1 /*
   2  * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
   3  *
   4  * This code is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 only, as
   6  * published by the Free Software Foundation.
   7  *
   8  * This code is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  11  * version 2 for more details (a copy is included in the LICENSE file that
  12  * accompanied this code).
  13  *
  14  * You should have received a copy of the GNU General Public License version
  15  * 2 along with this work; if not, write to the Free Software Foundation,
  16  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  17  *
  18  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  19  * or visit www.oracle.com if you need additional information or have any
  20  * questions.
  21  *
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "gc/shared/barrierSet.hpp"
  26 #include "gc/shenandoah/shenandoahForwarding.hpp"
  27 #include "gc/shenandoah/shenandoahHeap.hpp"
  28 #include "gc/shenandoah/shenandoahHeuristics.hpp"
  29 #include "gc/shenandoah/shenandoahRuntime.hpp"
  30 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  31 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  32 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
  33 #include "opto/arraycopynode.hpp"
  34 #include "opto/escape.hpp"
  35 #include "opto/graphKit.hpp"
  36 #include "opto/idealKit.hpp"
  37 #include "opto/macro.hpp"
  38 #include "opto/movenode.hpp"
  39 #include "opto/narrowptrnode.hpp"
  40 #include "opto/rootnode.hpp"
  41 
  42 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
  43   return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
  44 }
  45 
  46 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
  47   : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8,  0, NULL)),
  48     _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8,  0, NULL)) {
  49 }
  50 
  51 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const {
  52   return _enqueue_barriers->length();
  53 }
  54 
  55 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const {
  56   return _enqueue_barriers->at(idx);
  57 }
  58 
  59 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
  60   assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list");
  61   _enqueue_barriers->append(n);
  62 }
  63 
  64 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
  65   if (_enqueue_barriers->contains(n)) {
  66     _enqueue_barriers->remove(n);
  67   }
  68 }
  69 
  70 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
  71   return _load_reference_barriers->length();
  72 }
  73 
  74 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
  75   return _load_reference_barriers->at(idx);
  76 }
  77 
  78 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
  79   assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
  80   _load_reference_barriers->append(n);
  81 }
  82 
  83 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
  84   if (_load_reference_barriers->contains(n)) {
  85     _load_reference_barriers->remove(n);
  86   }
  87 }
  88 
  89 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const {
  90   if (ShenandoahStoreValEnqueueBarrier) {
  91     obj = shenandoah_enqueue_barrier(kit, obj);
  92   }
  93   return obj;
  94 }
  95 
  96 #define __ kit->
  97 
  98 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr,
  99                                                          BasicType bt, uint adr_idx) const {
 100   intptr_t offset = 0;
 101   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 102   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
 103 
 104   if (offset == Type::OffsetBot) {
 105     return false; // cannot unalias unless there are precise offsets
 106   }
 107 
 108   if (alloc == NULL) {
 109     return false; // No allocation found
 110   }
 111 
 112   intptr_t size_in_bytes = type2aelembytes(bt);
 113 
 114   Node* mem = __ memory(adr_idx); // start searching here...
 115 
 116   for (int cnt = 0; cnt < 50; cnt++) {
 117 
 118     if (mem->is_Store()) {
 119 
 120       Node* st_adr = mem->in(MemNode::Address);
 121       intptr_t st_offset = 0;
 122       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 123 
 124       if (st_base == NULL) {
 125         break; // inscrutable pointer
 126       }
 127 
 128       // Break we have found a store with same base and offset as ours so break
 129       if (st_base == base && st_offset == offset) {
 130         break;
 131       }
 132 
 133       if (st_offset != offset && st_offset != Type::OffsetBot) {
 134         const int MAX_STORE = BytesPerLong;
 135         if (st_offset >= offset + size_in_bytes ||
 136             st_offset <= offset - MAX_STORE ||
 137             st_offset <= offset - mem->as_Store()->memory_size()) {
 138           // Success:  The offsets are provably independent.
 139           // (You may ask, why not just test st_offset != offset and be done?
 140           // The answer is that stores of different sizes can co-exist
 141           // in the same sequence of RawMem effects.  We sometimes initialize
 142           // a whole 'tile' of array elements with a single jint or jlong.)
 143           mem = mem->in(MemNode::Memory);
 144           continue; // advance through independent store memory
 145         }
 146       }
 147 
 148       if (st_base != base
 149           && MemNode::detect_ptr_independence(base, alloc, st_base,
 150                                               AllocateNode::Ideal_allocation(st_base, phase),
 151                                               phase)) {
 152         // Success:  The bases are provably independent.
