< prev index next >

src/hotspot/share/opto/memnode.cpp

Print this page

   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "precompiled.hpp"

  27 #include "classfile/javaClasses.hpp"

  28 #include "compiler/compileLog.hpp"
  29 #include "gc/shared/barrierSet.hpp"
  30 #include "gc/shared/c2/barrierSetC2.hpp"
  31 #include "gc/shared/tlab_globals.hpp"
  32 #include "memory/allocation.inline.hpp"
  33 #include "memory/resourceArea.hpp"
  34 #include "oops/objArrayKlass.hpp"
  35 #include "opto/addnode.hpp"
  36 #include "opto/arraycopynode.hpp"
  37 #include "opto/cfgnode.hpp"
  38 #include "opto/regalloc.hpp"
  39 #include "opto/compile.hpp"
  40 #include "opto/connode.hpp"
  41 #include "opto/convertnode.hpp"

  42 #include "opto/loopnode.hpp"
  43 #include "opto/machnode.hpp"
  44 #include "opto/matcher.hpp"
  45 #include "opto/memnode.hpp"
  46 #include "opto/mulnode.hpp"
  47 #include "opto/narrowptrnode.hpp"
  48 #include "opto/phaseX.hpp"
  49 #include "opto/regmask.hpp"
  50 #include "opto/rootnode.hpp"
  51 #include "opto/vectornode.hpp"
  52 #include "utilities/align.hpp"
  53 #include "utilities/copy.hpp"
  54 #include "utilities/macros.hpp"
  55 #include "utilities/powerOfTwo.hpp"
  56 #include "utilities/vmError.hpp"
  57 
  58 // Portions of code courtesy of Clifford Click
  59 
  60 // Optimization - Graph Style
  61 

 215   bool is_instance = t_oop->is_known_instance_field();
 216   PhaseIterGVN *igvn = phase->is_IterGVN();
 217   if (is_instance && igvn != nullptr && result->is_Phi()) {
 218     PhiNode *mphi = result->as_Phi();
 219     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 220     const TypePtr *t = mphi->adr_type();
 221     bool do_split = false;
 222     // In the following cases, Load memory input can be further optimized based on
 223     // its precise address type
 224     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
 225       do_split = true;
 226     } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
 227       const TypeOopPtr* mem_t =
 228         t->is_oopptr()->cast_to_exactness(true)
 229         ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 230         ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
 231       if (t_oop->isa_aryptr()) {
 232         mem_t = mem_t->is_aryptr()
 233                      ->cast_to_stable(t_oop->is_aryptr()->is_stable())
 234                      ->cast_to_size(t_oop->is_aryptr()->size())


 235                      ->with_offset(t_oop->is_aryptr()->offset())
 236                      ->is_aryptr();
 237       }
 238       do_split = mem_t == t_oop;
 239     }
 240     if (do_split) {
 241       // clone the Phi with our address type
 242       result = mphi->split_out_instance(t_adr, igvn);
 243     } else {
 244       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 245     }
 246   }
 247   return result;
 248 }
 249 
 250 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 251   uint alias_idx = phase->C->get_alias_index(tp);
 252   Node *mem = mmem;
 253 #ifdef ASSERT
 254   {
 255     // Check that current type is consistent with the alias index used during graph construction
 256     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 257     bool consistent =  adr_check == nullptr || adr_check->empty() ||
 258                        phase->C->must_alias(adr_check, alias_idx );
 259     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 260     if( !consistent && adr_check != nullptr && !adr_check->empty() &&
 261                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 262         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 263         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 264           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 265           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 266       // don't assert if it is dead code.
 267       consistent = true;
 268     }
 269     if( !consistent ) {
 270       st->print("alias_idx==%d, adr_check==", alias_idx);
 271       if( adr_check == nullptr ) {
 272         st->print("null");
 273       } else {
 274         adr_check->dump();
 275       }
 276       st->cr();
 277       print_alias_types();
 278       assert(consistent, "adr_check must match alias idx");
 279     }
 280   }
 281 #endif

 994     Node* ld = gvn.transform(load);
 995     return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
 996   }
 997 
 998   return load;
 999 }
1000 
1001 //------------------------------hash-------------------------------------------
1002 uint LoadNode::hash() const {
1003   // unroll addition of interesting fields
1004   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1005 }
1006 
1007 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1008   if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1009     bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1010     bool is_stable_ary = FoldStableValues &&
1011                          (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1012                          tp->isa_aryptr()->is_stable();
1013 
1014     return (eliminate_boxing && non_volatile) || is_stable_ary;
1015   }
1016 
1017   return false;
1018 }
1019 
1020 LoadNode* LoadNode::pin_array_access_node() const {
1021   const TypePtr* adr_type = this->adr_type();
1022   if (adr_type != nullptr && adr_type->isa_aryptr()) {
1023     return clone_pinned();
1024   }
1025   return nullptr;
1026 }
1027 
1028 // Is the value loaded previously stored by an arraycopy? If so return
1029 // a load node that reads from the source array so we may be able to
1030 // optimize out the ArrayCopy node later.
1031 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1032   Node* ld_adr = in(MemNode::Address);
1033   intptr_t ld_off = 0;
1034   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);

1051       assert(ld_alloc != nullptr, "need an alloc");
1052       assert(addp->is_AddP(), "address must be addp");
1053       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1054       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1055       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1056       addp->set_req(AddPNode::Base, src);
1057       addp->set_req(AddPNode::Address, src);
1058     } else {
1059       assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1060              ac->as_ArrayCopy()->is_copyof_validated() ||
1061              ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1062       assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1063       addp->set_req(AddPNode::Base, src);
1064       addp->set_req(AddPNode::Address, src);
1065 
1066       const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1067       BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1068       if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1069 
1070       uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1071       uint shift  = exact_log2(type2aelembytes(ary_elem));
1072 
1073       Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1074 #ifdef _LP64
1075       diff = phase->transform(new ConvI2LNode(diff));
1076 #endif
1077       diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1078 
1079       Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1080       addp->set_req(AddPNode::Offset, offset);
1081     }
1082     addp = phase->transform(addp);
1083 #ifdef ASSERT
1084     const TypePtr* adr_type = phase->type(addp)->is_ptr();
1085     ld->_adr_type = adr_type;
1086 #endif
1087     ld->set_req(MemNode::Address, addp);
1088     ld->set_req(0, ctl);
1089     ld->set_req(MemNode::Memory, mem);
1090     return ld;
1091   }

1170         // Same base, same offset.
1171         // Possible improvement for arrays: check index value instead of absolute offset.
1172 
1173         // At this point we have proven something like this setup:
1174         //   B = << base >>
1175         //   L =  LoadQ(AddP(Check/CastPP(B), #Off))
1176         //   S = StoreQ(AddP(             B , #Off), V)
1177         // (Actually, we haven't yet proven the Q's are the same.)
1178         // In other words, we are loading from a casted version of
1179         // the same pointer-and-offset that we stored to.
1180         // Casted version may carry a dependency and it is respected.
1181         // Thus, we are able to replace L by V.
1182       }
1183       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1184       if (store_Opcode() != st->Opcode()) {
1185         return nullptr;
1186       }
1187       // LoadVector/StoreVector needs additional check to ensure the types match.
1188       if (st->is_StoreVector()) {
1189         const TypeVect*  in_vt = st->as_StoreVector()->vect_type();
1190         const TypeVect* out_vt = as_LoadVector()->vect_type();
1191         if (in_vt != out_vt) {
1192           return nullptr;
1193         }
1194       }
1195       return st->in(MemNode::ValueIn);
1196     }
1197 
1198     // A load from a freshly-created object always returns zero.
1199     // (This can happen after LoadNode::Ideal resets the load's memory input
1200     // to find_captured_store, which returned InitializeNode::zero_memory.)
1201     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1202         (st->in(0) == ld_alloc) &&
1203         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1204       // return a zero value for the load's basic type
1205       // (This is one of the few places where a generic PhaseTransform
1206       // can create new nodes.  Think of it as lazily manifesting
1207       // virtually pre-existing constants.)





1208       if (memory_type() != T_VOID) {
1209         if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1210           // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1211           // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1212           // by the ArrayCopyNode.
1213           return phase->zerocon(memory_type());
1214         }
1215       } else {
1216         // TODO: materialize all-zero vector constant
1217         assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1218       }
1219     }
1220 
1221     // A load from an initialization barrier can match a captured store.
1222     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1223       InitializeNode* init = st->in(0)->as_Initialize();
1224       AllocateNode* alloc = init->allocation();
1225       if ((alloc != nullptr) && (alloc == ld_alloc)) {
1226         // examine a captured store value
1227         st = init->find_captured_store(ld_off, memory_size(), phase);

1255 //----------------------is_instance_field_load_with_local_phi------------------
1256 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1257   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1258       in(Address)->is_AddP() ) {
1259     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1260     // Only instances and boxed values.
1261     if( t_oop != nullptr &&
1262         (t_oop->is_ptr_to_boxed_value() ||
1263          t_oop->is_known_instance_field()) &&
1264         t_oop->offset() != Type::OffsetBot &&
1265         t_oop->offset() != Type::OffsetTop) {
1266       return true;
1267     }
1268   }
1269   return false;
1270 }
1271 
1272 //------------------------------Identity---------------------------------------
1273 // Loads are identity if previous store is to same address
1274 Node* LoadNode::Identity(PhaseGVN* phase) {

















