src/share/vm/opto/memnode.cpp

Sat, 09 Feb 2013 12:55:09 -0800

author
drchase
date
Sat, 09 Feb 2013 12:55:09 -0800
changeset 4585
2c673161698a
parent 4479
b30b3c2a0cf2
child 4657
6931f425c517
permissions
-rw-r--r--

8007402: Code cleanup to remove Parfait false positive
Summary: add array access range check
Reviewed-by: kvn

     1 /*
     2  * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
    20  * or visit www.oracle.com if you need additional information or have any
    21  * questions.
    22  *
    23  */
    25 #include "precompiled.hpp"
    26 #include "classfile/systemDictionary.hpp"
    27 #include "compiler/compileLog.hpp"
    28 #include "memory/allocation.inline.hpp"
    29 #include "oops/objArrayKlass.hpp"
    30 #include "opto/addnode.hpp"
    31 #include "opto/cfgnode.hpp"
    32 #include "opto/compile.hpp"
    33 #include "opto/connode.hpp"
    34 #include "opto/loopnode.hpp"
    35 #include "opto/machnode.hpp"
    36 #include "opto/matcher.hpp"
    37 #include "opto/memnode.hpp"
    38 #include "opto/mulnode.hpp"
    39 #include "opto/phaseX.hpp"
    40 #include "opto/regmask.hpp"
    42 // Portions of code courtesy of Clifford Click
    44 // Optimization - Graph Style
    46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
    48 //=============================================================================
    49 uint MemNode::size_of() const { return sizeof(*this); }
    51 const TypePtr *MemNode::adr_type() const {
    52   Node* adr = in(Address);
    53   const TypePtr* cross_check = NULL;
    54   DEBUG_ONLY(cross_check = _adr_type);
    55   return calculate_adr_type(adr->bottom_type(), cross_check);
    56 }
    58 #ifndef PRODUCT
    59 void MemNode::dump_spec(outputStream *st) const {
    60   if (in(Address) == NULL)  return; // node is dead
    61 #ifndef ASSERT
    62   // fake the missing field
    63   const TypePtr* _adr_type = NULL;
    64   if (in(Address) != NULL)
    65     _adr_type = in(Address)->bottom_type()->isa_ptr();
    66 #endif
    67   dump_adr_type(this, _adr_type, st);
    69   Compile* C = Compile::current();
    70   if( C->alias_type(_adr_type)->is_volatile() )
    71     st->print(" Volatile!");
    72 }
    74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
    75   st->print(" @");
    76   if (adr_type == NULL) {
    77     st->print("NULL");
    78   } else {
    79     adr_type->dump_on(st);
    80     Compile* C = Compile::current();
    81     Compile::AliasType* atp = NULL;
    82     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
    83     if (atp == NULL)
    84       st->print(", idx=?\?;");
    85     else if (atp->index() == Compile::AliasIdxBot)
    86       st->print(", idx=Bot;");
    87     else if (atp->index() == Compile::AliasIdxTop)
    88       st->print(", idx=Top;");
    89     else if (atp->index() == Compile::AliasIdxRaw)
    90       st->print(", idx=Raw;");
    91     else {
    92       ciField* field = atp->field();
    93       if (field) {
    94         st->print(", name=");
    95         field->print_name_on(st);
    96       }
    97       st->print(", idx=%d;", atp->index());
    98     }
    99   }
   100 }
   102 extern void print_alias_types();
   104 #endif
   106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
   107   const TypeOopPtr *tinst = t_adr->isa_oopptr();
   108   if (tinst == NULL || !tinst->is_known_instance_field())
   109     return mchain;  // don't try to optimize non-instance types
   110   uint instance_id = tinst->instance_id();
   111   Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
   112   Node *prev = NULL;
   113   Node *result = mchain;
   114   while (prev != result) {
   115     prev = result;
   116     if (result == start_mem)
   117       break;  // hit one of our sentinels
   118     // skip over a call which does not affect this memory slice
   119     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
   120       Node *proj_in = result->in(0);
   121       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
   122         break;  // hit one of our sentinels
   123       } else if (proj_in->is_Call()) {
   124         CallNode *call = proj_in->as_Call();
   125         if (!call->may_modify(t_adr, phase)) {
   126           result = call->in(TypeFunc::Memory);
   127         }
   128       } else if (proj_in->is_Initialize()) {
   129         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
   130         // Stop if this is the initialization for the object instance which
   131         // which contains this memory slice, otherwise skip over it.
   132         if (alloc != NULL && alloc->_idx != instance_id) {
   133           result = proj_in->in(TypeFunc::Memory);
   134         }
   135       } else if (proj_in->is_MemBar()) {
   136         result = proj_in->in(TypeFunc::Memory);
   137       } else {
   138         assert(false, "unexpected projection");
   139       }
   140     } else if (result->is_ClearArray()) {
   141       if (!ClearArrayNode::step_through(&result, instance_id, phase)) {
   142         // Can not bypass initialization of the instance
   143         // we are looking for.
   144         break;
   145       }
   146       // Otherwise skip it (the call updated 'result' value).
   147     } else if (result->is_MergeMem()) {
   148       result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
   149     }
   150   }
   151   return result;
   152 }
   154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
   155   const TypeOopPtr *t_oop = t_adr->isa_oopptr();
   156   bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
   157   PhaseIterGVN *igvn = phase->is_IterGVN();
   158   Node *result = mchain;
   159   result = optimize_simple_memory_chain(result, t_adr, phase);
   160   if (is_instance && igvn != NULL  && result->is_Phi()) {
   161     PhiNode *mphi = result->as_Phi();
   162     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
   163     const TypePtr *t = mphi->adr_type();
   164     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
   165         t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
   166         t->is_oopptr()->cast_to_exactness(true)
   167          ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
   168          ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
   169       // clone the Phi with our address type
   170       result = mphi->split_out_instance(t_adr, igvn);
   171     } else {
   172       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
   173     }
   174   }
   175   return result;
   176 }
   178 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
   179   uint alias_idx = phase->C->get_alias_index(tp);
   180   Node *mem = mmem;
   181 #ifdef ASSERT
   182   {
   183     // Check that current type is consistent with the alias index used during graph construction
   184     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
   185     bool consistent =  adr_check == NULL || adr_check->empty() ||
   186                        phase->C->must_alias(adr_check, alias_idx );
   187     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
   188     if( !consistent && adr_check != NULL && !adr_check->empty() &&
   189                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
   190         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
   191         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
   192           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
   193           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
   194       // don't assert if it is dead code.
   195       consistent = true;
   196     }
   197     if( !consistent ) {
   198       st->print("alias_idx==%d, adr_check==", alias_idx);
   199       if( adr_check == NULL ) {
   200         st->print("NULL");
   201       } else {
   202         adr_check->dump();
   203       }
   204       st->cr();
   205       print_alias_types();
   206       assert(consistent, "adr_check must match alias idx");
   207     }
   208   }
   209 #endif
   210   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
   211   // means an array I have not precisely typed yet.  Do not do any
   212   // alias stuff with it any time soon.
   213   const TypeOopPtr *toop = tp->isa_oopptr();
   214   if( tp->base() != Type::AnyPtr &&
   215       !(toop &&
   216         toop->klass() != NULL &&
   217         toop->klass()->is_java_lang_Object() &&
   218         toop->offset() == Type::OffsetBot) ) {
   219     // compress paths and change unreachable cycles to TOP
   220     // If not, we can update the input infinitely along a MergeMem cycle
   221     // Equivalent code in PhiNode::Ideal
   222     Node* m  = phase->transform(mmem);
   223     // If transformed to a MergeMem, get the desired slice
   224     // Otherwise the returned node represents memory for every slice
   225     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
   226     // Update input if it is progress over what we have now
   227   }
   228   return mem;
   229 }
   231 //--------------------------Ideal_common---------------------------------------
   232 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
   233 // Unhook non-raw memories from complete (macro-expanded) initializations.
   234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
   235   // If our control input is a dead region, kill all below the region
   236   Node *ctl = in(MemNode::Control);
   237   if (ctl && remove_dead_region(phase, can_reshape))
   238     return this;
   239   ctl = in(MemNode::Control);
   240   // Don't bother trying to transform a dead node
   241   if( ctl && ctl->is_top() )  return NodeSentinel;
   243   PhaseIterGVN *igvn = phase->is_IterGVN();
   244   // Wait if control on the worklist.
   245   if (ctl && can_reshape && igvn != NULL) {
   246     Node* bol = NULL;
   247     Node* cmp = NULL;
   248     if (ctl->in(0)->is_If()) {
   249       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
   250       bol = ctl->in(0)->in(1);
   251       if (bol->is_Bool())
   252         cmp = ctl->in(0)->in(1)->in(1);
   253     }
   254     if (igvn->_worklist.member(ctl) ||
   255         (bol != NULL && igvn->_worklist.member(bol)) ||
   256         (cmp != NULL && igvn->_worklist.member(cmp)) ) {
   257       // This control path may be dead.
   258       // Delay this memory node transformation until the control is processed.
   259       phase->is_IterGVN()->_worklist.push(this);
   260       return NodeSentinel; // caller will return NULL
   261     }
   262   }
   263   // Ignore if memory is dead, or self-loop
   264   Node *mem = in(MemNode::Memory);
   265   if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
   266   assert( mem != this, "dead loop in MemNode::Ideal" );
   268   if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
   269     // This memory slice may be dead.
   270     // Delay this mem node transformation until the memory is processed.
   271     phase->is_IterGVN()->_worklist.push(this);
   272     return NodeSentinel; // caller will return NULL
   273   }
   275   Node *address = in(MemNode::Address);
   276   const Type *t_adr = phase->type( address );
   277   if( t_adr == Type::TOP )              return NodeSentinel; // caller will return NULL
   279   if( can_reshape && igvn != NULL &&
   280       (igvn->_worklist.member(address) ||
   281        igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) {
   282     // The address's base and type may change when the address is processed.
   283     // Delay this mem node transformation until the address is processed.
   284     phase->is_IterGVN()->_worklist.push(this);
   285     return NodeSentinel; // caller will return NULL
   286   }
   288   // Do NOT remove or optimize the next lines: ensure a new alias index
   289   // is allocated for an oop pointer type before Escape Analysis.
   290   // Note: C++ will not remove it since the call has side effect.
   291   if ( t_adr->isa_oopptr() ) {
   292     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
   293   }
   295 #ifdef ASSERT
   296   Node* base = NULL;
   297   if (address->is_AddP())
   298     base = address->in(AddPNode::Base);
   299   assert(base == NULL || t_adr->isa_rawptr() ||
   300         !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
   301 #endif
   303   // Avoid independent memory operations
   304   Node* old_mem = mem;
   306   // The code which unhooks non-raw memories from complete (macro-expanded)
   307   // initializations was removed. After macro-expansion all stores catched
   308   // by Initialize node became raw stores and there is no information
   309   // which memory slices they modify. So it is unsafe to move any memory
   310   // operation above these stores. Also in most cases hooked non-raw memories
   311   // were already unhooked by using information from detect_ptr_independence()
   312   // and find_previous_store().
   314   if (mem->is_MergeMem()) {
   315     MergeMemNode* mmem = mem->as_MergeMem();
   316     const TypePtr *tp = t_adr->is_ptr();
   318     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
   319   }
   321   if (mem != old_mem) {
   322     set_req(MemNode::Memory, mem);
   323     if (phase->type( mem ) == Type::TOP) return NodeSentinel;
   324     return this;
   325   }
   327   // let the subclass continue analyzing...
   328   return NULL;
   329 }
   331 // Helper function for proving some simple control dominations.
   332 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
   333 // Already assumes that 'dom' is available at 'sub', and that 'sub'
   334 // is not a constant (dominated by the method's StartNode).
   335 // Used by MemNode::find_previous_store to prove that the
   336 // control input of a memory operation predates (dominates)
   337 // an allocation it wants to look past.
   338 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
   339   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
   340     return false; // Conservative answer for dead code
   342   // Check 'dom'. Skip Proj and CatchProj nodes.
   343   dom = dom->find_exact_control(dom);
   344   if (dom == NULL || dom->is_top())
   345     return false; // Conservative answer for dead code
   347   if (dom == sub) {
   348     // For the case when, for example, 'sub' is Initialize and the original
   349     // 'dom' is Proj node of the 'sub'.
   350     return false;
   351   }
   353   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
   354     return true;
   356   // 'dom' dominates 'sub' if its control edge and control edges
   357   // of all its inputs dominate or equal to sub's control edge.
   359   // Currently 'sub' is either Allocate, Initialize or Start nodes.
   360   // Or Region for the check in LoadNode::Ideal();
   361   // 'sub' should have sub->in(0) != NULL.
   362   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
   363          sub->is_Region(), "expecting only these nodes");
   365   // Get control edge of 'sub'.
   366   Node* orig_sub = sub;
   367   sub = sub->find_exact_control(sub->in(0));
   368   if (sub == NULL || sub->is_top())
   369     return false; // Conservative answer for dead code
   371   assert(sub->is_CFG(), "expecting control");
   373   if (sub == dom)
   374     return true;
   376   if (sub->is_Start() || sub->is_Root())
   377     return false;
   379   {
   380     // Check all control edges of 'dom'.
   382     ResourceMark rm;
   383     Arena* arena = Thread::current()->resource_area();
   384     Node_List nlist(arena);
   385     Unique_Node_List dom_list(arena);
   387     dom_list.push(dom);
   388     bool only_dominating_controls = false;
   390     for (uint next = 0; next < dom_list.size(); next++) {
   391       Node* n = dom_list.at(next);
   392       if (n == orig_sub)
   393         return false; // One of dom's inputs dominated by sub.
   394       if (!n->is_CFG() && n->pinned()) {
   395         // Check only own control edge for pinned non-control nodes.
   396         n = n->find_exact_control(n->in(0));
   397         if (n == NULL || n->is_top())
   398           return false; // Conservative answer for dead code
   399         assert(n->is_CFG(), "expecting control");
   400         dom_list.push(n);
   401       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
   402         only_dominating_controls = true;
   403       } else if (n->is_CFG()) {
   404         if (n->dominates(sub, nlist))
   405           only_dominating_controls = true;
   406         else
   407           return false;
   408       } else {
   409         // First, own control edge.
   410         Node* m = n->find_exact_control(n->in(0));
   411         if (m != NULL) {
   412           if (m->is_top())
   413             return false; // Conservative answer for dead code
   414           dom_list.push(m);
   415         }
   416         // Now, the rest of edges.
   417         uint cnt = n->req();
   418         for (uint i = 1; i < cnt; i++) {
   419           m = n->find_exact_control(n->in(i));
   420           if (m == NULL || m->is_top())
   421             continue;
   422           dom_list.push(m);
   423         }
   424       }
   425     }
   426     return only_dominating_controls;
   427   }
   428 }
   430 //---------------------detect_ptr_independence---------------------------------
   431 // Used by MemNode::find_previous_store to prove that two base
   432 // pointers are never equal.
   433 // The pointers are accompanied by their associated allocations,
   434 // if any, which have been previously discovered by the caller.
   435 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
   436                                       Node* p2, AllocateNode* a2,
   437                                       PhaseTransform* phase) {
   438   // Attempt to prove that these two pointers cannot be aliased.
   439   // They may both manifestly be allocations, and they should differ.
   440   // Or, if they are not both allocations, they can be distinct constants.
   441   // Otherwise, one is an allocation and the other a pre-existing value.
   442   if (a1 == NULL && a2 == NULL) {           // neither an allocation
   443     return (p1 != p2) && p1->is_Con() && p2->is_Con();
   444   } else if (a1 != NULL && a2 != NULL) {    // both allocations
   445     return (a1 != a2);
   446   } else if (a1 != NULL) {                  // one allocation a1
   447     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
   448     return all_controls_dominate(p2, a1);
   449   } else { //(a2 != NULL)                   // one allocation a2
   450     return all_controls_dominate(p1, a2);
   451   }
   452   return false;
   453 }
   456 // The logic for reordering loads and stores uses four steps:
   457 // (a) Walk carefully past stores and initializations which we
   458 //     can prove are independent of this load.
   459 // (b) Observe that the next memory state makes an exact match
   460 //     with self (load or store), and locate the relevant store.
   461 // (c) Ensure that, if we were to wire self directly to the store,
   462 //     the optimizer would fold it up somehow.
   463 // (d) Do the rewiring, and return, depending on some other part of
   464 //     the optimizer to fold up the load.
   465 // This routine handles steps (a) and (b).  Steps (c) and (d) are
   466 // specific to loads and stores, so they are handled by the callers.
   467 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
   468 //
   469 Node* MemNode::find_previous_store(PhaseTransform* phase) {
   470   Node*         ctrl   = in(MemNode::Control);
   471   Node*         adr    = in(MemNode::Address);
   472   intptr_t      offset = 0;
   473   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
   474   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
   476   if (offset == Type::OffsetBot)
   477     return NULL;            // cannot unalias unless there are precise offsets
   479   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
   481   intptr_t size_in_bytes = memory_size();
   483   Node* mem = in(MemNode::Memory);   // start searching here...
   485   int cnt = 50;             // Cycle limiter
   486   for (;;) {                // While we can dance past unrelated stores...
   487     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
   489     if (mem->is_Store()) {
   490       Node* st_adr = mem->in(MemNode::Address);
   491       intptr_t st_offset = 0;
   492       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
   493       if (st_base == NULL)
   494         break;              // inscrutable pointer
   495       if (st_offset != offset && st_offset != Type::OffsetBot) {
   496         const int MAX_STORE = BytesPerLong;
   497         if (st_offset >= offset + size_in_bytes ||
   498             st_offset <= offset - MAX_STORE ||
   499             st_offset <= offset - mem->as_Store()->memory_size()) {
   500           // Success:  The offsets are provably independent.
   501           // (You may ask, why not just test st_offset != offset and be done?
   502           // The answer is that stores of different sizes can co-exist
   503           // in the same sequence of RawMem effects.  We sometimes initialize
   504           // a whole 'tile' of array elements with a single jint or jlong.)
   505           mem = mem->in(MemNode::Memory);
   506           continue;           // (a) advance through independent store memory
   507         }
   508       }
   509       if (st_base != base &&
   510           detect_ptr_independence(base, alloc,
   511                                   st_base,
   512                                   AllocateNode::Ideal_allocation(st_base, phase),
   513                                   phase)) {
   514         // Success:  The bases are provably independent.
   515         mem = mem->in(MemNode::Memory);
   516         continue;           // (a) advance through independent store memory
   517       }
   519       // (b) At this point, if the bases or offsets do not agree, we lose,
   520       // since we have not managed to prove 'this' and 'mem' independent.
   521       if (st_base == base && st_offset == offset) {
   522         return mem;         // let caller handle steps (c), (d)
   523       }
   525     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
   526       InitializeNode* st_init = mem->in(0)->as_Initialize();
   527       AllocateNode*  st_alloc = st_init->allocation();
   528       if (st_alloc == NULL)
   529         break;              // something degenerated
   530       bool known_identical = false;
   531       bool known_independent = false;
   532       if (alloc == st_alloc)
   533         known_identical = true;
   534       else if (alloc != NULL)
   535         known_independent = true;
   536       else if (all_controls_dominate(this, st_alloc))
   537         known_independent = true;
   539       if (known_independent) {
   540         // The bases are provably independent: Either they are
   541         // manifestly distinct allocations, or else the control
   542         // of this load dominates the store's allocation.
   543         int alias_idx = phase->C->get_alias_index(adr_type());
   544         if (alias_idx == Compile::AliasIdxRaw) {
   545           mem = st_alloc->in(TypeFunc::Memory);
   546         } else {
   547           mem = st_init->memory(alias_idx);
   548         }
   549         continue;           // (a) advance through independent store memory
   550       }
   552       // (b) at this point, if we are not looking at a store initializing
   553       // the same allocation we are loading from, we lose.
   554       if (known_identical) {
   555         // From caller, can_see_stored_value will consult find_captured_store.
