3483 // 3484 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3485 PhaseTransform* phase, bool can_reshape) { 3486 assert(stores_are_sane(phase), ""); 3487 3488 if (start < 0) return NULL; 3489 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3490 3491 Compile* C = phase->C; 3492 int size_in_bytes = st->memory_size(); 3493 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3494 if (i == 0) return NULL; // bail out 3495 Node* prev_mem = NULL; // raw memory for the captured store 3496 if (i > 0) { 3497 prev_mem = in(i); // there is a pre-existing store under this one 3498 set_req(i, C->top()); // temporarily disconnect it 3499 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3500 } else { 3501 i = -i; // no pre-existing store 3502 prev_mem = zero_memory(); // a slice of the newly allocated object 3503 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3504 set_req(--i, C->top()); // reuse this edge; it has been folded away 3505 else 3506 ins_req(i, C->top()); // build a new edge 3507 } 3508 Node* new_st = st->clone(); 3509 new_st->set_req(MemNode::Control, in(Control)); 3510 new_st->set_req(MemNode::Memory, prev_mem); 3511 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3512 new_st = phase->transform(new_st); 3513 3514 // At this point, new_st might have swallowed a pre-existing store 3515 // at the same offset, or perhaps new_st might have disappeared, 3516 // if it redundantly stored the same value (or zero to fresh memory). 3517 3518 // In any case, wire it in: 3519 phase->igvn_rehash_node_delayed(this); 3520 set_req(i, new_st); 3521 3522 // The caller may now kill the old guy. 3523 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3524 assert(check_st == new_st || check_st == NULL, "must be findable"); 3525 assert(!is_complete(), ""); 3526 return new_st; 3527 } 3528 3529 static bool store_constant(jlong* tiles, int num_tiles, 3530 intptr_t st_off, int st_size, 3531 jlong con) { 3532 if ((st_off & (st_size-1)) != 0) 3533 return false; // strange store offset (assume size==2**N) 3534 address addr = (address)tiles + st_off; 3535 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3536 switch (st_size) { 3537 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3538 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3539 case sizeof(jint): *(jint*) addr = (jint) con; break; 3540 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3541 default: return false; // strange store size (detect size!=2**N here) 3542 } 3543 return true; // return success to caller 3544 } 3545 3546 // Coalesce subword constants into int constants and possibly 3547 // into long constants. The goal, if the CPU permits, 3548 // is to initialize the object with a small number of 64-bit tiles. 3549 // Also, convert floating-point constants to bit patterns. 3550 // Non-constants are not relevant to this pass. 3551 // 3552 // In terms of the running example on InitializeNode::InitializeNode 3553 // and InitializeNode::capture_store, here is the transformation 3554 // of rawstore1 and rawstore2 into rawstore12: 3555 // alloc = (Allocate ...) 3556 // rawoop = alloc.RawAddress 3557 // tile12 = 0x00010002 3558 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3559 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3560 // 3561 void 3562 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3563 Node* size_in_bytes, 3564 PhaseGVN* phase) { 3565 Compile* C = phase->C; 3566 3567 assert(stores_are_sane(phase), ""); 3568 // Note: After this pass, they are not completely sane, 3569 // since there may be some overlaps. 3570 3571 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3572 3573 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3574 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3575 size_limit = MIN2(size_limit, ti_limit); 3576 size_limit = align_size_up(size_limit, BytesPerLong); 3577 int num_tiles = size_limit / BytesPerLong; 3578 3579 // allocate space for the tile map: 3580 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3581 jlong tiles_buf[small_len]; 3582 Node* nodes_buf[small_len]; 3583 jlong inits_buf[small_len]; 3584 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3585 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3586 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3587 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3588 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3589 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3590 // tiles: exact bitwise model of all primitive constants 3591 // nodes: last constant-storing node subsumed into the tiles model 3592 // inits: which bytes (in each tile) are touched by any initializations 3593 3594 //// Pass A: Fill in the tile model with any relevant stores. 