1275   // If the previous store-maker is the right kind of Store, and the store is
1276   // to the same address, then we are equal to the value stored.
1277   Node* mem = in(Memory);
1278   Node* value = can_see_stored_value(mem, phase);
1279   if( value ) {
1280     // byte, short & char stores truncate naturally.
1281     // A load has to load the truncated value which requires
1282     // some sort of masking operation and that requires an
1283     // Ideal call instead of an Identity call.
1284     if (memory_size() < BytesPerInt) {
1285       // If the input to the store does not fit with the load's result type,
1286       // it must be truncated via an Ideal call.
1287       if (!phase->type(value)->higher_equal(phase->type(this)))
1288         return this;
1289     }
1290     // (This works even when value is a Con, but LoadNode::Value
1291     // usually runs first, producing the singleton type of the Con.)
1292     if (!has_pinned_control_dependency() || value->is_Con()) {
1293       return value;
1294     } else {

2039       }
2040     }
2041 
2042     // Don't do this for integer types. There is only potential profit if
2043     // the element type t is lower than _type; that is, for int types, if _type is
2044     // more restrictive than t.  This only happens here if one is short and the other
2045     // char (both 16 bits), and in those cases we've made an intentional decision
2046     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2047     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2048     //
2049     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2050     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
2051     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2052     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
2053     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2054     // In fact, that could have been the original type of p1, and p1 could have
2055     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2056     // expression (LShiftL quux 3) independently optimized to the constant 8.
2057     if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2058         && (_type->isa_vect() == nullptr)

2059         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2060       // t might actually be lower than _type, if _type is a unique
2061       // concrete subclass of abstract class t.
2062       if (off_beyond_header || off == Type::OffsetBot) {  // is the offset beyond the header?
2063         const Type* jt = t->join_speculative(_type);
2064         // In any case, do not allow the join, per se, to empty out the type.
2065         if (jt->empty() && !t->empty()) {
2066           // This can happen if a interface-typed array narrows to a class type.
2067           jt = _type;
2068         }
2069 #ifdef ASSERT
2070         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2071           // The pointers in the autobox arrays are always non-null
2072           Node* base = adr->in(AddPNode::Base);
2073           if ((base != nullptr) && base->is_DecodeN()) {
2074             // Get LoadN node which loads IntegerCache.cache field
2075             base = base->in(1);
2076           }
2077           if ((base != nullptr) && base->is_Con()) {
2078             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2079             if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2080               // It could be narrow oop
2081               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2082             }
2083           }
2084         }
2085 #endif
2086         return jt;
2087       }
2088     }
2089   } else if (tp->base() == Type::InstPtr) {
2090     assert( off != Type::OffsetBot ||
2091             // arrays can be cast to Objects
2092             !tp->isa_instptr() ||
2093             tp->is_instptr()->instance_klass()->is_java_lang_Object() ||


2094             // unsafe field access may not have a constant offset
2095             C->has_unsafe_access(),
2096             "Field accesses must be precise" );
2097     // For oop loads, we expect the _type to be precise.
2098 
2099     // Optimize loads from constant fields.
2100     const TypeInstPtr* tinst = tp->is_instptr();



2101     ciObject* const_oop = tinst->const_oop();
2102     if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2103       const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
2104       if (con_type != nullptr) {
2105         return con_type;
2106       }
2107     }
2108   } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2109     assert(off != Type::OffsetBot ||
2110             !tp->isa_instklassptr() ||
2111            // arrays can be cast to Objects
2112            tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2113            // also allow array-loading from the primary supertype
2114            // array during subtype checks
2115            Opcode() == Op_LoadKlass,
2116            "Field accesses must be precise");
2117     // For klass/static loads, we expect the _type to be precise
2118   } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2119     /* With mirrors being an indirect in the Klass*
2120      * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2121      * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2122      *
2123      * So check the type and klass of the node before the LoadP.

2218   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2219     Node* value = can_see_stored_value(mem,phase);
2220     if (value != nullptr && value->is_Con()) {
2221       assert(value->bottom_type()->higher_equal(_type),"sanity");
2222       return value->bottom_type();
2223     }
2224   }
2225 
2226   bool is_vect = (_type->isa_vect() != nullptr);
2227   if (is_instance && !is_vect) {
2228     // If we have an instance type and our memory input is the
2229     // programs's initial memory state, there is no matching store,
2230     // so just return a zero of the appropriate type -
2231     // except if it is vectorized - then we have no zero constant.
2232     Node *mem = in(MemNode::Memory);
2233     if (mem->is_Parm() && mem->in(0)->is_Start()) {
2234       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2235       return Type::get_zero_type(_type->basic_type());
2236     }
2237   }
2238 
2239   Node* alloc = is_new_object_mark_load();
2240   if (alloc != nullptr) {
2241     return TypeX::make(markWord::prototype().value());









2242   }
2243 
2244   return _type;
2245 }
2246 
2247 //------------------------------match_edge-------------------------------------
2248 // Do we Match on this edge index or not?  Match only the address.
2249 uint LoadNode::match_edge(uint idx) const {
2250   return idx == MemNode::Address;
2251 }
2252 
2253 //--------------------------LoadBNode::Ideal--------------------------------------
2254 //
2255 //  If the previous store is to the same address as this load,
2256 //  and the value stored was larger than a byte, replace this load
2257 //  with the value stored truncated to a byte.  If no truncation is
2258 //  needed, the replacement is done in LoadNode::Identity().
2259 //
2260 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2261   Node* mem = in(MemNode::Memory);

2372   return LoadNode::Ideal(phase, can_reshape);
2373 }
2374 
2375 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2376   Node* mem = in(MemNode::Memory);
2377   Node* value = can_see_stored_value(mem,phase);
2378   if (value != nullptr && value->is_Con() &&
2379       !value->bottom_type()->higher_equal(_type)) {
2380     // If the input to the store does not fit with the load's result type,
2381     // it must be truncated. We can't delay until Ideal call since
2382     // a singleton Value is needed for split_thru_phi optimization.
2383     int con = value->get_int();
2384     return TypeInt::make((con << 16) >> 16);
2385   }
2386   return LoadNode::Value(phase);
2387 }
2388 
2389 //=============================================================================
2390 //----------------------------LoadKlassNode::make------------------------------
2391 // Polymorphic factory method:
2392 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {

2393   // sanity check the alias category against the created node type
2394   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2395   assert(adr_type != nullptr, "expecting TypeKlassPtr");
2396 #ifdef _LP64
2397   if (adr_type->is_ptr_to_narrowklass()) {
2398     assert(UseCompressedClassPointers, "no compressed klasses");
2399     Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2400     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2401   }
2402 #endif
2403   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2404   return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2405 }
2406 
2407 //------------------------------Value------------------------------------------
2408 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2409   return klass_value_common(phase);
2410 }
2411 
2412 // In most cases, LoadKlassNode does not have the control input set. If the control

2419   // Either input is TOP ==> the result is TOP
2420   const Type *t1 = phase->type( in(MemNode::Memory) );
2421   if (t1 == Type::TOP)  return Type::TOP;
2422   Node *adr = in(MemNode::Address);
2423   const Type *t2 = phase->type( adr );
2424   if (t2 == Type::TOP)  return Type::TOP;
2425   const TypePtr *tp = t2->is_ptr();
2426   if (TypePtr::above_centerline(tp->ptr()) ||
2427       tp->ptr() == TypePtr::Null)  return Type::TOP;
2428 
2429   // Return a more precise klass, if possible
2430   const TypeInstPtr *tinst = tp->isa_instptr();
2431   if (tinst != nullptr) {
2432     ciInstanceKlass* ik = tinst->instance_klass();
2433     int offset = tinst->offset();
2434     if (ik == phase->C->env()->Class_klass()
2435         && (offset == java_lang_Class::klass_offset() ||
2436             offset == java_lang_Class::array_klass_offset())) {
2437       // We are loading a special hidden field from a Class mirror object,
2438       // the field which points to the VM's Klass metaobject.
2439       ciType* t = tinst->java_mirror_type();

2440       // java_mirror_type returns non-null for compile-time Class constants.
2441       if (t != nullptr) {
2442         // constant oop => constant klass
2443         if (offset == java_lang_Class::array_klass_offset()) {
2444           if (t->is_void()) {
2445             // We cannot create a void array.  Since void is a primitive type return null
2446             // klass.  Users of this result need to do a null check on the returned klass.
2447             return TypePtr::NULL_PTR;
2448           }
2449           return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);




2450         }
2451         if (!t->is_klass()) {
2452           // a primitive Class (e.g., int.class) has null for a klass field
2453           return TypePtr::NULL_PTR;
2454         }
2455         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2456         return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);




2457       }
2458       // non-constant mirror, so we can't tell what's going on
2459     }
2460     if (!tinst->is_loaded())
2461       return _type;             // Bail out if not loaded
2462     if (offset == oopDesc::klass_offset_in_bytes()) {
2463       return tinst->as_klass_type(true);
2464     }
2465   }
2466 
2467   // Check for loading klass from an array
2468   const TypeAryPtr *tary = tp->isa_aryptr();
2469   if (tary != nullptr &&
2470       tary->offset() == oopDesc::klass_offset_in_bytes()) {
2471     return tary->as_klass_type(true);
2472   }
2473 
2474   // Check for loading klass from an array klass
2475   const TypeKlassPtr *tkls = tp->isa_klassptr();
2476   if (tkls != nullptr && !StressReflectiveCode) {
2477     if (!tkls->is_loaded())
2478      return _type;             // Bail out if not loaded
2479     if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2480         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2481       // // Always returning precise element type is incorrect,
2482       // // e.g., element type could be object and array may contain strings
2483       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2484 
2485       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2486       // according to the element type's subclassing.
2487       return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2488     }