   556         return mem;         // let caller handle steps (c), (d)
   557       }
   559     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
   560       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
   561       if (mem->is_Proj() && mem->in(0)->is_Call()) {
   562         CallNode *call = mem->in(0)->as_Call();
   563         if (!call->may_modify(addr_t, phase)) {
   564           mem = call->in(TypeFunc::Memory);
   565           continue;         // (a) advance through independent call memory
   566         }
   567       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
   568         mem = mem->in(0)->in(TypeFunc::Memory);
   569         continue;           // (a) advance through independent MemBar memory
   570       } else if (mem->is_ClearArray()) {
   571         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
   572           // (the call updated 'mem' value)
   573           continue;         // (a) advance through independent allocation memory
   574         } else {
   575           // Can not bypass initialization of the instance
   576           // we are looking for.
   577           return mem;
   578         }
   579       } else if (mem->is_MergeMem()) {
   580         int alias_idx = phase->C->get_alias_index(adr_type());
   581         mem = mem->as_MergeMem()->memory_at(alias_idx);
   582         continue;           // (a) advance through independent MergeMem memory
   583       }
   584     }
   586     // Unless there is an explicit 'continue', we must bail out here,
   587     // because 'mem' is an inscrutable memory state (e.g., a call).
   588     break;
   589   }
   591   return NULL;              // bail out
   592 }
   594 //----------------------calculate_adr_type-------------------------------------
   595 // Helper function.  Notices when the given type of address hits top or bottom.
   596 // Also, asserts a cross-check of the type against the expected address type.
   597 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
   598   if (t == Type::TOP)  return NULL; // does not touch memory any more?
   599   #ifdef PRODUCT
   600   cross_check = NULL;
   601   #else
   602   if (!VerifyAliases || is_error_reported() || Node::in_dump())  cross_check = NULL;
   603   #endif
   604   const TypePtr* tp = t->isa_ptr();
   605   if (tp == NULL) {
   606     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
   607     return TypePtr::BOTTOM;           // touches lots of memory
   608   } else {
   609     #ifdef ASSERT
   610     // %%%% [phh] We don't check the alias index if cross_check is
   611     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
   612     if (cross_check != NULL &&
   613         cross_check != TypePtr::BOTTOM &&
   614         cross_check != TypeRawPtr::BOTTOM) {
   615       // Recheck the alias index, to see if it has changed (due to a bug).
   616       Compile* C = Compile::current();
   617       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
   618              "must stay in the original alias category");
   619       // The type of the address must be contained in the adr_type,
   620       // disregarding "null"-ness.
   621       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
   622       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
   623       assert(cross_check->meet(tp_notnull) == cross_check,
   624              "real address must not escape from expected memory type");
   625     }
   626     #endif
   627     return tp;
   628   }
   629 }
   631 //------------------------adr_phi_is_loop_invariant----------------------------
   632 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
   633 // loop is loop invariant. Make a quick traversal of Phi and associated
   634 // CastPP nodes, looking to see if they are a closed group within the loop.
   635 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
   636   // The idea is that the phi-nest must boil down to only CastPP nodes
   637   // with the same data. This implies that any path into the loop already
   638   // includes such a CastPP, and so the original cast, whatever its input,
   639   // must be covered by an equivalent cast, with an earlier control input.
   640   ResourceMark rm;
   642   // The loop entry input of the phi should be the unique dominating
   643   // node for every Phi/CastPP in the loop.
   644   Unique_Node_List closure;
   645   closure.push(adr_phi->in(LoopNode::EntryControl));
   647   // Add the phi node and the cast to the worklist.
   648   Unique_Node_List worklist;
   649   worklist.push(adr_phi);
   650   if( cast != NULL ){
   651     if( !cast->is_ConstraintCast() ) return false;
   652     worklist.push(cast);
   653   }
   655   // Begin recursive walk of phi nodes.
   656   while( worklist.size() ){
   657     // Take a node off the worklist
   658     Node *n = worklist.pop();
   659     if( !closure.member(n) ){
   660       // Add it to the closure.
   661       closure.push(n);
   662       // Make a sanity check to ensure we don't waste too much time here.
   663       if( closure.size() > 20) return false;
   664       // This node is OK if:
   665       //  - it is a cast of an identical value
   666       //  - or it is a phi node (then we add its inputs to the worklist)
   667       // Otherwise, the node is not OK, and we presume the cast is not invariant
   668       if( n->is_ConstraintCast() ){
   669         worklist.push(n->in(1));
   670       } else if( n->is_Phi() ) {
   671         for( uint i = 1; i < n->req(); i++ ) {
   672           worklist.push(n->in(i));
   673         }
   674       } else {
   675         return false;
   676       }
   677     }
   678   }
   680   // Quit when the worklist is empty, and we've found no offending nodes.
   681   return true;
   682 }
   684 //------------------------------Ideal_DU_postCCP-------------------------------
   685 // Find any cast-away of null-ness and keep its control.  Null cast-aways are
   686 // going away in this pass and we need to make this memory op depend on the
   687 // gating null check.
   688 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
   689   return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
   690 }
   692 // I tried to leave the CastPP's in.  This makes the graph more accurate in
   693 // some sense; we get to keep around the knowledge that an oop is not-null
   694 // after some test.  Alas, the CastPP's interfere with GVN (some values are
   695 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
   696 // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed
   697 // some of the more trivial cases in the optimizer.  Removing more useless
   698 // Phi's started allowing Loads to illegally float above null checks.  I gave
   699 // up on this approach.  CNC 10/20/2000
   700 // This static method may be called not from MemNode (EncodePNode calls it).
   701 // Only the control edge of the node 'n' might be updated.
   702 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
   703   Node *skipped_cast = NULL;
   704   // Need a null check?  Regular static accesses do not because they are
   705   // from constant addresses.  Array ops are gated by the range check (which
   706   // always includes a NULL check).  Just check field ops.
   707   if( n->in(MemNode::Control) == NULL ) {
   708     // Scan upwards for the highest location we can place this memory op.
   709     while( true ) {
   710       switch( adr->Opcode() ) {
   712       case Op_AddP:             // No change to NULL-ness, so peek thru AddP's
   713         adr = adr->in(AddPNode::Base);
   714         continue;
   716       case Op_DecodeN:         // No change to NULL-ness, so peek thru
   717       case Op_DecodeNKlass:
   718         adr = adr->in(1);
   719         continue;
   721       case Op_EncodeP:
   722       case Op_EncodePKlass:
   723         // EncodeP node's control edge could be set by this method
   724         // when EncodeP node depends on CastPP node.
   725         //
   726         // Use its control edge for memory op because EncodeP may go away
   727         // later when it is folded with following or preceding DecodeN node.
   728         if (adr->in(0) == NULL) {
   729           // Keep looking for cast nodes.
   730           adr = adr->in(1);
   731           continue;
   732         }
   733         ccp->hash_delete(n);
   734         n->set_req(MemNode::Control, adr->in(0));
   735         ccp->hash_insert(n);
   736         return n;
   738       case Op_CastPP:
   739         // If the CastPP is useless, just peek on through it.
   740         if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
   741           // Remember the cast that we've peeked though. If we peek
   742           // through more than one, then we end up remembering the highest
   743           // one, that is, if in a loop, the one closest to the top.
   744           skipped_cast = adr;
   745           adr = adr->in(1);
   746           continue;
   747         }
   748         // CastPP is going away in this pass!  We need this memory op to be
   749         // control-dependent on the test that is guarding the CastPP.
   750         ccp->hash_delete(n);
   751         n->set_req(MemNode::Control, adr->in(0));
   752         ccp->hash_insert(n);
   753         return n;
   755       case Op_Phi:
   756         // Attempt to float above a Phi to some dominating point.
   757         if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
   758           // If we've already peeked through a Cast (which could have set the
   759           // control), we can't float above a Phi, because the skipped Cast
   760           // may not be loop invariant.
   761           if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
   762             adr = adr->in(1);
   763             continue;
   764           }
   765         }
   767         // Intentional fallthrough!
   769         // No obvious dominating point.  The mem op is pinned below the Phi
   770         // by the Phi itself.  If the Phi goes away (no true value is merged)
   771         // then the mem op can float, but not indefinitely.  It must be pinned
   772         // behind the controls leading to the Phi.
   773       case Op_CheckCastPP:
   774         // These usually stick around to change address type, however a
   775         // useless one can be elided and we still need to pick up a control edge
   776         if (adr->in(0) == NULL) {
   777           // This CheckCastPP node has NO control and is likely useless. But we
   778           // need check further up the ancestor chain for a control input to keep
   779           // the node in place. 4959717.
   780           skipped_cast = adr;
   781           adr = adr->in(1);
   782           continue;
   783         }
   784         ccp->hash_delete(n);
   785         n->set_req(MemNode::Control, adr->in(0));
   786         ccp->hash_insert(n);
   787         return n;
   789         // List of "safe" opcodes; those that implicitly block the memory
   790         // op below any null check.
   791       case Op_CastX2P:          // no null checks on native pointers
   792       case Op_Parm:             // 'this' pointer is not null
   793       case Op_LoadP:            // Loading from within a klass
   794       case Op_LoadN:            // Loading from within a klass
   795       case Op_LoadKlass:        // Loading from within a klass
   796       case Op_LoadNKlass:       // Loading from within a klass
   797       case Op_ConP:             // Loading from a klass
   798       case Op_ConN:             // Loading from a klass
   799       case Op_ConNKlass:        // Loading from a klass
   800       case Op_CreateEx:         // Sucking up the guts of an exception oop
   801       case Op_Con:              // Reading from TLS
   802       case Op_CMoveP:           // CMoveP is pinned
   803       case Op_CMoveN:           // CMoveN is pinned
   804         break;                  // No progress
   806       case Op_Proj:             // Direct call to an allocation routine
   807       case Op_SCMemProj:        // Memory state from store conditional ops
   808 #ifdef ASSERT
   809         {
   810           assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
   811           const Node* call = adr->in(0);
   812           if (call->is_CallJava()) {
   813             const CallJavaNode* call_java = call->as_CallJava();
   814             const TypeTuple *r = call_java->tf()->range();
   815             assert(r->cnt() > TypeFunc::Parms, "must return value");
   816             const Type* ret_type = r->field_at(TypeFunc::Parms);
   817             assert(ret_type && ret_type->isa_ptr(), "must return pointer");
   818             // We further presume that this is one of
   819             // new_instance_Java, new_array_Java, or
   820             // the like, but do not assert for this.
   821           } else if (call->is_Allocate()) {
   822             // similar case to new_instance_Java, etc.
   823           } else if (!call->is_CallLeaf()) {
   824             // Projections from fetch_oop (OSR) are allowed as well.
   825             ShouldNotReachHere();
   826           }
   827         }
   828 #endif
   829         break;
   830       default:
   831         ShouldNotReachHere();
   832       }
   833       break;
   834     }
   835   }
   837   return  NULL;               // No progress
   838 }
   841 //=============================================================================
   842 uint LoadNode::size_of() const { return sizeof(*this); }
   843 uint LoadNode::cmp( const Node &n ) const
   844 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
   845 const Type *LoadNode::bottom_type() const { return _type; }
   846 uint LoadNode::ideal_reg() const {
   847   return _type->ideal_reg();
   848 }
   850 #ifndef PRODUCT
   851 void LoadNode::dump_spec(outputStream *st) const {
   852   MemNode::dump_spec(st);
   853   if( !Verbose && !WizardMode ) {
   854     // standard dump does this in Verbose and WizardMode
   855     st->print(" #"); _type->dump_on(st);
   856   }
   857 }
   858 #endif
   860 #ifdef ASSERT
   861 //----------------------------is_immutable_value-------------------------------
   862 // Helper function to allow a raw load without control edge for some cases
   863 bool LoadNode::is_immutable_value(Node* adr) {
   864   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
   865           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
   866           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
   867            in_bytes(JavaThread::osthread_offset())));
   868 }
   869 #endif
   871 //----------------------------LoadNode::make-----------------------------------
   872 // Polymorphic factory method:
   873 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
   874   Compile* C = gvn.C;
   876   // sanity check the alias category against the created node type
   877   assert(!(adr_type->isa_oopptr() &&
   878            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
   879          "use LoadKlassNode instead");
   880   assert(!(adr_type->isa_aryptr() &&
   881            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
   882          "use LoadRangeNode instead");
   883   // Check control edge of raw loads
   884   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
   885           // oop will be recorded in oop map if load crosses safepoint
   886           rt->isa_oopptr() || is_immutable_value(adr),
   887           "raw memory operations should have control edge");
   888   switch (bt) {
   889   case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int()    );
   890   case T_BYTE:    return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int()    );
   891   case T_INT:     return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int()    );
   892   case T_CHAR:    return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int()    );
   893   case T_SHORT:   return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int()    );
   894   case T_LONG:    return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long()   );
   895   case T_FLOAT:   return new (C) LoadFNode (ctl, mem, adr, adr_type, rt              );
   896   case T_DOUBLE:  return new (C) LoadDNode (ctl, mem, adr, adr_type, rt              );
   897   case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr()    );
   898   case T_OBJECT:
   899 #ifdef _LP64
   900     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
   901       Node* load  = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
   902       return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
   903     } else
   904 #endif
   905     {
   906       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
   907       return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
   908     }
   909   }
   910   ShouldNotReachHere();
   911   return (LoadNode*)NULL;
   912 }
   914 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
   915   bool require_atomic = true;
   916   return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
   917 }
   922 //------------------------------hash-------------------------------------------
   923 uint LoadNode::hash() const {
   924   // unroll addition of interesting fields
   925   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
   926 }
   928 //---------------------------can_see_stored_value------------------------------
   929 // This routine exists to make sure this set of tests is done the same
   930 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
   931 // will change the graph shape in a way which makes memory alive twice at the
   932 // same time (uses the Oracle model of aliasing), then some
   933 // LoadXNode::Identity will fold things back to the equivalence-class model
   934 // of aliasing.
   935 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
   936   Node* ld_adr = in(MemNode::Address);
   938   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
   939   Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
   940   if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
   941       atp->field() != NULL && !atp->field()->is_volatile()) {
   942     uint alias_idx = atp->index();
   943     bool final = atp->field()->is_final();
   944     Node* result = NULL;
   945     Node* current = st;
   946     // Skip through chains of MemBarNodes checking the MergeMems for
   947     // new states for the slice of this load.  Stop once any other
   948     // kind of node is encountered.  Loads from final memory can skip
   949     // through any kind of MemBar but normal loads shouldn't skip
   950     // through MemBarAcquire since the could allow them to move out of
   951     // a synchronized region.
   952     while (current->is_Proj()) {
   953       int opc = current->in(0)->Opcode();
   954       if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) ||
   955           opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder ||
   956           opc == Op_MemBarReleaseLock) {
   957         Node* mem = current->in(0)->in(TypeFunc::Memory);
   958         if (mem->is_MergeMem()) {
   959           MergeMemNode* merge = mem->as_MergeMem();
   960           Node* new_st = merge->memory_at(alias_idx);
   961           if (new_st == merge->base_memory()) {
   962             // Keep searching
   963             current = merge->base_memory();
   964             continue;
   965           }
   966           // Save the new memory state for the slice and fall through
   967           // to exit.
   968           result = new_st;
   969         }
   970       }
   971       break;
   972     }
   973     if (result != NULL) {
   974       st = result;
   975     }
   976   }
   979   // Loop around twice in the case Load -> Initialize -> Store.
   980   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
   981   for (int trip = 0; trip <= 1; trip++) {
   983     if (st->is_Store()) {
   984       Node* st_adr = st->in(MemNode::Address);
   985       if (!phase->eqv(st_adr, ld_adr)) {
   986         // Try harder before giving up...  Match raw and non-raw pointers.
   987         intptr_t st_off = 0;
   988         AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
   989         if (alloc == NULL)       return NULL;
   990         intptr_t ld_off = 0;
   991         AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
   992         if (alloc != allo2)      return NULL;
   993         if (ld_off != st_off)    return NULL;
   994         // At this point we have proven something like this setup:
   995         //  A = Allocate(...)
   996         //  L = LoadQ(,  AddP(CastPP(, A.Parm),, #Off))
   997         //  S = StoreQ(, AddP(,        A.Parm  , #Off), V)
   998         // (Actually, we haven't yet proven the Q's are the same.)
   999         // In other words, we are loading from a casted version of
  1000         // the same pointer-and-offset that we stored to.
  1001         // Thus, we are able to replace L by V.
  1003       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
  1004       if (store_Opcode() != st->Opcode())
  1005         return NULL;
  1006       return st->in(MemNode::ValueIn);
  1009     intptr_t offset = 0;  // scratch
  1011     // A load from a freshly-created object always returns zero.
  1012     // (This can happen after LoadNode::Ideal resets the load's memory input
  1013     // to find_captured_store, which returned InitializeNode::zero_memory.)
  1014     if (st->is_Proj() && st->in(0)->is_Allocate() &&
  1015         st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
  1016         offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
  1017       // return a zero value for the load's basic type
  1018       // (This is one of the few places where a generic PhaseTransform
  1019       // can create new nodes.  Think of it as lazily manifesting
  1020       // virtually pre-existing constants.)
  1021       return phase->zerocon(memory_type());
  1024     // A load from an initialization barrier can match a captured store.
  1025     if (st->is_Proj() && st->in(0)->is_Initialize()) {
  1026       InitializeNode* init = st->in(0)->as_Initialize();
  1027       AllocateNode* alloc = init->allocation();
  1028       if (alloc != NULL &&
  1029           alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
  1030         // examine a captured store value
  1031         st = init->find_captured_store(offset, memory_size(), phase);
  1032         if (st != NULL)
  1033           continue;             // take one more trip around
  1037     break;
  1040   return NULL;
  1043 //----------------------is_instance_field_load_with_local_phi------------------
  1044 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
  1045   if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
  1046       in(MemNode::Address)->is_AddP() ) {
  1047     const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
  1048     // Only instances.
  1049     if( t_oop != NULL && t_oop->is_known_instance_field() &&
  1050         t_oop->offset() != Type::OffsetBot &&
  1051         t_oop->offset() != Type::OffsetTop) {
  1052       return true;
  1055   return false;
  1058 //------------------------------Identity---------------------------------------
  1059 // Loads are identity if previous store is to same address
  1060 Node *LoadNode::Identity( PhaseTransform *phase ) {
  1061   // If the previous store-maker is the right kind of Store, and the store is
  1062   // to the same address, then we are equal to the value stored.
  1063   Node* mem = in(MemNode::Memory);
  1064   Node* value = can_see_stored_value(mem, phase);
  1065   if( value ) {
  1066     // byte, short & char stores truncate naturally.
  1067     // A load has to load the truncated value which requires
  1068     // some sort of masking operation and that requires an
  1069     // Ideal call instead of an Identity call.
  1070     if (memory_size() < BytesPerInt) {
  1071       // If the input to the store does not fit with the load's result type,
  1072       // it must be truncated via an Ideal call.
  1073       if (!phase->type(value)->higher_equal(phase->type(this)))
  1074         return this;
  1076     // (This works even when value is a Con, but LoadNode::Value
  1077     // usually runs first, producing the singleton type of the Con.)
  1078     return value;
  1081   // Search for an existing data phi which was generated before for the same
  1082   // instance's field to avoid infinite generation of phis in a loop.
  1083   Node *region = mem->in(0);
  1084   if (is_instance_field_load_with_local_phi(region)) {
  1085     const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
  1086     int this_index  = phase->C->get_alias_index(addr_t);
  1087     int this_offset = addr_t->offset();
  1088     int this_id    = addr_t->is_oopptr()->instance_id();
  1089     const Type* this_type = bottom_type();
  1090     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
  1091       Node* phi = region->fast_out(i);
  1092       if (phi->is_Phi() && phi != mem &&
  1093           phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
  1094         return phi;
  1099   return this;
  1103 // Returns true if the AliasType refers to the field that holds the
  1104 // cached box array.  Currently only handles the IntegerCache case.