3595 3596 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3597 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3598 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3599 Node* zmem = zero_memory(); // initially zero memory state 3600 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3601 Node* st = in(i); 3602 intptr_t st_off = get_store_offset(st, phase); 3603 3604 // Figure out the store's offset and constant value: 3605 if (st_off < header_size) continue; //skip (ignore header) 3606 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3607 int st_size = st->as_Store()->memory_size(); 3608 if (st_off + st_size > size_limit) break; 3609 3610 // Record which bytes are touched, whether by constant or not. 3611 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3612 continue; // skip (strange store size) 3613 3614 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3615 if (!val->singleton()) continue; //skip (non-con store) 3616 BasicType type = val->basic_type(); 3617 3618 jlong con = 0; 3619 switch (type) { 3620 case T_INT: con = val->is_int()->get_con(); break; 3621 case T_LONG: con = val->is_long()->get_con(); break; 3622 case T_FLOAT: con = jint_cast(val->getf()); break; 3623 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3624 default: continue; //skip (odd store type) 3625 } 3626 3627 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3628 st->Opcode() == Op_StoreL) { 3629 continue; // This StoreL is already optimal. 3630 } 3631 3632 // Store down the constant. 3633 store_constant(tiles, num_tiles, st_off, st_size, con); 3634 3635 intptr_t j = st_off >> LogBytesPerLong; 3636 3637 if (type == T_INT && st_size == BytesPerInt 3638 && (st_off & BytesPerInt) == BytesPerInt) { 3639 jlong lcon = tiles[j]; 3640 if (!Matcher::isSimpleConstant64(lcon) && 3641 st->Opcode() == Op_StoreI) { 3642 // This StoreI is already optimal by itself. 3643 jint* intcon = (jint*) &tiles[j]; 3644 intcon[1] = 0; // undo the store_constant() 3645 3646 // If the previous store is also optimal by itself, back up and 3647 // undo the action of the previous loop iteration... if we can. 3648 // But if we can't, just let the previous half take care of itself. 3649 st = nodes[j]; 3650 st_off -= BytesPerInt; 3651 con = intcon[0]; 3652 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3653 assert(st_off >= header_size, "still ignoring header"); 3654 assert(get_store_offset(st, phase) == st_off, "must be"); 3655 assert(in(i-1) == zmem, "must be"); 3656 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3657 assert(con == tcon->is_int()->get_con(), "must be"); 3658 // Undo the effects of the previous loop trip, which swallowed st: 3659 intcon[0] = 0; // undo store_constant() 3660 set_req(i-1, st); // undo set_req(i, zmem) 3661 nodes[j] = NULL; // undo nodes[j] = st 3662 --old_subword; // undo ++old_subword 3663 } 3664 continue; // This StoreI is already optimal. 3665 } 3666 } 3667 3668 // This store is not needed. 3669 set_req(i, zmem); 3670 nodes[j] = st; // record for the moment 3671 if (st_size < BytesPerLong) // something has changed 3672 ++old_subword; // includes int/float, but who's counting... 3673 else ++old_long; 3674 } 3675 3676 if ((old_subword + old_long) == 0) 3677 return; // nothing more to do 3678 3679 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3680 // Be sure to insert them before overlapping non-constant stores. 3681 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3682 for (int j = 0; j < num_tiles; j++) { 3683 jlong con = tiles[j]; 3684 jlong init = inits[j]; 3685 if (con == 0) continue; 3686 jint con0, con1; // split the constant, address-wise 3687 jint init0, init1; // split the init map, address-wise 3688 { union { jlong con; jint intcon[2]; } u; 3689 u.con = con; 3690 con0 = u.intcon[0]; 3691 con1 = u.intcon[1]; 3692 u.con = init; 3693 init0 = u.intcon[0]; 3694 init1 = u.intcon[1]; 3695 } 3696 3697 Node* old = nodes[j]; 3698 assert(old != NULL, "need the prior store"); 3699 intptr_t offset = (j * BytesPerLong); 3700 3701 bool split = !Matcher::isSimpleConstant64(con); 3702 3703 if (offset < header_size) { 3704 assert(offset + BytesPerInt >= header_size, "second int counts"); 3705 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3706 split = true; // only the second word counts 3707 // Example: int a[] = { 42 ... } 3708 } else if (con0 == 0 && init0 == -1) { 3709 split = true; // first word is covered by full inits 3710 // Example: int a[] = { ... foo(), 42 ... } 3711 } else if (con1 == 0 && init1 == -1) { 3712 split = true; // second word is covered by full inits 3713 // Example: int a[] = { ... 42, foo() ... } 3714 } 3715 3716 // Here's a case where init0 is neither 0 nor -1: 3717 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3718 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3719 // In this case the tile is not split; it is (jlong)42. 3720 // The big tile is stored down, and then the foo() value is inserted. 3721 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3722 3723 Node* ctl = old->in(MemNode::Control); 3724 Node* adr = make_raw_address(offset, phase); 3725 const TypePtr* atp = TypeRawPtr::BOTTOM; 3726 3727 // One or two coalesced stores to plop down. 3728 Node* st[2]; 3729 intptr_t off[2]; 3730 int nst = 0; 3731 if (!split) { 3732 ++new_long; 3733 off[nst] = offset; 3734 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3735 phase->longcon(con), T_LONG, MemNode::unordered); 3736 } else { 3737 // Omit either if it is a zero. 3738 if (con0 != 0) { 3739 ++new_int; 3740 off[nst] = offset; 3741 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3742 phase->intcon(con0), T_INT, MemNode::unordered); 3743 } 3744 if (con1 != 0) { 3745 ++new_int; 3746 offset += BytesPerInt; 3747 adr = make_raw_address(offset, phase); 3748 off[nst] = offset; 3749 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3750 phase->intcon(con1), T_INT, MemNode::unordered); 3751 } 3752 } 3753 3754 // Insert second store first, then the first before the second. 