3445   }
3446   ss.print_cr("[TraceMergeStores]: with");
3447   merged_input_value->dump("\n", false, &ss);
3448   merged_store->dump("\n", false, &ss);
3449   tty->print("%s", ss.as_string());
3450 }
3451 #endif
3452 
3453 //------------------------------Ideal------------------------------------------
3454 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3455 // When a store immediately follows a relevant allocation/initialization,
3456 // try to capture it into the initialization, or hoist it above.
3457 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3458   Node* p = MemNode::Ideal_common(phase, can_reshape);
3459   if (p)  return (p == NodeSentinel) ? nullptr : p;
3460 
3461   Node* mem     = in(MemNode::Memory);
3462   Node* address = in(MemNode::Address);
3463   Node* value   = in(MemNode::ValueIn);
3464   // Back-to-back stores to same address?  Fold em up.  Generally
3465   // unsafe if I have intervening uses.
3466   {
3467     Node* st = mem;
3468     // If Store 'st' has more than one use, we cannot fold 'st' away.
3469     // For example, 'st' might be the final state at a conditional
3470     // return.  Or, 'st' might be used by some node which is live at
3471     // the same time 'st' is live, which might be unschedulable.  So,
3472     // require exactly ONE user until such time as we clone 'mem' for
3473     // each of 'mem's uses (thus making the exactly-1-user-rule hold
3474     // true).
3475     while (st->is_Store() && st->outcnt() == 1) {
3476       // Looking at a dead closed cycle of memory?
3477       assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3478       assert(Opcode() == st->Opcode() ||
3479              st->Opcode() == Op_StoreVector ||
3480              Opcode() == Op_StoreVector ||
3481              st->Opcode() == Op_StoreVectorScatter ||
3482              Opcode() == Op_StoreVectorScatter ||
3483              phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3484              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3485              (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy

3486              (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3487              "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3488 
3489       if (st->in(MemNode::Address)->eqv_uncast(address) &&
3490           st->as_Store()->memory_size() <= this->memory_size()) {
3491         Node* use = st->raw_out(0);
3492         if (phase->is_IterGVN()) {
3493           phase->is_IterGVN()->rehash_node_delayed(use);
3494         }
3495         // It's OK to do this in the parser, since DU info is always accurate,
3496         // and the parser always refers to nodes via SafePointNode maps.
3497         use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3498         return this;
3499       }
3500       st = st->in(MemNode::Memory);
3501     }
3502   }
3503 
3504 
3505   // Capture an unaliased, unconditional, simple store into an initializer.

3592       const StoreVectorNode* store_vector = as_StoreVector();
3593       const StoreVectorNode* mem_vector = mem->as_StoreVector();
3594       const Node* store_indices = store_vector->indices();
3595       const Node* mem_indices = mem_vector->indices();
3596       const Node* store_mask = store_vector->mask();
3597       const Node* mem_mask = mem_vector->mask();
3598       // Ensure types, indices, and masks match
3599       if (store_vector->vect_type() == mem_vector->vect_type() &&
3600           ((store_indices == nullptr) == (mem_indices == nullptr) &&
3601            (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3602           ((store_mask == nullptr) == (mem_mask == nullptr) &&
3603            (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3604         result = mem;
3605       }
3606     }
3607   }
3608 
3609   // Store of zero anywhere into a freshly-allocated object?
3610   // Then the store is useless.
3611   // (It must already have been captured by the InitializeNode.)
3612   if (result == this &&
3613       ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3614     // a newly allocated object is already all-zeroes everywhere
3615     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {

3616       result = mem;
3617     }
3618 
3619     if (result == this) {
3620       // the store may also apply to zero-bits in an earlier object
3621       Node* prev_mem = find_previous_store(phase);
3622       // Steps (a), (b):  Walk past independent stores to find an exact match.
3623       if (prev_mem != nullptr) {
3624         Node* prev_val = can_see_stored_value(prev_mem, phase);
3625         if (prev_val != nullptr && prev_val == val) {
3626           // prev_val and val might differ by a cast; it would be good
3627           // to keep the more informative of the two.
3628           result = mem;
3629         }
3630       }
3631     }
3632   }
3633 
3634   PhaseIterGVN* igvn = phase->is_IterGVN();
3635   if (result != this && igvn != nullptr) {
3636     MemBarNode* trailing = trailing_membar();
3637     if (trailing != nullptr) {
3638 #ifdef ASSERT
3639       const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();

3915 // Clearing a short array is faster with stores
3916 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3917   // Already know this is a large node, do not try to ideal it
3918   if (_is_large) return nullptr;
3919 
3920   const int unit = BytesPerLong;
3921   const TypeX* t = phase->type(in(2))->isa_intptr_t();
3922   if (!t)  return nullptr;
3923   if (!t->is_con())  return nullptr;
3924   intptr_t raw_count = t->get_con();
3925   intptr_t size = raw_count;
3926   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3927   // Clearing nothing uses the Identity call.
3928   // Negative clears are possible on dead ClearArrays
3929   // (see jck test stmt114.stmt11402.val).
3930   if (size <= 0 || size % unit != 0)  return nullptr;
3931   intptr_t count = size / unit;
3932   // Length too long; communicate this to matchers and assemblers.
3933   // Assemblers are responsible to produce fast hardware clears for it.
3934   if (size > InitArrayShortSize) {
3935     return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3936   } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3937     return nullptr;
3938   }
3939   if (!IdealizeClearArrayNode) return nullptr;
3940   Node *mem = in(1);
3941   if( phase->type(mem)==Type::TOP ) return nullptr;
3942   Node *adr = in(3);
3943   const Type* at = phase->type(adr);
3944   if( at==Type::TOP ) return nullptr;
3945   const TypePtr* atp = at->isa_ptr();
3946   // adjust atp to be the correct array element address type
3947   if (atp == nullptr)  atp = TypePtr::BOTTOM;
3948   else              atp = atp->add_offset(Type::OffsetBot);
3949   // Get base for derived pointer purposes
3950   if( adr->Opcode() != Op_AddP ) Unimplemented();
3951   Node *base = adr->in(1);
3952 
3953   Node *zero = phase->makecon(TypeLong::ZERO);
3954   Node *off  = phase->MakeConX(BytesPerLong);
3955   mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3956   count--;
3957   while( count-- ) {
3958     mem = phase->transform(mem);
3959     adr = phase->transform(new AddPNode(base,adr,off));
3960     mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3961   }
3962   return mem;
3963 }
3964 
3965 //----------------------------step_through----------------------------------
3966 // Return allocation input memory edge if it is different instance
3967 // or itself if it is the one we are looking for.
3968 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
3969   Node* n = *np;
3970   assert(n->is_ClearArray(), "sanity");
3971   intptr_t offset;
3972   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3973   // This method is called only before Allocate nodes are expanded
3974   // during macro nodes expansion. Before that ClearArray nodes are
3975   // only generated in PhaseMacroExpand::generate_arraycopy() (before
3976   // Allocate nodes are expanded) which follows allocations.
3977   assert(alloc != nullptr, "should have allocation");
3978   if (alloc->_idx == instance_id) {
3979     // Can not bypass initialization of the instance we are looking for.
3980     return false;
3981   }
3982   // Otherwise skip it.
3983   InitializeNode* init = alloc->initialization();
3984   if (init != nullptr)
3985     *np = init->in(TypeFunc::Memory);
3986   else
3987     *np = alloc->in(TypeFunc::Memory);
3988   return true;
3989 }
3990 
3991 //----------------------------clear_memory-------------------------------------
3992 // Generate code to initialize object storage to zero.
3993 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,


3994                                    intptr_t start_offset,
3995                                    Node* end_offset,
3996                                    PhaseGVN* phase) {
3997   intptr_t offset = start_offset;
3998 
3999   int unit = BytesPerLong;
4000   if ((offset % unit) != 0) {
4001     Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4002     adr = phase->transform(adr);
4003     const TypePtr* atp = TypeRawPtr::BOTTOM;
4004     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);






4005     mem = phase->transform(mem);
4006     offset += BytesPerInt;
4007   }
4008   assert((offset % unit) == 0, "");
4009 
4010   // Initialize the remaining stuff, if any, with a ClearArray.
4011   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
4012 }
4013 
4014 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,

4015                                    Node* start_offset,
4016                                    Node* end_offset,
4017                                    PhaseGVN* phase) {
4018   if (start_offset == end_offset) {
4019     // nothing to do
4020     return mem;
4021   }
4022 
4023   int unit = BytesPerLong;
4024   Node* zbase = start_offset;
4025   Node* zend  = end_offset;
4026 
4027   // Scale to the unit required by the CPU:
4028   if (!Matcher::init_array_count_is_in_bytes) {
4029     Node* shift = phase->intcon(exact_log2(unit));
4030     zbase = phase->transform(new URShiftXNode(zbase, shift) );
4031     zend  = phase->transform(new URShiftXNode(zend,  shift) );
4032   }
4033 
4034   // Bulk clear double-words
4035   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4036   Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4037   mem = new ClearArrayNode(ctl, mem, zsize, adr, false);



4038   return phase->transform(mem);
4039 }
4040 
4041 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,