  1105 static bool is_autobox_cache(Compile::AliasType* atp) {
  1106   if (atp != NULL && atp->field() != NULL) {
  1107     ciField* field = atp->field();
  1108     ciSymbol* klass = field->holder()->name();
  1109     if (field->name() == ciSymbol::cache_field_name() &&
  1110         field->holder()->uses_default_loader() &&
  1111         klass == ciSymbol::java_lang_Integer_IntegerCache()) {
  1112       return true;
  1115   return false;
  1118 // Fetch the base value in the autobox array
  1119 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
  1120   if (atp != NULL && atp->field() != NULL) {
  1121     ciField* field = atp->field();
  1122     ciSymbol* klass = field->holder()->name();
  1123     if (field->name() == ciSymbol::cache_field_name() &&
  1124         field->holder()->uses_default_loader() &&
  1125         klass == ciSymbol::java_lang_Integer_IntegerCache()) {
  1126       assert(field->is_constant(), "what?");
  1127       ciObjArray* array = field->constant_value().as_object()->as_obj_array();
  1128       // Fetch the box object at the base of the array and get its value
  1129       ciInstance* box = array->obj_at(0)->as_instance();
  1130       ciInstanceKlass* ik = box->klass()->as_instance_klass();
  1131       if (ik->nof_nonstatic_fields() == 1) {
  1132         // This should be true nonstatic_field_at requires calling
  1133         // nof_nonstatic_fields so check it anyway
  1134         ciConstant c = box->field_value(ik->nonstatic_field_at(0));
  1135         cache_offset = c.as_int();
  1137       return true;
  1140   return false;
  1143 // Returns true if the AliasType refers to the value field of an
  1144 // autobox object.  Currently only handles Integer.
  1145 static bool is_autobox_object(Compile::AliasType* atp) {
  1146   if (atp != NULL && atp->field() != NULL) {
  1147     ciField* field = atp->field();
  1148     ciSymbol* klass = field->holder()->name();
  1149     if (field->name() == ciSymbol::value_name() &&
  1150         field->holder()->uses_default_loader() &&
  1151         klass == ciSymbol::java_lang_Integer()) {
  1152       return true;
  1155   return false;
  1159 // We're loading from an object which has autobox behaviour.
  1160 // If this object is result of a valueOf call we'll have a phi
  1161 // merging a newly allocated object and a load from the cache.
  1162 // We want to replace this load with the original incoming
  1163 // argument to the valueOf call.
  1164 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
  1165   Node* base = in(Address)->in(AddPNode::Base);
  1166   if (base->is_Phi() && base->req() == 3) {
  1167     AllocateNode* allocation = NULL;
  1168     int allocation_index = -1;
  1169     int load_index = -1;
  1170     for (uint i = 1; i < base->req(); i++) {
  1171       allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
  1172       if (allocation != NULL) {
  1173         allocation_index = i;
  1174         load_index = 3 - allocation_index;
  1175         break;
  1178     bool has_load = ( allocation != NULL &&
  1179                       (base->in(load_index)->is_Load() ||
  1180                        base->in(load_index)->is_DecodeN() &&
  1181                        base->in(load_index)->in(1)->is_Load()) );
  1182     if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
  1183       // Push the loads from the phi that comes from valueOf up
  1184       // through it to allow elimination of the loads and the recovery
  1185       // of the original value.
  1186       Node* mem_phi = in(Memory);
  1187       Node* offset = in(Address)->in(AddPNode::Offset);
  1188       Node* region = base->in(0);
  1190       Node* in1 = clone();
  1191       Node* in1_addr = in1->in(Address)->clone();
  1192       in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
  1193       in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
  1194       in1_addr->set_req(AddPNode::Offset, offset);
  1195       in1->set_req(0, region->in(allocation_index));
  1196       in1->set_req(Address, in1_addr);
  1197       in1->set_req(Memory, mem_phi->in(allocation_index));
  1199       Node* in2 = clone();
  1200       Node* in2_addr = in2->in(Address)->clone();
  1201       in2_addr->set_req(AddPNode::Base, base->in(load_index));
  1202       in2_addr->set_req(AddPNode::Address, base->in(load_index));
  1203       in2_addr->set_req(AddPNode::Offset, offset);
  1204       in2->set_req(0, region->in(load_index));
  1205       in2->set_req(Address, in2_addr);
  1206       in2->set_req(Memory, mem_phi->in(load_index));
  1208       in1_addr = phase->transform(in1_addr);
  1209       in1 =      phase->transform(in1);
  1210       in2_addr = phase->transform(in2_addr);
  1211       in2 =      phase->transform(in2);
  1213       PhiNode* result = PhiNode::make_blank(region, this);
  1214       result->set_req(allocation_index, in1);
  1215       result->set_req(load_index, in2);
  1216       return result;
  1218   } else if (base->is_Load() ||
  1219              base->is_DecodeN() && base->in(1)->is_Load()) {
  1220     if (base->is_DecodeN()) {
  1221       // Get LoadN node which loads cached Integer object
  1222       base = base->in(1);
  1224     // Eliminate the load of Integer.value for integers from the cache
  1225     // array by deriving the value from the index into the array.
  1226     // Capture the offset of the load and then reverse the computation.
  1227     Node* load_base = base->in(Address)->in(AddPNode::Base);
  1228     if (load_base->is_DecodeN()) {
  1229       // Get LoadN node which loads IntegerCache.cache field
  1230       load_base = load_base->in(1);
  1232     if (load_base != NULL) {
  1233       Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
  1234       intptr_t cache_offset;
  1235       int shift = -1;
  1236       Node* cache = NULL;
  1237       if (is_autobox_cache(atp)) {
  1238         shift  = exact_log2(type2aelembytes(T_OBJECT));
  1239         cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
  1241       if (cache != NULL && base->in(Address)->is_AddP()) {
  1242         Node* elements[4];
  1243         int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
  1244         int cache_low;
  1245         if (count > 0 && fetch_autobox_base(atp, cache_low)) {
  1246           int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
  1247           // Add up all the offsets making of the address of the load
  1248           Node* result = elements[0];
  1249           for (int i = 1; i < count; i++) {
  1250             result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
  1252           // Remove the constant offset from the address and then
  1253           // remove the scaling of the offset to recover the original index.
  1254           result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-offset)));
  1255           if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
  1256             // Peel the shift off directly but wrap it in a dummy node
  1257             // since Ideal can't return existing nodes
  1258             result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
  1259           } else {
  1260             result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
  1262 #ifdef _LP64
  1263           result = new (phase->C) ConvL2INode(phase->transform(result));
  1264 #endif
  1265           return result;
  1270   return NULL;
  1273 //------------------------------split_through_phi------------------------------
  1274 // Split instance field load through Phi.
  1275 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
  1276   Node* mem     = in(MemNode::Memory);
  1277   Node* address = in(MemNode::Address);
  1278   const TypePtr *addr_t = phase->type(address)->isa_ptr();
  1279   const TypeOopPtr *t_oop = addr_t->isa_oopptr();
  1281   assert(mem->is_Phi() && (t_oop != NULL) &&
  1282          t_oop->is_known_instance_field(), "invalide conditions");
  1284   Node *region = mem->in(0);
  1285   if (region == NULL) {
  1286     return NULL; // Wait stable graph
  1288   uint cnt = mem->req();
  1289   for (uint i = 1; i < cnt; i++) {
  1290     Node* rc = region->in(i);
  1291     if (rc == NULL || phase->type(rc) == Type::TOP)
  1292       return NULL; // Wait stable graph
  1293     Node *in = mem->in(i);
  1294     if (in == NULL) {
  1295       return NULL; // Wait stable graph
  1298   // Check for loop invariant.
  1299   if (cnt == 3) {
  1300     for (uint i = 1; i < cnt; i++) {
  1301       Node *in = mem->in(i);
  1302       Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
  1303       if (m == mem) {
  1304         set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
  1305         return this;
  1309   // Split through Phi (see original code in loopopts.cpp).
  1310   assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
  1312   // Do nothing here if Identity will find a value
  1313   // (to avoid infinite chain of value phis generation).
  1314   if (!phase->eqv(this, this->Identity(phase)))
  1315     return NULL;
  1317   // Skip the split if the region dominates some control edge of the address.
  1318   if (!MemNode::all_controls_dominate(address, region))
  1319     return NULL;
  1321   const Type* this_type = this->bottom_type();
  1322   int this_index  = phase->C->get_alias_index(addr_t);
  1323   int this_offset = addr_t->offset();
  1324   int this_iid    = addr_t->is_oopptr()->instance_id();
  1325   PhaseIterGVN *igvn = phase->is_IterGVN();
  1326   Node *phi = new (igvn->C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
  1327   for (uint i = 1; i < region->req(); i++) {
  1328     Node *x;
  1329     Node* the_clone = NULL;
  1330     if (region->in(i) == phase->C->top()) {
  1331       x = phase->C->top();      // Dead path?  Use a dead data op
  1332     } else {
  1333       x = this->clone();        // Else clone up the data op
  1334       the_clone = x;            // Remember for possible deletion.
  1335       // Alter data node to use pre-phi inputs
  1336       if (this->in(0) == region) {
  1337         x->set_req(0, region->in(i));
  1338       } else {
  1339         x->set_req(0, NULL);
  1341       for (uint j = 1; j < this->req(); j++) {
  1342         Node *in = this->in(j);
  1343         if (in->is_Phi() && in->in(0) == region)
  1344           x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
  1347     // Check for a 'win' on some paths
  1348     const Type *t = x->Value(igvn);
  1350     bool singleton = t->singleton();
  1352     // See comments in PhaseIdealLoop::split_thru_phi().
  1353     if (singleton && t == Type::TOP) {
  1354       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
  1357     if (singleton) {
  1358       x = igvn->makecon(t);
  1359     } else {
  1360       // We now call Identity to try to simplify the cloned node.
  1361       // Note that some Identity methods call phase->type(this).
  1362       // Make sure that the type array is big enough for
  1363       // our new node, even though we may throw the node away.
  1364       // (This tweaking with igvn only works because x is a new node.)
  1365       igvn->set_type(x, t);
  1366       // If x is a TypeNode, capture any more-precise type permanently into Node
  1367       // otherwise it will be not updated during igvn->transform since
  1368       // igvn->type(x) is set to x->Value() already.
  1369       x->raise_bottom_type(t);
  1370       Node *y = x->Identity(igvn);
  1371       if (y != x) {
  1372         x = y;
  1373       } else {
  1374         y = igvn->hash_find(x);
  1375         if (y) {
  1376           x = y;
  1377         } else {
  1378           // Else x is a new node we are keeping
  1379           // We do not need register_new_node_with_optimizer
  1380           // because set_type has already been called.
  1381           igvn->_worklist.push(x);
  1385     if (x != the_clone && the_clone != NULL)
  1386       igvn->remove_dead_node(the_clone);
  1387     phi->set_req(i, x);
  1389   // Record Phi
  1390   igvn->register_new_node_with_optimizer(phi);
  1391   return phi;
  1394 //------------------------------Ideal------------------------------------------
  1395 // If the load is from Field memory and the pointer is non-null, we can
  1396 // zero out the control input.
  1397 // If the offset is constant and the base is an object allocation,
  1398 // try to hook me up to the exact initializing store.
  1399 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1400   Node* p = MemNode::Ideal_common(phase, can_reshape);
  1401   if (p)  return (p == NodeSentinel) ? NULL : p;
  1403   Node* ctrl    = in(MemNode::Control);
  1404   Node* address = in(MemNode::Address);
  1406   // Skip up past a SafePoint control.  Cannot do this for Stores because
  1407   // pointer stores & cardmarks must stay on the same side of a SafePoint.
  1408   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
  1409       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
  1410     ctrl = ctrl->in(0);
  1411     set_req(MemNode::Control,ctrl);
  1414   intptr_t ignore = 0;
  1415   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
  1416   if (base != NULL
  1417       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
  1418     // Check for useless control edge in some common special cases
  1419     if (in(MemNode::Control) != NULL
  1420         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
  1421         && all_controls_dominate(base, phase->C->start())) {
  1422       // A method-invariant, non-null address (constant or 'this' argument).
  1423       set_req(MemNode::Control, NULL);
  1426     if (EliminateAutoBox && can_reshape) {
  1427       assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
  1428       Compile::AliasType* atp = phase->C->alias_type(adr_type());
  1429       if (is_autobox_object(atp)) {
  1430         Node* result = eliminate_autobox(phase);
  1431         if (result != NULL) return result;
  1436   Node* mem = in(MemNode::Memory);
  1437   const TypePtr *addr_t = phase->type(address)->isa_ptr();
  1439   if (addr_t != NULL) {
  1440     // try to optimize our memory input
  1441     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
  1442     if (opt_mem != mem) {
  1443       set_req(MemNode::Memory, opt_mem);
  1444       if (phase->type( opt_mem ) == Type::TOP) return NULL;
  1445       return this;
  1447     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
  1448     if (can_reshape && opt_mem->is_Phi() &&
  1449         (t_oop != NULL) && t_oop->is_known_instance_field()) {
  1450       PhaseIterGVN *igvn = phase->is_IterGVN();
  1451       if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
  1452         // Delay this transformation until memory Phi is processed.
  1453         phase->is_IterGVN()->_worklist.push(this);
  1454         return NULL;
  1456       // Split instance field load through Phi.
  1457       Node* result = split_through_phi(phase);
  1458       if (result != NULL) return result;
  1462   // Check for prior store with a different base or offset; make Load
  1463   // independent.  Skip through any number of them.  Bail out if the stores
  1464   // are in an endless dead cycle and report no progress.  This is a key
  1465   // transform for Reflection.  However, if after skipping through the Stores
  1466   // we can't then fold up against a prior store do NOT do the transform as
  1467   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
  1468   // array memory alive twice: once for the hoisted Load and again after the
  1469   // bypassed Store.  This situation only works if EVERYBODY who does
  1470   // anti-dependence work knows how to bypass.  I.e. we need all
  1471   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
  1472   // the alias index stuff.  So instead, peek through Stores and IFF we can
  1473   // fold up, do so.
  1474   Node* prev_mem = find_previous_store(phase);
  1475   // Steps (a), (b):  Walk past independent stores to find an exact match.
  1476   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
  1477     // (c) See if we can fold up on the spot, but don't fold up here.
  1478     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
  1479     // just return a prior value, which is done by Identity calls.
  1480     if (can_see_stored_value(prev_mem, phase)) {
  1481       // Make ready for step (d):
  1482       set_req(MemNode::Memory, prev_mem);
  1483       return this;
  1487   return NULL;                  // No further progress
  1490 // Helper to recognize certain Klass fields which are invariant across
  1491 // some group of array types (e.g., int[] or all T[] where T < Object).
  1492 const Type*
  1493 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
  1494                                  ciKlass* klass) const {
  1495   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
  1496     // The field is Klass::_modifier_flags.  Return its (constant) value.
  1497     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
  1498     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
  1499     return TypeInt::make(klass->modifier_flags());
  1501   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
  1502     // The field is Klass::_access_flags.  Return its (constant) value.
  1503     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
  1504     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
  1505     return TypeInt::make(klass->access_flags());
  1507   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
  1508     // The field is Klass::_layout_helper.  Return its constant value if known.
  1509     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
  1510     return TypeInt::make(klass->layout_helper());
  1513   // No match.
  1514   return NULL;
  1517 //------------------------------Value-----------------------------------------
  1518 const Type *LoadNode::Value( PhaseTransform *phase ) const {
  1519   // Either input is TOP ==> the result is TOP
  1520   Node* mem = in(MemNode::Memory);
  1521   const Type *t1 = phase->type(mem);
  1522   if (t1 == Type::TOP)  return Type::TOP;
  1523   Node* adr = in(MemNode::Address);
  1524   const TypePtr* tp = phase->type(adr)->isa_ptr();
  1525   if (tp == NULL || tp->empty())  return Type::TOP;
  1526   int off = tp->offset();
  1527   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
  1528   Compile* C = phase->C;
  1530   // Try to guess loaded type from pointer type
  1531   if (tp->base() == Type::AryPtr) {
  1532     const Type *t = tp->is_aryptr()->elem();
  1533     // Don't do this for integer types. There is only potential profit if
  1534     // the element type t is lower than _type; that is, for int types, if _type is
  1535     // more restrictive than t.  This only happens here if one is short and the other
  1536     // char (both 16 bits), and in those cases we've made an intentional decision
  1537     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
  1538     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
  1539     //
  1540     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
  1541     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
  1542     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
  1543     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
  1544     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
  1545     // In fact, that could have been the original type of p1, and p1 could have
  1546     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
  1547     // expression (LShiftL quux 3) independently optimized to the constant 8.
  1548     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
  1549         && (_type->isa_vect() == NULL)
  1550         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
  1551       // t might actually be lower than _type, if _type is a unique
  1552       // concrete subclass of abstract class t.
  1553       // Make sure the reference is not into the header, by comparing
  1554       // the offset against the offset of the start of the array's data.
  1555       // Different array types begin at slightly different offsets (12 vs. 16).
  1556       // We choose T_BYTE as an example base type that is least restrictive
  1557       // as to alignment, which will therefore produce the smallest
  1558       // possible base offset.
  1559       const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
  1560       if ((uint)off >= (uint)min_base_off) {  // is the offset beyond the header?
  1561         const Type* jt = t->join(_type);
  1562         // In any case, do not allow the join, per se, to empty out the type.
  1563         if (jt->empty() && !t->empty()) {
  1564           // This can happen if a interface-typed array narrows to a class type.
  1565           jt = _type;
  1568         if (EliminateAutoBox && adr->is_AddP()) {
  1569           // The pointers in the autobox arrays are always non-null
  1570           Node* base = adr->in(AddPNode::Base);
  1571           if (base != NULL &&
  1572               !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
  1573             Compile::AliasType* atp = C->alias_type(base->adr_type());
  1574             if (is_autobox_cache(atp)) {
  1575               return jt->join(TypePtr::NOTNULL)->is_ptr();
  1579         return jt;
  1582   } else if (tp->base() == Type::InstPtr) {
  1583     ciEnv* env = C->env();
  1584     const TypeInstPtr* tinst = tp->is_instptr();
  1585     ciKlass* klass = tinst->klass();
  1586     assert( off != Type::OffsetBot ||
  1587             // arrays can be cast to Objects
  1588             tp->is_oopptr()->klass()->is_java_lang_Object() ||
  1589             // unsafe field access may not have a constant offset
  1590             C->has_unsafe_access(),
  1591             "Field accesses must be precise" );
  1592     // For oop loads, we expect the _type to be precise
  1593     if (klass == env->String_klass() &&
  1594         adr->is_AddP() && off != Type::OffsetBot) {
  1595       // For constant Strings treat the final fields as compile time constants.
  1596       Node* base = adr->in(AddPNode::Base);
  1597       const TypeOopPtr* t = phase->type(base)->isa_oopptr();
  1598       if (t != NULL && t->singleton()) {
  1599         ciField* field = env->String_klass()->get_field_by_offset(off, false);
  1600         if (field != NULL && field->is_final()) {
  1601           ciObject* string = t->const_oop();
  1602           ciConstant constant = string->as_instance()->field_value(field);
  1603           if (constant.basic_type() == T_INT) {
  1604             return TypeInt::make(constant.as_int());
  1605           } else if (constant.basic_type() == T_ARRAY) {
  1606             if (adr->bottom_type()->is_ptr_to_narrowoop()) {
  1607               return TypeNarrowOop::make_from_constant(constant.as_object(), true);
  1608             } else {
  1609               return TypeOopPtr::make_from_constant(constant.as_object(), true);
  1615     // Optimizations for constant objects
  1616     ciObject* const_oop = tinst->const_oop();
  1617     if (const_oop != NULL) {
  1618       // For constant CallSites treat the target field as a compile time constant.