3755 // Insert each one just before any overlapping non-constant stores. 3756 while (nst > 0) { 3757 Node* st1 = st[--nst]; 3758 C->copy_node_notes_to(st1, old); 3759 st1 = phase->transform(st1); 3760 offset = off[nst]; 3761 assert(offset >= header_size, "do not smash header"); 3762 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3763 guarantee(ins_idx != 0, "must re-insert constant store"); 3764 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3765 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3766 set_req(--ins_idx, st1); 3767 else 3768 ins_req(ins_idx, st1); 3769 } 3770 } 3771 3772 if (PrintCompilation && WizardMode) 3773 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3774 old_subword, old_long, new_int, new_long); 3775 if (C->log() != NULL) 3776 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3777 old_subword, old_long, new_int, new_long); 3778 3779 // Clean up any remaining occurrences of zmem: 3780 remove_extra_zeroes(); 3781 } 3782 3783 // Explore forward from in(start) to find the first fully initialized 3784 // word, and return its offset. Skip groups of subword stores which 3785 // together initialize full words. If in(start) is itself part of a 3786 // fully initialized word, return the offset of in(start). If there 3787 // are no following full-word stores, or if something is fishy, return 3788 // a negative value. 3789 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3790 int int_map = 0; 3791 intptr_t int_map_off = 0; 3792 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3793 3794 for (uint i = start, limit = req(); i < limit; i++) { 3795 Node* st = in(i); 3796 3797 intptr_t st_off = get_store_offset(st, phase); 3798 if (st_off < 0) break; // return conservative answer 3799 3800 int st_size = st->as_Store()->memory_size(); 3801 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3802 return st_off; // we found a complete word init 3803 } 3804 3805 // update the map: 3806 3807 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3808 if (this_int_off != int_map_off) { 3809 // reset the map: 3810 int_map = 0; 3811 int_map_off = this_int_off; 3812 } 3813 3814 int subword_off = st_off - this_int_off; 3815 int_map |= right_n_bits(st_size) << subword_off; 3816 if ((int_map & FULL_MAP) == FULL_MAP) { 3817 return this_int_off; // we found a complete word init 3818 } 3819 3820 // Did this store hit or cross the word boundary? 3821 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3822 if (next_int_off == this_int_off + BytesPerInt) { 3823 // We passed the current int, without fully initializing it. 3824 int_map_off = next_int_off; 3825 int_map >>= BytesPerInt; 3826 } else if (next_int_off > this_int_off + BytesPerInt) { 3827 // We passed the current and next int. 3828 return this_int_off + BytesPerInt; 3829 } 3830 } 3831 3832 return -1; 3833 } 3834 3835 3836 // Called when the associated AllocateNode is expanded into CFG. 3837 // At this point, we may perform additional optimizations. 3838 // Linearize the stores by ascending offset, to make memory 3839 // activity as coherent as possible. 3840 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3841 intptr_t header_size, 3842 Node* size_in_bytes, 3843 PhaseGVN* phase) { 3844 assert(!is_complete(), "not already complete"); 3845 assert(stores_are_sane(phase), ""); 3846 assert(allocation() != NULL, "must be present"); 3847 3848 remove_extra_zeroes(); 3849 3850 if (ReduceFieldZeroing || ReduceBulkZeroing) 3851 // reduce instruction count for common initialization patterns 3852 coalesce_subword_stores(header_size, size_in_bytes, phase); 3853 3854 Node* zmem = zero_memory(); // initially zero memory state 3855 Node* inits = zmem; // accumulating a linearized chain of inits 3856 #ifdef ASSERT 3857 intptr_t first_offset = allocation()->minimum_header_size(); 3858 intptr_t last_init_off = first_offset; // previous init offset 3859 intptr_t last_init_end = first_offset; // previous init offset+size 3860 intptr_t last_tile_end = first_offset; // previous tile offset+size 3861 #endif 3862 intptr_t zeroes_done = header_size; 3863 3864 bool do_zeroing = true; // we might give up if inits are very sparse 3865 int big_init_gaps = 0; // how many large gaps have we seen? 3866 3867 if (ZeroTLAB) do_zeroing = false; 3868 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3869 3870 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3871 Node* st = in(i); 3872 intptr_t st_off = get_store_offset(st, phase); 3873 if (st_off < 0) 3874 break; // unknown junk in the inits 3875 if (st->in(MemNode::Memory) != zmem) 3876 break; // complicated store chains somehow in list 3877 3878 int st_size = st->as_Store()->memory_size(); 3879 intptr_t next_init_off = st_off + st_size; 3880 3881 if (do_zeroing && zeroes_done < next_init_off) { 3882 // See if this store needs a zero before it or under it. 3883 intptr_t zeroes_needed = st_off; 3884 3885 if (st_size < BytesPerInt) { 3886 // Look for subword stores which only partially initialize words. 