4042                                    intptr_t start_offset,
4043                                    intptr_t end_offset,
4044                                    PhaseGVN* phase) {
4045   if (start_offset == end_offset) {
4046     // nothing to do
4047     return mem;
4048   }
4049 
4050   assert((end_offset % BytesPerInt) == 0, "odd end offset");
4051   intptr_t done_offset = end_offset;
4052   if ((done_offset % BytesPerLong) != 0) {
4053     done_offset -= BytesPerInt;
4054   }
4055   if (done_offset > start_offset) {
4056     mem = clear_memory(ctl, mem, dest,
4057                        start_offset, phase->MakeConX(done_offset), phase);
4058   }
4059   if (done_offset < end_offset) { // emit the final 32-bit store
4060     Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4061     adr = phase->transform(adr);
4062     const TypePtr* atp = TypeRawPtr::BOTTOM;
4063     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);






4064     mem = phase->transform(mem);
4065     done_offset += BytesPerInt;
4066   }
4067   assert(done_offset == end_offset, "");
4068   return mem;
4069 }
4070 
4071 //=============================================================================
4072 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4073   : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4074     _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4075 #ifdef ASSERT
4076   , _pair_idx(0)
4077 #endif
4078 {
4079   init_class_id(Class_MemBar);
4080   Node* top = C->top();
4081   init_req(TypeFunc::I_O,top);
4082   init_req(TypeFunc::FramePtr,top);
4083   init_req(TypeFunc::ReturnAdr,top);

4189       PhaseIterGVN* igvn = phase->is_IterGVN();
4190       remove(igvn);
4191       // Must return either the original node (now dead) or a new node
4192       // (Do not return a top here, since that would break the uniqueness of top.)
4193       return new ConINode(TypeInt::ZERO);
4194     }
4195   }
4196   return progress ? this : nullptr;
4197 }
4198 
4199 //------------------------------Value------------------------------------------
4200 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4201   if( !in(0) ) return Type::TOP;
4202   if( phase->type(in(0)) == Type::TOP )
4203     return Type::TOP;
4204   return TypeTuple::MEMBAR;
4205 }
4206 
4207 //------------------------------match------------------------------------------
4208 // Construct projections for memory.
4209 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4210   switch (proj->_con) {
4211   case TypeFunc::Control:
4212   case TypeFunc::Memory:
4213     return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4214   }
4215   ShouldNotReachHere();
4216   return nullptr;
4217 }
4218 
4219 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4220   trailing->_kind = TrailingStore;
4221   leading->_kind = LeadingStore;
4222 #ifdef ASSERT
4223   trailing->_pair_idx = leading->_idx;
4224   leading->_pair_idx = leading->_idx;
4225 #endif
4226 }
4227 
4228 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4229   trailing->_kind = TrailingLoadStore;

4476   return (req() > RawStores);
4477 }
4478 
4479 void InitializeNode::set_complete(PhaseGVN* phase) {
4480   assert(!is_complete(), "caller responsibility");
4481   _is_complete = Complete;
4482 
4483   // After this node is complete, it contains a bunch of
4484   // raw-memory initializations.  There is no need for
4485   // it to have anything to do with non-raw memory effects.
4486   // Therefore, tell all non-raw users to re-optimize themselves,
4487   // after skipping the memory effects of this initialization.
4488   PhaseIterGVN* igvn = phase->is_IterGVN();
4489   if (igvn)  igvn->add_users_to_worklist(this);
4490 }
4491 
4492 // convenience function
4493 // return false if the init contains any stores already
4494 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4495   InitializeNode* init = initialization();
4496   if (init == nullptr || init->is_complete())  return false;


4497   init->remove_extra_zeroes();
4498   // for now, if this allocation has already collected any inits, bail:
4499   if (init->is_non_zero())  return false;
4500   init->set_complete(phase);
4501   return true;
4502 }
4503 
4504 void InitializeNode::remove_extra_zeroes() {
4505   if (req() == RawStores)  return;
4506   Node* zmem = zero_memory();
4507   uint fill = RawStores;
4508   for (uint i = fill; i < req(); i++) {
4509     Node* n = in(i);
4510     if (n->is_top() || n == zmem)  continue;  // skip
4511     if (fill < i)  set_req(fill, n);          // compact
4512     ++fill;
4513   }
4514   // delete any empty spaces created:
4515   while (fill < req()) {
4516     del_req(fill);

4660             // store node that we'd like to capture. We need to check
4661             // the uses of the MergeMemNode.
4662             mems.push(n);
4663           }
4664         } else if (n->is_Mem()) {
4665           Node* other_adr = n->in(MemNode::Address);
4666           if (other_adr == adr) {
4667             failed = true;
4668             break;
4669           } else {
4670             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4671             if (other_t_adr != nullptr) {
4672               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4673               if (other_alias_idx == alias_idx) {
4674                 // A load from the same memory slice as the store right
4675                 // after the InitializeNode. We check the control of the
4676                 // object/array that is loaded from. If it's the same as
4677                 // the store control then we cannot capture the store.
4678                 assert(!n->is_Store(), "2 stores to same slice on same control?");
4679                 Node* base = other_adr;






4680                 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4681                 base = base->in(AddPNode::Base);
4682                 if (base != nullptr) {
4683                   base = base->uncast();
4684                   if (base->is_Proj() && base->in(0) == alloc) {
4685                     failed = true;
4686                     break;
4687                   }
4688                 }
4689               }
4690             }
4691           }
4692         } else {
4693           failed = true;
4694           break;
4695         }
4696       }
4697     }
4698   }
4699   if (failed) {

5246         //   z's_done      12  16  16  16    12  16    12
5247         //   z's_needed    12  16  16  16    16  16    16
5248         //   zsize          0   0   0   0     4   0     4
5249         if (next_full_store < 0) {
5250           // Conservative tack:  Zero to end of current word.
5251           zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5252         } else {
5253           // Zero to beginning of next fully initialized word.
5254           // Or, don't zero at all, if we are already in that word.
5255           assert(next_full_store >= zeroes_needed, "must go forward");
5256           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5257           zeroes_needed = next_full_store;
5258         }
5259       }
5260 
5261       if (zeroes_needed > zeroes_done) {
5262         intptr_t zsize = zeroes_needed - zeroes_done;
5263         // Do some incremental zeroing on rawmem, in parallel with inits.
5264         zeroes_done = align_down(zeroes_done, BytesPerInt);
5265         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,


5266                                               zeroes_done, zeroes_needed,
5267                                               phase);
5268         zeroes_done = zeroes_needed;
5269         if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5270           do_zeroing = false;   // leave the hole, next time
5271       }
5272     }
5273 
5274     // Collect the store and move on:
5275     phase->replace_input_of(st, MemNode::Memory, inits);
5276     inits = st;                 // put it on the linearized chain
5277     set_req(i, zmem);           // unhook from previous position
5278 
5279     if (zeroes_done == st_off)
5280       zeroes_done = next_init_off;
5281 
5282     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5283 
5284     #ifdef ASSERT
5285     // Various order invariants.  Weaker than stores_are_sane because

5305   remove_extra_zeroes();        // clear out all the zmems left over
5306   add_req(inits);
5307 
5308   if (!(UseTLAB && ZeroTLAB)) {
5309     // If anything remains to be zeroed, zero it all now.
5310     zeroes_done = align_down(zeroes_done, BytesPerInt);
5311     // if it is the last unused 4 bytes of an instance, forget about it
5312     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5313     if (zeroes_done + BytesPerLong >= size_limit) {
5314       AllocateNode* alloc = allocation();
5315       assert(alloc != nullptr, "must be present");
5316       if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5317         Node* klass_node = alloc->in(AllocateNode::KlassNode);
5318         ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5319         if (zeroes_done == k->layout_helper())
5320           zeroes_done = size_limit;
5321       }
5322     }
5323     if (zeroes_done < size_limit) {
5324       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,


5325                                             zeroes_done, size_in_bytes, phase);
5326     }
5327   }
5328 
5329   set_complete(phase);
5330   return rawmem;
5331 }
5332 
5333 
5334 #ifdef ASSERT
5335 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5336   if (is_complete())
5337     return true;                // stores could be anything at this point
5338   assert(allocation() != nullptr, "must be present");
5339   intptr_t last_off = allocation()->minimum_header_size();
5340   for (uint i = InitializeNode::RawStores; i < req(); i++) {
5341     Node* st = in(i);
5342     intptr_t st_off = get_store_offset(st, phase);
5343     if (st_off < 0)  continue;  // ignore dead garbage
5344     if (last_off > st_off) {

   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "precompiled.hpp"
  27 #include "ci/ciFlatArrayKlass.hpp"
  28 #include "classfile/javaClasses.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "compiler/compileLog.hpp"
  31 #include "gc/shared/barrierSet.hpp"
  32 #include "gc/shared/c2/barrierSetC2.hpp"
  33 #include "gc/shared/tlab_globals.hpp"
  34 #include "memory/allocation.inline.hpp"
  35 #include "memory/resourceArea.hpp"
  36 #include "oops/objArrayKlass.hpp"
  37 #include "opto/addnode.hpp"
  38 #include "opto/arraycopynode.hpp"
  39 #include "opto/cfgnode.hpp"
  40 #include "opto/regalloc.hpp"
  41 #include "opto/compile.hpp"
  42 #include "opto/connode.hpp"
  43 #include "opto/convertnode.hpp"
  44 #include "opto/inlinetypenode.hpp"
  45 #include "opto/loopnode.hpp"
  46 #include "opto/machnode.hpp"
  47 #include "opto/matcher.hpp"
  48 #include "opto/memnode.hpp"
  49 #include "opto/mulnode.hpp"
  50 #include "opto/narrowptrnode.hpp"
  51 #include "opto/phaseX.hpp"
  52 #include "opto/regmask.hpp"
  53 #include "opto/rootnode.hpp"
  54 #include "opto/vectornode.hpp"
  55 #include "utilities/align.hpp"
  56 #include "utilities/copy.hpp"
  57 #include "utilities/macros.hpp"
  58 #include "utilities/powerOfTwo.hpp"
  59 #include "utilities/vmError.hpp"
  60 
  61 // Portions of code courtesy of Clifford Click
  62 
  63 // Optimization - Graph Style
  64 