 153         mem = mem->in(MemNode::Memory);
 154         continue; // advance through independent store memory
 155       }
 156     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 157 
 158       InitializeNode* st_init = mem->in(0)->as_Initialize();
 159       AllocateNode* st_alloc = st_init->allocation();
 160 
 161       // Make sure that we are looking at the same allocation site.
 162       // The alloc variable is guaranteed to not be null here from earlier check.
 163       if (alloc == st_alloc) {
 164         // Check that the initialization is storing NULL so that no previous store
 165         // has been moved up and directly write a reference
 166         Node* captured_store = st_init->find_captured_store(offset,
 167                                                             type2aelembytes(T_OBJECT),
 168                                                             phase);
 169         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
 170           return true;
 171         }
 172       }
 173     }
 174 
 175     // Unless there is an explicit 'continue', we must bail out here,
 176     // because 'mem' is an inscrutable memory state (e.g., a call).
 177     break;
 178   }
 179 
 180   return false;
 181 }
 182 
 183 #undef __
 184 #define __ ideal.
 185 
 186 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
 187                                                     bool do_load,
 188                                                     Node* obj,
 189                                                     Node* adr,
 190                                                     uint alias_idx,
 191                                                     Node* val,
 192                                                     const TypeOopPtr* val_type,
 193                                                     Node* pre_val,
 194                                                     BasicType bt) const {
 195   // Some sanity checks
 196   // Note: val is unused in this routine.
 197 
 198   if (do_load) {
 199     // We need to generate the load of the previous value
 200     assert(obj != NULL, "must have a base");
 201     assert(adr != NULL, "where are loading from?");
 202     assert(pre_val == NULL, "loaded already?");
 203     assert(val_type != NULL, "need a type");
 204 
 205     if (ReduceInitialCardMarks
 206         && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
 207       return;
 208     }
 209 
 210   } else {
 211     // In this case both val_type and alias_idx are unused.
 212     assert(pre_val != NULL, "must be loaded already");
 213     // Nothing to be done if pre_val is null.
 214     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
 215     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
 216   }
 217   assert(bt == T_OBJECT, "or we shouldn't be here");
 218 
 219   IdealKit ideal(kit, true);
 220 
 221   Node* tls = __ thread(); // ThreadLocalStorage
 222 
 223   Node* no_base = __ top();
 224   Node* zero  = __ ConI(0);
 225   Node* zeroX = __ ConX(0);
 226 
 227   float likely  = PROB_LIKELY(0.999);
 228   float unlikely  = PROB_UNLIKELY(0.999);
 229 
 230   // Offsets into the thread
 231   const int index_offset   = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
 232   const int buffer_offset  = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
 233 
 234   // Now the actual pointers into the thread
 235   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
 236   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
 237 
 238   // Now some of the values
 239   Node* marking;
 240   Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
 241   Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
 242   marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
 243   assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
 244 
 245   // if (!marking)
 246   __ if_then(marking, BoolTest::ne, zero, unlikely); {
 247     BasicType index_bt = TypeX_X->basic_type();
 248     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
 249     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
 250 
 251     if (do_load) {
 252       // load original value
 253       // alias_idx correct??
 254       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
 255     }
 256 
 257     // if (pre_val != NULL)
 258     __ if_then(pre_val, BoolTest::ne, kit->null()); {
 259       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
 260 
 261       // is the queue for this thread full?
 262       __ if_then(index, BoolTest::ne, zeroX, likely); {
 263 
 264         // decrement the index
 265         Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
 266 
 267         // Now get the buffer location we will log the previous value into and store it
 268         Node *log_addr = __ AddP(no_base, buffer, next_index);
 269         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
 270         // update the index
 271         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
 272 
 273       } __ else_(); {
 274 
 275         // logging buffer is full, call the runtime
 276         const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
 277         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
 278       } __ end_if();  // (!index)
 279     } __ end_if();  // (pre_val != NULL)
 280   } __ end_if();  // (!marking)
 281 
 282   // Final sync IdealKit and GraphKit.