3887 // If we find some, we must lay down some word-level zeroes first, 3888 // underneath the subword stores. 3889 // 3890 // Examples: 3891 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3892 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3893 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3894 // 3895 // Note: coalesce_subword_stores may have already done this, 3896 // if it was prompted by constant non-zero subword initializers. 3897 // But this case can still arise with non-constant stores. 3898 3899 intptr_t next_full_store = find_next_fullword_store(i, phase); 3900 3901 // In the examples above: 3902 // in(i) p q r s x y z 3903 // st_off 12 13 14 15 12 13 14 3904 // st_size 1 1 1 1 1 1 1 3905 // next_full_s. 12 16 16 16 16 16 16 3906 // z's_done 12 16 16 16 12 16 12 3907 // z's_needed 12 16 16 16 16 16 16 3908 // zsize 0 0 0 0 4 0 4 3909 if (next_full_store < 0) { 3910 // Conservative tack: Zero to end of current word. 3911 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3912 } else { 3913 // Zero to beginning of next fully initialized word. 3914 // Or, don't zero at all, if we are already in that word. 3915 assert(next_full_store >= zeroes_needed, "must go forward"); 3916 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3917 zeroes_needed = next_full_store; 3918 } 3919 } 3920 3921 if (zeroes_needed > zeroes_done) { 3922 intptr_t zsize = zeroes_needed - zeroes_done; 3923 // Do some incremental zeroing on rawmem, in parallel with inits. 3924 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3925 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3926 zeroes_done, zeroes_needed, 3927 phase); 3928 zeroes_done = zeroes_needed; 3929 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3930 do_zeroing = false; // leave the hole, next time 3931 } 3932 } 3933 3934 // Collect the store and move on: 3935 st->set_req(MemNode::Memory, inits); 3936 inits = st; // put it on the linearized chain 3937 set_req(i, zmem); // unhook from previous position 3938 3939 if (zeroes_done == st_off) 3940 zeroes_done = next_init_off; 3941 3942 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3943 3944 #ifdef ASSERT 3945 // Various order invariants. Weaker than stores_are_sane because 3946 // a large constant tile can be filled in by smaller non-constant stores. 3947 assert(st_off >= last_init_off, "inits do not reverse"); 3948 last_init_off = st_off; 3949 const Type* val = NULL; 3950 if (st_size >= BytesPerInt && 3951 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3952 (int)val->basic_type() < (int)T_OBJECT) { 3953 assert(st_off >= last_tile_end, "tiles do not overlap"); 3954 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3955 last_tile_end = MAX2(last_tile_end, next_init_off); 3956 } else { 3957 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3958 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3959 assert(st_off >= last_init_end, "inits do not overlap"); 3960 last_init_end = next_init_off; // it's a non-tile 3961 } 3962 #endif //ASSERT 3963 } 3964 3965 remove_extra_zeroes(); // clear out all the zmems left over 3966 add_req(inits); 3967 3968 if (!ZeroTLAB) { 3969 // If anything remains to be zeroed, zero it all now. 3970 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3971 // if it is the last unused 4 bytes of an instance, forget about it 3972 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3973 if (zeroes_done + BytesPerLong >= size_limit) { 3974 assert(allocation() != NULL, ""); 3975 if (allocation()->Opcode() == Op_Allocate) { 3976 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3977 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3978 if (zeroes_done == k->layout_helper()) 3979 zeroes_done = size_limit; 3980 } 3981 } 3982 if (zeroes_done < size_limit) { 3983 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3984 zeroes_done, size_in_bytes, phase); 3985 } 3986 } 3987 3988 set_complete(phase); 3989 return rawmem; 3990 } 3991 3992 3993 #ifdef ASSERT 3994 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3995 if (is_complete()) 3996 return true; // stores could be anything at this point 3997 assert(allocation() != NULL, "must be present"); 3998 intptr_t last_off = allocation()->minimum_header_size(); 3999 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4000 Node* st = in(i); 4001 intptr_t st_off = get_store_offset(st, phase); 4002 if (st_off < 0) continue; // ignore dead garbage 4003 if (last_off > st_off) { 4004 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4005 this->dump(2); 4006 assert(false, "ascending store offsets"); 4007 return false; 4008 } 4009 last_off = st_off + st->as_Store()->memory_size(); 4010 } 4011 return true; 4012 } 4013 #endif //ASSERT 4014 4015 4016 4017 4018 //============================MergeMemNode===================================== 4019 // 4020 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4021 // contributing store or call operations. Each contributor provides the memory 4022 // state for a particular "alias type" (see Compile::alias_type). For example, 4023 // if a MergeMem has an input X for alias category #6, then any memory reference 4024 // to alias category #6 may use X as its memory state input, as an exact equivalent 4025 // to using the MergeMem as a whole. 