 218   bool is_instance = t_oop->is_known_instance_field();
 219   PhaseIterGVN *igvn = phase->is_IterGVN();
 220   if (is_instance && igvn != nullptr && result->is_Phi()) {
 221     PhiNode *mphi = result->as_Phi();
 222     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 223     const TypePtr *t = mphi->adr_type();
 224     bool do_split = false;
 225     // In the following cases, Load memory input can be further optimized based on
 226     // its precise address type
 227     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
 228       do_split = true;
 229     } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
 230       const TypeOopPtr* mem_t =
 231         t->is_oopptr()->cast_to_exactness(true)
 232         ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 233         ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
 234       if (t_oop->isa_aryptr()) {
 235         mem_t = mem_t->is_aryptr()
 236                      ->cast_to_stable(t_oop->is_aryptr()->is_stable())
 237                      ->cast_to_size(t_oop->is_aryptr()->size())
 238                      ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat())
 239                      ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free())
 240                      ->with_offset(t_oop->is_aryptr()->offset())
 241                      ->is_aryptr();
 242       }
 243       do_split = mem_t == t_oop;
 244     }
 245     if (do_split) {
 246       // clone the Phi with our address type
 247       result = mphi->split_out_instance(t_adr, igvn);
 248     } else {
 249       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 250     }
 251   }
 252   return result;
 253 }
 254 
 255 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 256   uint alias_idx = phase->C->get_alias_index(tp);
 257   Node *mem = mmem;
 258 #ifdef ASSERT
 259   {
 260     // Check that current type is consistent with the alias index used during graph construction
 261     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 262     bool consistent =  adr_check == nullptr || adr_check->empty() ||
 263                        phase->C->must_alias(adr_check, alias_idx );
 264     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 265     if( !consistent && adr_check != nullptr && !adr_check->empty() &&
 266         tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 267         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 268         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 269           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 270           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 271       // don't assert if it is dead code.
 272       consistent = true;
 273     }
 274     if( !consistent ) {
 275       st->print("alias_idx==%d, adr_check==", alias_idx);
 276       if( adr_check == nullptr ) {
 277         st->print("null");
 278       } else {
 279         adr_check->dump();
 280       }
 281       st->cr();
 282       print_alias_types();
 283       assert(consistent, "adr_check must match alias idx");
 284     }
 285   }
 286 #endif

 999     Node* ld = gvn.transform(load);
1000     return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1001   }
1002 
1003   return load;
1004 }
1005 
1006 //------------------------------hash-------------------------------------------
1007 uint LoadNode::hash() const {
1008   // unroll addition of interesting fields
1009   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1010 }
1011 
1012 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1013   if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1014     bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1015     bool is_stable_ary = FoldStableValues &&
1016                          (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1017                          tp->isa_aryptr()->is_stable();
1018 
1019     return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr();
1020   }
1021 
1022   return false;
1023 }
1024 
1025 LoadNode* LoadNode::pin_array_access_node() const {
1026   const TypePtr* adr_type = this->adr_type();
1027   if (adr_type != nullptr && adr_type->isa_aryptr()) {
1028     return clone_pinned();
1029   }
1030   return nullptr;
1031 }
1032 
1033 // Is the value loaded previously stored by an arraycopy? If so return
1034 // a load node that reads from the source array so we may be able to
1035 // optimize out the ArrayCopy node later.
1036 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1037   Node* ld_adr = in(MemNode::Address);
1038   intptr_t ld_off = 0;
1039   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);

1056       assert(ld_alloc != nullptr, "need an alloc");
1057       assert(addp->is_AddP(), "address must be addp");
1058       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1059       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1060       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1061       addp->set_req(AddPNode::Base, src);
1062       addp->set_req(AddPNode::Address, src);
1063     } else {
1064       assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1065              ac->as_ArrayCopy()->is_copyof_validated() ||
1066              ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1067       assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1068       addp->set_req(AddPNode::Base, src);
1069       addp->set_req(AddPNode::Address, src);
1070 
1071       const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1072       BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1073       if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1074 
1075       uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1076       uint shift  = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1077 
1078       Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1079 #ifdef _LP64
1080       diff = phase->transform(new ConvI2LNode(diff));
1081 #endif
1082       diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1083 
1084       Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1085       addp->set_req(AddPNode::Offset, offset);
1086     }
1087     addp = phase->transform(addp);
1088 #ifdef ASSERT
1089     const TypePtr* adr_type = phase->type(addp)->is_ptr();
1090     ld->_adr_type = adr_type;
1091 #endif
1092     ld->set_req(MemNode::Address, addp);
1093     ld->set_req(0, ctl);
1094     ld->set_req(MemNode::Memory, mem);
1095     return ld;
1096   }

1175         // Same base, same offset.
1176         // Possible improvement for arrays: check index value instead of absolute offset.
1177 
1178         // At this point we have proven something like this setup:
1179         //   B = << base >>
1180         //   L =  LoadQ(AddP(Check/CastPP(B), #Off))
1181         //   S = StoreQ(AddP(             B , #Off), V)
1182         // (Actually, we haven't yet proven the Q's are the same.)
1183         // In other words, we are loading from a casted version of
1184         // the same pointer-and-offset that we stored to.
1185         // Casted version may carry a dependency and it is respected.
1186         // Thus, we are able to replace L by V.
1187       }
1188       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1189       if (store_Opcode() != st->Opcode()) {
1190         return nullptr;
1191       }
1192       // LoadVector/StoreVector needs additional check to ensure the types match.
1193       if (st->is_StoreVector()) {
1194         const TypeVect*  in_vt = st->as_StoreVector()->vect_type();
1195         const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1196         if (in_vt != out_vt) {
1197           return nullptr;
1198         }
1199       }
1200       return st->in(MemNode::ValueIn);
1201     }
1202 
1203     // A load from a freshly-created object always returns zero.
1204     // (This can happen after LoadNode::Ideal resets the load's memory input
1205     // to find_captured_store, which returned InitializeNode::zero_memory.)
1206     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1207         (st->in(0) == ld_alloc) &&
1208         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1209       // return a zero value for the load's basic type
1210       // (This is one of the few places where a generic PhaseTransform
1211       // can create new nodes.  Think of it as lazily manifesting
1212       // virtually pre-existing constants.)
1213       Node* default_value = ld_alloc->in(AllocateNode::DefaultValue);
1214       if (default_value != nullptr) {
1215         return default_value;
1216       }
1217       assert(ld_alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
1218       if (memory_type() != T_VOID) {
1219         if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1220           // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1221           // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1222           // by the ArrayCopyNode.
1223           return phase->zerocon(memory_type());
1224         }
1225       } else {
1226         // TODO: materialize all-zero vector constant
1227         assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1228       }
1229     }
1230 
1231     // A load from an initialization barrier can match a captured store.
1232     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1233       InitializeNode* init = st->in(0)->as_Initialize();
1234       AllocateNode* alloc = init->allocation();
1235       if ((alloc != nullptr) && (alloc == ld_alloc)) {
1236         // examine a captured store value
1237         st = init->find_captured_store(ld_off, memory_size(), phase);

1265 //----------------------is_instance_field_load_with_local_phi------------------
1266 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1267   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1268       in(Address)->is_AddP() ) {
1269     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1270     // Only instances and boxed values.
1271     if( t_oop != nullptr &&
1272         (t_oop->is_ptr_to_boxed_value() ||
1273          t_oop->is_known_instance_field()) &&
1274         t_oop->offset() != Type::OffsetBot &&
1275         t_oop->offset() != Type::OffsetTop) {
1276       return true;
1277     }
1278   }
1279   return false;
1280 }
1281 
1282 //------------------------------Identity---------------------------------------
1283 // Loads are identity if previous store is to same address
1284 Node* LoadNode::Identity(PhaseGVN* phase) {
1285   // Loading from an InlineType? The InlineType has the values of
1286   // all fields as input. Look for the field with matching offset.
1287   Node* addr = in(Address);
1288   intptr_t offset;
1289   Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
1290   if (base != nullptr && base->is_InlineType() && offset > oopDesc::klass_offset_in_bytes()) {
1291     Node* value = base->as_InlineType()->field_value_by_offset((int)offset, true);
1292     if (value != nullptr) {
1293       if (Opcode() == Op_LoadN) {
1294         // Encode oop value if we are loading a narrow oop
1295         assert(!phase->type(value)->isa_narrowoop(), "should already be decoded");
1296         value = phase->transform(new EncodePNode(value, bottom_type()));
1297       }
1298       return value;
1299     }
1300   }
1301 
1302   // If the previous store-maker is the right kind of Store, and the store is
1303   // to the same address, then we are equal to the value stored.
1304   Node* mem = in(Memory);
1305   Node* value = can_see_stored_value(mem, phase);
1306   if( value ) {
1307     // byte, short & char stores truncate naturally.
1308     // A load has to load the truncated value which requires
1309     // some sort of masking operation and that requires an
1310     // Ideal call instead of an Identity call.
1311     if (memory_size() < BytesPerInt) {
1312       // If the input to the store does not fit with the load's result type,
1313       // it must be truncated via an Ideal call.
1314       if (!phase->type(value)->higher_equal(phase->type(this)))
1315         return this;
1316     }
1317     // (This works even when value is a Con, but LoadNode::Value
1318     // usually runs first, producing the singleton type of the Con.)
1319     if (!has_pinned_control_dependency() || value->is_Con()) {
1320       return value;
1321     } else {