  1619       if (const_oop->is_call_site()) {
  1620         ciCallSite* call_site = const_oop->as_call_site();
  1621         ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
  1622         if (field != NULL && field->is_call_site_target()) {
  1623           ciMethodHandle* target = call_site->get_target();
  1624           if (target != NULL) {  // just in case
  1625             ciConstant constant(T_OBJECT, target);
  1626             const Type* t;
  1627             if (adr->bottom_type()->is_ptr_to_narrowoop()) {
  1628               t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
  1629             } else {
  1630               t = TypeOopPtr::make_from_constant(constant.as_object(), true);
  1632             // Add a dependence for invalidation of the optimization.
  1633             if (!call_site->is_constant_call_site()) {
  1634               C->dependencies()->assert_call_site_target_value(call_site, target);
  1636             return t;
  1641   } else if (tp->base() == Type::KlassPtr) {
  1642     assert( off != Type::OffsetBot ||
  1643             // arrays can be cast to Objects
  1644             tp->is_klassptr()->klass()->is_java_lang_Object() ||
  1645             // also allow array-loading from the primary supertype
  1646             // array during subtype checks
  1647             Opcode() == Op_LoadKlass,
  1648             "Field accesses must be precise" );
  1649     // For klass/static loads, we expect the _type to be precise
  1652   const TypeKlassPtr *tkls = tp->isa_klassptr();
  1653   if (tkls != NULL && !StressReflectiveCode) {
  1654     ciKlass* klass = tkls->klass();
  1655     if (klass->is_loaded() && tkls->klass_is_exact()) {
  1656       // We are loading a field from a Klass metaobject whose identity
  1657       // is known at compile time (the type is "exact" or "precise").
  1658       // Check for fields we know are maintained as constants by the VM.
  1659       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
  1660         // The field is Klass::_super_check_offset.  Return its (constant) value.
  1661         // (Folds up type checking code.)
  1662         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
  1663         return TypeInt::make(klass->super_check_offset());
  1665       // Compute index into primary_supers array
  1666       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
  1667       // Check for overflowing; use unsigned compare to handle the negative case.
  1668       if( depth < ciKlass::primary_super_limit() ) {
  1669         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
  1670         // (Folds up type checking code.)
  1671         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
  1672         ciKlass *ss = klass->super_of_depth(depth);
  1673         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
  1675       const Type* aift = load_array_final_field(tkls, klass);
  1676       if (aift != NULL)  return aift;
  1677       if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
  1678           && klass->is_array_klass()) {
  1679         // The field is ArrayKlass::_component_mirror.  Return its (constant) value.
  1680         // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
  1681         assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
  1682         return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
  1684       if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
  1685         // The field is Klass::_java_mirror.  Return its (constant) value.
  1686         // (Folds up the 2nd indirection in anObjConstant.getClass().)
  1687         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
  1688         return TypeInstPtr::make(klass->java_mirror());
  1692     // We can still check if we are loading from the primary_supers array at a
  1693     // shallow enough depth.  Even though the klass is not exact, entries less
  1694     // than or equal to its super depth are correct.
  1695     if (klass->is_loaded() ) {
  1696       ciType *inner = klass;
  1697       while( inner->is_obj_array_klass() )
  1698         inner = inner->as_obj_array_klass()->base_element_type();
  1699       if( inner->is_instance_klass() &&
  1700           !inner->as_instance_klass()->flags().is_interface() ) {
  1701         // Compute index into primary_supers array
  1702         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
  1703         // Check for overflowing; use unsigned compare to handle the negative case.
  1704         if( depth < ciKlass::primary_super_limit() &&
  1705             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
  1706           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
  1707           // (Folds up type checking code.)
  1708           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
  1709           ciKlass *ss = klass->super_of_depth(depth);
  1710           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
  1715     // If the type is enough to determine that the thing is not an array,
  1716     // we can give the layout_helper a positive interval type.
  1717     // This will help short-circuit some reflective code.
  1718     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
  1719         && !klass->is_array_klass() // not directly typed as an array
  1720         && !klass->is_interface()  // specifically not Serializable & Cloneable
  1721         && !klass->is_java_lang_Object()   // not the supertype of all T[]
  1722         ) {
  1723       // Note:  When interfaces are reliable, we can narrow the interface
  1724       // test to (klass != Serializable && klass != Cloneable).
  1725       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
  1726       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
  1727       // The key property of this type is that it folds up tests
  1728       // for array-ness, since it proves that the layout_helper is positive.
  1729       // Thus, a generic value like the basic object layout helper works fine.
  1730       return TypeInt::make(min_size, max_jint, Type::WidenMin);
  1734   // If we are loading from a freshly-allocated object, produce a zero,
  1735   // if the load is provably beyond the header of the object.
  1736   // (Also allow a variable load from a fresh array to produce zero.)
  1737   const TypeOopPtr *tinst = tp->isa_oopptr();
  1738   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
  1739   if (ReduceFieldZeroing || is_instance) {
  1740     Node* value = can_see_stored_value(mem,phase);
  1741     if (value != NULL && value->is_Con()) {
  1742       assert(value->bottom_type()->higher_equal(_type),"sanity");
  1743       return value->bottom_type();
  1747   if (is_instance) {
  1748     // If we have an instance type and our memory input is the
  1749     // programs's initial memory state, there is no matching store,
  1750     // so just return a zero of the appropriate type
  1751     Node *mem = in(MemNode::Memory);
  1752     if (mem->is_Parm() && mem->in(0)->is_Start()) {
  1753       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
  1754       return Type::get_zero_type(_type->basic_type());
  1757   return _type;
  1760 //------------------------------match_edge-------------------------------------
  1761 // Do we Match on this edge index or not?  Match only the address.
  1762 uint LoadNode::match_edge(uint idx) const {
  1763   return idx == MemNode::Address;
  1766 //--------------------------LoadBNode::Ideal--------------------------------------
  1767 //
  1768 //  If the previous store is to the same address as this load,
  1769 //  and the value stored was larger than a byte, replace this load
  1770 //  with the value stored truncated to a byte.  If no truncation is
  1771 //  needed, the replacement is done in LoadNode::Identity().
  1772 //
  1773 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1774   Node* mem = in(MemNode::Memory);
  1775   Node* value = can_see_stored_value(mem,phase);
  1776   if( value && !phase->type(value)->higher_equal( _type ) ) {
  1777     Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
  1778     return new (phase->C) RShiftINode(result, phase->intcon(24));
  1780   // Identity call will handle the case where truncation is not needed.
  1781   return LoadNode::Ideal(phase, can_reshape);
  1784 const Type* LoadBNode::Value(PhaseTransform *phase) const {
  1785   Node* mem = in(MemNode::Memory);
  1786   Node* value = can_see_stored_value(mem,phase);
  1787   if (value != NULL && value->is_Con() &&
  1788       !value->bottom_type()->higher_equal(_type)) {
  1789     // If the input to the store does not fit with the load's result type,
  1790     // it must be truncated. We can't delay until Ideal call since
  1791     // a singleton Value is needed for split_thru_phi optimization.
  1792     int con = value->get_int();
  1793     return TypeInt::make((con << 24) >> 24);
  1795   return LoadNode::Value(phase);
  1798 //--------------------------LoadUBNode::Ideal-------------------------------------
  1799 //
  1800 //  If the previous store is to the same address as this load,
  1801 //  and the value stored was larger than a byte, replace this load
  1802 //  with the value stored truncated to a byte.  If no truncation is
  1803 //  needed, the replacement is done in LoadNode::Identity().
  1804 //
  1805 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
  1806   Node* mem = in(MemNode::Memory);
  1807   Node* value = can_see_stored_value(mem, phase);
  1808   if (value && !phase->type(value)->higher_equal(_type))
  1809     return new (phase->C) AndINode(value, phase->intcon(0xFF));
  1810   // Identity call will handle the case where truncation is not needed.
  1811   return LoadNode::Ideal(phase, can_reshape);
  1814 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
  1815   Node* mem = in(MemNode::Memory);
  1816   Node* value = can_see_stored_value(mem,phase);
  1817   if (value != NULL && value->is_Con() &&
  1818       !value->bottom_type()->higher_equal(_type)) {
  1819     // If the input to the store does not fit with the load's result type,
  1820     // it must be truncated. We can't delay until Ideal call since
  1821     // a singleton Value is needed for split_thru_phi optimization.
  1822     int con = value->get_int();
  1823     return TypeInt::make(con & 0xFF);
  1825   return LoadNode::Value(phase);
  1828 //--------------------------LoadUSNode::Ideal-------------------------------------
  1829 //
  1830 //  If the previous store is to the same address as this load,
  1831 //  and the value stored was larger than a char, replace this load
  1832 //  with the value stored truncated to a char.  If no truncation is
  1833 //  needed, the replacement is done in LoadNode::Identity().
  1834 //
  1835 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1836   Node* mem = in(MemNode::Memory);
  1837   Node* value = can_see_stored_value(mem,phase);
  1838   if( value && !phase->type(value)->higher_equal( _type ) )
  1839     return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
  1840   // Identity call will handle the case where truncation is not needed.
  1841   return LoadNode::Ideal(phase, can_reshape);
  1844 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
  1845   Node* mem = in(MemNode::Memory);
  1846   Node* value = can_see_stored_value(mem,phase);
  1847   if (value != NULL && value->is_Con() &&
  1848       !value->bottom_type()->higher_equal(_type)) {
  1849     // If the input to the store does not fit with the load's result type,
  1850     // it must be truncated. We can't delay until Ideal call since
  1851     // a singleton Value is needed for split_thru_phi optimization.
  1852     int con = value->get_int();
  1853     return TypeInt::make(con & 0xFFFF);
  1855   return LoadNode::Value(phase);
  1858 //--------------------------LoadSNode::Ideal--------------------------------------
  1859 //
  1860 //  If the previous store is to the same address as this load,
  1861 //  and the value stored was larger than a short, replace this load
  1862 //  with the value stored truncated to a short.  If no truncation is
  1863 //  needed, the replacement is done in LoadNode::Identity().
  1864 //
  1865 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  1866   Node* mem = in(MemNode::Memory);
  1867   Node* value = can_see_stored_value(mem,phase);
  1868   if( value && !phase->type(value)->higher_equal( _type ) ) {
  1869     Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
  1870     return new (phase->C) RShiftINode(result, phase->intcon(16));
  1872   // Identity call will handle the case where truncation is not needed.
  1873   return LoadNode::Ideal(phase, can_reshape);
  1876 const Type* LoadSNode::Value(PhaseTransform *phase) const {
  1877   Node* mem = in(MemNode::Memory);
  1878   Node* value = can_see_stored_value(mem,phase);
  1879   if (value != NULL && value->is_Con() &&
  1880       !value->bottom_type()->higher_equal(_type)) {
  1881     // If the input to the store does not fit with the load's result type,
  1882     // it must be truncated. We can't delay until Ideal call since
  1883     // a singleton Value is needed for split_thru_phi optimization.
  1884     int con = value->get_int();
  1885     return TypeInt::make((con << 16) >> 16);
  1887   return LoadNode::Value(phase);
  1890 //=============================================================================
  1891 //----------------------------LoadKlassNode::make------------------------------
  1892 // Polymorphic factory method:
  1893 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
  1894   Compile* C = gvn.C;
  1895   Node *ctl = NULL;
  1896   // sanity check the alias category against the created node type
  1897   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
  1898   assert(adr_type != NULL, "expecting TypeKlassPtr");
  1899 #ifdef _LP64
  1900   if (adr_type->is_ptr_to_narrowklass()) {
  1901     assert(UseCompressedKlassPointers, "no compressed klasses");
  1902     Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass()));
  1903     return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
  1905 #endif
  1906   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
  1907   return new (C) LoadKlassNode(ctl, mem, adr, at, tk);
  1910 //------------------------------Value------------------------------------------
  1911 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
  1912   return klass_value_common(phase);
  1915 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
  1916   // Either input is TOP ==> the result is TOP
  1917   const Type *t1 = phase->type( in(MemNode::Memory) );
  1918   if (t1 == Type::TOP)  return Type::TOP;
  1919   Node *adr = in(MemNode::Address);
  1920   const Type *t2 = phase->type( adr );
  1921   if (t2 == Type::TOP)  return Type::TOP;
  1922   const TypePtr *tp = t2->is_ptr();
  1923   if (TypePtr::above_centerline(tp->ptr()) ||
  1924       tp->ptr() == TypePtr::Null)  return Type::TOP;
  1926   // Return a more precise klass, if possible
  1927   const TypeInstPtr *tinst = tp->isa_instptr();
  1928   if (tinst != NULL) {
  1929     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
  1930     int offset = tinst->offset();
  1931     if (ik == phase->C->env()->Class_klass()
  1932         && (offset == java_lang_Class::klass_offset_in_bytes() ||
  1933             offset == java_lang_Class::array_klass_offset_in_bytes())) {
  1934       // We are loading a special hidden field from a Class mirror object,
  1935       // the field which points to the VM's Klass metaobject.
  1936       ciType* t = tinst->java_mirror_type();
  1937       // java_mirror_type returns non-null for compile-time Class constants.
  1938       if (t != NULL) {
  1939         // constant oop => constant klass
  1940         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
  1941           return TypeKlassPtr::make(ciArrayKlass::make(t));
  1943         if (!t->is_klass()) {
  1944           // a primitive Class (e.g., int.class) has NULL for a klass field
  1945           return TypePtr::NULL_PTR;
  1947         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
  1948         return TypeKlassPtr::make(t->as_klass());
  1950       // non-constant mirror, so we can't tell what's going on
  1952     if( !ik->is_loaded() )
  1953       return _type;             // Bail out if not loaded
  1954     if (offset == oopDesc::klass_offset_in_bytes()) {
  1955       if (tinst->klass_is_exact()) {
  1956         return TypeKlassPtr::make(ik);
  1958       // See if we can become precise: no subklasses and no interface
  1959       // (Note:  We need to support verified interfaces.)
  1960       if (!ik->is_interface() && !ik->has_subklass()) {
  1961         //assert(!UseExactTypes, "this code should be useless with exact types");
  1962         // Add a dependence; if any subclass added we need to recompile
  1963         if (!ik->is_final()) {
  1964           // %%% should use stronger assert_unique_concrete_subtype instead
  1965           phase->C->dependencies()->assert_leaf_type(ik);
  1967         // Return precise klass
  1968         return TypeKlassPtr::make(ik);
  1971       // Return root of possible klass
  1972       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
  1976   // Check for loading klass from an array
  1977   const TypeAryPtr *tary = tp->isa_aryptr();
  1978   if( tary != NULL ) {
  1979     ciKlass *tary_klass = tary->klass();
  1980     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
  1981         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
  1982       if (tary->klass_is_exact()) {
  1983         return TypeKlassPtr::make(tary_klass);
  1985       ciArrayKlass *ak = tary->klass()->as_array_klass();
  1986       // If the klass is an object array, we defer the question to the
  1987       // array component klass.
  1988       if( ak->is_obj_array_klass() ) {
  1989         assert( ak->is_loaded(), "" );
  1990         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
  1991         if( base_k->is_loaded() && base_k->is_instance_klass() ) {
  1992           ciInstanceKlass* ik = base_k->as_instance_klass();
  1993           // See if we can become precise: no subklasses and no interface
  1994           if (!ik->is_interface() && !ik->has_subklass()) {
  1995             //assert(!UseExactTypes, "this code should be useless with exact types");
  1996             // Add a dependence; if any subclass added we need to recompile
  1997             if (!ik->is_final()) {
  1998               phase->C->dependencies()->assert_leaf_type(ik);
  2000             // Return precise array klass
  2001             return TypeKlassPtr::make(ak);
  2004         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
  2005       } else {                  // Found a type-array?
  2006         //assert(!UseExactTypes, "this code should be useless with exact types");
  2007         assert( ak->is_type_array_klass(), "" );
  2008         return TypeKlassPtr::make(ak); // These are always precise
  2013   // Check for loading klass from an array klass
  2014   const TypeKlassPtr *tkls = tp->isa_klassptr();
  2015   if (tkls != NULL && !StressReflectiveCode) {
  2016     ciKlass* klass = tkls->klass();
  2017     if( !klass->is_loaded() )
  2018       return _type;             // Bail out if not loaded
  2019     if( klass->is_obj_array_klass() &&
  2020         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
  2021       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
  2022       // // Always returning precise element type is incorrect,
  2023       // // e.g., element type could be object and array may contain strings
  2024       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
  2026       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
  2027       // according to the element type's subclassing.
  2028       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
  2030     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
  2031         tkls->offset() == in_bytes(Klass::super_offset())) {
  2032       ciKlass* sup = klass->as_instance_klass()->super();
  2033       // The field is Klass::_super.  Return its (constant) value.
  2034       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
  2035       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
  2039   // Bailout case
  2040   return LoadNode::Value(phase);
  2043 //------------------------------Identity---------------------------------------
  2044 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
  2045 // Also feed through the klass in Allocate(...klass...)._klass.
  2046 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
  2047   return klass_identity_common(phase);
  2050 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
  2051   Node* x = LoadNode::Identity(phase);
  2052   if (x != this)  return x;
  2054   // Take apart the address into an oop and and offset.
  2055   // Return 'this' if we cannot.
  2056   Node*    adr    = in(MemNode::Address);
  2057   intptr_t offset = 0;
  2058   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
  2059   if (base == NULL)     return this;
  2060   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
  2061   if (toop == NULL)     return this;
  2063   // We can fetch the klass directly through an AllocateNode.
  2064   // This works even if the klass is not constant (clone or newArray).
  2065   if (offset == oopDesc::klass_offset_in_bytes()) {
  2066     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
  2067     if (allocated_klass != NULL) {
  2068       return allocated_klass;
  2072   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
  2073   // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
  2074   // See inline_native_Class_query for occurrences of these patterns.
  2075   // Java Example:  x.getClass().isAssignableFrom(y)
  2076   // Java Example:  Array.newInstance(x.getClass().getComponentType(), n)
  2077   //
  2078   // This improves reflective code, often making the Class
  2079   // mirror go completely dead.  (Current exception:  Class
  2080   // mirrors may appear in debug info, but we could clean them out by
  2081   // introducing a new debug info operator for Klass*.java_mirror).
  2082   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
  2083       && (offset == java_lang_Class::klass_offset_in_bytes() ||
  2084           offset == java_lang_Class::array_klass_offset_in_bytes())) {
  2085     // We are loading a special hidden field from a Class mirror,
  2086     // the field which points to its Klass or ArrayKlass metaobject.
  2087     if (base->is_Load()) {
  2088       Node* adr2 = base->in(MemNode::Address);
  2089       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
  2090       if (tkls != NULL && !tkls->empty()
  2091           && (tkls->klass()->is_instance_klass() ||
  2092               tkls->klass()->is_array_klass())
  2093           && adr2->is_AddP()
  2094           ) {
  2095         int mirror_field = in_bytes(Klass::java_mirror_offset());
  2096         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
  2097           mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
  2099         if (tkls->offset() == mirror_field) {
  2100           return adr2->in(AddPNode::Base);
  2106   return this;
  2110 //------------------------------Value------------------------------------------
  2111 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
  2112   const Type *t = klass_value_common(phase);
  2113   if (t == Type::TOP)
  2114     return t;
  2116   return t->make_narrowklass();
  2119 //------------------------------Identity---------------------------------------
  2120 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
  2121 // Also feed through the klass in Allocate(...klass...)._klass.
  2122 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
  2123   Node *x = klass_identity_common(phase);
  2125   const Type *t = phase->type( x );
  2126   if( t == Type::TOP ) return x;
  2127   if( t->isa_narrowklass()) return x;
  2128   assert (!t->isa_narrowoop(), "no narrow oop here");
  2130   return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
  2133 //------------------------------Value-----------------------------------------
  2134 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
  2135   // Either input is TOP ==> the result is TOP
  2136   const Type *t1 = phase->type( in(MemNode::Memory) );
  2137   if( t1 == Type::TOP ) return Type::TOP;
  2138   Node *adr = in(MemNode::Address);
  2139   const Type *t2 = phase->type( adr );
  2140   if( t2 == Type::TOP ) return Type::TOP;
  2141   const TypePtr *tp = t2->is_ptr();
  2142   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
  2143   const TypeAryPtr *tap = tp->isa_aryptr();
  2144   if( !tap ) return _type;
  2145   return tap->size();
  2148 //-------------------------------Ideal---------------------------------------
  2149 // Feed through the length in AllocateArray(...length...)._length.