4026 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4027 // 4028 // (Here, the <N> notation gives the index of the relevant adr_type.) 4029 // 4030 // In one special case (and more cases in the future), alias categories overlap. 4031 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4032 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4033 // it is exactly equivalent to that state W: 4034 // MergeMem(<Bot>: W) <==> W 4035 // 4036 // Usually, the merge has more than one input. In that case, where inputs 4037 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4038 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4039 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4040 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4041 // 4042 // A merge can take a "wide" memory state as one of its narrow inputs. 4043 // This simply means that the merge observes out only the relevant parts of 4044 // the wide input. That is, wide memory states arriving at narrow merge inputs 4045 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4046 // 4047 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4048 // and that memory slices "leak through": 4049 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4050 // 4051 // But, in such a cascade, repeated memory slices can "block the leak": 4052 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4053 // 4054 // In the last example, Y is not part of the combined memory state of the 4055 // outermost MergeMem. The system must, of course, prevent unschedulable 4056 // memory states from arising, so you can be sure that the state Y is somehow 4057 // a precursor to state Y'. 4058 // 4059 // 4060 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4061 // of each MergeMemNode array are exactly the numerical alias indexes, including 4062 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4063 // Compile::alias_type (and kin) produce and manage these indexes. 4064 // 4065 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4066 // (Note that this provides quick access to the top node inside MergeMem methods, 4067 // without the need to reach out via TLS to Compile::current.) 4068 // 4069 // As a consequence of what was just described, a MergeMem that represents a full 4070 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4071 // containing all alias categories. 4072 // 4073 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4074 // 4075 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4076 // a memory state for the alias type <N>, or else the top node, meaning that 4077 // there is no particular input for that alias type. Note that the length of 4078 // a MergeMem is variable, and may be extended at any time to accommodate new 4079 // memory states at larger alias indexes. When merges grow, they are of course 4080 // filled with "top" in the unused in() positions. 4081 // 4082 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4083 // (Top was chosen because it works smoothly with passes like GCM.) 4084 // 4085 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4086 // the type of random VM bits like TLS references.) Since it is always the 4087 // first non-Bot memory slice, some low-level loops use it to initialize an 4088 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4089 // 4090 // 4091 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4092 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4093 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4094 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4095 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4096 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4097 // 4098 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4099 // really that different from the other memory inputs. An abbreviation called 4100 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4101 // 4102 // 4103 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4104 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4105 // that "emerges though" the base memory will be marked as excluding the alias types 4106 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4107 // 4108 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4109 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4110 // 4111 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4112 // (It is currently unimplemented.) As you can see, the resulting merge is 4113 // actually a disjoint union of memory states, rather than an overlay. 4114 // 4115 4116 //------------------------------MergeMemNode----------------------------------- 4117 Node* MergeMemNode::make_empty_memory() { 4118 Node* empty_memory = (Node*) Compile::current()->top(); 4119 assert(empty_memory->is_top(), "correct sentinel identity"); 4120 return empty_memory; 4121 } 4122 4123 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4124 init_class_id(Class_MergeMem); 4125 // all inputs are nullified in Node::Node(int) 4126 // set_input(0, NULL); // no control input 4127 4128 // Initialize the edges uniformly to top, for starters. 