2066       }
2067     }
2068 
2069     // Don't do this for integer types. There is only potential profit if
2070     // the element type t is lower than _type; that is, for int types, if _type is
2071     // more restrictive than t.  This only happens here if one is short and the other
2072     // char (both 16 bits), and in those cases we've made an intentional decision
2073     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2074     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2075     //
2076     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2077     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
2078     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2079     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
2080     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2081     // In fact, that could have been the original type of p1, and p1 could have
2082     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2083     // expression (LShiftL quux 3) independently optimized to the constant 8.
2084     if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2085         && (_type->isa_vect() == nullptr)
2086         && !ary->is_flat()
2087         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2088       // t might actually be lower than _type, if _type is a unique
2089       // concrete subclass of abstract class t.
2090       if (off_beyond_header || off == Type::OffsetBot) {  // is the offset beyond the header?
2091         const Type* jt = t->join_speculative(_type);
2092         // In any case, do not allow the join, per se, to empty out the type.
2093         if (jt->empty() && !t->empty()) {
2094           // This can happen if a interface-typed array narrows to a class type.
2095           jt = _type;
2096         }
2097 #ifdef ASSERT
2098         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2099           // The pointers in the autobox arrays are always non-null
2100           Node* base = adr->in(AddPNode::Base);
2101           if ((base != nullptr) && base->is_DecodeN()) {
2102             // Get LoadN node which loads IntegerCache.cache field
2103             base = base->in(1);
2104           }
2105           if ((base != nullptr) && base->is_Con()) {
2106             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2107             if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2108               // It could be narrow oop
2109               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2110             }
2111           }
2112         }
2113 #endif
2114         return jt;
2115       }
2116     }
2117   } else if (tp->base() == Type::InstPtr) {
2118     assert( off != Type::OffsetBot ||
2119             // arrays can be cast to Objects
2120             !tp->isa_instptr() ||
2121             tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2122             // Default value load
2123             tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
2124             // unsafe field access may not have a constant offset
2125             C->has_unsafe_access(),
2126             "Field accesses must be precise" );
2127     // For oop loads, we expect the _type to be precise.
2128 

2129     const TypeInstPtr* tinst = tp->is_instptr();
2130     BasicType bt = memory_type();
2131 
2132     // Optimize loads from constant fields.
2133     ciObject* const_oop = tinst->const_oop();
2134     if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2135       const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2136       if (con_type != nullptr) {
2137         return con_type;
2138       }
2139     }
2140   } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2141     assert(off != Type::OffsetBot ||
2142             !tp->isa_instklassptr() ||
2143            // arrays can be cast to Objects
2144            tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2145            // also allow array-loading from the primary supertype
2146            // array during subtype checks
2147            Opcode() == Op_LoadKlass,
2148            "Field accesses must be precise");
2149     // For klass/static loads, we expect the _type to be precise
2150   } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2151     /* With mirrors being an indirect in the Klass*
2152      * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2153      * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2154      *
2155      * So check the type and klass of the node before the LoadP.

2250   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2251     Node* value = can_see_stored_value(mem,phase);
2252     if (value != nullptr && value->is_Con()) {
2253       assert(value->bottom_type()->higher_equal(_type),"sanity");
2254       return value->bottom_type();
2255     }
2256   }
2257 
2258   bool is_vect = (_type->isa_vect() != nullptr);
2259   if (is_instance && !is_vect) {
2260     // If we have an instance type and our memory input is the
2261     // programs's initial memory state, there is no matching store,
2262     // so just return a zero of the appropriate type -
2263     // except if it is vectorized - then we have no zero constant.
2264     Node *mem = in(MemNode::Memory);
2265     if (mem->is_Parm() && mem->in(0)->is_Start()) {
2266       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2267       return Type::get_zero_type(_type->basic_type());
2268     }
2269   }

2270   Node* alloc = is_new_object_mark_load();
2271   if (alloc != nullptr) {
2272     if (EnableValhalla) {
2273       // The mark word may contain property bits (inline, flat, null-free)
2274       Node* klass_node = alloc->in(AllocateNode::KlassNode);
2275       const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr();
2276       if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) {
2277         return TypeX::make(tkls->exact_klass()->prototype_header().value());
2278       }
2279     } else {
2280       return TypeX::make(markWord::prototype().value());
2281     }
2282   }
2283 
2284   return _type;
2285 }
2286 
2287 //------------------------------match_edge-------------------------------------
2288 // Do we Match on this edge index or not?  Match only the address.
2289 uint LoadNode::match_edge(uint idx) const {
2290   return idx == MemNode::Address;
2291 }
2292 
2293 //--------------------------LoadBNode::Ideal--------------------------------------
2294 //
2295 //  If the previous store is to the same address as this load,
2296 //  and the value stored was larger than a byte, replace this load
2297 //  with the value stored truncated to a byte.  If no truncation is
2298 //  needed, the replacement is done in LoadNode::Identity().
2299 //
2300 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2301   Node* mem = in(MemNode::Memory);

2412   return LoadNode::Ideal(phase, can_reshape);
2413 }
2414 
2415 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2416   Node* mem = in(MemNode::Memory);
2417   Node* value = can_see_stored_value(mem,phase);
2418   if (value != nullptr && value->is_Con() &&
2419       !value->bottom_type()->higher_equal(_type)) {
2420     // If the input to the store does not fit with the load's result type,
2421     // it must be truncated. We can't delay until Ideal call since
2422     // a singleton Value is needed for split_thru_phi optimization.
2423     int con = value->get_int();
2424     return TypeInt::make((con << 16) >> 16);
2425   }
2426   return LoadNode::Value(phase);
2427 }
2428 
2429 //=============================================================================
2430 //----------------------------LoadKlassNode::make------------------------------
2431 // Polymorphic factory method:
2432 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at,
2433                           const TypeKlassPtr* tk) {
2434   // sanity check the alias category against the created node type
2435   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2436   assert(adr_type != nullptr, "expecting TypeKlassPtr");
2437 #ifdef _LP64
2438   if (adr_type->is_ptr_to_narrowklass()) {
2439     assert(UseCompressedClassPointers, "no compressed klasses");
2440     Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2441     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2442   }
2443 #endif
2444   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2445   return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2446 }
2447 
2448 //------------------------------Value------------------------------------------
2449 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2450   return klass_value_common(phase);
2451 }
2452 
2453 // In most cases, LoadKlassNode does not have the control input set. If the control

2460   // Either input is TOP ==> the result is TOP
2461   const Type *t1 = phase->type( in(MemNode::Memory) );
2462   if (t1 == Type::TOP)  return Type::TOP;
2463   Node *adr = in(MemNode::Address);
2464   const Type *t2 = phase->type( adr );
2465   if (t2 == Type::TOP)  return Type::TOP;
2466   const TypePtr *tp = t2->is_ptr();
2467   if (TypePtr::above_centerline(tp->ptr()) ||
2468       tp->ptr() == TypePtr::Null)  return Type::TOP;
2469 
2470   // Return a more precise klass, if possible
2471   const TypeInstPtr *tinst = tp->isa_instptr();
2472   if (tinst != nullptr) {
2473     ciInstanceKlass* ik = tinst->instance_klass();
2474     int offset = tinst->offset();
2475     if (ik == phase->C->env()->Class_klass()
2476         && (offset == java_lang_Class::klass_offset() ||
2477             offset == java_lang_Class::array_klass_offset())) {
2478       // We are loading a special hidden field from a Class mirror object,
2479       // the field which points to the VM's Klass metaobject.
2480       bool is_null_free_array = false;
2481       ciType* t = tinst->java_mirror_type(&is_null_free_array);
2482       // java_mirror_type returns non-null for compile-time Class constants.
2483       if (t != nullptr) {
2484         // constant oop => constant klass
2485         if (offset == java_lang_Class::array_klass_offset()) {
2486           if (t->is_void()) {
2487             // We cannot create a void array.  Since void is a primitive type return null
2488             // klass.  Users of this result need to do a null check on the returned klass.
2489             return TypePtr::NULL_PTR;
2490           }
2491           const TypeKlassPtr* tklass = TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2492           if (is_null_free_array) {
2493             tklass = tklass->is_aryklassptr()->cast_to_null_free();
2494           }
2495           return tklass;
2496         }
2497         if (!t->is_klass()) {
2498           // a primitive Class (e.g., int.class) has null for a klass field
2499           return TypePtr::NULL_PTR;
2500         }
2501         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2502         const TypeKlassPtr* tklass = TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2503         if (is_null_free_array) {
2504           tklass = tklass->is_aryklassptr()->cast_to_null_free();
2505         }
2506         return tklass;
2507       }
2508       // non-constant mirror, so we can't tell what's going on
2509     }
2510     if (!tinst->is_loaded())
2511       return _type;             // Bail out if not loaded
2512     if (offset == oopDesc::klass_offset_in_bytes()) {
2513       return tinst->as_klass_type(true);
2514     }
2515   }
2516 
2517   // Check for loading klass from an array
2518   const TypeAryPtr* tary = tp->isa_aryptr();
2519   if (tary != nullptr &&
2520       tary->offset() == oopDesc::klass_offset_in_bytes()) {
2521     return tary->as_klass_type(true);
2522   }
2523 
2524   // Check for loading klass from an array klass
2525   const TypeKlassPtr *tkls = tp->isa_klassptr();
2526   if (tkls != nullptr && !StressReflectiveCode) {
2527     if (!tkls->is_loaded())
2528      return _type;             // Bail out if not loaded
2529     if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2530         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2531       // // Always returning precise element type is incorrect,
2532       // // e.g., element type could be object and array may contain strings
2533       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2534 
2535       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2536       // according to the element type's subclassing.
2537       return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2538     }