  2150 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  2151   Node* p = MemNode::Ideal_common(phase, can_reshape);
  2152   if (p)  return (p == NodeSentinel) ? NULL : p;
  2154   // Take apart the address into an oop and and offset.
  2155   // Return 'this' if we cannot.
  2156   Node*    adr    = in(MemNode::Address);
  2157   intptr_t offset = 0;
  2158   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
  2159   if (base == NULL)     return NULL;
  2160   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
  2161   if (tary == NULL)     return NULL;
  2163   // We can fetch the length directly through an AllocateArrayNode.
  2164   // This works even if the length is not constant (clone or newArray).
  2165   if (offset == arrayOopDesc::length_offset_in_bytes()) {
  2166     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
  2167     if (alloc != NULL) {
  2168       Node* allocated_length = alloc->Ideal_length();
  2169       Node* len = alloc->make_ideal_length(tary, phase);
  2170       if (allocated_length != len) {
  2171         // New CastII improves on this.
  2172         return len;
  2177   return NULL;
  2180 //------------------------------Identity---------------------------------------
  2181 // Feed through the length in AllocateArray(...length...)._length.
  2182 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
  2183   Node* x = LoadINode::Identity(phase);
  2184   if (x != this)  return x;
  2186   // Take apart the address into an oop and and offset.
  2187   // Return 'this' if we cannot.
  2188   Node*    adr    = in(MemNode::Address);
  2189   intptr_t offset = 0;
  2190   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
  2191   if (base == NULL)     return this;
  2192   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
  2193   if (tary == NULL)     return this;
  2195   // We can fetch the length directly through an AllocateArrayNode.
  2196   // This works even if the length is not constant (clone or newArray).
  2197   if (offset == arrayOopDesc::length_offset_in_bytes()) {
  2198     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
  2199     if (alloc != NULL) {
  2200       Node* allocated_length = alloc->Ideal_length();
  2201       // Do not allow make_ideal_length to allocate a CastII node.
  2202       Node* len = alloc->make_ideal_length(tary, phase, false);
  2203       if (allocated_length == len) {
  2204         // Return allocated_length only if it would not be improved by a CastII.
  2205         return allocated_length;
  2210   return this;
  2214 //=============================================================================
  2215 //---------------------------StoreNode::make-----------------------------------
  2216 // Polymorphic factory method:
  2217 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
  2218   Compile* C = gvn.C;
  2219   assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
  2220           ctl != NULL, "raw memory operations should have control edge");
  2222   switch (bt) {
  2223   case T_BOOLEAN:
  2224   case T_BYTE:    return new (C) StoreBNode(ctl, mem, adr, adr_type, val);
  2225   case T_INT:     return new (C) StoreINode(ctl, mem, adr, adr_type, val);
  2226   case T_CHAR:
  2227   case T_SHORT:   return new (C) StoreCNode(ctl, mem, adr, adr_type, val);
  2228   case T_LONG:    return new (C) StoreLNode(ctl, mem, adr, adr_type, val);
  2229   case T_FLOAT:   return new (C) StoreFNode(ctl, mem, adr, adr_type, val);
  2230   case T_DOUBLE:  return new (C) StoreDNode(ctl, mem, adr, adr_type, val);
  2231   case T_METADATA:
  2232   case T_ADDRESS:
  2233   case T_OBJECT:
  2234 #ifdef _LP64
  2235     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
  2236       val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
  2237       return new (C) StoreNNode(ctl, mem, adr, adr_type, val);
  2238     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
  2239                (UseCompressedKlassPointers && val->bottom_type()->isa_klassptr() &&
  2240                 adr->bottom_type()->isa_rawptr())) {
  2241       val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
  2242       return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val);
  2244 #endif
  2246       return new (C) StorePNode(ctl, mem, adr, adr_type, val);
  2249   ShouldNotReachHere();
  2250   return (StoreNode*)NULL;
  2253 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
  2254   bool require_atomic = true;
  2255   return new (C) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
  2259 //--------------------------bottom_type----------------------------------------
  2260 const Type *StoreNode::bottom_type() const {
  2261   return Type::MEMORY;
  2264 //------------------------------hash-------------------------------------------
  2265 uint StoreNode::hash() const {
  2266   // unroll addition of interesting fields
  2267   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
  2269   // Since they are not commoned, do not hash them:
  2270   return NO_HASH;
  2273 //------------------------------Ideal------------------------------------------
  2274 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
  2275 // When a store immediately follows a relevant allocation/initialization,
  2276 // try to capture it into the initialization, or hoist it above.
  2277 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  2278   Node* p = MemNode::Ideal_common(phase, can_reshape);
  2279   if (p)  return (p == NodeSentinel) ? NULL : p;
  2281   Node* mem     = in(MemNode::Memory);
  2282   Node* address = in(MemNode::Address);
  2284   // Back-to-back stores to same address?  Fold em up.  Generally
  2285   // unsafe if I have intervening uses...  Also disallowed for StoreCM
  2286   // since they must follow each StoreP operation.  Redundant StoreCMs
  2287   // are eliminated just before matching in final_graph_reshape.
  2288   if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
  2289       mem->Opcode() != Op_StoreCM) {
  2290     // Looking at a dead closed cycle of memory?
  2291     assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
  2293     assert(Opcode() == mem->Opcode() ||
  2294            phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
  2295            "no mismatched stores, except on raw memory");
  2297     if (mem->outcnt() == 1 &&           // check for intervening uses
  2298         mem->as_Store()->memory_size() <= this->memory_size()) {
  2299       // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
  2300       // For example, 'mem' might be the final state at a conditional return.
  2301       // Or, 'mem' might be used by some node which is live at the same time
  2302       // 'this' is live, which might be unschedulable.  So, require exactly
  2303       // ONE user, the 'this' store, until such time as we clone 'mem' for
  2304       // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
  2305       if (can_reshape) {  // (%%% is this an anachronism?)
  2306         set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
  2307                   phase->is_IterGVN());
  2308       } else {
  2309         // It's OK to do this in the parser, since DU info is always accurate,
  2310         // and the parser always refers to nodes via SafePointNode maps.
  2311         set_req(MemNode::Memory, mem->in(MemNode::Memory));
  2313       return this;
  2317   // Capture an unaliased, unconditional, simple store into an initializer.
  2318   // Or, if it is independent of the allocation, hoist it above the allocation.
  2319   if (ReduceFieldZeroing && /*can_reshape &&*/
  2320       mem->is_Proj() && mem->in(0)->is_Initialize()) {
  2321     InitializeNode* init = mem->in(0)->as_Initialize();
  2322     intptr_t offset = init->can_capture_store(this, phase);
  2323     if (offset > 0) {
  2324       Node* moved = init->capture_store(this, offset, phase);
  2325       // If the InitializeNode captured me, it made a raw copy of me,
  2326       // and I need to disappear.
  2327       if (moved != NULL) {
  2328         // %%% hack to ensure that Ideal returns a new node:
  2329         mem = MergeMemNode::make(phase->C, mem);
  2330         return mem;             // fold me away
  2335   return NULL;                  // No further progress
  2338 //------------------------------Value-----------------------------------------
  2339 const Type *StoreNode::Value( PhaseTransform *phase ) const {
  2340   // Either input is TOP ==> the result is TOP
  2341   const Type *t1 = phase->type( in(MemNode::Memory) );
  2342   if( t1 == Type::TOP ) return Type::TOP;
  2343   const Type *t2 = phase->type( in(MemNode::Address) );
  2344   if( t2 == Type::TOP ) return Type::TOP;
  2345   const Type *t3 = phase->type( in(MemNode::ValueIn) );
  2346   if( t3 == Type::TOP ) return Type::TOP;
  2347   return Type::MEMORY;
  2350 //------------------------------Identity---------------------------------------
  2351 // Remove redundant stores:
  2352 //   Store(m, p, Load(m, p)) changes to m.
  2353 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
  2354 Node *StoreNode::Identity( PhaseTransform *phase ) {
  2355   Node* mem = in(MemNode::Memory);
  2356   Node* adr = in(MemNode::Address);
  2357   Node* val = in(MemNode::ValueIn);
  2359   // Load then Store?  Then the Store is useless
  2360   if (val->is_Load() &&
  2361       val->in(MemNode::Address)->eqv_uncast(adr) &&
  2362       val->in(MemNode::Memory )->eqv_uncast(mem) &&
  2363       val->as_Load()->store_Opcode() == Opcode()) {
  2364     return mem;
  2367   // Two stores in a row of the same value?
  2368   if (mem->is_Store() &&
  2369       mem->in(MemNode::Address)->eqv_uncast(adr) &&
  2370       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
  2371       mem->Opcode() == Opcode()) {
  2372     return mem;
  2375   // Store of zero anywhere into a freshly-allocated object?
  2376   // Then the store is useless.
  2377   // (It must already have been captured by the InitializeNode.)
  2378   if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
  2379     // a newly allocated object is already all-zeroes everywhere
  2380     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
  2381       return mem;
  2384     // the store may also apply to zero-bits in an earlier object
  2385     Node* prev_mem = find_previous_store(phase);
  2386     // Steps (a), (b):  Walk past independent stores to find an exact match.
  2387     if (prev_mem != NULL) {
  2388       Node* prev_val = can_see_stored_value(prev_mem, phase);
  2389       if (prev_val != NULL && phase->eqv(prev_val, val)) {
  2390         // prev_val and val might differ by a cast; it would be good
  2391         // to keep the more informative of the two.
  2392         return mem;
  2397   return this;
  2400 //------------------------------match_edge-------------------------------------
  2401 // Do we Match on this edge index or not?  Match only memory & value
  2402 uint StoreNode::match_edge(uint idx) const {
  2403   return idx == MemNode::Address || idx == MemNode::ValueIn;
  2406 //------------------------------cmp--------------------------------------------
  2407 // Do not common stores up together.  They generally have to be split
  2408 // back up anyways, so do not bother.
  2409 uint StoreNode::cmp( const Node &n ) const {
  2410   return (&n == this);          // Always fail except on self
  2413 //------------------------------Ideal_masked_input-----------------------------
  2414 // Check for a useless mask before a partial-word store
  2415 // (StoreB ... (AndI valIn conIa) )
  2416 // If (conIa & mask == mask) this simplifies to
  2417 // (StoreB ... (valIn) )
  2418 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
  2419   Node *val = in(MemNode::ValueIn);
  2420   if( val->Opcode() == Op_AndI ) {
  2421     const TypeInt *t = phase->type( val->in(2) )->isa_int();
  2422     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
  2423       set_req(MemNode::ValueIn, val->in(1));
  2424       return this;
  2427   return NULL;
  2431 //------------------------------Ideal_sign_extended_input----------------------
  2432 // Check for useless sign-extension before a partial-word store
  2433 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
  2434 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
  2435 // (StoreB ... (valIn) )
  2436 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
  2437   Node *val = in(MemNode::ValueIn);
  2438   if( val->Opcode() == Op_RShiftI ) {
  2439     const TypeInt *t = phase->type( val->in(2) )->isa_int();
  2440     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
  2441       Node *shl = val->in(1);
  2442       if( shl->Opcode() == Op_LShiftI ) {
  2443         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
  2444         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
  2445           set_req(MemNode::ValueIn, shl->in(1));
  2446           return this;
  2451   return NULL;
  2454 //------------------------------value_never_loaded-----------------------------------
  2455 // Determine whether there are any possible loads of the value stored.
  2456 // For simplicity, we actually check if there are any loads from the
  2457 // address stored to, not just for loads of the value stored by this node.
  2458 //
  2459 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
  2460   Node *adr = in(Address);
  2461   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
  2462   if (adr_oop == NULL)
  2463     return false;
  2464   if (!adr_oop->is_known_instance_field())
  2465     return false; // if not a distinct instance, there may be aliases of the address
  2466   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
  2467     Node *use = adr->fast_out(i);
  2468     int opc = use->Opcode();
  2469     if (use->is_Load() || use->is_LoadStore()) {
  2470       return false;
  2473   return true;
  2476 //=============================================================================
  2477 //------------------------------Ideal------------------------------------------
  2478 // If the store is from an AND mask that leaves the low bits untouched, then
  2479 // we can skip the AND operation.  If the store is from a sign-extension
  2480 // (a left shift, then right shift) we can skip both.
  2481 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2482   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
  2483   if( progress != NULL ) return progress;
  2485   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
  2486   if( progress != NULL ) return progress;
  2488   // Finally check the default case
  2489   return StoreNode::Ideal(phase, can_reshape);
  2492 //=============================================================================
  2493 //------------------------------Ideal------------------------------------------
  2494 // If the store is from an AND mask that leaves the low bits untouched, then
  2495 // we can skip the AND operation
  2496 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2497   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
  2498   if( progress != NULL ) return progress;
  2500   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
  2501   if( progress != NULL ) return progress;
  2503   // Finally check the default case
  2504   return StoreNode::Ideal(phase, can_reshape);
  2507 //=============================================================================
  2508 //------------------------------Identity---------------------------------------
  2509 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
  2510   // No need to card mark when storing a null ptr
  2511   Node* my_store = in(MemNode::OopStore);
  2512   if (my_store->is_Store()) {
  2513     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
  2514     if( t1 == TypePtr::NULL_PTR ) {
  2515       return in(MemNode::Memory);
  2518   return this;
  2521 //=============================================================================
  2522 //------------------------------Ideal---------------------------------------
  2523 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2524   Node* progress = StoreNode::Ideal(phase, can_reshape);
  2525   if (progress != NULL) return progress;
  2527   Node* my_store = in(MemNode::OopStore);
  2528   if (my_store->is_MergeMem()) {
  2529     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
  2530     set_req(MemNode::OopStore, mem);
  2531     return this;
  2534   return NULL;
  2537 //------------------------------Value-----------------------------------------
  2538 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
  2539   // Either input is TOP ==> the result is TOP
  2540   const Type *t = phase->type( in(MemNode::Memory) );
  2541   if( t == Type::TOP ) return Type::TOP;
  2542   t = phase->type( in(MemNode::Address) );
  2543   if( t == Type::TOP ) return Type::TOP;
  2544   t = phase->type( in(MemNode::ValueIn) );
  2545   if( t == Type::TOP ) return Type::TOP;
  2546   // If extra input is TOP ==> the result is TOP
  2547   t = phase->type( in(MemNode::OopStore) );
  2548   if( t == Type::TOP ) return Type::TOP;
  2550   return StoreNode::Value( phase );
  2554 //=============================================================================
  2555 //----------------------------------SCMemProjNode------------------------------
  2556 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
  2558   return bottom_type();
  2561 //=============================================================================
  2562 //----------------------------------LoadStoreNode------------------------------
  2563 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
  2564   : Node(required),
  2565     _type(rt),
  2566     _adr_type(at)
  2568   init_req(MemNode::Control, c  );
  2569   init_req(MemNode::Memory , mem);
  2570   init_req(MemNode::Address, adr);
  2571   init_req(MemNode::ValueIn, val);
  2572   init_class_id(Class_LoadStore);
  2575 uint LoadStoreNode::ideal_reg() const {
  2576   return _type->ideal_reg();
  2579 bool LoadStoreNode::result_not_used() const {
  2580   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
  2581     Node *x = fast_out(i);
  2582     if (x->Opcode() == Op_SCMemProj) continue;
  2583     return false;
  2585   return true;
  2588 uint LoadStoreNode::size_of() const { return sizeof(*this); }
  2590 //=============================================================================
  2591 //----------------------------------LoadStoreConditionalNode--------------------
  2592 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
  2593   init_req(ExpectedIn, ex );
  2596 //=============================================================================
  2597 //-------------------------------adr_type--------------------------------------
  2598 // Do we Match on this edge index or not?  Do not match memory
  2599 const TypePtr* ClearArrayNode::adr_type() const {
  2600   Node *adr = in(3);
  2601   return MemNode::calculate_adr_type(adr->bottom_type());
  2604 //------------------------------match_edge-------------------------------------
  2605 // Do we Match on this edge index or not?  Do not match memory
  2606 uint ClearArrayNode::match_edge(uint idx) const {
  2607   return idx > 1;
  2610 //------------------------------Identity---------------------------------------
  2611 // Clearing a zero length array does nothing
  2612 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
  2613   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
  2616 //------------------------------Idealize---------------------------------------
  2617 // Clearing a short array is faster with stores
  2618 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
  2619   const int unit = BytesPerLong;
  2620   const TypeX* t = phase->type(in(2))->isa_intptr_t();
  2621   if (!t)  return NULL;
  2622   if (!t->is_con())  return NULL;
  2623   intptr_t raw_count = t->get_con();
  2624   intptr_t size = raw_count;
  2625   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
  2626   // Clearing nothing uses the Identity call.
  2627   // Negative clears are possible on dead ClearArrays
  2628   // (see jck test stmt114.stmt11402.val).
  2629   if (size <= 0 || size % unit != 0)  return NULL;
  2630   intptr_t count = size / unit;
  2631   // Length too long; use fast hardware clear
  2632   if (size > Matcher::init_array_short_size)  return NULL;
  2633   Node *mem = in(1);
  2634   if( phase->type(mem)==Type::TOP ) return NULL;
  2635   Node *adr = in(3);
  2636   const Type* at = phase->type(adr);
  2637   if( at==Type::TOP ) return NULL;
  2638   const TypePtr* atp = at->isa_ptr();
  2639   // adjust atp to be the correct array element address type
  2640   if (atp == NULL)  atp = TypePtr::BOTTOM;
  2641   else              atp = atp->add_offset(Type::OffsetBot);
  2642   // Get base for derived pointer purposes
  2643   if( adr->Opcode() != Op_AddP ) Unimplemented();
  2644   Node *base = adr->in(1);
  2646   Node *zero = phase->makecon(TypeLong::ZERO);
  2647   Node *off  = phase->MakeConX(BytesPerLong);
  2648   mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero);
  2649   count--;
  2650   while( count-- ) {
  2651     mem = phase->transform(mem);
  2652     adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
  2653     mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero);
  2655   return mem;
  2658 //----------------------------step_through----------------------------------
  2659 // Return allocation input memory edge if it is different instance
  2660 // or itself if it is the one we are looking for.
  2661 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
  2662   Node* n = *np;
  2663   assert(n->is_ClearArray(), "sanity");
  2664   intptr_t offset;
  2665   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
  2666   // This method is called only before Allocate nodes are expanded during
  2667   // macro nodes expansion. Before that ClearArray nodes are only generated
  2668   // in LibraryCallKit::generate_arraycopy() which follows allocations.
  2669   assert(alloc != NULL, "should have allocation");
  2670   if (alloc->_idx == instance_id) {
  2671     // Can not bypass initialization of the instance we are looking for.
  2672     return false;
  2674   // Otherwise skip it.
  2675   InitializeNode* init = alloc->initialization();
  2676   if (init != NULL)
  2677     *np = init->in(TypeFunc::Memory);
  2678   else
  2679     *np = alloc->in(TypeFunc::Memory);
  2680   return true;
  2683 //----------------------------clear_memory-------------------------------------
  2684 // Generate code to initialize object storage to zero.
  2685 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
  2686                                    intptr_t start_offset,
  2687                                    Node* end_offset,
  2688                                    PhaseGVN* phase) {
  2689   Compile* C = phase->C;
  2690   intptr_t offset = start_offset;
  2692   int unit = BytesPerLong;
  2693   if ((offset % unit) != 0) {
  2694     Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
  2695     adr = phase->transform(adr);
  2696     const TypePtr* atp = TypeRawPtr::BOTTOM;
  2697     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
  2698     mem = phase->transform(mem);
  2699     offset += BytesPerInt;
  2701   assert((offset % unit) == 0, "");
  2703   // Initialize the remaining stuff, if any, with a ClearArray.