 283   kit->final_sync(ideal);
 284 
 285   if (ShenandoahSATBBarrier && adr != NULL) {
 286     Node* c = kit->control();
 287     Node* call = c->in(1)->in(1)->in(1)->in(0);
 288     assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
 289     call->add_req(adr);
 290   }
 291 }
 292 
 293 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
 294   return call->is_CallLeaf() &&
 295          call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
 296 }
 297 
 298 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
 299   return call->is_CallLeaf() &&
 300           call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier);
 301 }
 302 
 303 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
 304   if (n->Opcode() != Op_If) {
 305     return false;
 306   }
 307 
 308   Node* bol = n->in(1);
 309   assert(bol->is_Bool(), "");
 310   Node* cmpx = bol->in(1);
 311   if (bol->as_Bool()->_test._test == BoolTest::ne &&
 312       cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
 313       is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
 314       cmpx->in(1)->in(2)->is_Con() &&
 315       cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
 316     return true;
 317   }
 318 
 319   return false;
 320 }
 321 
 322 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
 323   if (!n->is_Load()) return false;
 324   const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
 325   return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
 326          && n->in(2)->in(3)->is_Con()
 327          && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
 328 }
 329 
 330 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
 331                                                           bool do_load,
 332                                                           Node* obj,
 333                                                           Node* adr,
 334                                                           uint alias_idx,
 335                                                           Node* val,
 336                                                           const TypeOopPtr* val_type,
 337                                                           Node* pre_val,
 338                                                           BasicType bt) const {
 339   if (ShenandoahSATBBarrier) {
 340     IdealKit ideal(kit);
 341     kit->sync_kit(ideal);
 342 
 343     satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
 344 
 345     ideal.sync_kit(kit);
 346     kit->final_sync(ideal);
 347   }
 348 }
 349 
 350 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const {
 351   return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val));
 352 }
 353 
 354 // Helper that guards and inserts a pre-barrier.
 355 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
 356                                                 Node* pre_val, bool need_mem_bar) const {
 357   // We could be accessing the referent field of a reference object. If so, when G1
 358   // is enabled, we need to log the value in the referent field in an SATB buffer.
 359   // This routine performs some compile time filters and generates suitable
 360   // runtime filters that guard the pre-barrier code.
 361   // Also add memory barrier for non volatile load from the referent field
 362   // to prevent commoning of loads across safepoint.
 363 
 364   // Some compile time checks.
 365 
 366   // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
 367   const TypeX* otype = offset->find_intptr_t_type();
 368   if (otype != NULL && otype->is_con() &&
 369       otype->get_con() != java_lang_ref_Reference::referent_offset) {
 370     // Constant offset but not the reference_offset so just return
 371     return;
 372   }
 373 
 374   // We only need to generate the runtime guards for instances.
 375   const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
 376   if (btype != NULL) {
 377     if (btype->isa_aryptr()) {
 378       // Array type so nothing to do
 379       return;
 380     }
 381 
 382     const TypeInstPtr* itype = btype->isa_instptr();
 383     if (itype != NULL) {
 384       // Can the klass of base_oop be statically determined to be
 385       // _not_ a sub-class of Reference and _not_ Object?
 386       ciKlass* klass = itype->klass();
 387       if ( klass->is_loaded() &&
 388           !klass->is_subtype_of(kit->env()->Reference_klass()) &&
 389           !kit->env()->Object_klass()->is_subtype_of(klass)) {
 390         return;
 391       }
 392     }
 393   }
 394 
 395   // The compile time filters did not reject base_oop/offset so
 396   // we need to generate the following runtime filters
 397   //
 398   // if (offset == java_lang_ref_Reference::_reference_offset) {
 399   //   if (instance_of(base, java.lang.ref.Reference)) {
 400   //     pre_barrier(_, pre_val, ...);
 401   //   }
 402   // }
 403 
 404   float likely   = PROB_LIKELY(  0.999);
 405   float unlikely = PROB_UNLIKELY(0.999);
 406 
 407   IdealKit ideal(kit);
 408 
 409   Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
 410 
 411   __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
 412       // Update graphKit memory and control from IdealKit.
 413       kit->sync_kit(ideal);
 414 
 415       Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
 416       Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
 417 
 418       // Update IdealKit memory and control from graphKit.
 419       __ sync_kit(kit);
 420 
 421       Node* one = __ ConI(1);
 422       // is_instof == 0 if base_oop == NULL
 423       __ if_then(is_instof, BoolTest::eq, one, unlikely); {
 424 
 425         // Update graphKit from IdeakKit.
 426         kit->sync_kit(ideal);
 427 
 428         // Use the pre-barrier to record the value in the referent field
 429         satb_write_barrier_pre(kit, false /* do_load */,
 430                                NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
 431                                pre_val /* pre_val */,
 432                                T_OBJECT);
 433         if (need_mem_bar) {
 434           // Add memory barrier to prevent commoning reads from this field
 435           // across safepoint since GC can change its value.
 436           kit->insert_mem_bar(Op_MemBarCPUOrder);
 437         }
 438         // Update IdealKit from graphKit.
 439         __ sync_kit(kit);
 440 
 441       } __ end_if(); // _ref_type != ref_none
 442   } __ end_if(); // offset == referent_offset
 443 
 444   // Final sync IdealKit and GraphKit.