3495   }
3496   ss.print_cr("[TraceMergeStores]: with");
3497   merged_input_value->dump("\n", false, &ss);
3498   merged_store->dump("\n", false, &ss);
3499   tty->print("%s", ss.as_string());
3500 }
3501 #endif
3502 
3503 //------------------------------Ideal------------------------------------------
3504 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3505 // When a store immediately follows a relevant allocation/initialization,
3506 // try to capture it into the initialization, or hoist it above.
3507 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3508   Node* p = MemNode::Ideal_common(phase, can_reshape);
3509   if (p)  return (p == NodeSentinel) ? nullptr : p;
3510 
3511   Node* mem     = in(MemNode::Memory);
3512   Node* address = in(MemNode::Address);
3513   Node* value   = in(MemNode::ValueIn);
3514   // Back-to-back stores to same address?  Fold em up.  Generally
3515   // unsafe if I have intervening uses...
3516   if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
3517     Node* st = mem;
3518     // If Store 'st' has more than one use, we cannot fold 'st' away.
3519     // For example, 'st' might be the final state at a conditional
3520     // return.  Or, 'st' might be used by some node which is live at
3521     // the same time 'st' is live, which might be unschedulable.  So,
3522     // require exactly ONE user until such time as we clone 'mem' for
3523     // each of 'mem's uses (thus making the exactly-1-user-rule hold
3524     // true).
3525     while (st->is_Store() && st->outcnt() == 1) {
3526       // Looking at a dead closed cycle of memory?
3527       assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3528       assert(Opcode() == st->Opcode() ||
3529              st->Opcode() == Op_StoreVector ||
3530              Opcode() == Op_StoreVector ||
3531              st->Opcode() == Op_StoreVectorScatter ||
3532              Opcode() == Op_StoreVectorScatter ||
3533              phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3534              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3535              (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3536              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
3537              (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3538              "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3539 
3540       if (st->in(MemNode::Address)->eqv_uncast(address) &&
3541           st->as_Store()->memory_size() <= this->memory_size()) {
3542         Node* use = st->raw_out(0);
3543         if (phase->is_IterGVN()) {
3544           phase->is_IterGVN()->rehash_node_delayed(use);
3545         }
3546         // It's OK to do this in the parser, since DU info is always accurate,
3547         // and the parser always refers to nodes via SafePointNode maps.
3548         use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3549         return this;
3550       }
3551       st = st->in(MemNode::Memory);
3552     }
3553   }
3554 
3555 
3556   // Capture an unaliased, unconditional, simple store into an initializer.

3643       const StoreVectorNode* store_vector = as_StoreVector();
3644       const StoreVectorNode* mem_vector = mem->as_StoreVector();
3645       const Node* store_indices = store_vector->indices();
3646       const Node* mem_indices = mem_vector->indices();
3647       const Node* store_mask = store_vector->mask();
3648       const Node* mem_mask = mem_vector->mask();
3649       // Ensure types, indices, and masks match
3650       if (store_vector->vect_type() == mem_vector->vect_type() &&
3651           ((store_indices == nullptr) == (mem_indices == nullptr) &&
3652            (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3653           ((store_mask == nullptr) == (mem_mask == nullptr) &&
3654            (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3655         result = mem;
3656       }
3657     }
3658   }
3659 
3660   // Store of zero anywhere into a freshly-allocated object?
3661   // Then the store is useless.
3662   // (It must already have been captured by the InitializeNode.)
3663   if (result == this && ReduceFieldZeroing) {

3664     // a newly allocated object is already all-zeroes everywhere
3665     if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
3666         (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) {
3667       result = mem;
3668     }
3669 
3670     if (result == this && phase->type(val)->is_zero_type()) {
3671       // the store may also apply to zero-bits in an earlier object
3672       Node* prev_mem = find_previous_store(phase);
3673       // Steps (a), (b):  Walk past independent stores to find an exact match.
3674       if (prev_mem != nullptr) {
3675         Node* prev_val = can_see_stored_value(prev_mem, phase);
3676         if (prev_val != nullptr && prev_val == val) {
3677           // prev_val and val might differ by a cast; it would be good
3678           // to keep the more informative of the two.
3679           result = mem;
3680         }
3681       }
3682     }
3683   }
3684 
3685   PhaseIterGVN* igvn = phase->is_IterGVN();
3686   if (result != this && igvn != nullptr) {
3687     MemBarNode* trailing = trailing_membar();
3688     if (trailing != nullptr) {
3689 #ifdef ASSERT
3690       const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();

3966 // Clearing a short array is faster with stores
3967 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3968   // Already know this is a large node, do not try to ideal it
3969   if (_is_large) return nullptr;
3970 
3971   const int unit = BytesPerLong;
3972   const TypeX* t = phase->type(in(2))->isa_intptr_t();
3973   if (!t)  return nullptr;
3974   if (!t->is_con())  return nullptr;
3975   intptr_t raw_count = t->get_con();
3976   intptr_t size = raw_count;
3977   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3978   // Clearing nothing uses the Identity call.
3979   // Negative clears are possible on dead ClearArrays
3980   // (see jck test stmt114.stmt11402.val).
3981   if (size <= 0 || size % unit != 0)  return nullptr;
3982   intptr_t count = size / unit;
3983   // Length too long; communicate this to matchers and assemblers.
3984   // Assemblers are responsible to produce fast hardware clears for it.
3985   if (size > InitArrayShortSize) {
3986     return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
3987   } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3988     return nullptr;
3989   }
3990   if (!IdealizeClearArrayNode) return nullptr;
3991   Node *mem = in(1);
3992   if( phase->type(mem)==Type::TOP ) return nullptr;
3993   Node *adr = in(3);
3994   const Type* at = phase->type(adr);
3995   if( at==Type::TOP ) return nullptr;
3996   const TypePtr* atp = at->isa_ptr();
3997   // adjust atp to be the correct array element address type
3998   if (atp == nullptr)  atp = TypePtr::BOTTOM;
3999   else              atp = atp->add_offset(Type::OffsetBot);
4000   // Get base for derived pointer purposes
4001   if( adr->Opcode() != Op_AddP ) Unimplemented();
4002   Node *base = adr->in(1);
4003 
4004   Node *val = in(4);
4005   Node *off  = phase->MakeConX(BytesPerLong);
4006   mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4007   count--;
4008   while( count-- ) {
4009     mem = phase->transform(mem);
4010     adr = phase->transform(new AddPNode(base,adr,off));
4011     mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4012   }
4013   return mem;
4014 }
4015 
4016 //----------------------------step_through----------------------------------
4017 // Return allocation input memory edge if it is different instance
4018 // or itself if it is the one we are looking for.
4019 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4020   Node* n = *np;
4021   assert(n->is_ClearArray(), "sanity");
4022   intptr_t offset;
4023   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4024   // This method is called only before Allocate nodes are expanded
4025   // during macro nodes expansion. Before that ClearArray nodes are
4026   // only generated in PhaseMacroExpand::generate_arraycopy() (before
4027   // Allocate nodes are expanded) which follows allocations.
4028   assert(alloc != nullptr, "should have allocation");
4029   if (alloc->_idx == instance_id) {
4030     // Can not bypass initialization of the instance we are looking for.
4031     return false;
4032   }
4033   // Otherwise skip it.
4034   InitializeNode* init = alloc->initialization();
4035   if (init != nullptr)
4036     *np = init->in(TypeFunc::Memory);
4037   else
4038     *np = alloc->in(TypeFunc::Memory);
4039   return true;
4040 }
4041 
4042 //----------------------------clear_memory-------------------------------------
4043 // Generate code to initialize object storage to zero.
4044 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4045                                    Node* val,
4046                                    Node* raw_val,
4047                                    intptr_t start_offset,
4048                                    Node* end_offset,
4049                                    PhaseGVN* phase) {
4050   intptr_t offset = start_offset;
4051 
4052   int unit = BytesPerLong;
4053   if ((offset % unit) != 0) {
4054     Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4055     adr = phase->transform(adr);
4056     const TypePtr* atp = TypeRawPtr::BOTTOM;
4057     if (val != nullptr) {
4058       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4059       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4060     } else {
4061       assert(raw_val == nullptr, "val may not be null");
4062       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4063     }
4064     mem = phase->transform(mem);
4065     offset += BytesPerInt;
4066   }
4067   assert((offset % unit) == 0, "");
4068 
4069   // Initialize the remaining stuff, if any, with a ClearArray.
4070   return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
4071 }
4072 
4073 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4074                                    Node* raw_val,
4075                                    Node* start_offset,
4076                                    Node* end_offset,
4077                                    PhaseGVN* phase) {
4078   if (start_offset == end_offset) {
4079     // nothing to do
4080     return mem;
4081   }
4082 
4083   int unit = BytesPerLong;
4084   Node* zbase = start_offset;
4085   Node* zend  = end_offset;
4086 
4087   // Scale to the unit required by the CPU:
4088   if (!Matcher::init_array_count_is_in_bytes) {
4089     Node* shift = phase->intcon(exact_log2(unit));
4090     zbase = phase->transform(new URShiftXNode(zbase, shift) );
4091     zend  = phase->transform(new URShiftXNode(zend,  shift) );
4092   }
4093 
4094   // Bulk clear double-words
4095   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4096   Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4097   if (raw_val == nullptr) {
4098     raw_val = phase->MakeConX(0);
4099   }
4100   mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
4101   return phase->transform(mem);
4102 }
4103 
4104 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4105                                    Node* val,
4106                                    Node* raw_val,
4107                                    intptr_t start_offset,
4108                                    intptr_t end_offset,
4109                                    PhaseGVN* phase) {
4110   if (start_offset == end_offset) {
4111     // nothing to do
4112     return mem;
4113   }
4114 
4115   assert((end_offset % BytesPerInt) == 0, "odd end offset");
4116   intptr_t done_offset = end_offset;
4117   if ((done_offset % BytesPerLong) != 0) {
4118     done_offset -= BytesPerInt;
4119   }
4120   if (done_offset > start_offset) {
4121     mem = clear_memory(ctl, mem, dest, val, raw_val,
4122                        start_offset, phase->MakeConX(done_offset), phase);
4123   }
4124   if (done_offset < end_offset) { // emit the final 32-bit store
4125     Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4126     adr = phase->transform(adr);
4127     const TypePtr* atp = TypeRawPtr::BOTTOM;
4128     if (val != nullptr) {
4129       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4130       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4131     } else {
4132       assert(raw_val == nullptr, "val may not be null");
4133       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4134     }
4135     mem = phase->transform(mem);
4136     done_offset += BytesPerInt;
4137   }
4138   assert(done_offset == end_offset, "");
4139   return mem;
4140 }
4141 
4142 //=============================================================================
4143 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4144   : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4145     _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4146 #ifdef ASSERT
4147   , _pair_idx(0)
4148 #endif
4149 {
4150   init_class_id(Class_MemBar);
4151   Node* top = C->top();
4152   init_req(TypeFunc::I_O,top);
4153   init_req(TypeFunc::FramePtr,top);
4154   init_req(TypeFunc::ReturnAdr,top);