  2704   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
  2707 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
  2708                                    Node* start_offset,
  2709                                    Node* end_offset,
  2710                                    PhaseGVN* phase) {
  2711   if (start_offset == end_offset) {
  2712     // nothing to do
  2713     return mem;
  2716   Compile* C = phase->C;
  2717   int unit = BytesPerLong;
  2718   Node* zbase = start_offset;
  2719   Node* zend  = end_offset;
  2721   // Scale to the unit required by the CPU:
  2722   if (!Matcher::init_array_count_is_in_bytes) {
  2723     Node* shift = phase->intcon(exact_log2(unit));
  2724     zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
  2725     zend  = phase->transform( new(C) URShiftXNode(zend,  shift) );
  2728   // Bulk clear double-words
  2729   Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
  2730   Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
  2731   mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
  2732   return phase->transform(mem);
  2735 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
  2736                                    intptr_t start_offset,
  2737                                    intptr_t end_offset,
  2738                                    PhaseGVN* phase) {
  2739   if (start_offset == end_offset) {
  2740     // nothing to do
  2741     return mem;
  2744   Compile* C = phase->C;
  2745   assert((end_offset % BytesPerInt) == 0, "odd end offset");
  2746   intptr_t done_offset = end_offset;
  2747   if ((done_offset % BytesPerLong) != 0) {
  2748     done_offset -= BytesPerInt;
  2750   if (done_offset > start_offset) {
  2751     mem = clear_memory(ctl, mem, dest,
  2752                        start_offset, phase->MakeConX(done_offset), phase);
  2754   if (done_offset < end_offset) { // emit the final 32-bit store
  2755     Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
  2756     adr = phase->transform(adr);
  2757     const TypePtr* atp = TypeRawPtr::BOTTOM;
  2758     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
  2759     mem = phase->transform(mem);
  2760     done_offset += BytesPerInt;
  2762   assert(done_offset == end_offset, "");
  2763   return mem;
  2766 //=============================================================================
  2767 // Do not match memory edge.
  2768 uint StrIntrinsicNode::match_edge(uint idx) const {
  2769   return idx == 2 || idx == 3;
  2772 //------------------------------Ideal------------------------------------------
  2773 // Return a node which is more "ideal" than the current node.  Strip out
  2774 // control copies
  2775 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  2776   if (remove_dead_region(phase, can_reshape)) return this;
  2777   // Don't bother trying to transform a dead node
  2778   if (in(0) && in(0)->is_top())  return NULL;
  2780   if (can_reshape) {
  2781     Node* mem = phase->transform(in(MemNode::Memory));
  2782     // If transformed to a MergeMem, get the desired slice
  2783     uint alias_idx = phase->C->get_alias_index(adr_type());
  2784     mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
  2785     if (mem != in(MemNode::Memory)) {
  2786       set_req(MemNode::Memory, mem);
  2787       return this;
  2790   return NULL;
  2793 //------------------------------Value------------------------------------------
  2794 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
  2795   if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
  2796   return bottom_type();
  2799 //=============================================================================
  2800 //------------------------------match_edge-------------------------------------
  2801 // Do not match memory edge
  2802 uint EncodeISOArrayNode::match_edge(uint idx) const {
  2803   return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len)
  2806 //------------------------------Ideal------------------------------------------
  2807 // Return a node which is more "ideal" than the current node.  Strip out
  2808 // control copies
  2809 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  2810   return remove_dead_region(phase, can_reshape) ? this : NULL;
  2813 //------------------------------Value------------------------------------------
  2814 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const {
  2815   if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
  2816   return bottom_type();
  2819 //=============================================================================
  2820 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
  2821   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
  2822     _adr_type(C->get_adr_type(alias_idx))
  2824   init_class_id(Class_MemBar);
  2825   Node* top = C->top();
  2826   init_req(TypeFunc::I_O,top);
  2827   init_req(TypeFunc::FramePtr,top);
  2828   init_req(TypeFunc::ReturnAdr,top);
  2829   if (precedent != NULL)
  2830     init_req(TypeFunc::Parms, precedent);
  2833 //------------------------------cmp--------------------------------------------
  2834 uint MemBarNode::hash() const { return NO_HASH; }
  2835 uint MemBarNode::cmp( const Node &n ) const {
  2836   return (&n == this);          // Always fail except on self
  2839 //------------------------------make-------------------------------------------
  2840 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
  2841   switch (opcode) {
  2842   case Op_MemBarAcquire:   return new(C) MemBarAcquireNode(C,  atp, pn);
  2843   case Op_MemBarRelease:   return new(C) MemBarReleaseNode(C,  atp, pn);
  2844   case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C,  atp, pn);
  2845   case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C,  atp, pn);
  2846   case Op_MemBarVolatile:  return new(C) MemBarVolatileNode(C, atp, pn);
  2847   case Op_MemBarCPUOrder:  return new(C) MemBarCPUOrderNode(C, atp, pn);
  2848   case Op_Initialize:      return new(C) InitializeNode(C,     atp, pn);
  2849   case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C,  atp, pn);
  2850   default:                 ShouldNotReachHere(); return NULL;
  2854 //------------------------------Ideal------------------------------------------
  2855 // Return a node which is more "ideal" than the current node.  Strip out
  2856 // control copies
  2857 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  2858   if (remove_dead_region(phase, can_reshape)) return this;
  2859   // Don't bother trying to transform a dead node
  2860   if (in(0) && in(0)->is_top())  return NULL;
  2862   // Eliminate volatile MemBars for scalar replaced objects.
  2863   if (can_reshape && req() == (Precedent+1) &&
  2864       (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) {
  2865     // Volatile field loads and stores.
  2866     Node* my_mem = in(MemBarNode::Precedent);
  2867     if (my_mem != NULL && my_mem->is_Mem()) {
  2868       const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
  2869       // Check for scalar replaced object reference.
  2870       if( t_oop != NULL && t_oop->is_known_instance_field() &&
  2871           t_oop->offset() != Type::OffsetBot &&
  2872           t_oop->offset() != Type::OffsetTop) {
  2873         // Replace MemBar projections by its inputs.
  2874         PhaseIterGVN* igvn = phase->is_IterGVN();
  2875         igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
  2876         igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
  2877         // Must return either the original node (now dead) or a new node
  2878         // (Do not return a top here, since that would break the uniqueness of top.)
  2879         return new (phase->C) ConINode(TypeInt::ZERO);
  2883   return NULL;
  2886 //------------------------------Value------------------------------------------
  2887 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
  2888   if( !in(0) ) return Type::TOP;
  2889   if( phase->type(in(0)) == Type::TOP )
  2890     return Type::TOP;
  2891   return TypeTuple::MEMBAR;
  2894 //------------------------------match------------------------------------------
  2895 // Construct projections for memory.
  2896 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
  2897   switch (proj->_con) {
  2898   case TypeFunc::Control:
  2899   case TypeFunc::Memory:
  2900     return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
  2902   ShouldNotReachHere();
  2903   return NULL;
  2906 //===========================InitializeNode====================================
  2907 // SUMMARY:
  2908 // This node acts as a memory barrier on raw memory, after some raw stores.
  2909 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
  2910 // The Initialize can 'capture' suitably constrained stores as raw inits.
  2911 // It can coalesce related raw stores into larger units (called 'tiles').
  2912 // It can avoid zeroing new storage for memory units which have raw inits.
  2913 // At macro-expansion, it is marked 'complete', and does not optimize further.
  2914 //
  2915 // EXAMPLE:
  2916 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
  2917 //   ctl = incoming control; mem* = incoming memory
  2918 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
  2919 // First allocate uninitialized memory and fill in the header:
  2920 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
  2921 //   ctl := alloc.Control; mem* := alloc.Memory*
  2922 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
  2923 // Then initialize to zero the non-header parts of the raw memory block:
  2924 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
  2925 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
  2926 // After the initialize node executes, the object is ready for service:
  2927 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
  2928 // Suppose its body is immediately initialized as {1,2}:
  2929 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
  2930 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
  2931 //   mem.SLICE(#short[*]) := store2
  2932 //
  2933 // DETAILS:
  2934 // An InitializeNode collects and isolates object initialization after
  2935 // an AllocateNode and before the next possible safepoint.  As a
  2936 // memory barrier (MemBarNode), it keeps critical stores from drifting
  2937 // down past any safepoint or any publication of the allocation.
  2938 // Before this barrier, a newly-allocated object may have uninitialized bits.
  2939 // After this barrier, it may be treated as a real oop, and GC is allowed.
  2940 //
  2941 // The semantics of the InitializeNode include an implicit zeroing of
  2942 // the new object from object header to the end of the object.
  2943 // (The object header and end are determined by the AllocateNode.)
  2944 //
  2945 // Certain stores may be added as direct inputs to the InitializeNode.
  2946 // These stores must update raw memory, and they must be to addresses
  2947 // derived from the raw address produced by AllocateNode, and with
  2948 // a constant offset.  They must be ordered by increasing offset.
  2949 // The first one is at in(RawStores), the last at in(req()-1).
  2950 // Unlike most memory operations, they are not linked in a chain,
  2951 // but are displayed in parallel as users of the rawmem output of
  2952 // the allocation.
  2953 //
  2954 // (See comments in InitializeNode::capture_store, which continue
  2955 // the example given above.)
  2956 //
  2957 // When the associated Allocate is macro-expanded, the InitializeNode
  2958 // may be rewritten to optimize collected stores.  A ClearArrayNode
  2959 // may also be created at that point to represent any required zeroing.
  2960 // The InitializeNode is then marked 'complete', prohibiting further
  2961 // capturing of nearby memory operations.
  2962 //
  2963 // During macro-expansion, all captured initializations which store
  2964 // constant values of 32 bits or smaller are coalesced (if advantageous)
  2965 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
  2966 // initialized in fewer memory operations.  Memory words which are
  2967 // covered by neither tiles nor non-constant stores are pre-zeroed
  2968 // by explicit stores of zero.  (The code shape happens to do all
  2969 // zeroing first, then all other stores, with both sequences occurring
  2970 // in order of ascending offsets.)
  2971 //
  2972 // Alternatively, code may be inserted between an AllocateNode and its
  2973 // InitializeNode, to perform arbitrary initialization of the new object.
  2974 // E.g., the object copying intrinsics insert complex data transfers here.
  2975 // The initialization must then be marked as 'complete' disable the
  2976 // built-in zeroing semantics and the collection of initializing stores.
  2977 //
  2978 // While an InitializeNode is incomplete, reads from the memory state
  2979 // produced by it are optimizable if they match the control edge and
  2980 // new oop address associated with the allocation/initialization.
  2981 // They return a stored value (if the offset matches) or else zero.
  2982 // A write to the memory state, if it matches control and address,
  2983 // and if it is to a constant offset, may be 'captured' by the
  2984 // InitializeNode.  It is cloned as a raw memory operation and rewired
  2985 // inside the initialization, to the raw oop produced by the allocation.
  2986 // Operations on addresses which are provably distinct (e.g., to
  2987 // other AllocateNodes) are allowed to bypass the initialization.
  2988 //
  2989 // The effect of all this is to consolidate object initialization
  2990 // (both arrays and non-arrays, both piecewise and bulk) into a
  2991 // single location, where it can be optimized as a unit.
  2992 //
  2993 // Only stores with an offset less than TrackedInitializationLimit words
  2994 // will be considered for capture by an InitializeNode.  This puts a
  2995 // reasonable limit on the complexity of optimized initializations.
  2997 //---------------------------InitializeNode------------------------------------
  2998 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
  2999   : _is_complete(Incomplete), _does_not_escape(false),
  3000     MemBarNode(C, adr_type, rawoop)
  3002   init_class_id(Class_Initialize);
  3004   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
  3005   assert(in(RawAddress) == rawoop, "proper init");
  3006   // Note:  allocation() can be NULL, for secondary initialization barriers
  3009 // Since this node is not matched, it will be processed by the
  3010 // register allocator.  Declare that there are no constraints
  3011 // on the allocation of the RawAddress edge.
  3012 const RegMask &InitializeNode::in_RegMask(uint idx) const {
  3013   // This edge should be set to top, by the set_complete.  But be conservative.
  3014   if (idx == InitializeNode::RawAddress)
  3015     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
  3016   return RegMask::Empty;
  3019 Node* InitializeNode::memory(uint alias_idx) {
  3020   Node* mem = in(Memory);
  3021   if (mem->is_MergeMem()) {
  3022     return mem->as_MergeMem()->memory_at(alias_idx);
  3023   } else {
  3024     // incoming raw memory is not split
  3025     return mem;
  3029 bool InitializeNode::is_non_zero() {
  3030   if (is_complete())  return false;
  3031   remove_extra_zeroes();
  3032   return (req() > RawStores);
  3035 void InitializeNode::set_complete(PhaseGVN* phase) {
  3036   assert(!is_complete(), "caller responsibility");
  3037   _is_complete = Complete;
  3039   // After this node is complete, it contains a bunch of
  3040   // raw-memory initializations.  There is no need for
  3041   // it to have anything to do with non-raw memory effects.
  3042   // Therefore, tell all non-raw users to re-optimize themselves,
  3043   // after skipping the memory effects of this initialization.
  3044   PhaseIterGVN* igvn = phase->is_IterGVN();
  3045   if (igvn)  igvn->add_users_to_worklist(this);
  3048 // convenience function
  3049 // return false if the init contains any stores already
  3050 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
  3051   InitializeNode* init = initialization();
  3052   if (init == NULL || init->is_complete())  return false;
  3053   init->remove_extra_zeroes();
  3054   // for now, if this allocation has already collected any inits, bail:
  3055   if (init->is_non_zero())  return false;
  3056   init->set_complete(phase);
  3057   return true;
  3060 void InitializeNode::remove_extra_zeroes() {
  3061   if (req() == RawStores)  return;
  3062   Node* zmem = zero_memory();
  3063   uint fill = RawStores;
  3064   for (uint i = fill; i < req(); i++) {
  3065     Node* n = in(i);
  3066     if (n->is_top() || n == zmem)  continue;  // skip
  3067     if (fill < i)  set_req(fill, n);          // compact
  3068     ++fill;
  3070   // delete any empty spaces created:
  3071   while (fill < req()) {
  3072     del_req(fill);
  3076 // Helper for remembering which stores go with which offsets.
  3077 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
  3078   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
  3079   intptr_t offset = -1;
  3080   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
  3081                                                phase, offset);
  3082   if (base == NULL)     return -1;  // something is dead,
  3083   if (offset < 0)       return -1;  //        dead, dead
  3084   return offset;
  3087 // Helper for proving that an initialization expression is
  3088 // "simple enough" to be folded into an object initialization.
  3089 // Attempts to prove that a store's initial value 'n' can be captured
  3090 // within the initialization without creating a vicious cycle, such as:
  3091 //     { Foo p = new Foo(); p.next = p; }
  3092 // True for constants and parameters and small combinations thereof.
  3093 bool InitializeNode::detect_init_independence(Node* n,
  3094                                               bool st_is_pinned,
  3095                                               int& count) {
  3096   if (n == NULL)      return true;   // (can this really happen?)
  3097   if (n->is_Proj())   n = n->in(0);
  3098   if (n == this)      return false;  // found a cycle
  3099   if (n->is_Con())    return true;
  3100   if (n->is_Start())  return true;   // params, etc., are OK
  3101   if (n->is_Root())   return true;   // even better
  3103   Node* ctl = n->in(0);
  3104   if (ctl != NULL && !ctl->is_top()) {
  3105     if (ctl->is_Proj())  ctl = ctl->in(0);
  3106     if (ctl == this)  return false;
  3108     // If we already know that the enclosing memory op is pinned right after
  3109     // the init, then any control flow that the store has picked up
  3110     // must have preceded the init, or else be equal to the init.
  3111     // Even after loop optimizations (which might change control edges)
  3112     // a store is never pinned *before* the availability of its inputs.
  3113     if (!MemNode::all_controls_dominate(n, this))
  3114       return false;                  // failed to prove a good control
  3118   // Check data edges for possible dependencies on 'this'.
  3119   if ((count += 1) > 20)  return false;  // complexity limit
  3120   for (uint i = 1; i < n->req(); i++) {
  3121     Node* m = n->in(i);
  3122     if (m == NULL || m == n || m->is_top())  continue;
  3123     uint first_i = n->find_edge(m);
  3124     if (i != first_i)  continue;  // process duplicate edge just once
  3125     if (!detect_init_independence(m, st_is_pinned, count)) {
  3126       return false;
  3130   return true;
  3133 // Here are all the checks a Store must pass before it can be moved into
  3134 // an initialization.  Returns zero if a check fails.
  3135 // On success, returns the (constant) offset to which the store applies,
  3136 // within the initialized memory.
  3137 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
  3138   const int FAIL = 0;
  3139   if (st->req() != MemNode::ValueIn + 1)
  3140     return FAIL;                // an inscrutable StoreNode (card mark?)
  3141   Node* ctl = st->in(MemNode::Control);
  3142   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
  3143     return FAIL;                // must be unconditional after the initialization
  3144   Node* mem = st->in(MemNode::Memory);
  3145   if (!(mem->is_Proj() && mem->in(0) == this))
  3146     return FAIL;                // must not be preceded by other stores
  3147   Node* adr = st->in(MemNode::Address);
  3148   intptr_t offset;
  3149   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
  3150   if (alloc == NULL)
  3151     return FAIL;                // inscrutable address
  3152   if (alloc != allocation())
  3153     return FAIL;                // wrong allocation!  (store needs to float up)
  3154   Node* val = st->in(MemNode::ValueIn);
  3155   int complexity_count = 0;
  3156   if (!detect_init_independence(val, true, complexity_count))
  3157     return FAIL;                // stored value must be 'simple enough'
  3159   return offset;                // success
  3162 // Find the captured store in(i) which corresponds to the range
  3163 // [start..start+size) in the initialized object.
  3164 // If there is one, return its index i.  If there isn't, return the
  3165 // negative of the index where it should be inserted.
  3166 // Return 0 if the queried range overlaps an initialization boundary
  3167 // or if dead code is encountered.
  3168 // If size_in_bytes is zero, do not bother with overlap checks.
  3169 int InitializeNode::captured_store_insertion_point(intptr_t start,
  3170                                                    int size_in_bytes,
  3171                                                    PhaseTransform* phase) {
  3172   const int FAIL = 0, MAX_STORE = BytesPerLong;
  3174   if (is_complete())
  3175     return FAIL;                // arraycopy got here first; punt
  3177   assert(allocation() != NULL, "must be present");
  3179   // no negatives, no header fields:
  3180   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
  3182   // after a certain size, we bail out on tracking all the stores:
  3183   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
  3184   if (start >= ti_limit)  return FAIL;
  3186   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
  3187     if (i >= limit)  return -(int)i; // not found; here is where to put it
  3189     Node*    st     = in(i);
  3190     intptr_t st_off = get_store_offset(st, phase);
  3191     if (st_off < 0) {
  3192       if (st != zero_memory()) {
  3193         return FAIL;            // bail out if there is dead garbage
  3195     } else if (st_off > start) {
  3196       // ...we are done, since stores are ordered
  3197       if (st_off < start + size_in_bytes) {
  3198         return FAIL;            // the next store overlaps
  3200       return -(int)i;           // not found; here is where to put it
  3201     } else if (st_off < start) {
  3202       if (size_in_bytes != 0 &&
  3203           start < st_off + MAX_STORE &&
  3204           start < st_off + st->as_Store()->memory_size()) {
  3205         return FAIL;            // the previous store overlaps
  3207     } else {
  3208       if (size_in_bytes != 0 &&
  3209           st->as_Store()->memory_size() != size_in_bytes) {
  3210         return FAIL;            // mismatched store size
  3212       return i;
  3215     ++i;
  3219 // Look for a captured store which initializes at the offset 'start'
  3220 // with the given size.  If there is no such store, and no other
  3221 // initialization interferes, then return zero_memory (the memory
  3222 // projection of the AllocateNode).