 445   kit->final_sync(ideal);
 446 }
 447 
 448 #undef __
 449 
 450 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
 451   const Type **fields = TypeTuple::fields(2);
 452   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
 453   fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
 454   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
 455 
 456   // create result type (range)
 457   fields = TypeTuple::fields(0);
 458   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
 459 
 460   return TypeFunc::make(domain, range);
 461 }
 462 
 463 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
 464   const Type **fields = TypeTuple::fields(3);
 465   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // src
 466   fields[TypeFunc::Parms+1] = TypeInstPtr::NOTNULL; // dst
 467   fields[TypeFunc::Parms+2] = TypeInt::INT; // length
 468   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+3, fields);
 469 
 470   // create result type (range)
 471   fields = TypeTuple::fields(0);
 472   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
 473 
 474   return TypeFunc::make(domain, range);
 475 }
 476 
 477 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
 478   const Type **fields = TypeTuple::fields(2);
 479   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
 480   fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM;   // original load address
 481 
 482   const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
 483 
 484   // create result type (range)
 485   fields = TypeTuple::fields(1);
 486   fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL;
 487   const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
 488 
 489   return TypeFunc::make(domain, range);
 490 }
 491 
 492 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
 493   DecoratorSet decorators = access.decorators();
 494 
 495   const TypePtr* adr_type = access.addr().type();
 496   Node* adr = access.addr().node();
 497 
 498   bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
 499   bool on_heap = (decorators & IN_HEAP) != 0;
 500 
 501   if (!access.is_oop() || (!on_heap && !anonymous)) {
 502     return BarrierSetC2::store_at_resolved(access, val);
 503   }
 504 
 505   if (access.is_parse_access()) {
 506     C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
 507     GraphKit* kit = parse_access.kit();
 508 
 509     uint adr_idx = kit->C->get_alias_index(adr_type);
 510     assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
 511     Node* value = val.node();
 512     value = shenandoah_storeval_barrier(kit, value);
 513     val.set_node(value);
 514     shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
 515                                  static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
 516   } else {
 517     assert(access.is_opt_access(), "only for optimization passes");
 518     assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
 519     C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
 520     PhaseGVN& gvn =  opt_access.gvn();
 521     MergeMemNode* mm = opt_access.mem();
 522 
 523     if (ShenandoahStoreValEnqueueBarrier) {
 524       Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(val.node()));
 525       val.set_node(enqueue);
 526     }
 527   }
 528   return BarrierSetC2::store_at_resolved(access, val);
 529 }
 530 
 531 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
 532   DecoratorSet decorators = access.decorators();
 533 
 534   Node* adr = access.addr().node();
 535   Node* obj = access.base();
 536 
 537   bool mismatched = (decorators & C2_MISMATCHED) != 0;
 538   bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
 539   bool on_heap = (decorators & IN_HEAP) != 0;
 540   bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
 541   bool is_unordered = (decorators & MO_UNORDERED) != 0;
 542   bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap;
 543 
 544   Node* top = Compile::current()->top();
 545 
 546   Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
 547   Node* load = BarrierSetC2::load_at_resolved(access, val_type);
 548 
 549   if (access.is_oop()) {
 550     if (ShenandoahLoadRefBarrier) {
 551       load = new ShenandoahLoadReferenceBarrierNode(NULL, load, (decorators & IN_NATIVE) != 0);
 552       if (access.is_parse_access()) {
 553         load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
 554       } else {
 555         load = static_cast<C2OptAccess &>(access).gvn().transform(load);
 556       }
 557     }
 558   }
 559 
 560   // If we are reading the value of the referent field of a Reference
 561   // object (either by using Unsafe directly or through reflection)
 562   // then, if SATB is enabled, we need to record the referent in an
 563   // SATB log buffer using the pre-barrier mechanism.
 564   // Also we need to add memory barrier to prevent commoning reads
 565   // from this field across safepoint since GC can change its value.
 566   bool need_read_barrier = ShenandoahKeepAliveBarrier &&
 567     (on_heap && (on_weak || (unknown && offset != top && obj != top)));
 568 
 569   if (!access.is_oop() || !need_read_barrier) {
 570     return load;
 571   }
 572 
 573   assert(access.is_parse_access(), "entry not supported at optimization time");
 574   C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
 575   GraphKit* kit = parse_access.kit();
 576 
 577   if (on_weak) {
 578     // Use the pre-barrier to record the value in the referent field
 579     satb_write_barrier_pre(kit, false /* do_load */,
 580                            NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
 581                            load /* pre_val */, T_OBJECT);
 582     // Add memory barrier to prevent commoning reads from this field
 583     // across safepoint since GC can change its value.
 584     kit->insert_mem_bar(Op_MemBarCPUOrder);
 585   } else if (unknown) {
 586     // We do not require a mem bar inside pre_barrier if need_mem_bar
 587     // is set: the barriers would be emitted by us.