4260       PhaseIterGVN* igvn = phase->is_IterGVN();
4261       remove(igvn);
4262       // Must return either the original node (now dead) or a new node
4263       // (Do not return a top here, since that would break the uniqueness of top.)
4264       return new ConINode(TypeInt::ZERO);
4265     }
4266   }
4267   return progress ? this : nullptr;
4268 }
4269 
4270 //------------------------------Value------------------------------------------
4271 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4272   if( !in(0) ) return Type::TOP;
4273   if( phase->type(in(0)) == Type::TOP )
4274     return Type::TOP;
4275   return TypeTuple::MEMBAR;
4276 }
4277 
4278 //------------------------------match------------------------------------------
4279 // Construct projections for memory.
4280 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
4281   switch (proj->_con) {
4282   case TypeFunc::Control:
4283   case TypeFunc::Memory:
4284     return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4285   }
4286   ShouldNotReachHere();
4287   return nullptr;
4288 }
4289 
4290 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4291   trailing->_kind = TrailingStore;
4292   leading->_kind = LeadingStore;
4293 #ifdef ASSERT
4294   trailing->_pair_idx = leading->_idx;
4295   leading->_pair_idx = leading->_idx;
4296 #endif
4297 }
4298 
4299 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4300   trailing->_kind = TrailingLoadStore;

4547   return (req() > RawStores);
4548 }
4549 
4550 void InitializeNode::set_complete(PhaseGVN* phase) {
4551   assert(!is_complete(), "caller responsibility");
4552   _is_complete = Complete;
4553 
4554   // After this node is complete, it contains a bunch of
4555   // raw-memory initializations.  There is no need for
4556   // it to have anything to do with non-raw memory effects.
4557   // Therefore, tell all non-raw users to re-optimize themselves,
4558   // after skipping the memory effects of this initialization.
4559   PhaseIterGVN* igvn = phase->is_IterGVN();
4560   if (igvn)  igvn->add_users_to_worklist(this);
4561 }
4562 
4563 // convenience function
4564 // return false if the init contains any stores already
4565 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4566   InitializeNode* init = initialization();
4567   if (init == nullptr || init->is_complete()) {
4568     return false;
4569   }
4570   init->remove_extra_zeroes();
4571   // for now, if this allocation has already collected any inits, bail:
4572   if (init->is_non_zero())  return false;
4573   init->set_complete(phase);
4574   return true;
4575 }
4576 
4577 void InitializeNode::remove_extra_zeroes() {
4578   if (req() == RawStores)  return;
4579   Node* zmem = zero_memory();
4580   uint fill = RawStores;
4581   for (uint i = fill; i < req(); i++) {
4582     Node* n = in(i);
4583     if (n->is_top() || n == zmem)  continue;  // skip
4584     if (fill < i)  set_req(fill, n);          // compact
4585     ++fill;
4586   }
4587   // delete any empty spaces created:
4588   while (fill < req()) {
4589     del_req(fill);

4733             // store node that we'd like to capture. We need to check
4734             // the uses of the MergeMemNode.
4735             mems.push(n);
4736           }
4737         } else if (n->is_Mem()) {
4738           Node* other_adr = n->in(MemNode::Address);
4739           if (other_adr == adr) {
4740             failed = true;
4741             break;
4742           } else {
4743             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4744             if (other_t_adr != nullptr) {
4745               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4746               if (other_alias_idx == alias_idx) {
4747                 // A load from the same memory slice as the store right
4748                 // after the InitializeNode. We check the control of the
4749                 // object/array that is loaded from. If it's the same as
4750                 // the store control then we cannot capture the store.
4751                 assert(!n->is_Store(), "2 stores to same slice on same control?");
4752                 Node* base = other_adr;
4753                 if (base->is_Phi()) {
4754                   // In rare case, base may be a PhiNode and it may read
4755                   // the same memory slice between InitializeNode and store.
4756                   failed = true;
4757                   break;
4758                 }
4759                 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4760                 base = base->in(AddPNode::Base);
4761                 if (base != nullptr) {
4762                   base = base->uncast();
4763                   if (base->is_Proj() && base->in(0) == alloc) {
4764                     failed = true;
4765                     break;
4766                   }
4767                 }
4768               }
4769             }
4770           }
4771         } else {
4772           failed = true;
4773           break;
4774         }
4775       }
4776     }
4777   }
4778   if (failed) {

5325         //   z's_done      12  16  16  16    12  16    12
5326         //   z's_needed    12  16  16  16    16  16    16
5327         //   zsize          0   0   0   0     4   0     4
5328         if (next_full_store < 0) {
5329           // Conservative tack:  Zero to end of current word.
5330           zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5331         } else {
5332           // Zero to beginning of next fully initialized word.
5333           // Or, don't zero at all, if we are already in that word.
5334           assert(next_full_store >= zeroes_needed, "must go forward");
5335           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5336           zeroes_needed = next_full_store;
5337         }
5338       }
5339 
5340       if (zeroes_needed > zeroes_done) {
5341         intptr_t zsize = zeroes_needed - zeroes_done;
5342         // Do some incremental zeroing on rawmem, in parallel with inits.
5343         zeroes_done = align_down(zeroes_done, BytesPerInt);
5344         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5345                                               allocation()->in(AllocateNode::DefaultValue),
5346                                               allocation()->in(AllocateNode::RawDefaultValue),
5347                                               zeroes_done, zeroes_needed,
5348                                               phase);
5349         zeroes_done = zeroes_needed;
5350         if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5351           do_zeroing = false;   // leave the hole, next time
5352       }
5353     }
5354 
5355     // Collect the store and move on:
5356     phase->replace_input_of(st, MemNode::Memory, inits);
5357     inits = st;                 // put it on the linearized chain
5358     set_req(i, zmem);           // unhook from previous position
5359 
5360     if (zeroes_done == st_off)
5361       zeroes_done = next_init_off;
5362 
5363     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5364 
5365     #ifdef ASSERT
5366     // Various order invariants.  Weaker than stores_are_sane because

5386   remove_extra_zeroes();        // clear out all the zmems left over
5387   add_req(inits);
5388 
5389   if (!(UseTLAB && ZeroTLAB)) {
5390     // If anything remains to be zeroed, zero it all now.
5391     zeroes_done = align_down(zeroes_done, BytesPerInt);
5392     // if it is the last unused 4 bytes of an instance, forget about it
5393     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5394     if (zeroes_done + BytesPerLong >= size_limit) {
5395       AllocateNode* alloc = allocation();
5396       assert(alloc != nullptr, "must be present");
5397       if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5398         Node* klass_node = alloc->in(AllocateNode::KlassNode);
5399         ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5400         if (zeroes_done == k->layout_helper())
5401           zeroes_done = size_limit;
5402       }
5403     }
5404     if (zeroes_done < size_limit) {
5405       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5406                                             allocation()->in(AllocateNode::DefaultValue),
5407                                             allocation()->in(AllocateNode::RawDefaultValue),
5408                                             zeroes_done, size_in_bytes, phase);
5409     }
5410   }
5411 
5412   set_complete(phase);
5413   return rawmem;
5414 }
5415 
5416 
5417 #ifdef ASSERT
5418 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5419   if (is_complete())
5420     return true;                // stores could be anything at this point
5421   assert(allocation() != nullptr, "must be present");
5422   intptr_t last_off = allocation()->minimum_header_size();
5423   for (uint i = InitializeNode::RawStores; i < req(); i++) {
5424     Node* st = in(i);
5425     intptr_t st_off = get_store_offset(st, phase);
5426     if (st_off < 0)  continue;  // ignore dead garbage
5427     if (last_off > st_off) {
< prev index next >