  3223 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
  3224                                           PhaseTransform* phase) {
  3225   assert(stores_are_sane(phase), "");
  3226   int i = captured_store_insertion_point(start, size_in_bytes, phase);
  3227   if (i == 0) {
  3228     return NULL;                // something is dead
  3229   } else if (i < 0) {
  3230     return zero_memory();       // just primordial zero bits here
  3231   } else {
  3232     Node* st = in(i);           // here is the store at this position
  3233     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
  3234     return st;
  3238 // Create, as a raw pointer, an address within my new object at 'offset'.
  3239 Node* InitializeNode::make_raw_address(intptr_t offset,
  3240                                        PhaseTransform* phase) {
  3241   Node* addr = in(RawAddress);
  3242   if (offset != 0) {
  3243     Compile* C = phase->C;
  3244     addr = phase->transform( new (C) AddPNode(C->top(), addr,
  3245                                                  phase->MakeConX(offset)) );
  3247   return addr;
  3250 // Clone the given store, converting it into a raw store
  3251 // initializing a field or element of my new object.
  3252 // Caller is responsible for retiring the original store,
  3253 // with subsume_node or the like.
  3254 //
  3255 // From the example above InitializeNode::InitializeNode,
  3256 // here are the old stores to be captured:
  3257 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
  3258 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
  3259 //
  3260 // Here is the changed code; note the extra edges on init:
  3261 //   alloc = (Allocate ...)
  3262 //   rawoop = alloc.RawAddress
  3263 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
  3264 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
  3265 //   init = (Initialize alloc.Control alloc.Memory rawoop
  3266 //                      rawstore1 rawstore2)
  3267 //
  3268 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
  3269                                     PhaseTransform* phase) {
  3270   assert(stores_are_sane(phase), "");
  3272   if (start < 0)  return NULL;
  3273   assert(can_capture_store(st, phase) == start, "sanity");
  3275   Compile* C = phase->C;
  3276   int size_in_bytes = st->memory_size();
  3277   int i = captured_store_insertion_point(start, size_in_bytes, phase);
  3278   if (i == 0)  return NULL;     // bail out
  3279   Node* prev_mem = NULL;        // raw memory for the captured store
  3280   if (i > 0) {
  3281     prev_mem = in(i);           // there is a pre-existing store under this one
  3282     set_req(i, C->top());       // temporarily disconnect it
  3283     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
  3284   } else {
  3285     i = -i;                     // no pre-existing store
  3286     prev_mem = zero_memory();   // a slice of the newly allocated object
  3287     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
  3288       set_req(--i, C->top());   // reuse this edge; it has been folded away
  3289     else
  3290       ins_req(i, C->top());     // build a new edge
  3292   Node* new_st = st->clone();
  3293   new_st->set_req(MemNode::Control, in(Control));
  3294   new_st->set_req(MemNode::Memory,  prev_mem);
  3295   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
  3296   new_st = phase->transform(new_st);
  3298   // At this point, new_st might have swallowed a pre-existing store
  3299   // at the same offset, or perhaps new_st might have disappeared,
  3300   // if it redundantly stored the same value (or zero to fresh memory).
  3302   // In any case, wire it in:
  3303   set_req(i, new_st);
  3305   // The caller may now kill the old guy.
  3306   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
  3307   assert(check_st == new_st || check_st == NULL, "must be findable");
  3308   assert(!is_complete(), "");
  3309   return new_st;
  3312 static bool store_constant(jlong* tiles, int num_tiles,
  3313                            intptr_t st_off, int st_size,
  3314                            jlong con) {
  3315   if ((st_off & (st_size-1)) != 0)
  3316     return false;               // strange store offset (assume size==2**N)
  3317   address addr = (address)tiles + st_off;
  3318   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
  3319   switch (st_size) {
  3320   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
  3321   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
  3322   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
  3323   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
  3324   default: return false;        // strange store size (detect size!=2**N here)
  3326   return true;                  // return success to caller
  3329 // Coalesce subword constants into int constants and possibly
  3330 // into long constants.  The goal, if the CPU permits,
  3331 // is to initialize the object with a small number of 64-bit tiles.
  3332 // Also, convert floating-point constants to bit patterns.
  3333 // Non-constants are not relevant to this pass.
  3334 //
  3335 // In terms of the running example on InitializeNode::InitializeNode
  3336 // and InitializeNode::capture_store, here is the transformation
  3337 // of rawstore1 and rawstore2 into rawstore12:
  3338 //   alloc = (Allocate ...)
  3339 //   rawoop = alloc.RawAddress
  3340 //   tile12 = 0x00010002
  3341 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
  3342 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
  3343 //
  3344 void
  3345 InitializeNode::coalesce_subword_stores(intptr_t header_size,
  3346                                         Node* size_in_bytes,
  3347                                         PhaseGVN* phase) {
  3348   Compile* C = phase->C;
  3350   assert(stores_are_sane(phase), "");
  3351   // Note:  After this pass, they are not completely sane,
  3352   // since there may be some overlaps.
  3354   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
  3356   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
  3357   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
  3358   size_limit = MIN2(size_limit, ti_limit);
  3359   size_limit = align_size_up(size_limit, BytesPerLong);
  3360   int num_tiles = size_limit / BytesPerLong;
  3362   // allocate space for the tile map:
  3363   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
  3364   jlong  tiles_buf[small_len];
  3365   Node*  nodes_buf[small_len];
  3366   jlong  inits_buf[small_len];
  3367   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
  3368                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
  3369   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
  3370                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
  3371   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
  3372                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
  3373   // tiles: exact bitwise model of all primitive constants
  3374   // nodes: last constant-storing node subsumed into the tiles model
  3375   // inits: which bytes (in each tile) are touched by any initializations
  3377   //// Pass A: Fill in the tile model with any relevant stores.
  3379   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
  3380   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
  3381   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
  3382   Node* zmem = zero_memory(); // initially zero memory state
  3383   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
  3384     Node* st = in(i);
  3385     intptr_t st_off = get_store_offset(st, phase);
  3387     // Figure out the store's offset and constant value:
  3388     if (st_off < header_size)             continue; //skip (ignore header)
  3389     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
  3390     int st_size = st->as_Store()->memory_size();
  3391     if (st_off + st_size > size_limit)    break;
  3393     // Record which bytes are touched, whether by constant or not.
  3394     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
  3395       continue;                 // skip (strange store size)
  3397     const Type* val = phase->type(st->in(MemNode::ValueIn));
  3398     if (!val->singleton())                continue; //skip (non-con store)
  3399     BasicType type = val->basic_type();
  3401     jlong con = 0;
  3402     switch (type) {
  3403     case T_INT:    con = val->is_int()->get_con();  break;
  3404     case T_LONG:   con = val->is_long()->get_con(); break;
  3405     case T_FLOAT:  con = jint_cast(val->getf());    break;
  3406     case T_DOUBLE: con = jlong_cast(val->getd());   break;
  3407     default:                              continue; //skip (odd store type)
  3410     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
  3411         st->Opcode() == Op_StoreL) {
  3412       continue;                 // This StoreL is already optimal.
  3415     // Store down the constant.
  3416     store_constant(tiles, num_tiles, st_off, st_size, con);
  3418     intptr_t j = st_off >> LogBytesPerLong;
  3420     if (type == T_INT && st_size == BytesPerInt
  3421         && (st_off & BytesPerInt) == BytesPerInt) {
  3422       jlong lcon = tiles[j];
  3423       if (!Matcher::isSimpleConstant64(lcon) &&
  3424           st->Opcode() == Op_StoreI) {
  3425         // This StoreI is already optimal by itself.
  3426         jint* intcon = (jint*) &tiles[j];
  3427         intcon[1] = 0;  // undo the store_constant()
  3429         // If the previous store is also optimal by itself, back up and
  3430         // undo the action of the previous loop iteration... if we can.
  3431         // But if we can't, just let the previous half take care of itself.
  3432         st = nodes[j];
  3433         st_off -= BytesPerInt;
  3434         con = intcon[0];
  3435         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
  3436           assert(st_off >= header_size, "still ignoring header");
  3437           assert(get_store_offset(st, phase) == st_off, "must be");
  3438           assert(in(i-1) == zmem, "must be");
  3439           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
  3440           assert(con == tcon->is_int()->get_con(), "must be");
  3441           // Undo the effects of the previous loop trip, which swallowed st:
  3442           intcon[0] = 0;        // undo store_constant()
  3443           set_req(i-1, st);     // undo set_req(i, zmem)
  3444           nodes[j] = NULL;      // undo nodes[j] = st
  3445           --old_subword;        // undo ++old_subword
  3447         continue;               // This StoreI is already optimal.
  3451     // This store is not needed.
  3452     set_req(i, zmem);
  3453     nodes[j] = st;              // record for the moment
  3454     if (st_size < BytesPerLong) // something has changed
  3455           ++old_subword;        // includes int/float, but who's counting...
  3456     else  ++old_long;
  3459   if ((old_subword + old_long) == 0)
  3460     return;                     // nothing more to do
  3462   //// Pass B: Convert any non-zero tiles into optimal constant stores.
  3463   // Be sure to insert them before overlapping non-constant stores.
  3464   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
  3465   for (int j = 0; j < num_tiles; j++) {
  3466     jlong con  = tiles[j];
  3467     jlong init = inits[j];
  3468     if (con == 0)  continue;
  3469     jint con0,  con1;           // split the constant, address-wise
  3470     jint init0, init1;          // split the init map, address-wise
  3471     { union { jlong con; jint intcon[2]; } u;
  3472       u.con = con;
  3473       con0  = u.intcon[0];
  3474       con1  = u.intcon[1];
  3475       u.con = init;
  3476       init0 = u.intcon[0];
  3477       init1 = u.intcon[1];
  3480     Node* old = nodes[j];
  3481     assert(old != NULL, "need the prior store");
  3482     intptr_t offset = (j * BytesPerLong);
  3484     bool split = !Matcher::isSimpleConstant64(con);
  3486     if (offset < header_size) {
  3487       assert(offset + BytesPerInt >= header_size, "second int counts");
  3488       assert(*(jint*)&tiles[j] == 0, "junk in header");
  3489       split = true;             // only the second word counts
  3490       // Example:  int a[] = { 42 ... }
  3491     } else if (con0 == 0 && init0 == -1) {
  3492       split = true;             // first word is covered by full inits
  3493       // Example:  int a[] = { ... foo(), 42 ... }
  3494     } else if (con1 == 0 && init1 == -1) {
  3495       split = true;             // second word is covered by full inits
  3496       // Example:  int a[] = { ... 42, foo() ... }
  3499     // Here's a case where init0 is neither 0 nor -1:
  3500     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
  3501     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
  3502     // In this case the tile is not split; it is (jlong)42.
  3503     // The big tile is stored down, and then the foo() value is inserted.
  3504     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
  3506     Node* ctl = old->in(MemNode::Control);
  3507     Node* adr = make_raw_address(offset, phase);
  3508     const TypePtr* atp = TypeRawPtr::BOTTOM;
  3510     // One or two coalesced stores to plop down.
  3511     Node*    st[2];
  3512     intptr_t off[2];
  3513     int  nst = 0;
  3514     if (!split) {
  3515       ++new_long;
  3516       off[nst] = offset;
  3517       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
  3518                                   phase->longcon(con), T_LONG);
  3519     } else {
  3520       // Omit either if it is a zero.
  3521       if (con0 != 0) {
  3522         ++new_int;
  3523         off[nst]  = offset;
  3524         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
  3525                                     phase->intcon(con0), T_INT);
  3527       if (con1 != 0) {
  3528         ++new_int;
  3529         offset += BytesPerInt;
  3530         adr = make_raw_address(offset, phase);
  3531         off[nst]  = offset;
  3532         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
  3533                                     phase->intcon(con1), T_INT);
  3537     // Insert second store first, then the first before the second.
  3538     // Insert each one just before any overlapping non-constant stores.
  3539     while (nst > 0) {
  3540       Node* st1 = st[--nst];
  3541       C->copy_node_notes_to(st1, old);
  3542       st1 = phase->transform(st1);
  3543       offset = off[nst];
  3544       assert(offset >= header_size, "do not smash header");
  3545       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
  3546       guarantee(ins_idx != 0, "must re-insert constant store");
  3547       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
  3548       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
  3549         set_req(--ins_idx, st1);
  3550       else
  3551         ins_req(ins_idx, st1);
  3555   if (PrintCompilation && WizardMode)
  3556     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
  3557                   old_subword, old_long, new_int, new_long);
  3558   if (C->log() != NULL)
  3559     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
  3560                    old_subword, old_long, new_int, new_long);
  3562   // Clean up any remaining occurrences of zmem:
  3563   remove_extra_zeroes();
  3566 // Explore forward from in(start) to find the first fully initialized
  3567 // word, and return its offset.  Skip groups of subword stores which
  3568 // together initialize full words.  If in(start) is itself part of a
  3569 // fully initialized word, return the offset of in(start).  If there
  3570 // are no following full-word stores, or if something is fishy, return
  3571 // a negative value.
  3572 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
  3573   int       int_map = 0;
  3574   intptr_t  int_map_off = 0;
  3575   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
  3577   for (uint i = start, limit = req(); i < limit; i++) {
  3578     Node* st = in(i);
  3580     intptr_t st_off = get_store_offset(st, phase);
  3581     if (st_off < 0)  break;  // return conservative answer
  3583     int st_size = st->as_Store()->memory_size();
  3584     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
  3585       return st_off;            // we found a complete word init
  3588     // update the map:
  3590     intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
  3591     if (this_int_off != int_map_off) {
  3592       // reset the map:
  3593       int_map = 0;
  3594       int_map_off = this_int_off;
  3597     int subword_off = st_off - this_int_off;
  3598     int_map |= right_n_bits(st_size) << subword_off;
  3599     if ((int_map & FULL_MAP) == FULL_MAP) {
  3600       return this_int_off;      // we found a complete word init
  3603     // Did this store hit or cross the word boundary?
  3604     intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
  3605     if (next_int_off == this_int_off + BytesPerInt) {
  3606       // We passed the current int, without fully initializing it.
  3607       int_map_off = next_int_off;
  3608       int_map >>= BytesPerInt;
  3609     } else if (next_int_off > this_int_off + BytesPerInt) {
  3610       // We passed the current and next int.
  3611       return this_int_off + BytesPerInt;
  3615   return -1;
  3619 // Called when the associated AllocateNode is expanded into CFG.
  3620 // At this point, we may perform additional optimizations.
  3621 // Linearize the stores by ascending offset, to make memory
  3622 // activity as coherent as possible.
  3623 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
  3624                                       intptr_t header_size,
  3625                                       Node* size_in_bytes,
  3626                                       PhaseGVN* phase) {
  3627   assert(!is_complete(), "not already complete");
  3628   assert(stores_are_sane(phase), "");
  3629   assert(allocation() != NULL, "must be present");
  3631   remove_extra_zeroes();
  3633   if (ReduceFieldZeroing || ReduceBulkZeroing)
  3634     // reduce instruction count for common initialization patterns
  3635     coalesce_subword_stores(header_size, size_in_bytes, phase);
  3637   Node* zmem = zero_memory();   // initially zero memory state
  3638   Node* inits = zmem;           // accumulating a linearized chain of inits
  3639   #ifdef ASSERT
  3640   intptr_t first_offset = allocation()->minimum_header_size();
  3641   intptr_t last_init_off = first_offset;  // previous init offset
  3642   intptr_t last_init_end = first_offset;  // previous init offset+size
  3643   intptr_t last_tile_end = first_offset;  // previous tile offset+size
  3644   #endif
  3645   intptr_t zeroes_done = header_size;
  3647   bool do_zeroing = true;       // we might give up if inits are very sparse
  3648   int  big_init_gaps = 0;       // how many large gaps have we seen?
  3650   if (ZeroTLAB)  do_zeroing = false;
  3651   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
  3653   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
  3654     Node* st = in(i);
  3655     intptr_t st_off = get_store_offset(st, phase);
  3656     if (st_off < 0)
  3657       break;                    // unknown junk in the inits
  3658     if (st->in(MemNode::Memory) != zmem)
  3659       break;                    // complicated store chains somehow in list
  3661     int st_size = st->as_Store()->memory_size();
  3662     intptr_t next_init_off = st_off + st_size;
  3664     if (do_zeroing && zeroes_done < next_init_off) {
  3665       // See if this store needs a zero before it or under it.
  3666       intptr_t zeroes_needed = st_off;
  3668       if (st_size < BytesPerInt) {
  3669         // Look for subword stores which only partially initialize words.
  3670         // If we find some, we must lay down some word-level zeroes first,
  3671         // underneath the subword stores.
  3672         //
  3673         // Examples:
  3674         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
  3675         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
  3676         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
  3677         //
  3678         // Note:  coalesce_subword_stores may have already done this,
  3679         // if it was prompted by constant non-zero subword initializers.
  3680         // But this case can still arise with non-constant stores.
  3682         intptr_t next_full_store = find_next_fullword_store(i, phase);
  3684         // In the examples above:
  3685         //   in(i)          p   q   r   s     x   y     z
  3686         //   st_off        12  13  14  15    12  13    14
  3687         //   st_size        1   1   1   1     1   1     1
  3688         //   next_full_s.  12  16  16  16    16  16    16
  3689         //   z's_done      12  16  16  16    12  16    12
  3690         //   z's_needed    12  16  16  16    16  16    16
  3691         //   zsize          0   0   0   0     4   0     4
  3692         if (next_full_store < 0) {
  3693           // Conservative tack:  Zero to end of current word.
  3694           zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
  3695         } else {
  3696           // Zero to beginning of next fully initialized word.
  3697           // Or, don't zero at all, if we are already in that word.
  3698           assert(next_full_store >= zeroes_needed, "must go forward");
  3699           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
  3700           zeroes_needed = next_full_store;
  3704       if (zeroes_needed > zeroes_done) {
  3705         intptr_t zsize = zeroes_needed - zeroes_done;
  3706         // Do some incremental zeroing on rawmem, in parallel with inits.
  3707         zeroes_done = align_size_down(zeroes_done, BytesPerInt);
  3708         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
  3709                                               zeroes_done, zeroes_needed,
  3710                                               phase);
  3711         zeroes_done = zeroes_needed;
  3712         if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
  3713           do_zeroing = false;   // leave the hole, next time
  3717     // Collect the store and move on:
  3718     st->set_req(MemNode::Memory, inits);
  3719     inits = st;                 // put it on the linearized chain
  3720     set_req(i, zmem);           // unhook from previous position
  3722     if (zeroes_done == st_off)
  3723       zeroes_done = next_init_off;
  3725     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
  3727     #ifdef ASSERT
  3728     // Various order invariants.  Weaker than stores_are_sane because
  3729     // a large constant tile can be filled in by smaller non-constant stores.
  3730     assert(st_off >= last_init_off, "inits do not reverse");
  3731     last_init_off = st_off;
  3732     const Type* val = NULL;
  3733     if (st_size >= BytesPerInt &&
  3734         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
  3735         (int)val->basic_type() < (int)T_OBJECT) {
  3736       assert(st_off >= last_tile_end, "tiles do not overlap");
  3737       assert(st_off >= last_init_end, "tiles do not overwrite inits");
  3738       last_tile_end = MAX2(last_tile_end, next_init_off);
  3739     } else {
  3740       intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
  3741       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
  3742       assert(st_off      >= last_init_end, "inits do not overlap");
  3743       last_init_end = next_init_off;  // it's a non-tile
  3745     #endif //ASSERT
  3748   remove_extra_zeroes();        // clear out all the zmems left over
  3749   add_req(inits);
  3751   if (!ZeroTLAB) {
  3752     // If anything remains to be zeroed, zero it all now.