 588     insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
 589   }
 590 
 591   return load;
 592 }
 593 
 594 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
 595                                                    Node* new_val, const Type* value_type) const {
 596   GraphKit* kit = access.kit();
 597   if (access.is_oop()) {
 598     new_val = shenandoah_storeval_barrier(kit, new_val);
 599     shenandoah_write_barrier_pre(kit, false /* do_load */,
 600                                  NULL, NULL, max_juint, NULL, NULL,
 601                                  expected_val /* pre_val */, T_OBJECT);
 602 
 603     MemNode::MemOrd mo = access.mem_node_mo();
 604     Node* mem = access.memory();
 605     Node* adr = access.addr().node();
 606     const TypePtr* adr_type = access.addr().type();
 607     Node* load_store = NULL;
 608 
 609 #ifdef _LP64
 610     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 611       Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
 612       Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
 613       if (ShenandoahCASBarrier) {
 614         load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
 615       } else {
 616         load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
 617       }
 618     } else
 619 #endif
 620     {
 621       if (ShenandoahCASBarrier) {
 622         load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
 623       } else {
 624         load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
 625       }
 626     }
 627 
 628     access.set_raw_access(load_store);
 629     pin_atomic_op(access);
 630 
 631 #ifdef _LP64
 632     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 633       load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
 634     }
 635 #endif
 636     load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store, false));
 637     return load_store;
 638   }
 639   return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
 640 }
 641 
 642 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
 643                                                               Node* new_val, const Type* value_type) const {
 644   GraphKit* kit = access.kit();
 645   if (access.is_oop()) {
 646     new_val = shenandoah_storeval_barrier(kit, new_val);
 647     shenandoah_write_barrier_pre(kit, false /* do_load */,
 648                                  NULL, NULL, max_juint, NULL, NULL,
 649                                  expected_val /* pre_val */, T_OBJECT);
 650     DecoratorSet decorators = access.decorators();
 651     MemNode::MemOrd mo = access.mem_node_mo();
 652     Node* mem = access.memory();
 653     bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
 654     Node* load_store = NULL;
 655     Node* adr = access.addr().node();
 656 #ifdef _LP64
 657     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 658       Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
 659       Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
 660       if (ShenandoahCASBarrier) {
 661         if (is_weak_cas) {
 662           load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 663         } else {
 664           load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 665         }
 666       } else {
 667         if (is_weak_cas) {
 668           load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 669         } else {
 670           load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
 671         }
 672       }
 673     } else
 674 #endif
 675     {
 676       if (ShenandoahCASBarrier) {
 677         if (is_weak_cas) {
 678           load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 679         } else {
 680           load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 681         }
 682       } else {
 683         if (is_weak_cas) {
 684           load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 685         } else {
 686           load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
 687         }
 688       }
 689     }
 690     access.set_raw_access(load_store);
 691     pin_atomic_op(access);
 692     return load_store;
 693   }
 694   return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
 695 }
 696 
 697 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
 698   GraphKit* kit = access.kit();
 699   if (access.is_oop()) {
 700     val = shenandoah_storeval_barrier(kit, val);
 701   }
 702   Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
 703   if (access.is_oop()) {
 704     result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result, false));
 705     shenandoah_write_barrier_pre(kit, false /* do_load */,
 706                                  NULL, NULL, max_juint, NULL, NULL,
 707                                  result /* pre_val */, T_OBJECT);
 708   }
 709   return result;
 710 }
 711 
 712 // Support for GC barriers emitted during parsing
 713 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
 714   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
 715   if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
 716     return false;
 717   }
 718   CallLeafNode *call = node->as_CallLeaf();
 719   if (call->_name == NULL) {
 720     return false;
 721   }
 722 
 723   return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
 724          strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
 725          strcmp(call->_name, "shenandoah_wb_pre") == 0;
 726 }
 727 
 728 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
 729   if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 730     return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
 731   }
 732   if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
 733     c = c->in(1);
 734   }
 735   return c;
 736 }
 737 
 738 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
 739   return !ShenandoahBarrierC2Support::expand(C, igvn);
 740 }
 741 
 742 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
 743   if (mode == LoopOptsShenandoahExpand) {
 744     assert(UseShenandoahGC, "only for shenandoah");