4129 Node* empty_mem = make_empty_memory(); 4130 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4131 init_req(i,empty_mem); 4132 } 4133 assert(empty_memory() == empty_mem, ""); 4134 4135 if( new_base != NULL && new_base->is_MergeMem() ) { 4136 MergeMemNode* mdef = new_base->as_MergeMem(); 4137 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4138 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4139 mms.set_memory(mms.memory2()); 4140 } 4141 assert(base_memory() == mdef->base_memory(), ""); 4142 } else { 4143 set_base_memory(new_base); 4144 } 4145 } 4146 4147 // Make a new, untransformed MergeMem with the same base as 'mem'. 4148 // If mem is itself a MergeMem, populate the result with the same edges. 4149 MergeMemNode* MergeMemNode::make(Node* mem) { 4150 return new MergeMemNode(mem); 4151 } 4152 4153 //------------------------------cmp-------------------------------------------- 4154 uint MergeMemNode::hash() const { return NO_HASH; } 4155 uint MergeMemNode::cmp( const Node &n ) const { 4156 return (&n == this); // Always fail except on self 4157 } 4158 4159 //------------------------------Identity--------------------------------------- 4160 Node* MergeMemNode::Identity(PhaseTransform *phase) { 4161 // Identity if this merge point does not record any interesting memory 4162 // disambiguations. 4163 Node* base_mem = base_memory(); 4164 Node* empty_mem = empty_memory(); 4165 if (base_mem != empty_mem) { // Memory path is not dead? 4166 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4167 Node* mem = in(i); 4168 if (mem != empty_mem && mem != base_mem) { 4169 return this; // Many memory splits; no change 4170 } 4171 } 4172 } 4173 return base_mem; // No memory splits; ID on the one true input 4174 } 4175 4176 //------------------------------Ideal------------------------------------------ 4177 // This method is invoked recursively on chains of MergeMem nodes 4178 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4179 // Remove chain'd MergeMems 4180 // 4181 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4182 // relative to the "in(Bot)". Since we are patching both at the same time, 4183 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4184 // but rewrite each "in(i)" relative to the new "in(Bot)". 4185 Node *progress = NULL; 4186 4187 4188 Node* old_base = base_memory(); 4189 Node* empty_mem = empty_memory(); 4190 if (old_base == empty_mem) 4191 return NULL; // Dead memory path. 4192 4193 MergeMemNode* old_mbase; 4194 if (old_base != NULL && old_base->is_MergeMem()) 4195 old_mbase = old_base->as_MergeMem(); 4196 else 4197 old_mbase = NULL; 4198 Node* new_base = old_base; 4199 4200 // simplify stacked MergeMems in base memory 4201 if (old_mbase) new_base = old_mbase->base_memory(); 4202 4203 // the base memory might contribute new slices beyond my req() 4204 if (old_mbase) grow_to_match(old_mbase); 4205 4206 // Look carefully at the base node if it is a phi. 4207 PhiNode* phi_base; 4208 if (new_base != NULL && new_base->is_Phi()) 4209 phi_base = new_base->as_Phi(); 4210 else 4211 phi_base = NULL; 4212 4213 Node* phi_reg = NULL; 4214 uint phi_len = (uint)-1; 4215 if (phi_base != NULL && !phi_base->is_copy()) { 4216 // do not examine phi if degraded to a copy 4217 phi_reg = phi_base->region(); 4218 phi_len = phi_base->req(); 4219 // see if the phi is unfinished 4220 for (uint i = 1; i < phi_len; i++) { 4221 if (phi_base->in(i) == NULL) { 4222 // incomplete phi; do not look at it yet! 4223 phi_reg = NULL; 4224 phi_len = (uint)-1; 4225 break; 4226 } 4227 } 4228 } 4229 4230 // Note: We do not call verify_sparse on entry, because inputs 4231 // can normalize to the base_memory via subsume_node or similar 4232 // mechanisms. This method repairs that damage. 4233 4234 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4235 4236 // Look at each slice. 4237 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4238 Node* old_in = in(i); 4239 // calculate the old memory value 4240 Node* old_mem = old_in; 4241 if (old_mem == empty_mem) old_mem = old_base; 4242 assert(old_mem == memory_at(i), ""); 4243 4244 // maybe update (reslice) the old memory value 4245 4246 // simplify stacked MergeMems 4247 Node* new_mem = old_mem; 4248 MergeMemNode* old_mmem; 4249 if (old_mem != NULL && old_mem->is_MergeMem()) 4250 old_mmem = old_mem->as_MergeMem(); 4251 else 4252 old_mmem = NULL; 4253 if (old_mmem == this) { 4254 // This can happen if loops break up and safepoints disappear. 4255 // A merge of BotPtr (default) with a RawPtr memory derived from a 4256 // safepoint can be rewritten to a merge of the same BotPtr with 4257 // the BotPtr phi coming into the loop. If that phi disappears 4258 // also, we can end up with a self-loop of the mergemem. 4259 // In general, if loops degenerate and memory effects disappear, 4260 // a mergemem can be left looking at itself. This simply means 4261 // that the mergemem's default should be used, since there is 4262 // no longer any apparent effect on this slice. 