  3753     zeroes_done = align_size_down(zeroes_done, BytesPerInt);
  3754     // if it is the last unused 4 bytes of an instance, forget about it
  3755     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
  3756     if (zeroes_done + BytesPerLong >= size_limit) {
  3757       assert(allocation() != NULL, "");
  3758       if (allocation()->Opcode() == Op_Allocate) {
  3759         Node* klass_node = allocation()->in(AllocateNode::KlassNode);
  3760         ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
  3761         if (zeroes_done == k->layout_helper())
  3762           zeroes_done = size_limit;
  3765     if (zeroes_done < size_limit) {
  3766       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
  3767                                             zeroes_done, size_in_bytes, phase);
  3771   set_complete(phase);
  3772   return rawmem;
  3776 #ifdef ASSERT
  3777 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
  3778   if (is_complete())
  3779     return true;                // stores could be anything at this point
  3780   assert(allocation() != NULL, "must be present");
  3781   intptr_t last_off = allocation()->minimum_header_size();
  3782   for (uint i = InitializeNode::RawStores; i < req(); i++) {
  3783     Node* st = in(i);
  3784     intptr_t st_off = get_store_offset(st, phase);
  3785     if (st_off < 0)  continue;  // ignore dead garbage
  3786     if (last_off > st_off) {
  3787       tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
  3788       this->dump(2);
  3789       assert(false, "ascending store offsets");
  3790       return false;
  3792     last_off = st_off + st->as_Store()->memory_size();
  3794   return true;
  3796 #endif //ASSERT
  3801 //============================MergeMemNode=====================================
  3802 //
  3803 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
  3804 // contributing store or call operations.  Each contributor provides the memory
  3805 // state for a particular "alias type" (see Compile::alias_type).  For example,
  3806 // if a MergeMem has an input X for alias category #6, then any memory reference
  3807 // to alias category #6 may use X as its memory state input, as an exact equivalent
  3808 // to using the MergeMem as a whole.
  3809 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
  3810 //
  3811 // (Here, the <N> notation gives the index of the relevant adr_type.)
  3812 //
  3813 // In one special case (and more cases in the future), alias categories overlap.
  3814 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
  3815 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
  3816 // it is exactly equivalent to that state W:
  3817 //   MergeMem(<Bot>: W) <==> W
  3818 //
  3819 // Usually, the merge has more than one input.  In that case, where inputs
  3820 // overlap (i.e., one is Bot), the narrower alias type determines the memory
  3821 // state for that type, and the wider alias type (Bot) fills in everywhere else:
  3822 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
  3823 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
  3824 //
  3825 // A merge can take a "wide" memory state as one of its narrow inputs.
  3826 // This simply means that the merge observes out only the relevant parts of
  3827 // the wide input.  That is, wide memory states arriving at narrow merge inputs
  3828 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
  3829 //
  3830 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
  3831 // and that memory slices "leak through":
  3832 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
  3833 //
  3834 // But, in such a cascade, repeated memory slices can "block the leak":
  3835 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
  3836 //
  3837 // In the last example, Y is not part of the combined memory state of the
  3838 // outermost MergeMem.  The system must, of course, prevent unschedulable
  3839 // memory states from arising, so you can be sure that the state Y is somehow
  3840 // a precursor to state Y'.
  3841 //
  3842 //
  3843 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
  3844 // of each MergeMemNode array are exactly the numerical alias indexes, including
  3845 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
  3846 // Compile::alias_type (and kin) produce and manage these indexes.
  3847 //
  3848 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
  3849 // (Note that this provides quick access to the top node inside MergeMem methods,
  3850 // without the need to reach out via TLS to Compile::current.)
  3851 //
  3852 // As a consequence of what was just described, a MergeMem that represents a full
  3853 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
  3854 // containing all alias categories.
  3855 //
  3856 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
  3857 //
  3858 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
  3859 // a memory state for the alias type <N>, or else the top node, meaning that
  3860 // there is no particular input for that alias type.  Note that the length of
  3861 // a MergeMem is variable, and may be extended at any time to accommodate new
  3862 // memory states at larger alias indexes.  When merges grow, they are of course
  3863 // filled with "top" in the unused in() positions.
  3864 //
  3865 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
  3866 // (Top was chosen because it works smoothly with passes like GCM.)
  3867 //
  3868 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
  3869 // the type of random VM bits like TLS references.)  Since it is always the
  3870 // first non-Bot memory slice, some low-level loops use it to initialize an
  3871 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
  3872 //
  3873 //
  3874 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
  3875 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
  3876 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
  3877 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
  3878 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
  3879 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
  3880 //
  3881 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
  3882 // really that different from the other memory inputs.  An abbreviation called
  3883 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
  3884 //
  3885 //
  3886 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
  3887 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
  3888 // that "emerges though" the base memory will be marked as excluding the alias types
  3889 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
  3890 //
  3891 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
  3892 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
  3893 //
  3894 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
  3895 // (It is currently unimplemented.)  As you can see, the resulting merge is
  3896 // actually a disjoint union of memory states, rather than an overlay.
  3897 //
  3899 //------------------------------MergeMemNode-----------------------------------
  3900 Node* MergeMemNode::make_empty_memory() {
  3901   Node* empty_memory = (Node*) Compile::current()->top();
  3902   assert(empty_memory->is_top(), "correct sentinel identity");
  3903   return empty_memory;
  3906 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
  3907   init_class_id(Class_MergeMem);
  3908   // all inputs are nullified in Node::Node(int)
  3909   // set_input(0, NULL);  // no control input
  3911   // Initialize the edges uniformly to top, for starters.
  3912   Node* empty_mem = make_empty_memory();
  3913   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
  3914     init_req(i,empty_mem);
  3916   assert(empty_memory() == empty_mem, "");
  3918   if( new_base != NULL && new_base->is_MergeMem() ) {
  3919     MergeMemNode* mdef = new_base->as_MergeMem();
  3920     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
  3921     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
  3922       mms.set_memory(mms.memory2());
  3924     assert(base_memory() == mdef->base_memory(), "");
  3925   } else {
  3926     set_base_memory(new_base);
  3930 // Make a new, untransformed MergeMem with the same base as 'mem'.
  3931 // If mem is itself a MergeMem, populate the result with the same edges.
  3932 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
  3933   return new(C) MergeMemNode(mem);
  3936 //------------------------------cmp--------------------------------------------
  3937 uint MergeMemNode::hash() const { return NO_HASH; }
  3938 uint MergeMemNode::cmp( const Node &n ) const {
  3939   return (&n == this);          // Always fail except on self
  3942 //------------------------------Identity---------------------------------------
  3943 Node* MergeMemNode::Identity(PhaseTransform *phase) {
  3944   // Identity if this merge point does not record any interesting memory
  3945   // disambiguations.
  3946   Node* base_mem = base_memory();
  3947   Node* empty_mem = empty_memory();
  3948   if (base_mem != empty_mem) {  // Memory path is not dead?
  3949     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  3950       Node* mem = in(i);
  3951       if (mem != empty_mem && mem != base_mem) {
  3952         return this;            // Many memory splits; no change
  3956   return base_mem;              // No memory splits; ID on the one true input
  3959 //------------------------------Ideal------------------------------------------
  3960 // This method is invoked recursively on chains of MergeMem nodes
  3961 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  3962   // Remove chain'd MergeMems
  3963   //
  3964   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
  3965   // relative to the "in(Bot)".  Since we are patching both at the same time,
  3966   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
  3967   // but rewrite each "in(i)" relative to the new "in(Bot)".
  3968   Node *progress = NULL;
  3971   Node* old_base = base_memory();
  3972   Node* empty_mem = empty_memory();
  3973   if (old_base == empty_mem)
  3974     return NULL; // Dead memory path.
  3976   MergeMemNode* old_mbase;
  3977   if (old_base != NULL && old_base->is_MergeMem())
  3978     old_mbase = old_base->as_MergeMem();
  3979   else
  3980     old_mbase = NULL;
  3981   Node* new_base = old_base;
  3983   // simplify stacked MergeMems in base memory
  3984   if (old_mbase)  new_base = old_mbase->base_memory();
  3986   // the base memory might contribute new slices beyond my req()
  3987   if (old_mbase)  grow_to_match(old_mbase);
  3989   // Look carefully at the base node if it is a phi.
  3990   PhiNode* phi_base;
  3991   if (new_base != NULL && new_base->is_Phi())
  3992     phi_base = new_base->as_Phi();
  3993   else
  3994     phi_base = NULL;
  3996   Node*    phi_reg = NULL;
  3997   uint     phi_len = (uint)-1;
  3998   if (phi_base != NULL && !phi_base->is_copy()) {
  3999     // do not examine phi if degraded to a copy
  4000     phi_reg = phi_base->region();
  4001     phi_len = phi_base->req();
  4002     // see if the phi is unfinished
  4003     for (uint i = 1; i < phi_len; i++) {
  4004       if (phi_base->in(i) == NULL) {
  4005         // incomplete phi; do not look at it yet!
  4006         phi_reg = NULL;
  4007         phi_len = (uint)-1;
  4008         break;
  4013   // Note:  We do not call verify_sparse on entry, because inputs
  4014   // can normalize to the base_memory via subsume_node or similar
  4015   // mechanisms.  This method repairs that damage.
  4017   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
  4019   // Look at each slice.
  4020   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  4021     Node* old_in = in(i);
  4022     // calculate the old memory value
  4023     Node* old_mem = old_in;
  4024     if (old_mem == empty_mem)  old_mem = old_base;
  4025     assert(old_mem == memory_at(i), "");
  4027     // maybe update (reslice) the old memory value
  4029     // simplify stacked MergeMems
  4030     Node* new_mem = old_mem;
  4031     MergeMemNode* old_mmem;
  4032     if (old_mem != NULL && old_mem->is_MergeMem())
  4033       old_mmem = old_mem->as_MergeMem();
  4034     else
  4035       old_mmem = NULL;
  4036     if (old_mmem == this) {
  4037       // This can happen if loops break up and safepoints disappear.
  4038       // A merge of BotPtr (default) with a RawPtr memory derived from a
  4039       // safepoint can be rewritten to a merge of the same BotPtr with
  4040       // the BotPtr phi coming into the loop.  If that phi disappears
  4041       // also, we can end up with a self-loop of the mergemem.
  4042       // In general, if loops degenerate and memory effects disappear,
  4043       // a mergemem can be left looking at itself.  This simply means
  4044       // that the mergemem's default should be used, since there is
  4045       // no longer any apparent effect on this slice.
  4046       // Note: If a memory slice is a MergeMem cycle, it is unreachable
  4047       //       from start.  Update the input to TOP.
  4048       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
  4050     else if (old_mmem != NULL) {
  4051       new_mem = old_mmem->memory_at(i);
  4053     // else preceding memory was not a MergeMem
  4055     // replace equivalent phis (unfortunately, they do not GVN together)
  4056     if (new_mem != NULL && new_mem != new_base &&
  4057         new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
  4058       if (new_mem->is_Phi()) {
  4059         PhiNode* phi_mem = new_mem->as_Phi();
  4060         for (uint i = 1; i < phi_len; i++) {
  4061           if (phi_base->in(i) != phi_mem->in(i)) {
  4062             phi_mem = NULL;
  4063             break;
  4066         if (phi_mem != NULL) {
  4067           // equivalent phi nodes; revert to the def
  4068           new_mem = new_base;
  4073     // maybe store down a new value
  4074     Node* new_in = new_mem;
  4075     if (new_in == new_base)  new_in = empty_mem;
  4077     if (new_in != old_in) {
  4078       // Warning:  Do not combine this "if" with the previous "if"
  4079       // A memory slice might have be be rewritten even if it is semantically
  4080       // unchanged, if the base_memory value has changed.
  4081       set_req(i, new_in);
  4082       progress = this;          // Report progress
  4086   if (new_base != old_base) {
  4087     set_req(Compile::AliasIdxBot, new_base);
  4088     // Don't use set_base_memory(new_base), because we need to update du.
  4089     assert(base_memory() == new_base, "");
  4090     progress = this;
  4093   if( base_memory() == this ) {
  4094     // a self cycle indicates this memory path is dead
  4095     set_req(Compile::AliasIdxBot, empty_mem);
  4098   // Resolve external cycles by calling Ideal on a MergeMem base_memory
  4099   // Recursion must occur after the self cycle check above
  4100   if( base_memory()->is_MergeMem() ) {
  4101     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
  4102     Node *m = phase->transform(new_mbase);  // Rollup any cycles
  4103     if( m != NULL && (m->is_top() ||
  4104         m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
  4105       // propagate rollup of dead cycle to self
  4106       set_req(Compile::AliasIdxBot, empty_mem);
  4110   if( base_memory() == empty_mem ) {
  4111     progress = this;
  4112     // Cut inputs during Parse phase only.
  4113     // During Optimize phase a dead MergeMem node will be subsumed by Top.
  4114     if( !can_reshape ) {
  4115       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  4116         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
  4121   if( !progress && base_memory()->is_Phi() && can_reshape ) {
  4122     // Check if PhiNode::Ideal's "Split phis through memory merges"
  4123     // transform should be attempted. Look for this->phi->this cycle.
  4124     uint merge_width = req();
  4125     if (merge_width > Compile::AliasIdxRaw) {
  4126       PhiNode* phi = base_memory()->as_Phi();
  4127       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
  4128         if (phi->in(i) == this) {
  4129           phase->is_IterGVN()->_worklist.push(phi);
  4130           break;
  4136   assert(progress || verify_sparse(), "please, no dups of base");
  4137   return progress;
  4140 //-------------------------set_base_memory-------------------------------------
  4141 void MergeMemNode::set_base_memory(Node *new_base) {
  4142   Node* empty_mem = empty_memory();
  4143   set_req(Compile::AliasIdxBot, new_base);
  4144   assert(memory_at(req()) == new_base, "must set default memory");
  4145   // Clear out other occurrences of new_base:
  4146   if (new_base != empty_mem) {
  4147     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  4148       if (in(i) == new_base)  set_req(i, empty_mem);
  4153 //------------------------------out_RegMask------------------------------------
  4154 const RegMask &MergeMemNode::out_RegMask() const {
  4155   return RegMask::Empty;
  4158 //------------------------------dump_spec--------------------------------------
  4159 #ifndef PRODUCT
  4160 void MergeMemNode::dump_spec(outputStream *st) const {
  4161   st->print(" {");
  4162   Node* base_mem = base_memory();
  4163   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
  4164     Node* mem = memory_at(i);
  4165     if (mem == base_mem) { st->print(" -"); continue; }
  4166     st->print( " N%d:", mem->_idx );
  4167     Compile::current()->get_adr_type(i)->dump_on(st);
  4169   st->print(" }");
  4171 #endif // !PRODUCT
  4174 #ifdef ASSERT
  4175 static bool might_be_same(Node* a, Node* b) {
  4176   if (a == b)  return true;
  4177   if (!(a->is_Phi() || b->is_Phi()))  return false;
  4178   // phis shift around during optimization
  4179   return true;  // pretty stupid...
  4182 // verify a narrow slice (either incoming or outgoing)
  4183 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
  4184   if (!VerifyAliases)       return;  // don't bother to verify unless requested
  4185   if (is_error_reported())  return;  // muzzle asserts when debugging an error
  4186   if (Node::in_dump())      return;  // muzzle asserts when printing
  4187   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
  4188   assert(n != NULL, "");
  4189   // Elide intervening MergeMem's
  4190   while (n->is_MergeMem()) {
  4191     n = n->as_MergeMem()->memory_at(alias_idx);
  4193   Compile* C = Compile::current();
  4194   const TypePtr* n_adr_type = n->adr_type();
  4195   if (n == m->empty_memory()) {
  4196     // Implicit copy of base_memory()
  4197   } else if (n_adr_type != TypePtr::BOTTOM) {
  4198     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
  4199     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
  4200   } else {
  4201     // A few places like make_runtime_call "know" that VM calls are narrow,
  4202     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
  4203     bool expected_wide_mem = false;
  4204     if (n == m->base_memory()) {
  4205       expected_wide_mem = true;
  4206     } else if (alias_idx == Compile::AliasIdxRaw ||
  4207                n == m->memory_at(Compile::AliasIdxRaw)) {
  4208       expected_wide_mem = true;
  4209     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
  4210       // memory can "leak through" calls on channels that
  4211       // are write-once.  Allow this also.
  4212       expected_wide_mem = true;
  4214     assert(expected_wide_mem, "expected narrow slice replacement");
  4217 #else // !ASSERT
  4218 #define verify_memory_slice(m,i,n) (0)  // PRODUCT version is no-op
  4219 #endif
  4222 //-----------------------------memory_at---------------------------------------
  4223 Node* MergeMemNode::memory_at(uint alias_idx) const {
  4224   assert(alias_idx >= Compile::AliasIdxRaw ||
  4225          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
  4226          "must avoid base_memory and AliasIdxTop");
  4228   // Otherwise, it is a narrow slice.
  4229   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
  4230   Compile *C = Compile::current();
  4231   if (is_empty_memory(n)) {
  4232     // the array is sparse; empty slots are the "top" node
  4233     n = base_memory();
  4234     assert(Node::in_dump()
  4235            || n == NULL || n->bottom_type() == Type::TOP
  4236            || n->adr_type() == NULL // address is TOP
  4237            || n->adr_type() == TypePtr::BOTTOM
  4238            || n->adr_type() == TypeRawPtr::BOTTOM
  4239            || Compile::current()->AliasLevel() == 0,
  4240            "must be a wide memory");
  4241     // AliasLevel == 0 if we are organizing the memory states manually.
  4242     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
  4243   } else {
  4244     // make sure the stored slice is sane
  4245     #ifdef ASSERT
  4246     if (is_error_reported() || Node::in_dump()) {
  4247     } else if (might_be_same(n, base_memory())) {
  4248       // Give it a pass:  It is a mostly harmless repetition of the base.
  4249       // This can arise normally from node subsumption during optimization.
  4250     } else {
  4251       verify_memory_slice(this, alias_idx, n);
  4253     #endif
  4255   return n;
  4258 //---------------------------set_memory_at-------------------------------------
  4259 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
  4260   verify_memory_slice(this, alias_idx, n);
  4261   Node* empty_mem = empty_memory();
  4262   if (n == base_memory())  n = empty_mem;  // collapse default
  4263   uint need_req = alias_idx+1;
  4264   if (req() < need_req) {
  4265     if (n == empty_mem)  return;  // already the default, so do not grow me
  4266     // grow the sparse array
  4267     do {
  4268       add_req(empty_mem);
  4269     } while (req() < need_req);
  4271   set_req( alias_idx, n );
  4276 //--------------------------iteration_setup------------------------------------
  4277 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
  4278   if (other != NULL) {
  4279     grow_to_match(other);
  4280     // invariant:  the finite support of mm2 is within mm->req()
  4281     #ifdef ASSERT
  4282     for (uint i = req(); i < other->req(); i++) {
  4283       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
  4285     #endif
  4287   // Replace spurious copies of base_memory by top.
  4288   Node* base_mem = base_memory();
  4289   if (base_mem != NULL && !base_mem->is_top()) {
  4290     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
  4291       if (in(i) == base_mem)
  4292         set_req(i, empty_memory());
  4297 //---------------------------grow_to_match-------------------------------------
  4298 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
  4299   Node* empty_mem = empty_memory();
  4300   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
  4301   // look for the finite support of the other memory
  4302   for (uint i = other->req(); --i >= req(); ) {
  4303     if (other->in(i) != empty_mem) {
  4304       uint new_len = i+1;
  4305       while (req() < new_len)  add_req(empty_mem);
  4306       break;
  4311 //---------------------------verify_sparse-------------------------------------
  4312 #ifndef PRODUCT
  4313 bool MergeMemNode::verify_sparse() const {
  4314   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
  4315   Node* base_mem = base_memory();
  4316   // The following can happen in degenerate cases, since empty==top.
  4317   if (is_empty_memory(base_mem))  return true;
  4318   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
  4319     assert(in(i) != NULL, "sane slice");
  4320     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
  4322   return true;
  4325 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
  4326   Node* n;
  4327   n = mm->in(idx);
  4328   if (mem == n)  return true;  // might be empty_memory()
  4329   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
  4330   if (mem == n)  return true;
  4331   while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
  4332     if (mem == n)  return true;
  4333     if (n == NULL)  break;
  4335   return false;
  4337 #endif // !PRODUCT

mercurial