 745     ShenandoahBarrierC2Support::pin_and_expand(phase);
 746     return true;
 747   } else if (mode == LoopOptsShenandoahPostExpand) {
 748     assert(UseShenandoahGC, "only for shenandoah");
 749     visited.Clear();
 750     ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
 751     return true;
 752   }
 753   return false;
 754 }
 755 
 756 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const {
 757   bool is_oop = is_reference_type(type);
 758   if (!is_oop) {
 759     return false;
 760   }
 761   if (tightly_coupled_alloc) {
 762     if (phase == Optimization) {
 763       return false;
 764     }
 765     return !is_clone;
 766   }
 767   if (phase == Optimization) {
 768     return !ShenandoahStoreValEnqueueBarrier;
 769   }
 770   return true;
 771 }
 772 
 773 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
 774   const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
 775   if (src_type->isa_instptr() != NULL) {
 776     ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
 777     if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
 778       if (ik->has_object_fields()) {
 779         return true;
 780       } else {
 781         if (!src_type->klass_is_exact()) {
 782           Compile::current()->dependencies()->assert_leaf_type(ik);
 783         }
 784       }
 785     } else {
 786       return true;
 787         }
 788   } else if (src_type->isa_aryptr()) {
 789     BasicType src_elem  = src_type->klass()->as_array_klass()->element_type()->basic_type();
 790     if (is_reference_type(src_elem)) {
 791       return true;
 792     }
 793   } else {
 794     return true;
 795   }
 796   return false;
 797 }
 798 
 799 #define XTOP LP64_ONLY(COMMA phase->top())
 800 
 801 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
 802   Node* ctrl = ac->in(TypeFunc::Control);
 803   Node* mem = ac->in(TypeFunc::Memory);
 804   Node* src = ac->in(ArrayCopyNode::Src);
 805   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
 806   Node* dest = ac->in(ArrayCopyNode::Dest);
 807   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
 808   Node* length = ac->in(ArrayCopyNode::Length);
 809   assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
 810   if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
 811     Node* call = phase->make_leaf_call(ctrl, mem,
 812                     ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
 813                     CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
 814                     "shenandoah_clone",
 815                     TypeRawPtr::BOTTOM,
 816                     src, dest, length);
 817     call = phase->transform_later(call);
 818     phase->igvn().replace_node(ac, call);
 819   } else {
 820     BarrierSetC2::clone_at_expansion(phase, ac);
 821   }
 822 }
 823 
 824 
 825 // Support for macro expanded GC barriers
 826 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
 827   if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
 828     state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
 829   }
 830   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 831     state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
 832   }
 833 }
 834 
 835 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
 836   if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
 837     state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
 838   }
 839   if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 840     state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
 841   }
 842 }
 843 
 844 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
 845   if (is_shenandoah_wb_pre_call(n)) {
 846     shenandoah_eliminate_wb_pre(n, &macro->igvn());
 847   }
 848 }
 849 
 850 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
 851   assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
 852   Node* c = call->as_Call()->proj_out(TypeFunc::Control);
 853   c = c->unique_ctrl_out();
 854   assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 855   c = c->unique_ctrl_out();
 856   assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 857   Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
 858   assert(iff->is_If(), "expect test");
 859   if (!is_shenandoah_marking_if(igvn, iff)) {
 860     c = c->unique_ctrl_out();
 861     assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
 862     iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
 863     assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
 864   }
 865   Node* cmpx = iff->in(1)->in(1);
 866   igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
 867   igvn->rehash_node_delayed(call);
 868   call->del_req(call->req()-1);
 869 }
 870 
 871 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
 872   if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
 873     igvn->add_users_to_worklist(node);
 874   }
 875 }
 876 
 877 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
 878   for (uint i = 0; i < useful.size(); i++) {
 879     Node* n = useful.at(i);
 880     if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
 881       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 882         C->record_for_igvn(n->fast_out(i));
 883       }
 884     }
 885   }
 886   for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
 887     ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
 888     if (!useful.member(n)) {
 889       state()->remove_enqueue_barrier(n);
 890     }
 891   }
 892   for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
 893     ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
 894     if (!useful.member(n)) {
 895       state()->remove_load_reference_barrier(n);
 896     }
 897   }
 898 }
 899 
 900 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
 901   return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
 902 }
 903 
 904 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
 905   return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
 906 }
 907 
 908 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
 909 // expanded later, then now is the time to do so.