4263 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4264 // from start. Update the input to TOP. 4265 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4266 } 4267 else if (old_mmem != NULL) { 4268 new_mem = old_mmem->memory_at(i); 4269 } 4270 // else preceding memory was not a MergeMem 4271 4272 // replace equivalent phis (unfortunately, they do not GVN together) 4273 if (new_mem != NULL && new_mem != new_base && 4274 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4275 if (new_mem->is_Phi()) { 4276 PhiNode* phi_mem = new_mem->as_Phi(); 4277 for (uint i = 1; i < phi_len; i++) { 4278 if (phi_base->in(i) != phi_mem->in(i)) { 4279 phi_mem = NULL; 4280 break; 4281 } 4282 } 4283 if (phi_mem != NULL) { 4284 // equivalent phi nodes; revert to the def 4285 new_mem = new_base; 4286 } 4287 } 4288 } 4289 4290 // maybe store down a new value 4291 Node* new_in = new_mem; 4292 if (new_in == new_base) new_in = empty_mem; 4293 4294 if (new_in != old_in) { 4295 // Warning: Do not combine this "if" with the previous "if" 4296 // A memory slice might have be be rewritten even if it is semantically 4297 // unchanged, if the base_memory value has changed. 4298 set_req(i, new_in); 4299 progress = this; // Report progress 4300 } 4301 } 4302 4303 if (new_base != old_base) { 4304 set_req(Compile::AliasIdxBot, new_base); 4305 // Don't use set_base_memory(new_base), because we need to update du. 4306 assert(base_memory() == new_base, ""); 4307 progress = this; 4308 } 4309 4310 if( base_memory() == this ) { 4311 // a self cycle indicates this memory path is dead 4312 set_req(Compile::AliasIdxBot, empty_mem); 4313 } 4314 4315 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4316 // Recursion must occur after the self cycle check above 4317 if( base_memory()->is_MergeMem() ) { 4318 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4319 Node *m = phase->transform(new_mbase); // Rollup any cycles 4320 if( m != NULL && (m->is_top() || 4321 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4322 // propagate rollup of dead cycle to self 4323 set_req(Compile::AliasIdxBot, empty_mem); 4324 } 4325 } 4326 4327 if( base_memory() == empty_mem ) { 4328 progress = this; 4329 // Cut inputs during Parse phase only. 4330 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4331 if( !can_reshape ) { 4332 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4333 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4334 } 4335 } 4336 } 4337 4338 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4339 // Check if PhiNode::Ideal's "Split phis through memory merges" 4340 // transform should be attempted. Look for this->phi->this cycle. 4341 uint merge_width = req(); 4342 if (merge_width > Compile::AliasIdxRaw) { 4343 PhiNode* phi = base_memory()->as_Phi(); 4344 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4345 if (phi->in(i) == this) { 4346 phase->is_IterGVN()->_worklist.push(phi); 4347 break; 4348 } 4349 } 4350 } 4351 } 4352 4353 assert(progress || verify_sparse(), "please, no dups of base"); 4354 return progress; 4355 } 4356 4357 //-------------------------set_base_memory------------------------------------- 4358 void MergeMemNode::set_base_memory(Node *new_base) { 4359 Node* empty_mem = empty_memory(); 4360 set_req(Compile::AliasIdxBot, new_base); 4361 assert(memory_at(req()) == new_base, "must set default memory"); 4362 // Clear out other occurrences of new_base: 4363 if (new_base != empty_mem) { 4364 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4365 if (in(i) == new_base) set_req(i, empty_mem); 4366 } 4367 } 4368 } 4369 4370 //------------------------------out_RegMask------------------------------------ 4371 const RegMask &MergeMemNode::out_RegMask() const { 4372 return RegMask::Empty; 4373 } 4374 4375 //------------------------------dump_spec-------------------------------------- 4376 #ifndef PRODUCT 4377 void MergeMemNode::dump_spec(outputStream *st) const { 4378 st->print(" {"); 4379 Node* base_mem = base_memory(); 4380 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4381 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem; 4382 if (mem == base_mem) { st->print(" -"); continue; } 4383 st->print( " N%d:", mem->_idx ); 4384 Compile::current()->get_adr_type(i)->dump_on(st); 4385 } 4386 st->print(" }"); 4387 } 4388 #endif // !PRODUCT 4389 4390 4391 #ifdef ASSERT 4392 static bool might_be_same(Node* a, Node* b) { 4393 if (a == b) return true; 4394 if (!(a->is_Phi() || b->is_Phi())) return false; 4395 // phis shift around during optimization 4396 return true; // pretty stupid... 