 910 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
 911 
 912 #ifdef ASSERT
 913 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
 914   if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
 915     ShenandoahBarrierC2Support::verify(Compile::current()->root());
 916   } else if (phase == BarrierSetC2::BeforeCodeGen) {
 917     // Verify G1 pre-barriers
 918     const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
 919 
 920     ResourceArea *area = Thread::current()->resource_area();
 921     Unique_Node_List visited(area);
 922     Node_List worklist(area);
 923     // We're going to walk control flow backwards starting from the Root
 924     worklist.push(compile->root());
 925     while (worklist.size() > 0) {
 926       Node *x = worklist.pop();
 927       if (x == NULL || x == compile->top()) continue;
 928       if (visited.member(x)) {
 929         continue;
 930       } else {
 931         visited.push(x);
 932       }
 933 
 934       if (x->is_Region()) {
 935         for (uint i = 1; i < x->req(); i++) {
 936           worklist.push(x->in(i));
 937         }
 938       } else {
 939         worklist.push(x->in(0));
 940         // We are looking for the pattern:
 941         //                            /->ThreadLocal
 942         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
 943         //              \->ConI(0)
 944         // We want to verify that the If and the LoadB have the same control
 945         // See GraphKit::g1_write_barrier_pre()
 946         if (x->is_If()) {
 947           IfNode *iff = x->as_If();
 948           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
 949             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
 950             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
 951                 && cmp->in(1)->is_Load()) {
 952               LoadNode *load = cmp->in(1)->as_Load();
 953               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
 954                   && load->in(2)->in(3)->is_Con()
 955                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
 956 
 957                 Node *if_ctrl = iff->in(0);
 958                 Node *load_ctrl = load->in(0);
 959 
 960                 if (if_ctrl != load_ctrl) {
 961                   // Skip possible CProj->NeverBranch in infinite loops
 962                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
 963                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
 964                     if_ctrl = if_ctrl->in(0)->in(0);
 965                   }
 966                 }
 967                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
 968               }
 969             }
 970           }
 971         }
 972       }
 973     }
 974   }
 975 }
 976 #endif
 977 
 978 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
 979   if (is_shenandoah_wb_pre_call(n)) {
 980     uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
 981     if (n->req() > cnt) {
 982       Node* addp = n->in(cnt);
 983       if (has_only_shenandoah_wb_pre_uses(addp)) {
 984         n->del_req(cnt);
 985         if (can_reshape) {
 986           phase->is_IterGVN()->_worklist.push(addp);
 987         }
 988         return n;
 989       }
 990     }
 991   }
 992   if (n->Opcode() == Op_CmpP) {
 993     Node* in1 = n->in(1);
 994     Node* in2 = n->in(2);
 995     if (in1->bottom_type() == TypePtr::NULL_PTR) {
 996       in2 = step_over_gc_barrier(in2);
 997     }
 998     if (in2->bottom_type() == TypePtr::NULL_PTR) {
 999       in1 = step_over_gc_barrier(in1);
1000     }
1001     PhaseIterGVN* igvn = phase->is_IterGVN();
1002     if (in1 != n->in(1)) {
1003       if (igvn != NULL) {
1004         n->set_req_X(1, in1, igvn);
1005       } else {
1006         n->set_req(1, in1);
1007       }
1008       assert(in2 == n->in(2), "only one change");
1009       return n;
1010     }
1011     if (in2 != n->in(2)) {
1012       if (igvn != NULL) {
1013         n->set_req_X(2, in2, igvn);
1014       } else {
1015         n->set_req(2, in2);
1016       }
1017       return n;
1018     }
1019   } else if (can_reshape &&
1020              n->Opcode() == Op_If &&
1021              ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1022              n->in(0) != NULL) {
1023     Node* dom = n->in(0);
1024     Node* prev_dom = n;
1025     int op = n->Opcode();
1026     int dist = 16;
1027     // Search up the dominator tree for another heap stable test
1028     while (dom->Opcode() != op    ||  // Not same opcode?
1029            !ShenandoahBarrierC2Support::is_heap_stable_test(dom) ||  // Not same input 1?
1030            prev_dom->in(0) != dom) {  // One path of test does not dominate?
1031       if (dist < 0) return NULL;
1032 
1033       dist--;
1034       prev_dom = dom;
1035       dom = IfNode::up_one_dom(dom);
1036       if (!dom) return NULL;
1037     }
1038 
1039     // Check that we did not follow a loop back to ourselves
1040     if (n == dom) {
1041       return NULL;
1042     }
1043 
1044     return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1045   }
1046 
1047   return NULL;
1048 }
1049 
1050 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1051   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1052     Node* u = n->fast_out(i);
1053     if (!is_shenandoah_wb_pre_call(u)) {
1054       return false;
1055     }
1056   }
1057   return n->outcnt() > 0;
1058 }
1059 
1060 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const {
1061   switch (opcode) {
1062     case Op_CallLeaf:
1063     case Op_CallLeafNoFP: {
1064       assert (n->is_Call(), "");
1065       CallNode *call = n->as_Call();
1066       if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1067         uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1068         if (call->req() > cnt) {
1069           assert(call->req() == cnt + 1, "only one extra input");
1070           Node *addp = call->in(cnt);
1071           assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1072           call->del_req(cnt);
1073         }
1074       }
1075       return false;
1076     }
1077     case Op_ShenandoahCompareAndSwapP:
1078     case Op_ShenandoahCompareAndSwapN:
1079     case Op_ShenandoahWeakCompareAndSwapN:
1080     case Op_ShenandoahWeakCompareAndSwapP:
1081     case Op_ShenandoahCompareAndExchangeP:
1082     case Op_ShenandoahCompareAndExchangeN:
1083 #ifdef ASSERT
1084       if( VerifyOptoOopOffsets ) {
1085         MemNode* mem  = n->as_Mem();
1086         // Check to see if address types have grounded out somehow.