4397 } 4398 4399 // verify a narrow slice (either incoming or outgoing) 4400 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4401 if (!VerifyAliases) return; // don't bother to verify unless requested 4402 if (is_error_reported()) return; // muzzle asserts when debugging an error 4403 if (Node::in_dump()) return; // muzzle asserts when printing 4404 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4405 assert(n != NULL, ""); 4406 // Elide intervening MergeMem's 4407 while (n->is_MergeMem()) { 4408 n = n->as_MergeMem()->memory_at(alias_idx); 4409 } 4410 Compile* C = Compile::current(); 4411 const TypePtr* n_adr_type = n->adr_type(); 4412 if (n == m->empty_memory()) { 4413 // Implicit copy of base_memory() 4414 } else if (n_adr_type != TypePtr::BOTTOM) { 4415 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4416 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4417 } else { 4418 // A few places like make_runtime_call "know" that VM calls are narrow, 4419 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4420 bool expected_wide_mem = false; 4421 if (n == m->base_memory()) { 4422 expected_wide_mem = true; 4423 } else if (alias_idx == Compile::AliasIdxRaw || 4424 n == m->memory_at(Compile::AliasIdxRaw)) { 4425 expected_wide_mem = true; 4426 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4427 // memory can "leak through" calls on channels that 4428 // are write-once. Allow this also. 4429 expected_wide_mem = true; 4430 } 4431 assert(expected_wide_mem, "expected narrow slice replacement"); 4432 } 4433 } 4434 #else // !ASSERT 4435 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4436 #endif 4437 4438 4439 //-----------------------------memory_at--------------------------------------- 4440 Node* MergeMemNode::memory_at(uint alias_idx) const { 4441 assert(alias_idx >= Compile::AliasIdxRaw || 4442 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4443 "must avoid base_memory and AliasIdxTop"); 4444 4445 // Otherwise, it is a narrow slice. 4446 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4447 Compile *C = Compile::current(); 4448 if (is_empty_memory(n)) { 4449 // the array is sparse; empty slots are the "top" node 4450 n = base_memory(); 4451 assert(Node::in_dump() 4452 || n == NULL || n->bottom_type() == Type::TOP 4453 || n->adr_type() == NULL // address is TOP 4454 || n->adr_type() == TypePtr::BOTTOM 4455 || n->adr_type() == TypeRawPtr::BOTTOM 4456 || Compile::current()->AliasLevel() == 0, 4457 "must be a wide memory"); 4458 // AliasLevel == 0 if we are organizing the memory states manually. 4459 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4460 } else { 4461 // make sure the stored slice is sane 4462 #ifdef ASSERT 4463 if (is_error_reported() || Node::in_dump()) { 4464 } else if (might_be_same(n, base_memory())) { 4465 // Give it a pass: It is a mostly harmless repetition of the base. 4466 // This can arise normally from node subsumption during optimization. 4467 } else { 4468 verify_memory_slice(this, alias_idx, n); 4469 } 4470 #endif 4471 } 4472 return n; 4473 } 4474 4475 //---------------------------set_memory_at------------------------------------- 4476 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4477 verify_memory_slice(this, alias_idx, n); 4478 Node* empty_mem = empty_memory(); 4479 if (n == base_memory()) n = empty_mem; // collapse default 4480 uint need_req = alias_idx+1; 4481 if (req() < need_req) { 4482 if (n == empty_mem) return; // already the default, so do not grow me 4483 // grow the sparse array 4484 do { 4485 add_req(empty_mem); 4486 } while (req() < need_req); 4487 } 4488 set_req( alias_idx, n ); 4489 } 4490 4491 4492 4493 //--------------------------iteration_setup------------------------------------ 4494 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4495 if (other != NULL) { 4496 grow_to_match(other); 4497 // invariant: the finite support of mm2 is within mm->req() 4498 #ifdef ASSERT 4499 for (uint i = req(); i < other->req(); i++) { 4500 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4501 } 4502 #endif 4503 } 4504 // Replace spurious copies of base_memory by top. 4505 Node* base_mem = base_memory(); 4506 if (base_mem != NULL && !base_mem->is_top()) { 4507 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4508 if (in(i) == base_mem) 4509 set_req(i, empty_memory()); 4510 } 4511 } 4512 } 4513 4514 //---------------------------grow_to_match------------------------------------- 4515 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4516 Node* empty_mem = empty_memory(); 4517 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4518 // look for the finite support of the other memory 4519 for (uint i = other->req(); --i >= req(); ) { 4520 if (other->in(i) != empty_mem) { 4521 uint new_len = i+1; 4522 while (req() < new_len) add_req(empty_mem); 4523 break; 4524 } 4525 } 4526 } 4527 4528 //---------------------------verify_sparse------------------------------------- 4529 #ifndef PRODUCT 4530 bool MergeMemNode::verify_sparse() const { 4531 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4532 Node* base_mem = base_memory(); 4533 // The following can happen in degenerate cases, since empty==top. 4534 if (is_empty_memory(base_mem)) return true; 4535 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4536 assert(in(i) != NULL, "sane slice"); 4537 if (in(i) == base_mem) return false; // should have been the sentinel value! 4538 } 4539 return true; 4540 } 4541 4542 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4543 Node* n; 4544 n = mm->in(idx); 4545 if (mem == n) return true; // might be empty_memory() 4546 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4547 if (mem == n) return true; 4548 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4549 if (mem == n) return true; 4550 if (n == NULL) break; 4551 } 4552 return false; 4553 } 4554 #endif // !PRODUCT