1087         const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
1088         ciInstanceKlass *k = tp->klass()->as_instance_klass();
1089         bool oop_offset_is_sane = k->contains_field_offset(tp->offset());
1090         assert( !tp || oop_offset_is_sane, "" );
1091       }
1092 #endif
1093       return true;
1094     case Op_ShenandoahLoadReferenceBarrier:
1095       assert(false, "should have been expanded already");
1096       return true;
1097     default:
1098       return false;
1099   }
1100 }
1101 
1102 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1103   switch (opcode) {
1104     case Op_ShenandoahCompareAndExchangeP:
1105     case Op_ShenandoahCompareAndExchangeN:
1106       conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1107       // fallthrough
1108     case Op_ShenandoahWeakCompareAndSwapP:
1109     case Op_ShenandoahWeakCompareAndSwapN:
1110     case Op_ShenandoahCompareAndSwapP:
1111     case Op_ShenandoahCompareAndSwapN:
1112       conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1113       return true;
1114     case Op_StoreP: {
1115       Node* adr = n->in(MemNode::Address);
1116       const Type* adr_type = gvn->type(adr);
1117       // Pointer stores in G1 barriers looks like unsafe access.
1118       // Ignore such stores to be able scalar replace non-escaping
1119       // allocations.
1120       if (adr_type->isa_rawptr() && adr->is_AddP()) {
1121         Node* base = conn_graph->get_addp_base(adr);
1122         if (base->Opcode() == Op_LoadP &&
1123           base->in(MemNode::Address)->is_AddP()) {
1124           adr = base->in(MemNode::Address);
1125           Node* tls = conn_graph->get_addp_base(adr);
1126           if (tls->Opcode() == Op_ThreadLocal) {
1127              int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1128              const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1129              if (offs == buf_offset) {
1130                return true; // Pre barrier previous oop value store.
1131              }
1132           }
1133         }
1134       }
1135       return false;
1136     }
1137     case Op_ShenandoahEnqueueBarrier:
1138       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1139       break;
1140     case Op_ShenandoahLoadReferenceBarrier:
1141       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1142       return true;
1143     default:
1144       // Nothing
1145       break;
1146   }
1147   return false;
1148 }
1149 
1150 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1151   switch (opcode) {
1152     case Op_ShenandoahCompareAndExchangeP:
1153     case Op_ShenandoahCompareAndExchangeN: {
1154       Node *adr = n->in(MemNode::Address);
1155       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
1156       // fallthrough
1157     }
1158     case Op_ShenandoahCompareAndSwapP:
1159     case Op_ShenandoahCompareAndSwapN:
1160     case Op_ShenandoahWeakCompareAndSwapP:
1161     case Op_ShenandoahWeakCompareAndSwapN:
1162       return conn_graph->add_final_edges_unsafe_access(n, opcode);
1163     case Op_ShenandoahEnqueueBarrier:
1164       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL);
1165       return true;
1166     case Op_ShenandoahLoadReferenceBarrier:
1167       conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), NULL);
1168       return true;
1169     default:
1170       // Nothing
1171       break;
1172   }
1173   return false;
1174 }
1175 
1176 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1177   return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1178          n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1179 
1180 }
1181 
1182 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1183   switch (opcode) {
1184     case Op_ShenandoahCompareAndExchangeP:
1185     case Op_ShenandoahCompareAndExchangeN:
1186     case Op_ShenandoahWeakCompareAndSwapP:
1187     case Op_ShenandoahWeakCompareAndSwapN:
1188     case Op_ShenandoahCompareAndSwapP:
1189     case Op_ShenandoahCompareAndSwapN: {   // Convert trinary to binary-tree
1190       Node* newval = n->in(MemNode::ValueIn);
1191       Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1192       Node* pair = new BinaryNode(oldval, newval);
1193       n->set_req(MemNode::ValueIn,pair);
1194       n->del_req(LoadStoreConditionalNode::ExpectedIn);
1195       return true;
1196     }
1197     default:
1198       break;
1199   }
1200   return false;
1201 }
1202 
1203 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1204   return xop == Op_ShenandoahCompareAndExchangeP ||
1205          xop == Op_ShenandoahCompareAndExchangeN ||
1206          xop == Op_ShenandoahWeakCompareAndSwapP ||
1207          xop == Op_ShenandoahWeakCompareAndSwapN ||
1208          xop == Op_ShenandoahCompareAndSwapN ||
1209          xop == Op_ShenandoahCompareAndSwapP;
1210 }